What GMO, Acadian, Arrowstreet, and Selected Quantitative Managers Are Showing About AI: Strategy, Talent, and Public Signals

September 21 delivered-talk archive update: finance systems behind the roster

The conference follow-up now includes the organizer’s full AI Engineer talk atlas and organization directory, not only channel titles or the New York schedule. The talk catalog contains 1,145 talks, 1,125 speakers, 630 reported organizations, and seven events; the separately fetched organization directory currently contains 629 records. Broad discovery initially surfaced 115 records through finance topics, employer and investor classifications, text matches, and four audited workflow tags. Those routes were useful for recall but not sufficient for inclusion: a financial employer does not make a generic coding talk a finance-system disclosure, and an investor affiliation does not make a product or hiring talk an investment workflow.

A full-organization-catalog audit and 1,030-talk unselected-corpus audit added five talk-scoped records. Best Buy maps merchant feeds, agent protocols, delegated payment mandates, and evaluation categories, but does not disclose a customer-facing autonomous checkout. Temporal publishes an executable, LLM-free durable-MCP invoice demonstration with retries and human gates; its “paid” state is not evidence of funds movement. New Generation connects product normalization, generated interfaces, checkout, wallets, and delegated cards without naming customer denominators. Kiduna documents a proposed legal-identity, permission, privacy, and decision-market design, not a measured production network. RISA Labs describes payer-data collection, deterministic eligibility gates, LLM extraction, corroboration, clinician escalation, and production-hour recovery, while withholding model revisions, thresholds, and evaluation denominators. The five-person technical and lineage audit keeps each system, speaker, employer, code artifact, and academic edge separate.

The subsequent 115-talk precision audit reviewed every prior record against its delivered transcript. It retained 69 and removed 46. Thirty-one removals were affiliation-only or generic systems; nine were text/topic collisions or passing examples; four were personal or historical records without material finance mechanics; and two were generic vendor talks whose separate finance products were outside the delivered session. The audit also reversed two provisional removals: Alithea Bio’s session materially covers agent procurement, payment rails, budgets, invoices, and receipts, while Allos AI’s session contains a sustained account of prior hedge-fund research agents, proprietary financial data, expert evaluation, and trader authority. Neither record establishes production settlement or a current-firm financial deployment.

The current selection-version 6 snapshot therefore contains 74 talks and 82 distinct speakers, with zero pending candidates. Its transcript manifest contains 74 captured records and 11,925 timestamped segments, with zero capture failures; 63 transcripts are labeled needs_review and 11 are labeled caption-unreviewed. The selected set includes 34 talks linked to organizer-classified finance organizations, three linked to investor organizations, 41 carrying a finance topic, and eight audited talk-scoped cross-tags; those routes overlap. All 46 excluded records remain preserved with stable IDs and reasons, so the corpus can be reproduced without allowing known false positives to return. New broad-route matches are emitted as pending review rather than admitted automatically. These are source-coverage counts, not capability scores. The organizer does not assert reusable transcript rights, so this article links to the delivered source while retained text supports verification rather than republication.

The table below is intentionally broader than the 74-talk finance-mechanics corpus. It preserves firm-associated AI architecture that matters to the wider landscape—such as coding agents, internal platforms, security controls, and procurement guidance—even when the precision audit excludes the delivered talk from the narrow finance-workflow count.

Organization	Narrow v6 corpus status	Publicly disclosed mechanism	What the evidence establishes—and does not
Two Sigma	Excluded; retained as broader firm-AI evidence	Shu Fang’s Tethered talk describes remote agents in per-user Kubernetes namespaces, running under employee identity, with an X-LLM-Agent attribution header propagated across downstream calls.	A delivered identity-and-provenance architecture. Fang says the header is attribution rather than authentication; the talk gives no user count, penetration result, incident rate, or investment outcome (timestamped audit).
Jane Street	Excluded; retained as broader firm-AI evidence	John Crepezzi’s 2025 developer-tools talk identifies an AI Assistance team building around Jane Street’s OCaml-heavy environment, proprietary infrastructure, Mercurial monorepo, code-review system, and editor mix. He describes representative developer-workflow data, domain-specific model training inspired by Meta’s CodeCompose, editor integrations, and task-specific evaluations.	A dated, delivered engineering account that helps explain why the firm builds custom AI tooling. It does not disclose model weights, corpus size, current team membership, rollout denominator, production error rate, trading-model use, or investment outcomes.
Morgan Stanley	Selected v6	AlphaLab turns a dataset and research objective into literature review, adversarial evaluation/backtest construction, and parallel experiments. Workers can write Slurm jobs and submit GPU work; separate critics check conceptual leakage and implementation correctness. The accompanying paper names GPT-5.2, Claude Opus 4.6, Claude Sonnet 4.6, four H100 NVL GPUs, and a persistent playbook.	A public research harness and paper, not a reproduced result. The code-and-lineage audit found that the release lacks the immutable datasets, generated experiments, checkpoints, baseline artifacts, traces, environment lock, and campaign databases needed to recompute the headline results. A fresh test run at the pinned commit produced 249 passes, five failures, and 21 errors. The speaker reports several unnamed models moving through internal risk review, but does not identify their desks, approvals, live use, or returns.
Millennium	Excluded; retained as broader firm-AI evidence	Brian Lewis’s enterprise-procurement talk names zero-data-retention or customer-managed encryption, SCIM-linked entitlements, admin APIs, configuration audit logs, controlled rollout, support engineering, SLAs, subprocessor transparency, and IP indemnity as buying requirements.	Personal procurement guidance from a speaker the organizer associates with Millennium. Lewis expressly says he speaks personally; the talk discloses no Millennium vendor, model, contract, dataset, or deployment.
JPMorgan Chase	Selected v6	A learned-execution-graph talk represents each payment-API request as a short-lived DAG, learns client- and payment-specific baselines, and discusses KL divergence, exponential moving averages, OpenTelemetry, Kafka, and tail-based sampling.	A proposed or experimental observability module. Its public example uses DeathStarBench and injected anomalies; the talk does not identify a production JPMorgan service, false-positive rate, alert volume, or adoption count.
Nubank	Selected v6	Skill Vector treats skills, plugins, MCP servers, and rules as a software supply chain. It combines deterministic and contextual-model checks, emits SARIF, comments on pull requests, and gates admission to an internal marketplace.	Nubank’s September 2 employer account reports reviewing more than 2,000 skills, identifying approximately 1,600 potential risks, producing about 1,000 remediations, and escalating roughly 90 priority reviews. Those are workflow counts—not precision, recall, confirmed-vulnerability, or prevented-incident measures. A separate Nubank/Snowglobe talk reports simulation outcomes without sample sizes or confidence intervals.
FactSet	Selected v6	Yogendra Miraje’s skill-centric financial-applications talk describes a move from proprietary blueprints to the Agent Skills format, selective skill loading, a governed registry, named maintainers, automated gates, and human admission review.	A described migration, governance model, and financial-research/report-generation demonstration. It does not establish customer deployment, adoption breadth, or investment performance.
Intuit	Selected v6	Udi Menkes’s financial-advice talk argues for causal treatment-effect estimation over business trajectories represented as state, action, and outcome sequences, followed by reinforcement learning.	That remains a research direction rather than a disclosed product. Separately, Intuit’s June 2026 engineering account says its TurboTax explanation system grounds Anthropic Claude through Amazon Bedrock in Intuit’s tax engine, tested thousands of use-case examples, and delivered contextual explanations to millions of individual filers with tax-expert support. This does not establish that Claude calculates taxes or that the causal/RL design was deployed.
Writer	Mixed: FailSafeQA selected; graph-RAG talk excluded	Sam Julien’s 2025 graph-RAG talk names Palmyra Fin and FailSafeQA and sketches context-aware splitting, text-to-graph extraction, graph-derived JSON in a Lucene-based search engine, and FiD-style answer generation. Writer’s Palmyra Fin post says the model used curated financial and finance-instruction data. Its FailSafeQA release, dataset, and paper expose a long-context finance evaluation covering malformed, missing, OCR-corrupted, irrelevant, and out-of-domain context.	Vendor research and a delivered historical architecture account, not a named financial-institution deployment. Writer evaluated 24 models with Qwen2.5-72B-Instruct as judge and reports vendor/paper metrics rather than production error rates. Current migration guidance lists palmyra-fin as deprecated and unavailable. Julien’s current site places him at CopilotKit, and he is not an author of Writer’s graph-generation paper.
Form3	Excluded; retained as broader firm-AI evidence	Moritz Johner’s PatchPilot talk describes deterministic Go orchestration around constrained agentic remediation for vulnerabilities spanning base images, runtimes, dependencies, and CI. The design treats an agent with production credentials as a software-supply-chain actor, restricts network access, minimizes changes, and ends in human-reviewed pull requests.	A delivered production-engineering architecture and explicit permission boundary. It is not evidence of autonomous merging, a published benchmark, an incident-rate reduction, or adoption by another institution.
Navan	Selected v6	Roberto Milev and Uday Kanagala’s operating-agents talk describes persistent sessions, memory, progressively loaded skills, tool-call tracing, trajectory evaluation, and fine-grained authorization.	A firm-presented operating pattern. The speakers identify cost control, replay, debugging, and standards as unresolved; no user denominator, reliability series, or financial outcome is disclosed.
Checkout.com	Excluded; retained as broader firm-AI evidence	Talha Sheikh’s Vector Harness talk describes Claude Code hooks, deterministic checks, asynchronous validation, model-judged evaluation, and code-review feedback intended to catch agents that report completion without satisfying requirements. A separate firm account says Agent HAL generates 18% of pull requests through narrow event-driven triggers, fresh-context review, tests, security scans, named approval, and no automatic merge.	The talk and firm page describe two bounded engineering systems. The public record does not connect HAL to Sheikh’s Vector Harness, publish the evaluation corpus or false-pass rate, or report quality outcomes for the 18% pull-request share.

The broader firm-associated archive also contains operator disclosures that are not investment-model claims. Northwestern Mutual’s GenBI talk starts from verified reports, dashboards, metadata, and BI practitioners before expanding to managers. A joint Freeplay and Chime session assigns customer-experience and quality-assurance teams roles in iterative evaluation, human review, hallucination monitoring, and feedback; the precision audit excludes it from the narrow finance-mechanics count because it discloses no Chime financial data, transaction, or banking control. A Wisedocs speaker compares coding agents on a legacy multi-repository software benchmark associated with the company’s claims domain; the public record does not show agents processing live claims, using customer data, or making claims decisions. Ramp discusses AI-supported scoping and implementation in forward-deployed engineering; and Filed frames tax work as supervised delegation rather than chat in its vertical-agent talk. Filed’s later integration disclosure says read and write permissions are separately controlled and write access is off by default. These accounts expose workflow choices and control surfaces, but their company-reported outcomes are not independent evaluations. Investor-affiliated exclusions remain preserved as ecosystem or thesis evidence where relevant; they are not counted as investment workflows without material investing, fundraising, portfolio, or capital-allocation mechanics.

One investor-labelled route does expose an internal research substrate. SignalFire says Beacon AI tracks 650 million people, 80 million companies, and roughly half a trillion data points for recruiting, go-to-market work, and investment research. Its technical account names 40 datasets, proprietary machine-learning ranking, GPT summarization and classification, and feedback from investment, talent, and data teams. These are changing firm-reported inventory counts and architecture claims, not a released dataset, model-validation result, bias audit, false-positive rate, or investment-outcome study. The conference talk does not establish personal ownership by its SignalFire speaker.

Payment providers expose several enforceable control primitives outside the model. Circle’s Agent Stack documents human-defined global and per-service spending limits, time-bounded sessions, and address, contract, and chain allowlists. Stripe’s Machine Payments Protocol release documents HTTP 402 requests, PaymentIntents settlement paths, and fiat or stablecoin payment options; its agentic-commerce guidance adds seller-, amount-, and time-scoped credentials enforced outside the agent. PayPal’s Agent Ready and Agentic Commerce documentation exposes customer-approval and tokenized payment-method paths. These are provider-documented capabilities, not evidence of transaction volume, loss reduction, customer adoption, production escrow, or removal of human approval.

Institution-controlled governance and evaluation records add context without validating the conference implementations. BlackRock’s October 2024 regulator response places AI inside human involvement and a three-lines-of-defense model spanning deploying teams, risk management, and internal audit; it does not name the custom knowledge application or its speakers. AXA’s July 2023 release documents Secure GPT on Azure OpenAI for an initial 1,000 Group Operations employees, but does not connect that deployment to the later DSPy optimization talk or prove its stated 140,000-employee rollout aim was reached. Bloomberg’s April 2025 responsible-AI release reports an 11-model, 16-category RAG safety study and a finance-specific risk taxonomy covering confidential disclosure, counterfactual narrative, impartiality, and financial-services misconduct. That is Bloomberg-authored research, not a production incident rate or evidence that every Bloomberg agent uses the taxonomy.

One implementation pattern repeats across otherwise different organizations: model work sits inside a deterministic or policy-enforced envelope. Jane Street describes executable parsing, type-checking, compilation, and tests; Form3 separates agent reasoning from a privileged orchestrator and microVM verification; Checkout.com feeds failed deterministic checks back into the coding loop; Ramp combines durable Temporal activities with scoped tools and approval pauses; and Navan names trajectory evaluation and fine-grained authorization. Block’s Agent Client Protocol and MCP address client/harness and tool interfaces; that talk is excluded from the narrow v6 corpus and retained here as broader protocol evidence. Edge & Node’s selected agent-spending talk separates paid-tool access from merchant-side transaction screening. Protocol support alone does not supply organization policy, identity assurance, budget authority, or investment permission.

The personnel expansion also shows why talk-time attribution and current ownership must remain separate. The organizer’s John Crepezzi biography places him on Jane Street’s AI Assistance team, while the delivered talk supports the system description. A Ramp cofounder’s appointment announcement says Rahul Sengottuvelu became CTO after three years leading Applied AI; his Scaffold Wisely talk separately describes code execution, parallel attempts, and a verifier. Form3 speaker Moritz Johner has an inspectable open-source bridge through the External Secrets Operator, but that project does not prove ownership of PatchPilot. Checkout.com’s public Agent HAL account is company context and does not assign HAL to conference speaker Talha Sheikh. Those distinctions prevent a conference affiliation from becoming an unsupported personnel or system-ownership claim.

The 137-person predecessor-roster source audit also recovered authorship and chronology edges for people who remain in the current corpus. Aman Gupta’s author-controlled profile connects his Nubank role to agent evaluation, alignment, training, and inference, while a 2026 Nubank paper documents offline simulation, human judgment, and online experiments for customer-support agents. Anju Kambadur is a named BloombergGPT coauthor, which establishes paper authorship rather than ownership of Bloomberg’s later agent systems. Lucas Palma’s employer-published Skill Vector article supplies a direct identity bridge to the skill-admission controls described in his talk. Waseem Alshikh’s FailSafeQA paper and released dataset connect him to finance-domain robustness evaluation. FactSet’s Yogendra Miraje biography and Mercury architecture article add employer-controlled role and system context, while leaving a title discrepancy with the later conference introduction. These links establish identity, authorship, and dated technical context; they do not establish sole system ownership, current firmwide deployment, or investment outcomes.

The current-person deep-gap pass adds several more bounded links. Cornell adviser Carla Gomes’s research page identifies Brendan Rappazzo as a 2025 computer-science graduate who moved to an AI/ML research role at Morgan Stanley; his earlier GEM-RAG paper establishes retrieval research and an adviser/coauthor network, not ownership of AlphaLab. Infant Vasanth’s author-controlled site gives the current title Senior Director of Engineering, Quant Research Platform, Data & AI Acceleration at BlackRock, while the joint BlackRock talk does not allocate platform ownership between its speakers. Anna Spysz’s résumé dates her move from AWS to Stripe and resolves the chronology behind her payment-agent demonstration. Rajiv Shah’s GitHub and agent-harness repositories place his current public work at OpenHands rather than the Contextual AI affiliation on the older talk. Klarity’s event page and an MIT innovation profile identify Nischal Nadhamuni as cofounder and CTO and connect him to MIT; neither source independently validates the performance of the E-Values framework discussed on stage.

The 109-speaker prior-media audit joins 29 future-roster speakers to 53 distinct earlier organizer talks. All 53 now have durable local transcript text after 11 missing organizer transcript objects were recovered; 22 retain timestamped segments and 31 retain cleaned text without cue timing. Five of the 53 are selected v6 records, three are excluded-but-preserved records, and 45 sit outside the narrow finance corpus. Preservation status does not change talk materiality or editorial review state.

Three of the eight current audited workflow cross-tags add controls that a firm-name filter would miss; the other five are summarized above. OpenGov engineer Gabe De Mesa says he helped build OG Assist and describes an Effect-native TypeScript loop, deterministic interruption for approval-required tools, rolling context summarization, and tracing in the delivered talk. Eliza Cabrera’s Workday-era session identifies product-management work on a financial-audit agent, policy agent, and assistant; her current author-controlled record places her at Smarsh on an agentic compliance-investigation and e-discovery product, so the Workday title must remain historical. Onlay founder Vasant Kearney’s healthcare-agent talk uses X12 insurance transactions as a constraint and validation structure and describes database-backed scoped memory, permissions, handoffs, and evaluations. These are public-sector, enterprise-finance, and healthcare revenue-cycle workflows—not hedge-fund systems—and no customer-controlled adoption or outcome evidence was located.

The archive changes the evidentiary shape of this research. A schedule says what someone intends to discuss; a delivered talk can expose controls, data flow, models, evaluation, and deployment boundaries; a paper or repository can make part of the claim inspectable. None of those layers automatically establishes firmwide adoption or investment value. The initial 75-talk transcript audit, 18-institution architecture delta, 29-speaker identity resolution, initial 81-person one-hop expansion, and corrected-taxonomy ledgers retain those layers separately. The version-4 137-person reconciliation records the predecessor discovery cohort and its evidence states. The current version-6 corpus contains 82 people after the five recall additions and the full 115-talk materiality audit. The two 68/69-person source-surface audits remain valid as audits of the frozen predecessor roster; they are not silently relabeled as the current corpus. These are provenance and selection states, not person or firm assessments. Earlier ledgers remain pinned to their original cohorts so later corrections cannot rewrite their denominators.

September 20 conference update: AI Engineer NYC names the people behind the finance preview

AI Engineer’s New York conference page first published a finance-heavy preview using company and role labels rather than names. Its September 20 schedule adds named evidence: the organizer identifies 109 speakers across 141 proposed sessions and supplies titles, companies, biographies, session descriptions, dates, and rooms. A reproducible local schedule snapshot preserves the underlying Sessionboard payload and marks 69 sessions as finance-relevant for review. The filter is a discovery aid, not a claim that every selected talk concerns investment management.

The bounded-exhaustion audit reconciles all 17 finance-relevant role/company pairs preserved from the post text, 19 finance/adjacent logo leads, 109 named speakers, 141 session records, and 70 dated repository ledgers. It finds no additional identity-safe current artifact outside the existing evidence set. That is not absolute exhaustion: 50 sessions still lack named speakers, 15 remain tentative, two named identities are unresolved, several preview roles remain anonymous, and the October 12–14 event has not produced recordings, slides, transcripts, repositories, corrected lower-thirds, or stage introductions.

The organizer’s later ticket announcement describes every talk as an in-production use case and says there will be no vendor pitches. This article treats that as an event-level marketing statement, not talk-level deployment evidence. The schedule still includes tentative assignments, vendors, demonstrations, proposals, and sessions that have not yet been delivered. Scheduled, delivered, speaker-reported production, customer-corroborated, and independently reproduced therefore remain separate evidence states.

The schedule-to-archive identity join finds six returning speakers with already delivered organizer records. Four remain in the narrow v6 corpus: Denys Linkov’s coding-agent comparison, Jan Curn’s sessions on the agentic economy and MCP and x402 payment limitations, Rahul Sengottuvelu’s Scaffold Wisely, and Stephen Chin’s context-graph talk. Three other delivered records are preserved but excluded from the narrow count: Linkov’s generic AI-team-structure talk, Max Kanat-Alexander’s developer-experience account, and Sai Krishna Rallabandi’s personal family-and-friends agent talk. The identity join supplies dated context; it does not establish that an older system will be discussed in New York or that an older affiliation remains current.

The named schedule corrects three plausible title matches from the initial announcement pass. Apollo’s Head of Data, Digital, and AI slot is Adam Nahari, not Vikram Mahidhar. Point72’s graph session names Amey Mahajan, not François Scharffe. Franklin Templeton’s session names Max Gokhman, while Deep Ratna Srivastav remains a separate publicly documented AI leader at the firm. This is a practical warning against promoting title matches to event identities before an agenda appears.

Firm	Named participant	Proposed public disclosure	Evidence boundary
Bridgewater	Aaron Linsky; Sam Green	“Vault: When Fifty Years of Structured Data Isn’t Enough”; a second session on keeping an AI analyst’s data private	Named, non-tentative schedule entries; not yet delivered talks or audited system documentation
Coatue	Frank Long	Featured keynote slot; title not yet published	Named schedule entry, but the organizer marks the session tentative
Point72	Amey Mahajan	Knowledge-graph patterns for connected financial data, provenance, access control, and production readiness	Joint session with Neo4j; proposed description does not prove Point72’s complete production stack
Citadel Securities	Joao Fiadeiro Wenzel	A proposed enterprise “world model” built from Slack, decks, meetings, screenshots, PDFs, and voice notes	Organizer biography says his remit includes firm-wide agentic AI; this remains speaker/organizer-provided disclosure pending the talk
Apollo	Adam Nahari; Rob Bittencourt; Jack McDonald	Private-equity operating-margin field report and an investor/operator panel	Intended conference disclosure, not an independent operating-margin or investment-performance audit
Wells Fargo	Freddy Lecue	A Wells Fargo foundation model applied to fraud detection at scale	Title is concrete; model architecture, data, denominator, and results remain unknown until the session is captured
Contour Asset Management	Hari Kumar	How the firm evaluates AI’s value in fundamental investing	Session description says Kumar leads internal AI systems; no public performance attribution is available yet
Fidelity Investments	Sai Krishna Rallabandi	Finance agents spanning earnings calls and investment decisions	Proposed talk, not proof of autonomous investment authority or realized alpha
Vanguard	Gareth Yoder	Cash-ledger modernization	Finance-infrastructure disclosure; the title alone does not establish an AI system
Capital One	Max Kanat-Alexander	AI-assisted software development and code quality	The schedule calls him Principal Software Engineer, while the preview said Distinguished Engineer; both labels are retained
New York Life	Vishal Srivastava	A proposed dataset of 1,368 AI failures and risk-based evaluation design	The denominator, sampling frame, failure taxonomy, and formal corporate title still require verification
JPMorgan Chase	Michael A. Davis	Intent-based authorization for agents	Security-architecture panel, not an investment-model disclosure
Bloomberg	Andrey Rybka	Recent AI attacks, incidents, and defenses	Security and architecture disclosure; the preview and schedule use different role labels

The proposed descriptions add system-level detail that the announcement did not contain. Bridgewater says Vault grew from a buy-versus-build decision into a centralized repository for third-party research, news, transcripts, and internal writing. The proposal says Vault serves people and agents, powers PAT for hundreds of daily users, and supports a second-generation system called Researcher. This is a speaker-submitted description of intended conference content, not a data-inventory, rights, security, or usage audit.

Wells Fargo proposes a transaction foundation model trained on billions of banking transactions, with heterogeneous-data representation, tokenization and sequence design, adaptive fine-tuning, multi-task learning, and applications in fraud detection and transaction decisioning. Contour Asset Management proposes a separate measurement layer: log each agent-generated investment idea with its source and reasoning, track whether it reaches a portfolio manager and becomes a position, and compare subsequent performance with benchmarks. Those are unusually specific intended disclosures, but neither session has yet supplied the model specification, evaluation sample, attribution method, or independent validation needed to assess the claim.

Citadel Securities’ proposal describes an enterprise state estimator built from more than 500,000 messages, 5,217 screenshots and PDFs, executive decks, and voice notes. The proposed system tracks beliefs and staleness over time and predicts operating events such as project slippage or client trading behavior. Point72’s joint session with Neo4j proposes a taxonomy spanning lexical, domain, and meta graphs; guided traversal; graph-filtered search; Cypher templates; Text2Cypher; graph embeddings; query-focused summarization; and agentic retrieval. Fidelity’s proposal names Earnings2Insights, multi-agent report generation, retrieval and evidence grounding, intermediate calculations, length-regularized DPO, and evaluation against human investment usefulness. These descriptions establish a post-event verification queue. They do not establish delivered production systems, firmwide scope, investment authority, or performance.

Two adjacent sessions add different signals. Coatue’s Frank Long is listed for a tentative keynote without a published abstract, but Long and Nick Gagnet separately published a July 2026 infrastructure thesis: GPUs generate agent actions as tokens while CPUs execute those actions as code, so repeated agent loops could create additional CPU demand. A separate Q2 2026 Coatue presentation names a Meeting Agent, Earnings Agent, Brainstorm Agent, and Expert Agent above a data and orchestration layer, while expressly reserving investment decisions to people. It does not attribute that system to Long or publish its models, evaluations, permissions, use counts, or investment results. Prior Labs co-founder Sauraj Gambhir is scheduled to discuss tabular foundation models in finance, including trading, lending, and risk examples. SAP’s September 2026 TabPFN-3.5 Plus release confirms product availability in SAP AI Core; the agenda’s comparative-performance language still requires defined benchmarks and independent verification.

Several preview slots remain unresolved as event identities, even where a current role holder can be matched. The organizer does not name any person for the anonymous BlackRock, Two Sigma, Bridgewater or Mastercard cards in the checked schedule. David Hefter is a current-role match for BlackRock’s investment-AI card, but the evidence comes from Constellation Research and a CalSTRS event biography, not an AI Engineer announcement. Two Sigma’s own 2026 material identifies Mike Schuster as Head of AI Core, while Google Cloud identifies Matt Chesler as SVP, Enterprise Platform Engineering Architecture. Each matches a different preview function; neither is confirmed for this event. Bridgewater’s Igor Tsyganskiy is a historical CTO match, while the current directory lists Oliver Radwan as Head of Technology and no CTO. Mastercard’s current exact-role match is Alissa Abdullah, deputy chief security officer. Ann Johnson is not the current-role match: Mastercard identifies her as EVP, Security Solutions and describes deputy CISO as her former Microsoft role. OpenAI and Anthropic remain TBA. The identity deep dive and role-resolution ledger retain current-role matches, stale matches and event participation as separate claims.

A current-state check of all 15 buy-side or trading-priority speakers found five chronology or title issues that should travel with the schedule. Gareth Yoder’s own public profile says he retired from Vanguard in February 2026, while the organizer describes him in the present tense. Apollo has not confirmed Jack McDonald’s organizer-supplied Principal title. Citadel Securities announced Joao Fiadeiro Wenzel’s arrival, but no firm-controlled page confirms the organizer’s COO label. John Feminella’s current biography places Two Sigma among prior affiliations despite two 2026 conference programs presenting him as an SVP there. Bridgewater’s current directory calls Aaron Linsky Head of Engineering, Alpha Engine, while earlier AWS material and the October schedule use different AIA Labs and product-engineering labels. These are source-specific title states, not interchangeable aliases. The 15-person current-state ledger preserves the dated evidence, collisions and unresolved employer confirmations.

The schedule also resolves one logo-wall organization only to the exhibitor level. It lists a non-tentative “Optiver Expo Session 1” on October 13 from 2:55–3:15 p.m., but provides no speaker, abstract or topic. That supports an expo-session relationship, not a named Optiver participant, a technical disclosure or use of any system. D. E. Shaw, AXQ Capital, RELAI and Extend remain artwork-level discovery leads in the preserved snapshot rather than named schedule participants.

The schedule also expands the monitoring vocabulary beyond firm names. Proposed sessions cover investor-context capture, foundation models for fraud, earnings-call agents, fundamental-investing evaluation, knowledge graphs, agent data entitlements, model-risk evaluation, AI-assisted software development, and enterprise state estimation. Those phrases now belong in recurring conference, job, podcast, paper, GitHub, and social-profile queries. The capture ledger preserves confirmed identities, superseded hypotheses, unresolved roles, logo-only leads, and prior-event cards as separate evidence classes. The first generator version checked the 81 people selected by the broad finance-session filter and therefore was not a complete roster audit. A second comparison against all 109 named speakers recovered 28 omitted people. The corrected all-speaker coverage audit now iterates the complete roster and retains selected/omitted status only as a triage field; runtime invariants fail if the two cohorts do not reconcile to the schedule total. The full-roster correction documents the false negatives, exact sources, unresolved identities, and revised ingestion method. These counts measure source coverage, not person or firm capability, and the priority classes route work without ranking people or firms.

The first personnel follow-up adds lineage and operating-system evidence without collapsing prior work into current-employer claims. Before joining Citadel Securities, Joao Fiadeiro Wenzel worked on Catena Labs’ open-source mixture-of-agents library, which names Llama 3.1, Mixtral, and Qwen proposal and aggregation models; that archived project is not Citadel technology. Long Lake founder Prathik Naidu’s earlier record includes the DataWig missing-value system and Scale Document, while current public descriptions name workflow-derived evaluations and post-training open models on proprietary operating data without disclosing the models, training method, evaluation set, or portfolio-company results.

Several non-investment firms expose controls directly relevant to a quant shop’s software factory. Monzo says Agent Chip authors about 10% of merged pull requests and runs more than 1,800 daily tasks through isolated namespaces, a centralized MCP gateway, per-task access controls, constrained proxies, team attribution, and human review. Coinbase engineering manager Evan Kormos’ Interrupt recording describes customer and internal support agents using self-hosted LangSmith, a remote documentation MCP server, RAG fallback, deterministic safeguards, output review, and production-trace backtesting. Ramp’s Stack benchmark runs accounting-agent tasks five times across synthetic business worlds and reports pass, consistency, and criteria-accuracy measures; its ablations found that some skills reduced performance and that period-specific memory could contaminate later work. A Credal case study of Wise describes model-independent access, audit logs, cost tracking, employee RAG, a 14-criterion support QA agent, and human-reviewed financial-crime paperwork; its productivity figures are vendor-reported and lack a disclosed sample or control. The technical follow-up ledger preserves these boundaries and the timestamped findings from all 13 previously missing archive recordings.

Wise’s own filings now provide a first-party boundary alongside the vendor case. Its April 2026 UK prospectus says machine-learning models use patterns across millions of customer profiles and transactions for real-time alerts and that LLM copilots have launched for Financial Crime and Customer Support teams. Wise’s current investor page separately names AML/fraud, treasury and liquidity management, customer-support automation, and AI access across product, marketing and financial-account preparation. Its customer-facing AI assistant page preserves a human-contact route. These sources still omit model identities, evaluation sets, false-positive and escalation rates, user counts and measured outcomes. They also do not resolve scheduled speaker Emil Chetty’s title chronology or make him the owner of these systems; the follow-up ledger keeps the personnel and firm evidence separate.

Revolut’s current record separates an employee enablement layer from a proprietary predictive-model program. The jointly authored PRAGMA paper describes 10-million, 100-million and 1-billion-parameter encoder models trained with masked modeling over 26 million anonymized user histories, 24 billion events, 207 billion tokens and 111 countries. Separate profile-state, event and history encoders produce representations adapted through frozen probes or LoRA updates over roughly 2%–4% of parameters for credit, fraud, engagement, recurring-transaction, lifetime-value and recommendation tasks. The paper withholds absolute scores and also reports the critical failure boundary: PRAGMA-L trails a network-aware AML baseline by 47.1% on F0.5 because isolated account histories omit cross-account relationships. Revolut’s research-division announcement names Anton Repushko as Head of Revolut Research and Pavel Nesterov as Head of AI; NVIDIA’s customer case supplies H100, Nebius, packing and batching details but its production-lift figures remain joint customer/vendor claims.

Revolut separately describes AIR as a customer assistant spanning financial analysis and bounded in-app actions, with a stated zero-retention/no-external-training policy for third-party AI partners. Its 2025 annual report says a GenAI chatbot resolved more than 75% of support interactions by year-end, without defining resolution, sample, escalation, repeat-contact or quality measures. Scheduled speaker David Abitbol is not a named author or owner in any of these artifacts; the follow-up ledger keeps his shared-platform statements separate from PRAGMA, AIR and firm-level support automation.

A dated delivered talk now clarifies Abitbol’s own remit. France FinTech listed him as Revolut’s AI & Infrastructure Lead for an April 22, 2026 session with Google, a more specific title than the October organizer’s “Ops & Infrastructure Leader.” In his public recap, he says Revolut moved from a centralized AI team that had become a bottleneck to shared infrastructure—an LLM gateway, MLOps platform, frameworks, evaluation tools and deployment playbooks—used by product teams. France FinTech’s event summary attributes to Revolut’s Rita support agent a move from 35% to approximately 80% resolution without human escalation, but provides no time window, case mix, query denominator, quality threshold or independent audit. The annual report independently corroborates controlled releases, versioning, validation, anomaly monitoring, evaluation, provider routing and proprietary fine-tuned models. It does not assign each component or Rita’s implementation to Abitbol. The concentrated institutional ledger preserves the delivered-talk chronology and paper-contribution boundaries.

Wise’s current hiring record supplies another first-party architecture view without proving the proposed system called LOOP. A Machine Learning Platform role describes model serving, training pipelines, a registry, experiment tracking, feature management and monitoring across hundreds of models and billions of events, with fraud, treasury and personalization among the use cases. A Fraud Automation role explicitly combines deterministic logic, ML, GenAI, agentic workflows and human judgment, including prompt systems, evaluation, failure monitoring, feedback loops and safe rollout. A Servicing data-science role names computer vision for identity verification and LLM workflows alongside fraud and money-laundering detection; an AML Handling role names fine-tuning, reinforcement-learning alignment, LLM evaluation and test suites. These are current mandate and platform signals. They neither verify LOOP’s name or production status nor establish Emil Chetty as the owner of every listed component.

A second person-by-person pass resolved or bounded all 33 finance-adjacent speakers that had no substantive repository coverage. The public artifacts are useful reference designs, not evidence that a tracked manager uses them. Epsilab’s SDK exposes immutable tasks and verifiers, deterministic seeds, complete model/tool/cost traces, trajectory replay, batch regression, and GRPO/DPO/SFT/KTO exports; the scheduled investment-banking environment itself is not public. LiveKit’s human-review pattern separates proposal from commitment for high-value financial actions, while its asynchronous-tool design separates cancellable reads from non-cancellable writes and adds duplicate-call controls. turbopuffer’s permission contract enforces user/group filters during retrieval but leaves authenticated identity-to-filter construction to the application. Composo’s OmissionBench shows that judge designs can detect added or altered controls much more reliably than missing information, a directly relevant warning for finance-agent evaluation.

Composo’s separate RewardBench 2 record illustrates a configuration distinction that can look like a version conflict when only headline scores are captured. Its public repository reports up to 85.8% accuracy for criteria injection plus ensembling. The current paper resolves that figure as the mini-model, eight-sample condition at 1.3 times baseline cost; its full-model, eight-sample condition reaches 83.6%, an 11.9-point increase over the 71.7% full-model baseline at 5.3 times cost. Both numbers can be correct under different model classes. Any use of the score therefore needs the model tier, sampling count, sample set, and cost basis rather than a bare “RewardBench result.” Neither configuration establishes a financial-institution deployment.

The same pass found more application-specific disclosure and an important denominator distinction. Tidalwave’s March company/HousingWire summary calls its 90-question, 10-scenario SOLO-versus-Claude test a first iteration. The August 25 version of the jointly authored MortarBench paper is a later research artifact with 188 test cases across 47 unique questions, three trials per model, four Tidalwave authors and five Columbia authors. It does not report a SOLO row. The paper says the synthetic records were calibrated from anonymous summary statistics over 300 real bank statements and that questions were adapted from an in-production loan-assistant chatbot. Its CRIT confidence-and-threshold layer moved Gemini 3.1 Pro from 80.8% to 83.6% F1, while a RAG experiment using the Fannie Mae selling guide stayed flat or degraded. Its name-language probe reports a 15.7% foreign-origin classification rate for English names and 74.0% for non-English names across seven baseline models. The code and dataset record make the construct inspectable, but the documents are synthetic, the paper authors ran the evaluation, and scheduled Tidalwave CEO Diane Yu is not an author. The first-iteration SOLO results and the later MortarBench results should remain separate rather than being treated as one leaderboard or one denominator.

Monzo supplies a separate live decisioning example relevant to the two scheduled Monzo analytics speakers but not personally attributable to them. The July 2026 Fraud Prevent disclosure, authored by Tom Turner and Robin Dhamankar, says every transaction first receives calibrated fraud probabilities; higher-risk payments trigger a questionnaire and customer evidence; a multimodal model converts that material into structured semantic features; another supervised model produces a questionnaire-specific probability; and a meta-stacking ensemble combines that result with the transaction-time score. Monzo says the system runs in under ten seconds in the live payment path, increased both authorized-fraud value and case count prevented by more than 20%, and reduced legitimate-payment referrals to human review. The firm intentionally withholds fraud-sensitive implementation details, denominators and model identities, and the post does not make scheduled speakers Francesco Galletta or Felippe Felisola Caso its authors.

Monzo’s June 2026 Ops Agent engineering account exposes the rollout path from conversational support to state-changing bank operations. It starts with a 100-conversation expert-approved golden set—which the authors explicitly call too small—then adds daily replay, novel expert scenarios, component tests, semantic-similarity and LLM-judge checks, and end-to-end tests over anonymized real conversations. Every outbound message initially required human review; all conversations were first sampled for QA before the rate declined. Intent-gated transaction context reportedly raised transaction-query resolution by ten percentage points. Later versions encode processes as Markdown instructions with tool references, evaluate them with simulated users and stateful tool environments, and execute missing-refund and replacement-card workflows. The page does not publish the baseline, sample, model names, error rate, escalation rate or audit. Its named authors are Jamie McDonald-Gibson, Joost van Oorschot, Robin Dhamankar and Tom Leitch—not the two scheduled speakers. The follow-up ledger preserves that attribution boundary.

Two scheduled practitioners expose software-factory and voice-agent evidence outside investment management. A March 2026 GEICO Prompt-to-Prod account, written by Paul Devitt with Ford Prior acknowledging collaboration, describes prompt-authored requirements and acceptance criteria, AI-generated code, tests, documentation and work items, retry loops, separate code-generation and evaluation models, specialized AI review, staged promotion and a human final production gate for regulated environments. Devitt explicitly withholds GEICO-specific technologies, and Prior’s public profile separately describes a human-supervised claims “war room” applied to vehicle salvage. No task count, defect rate, audit sample or savings result is published. An organizer’s Observe 2026 announcement associates Upstart’s Shiv Indap with a production voice bot and names interruptions, background noise and multilingual users as failure classes; that corroborates a deployed voice-agent context but not the future AI Engineer session’s judge calibration, self-healing loop, model, rubric or call-volume claims.

Google Cloud’s Rogo case study names Gemini 2.5, RAG, visual processing, Spanner hybrid retrieval, provisioned throughput, and multiple licensed financial-data sources, but its performance and scale figures remain joint vendor/customer claims. Rowspace describes a firm-specific data and judgment layer spanning memos, models, positions, trades, accounting, and covenant documents, while its security page claims customer-environment deployment and inherited access controls; no public artifact shows autonomous trade execution. Affiliation checks also prevent false attribution: Ishween Kaur’s Salesforce author page conflicts with the schedule’s SoFi field; Richmond Alake’s 2026 Oracle work makes MongoDB a historical affiliation; and Stephanie Jarmak’s current site and Sourcegraph webinar conflict with the schedule’s Omni Analytics label. Fengjiao Peng remains unresolved, and the Mintlify/Coinbase session’s structured speakers conflict with its prose. The 33-person verification ledger records every reviewed person, direct source, negative finding, and proposed-versus-deployed boundary.

The remaining 84 non-established records were then split into four fixed, non-overlapping cohorts and preserved as complete person-level ledgers: entries 0–20, 21–41, 42–62, and 63–83. Each record separates direct identity evidence, earlier academic or employer lineage, current system evidence, conference proposals, contradictions, negative searches, and attribution limits. After correcting exact-name aliases for middle initials and expanded names, the regenerated audit reports zero schedule-only people, 82 people represented in three or more substantive repository files, and 27 represented in one or two. Those file counts measure evidence spread, not research quality, importance, or capability. The split is a queueing device, not a capability classification.

A subsequent current-state pass rechecked all 94 P2 and P3 records against current employer pages, personal sites, papers, patents, repositories, jobs and delivered media. The 18-person financial-institution ledger retains title conflicts for Andrey Rybka, Emil Chetty, Denys Linkov and Max Kanat-Alexander and keeps the Monzo, Coinbase, Revolut, Wise, Wells Fargo, Ramp and Capital One firm records separate from speaker ownership. The P3-A ledger records Galileo’s completed Cisco acquisition, Oracle’s current Alake record, Jarmak’s Sourcegraph and NASA SciX affiliations, GEICO’s human-supervised Prompt-to-Prod account, and delivered Oracle, AAuth, Apify, Brainbase and Rowspace artifacts. The P3-B ledger separates authorization, policy, identity and durable-execution claims across Keycard, Speakeasy, WorkOS and Temporal and preserves version boundaries for Vals, Coval, Greptile and other evaluation artifacts. Fengjiao Peng and Isabelle Williams remain identity-unresolved; Ramakrishnan Lokanathan remains organizer-biography-only. These ledgers establish source coverage and contradiction handling, not deployment or capability scores.

A parallel first-party finance-firm graph rechecked 22 event-exposed organizations plus a cross-firm monitoring map. It admits named people only through firm pages, papers, patents, repositories, official jobs or inspectable media; coauthorship is retained as an artifact relationship rather than a reporting line. The resulting graph adds or consolidates Bridgewater’s expert-judgment coauthors, BlackRock AI Labs advisers, Capital One’s Context Specs and VulnHunter contributors, JPMorgan research authors, BloombergGPT authors, Fidelity financial-NLP coauthors and RELAI paper/code contributors. It also records dead job and biography URLs as historical provenance rather than presenting them as current openings.

The buy-side and institutional follow-up adds four firm-controlled records that were absent from the first schedule pass. Apollo’s 2025 Form 10-K says employees use internal GenAI applications for summarization, search, translation, and information gathering and identifies risks involving bias, confidential information, third-party controls, and model or training-data compromise; it does not attribute those systems to scheduled Principal Jack McDonald. Franklin Templeton’s current Max Gokhman biography assigns him AI-agent, model, MosaiQ, investment-outcome, and digital-asset responsibilities, while a MosaiQ brochure names portfolio construction, heterogeneous-data validation, risk/order-management APIs, and the Pixel natural-language copilot. Gokhman also publicly describes analyst, sector-lead, and risk-agent dependencies with humans initiating requests (self-authored post); the implementation and results are not public.

Point72’s live careers inventory confirms a Knowledge Graph Intelligence organization spanning data engineering, research engineering, and product design, with one public role placing the work inside Compliance for investment operations. That corroborates an organizational unit, not the proposed Neo4j architecture or an investment result. Oracle supplies a more explicit adjacent control reference: SQLcl, ORDS, and OCI Database Tools expose different local, self-managed, and managed MCP models. Public documentation describes JWT validation, role/scope-to-pool authorization, database-native row/column/cell controls, execution logging, and bounded typed tools; no named financial customer or controlled authorization-security result was found. The institutional-personnel ledger records the role histories, direct sources, and unresolved claims without treating the conference as completed evidence.

The follow-up on the 22 people with established repository coverage adds deployable architecture around several proposed finance sessions. Two Sigma’s Shu Fang describes employee-bound agents running in per-user Kubernetes namespaces, with employee identity and explicit agent-action lineage; the delivered-talk archive above now preserves the timestamped control and attribution boundaries. This is firm-level evidence but is not attributed to scheduled speaker John Feminella. Ramp’s Rahul Sengottuvelu separately exposes a coding and operations lifecycle spanning automated agent triggers, review, production observation, incident analysis, internal coworkers, least-privilege controls, and current roles for an Agent Developer Platform and continuously operating agents. The reported CI and adoption measurements remain company/vendor accounts rather than independent evaluations.

Ramp’s engineering archive makes that software-factory pattern more concrete, while preserving a human production boundary. A March 2026 Ramp Labs account says its internal Inspect agent runs in sandboxed development environments, responds to Datadog alerts, reproduces faults, proposes fixes and posts to Slack, but cannot merge code without engineer review. Ramp reports expanding from ten hand-written monitors to more than 1,000 AI-generated monitors—about one per 75 lines of code—and catching 40 bugs in the first week; it also records noisy thresholds, duplicate alerts and weak prioritization from unfocused nightly runs. A separate May 2026 security-scan disclosure reports roughly 10,000 parallel Inspect sessions over eight hours, agent-based deduplication and adversarial regrading, seven retained high-severity findings, and an estimated full-price GPT-5.5 run cost above $20,000. These are firm-reported counts over a private codebase, not an external audit. Scheduled speaker Ryan Stevens directly authors the separate Stack accounting-agent benchmark, while a retained February 2025 interview establishes his earlier lending-risk ML remit and lifecycle discussion. Neither source makes him an author of the Inspect articles; the follow-up ledger keeps personal authorship, firm architecture and proposed-session evidence separate.

Stevens’s prior delivered material reveals a second, narrower Ramp agent and the production ML substrate around it. His July 2025 AI in Production slides describe an analytics-engineering on-call agent that generates a candidate change, runs or rebuilds with dbt and checks execution. The evaluation model separates syntax, architecture and business correctness; code review supplies structured labels for prompt feedback, while people retain the business-context and maintainability judgment. His later public profile account describes almost 100 production models, thousands of upstream tables, hundreds of AWS batch-inference jobs, custom sensors, dataset-aware triggers, dynamic DAGs and Slack debugging. This evidence belongs to Stevens and Ramp, but it does not transfer authorship of the Inspect systems to him or establish a portfolio-management use case.

Capital One exposes a different regulated software-factory design around the scheduled code-quality session. Its open-source Context Specs repository places a deterministic dispatcher around probabilistic coding steps, feeds agents short implementation slices, separates disposable feature specifications from curated project memory, runs executable checks and returns either a pull request or an explicit STUCK diagnosis. Its separately released VulnHunter repository uses attacker-first path analysis, a stage that tries to falsify its own vulnerability claim, targeted repair proposals, a restricted fix verifier, headless execution and a public benchmark harness. Capital One also reports an enterprise AI learning hub for more than 60,000 associates and a month-long agentic-coding program that began with more than 10,000 engineers, but publishes no controlled productivity or proficiency result. The artifacts are inspectable; enterprise adoption and outcome denominators are not. Scheduled speaker Max Kanat-Alexander is not a named author of these reviewed artifacts, so the follow-up ledger treats them as firm context rather than personal ownership.

The completed roster audit now has four complementary evidence layers. The buy-side and institutional deep dive resolves delivered talks, current roles, prior papers and firm-versus-person attribution. The vendor deployment deep dive checks named customers, permissions, data, evaluation and adoption denominators. The research and identity deep dive resolves academic lineage, repositories, model versions and title conflicts. Finally, a title-blind media graph compares all 109 names with the official channel’s 803-video manifest and 801 retained caption files. It initially found 32 exact-name prior recordings across 18 people, all with local captions. The other 91 people were then searched individually across publisher pages, podcasts, video, slides, papers and code in waves A, B, C, and D. That pass recovered official recordings whose names appeared only in canonical talk-page or embedded-player metadata, so future archive checks now union channel titles, captions, talk pages, and player IDs. The finance, wave-B, controls, and general manifests preserve canonical URLs, hashes, durations, transcript provenance, and unresolved gaps.

Samaya publishes a separate investment-research training and evaluation path. Criteria-Eval uses dated financial questions, expert-authored binary criteria, public-information cutoffs, daily regression runs, and reinforcement-learning signals. Its June 2026 query-understanding system reports supervised fine-tuning followed by GRPO-style reinforcement learning on Qwen 3, model-size experiments, held-out entity/time/abstention metrics, and an 8B deployment choice. Those artifacts support retrieval, reasoning, and analyst-workflow claims; they do not establish trading authority or a released result for the conference’s proposed time-rewound stock-selection experiment.

Basis and WorkOS expose two additional control patterns. Basis’s behavior specs are hidden Markdown contracts judged against complete accounting-agent trajectories, while its internal “Satellite” account describes a 36-provider MCP gateway with user and service identities, provider allowlists, per-user OAuth, encrypted refresh tokens, request-time tool filtering, and structured telemetry. WorkOS’s Blog Bot uses a fact ledger, fresh-context critic, version checks, canonical author resolution, broken-link blocking, and a draft-only publication boundary; Nick Nisi’s demonstration separately gates progress on a SHA-256 hash of actual test output. These are adjacent enterprise controls, not proof that a named manager has adopted them. The established-personnel ledger preserves the full 22-person verification record, role conflicts, prior-employer boundaries, direct sources, and negative findings.

A newly retained long-form Basis interview makes the operating model around those behavior specs more concrete. Troyanovsky describes a dedicated accounting-product-operations team writing the specifications, evaluators reading complete multi-agent trajectories, and ten-hour runs producing inspectable traces rather than only a final answer (team and trajectory discussion; ten-hour-run discussion). These are founder descriptions preserved from automatic captions, not an independent reliability audit. They nonetheless separate three layers that should be tested independently in a financial software factory: the work product, the hidden behavioral contract, and the execution trace.

Neo4j’s May 2026 decision-traces talk supplies a different architecture reference. Its worked financial-analyst example joins customer records, transactions, policies, risk factors, prior approvals, precedents, and the reasons behind earlier decisions; it then retrieves candidate precedents through graph structure and vector similarity. The speaker also says that automated writing of new traces and trace-quality scoring were still open implementation questions (closing discussion). This is a vendor demonstration using a financial scenario, not evidence that Point72, Bridgewater, or another named institution deployed the design.

All-speaker correction: what the finance-session filter omitted

The full 109-speaker comparison found a material false negative in the core firm set. The agenda lists Frank Long as Coatue’s Head of AI, but the name-only tentative keynote carried no finance vocabulary. Coatue’s own January announcement gives the fuller title Head of AI & Partner on the Hedge Fund team. A November 2024 SEC filing dates Mosaic investment to 2015 and the Coatue Brain launch to 2023, before Long joined; the public record therefore supports organizational proximity, not authorship of the Brain. A February 2026 Coatue disclosure separately reports 35 data scientists and engineers, current Claude use for research-report compilation and long-idea generation, Claude Code agents running data-analysis scripts for hours, and Claude Skills being built to fetch earnings transcripts. Coatue co-led Anthropic’s financing and discloses that it may benefit from that holding, so the vendor, customer, and investor relationships must be read together.

The name-only Rayan Krishnan / Vals AI slot exposed a second direct finance branch. Krishnan coauthored the original Finance Agent Benchmark: 537 expert-authored questions across nine categories, an agent harness with Google Search and EDGAR access, and paper-reported o3 accuracy of 46.8% at an average $3.79 per question. The later Finance Agent v2 harness adds Tavily-backed web search, EDGAR search, HTML parsing, stored-information retrieval, price history, and per-question logs for tool use, token counts, and errors; Vals’s release material describes a changed taxonomy and more than 900 questions. The original and v2 result tables should not be treated as one time series because both the records and harness changed. Vals also lists private finance benchmarks for agent research, long credit agreements, mortgage-tax images, and tax questions, while an August 2026 policy interview describes benchmark decay and domain variance. The private records prevent independent record-level audit, benchmark scores are not deployment outcomes, and the future New York session adds no delivered evidence yet.

Horizontal vendor sessions also contained finance deployments. Browserbase’s Ramp case describes browser agents for receipt collection and procurement research across merchant sites; its Parcha case covers bank and fintech onboarding and due diligence. Those are company-level cases and should not be attributed personally to scheduled Browserbase Growth Engineer Jay Sahnan. Speakeasy’s MCP authorization release gives a concrete permission example in which a finance analyst may read a warehouse but not mutate it, backed by SSO, SCIM, role scope and gateway enforcement. Oracle’s Wojciech Pluta publishes runnable patterns for scoped agent memory, retrieval regression, provenance, deletion, ACLs and freshness. These are control references, not named fund deployments.

Five omitted software-engineering sessions form a useful reference chain for a financial software factory. Warp’s Skill Doctor scores historical agent traces and proposes reviewed skill changes. Greptile’s v4 account uses live behavior proxies while disclosing an LLM-judged metric. Cursor describes model-specific harness tuning, keep-rate measurement, error taxonomies and automated remediation. Sonar’s post-acquisition Gitar/SonarQube demonstration preserves independent AI and deterministic verification paths and says they did not yet share findings or a common gate. Linear’s coding sessions combine managed sandboxes, repository scope, reviewable diffs and human merge approval. Together they expose versioned instructions, trace-based evaluation, independent verification, bounded execution and human decision gates; none establishes adoption by a tracked manager without a customer-controlled source.

Customer-side checking narrows several vendor claims. Ramp’s own Browserbase page confirms browser infrastructure behind procurement, receipt and bill-pay agents, while Ramp’s agentic-payments guide requires scoped permissions and human confirmation before purchase. NAB’s AI archive confirms employee use of Cursor, but does not independently establish the scale or speed figures in Cursor’s NAB case. A Deutsche Bank team published a Temporal payments-architecture deck that keeps business data outside Temporal; no customer-controlled source was located for the separate ANZ outcome claims. These distinctions matter: confirmed use, rollout size, architectural design and measured business effect are four different claims.

Oracle’s product notes expose decision gates at a more operational level. The 26C Cash Processing Agent ingests email, bank-file and remittance documents, handles exceptions and invoice matching, and requires user confirmation before applying a receipt under existing authorization, segregation-of-duties and audit controls. The 26D cash-positioning and transfer workflow recommends transfers, waits for user acceptance, then generates, submits and monitors them. These are shipped capabilities without a named adopting institution. They provide a public reference for separating document extraction, recommendation, authorization, execution and settlement monitoring.

The omitted-person pass also recovered inspectable research lineage. Manimala Kumar is a named inventor on Honeywell patents for an industrial knowledge graph and operating-window optimization; those patents establish prior industrial work, not NVIDIA or finance deployment. Pim de Witte is one of 29 authors of MIRA, whose repository exposes a multi-actor world-model implementation trained on public-bot game play; market simulation is an analogy, not a demonstrated use. Sarah Chieng’s autoresearch harness exposes looping model comparisons, while Rayan Krishnan’s Finance Agent repository exposes the harness around a restricted record-level benchmark. The second-wave ledger retains recovered media, collaborators, schools, identity conflicts and negative customer findings.

Joint conference appearances create a relationship queue rather than a deployment map. Bridgewater’s Aaron Linsky is paired with turbopuffer’s Nikhil Benesch; Point72’s Amey Mahajan with Neo4j’s Stephen Chin; and JPMorgan Chase’s Michael A. Davis with Keycard’s Ian Livingstone and AAuth author Dick Hardt. The AAuth draft and repository expose intent-bound agent authorization. Davis’s separate September 2026 Insecure Agents interview discusses runtime tool authorization, compositional permissions, memory, exfiltration and uncertainty, but does not establish bank-wide deployment of the discussed reference architecture. Likewise, the Point72 and Bridgewater pairings identify technologies worth checking after the event, not procurement, production status, or investment outcomes.

The first coauthor and collaborator expansion adds inspectable control surfaces without turning vendors into financial-institution deployments. Finance Agent coauthor Langston Nashold is also a named contributor to Valkyrie, whose public implementation exposes hosted and self-hosted benchmark orchestration, sandboxing, artifact storage, secret handling, masked credentials, service authorization headers and run tracking. Samaya’s Criteria-Eval contributor statement separates design, annotation leadership, engineering and analysis across Ashwin Paranjape, Christos Baziotis, Jack Hessel, Jack Silva and Mingyi Yang; the company reports dated financial queries, expert-authored criteria, held-out human labels and daily regression runs, but does not release the full corpus needed to reproduce its measurements. Oracle’s agent-memory proof of concept, database-enforced authorization design, unified memory reference architecture and fraud-detection tutorial expose named authors and concrete retrieval, row/column/cell policy, tenant isolation, audit and graph/vector patterns. They remain vendor-authored proofs of concept, product designs or tutorials without a named adopting institution. The one-hop personnel ledger records qualified and rejected graph edges separately.

A title-blind media search recovered a dated baseline for Arena Investors. In a 2023 Linedata panel, CTO Ryan Houser said Arena combined substantial proprietary development with trusted partners for scale. At 22:36, after identifying the other manager by name, Houser described offshore or outsourced work spanning research, marketing outreach, initial deal-flow funnel work, fund accounting, operations and senior software development. GoldenTree co-CTO Chris Beels separately connected external capacity to front-office credit analysis. The managers also advised treating public GenAI inputs as potentially public information. The retained timestamped transcript note is a 2023 operating-model disclosure, not evidence of either firm’s current models, controls or results; it creates a concrete comparison point for Houser’s proposed 2026 “assembly line” session.

The same title-blind method recovered Linedata’s complete London and New York 2024 replay pages. In the London recording, Man Group Head of Enterprise Engineering Tom Price says the firm had built a security-approved internal ChatGPT-like proxy and connected it to JupyterHub. At 29:09–31:01, he describes quantitative researchers using it to write Python against internal market-data libraries and infrastructure, along with use in sales, marketing, legal review, due-diligence questionnaires, client-letter obligations, and regional regulatory analysis. Price also says GenAI represented a small share of Man’s technology headcount and spending at that point and was not yet used in client servicing. The recording does not disclose model providers, user counts, entitlement design, evaluations, production-code controls, or investment outcomes.

GoldenTree’s 2024 disclosures separate what was in daily use from what its co-CTO wanted next. Chris Beels says at 05:23–06:34 that the software team used GitHub Copilot daily and viewed it as speeding work, while Microsoft copilots received mixed reactions and BloombergGPT had not yet produced the use cases he was awaiting. At 35:32–36:30, Beels describes a desired interface that records the morning CIO call, lets a listener select a question, and presents an AI-generated answer while traders or analysts respond; he also describes document summarization and legal or regulatory point extraction as an area where GoldenTree was beginning to engage. A separate customer panel connects this to data architecture: Beels says AI increased the need for Snowflake-like infrastructure and natural-language questioning of firm data. None of these passages establishes the CIO-call interface as deployed or links the tools to portfolio performance.

That customer panel also identifies Sands Capital CTO and CISO Ryan Bateman. At 05:28–07:14, he says Sands began rebuilding its enterprise data warehouse in 2018, had reached a third cloud-based version, and was applying a DataOps model around cloud-native infrastructure, automated metadata tagging, lineage, and business intelligence. He names Databricks, Snowflake, and dbt as examples of the surrounding tool class and expressly says the cloud work had not yielded cost savings. The passage exposes a data foundation rather than a GenAI investment model; it does not confirm which example products were contracted or quantify the automated-lineage coverage.

The New York innovation panel supplies a smaller-manager control case. The recording identifies Ionic Capital Management’s CTO at 02:28–02:39, while an SEC filing resolves the ASR-mangled name as Arthur Vaccarino. At 20:27–22:04, Vaccarino says Ionic had explored tools for detecting AI use, found the then-available Microsoft Copilot and Purview controls insufficient for its needs, and deliberately limited AI use because the firm lacked the development capacity to build internal LLMs. A Jennison Associates technology participant in the same recording describes non-investment POCs and a four-to-five-month policy and permissions path from a successful POC to production, but the caption track does not preserve that participant’s name reliably. These are dated 2024 operating disclosures, not current-state or capability assessments. The full event note and retained transcripts preserve all 15 recordings, exact turn boundaries, and exclusions.

The same sweep found identity defects that need to remain visible. The schedule duplicated Killian Carlsen-Phelan’s first name; left Isabelle Williams, Jack Wotherspoon, Jay Sahnan, Daksh Gupta and Wojciech Pluta without role/company metadata; mixed Salesforce biography fields with a SoFi company label for Ishween Kaur; and combined founder history with an unverified NVIDIA reference for Ramakrishnan Lokanathan. The full-roster correction ledger records all 28 omissions, evidence classes, primary links and negative findings. The resulting process rule is simple: preserve and audit the complete named roster first; use topic filters only to order the queue.

Focused recovery replaces the clipped Wells Fargo and Fidelity records with published technical evidence. Freddy Lecue’s self-authored foundation-model recruiting post corroborates a Wells Fargo program exercised on billions of debit-card transactions and applied to fraud detection, but leaves the model, labels, data period, metrics, traffic share, and production controls undisclosed. His co-authored LLM Jury-on-Demand paper separately exposes a ten-model judge pool, instance-level XGBoost reliability prediction, dynamic judge selection and weighting, summarization and RAG datasets, Kendall correlation against human labels, repeated splits, Wilcoxon tests, and Cliff’s delta. That is evaluation research; the paper does not show that the jury evaluates the transaction model or runs in production.

The collaborator graph around that work resolves several distinct Wells Fargo surfaces. Pashmeen Mistry publicly names Tachyon as the enterprise AI/ML platform whose product remit includes generative and agentic AI; his Google Cloud Next session documents Wells Fargo participation in a Model Armor session covering prompt injection, jailbreaks, sensitive-data leakage, and runtime guardrails. Amrith Kumar’s GPU-platform recruitment names H100, H200, GB200, DGX SuperPod, vLLM, SGLang, Triton, TensorRT-LLM, FP8/INT4, KV-cache, prefill/decode, NVLink, NVSwitch, MIG, and NCCL as the engineering remit. A FHFA biography separately identifies Kiran Yalavarthy as head of enterprise model risk, while Wells Fargo and Google describe Swarup Pogalur’s horizontal AI-capability organization, reusable APIs, human oversight, monitoring, fallbacks, and branch-policy retrieval (Wells Fargo; Google Cloud). These records expose research, platform, governance, and application layers. They do not establish that Lecue manages those people, that every recruited component is operational, or that the transaction model uses Tachyon, Model Armor, or the named GPU stack.

The same pass separates older model-risk research from newer GenAI claims. Linwei Hu’s University of Michigan biography describes interpretable-ML research and an internal Python validation library in Corporate Model Risk; a later Wells Fargo patent with Ke Wang covers structured SHAP computation. Former model-risk head Agus Sudjianto appears on Wells Fargo-era patents for automated model-validation workflows and graph features for financial-crime detection, but his current external work must not be presented as a Wells Fargo system. Anuj Dimri’s self-authored banker-productivity post names a customer-summary tool and collaborators Shuyu Ding, Sarthak Sharma, and Balaji Gopalakrishnan; no source equates that tool with Lecue’s “Intelligent Banker Book.” Likewise, Fargo virtual-assistant scale, the GNN fraud demonstration, the network-feature patent, AutoX, Jury-on-Demand, and the transaction foundation model remain separate records unless a source explicitly joins them.

Fidelity’s Earnings2Insights paper discloses GPT-4o writer and specialist-review agents, Alpha Vantage financial inputs, Claude Sonnet 4 judging, a 64-call historical shared task, and human hypothetical-investment and usefulness ratings. A separate ConvFinQA paper uses Mistral-7B, Llama-3.2-1B, and Phi-3 with 4-bit LoRA, supervised fine-tuning, DPO, iterative DPO, and a length penalty. Its preferred/rejected pairs are gold calculations versus erroneous SFT outputs, not human preference rankings; the experiment is not an RLHF pipeline based on human feedback. Neither paper establishes Fidelity analyst adoption, autonomous investment authority, live returns, or production use.

The papers expose a recurring Fidelity AI Center research cohort rather than a disclosed org chart. Preethi Raghavan’s public biography describes a vice-president-level NLP/ML remit and earlier IBM and MIT–IBM AI Lab work; Parag Pravin Dakle’s personal research record recurs across the financial-NLP, hybrid-QA, optimization, and agent papers; Sai Krishna Rallabandi, Alolika Gon, Nikhil Kohli, Sihan Zha, and other collaborators recur in different combinations. The public sequence includes multilingual ESG classification, natural-language repair of text-to-SQL parses, BlendSQL, multilingual ESG-impact classification, and sentiment self-training. BlendSQL’s public implementation makes its hybrid structured/unstructured query design inspectable; NER4OPT, MABWiser, Mab2Rec, and Seq2Pat expose adjacent optimization, bandit, recommendation, and pattern-mining work. Those artifacts establish sustained public research and code output. They do not prove one shared runtime, common reporting lines, analyst adoption, portfolio use, or a path from shared-task scores to investment returns. FCAT, Fidelity Labs, Fidelity Institutional AI, the AI Center, and enterprise AI-risk roles remain distinct labels unless a primary source connects them.

The identity-recovery pass also prevents an unrelated attribution. Milind Sunil Shah is Brainbase Labs Founding MTS, not a Snowflake employee. Pipelines remains a separate company under Piper Cyterski; no primary source supports a rename, acquisition, merger, or product transfer into Brainbase. Brainbase publishes agent, task, sandbox, harness, tool, memory, secret, deployment, log, and evaluation-verdict surfaces. Its later Domino’s case study supersedes the conference abstract’s endpoint figures: it reports 11 versions, manually reviewed accuracy moving from 41% to 97%, cost per source from $2.11 to $1.01, and an average 79-minute run against a manual baseline of at least two hours. The page also says some runs lacked corrected workbooks and scope changed between versions. It remains a Brainbase-hosted case study without a released task set, item-level labels, assessor-agreement measure, complete traces, code, or independent reproduction.

OpenProse exposes a different architecture: structured Markdown contracts compile into a dependency graph, while fingerprint comparisons determine wake, skip, commit, and propagation; model sessions still generate the content. Its repository, runtime-contract release, Rust conformance tooling, and Raymond Weitekamp’s cross-framework evaluation repository make more of the control surface inspectable. The project also documents limits: receipt-chain verification does not yet bind published world-model bytes cryptographically; receipts lack a complete authenticated actor/time contract; declared CLI and MCP dependencies do not prove executable or service identity; host authorization remains external; and the runtime-contract-1 to runtime-contract-2 transition does not migrate prior ledgers or state. The scheduled large-bank workflow remains anonymized and future conference copy, with no bank confirmation, dataset, permission map, evaluation result, or production denominator.

The surrounding adoption evidence is also date-sensitive. OpenClaw first bundled an OpenProse plugin, but public issues record model-selection and asynchronous-subagent mismatches; OpenClaw removed the plugin and /prose command in August 2026 and now directs users to the upstream skill. A community fork preserves the older runtime, but it is not the current official distribution. Barrett’s Reactor account attributes a $10,000 hosted pilot to an unnamed university medical-research lab, while the public grant-finder repository shows a technically matching workflow. No university or lab confirmation, named customer, procurement record, or verified expansion was found. OpenProse Research’s classifier experiment supplies a bounded internal evaluation with disclosed cohort and retained-record limitations; it is not a finance, bank, or medical deployment benchmark.

August 29 title-blind W&B channel recovery: public research-agent vocabulary

The Weights & Biases YouTube channel was searched independently of podcast titles and the BrightTALK archive. Its large-scale agentic quant-research recording shows a vendor demonstration that routes a market event to macro, historical-analogs, sentiment, and quant agents before a synthesis forecast; the recording also shows trace-level cost/latency visibility, outcome evaluation, and an LLM-controlled agent-weight optimization loop. Related captioned routes cover LG Exaone Deep market forecasting, CoreWeave ARIA autoresearch, and NVIDIA/W&B agent evaluation. These are vendor demonstrations and product descriptions. They are useful for tracking public research-agent and evaluation patterns, but they do not show that GMO, Acadian, Arrowstreet, or another named manager uses them. They do not establish model weights, training data, data rights, permissions, live portfolio authority, or performance. See the caption recovery note.

---

September 1 Fasanara Capital: a named AI-Lab appointment spanning credit, trading, and operations

Fasanara’s July 21, 2026 first-party announcement names James Hylands as Head of Artificial Intelligence and says he joined from Man Group, where he led generative-AI strategy and delivery for the Discretionary division. Fasanara says Hylands will lead a new AI Lab intended to span investment decision-making, portfolio management, lending operations, risk management, and enterprise productivity. The release lists AI-enhanced credit underwriting, multi-asset trading and portfolio optimisation, autonomous investment research and market intelligence, predictive risk monitoring, and enterprise AI agents as intended workstreams. It also says proprietary systems will be trained on Fasanara’s data ecosystem and that the lab will work with AI companies, universities, and technology partners.

This is unusually explicit public scope for a credit-and-digital-assets manager, but it remains a firm announcement. It does not disclose the lab’s staff beyond Hylands, model or provider inventory, training corpus, data rights, evaluation protocol, agent permissions, production endpoints, or AI-attributed performance. The release reports approximately $6 billion AUM, 130 employees, and 141 fintech lenders integrated from more than 60 countries; those are company-reported scale figures without a published measurement method. Fasanara’s team page shows quant, technology, data, risk, and operations roles but does not enumerate the AI Lab. The capture note keeps the appointment, former-employer, scale, and implementation boundaries separate. This does not establish a relative position against Man Group or any other manager.

Fasanara’s August 12, 2026 quant-division announcement also names Dr Darran Specter, CFA as Head of Quant Equity. The firm says he works with Vinicius Karam, Managing Director and Head of Fasanara Quant, on Open Quant, with responsibility for identifying and evaluating external quantitative- equity managers globally, portfolio construction, and the investment process. The release gives Specter’s prior Brevan Howard and Abu Dhabi Investment Authority roles and an Imperial College London PhD/MEng lineage. This expands the personnel and manager-selection map, but does not establish an AI system, automated selection process, or a connection between Open Quant and the newly announced AI Lab. It is retained in the capture note as a separate remit.

The same expansion exposes a company-posted Quant Technology Lead vacancy for Open Quant. A contemporaneous job-description mirror describes a cloud-native platform built around a central position ledger, portfolio and cross-manager risk analytics, security-master and pricing data, broker/custodian/administrator integrations, resilient data pipelines, APIs, and production monitoring. It assigns the role to practical AI and automation across software-development and investment workflows while requiring controls, auditability, and collaboration across PM, research, risk, operations, and technology. The company post confirms the vacancy; the detailed description is secondary hiring evidence and does not prove that the role was filled or that each listed component is live.

The vacancy also describes Open Quant as a shared-capital and shared-infrastructure platform for independent quantitative teams, with common execution, risk, and operations support. That points to a platform-normalisation problem—position ledgers, broker and administrator integrations, reproducible analytics, and cross-manager monitoring—rather than revealing an alpha-model architecture. The role is a hiring signal, not proof that the platform or its AI workflows are fully deployed.

Executive Summary

• GMO’s public record spans the investment thesis and a named research initiative. Public material describes AI through valuation, quality, capital intensity, systematic signals, and research augmentation. An August 2026 Systematic Equity page names “Super Analyst” as a current agenda item that uses GenAI to add qualitative depth to quantitative breadth. The public record still does not identify a proprietary LLM platform, agent fleet, or GenAI engineering leader.
• Acadian is exposing the investment workflow. Its public materials connect AI modules to text, earnings, analyst-bias correction, expected returns, portfolio construction, and execution. A current Vice President, Investment AI Engineer role adds a direct agentic hiring signal: reusable investment skills, sub-agents, controlled execution, testing, and adoption metrics.
• Arrowstreet’s public record focuses on an enterprise control plane. Its Senior AI Platform Engineer role describes Bedrock-based model access, routing, MCP gateways, RAG, agent frameworks, cost telemetry, and default-deny execution controls. It says little publicly about LLMs creating alpha.
• Man AHL’s public record focuses on agentic research workflows. Its 2025 and 2026 Man Institute articles describe Alpha Assistant and AlphaTrend, including proprietary research context, internal tools, approval steps, signal generation, and evaluation of trend-following ideas.
• CFM’s public record spans AI research, model training, and production prediction infrastructure. Its public material names a narrow financial NER fine-tune, a new internal AI effort, LLM work on text data, and AI agents connected to code generation and predictive-model services.
• Balyasny’s public record describes a centralized Applied AI program with local workflow customization. The OpenAI customer case study names model evaluation, internal benchmarks, scoped tools, agent orchestration, compliance guardrails, named use cases, and firm-reported adoption; other public profiles add market-event forecasting and internal-assistant signals.
• Bridgewater’s newly surfaced Interrupt 2026 keynote exposes PAT’s operating details. The official recording describes an internally deployed research analyst used by hundreds of investors, per-user data entitlements, millions of documents, tens of millions of time series, LangGraph orchestration, deterministic Python execution, parallel sub-agents, human-audited benchmarks, and a context-and-harness pull-request loop. PAT is explicitly described as exploratory research tooling, not a trading system; broader artificial-investor claims remain separate and firm-reported.
• BlackRock adds a large traditional asset-manager comparison point. Its public record separates BlackRock AI Labs, long-running systematic alpha-model work, a proprietary equities research tool called Asimov, and current LLM-based macro-research material. Current AI Labs leadership and hiring are public; system-level ownership and attribution are not.
• WorldQuant exposes a detailed agentic-portfolio-management hiring specification. A 2026 role artifact describes a live trading book, planning, tool use, memory, reflection, collaborative reasoning, reinforcement-learning tuning, custom agentic workflows, and human-in-the-loop checks. The direct role URL no longer resolves as a live vacancy on the August 10 check, so it is retained as dated hiring evidence—not current headcount or deployment proof.
• Numerai adds a crowdsourced hedge-fund control case. Its public record now includes an AI-scientist homepage workflow, an 8-billion-parameter Predictive LLM for a released feature set, open-source agent skills, MCP access for tournament research and submissions, and an official careers-host AI Scientist role. This is a public platform-and-tournament signal, not evidence that any agent has autonomous authority over firm capital.
• The extended peer pass separates public disclosure layers. Point72/Cubist and G-Research expose current AI hiring and runtime signals; Schonfeld and Millennium describe internal platforms and adoption controls; Two Sigma exposes LLM-assisted feature research; Citadel separates research assistance from systematic ML; the remaining reviewed firms primarily expose predictive-ML, careers, or governance material.
• The source-expansion pass found several new disclosure surfaces. Official investor-relations archives, vendor customer pages, conference previews, technical-society pages, current leadership pages, company LinkedIn feeds, and firm-hosted careers APIs add temporal signals that the wire sweep cannot provide. New material includes Millennium’s AI Lab, Two Sigma’s current AI titles, CFM’s multimodal AI-lab language, Walleye’s Claude Code and model-testing disclosures, Acadian’s July 2026 staffing releases, XTX’s XTY Labs foundation-model hiring plus ML acceleration roles, Longqi’s current mainland/Hong Kong official page, and RQI Investors’ Australian long-short AI/ML release.
• The expanded podcast/video pass adds date-scoped practitioner evidence. HRT, Jane Street, Balyasny, Numerai, Versor, Point72, Two Sigma, CFM, Voleon, and Susquehanna now have source-reviewed media or firm-published video items in the research matrix. These sources add workflow, infrastructure, personnel, and research-culture detail; they do not establish relative capability, return attribution, or autonomous investment authority.
• Model-use evidence is more specific in several firms, but remains bounded. Man reports comparing Claude 4 Sonnet and GPT-5 for AlphaTrend signal proposals; Balyasny reports evaluation dimensions and task-by-task selection between GPT-5.4 and internal models; CFM reports financial-NER fine-tuning and financial-news retraining; Two Sigma reports post-training and leakage/overfitting concerns in current technical hiring and outlook material. None of these disclosures supplies a complete model inventory or independent performance audit.
• AI Engineer NYC turns an anonymous finance preview into a dated personnel and systems queue. The September 20 schedule confirms speakers from Bridgewater, Coatue, Point72, Citadel Securities, Apollo, Wells Fargo, Vanguard, JPMorgan Chase, Capital One, Fidelity, Franklin Templeton, Bloomberg, New York Life, Contour, and Arena Investors. The sessions are proposed public disclosures, not completed talks or audited deployment evidence.
• The first 81-person conference audit was a selected cohort, not the full roster. An all-109-speaker comparison recovered 28 omissions, including Coatue’s Frank Long, Vals AI’s finance benchmark, Browserbase’s Ramp and Parcha cases, Speakeasy’s permission control plane, Oracle’s governed-memory examples, and several software-factory control patterns. This is a methodology correction as well as a research finding.
• Coatue’s public record now connects the Brain, named analyst agents, and a current model provider without collapsing their ownership. A 2024 SEC filing dates the Brain to 2023; the Q2 2026 material names Meeting, Earnings, Brainstorm, and Expert Agents; and a February 2026 firm post reports Claude use for research, idea generation, data-analysis agents, and earnings-transcript skills. Frank Long joined in January 2026, so his current Head of AI & Partner title does not establish authorship of the pre-existing platform. Coatue reserves investment decisions to people and does not publish the agents’ evaluations, permissions, use counts, incidents, or investment effects.
• These public records address different questions. Acadian describes AI inside investment research; Arrowstreet describes GenAI platform architecture; GMO describes AI investment judgment; Man AHL describes agentic strategy-design workflows.
• A published academic baseline adds an industry-level adoption proxy. Sheng, Sun, Yang, and Zhang’s 2026 Review of Financial Studies paper infers GenAI reliance from ChatGPT-generated earnings-call signals and 13F portfolio changes, and reports version-specific adoption and performance associations. Morningstar’s 2026 public report page supplies a separate proprietary cross-check across approximately 3,200 strategies and manager interviews. Neither source identifies a deployed model or establishes the AI architecture of a named firm in this article.
• A recurring non-disclosure is consistent: none of these public sources gives an outside reader a complete model inventory, evaluation record, live-user count, production-stage map, or audited return attribution. Public signals are architectural and organizational, not a transparent view of the alpha engine.

August 28 academic and industry baseline: inferred GenAI reliance is not deployment evidence

The published Review of Financial Studies article by Sheng, Sun, Yang, and Zhang adds an industry-level empirical baseline to the firm-specific source map. The authors construct a “GenAI Reliance” measure from the incremental contribution of 14 ChatGPT-generated scores based on earnings-call transcripts to changes in quarterly 13F portfolios, alongside conventional fundamental and market controls. Their August 2025 full-text version reports 644 unique hedge-fund companies and 11,921 company-quarter observations from 2016 Q1 through 2024 Q3, and estimates inferred adoption of approximately 21% in 2022, above 40% in 2023, and close to 60% in 2024 after an estimated false-positive adjustment. These are authors’ estimates from an econometric proxy, not a census of named firms’ tools or deployments. See the method and version note.

The same paper’s published abstract and August 2025 version report a 2–4 percentage-point annualized abnormal-return association for an interquartile increase in GenAI Reliance in post-ChatGPT tests, with stronger relationships in subsamples containing pre-existing AI-skilled human capital and firm-policy or firm-performance information. The paper also reports differences between trades aligned with and contrary to generated signals. The article records these as paper-reported statistical estimates, not as evidence about GMO, Acadian, Arrowstreet, or any other named manager; 13F data also omit short positions, derivatives, and other non-reportable activity. Earlier conference versions report different estimate ranges, so the version and date must accompany any quotation.

The Morningstar 2026 report page supplies a separate, proprietary industry-context route. Its public description says the report covers approximately 3,200 strategies globally and interviews with asset managers, and describes GenAI effects as primarily visible in research while investment decisions remain a frontier. The page names BlackRock, Goldman Sachs, and Vanguard as examples but does not publish the complete strategy universe, interview roster, coding rules, or underlying observations. This is a qualitative cross-check, not a reproducible firm-level adoption estimate or a ranking.

The companion Morningstar Beyond the Stars episode, “Alpha Courtesy of AI”, published June 23, 2026, adds a distinct observer-method and workflow layer. The guests describe a global manager-research function of approximately 130 analysts, an internal historical database of rating notes, and a semantic-search/AI-tool route used to analyze those notes. They report reviewing approximately 3,200 strategies and finding AI integration in about 20 firms, but do not publish the firm list, coding rubric, or underlying notes. The timestamped capture note records the Apple, Spotify, RSS, and local-transcript routes.

The episode also reports several bounded examples: an asset-manager CounterPoint team experimenting with more than 50 persona-based agents; an unnamed manager estimating that an AI assistant handles about 70% of initial company research; and a claim that an agentic system reduced one unnamed academic-paper replication task from an intern-scale project to minutes. The speakers identify nondeterminism, look-ahead bias, and expert oversight as barriers to using generative AI as an autonomous portfolio decision-maker, and keep accountability with human portfolio managers. These are attributed speaker accounts, not evidence of a named hedge fund’s model, data rights, permissions, production deployment, or performance.

August 22 expansion: stack, permissions, partners, and modalities

The second-pass research changes the unit of analysis from a firm’s AI slogan to its publicly visible lifecycle: data access, model development, evaluation, agent permissions, human review, deployment, and monitoring. The new raw notes are stack dossiers for Acadian, Man AHL, Balyasny, Two Sigma, and WorldQuant, stack dossiers for GMO, Arrowstreet, CFM, Jane Street, and Bridgewater, personnel and academic lineage group A, personnel and academic lineage group B, the agent-permission matrix, and the partnership and data-provenance sweep.

Partnerships expose the infrastructure layer

New public relationship evidence includes HRT’s announced CoreWeave/NVIDIA research-platform agreement, IMC’s CoreWeave expansion, Flow Traders’ stated selection of CoreWeave for foundation-model training, Millennium’s Google Cloud relationship, Tower Research Capital’s IIT Bombay machine-learning-for-social- good lab, and DRW Foundation’s Caltech mathematics-and-AI research partnership. These sources expose compute, hybrid connectivity, training scale, educational and research collaboration, or data/agent workflows. They do not disclose training corpora, model weights, permissions, evaluation fixtures, live trading authority, or AI-attributed performance. See the HRT announcement, IMC announcement, Flow Traders announcement, Millennium–Google Cloud release, IIT Bombay collaboration, and Caltech–DRW announcement.

The counterparty sweep also found a broader vendor-side development. LSEG, Macrobond, Morningstar, Third Bridge, MSCI, Maybern, FactSet, and M Science are publicly describing agent interfaces for licensed financial data, expert transcripts, macroeconomic series, private-market diligence, or portfolio operations. Recurring controls include entitlement inheritance, provenance, revision-aware data, deterministic calculations, as-of dates, and audit trails. These are vendor product disclosures, not evidence of named hedge-fund deployment. Examples include LSEG MCP, Macrobond’s AI Data Feed, Third Bridge MCP, and FactSet–Google Cloud.

The title-blind pass also recovered FinTech Focus TV’s August 17, 2026 TradingTech Summit NYC episode, whose publisher names Diane Levich (HPR), Joshua Carroll (Solace), Martyn Snook (Genesis Global), and Michael Riebling (Adaptive). The vendor panel describes a production boundary relevant to investment-technology diligence: eventing and stream processing can filter or summarize high-volume inputs before an agent layer, reducing the amount of historical event data sent to a model (25:04–25:48). It also links rapid Claude-assisted prototyping to auditability, scalability, compliance, data quality, and in-house review requirements (12:00–12:54, 20:58–22:48). This is vendor/event evidence, not evidence that a named hedge fund uses the architecture; the episode discloses no customer identity, model evaluation, permissions, investment authority, or performance. See the timestamped capture note.

The same archive surfaced Genesis Global’s August 5, 2026 “What Happens When AI Stops Being an Experiment” episode, featuring President and Chief Product Officer Tej Sidhu and Head of Growth and Sales Strategy David Perkins. Genesis describes the bottleneck as moving experiments into governed, scalable capital-markets production (03:35–04:12), and describes an LLM working inside a prebuilt application and microservice container rather than generating an entire regulated system from scratch (13:33–15:03). The speakers also frame human-provided structure and use-case prioritization as controls on agentic work (22:56–23:52, 27:14–28:14). This is vendor self-report and adjacent-technology evidence, not proof of a tracked hedge fund’s Genesis deployment, model choice, data rights, evaluation, permissions, or investment authority. See the timestamped capture note.

A separate AIAgentStore episode, published August 4, 2026, adds a title-blind vendor-media route into the research-agent control surface. Its proposed workflow requires correct filing period and accounting definition, claim-level evidence links, model updates, point-in-time consensus comparison, contradiction and restatement checks, preserved prompts/model versions/approvals, and controls around material non-public information (00:00–03:36). The episode then separates document retrieval, deterministic calculation, model population, work-product generation, auditability, and governance, and proposes measuring retrieval errors, consensus variance, reviewer correction time, unsupported claims, and as-filed versus restated history (20:25–27:18). This is a vendor ecosystem taxonomy, not a named hedge fund disclosure or independent product comparison; the publisher’s product descriptions and reported figures do not establish customer adoption, model accuracy, permissions, investment authority, or performance. See the capture note.

The title-blind pass then recovered a May 5, 2026 Zeus Capital interview with Alistair Smallwood, identified by the publisher as Head of Applied AI at Primer and as having previously moved through sell-side roles to technology-focused hedge fund Lynott Partners. Smallwood describes an agent-oriented equity-research workflow built around source and company-document ingestion, user-specific rules, retained context, model updates, and analyst correction loops (05:49–10:36; 20:00–30:45). He also describes a roadmap involving alternative data, monitoring, sell-side research, and expert networks, while keeping judgment and skeptical review with the analyst (38:14–43:53). Primer’s first-party essays and product page corroborate the public positioning. This is named practitioner and vendor-product evidence, not proof of Lynott or customer deployment, data permissions, model inventory, autonomous authority, or performance; the capture note keeps those boundaries explicit.

The linked Primer author archive adds five separate dated routes. Smallwood defines prospective autonomous equity analysis around source selection, tools, model/thesis updates, continuous checking, and human escalation in a May 15 post; reports an explicitly nonrepresentative field note from more than 50 conversations in a May 21 post; proposes logging the research trajectory so process quality can be replayed in The Fundamental Backtest; and publishes vendor-run benchmark methodology and answer-key critiques in the BigFinanceBench report and the FrontierFinance audit. These posts expand the evaluation vocabulary—point-in-time data access, repeated stochastic runs, versioned keys, source lineage, rejected alternatives, analyst overrides, and process logs—but their adoption observations, product scores, and benchmark-defect counts remain author/vendor reports rather than independent comparisons or evidence of named-fund deployment. See the archive capture note.

The archive also contains a distinct July 21, 2026 annual-report agent experiment. It reports approximately 290 runs on a fixed historical task: six specialist roles repeatedly surfaced a staff-number issue, but a boss agent repeatedly discarded it when comparing short reports. A later “champion” pass developed each finding into a full argument before a judge selected among them; the author reports 12/13 final hits in that setup. This is useful evidence about a proposed evaluation and orchestration pattern—preserve rare candidates, develop them before selection, and record rejected alternatives—but it remains a self-reported, task-specific experiment. It does not establish any named fund’s production use, model policy, data rights, investment authority, or performance. See the capture note.

Personnel research connects firms to academic lineages

The personnel pass resolved public research routes without treating education as proof of a firm’s production system. Examples include Acadian’s Andy Moniz and his NLP/IR dissertation route, Balyasny’s Peter Anderson and Su J. Wang with public applied-AI research artifacts, Two Sigma’s Ben Wellington and current feature-engineering role, Bridgewater’s Oliver Simon as a current AI/ML investment-strategy leader, Voleon’s dated Simons Institute research-scientist affiliation, Vinva’s first-party research biographies, and Janus Henderson’s Zeyu Gao with an explicitly described agentic-AI quantitative fixed-income role. The underlying sources are Acadian’s Moniz announcement, Balyasny’s MIT event, Two Sigma’s Wellington profile, Bridgewater’s Oliver Simon profile, the Simons Institute programme, Vinva’s team page, and Gao’s Janus Henderson biography.

These links support dated identity, title, education, or research-lineage claims. They do not establish reporting lines beyond the cited page, current model ownership, advisor relationships unless explicitly stated, or a paper’s use in a live portfolio.

Agent permissions are more observable than model quality

Across the 41-firm permission matrix, public evidence is materially richer for what an agent may read, write, execute, backtest, or submit than for model quality. The recurring public controls are scoped data access, role-based permissions, sandboxed code execution, evaluation gates, human review, audit logs, and deterministic calculation layers. A job description remains hiring intent; a vendor case remains vendor/customer evidence; and a production infrastructure claim does not establish that an agent can place orders.

New modality routes worth testing

The modality pass identified several research-backed routes that are distinct from ordinary news sentiment. DeepVoice uses verbal and vocal features from 6,047 S&P 500 earnings-call samples; Engagement in Earnings Conference Calls studies 2,400 S&P 1500 calls and manager–analyst interactions; the Clinical Trial Outcome benchmark covers approximately 125,000 trials; and the CTO dataset and code separates phase progression from regulatory approval. Additional routes include FinDKG for financial knowledge graphs, satellite imagery of container ports for supply-chain activity, and a Renaissance Technologies execution patent.

These are research opportunities, not claims of hedge-fund deployment. The main technical risks are feature-release timing, revision history, label definition, survivorship, weather and cloud contamination in imagery, and the difference between a contemporaneous price association and a tradable holding-period signal.

Small-manager and economics routes

The legal-identity pass surfaced public AI or ML descriptions at PharVision, Runtime Fund, One Eleven Capital, JCube Capital Partners, Aggregate Asset Management, Golden Hen, and Constella. It also separated Noax and QuantOptimus as proprietary/SMA or software-platform structures rather than confirmed pooled hedge funds. These records remain a discovery cohort, not a ranking.

The infrastructure-economics pass found an anonymous WWT case describing an on-premises GPU data center, an IMC ML-engineering role mentioning distributed training and low-latency inference, and an Anthelion infrastructure role mentioning Azure, GPU provisioning, retraining, and batch/real-time inference. The WWT customer is unnamed and the job postings are hiring evidence; neither supports a model-performance or return claim. See the WWT case, IMC role, and Anthelion role.

What the public record can and cannot prove

This is a public-signal audit, not an attempt to obtain confidential information. “Public signal” here means information that firms reveal through legitimate public surfaces: investment letters, regulatory filings, careers pages, job descriptions, podcasts, videos, LinkedIn activity, and conference appearances.

A first-party source that names a workflow, role, system boundary, or control is direct evidence for that specific claim. A job description is evidence of hiring intent and desired architecture. It is not proof that the role is filled or that the system is live. A podcast can reveal process detail that a marketing page omits, but it remains a named practitioner’s account. LinkedIn is useful for timing and discovery; it is weaker evidence unless the post links back to a firm-controlled source.

Regulatory filings are useful for entity identity, business scope, risk language, and legal-vehicle checks, but they are not a complete AI observability layer. The SEC describes Form PF as confidential and separately discusses the disclosure and investor-protection questions raised by AI agents. Therefore, a missing AI reference in a public filing is treated as an evidence gap—not as evidence that a manager has no AI capability.

Across firms, public workflow and control descriptions are not public proof of alpha. Any claim that moves from research assistance to investment decision, backtest, execution, monitoring, or compliance would require separate evidence for source provenance, temporal-leakage controls, overfitting checks, data rights, runtime constraints, permissions, audit logs, and kill switches. The reviewed public material does not supply that complete chain for any firm.

Podcast and video cross-check

The new media pass should be read as a second evidence family, not as a ranking layer. HRT and Jane Street provide predictive-market and infrastructure signals; Balyasny, Versor, Numerai, and Point72 provide research-workflow and agent vocabulary; CFM, Two Sigma, and Voleon provide scientific and systematic research context. The expanded practitioner matrix links each item to its canonical page and records what the media does not prove.

The August 19 title-blind pass adds a firm-controlled Goldman Sachs QIS transcript with Osman Ali, a separate Versor founder account from CFI’s FinPod, and a historical Citadel personnel record for Nimit Sohoni linked to current Cartesia AI research. Goldman’s page describes large and small language models, smaller-model fine-tuning for Japanese management-disclosure sentiment, and cross-asset QIS research. The Versor transcript describes a proprietary event database, AI/NLP use since 2018, model research, portfolio construction, risk, simulation, and AI-supported non-investment workflows. The Sohoni episode adds personnel and research-culture context but does not establish a current Citadel role or a Citadel model assignment. These remain attributed media and personnel signals, not a capability ranking or performance claim.

A current Money Maze compilation, published August 20, 2026, adds three title-blind media routes. The publisher identifies Andrew Veglio as former CEO of Vantage Investment Management, Kristian West as Head of Investment Platform at J.P. Morgan Asset Management, and Lucia Soares as CIO and Head of Technology Transformation at Carlyle. Recovered YouTube captions record West describing internal and third-party data foundations, filings, earnings reports, central-bank minutes, web data, roughly 40 years of proprietary research data, and AI-assisted consumption of approximately 7,000 broker reports per day through an agentic framework keyed to portfolio holdings (24:25–26:38). The same recording captures discussion of agent context and employee adoption at Carlyle and a replacement thesis from Veglio. These are attributed speaker accounts and asset-manager/private-market control cases, not evidence of a hedge fund’s model inventory, permissions, autonomous trading, or performance. Automatic captions contain recognition errors; the capture note preserves the boundaries.

A separate Hedge Fund Huddle episode, published August 4, 2026, identifies Vik Bansal as a Systematic Portfolio Manager at Centiva Capital. The recovered timestamped capture records Bansal describing trials of multiple news and alternative-data sources, preference for differentiated properties such as local-language modeling or speed, and review by experienced data engineers; he says AI-generated signals would be treated like a data provider and checked out of sample. He also describes LLM-based sentiment models, AI-assisted code prototyping, and an approximately 40% speed-up in one live-trading engineering process after an unknown bottleneck was identified (03:37–07:26, 24:15–26:12). This is a named practitioner account and the speed figure is not independently audited; the source does not disclose Centiva’s model names, datasets, factor definitions, permissions, or AI-attributed returns.

A separate Odds on Open episode with Versor partners Nishant Gurnani and DeWayne Louis, published November 10, 2025, adds a different firm-controlled media route. The discussion describes alternative data outside tidy tables—text, audio, footfall traffic, and credit-card receipts—plus human-defined research specifications, model evaluation, point-in-time and coverage checks, and a merger-arbitrage framing with completion, failure, and competing-bid outcomes. Versor’s current careers page separately describes 50+ researchers, engineers, and financial professionals across New York and Mumbai and lists a machine-learning quantitative-research role. These sources expose a research philosophy, hiring surface, and media account; they do not disclose a current model inventory, data licenses, permissions, autonomous trade authority, or audited performance. See the capture note.

A cross-check of Versor’s first-party Athenaeum archive surfaced an April 21, 2026 firm-distributed announcement that was not in the prior ledger. It describes a partnership with an unnamed global multi-manager platform for an AI/ML-based Event-Driven strategy covering developed-market corporate events, including competing bids and dynamic exposure management. The release says approximately half of a stated $1 billion capacity had been allocated; that is a self-reported commercial claim, not an independently verified AUM, mandate, or performance result. Its disclosure section also makes AI governance risks visible by naming data quality, confidentiality, copyright/trade-secret, cybersecurity, privacy, and insider-trading risks. The announcement does not identify the partner, name models, disclose permissions, or establish whether the described components are live or experimental. See the release capture note.

The archive also links to a distinct June 4, 2026 Odds on Open follow-up with DeWayne Louis. Recovered captions describe a roughly 26-year hard-catalyst event dataset across North America, Europe, Japan, and Australia; a constructed database combining fundamental, market, news, event, and other alternative data; thousands of features; forecast scores for deal outcomes; and regime-sensitive volatility, correlation, and exposure analysis. This supplies more process detail around the event-driven strategy, but not a model inventory, vendor list, feature definitions, evaluation split, live permissions, or performance attribution. See the timestamped capture note.

Versor’s first-party YouTube playlist, linked from its Q1 2026 LinkedIn recap, returns eight short videos spanning M&A, long/short, alternative risk premia, portable alpha, outlook, and AI. The June 18, 2026 “How AI Is Actually Used in Investing” clip identifies Nishant Gurnani in its description and links the longer Hedge Fund Huddle episode. Its compressed framing describes agents as junior-researcher aids and stresses disciplined process integration; automated captions misidentify the guest and are not used for the name. This adds a firm-controlled distribution surface, not a new model or performance disclosure. See the playlist capture note.

The same first-party playlist had seven clips previously listed as metadata-only. A caption-recovery pass now supplies public automatic English timestamp layers for the entire set. The M&A clip links AI-related deal activity with data-center and infrastructure investment (00:39–00:54) and presents systematic event investing as selective risk-reward analysis (01:17–01:26). The first outlook clip briefly mentions AI-driven deal-risk assessment (01:03–01:18); the two transaction clips describe deal structure, bidding behavior, break dynamics, regulatory scrutiny, and pre-existing holdings as review dimensions. The portable-alpha, alternative-risk-premia, and long/short clips add portfolio, liquidity, common-exposure, stress, and index-level context. This is useful evidence about Versor’s public research packaging and vocabulary, but not about model architecture, data rights, deployment stage, permission boundaries, or outcomes. See the updated capture note.

A separate January 15, 2026 Odds on Open fireside conversation with Deepak Gurnani, linked from Versor’s Q1 recap, adds a founder-level historical account. Gurnani describes cloud adoption during the 2013 launch period, alternative-data expansion after a 2017 conference, AI analysis of earnings-call transcripts into quantitative scores, and use cases involving credit-card data, satellite imagery, and weather patterns. He describes AI/ML as developing alongside alternative data and cloud infrastructure and says expanding those methods remains a focus. These are attributed historical statements; the episode does not disclose the model, labels, data rights, evaluation record, permissions, or performance attribution. See the capture note.

The personnel/media sweep also surfaced a Middlemark Partners webinar announcement naming Deepak Gurnani, Luke Hinshelwood, and Angelo Calvello. The announcement describes hypothesis-driven AI frameworks, generative AI, explainability, domain knowledge, and AI as a tool versus the starting point; Versor’s later Q1 recap labels the route “The Great M&A Reset & Role of AI in Equity Event Investing.” The event year is unresolved, the direct page returned 403, and no recording or transcript was recovered. This is a verified discovery lead, not evidence of what the speakers ultimately said or implemented. See the recovery note.

The queue also surfaced a J.P. Morgan Market Matters episode with allocator Pierre Chabran, recorded April 24, 2023. Chabran describes a quantitative process broken into data management, idea generation and validation, production, execution, risk, and monitoring; stresses point-in-time availability and revision controls; and describes ML use in execution, data cleaning, and nonlinear alpha combination. He distinguishes open-source algorithms such as PyTorch from bespoke tools for noisy financial data. This is allocator-side, date-scoped context rather than evidence about any particular hedge fund’s current stack or permissions. See the capture note.

The current The Derivative feed also adds three title-blind records. Standpoint Asset Management founder and CIO Eric Crittenden describes personal use of Claude, ChatGPT, and Grok for external research and data analysis, including comparing model outputs (46:31–48:39). The interview does not establish that these tools enter Standpoint’s portfolio process, and the publisher labels the transcript as automatically compiled. See the capture note.

Center15 Capital founder Ian Winer discusses AI-driven analysis paired with space-based sensing, then names Shield AI, Epirus, Shift5, and HawkEye 360 as companies Center15 discusses or backs (37:32–40:56; 46:54–48:49). This is defense-tech and alternative-data investment evidence, not a disclosure of Center15’s internal financial models. See the capture note.

OneRiver’s Patrick Kazley interview covers long volatility, convexity, trend following, systematic macro, and QIS integration. It contains no AI or GenAI disclosure and is retained as a negative control: systematic-investing language alone is not treated as evidence of AI use. See the capture note.

The Risk.net Quantcast interview with Pietro Rossi adds a separate finance-ML route. Rossi describes Prometeia work originating from an insurance-company request for credit-rating-transition scenarios and portfolio-risk simulation (01:40–03:50); he also describes neural networks learning SPX and VIX volatility quantities to reduce calibration cost relative to Monte Carlo-heavy workflows (15:13–20:30). The linked credit paper and neural-calibration paper provide the formal research route. This is consulting and academic evidence, not proof of hedge-fund use or investment performance. See the capture note.

The Risk.net Quantcast interview with Eduardo Abi-Jaber and Shaun Li adds a separate derivatives-research lineage. Li is introduced as a Morgan Stanley exotic-equity-derivatives quant strategist, but says the work was completed during his PhD before joining Morgan Stanley and is not a representation of Morgan Stanley (01:46–02:05). Abi-Jaber says the PhD was sponsored by BNP Paribas and developed with BNP Paribas quant Camille Illand (02:49–03:20). The discussion describes a two-factor quintic Ornstein–Uhlenbeck model for jointly calibrating SPX and VIX volatility surfaces and the skew-stickiness ratio, with implications for Greeks, P&L interpretation, and hedging (05:39–07:45; 08:29–10:46). The paired paper page provides the formal model description. This is a research and career-lineage signal, not evidence of Morgan Stanley deployment, hedge-fund use, or performance. See the capture note.

The Risk.net Quantcast episode with Barclays XVA quants Ben Burnett and Benjamin Piau adds a bank-model-risk and production-boundary route. The speakers describe a framework for estimating the effect of simplifying stochastic XVA models and for decomposing model, discounting, payout, and meta-adjustment components (01:34–04:18). Piau says the framework is used in practice at Barclays (25:03–25:20); Risk.net separately describes it as in production for Barclays’ XVA calculations. Later, Burnett describes AI/ML as an emerging question for XVA, while Piau characterizes applications to XVA theory as ongoing research (35:06–37:42). The paired paper supplies the formal route. This is bank-reported model-use and research-direction evidence, not proof of an AI deployment, hedge-fund use, or investment performance. See the capture note.

The same title-blind archive screen recovered a Lipton–López de Prado interview that the filtered RSS search omitted. The discussion covers illiquidity-aware private-equity valuation, two-stage allocation, uncertainty ranges, tokenization, and a forward-looking quantum-optimization possibility (02:42–05:09; 21:45–24:28; 30:42–31:57). No AI or GenAI use is disclosed. The record is retained as a negative control and methodology/lineage route, not as evidence of AI adoption by ADIA, Adia Lab, or a hedge fund. See the capture note.

Risk.net’s paired agentic-tools feature and quant-jobs feature add an anonymous practitioner route. The publisher says the two June 2026 features draw on interviews with 13 senior quants, but withholds identities because of topic sensitivity; the publisher index names Claude Code and Aristotle as tools in the discussion. This establishes a public source family and vocabulary, not firm or personnel attribution, model ownership, deployment, permissions, or performance. See the capture note.

New source note: title-blind finance-media records — Goldman QIS, former Citadel AI researcher, and Versor

Regulatory boundary sources: SEC private-fund adviser overview, Form PF FAQ, and SEC statement on AI and investment management.

Source files: sources/06-industry-verticals/gmo-acadian-arrowstreet-ai-public-signals-2026-raw.md; sources/06-industry-verticals/hedge-fund-ai-source-expansion-2026-08-14-raw.md; sources/06-industry-verticals/hedge-fund-ai-labs-academic-lineage-2026-08-14-raw.md; sources/06-industry-verticals/hedge-fund-ai-leadership-resolution-2026-08-14-raw.md; sources/06-industry-verticals/hedge-fund-ai-conference-video-mining-2026-08-14-raw.md; sources/06-industry-verticals/hedge-fund-ai-provider-data-partner-registry-2026-08-14-raw.md; sources/06-industry-verticals/hedge-fund-ai-role-history-2026-08-14-raw.md; sources/06-industry-verticals/hedge-fund-ai-quiet-firm-artifacts-2026-08-14-raw.md; sources/06-industry-verticals/asia-acadian-scale-ai-public-signals-2026-08-14-raw.md; sources/06-industry-verticals/hedge-fund-league-tables-manager-directories-2026-raw.md; research/06-industry-verticals/tc43-emerging-quant-manager-intake-2026.md; research/13-multimodal-sources/fear-greed-interviews/ (named-practitioner source); research/06-industry-verticals/hedge-fund-research-signal-landscape-2026.md; research/06-industry-verticals/hedge-fund-domain-model-agentic-capability-audit-2026.md; sources/06-industry-verticals/hedge-fund-ai-worker-layer-linkedin-2026-raw.md; research/06-industry-verticals/hedge-fund-ai-personnel-title-sweep-2026.md; sources/06-industry-verticals/hedge-fund-ai-personnel-title-sweep-2026-raw.md; sources/13-multimodal-sources/robeco-quant-street-archive-capture-2026-08-20-raw.md; sources/13-multimodal-sources/blackrock-systematic-archive-episode-capture-2026-08-20-raw.md; sources/13-multimodal-sources/numercon-replay-episode-capture-2026-08-20-raw.md; sources/13-multimodal-sources/acadian-systematic-methods-api-expansion-2026-08-20-raw.md; sources/13-multimodal-sources/marshall-wace-paul-marshall-nbim-ai-tops-2026-raw.md; sources/13-multimodal-sources/hrt-marc-khoury-dwarkesh-order-book-signal-2025-raw.md; sources/13-multimodal-sources/israel-uae-titleblind-ai-manager-routes-2026-09-03-raw.md; sources/13-multimodal-sources/russia-indonesia-ai-fund-product-routes-2026-09-03-raw.md.

Latest regional source note: sources/13-multimodal-sources/swiss-french-titleblind-ml-allocation-routes-2026-09-03-raw.md

Newly added source family: conferences, keynotes, and vendor showcases

The research pass now treats conference archives and vendor-hosted customer sessions as a separate discovery layer. LangChain’s Interrupt 2026 archive exposes 23 sessions, including Bridgewater’s PAT presentation. Similar public surfaces include Anthropic’s financial-services agents page, OpenAI’s Balyasny case study, AWS’s Bridgewater customer video, Databricks sessions naming Ares and Waterfall Asset Management, and Bloomberg’s investment-research panel. These sources can reveal speakers, system names, architecture vocabulary, and deployment narratives that do not appear in formal reports.

The title-blind conference sweep also located AI Trader 2026, an inaugural San Francisco practitioner event advertised for August 27, 2026 around agentic trading. Its organizer lists tracks for equities alpha research, macro, options, market making, and prediction markets, and solicits demonstrations of production architectures, evaluations/backtests, risk controls, kill switches, and unattended deployment. This is an organizer-level discovery route, not evidence that a named hedge fund attended or that any advertised live system or P&L claim is verified. No firm is assigned from the event page alone; the capture note records the follow-up targets. The Waterfall route now has additional detail in the current Databricks listing. It names Shehzad Nabi as CTO and describes a Prophecy-built specialized Claude Code harmonization agent for inconsistent loan tapes, with users reviewing confidence, reasoning, and lineage before final validation. The listing reports 93% first-pass accuracy in production, approaching 100%; that is a vendor/customer claim with no public denominator, task definition, split, independent audit, or investment attribution. The June 21–24, 2027 date means this remains an event-listing route until a recording or slides are recovered. See the capture note. The title-blind archive sweep added firm-controlled media indexes that should be mined episode by episode: Citadel’s EQR archive, Jane Street’s Tech Talks and Real Numbers, Bridgewater’s media archive and AIA Labs video page, BlackRock Systematic’s podcast archive, Numerai’s NumerCon replay index, Robeco’s 11-episode Quant Street archive, and CFM’s Cumulus strategy page. Archive existence is a discovery signal only; the public record must still establish the episode, speaker, date, transcript, and evidence boundary before any claim is promoted.

The evidence boundary is important. A recording is primary evidence for what a presenter said; a vendor customer story is evidence of a vendor-reported or customer-provided claim; an event catalog is only a lead until the session content is retrieved. None of these surfaces, by itself, proves model ownership, current employment, production permissions, investment authority, or performance. Transcript mirrors and automated summaries are used to locate timestamps and candidates, then checked against the original recording or firm-controlled page.

A new title-blind pass adds a particularly concrete J.P. Morgan Asset Management model-inventory route. Its U.S. Value Strategy presentation, marked 4Q 2025 with data as of December 31, 2025, names Analyst Moneyball for machine-learning analysis of decades of fundamental research and analyst bias, Trader Moneyball for equity-trading optimization, SpectrumIQ for searching proprietary research, meeting transcripts, and broker notes, and IRIS, CAS, and EFM for outcome forecasting, analyst sentiment, and financial-distress signals. It also labels Smart Monitor as a pilot and PM Moneyball as underway/in development. The same presentation says JPMorgan Asset Management uses LLMs for internal operational scalability but does not rely on them for the portfolio manager’s final investment decision. This is firm-published tool and control evidence; it does not identify model weights, vendors, evaluation denominators, permissions, or investment contribution. See the capture note.

The Ask D.A.V.I.D. presentation identifies two JPMorgan Chase team members and describes a supervisor agent routing structured-data, unstructured RAG, and analytics agents backed by proprietary models and APIs. It also describes role-based personalization, short- and long-term memory, an LLM reflection/judge node, retries, text-to-code for complex queries, evaluation-driven development, and human supervision for higher-risk execution. The public mirror includes language about production operation and multiple models, but does not establish deployment scale, current providers, training data, permissions, or portfolio authority. It is a named bank architecture route, not proof of autonomous investment decisions or a hedge-fund system. The same capture note keeps the evidence classes separate.

The NVIDIA GTC 2026 session identifies Jianchi “JC” Chen as CIO of Beijing-based quantitative hedge fund Kendall Square Capital, with an office in Shanghai. Chen says the firm uses end-to-end deep learning and multi-period alpha strategies for Chinese A-shares and describes a GPU-accelerated portfolio optimizer built with NVIDIA technology. The session names CUDA, CUDA-X, cuDF, cuML, and cuOpt and frames optimization as part of the loop from prediction and risk modelling through constrained portfolio construction and execution. Chen’s reported RMB 12 billion pre-leverage and annual doubling of AUM and daily volume since 2023 are speaker-reported, not independently audited. The source does not disclose weights, training corpus, data rights, validation design, order permissions, or AI-attributed returns.

Two controls help calibrate the wider search. An AWS reference workflow describes a vendor architecture using market data, SEC EDGAR, PDFs, web/news, S3, Redshift, SageMaker Lakehouse, AWS Batch, Step Functions, and Bedrock, but names no customer fund. The QuantMinds International 2026 roster lists November 16–19 London speakers from Verition, Schonfeld, Jain Global, Crabel, and other buy-side and academic organizations, alongside Quant Tech and AI, Machine Learning in Finance, and Quant Invest streams. The AWS page is vendor evidence; the QuantMinds page is event metadata. Neither establishes a named firm’s model or deployment. See the capture note.

Newly added source family: emerging managers and league-table infrastructure

The TC43 search widened the universe to smaller and newer managers that publicly describe machine learning, systematic research, or AI-assisted investment workflows. The intake file records TC43, Machina Capital, Seldon Capital, Castle Ridge Asset Management, Machine Capital, TradeWell Capital, Bayswater, Sparkline, AXQ, CapitalsAI, Cedalion Capital, EMJ Capital, Hippocampus Capital Management, Marduci Capital Management, Prometei/FireSight, and Hydra Fund. These are search and verification leads, not peer rankings. The first disqualification loop is legal identity, current regulatory record, and an identifiable active vehicle; AI claims are then recorded separately from strategy and performance evidence.

The accompanying league-table and manager-directory source ledger adds HFR, Preqin, With Intelligence/Eurekahedge, Aurum, BarclayHedge, HFM/With Intelligence awards, Institutional Investor, Pensions & Investments, LCH, and regulatory registers. The sources measure different things: strategy indices, reported performance, AUM, lifetime net gains, awards, legal identity, or manager discovery. None is an AI-depth measure, and no composite ranking is created here.

AUM, strategy families, and foreign vehicles

The article’s manager universe is broader than hedge funds. It includes institutional systematic managers, multi-strategy funds, proprietary trading firms, and emerging managers. The table below keeps those categories separate. AUM figures are dated public observations, not a ranking of AI capability.

Firm	Public scale observation	Strategy families visible in public sources	Foreign/offshore figure	Evidence boundary
GMO	$76.8B discretionary AUM, Dec. 2025	Global equity, fixed income, multi-asset, alternatives	Not disclosed	Firmwide figure; no vehicle-level offshore total
Acadian	$233B, Jun. 2026	Systematic equity, credit, alternatives, fixed income, wealth/model advisory	$24B APAC AUM reported for Jun. 2025; not an offshore-vehicle figure	Regional AUM is not foreign-fund AUM
Arrowstreet	$344B+, Jun. 30, 2026	Global quantitative equity and systematic mandates	Not disclosed	Current official homepage observation; not an AI-capability measure
Man Group	$253.6B total, Jun. 2026; $66.5B hedge-fund AUM, Jun. 2025	AHL, Numeric, discretionary, private markets, solutions	Not disclosed	Total and hedge-fund figures have different scopes
CFM	$30B, Jul. 2026	Systematic macro, trend, multi-strategy, alternative-data/ML research	Not disclosed	Firm-reported total; official site also reports data and AI-lab activity
Balyasny	$38B, Aug. 2026	Multi-strategy, fundamental, quant, macro, tactical trading	Not disclosed	Firm-reported total
Bridgewater	$92B firm AUM, Sep. 2025; $78B hedge-fund AUM, Jun. 2025; $150.2B regulatory AUM, Mar. 2026	Pure Alpha, Pure Alpha Major Markets, All Weather, AIA Labs, Asia/China	$2.8B in a named Pure Alpha Major Markets Cayman vehicle, Mar. 2024; $7.5B China-management estimate from secondary reporting	These are not additive; the scopes differ
BlackRock	$11.55T total worldwide AUM, Dec. 2024 P&I profile	Systematic investing, active equities, fixed income, alternatives, ETFs, Aladdin	Not disclosed	Broad asset-manager total
WorldQuant	No comparable public AUM located	Systematic global equity and quantitative research	Not disclosed	Do not infer scale from recruiting or AI claims
Numerai	No conventional firmwide AUM located	Tournament signals, predictive models, Numerai fund ecosystem	Not disclosed	Not directly comparable to a conventional manager
Point72 / Cubist	About $50B firmwide in 2026 media reporting; $37.7B hedge-fund AUM, Jun. 2025	Fundamental equity, macro, Cubist systematic, private investments	Not disclosed	Firmwide figure is media-reported
Schonfeld	$22B+, Jul. 2026	Quantitative, fundamental equity, tactical trading, discretionary macro/fixed income	Not disclosed	Firm-reported total
G-Research	No external AUM; proprietary capital	Predictive ML, systematic research, engineering infrastructure	Not applicable	Prop/research firm rather than external-AUM manager
AQR	$207.1B, Mar. 2026; $77.6B hedge-fund AUM, Jun. 2025	Long-only quant, alternatives, multi-asset, hedge funds	Not disclosed	Total and hedge-fund figures have different scopes
Two Sigma	$75B+ firm-reported; $64.8B hedge-fund AUM, Jun. 2025	Quantitative hedge funds, systematic research, private investments	Not disclosed	Firm-reported and survey figures differ
Citadel	$66B hedge-fund AUM, Jun. 2025	Global quantitative strategies, fundamental equities, fixed income/macro, commodities, credit	Not disclosed	Hedge-fund series, not total corporate capital
Millennium	$77.5B hedge-fund AUM, Jun. 2025	Multi-manager pods across equity, macro, fixed income, relative value and quant	Not disclosed	Hedge-fund series
Capstone	No comparable public AUM located	Quantitative macro, systematic, relative value	Not disclosed	No verified current figure
Jump, Hudson River Trading, Optiver, DRW, IMC, XTX	Proprietary trading capital; external AUM not applicable or not disclosed	Market making, HFT, systematic execution, predictive ML	Not applicable	Prop firms should not be ranked with external-AUM managers
D. E. Shaw	$65.8B hedge-fund AUM, Jun. 2025	Systematic, fundamental, multi-strategy, private investments	Not disclosed	Hedge-fund series
PDT Partners	No comparable public AUM located	Quantitative equities, statistical arbitrage, systematic research	Not disclosed	No verified current figure
Aspect, Winton	No comparable current public AUM located in this pass	Managed futures, systematic macro, trend, alternative risk premia	Not disclosed	Strategy is visible; current scale is not
Systematica	$13B+ firm-reported homepage figure; measurement date not stated	Trend following, macro non-trend, multi-strategy, equity market neutral, customised solutions	Not disclosed	Current homepage figure; scope and measurement date are not defined
Marshall Wace	$43B hedge-fund AUM, Jun. 2025	Equity long/short, systematic and alternative strategies	Not disclosed	Hedge-fund series
Brevan Howard	$32B hedge-fund AUM, Dec. 2024	Global macro, systematic macro, relative value	Not disclosed	Dated hedge-fund series
Caxton	More than US$10B group client assets; information stated as at Dec. 31, 2024	Global macro across currency, financial, commodities, and securities markets	Not disclosed	FCA disclosure; delegated entity and group scope are not comparable to hedge-fund survey AUM
Voleon, Squarepoint	No comparable current public AUM located	ML equities, systematic multi-asset	Not disclosed	No verified current figure

Bridgewater detail: A publicly reposted March 2024 strategy document reports $111.8B firmwide, $7.7B in Pure Alpha Major Markets at 12% volatility, and $2.8B in Bridgewater Pure Alpha Major Markets Fund, Ltd. The $2.8B is one offshore vehicle, not the complete offshore complex. The Cayman registry confirms multiple historical Pure Alpha, Major Markets, All Weather/optimal-portfolio, inflation-linked-bond, and special-opportunities vehicles, but it does not report current assets. Strategy document · Cayman registry · secondary China estimate.

Interpretation rule: do not add firmwide AUM, hedge-fund AUM, regulatory AUM, gross assets, regional assets, and offshore vehicle assets. A foreign registration proves vehicle existence or registration status, not current assets, live activity, or investment authority. The full source ledger is sources/06-industry-verticals/hedge-fund-firmwide-aum-strategy-offshore-2026-raw.md.

The emerging-manager and regional intake names in this article—TC43, Machina, Seldon, Castle Ridge, Machine Capital, TradeWell, Bayswater, Sparkline, AXQ, CapitalsAI, Cedalion, EMJ, Hippocampus, Marduci, Prometei/FireSight, Hydra, Nujum, Major Capital, AYVID, Avangard, Index Solutions, High-Flyer, Lingjun, Mingshi, WizardQuant, DeepWin, JoinQuant, Zhongliang Wealth, QTS, Evovest, Maritime Fund, Dynamic Funds, GSA, Quantellence, and Ubiquant—remain discovery or verification leads. No current comparable firmwide AUM and offshore-vehicle figure was verified for that intake group in this pass.

Two additional title-blind manager surfaces sit in the formation and self-description lane. Lumenai Investments identifies itself as an SEC-registered investment adviser, names John Bailey as Founder and Managing Partner and Chris Messiana as Partner/COO/Head of Trading, and describes supervised and unsupervised algorithms, regime-conditioned security grouping, an OMS/PMS, daily monitoring, and systematic rebalancing. Its April 2026 announcement describes a planned Lumenai Innovation Fund using an agentic architecture developed with ETS Asset Management Factory. Lumenai’s disclosure explicitly says certain referenced ETS individuals are not Lumenai employees; that personnel boundary is preserved here. The public pages do not establish that the planned fund began operations, disclose model weights or providers, or provide independently audited performance.

Ventium Capital describes an AI-native event-driven strategy focused initially on index rebalances and names Alexandru Runcianu as Founder/CIO, with a biography citing Bank of America’s systematic-risk/portfolio-trading desk and prior Arini Capital experience. Its founding thesis describes agents for filing and flow research, 13F-based crowding, sizing, adaptive execution, liquidity and impact modeling, reversal capture, and a trade-level feedback loop. It also states that campaigns require human sign-off, hard limits are independent of the trading system, and positions can be overridden. The page displays “LIVE PILOT: ACTIVE,” but the public record reviewed does not independently verify a live fund, capital deployed, model architecture, data permissions, execution history, or performance. See the capture note. These are two public operating narratives, not a comparison or ranking.

The firms are revealing different layers

Firm	What it is showing	Public clue used in this review	Evidence boundary
GMO	AI as an investment object, research input, and named research initiative	2025 systematic letter, 2025 Quality letter, 2026 Tom Hancock video, Form ADV AI-risk language, and the August 2026 Systematic Equity “Super Analyst” page	Investment-thesis and research-agenda evidence; model, owner, permissions, and production stage are not specified
Acadian	AI as part of the investment process, with current agentic hiring language	2026 Investment AI Engineer role plus named-practitioner and management disclosures	Workflow direction is described; production status and reporting lines are not established
Arrowstreet	AI as a firm-wide platform with model access, tools, telemetry, and execution boundaries	Senior AI Platform Engineer role	Hiring artifact describes intended architecture; investment use and deployment are not established
Man AHL	GenAI as a research assistant and specialized workflow for trend-following strategy design	2025 Alpha Assistant and 2026 AlphaTrend publications	Disclosed research workflow; broader production scope and attribution are not public
CFM	AI/ML as part of a scientific research, text-analysis, model-training, multimodal-extraction, and prediction-services stack	CFM approach/strategy pages, 2026 AI-lab articles, Hugging Face case study, and CFM AI-hiring materials	Public AI/ML direction and selected modalities are described; current GenAI production scope is not established
Balyasny	Centralized Applied AI platform with team-specific agents and model-evaluation infrastructure	OpenAI customer case study, public personnel profiles, and media reporting	Firm/vendor-reported workflow; independent performance attribution is not public
BlackRock	Firmwide AI Labs, systematic alpha-model research, an equities research assistant, and LLM-based macro research	AI Labs, Augmented Investment Management, Asimov disclosure, current AI Labs role, and Systematic Investing pages	Multiple public AI layers are visible; model inventory, ownership, evaluation results, and attribution are not public
WorldQuant	Agentic systems connected directly to portfolio-management hiring language	Dated Portfolio Manager, Agentic Systems role artifact and current AI/LLM careers	Detailed hiring intent; the dated role is no longer live at the checked URL; system deployment and performance are not established
Numerai	Public AI-scientist workflow, predictive language-model feature generation, open agent skills, and MCP-enabled tournament automation	Numerai homepage, NumerCon 2026 recap, official docs, GitHub example scripts, and official careers host	Public platform, dataset, API, and hiring evidence; no proof of autonomous capital allocation, internal portfolio authority, or independently audited model contribution
Point72 / Cubist	AI hiring across fundamental equities, investment services, NLP, fine-tuning, agents, RAG, and systematic ML	Current firm-owned job listings, Cubist page, Market Intelligence page, CSP repository	Intended architecture and research scope are described; filled roles, live systems, and ownership are not public
Schonfeld	Internal multi-model platform, investment-team AI lab, agentic retrieval, MCP, and pilot/evaluation controls	FE AI Lab, current engineering roles, AI Q&A, and leadership disclosures	Workflow direction is described; exact deployment and authority are not public
G-Research	Production engineering LLM application, centralized open-model/MCP runtime, and NLP/agent hiring	Official code-review article and current roles	Engineering production and infrastructure are described; live GenAI investment use is not public
AQR	ML/NLP investment methods and affiliated point-in-time language-model research	Official research, product disclosure, leadership, and academic paper	ML/NLP and affiliated LM research are public; AQR-owned GenAI deployment is not public
Two Sigma	LLM-assisted feature research, company-aware tools, production LLM/NLP hiring, agentic research libraries, post-training hiring, and named AI leadership	2026 outlooks, feature research, CAIS material, current official roles, and current leadership titles	Workflow direction and titles are described; model weights, permissions, and trading authority are not public
Citadel	Equity-research assistant, research-validation agent account, Claude-for-Excel signal, systematic ML, and current LLM/fine-tuning hiring	Official pages, Anthropic customer page, current roles, Reuters NEXT, and named interviews	Separated workflow signals are public; exact assistant architecture and GenAI-to-trading link are not public
Millennium	AI advisory, a dedicated AI Lab, agentic infrastructure, internal search, end-user tools, and human-oversight policy	Official technology, leadership, June 2026 AI Lab Q&A, people, and privacy pages	Firm-reported operating model; partner projects, model inventory, and investment authority are not public
Capstone Investment Advisors, Jump Trading, Hudson River Trading, Optiver, DRW, and IMC Trading	Adjacent asset-manager, prop-trading, and market-making AI control cases	First-party AI/ML, AI Labs, research, AI/RAG career, and agentic developer-workflow pages checked on August 12, 2026	Current public operating and hiring evidence; no autonomous capital allocation, model inventory, or AI-attributed return evidence
D. E. Shaw, PDT, XTX, Aspect, Winton, and Systematica	Predictive ML, systematic research, data/compute, and selected AI/LLM practitioner signals	Official careers, research, podcast, and firm pages	The cited layer is public; a current GenAI investment workflow is not established
Marshall Wace, Brevan Howard, Caxton, Voleon, and Squarepoint	General technology, quantitative, macro, ML, regulatory, or personnel signals	Official firm, regulatory, public-profile, and hiring surfaces reviewed in this pass	Firm-controlled GenAI specifics were not located for most of this group; Voleon has one current personal/Scholar language-model personnel signal, but no firm-controlled LLM program disclosure

The distinction matters. An investment thesis, a research workflow, a platform-control design, and an agentic strategy-design system are different kinds of public evidence. The table describes evidence boundaries; it is not a score or ranking.

Fresh public-source verification: August 8, 2026

This verification pass added current or newly checked sources. Each item records what the source supports and what it does not. Hiring language supports hiring intent; firm or vendor statements remain attributed claims; none of these items is treated as proof of investment performance or autonomous capital allocation.

• Acadian: The Q1 2026 earnings-call transcript describes broader generative-AI use across research and operating workflows, AI-assisted coding, selected research services, shared guardrails, security, and investment in computer-science and machine-learning personnel. The investor-forum transcript adds Claude Code/coding assistance, infrastructure modernization, agentic-AI investment, and discussion of security and out-of-sample robustness. These are management disclosures; they do not name models, permissions, adoption metrics, or return attribution. A Michael Doros interview and Acadian’s related credit note describe LLMs for data processing, coding, idea testing, and productivity while retaining human judgment, testing, and portfolio controls.

• Man Group: Tushara Fernando’s current profile identifies her as Head of Data and AI and assigns responsibility for developing and implementing the firm’s generative-AI strategy. Man’s Anthropic partnership announcement describes Claude, Claude Skills, and Claude Code across investment, distribution, HR, and coding workflows. These sources establish executive remit and a company-reported partnership; they do not identify model-level ownership, permissions, or performance.

• Bridgewater: Oliver Simon’s profile describes end-to-end equities-system work and the AIA core investment system. Nina Lozinski’s profile describes the launch of AIA Labs and firm-reported growth in its technical team. Bridgewater’s current partnership page separately lists Nina Lozinski as Head of AIA, Aaron Linsky as Head of Engineering, Alpha Engine, Blake Cecil as Deputy Chief Investment Officer, Alpha Engine and AIA Labs, and Oliver Simon as Head of AI & ML Investment Strategy. The PAT case study names the tool’s contributors and describes codified knowledge, LLMs, agentic workflows, and investor feedback. Bridgewater’s Jasjeet Sekhon profile now records his former Chief Scientist/Head of AI role and current Google DeepMind position, so older rosters are date-scoped. These sources describe systems and staffing; Bridgewater’s claims about capital use and alpha remain firm-reported and unaudited here.

• Jane Street: The CoreWeave agreement and related SEC exhibit describe approximately $6 billion of AI-cloud capacity and a $1 billion equity investment. Jane Street’s data-center account describes a 4,032-GPU facility, while its AIDE engineering article describes an internal Claude-Code-like developer tool. A CoreWeave GTC 2025 recap, accessed August 27, 2026, separately identifies a GTC session with Jane Street research contributor Adam Canady and says the session discussed training and fine-tuning quantitative-trading AI models on CoreWeave infrastructure. That is vendor-conference evidence, not a Jane Street-authored model inventory, workload denominator, permission map, production endpoint, or AI-attributed return record. A separate 2017 Yaron Minsky recording and its talk listing provide historical engineering context: Minsky describes transparent, shared, streaming quantitative models; spreadsheet replacement; incremental computation; and production use across multiple desks, with an approximately tenfold internal speed improvement over prior iterations. The recording uses automatic captions and reports no model name, benchmark protocol, desk list, current deployment status, GenAI workflow, or investment-return attribution. A separate YOW! 2018 recording adds a named internal “Incremental” system, roughly eight historical implementations, and speaker-reported market-data engineering figures; it likewise predates current GenAI and supplies no current deployment or return evidence. A 2012 OCaml workshop recording adds historical Core-library context: Minsky describes a reusable standard-library layer, public release of non-proprietary infrastructure, and internal review/testing around the software foundation. A 2025 Jane Street Signals & Threads episode with Ian Henry adds a firm-controlled account of options-desk “Trader Tools” for market understanding, configuration of automated systems, manual orders, visualization, alerting, and monitoring, plus custom Bonsai-based interfaces. A June 2026 episode with Jacob Baskin adds public compute-platform detail: Jane Street’s named Superstore data layer and Hive cluster are described as supporting neural-network training, feature-data production, historical trading analysis, and large simulations; the speaker reports hundreds of thousands of CPU cores and more than 10,000 GPUs, research/trading-system isolation, and CPU/GPU-hour prioritization through a second-price auction. These are first-party practitioner disclosures, not a complete model inventory, data-rights map, autonomous trading permission, or performance record. See the 2017 capture note, YOW! capture note, Core capture note, Ian Henry capture note, and Baskin capture note.

• Jane Street / John Crepezzi: The AI Engineer talk identifies Crepezzi as working on Jane Street’s AI Assistance team and describes custom coding-assistant work for the firm’s OCaml-heavy environment. The recording describes workspace snapshots and build-status transitions as training examples (06:08–09:23), a Code Evaluation Service that tests whether generated diffs compile, typecheck, and pass tests (09:23–11:37), and an AIDE sidecar that centralizes context construction, prompt strategies, model swapping, editor integrations, telemetry, and A/B tests (12:39–15:50). The same talk mentions retrieval, multi-agent workflows, and reasoning models as ongoing areas (15:50–16:16). This is unusually specific developer-tool and evaluation evidence, but it is not evidence of a trading model, trading authority, model weights, data rights, or AI-attributed performance. See the capture note.

• Point72/Cubist and Schonfeld: Point72’s Cubist page continues to describe systematic ML and current research infrastructure; its Cubist AI-chess hackathon is an agentic evaluation signal outside investment deployment. Schonfeld’s Hannah Jiang Q&A describes an AI engineering team and a PM deployment exercise; current AI Strategy Analyst and AI Data Engineer roles add hiring intent around research, trading, risk systems, embeddings, and vector stores. The cited material does not establish production authority or investment performance.

• G-Research and Two Sigma: G-Research’s Mia profile identifies an AI Engineer in its GenAI Engineering team delivering AI capabilities across the business; the page does not establish finance-LLM or live-trading use. The G-Research personnel capture note records the partial-name boundary. Two Sigma’s 2026 outlook and event page provide current role and event context; the event page is dated October 29, 2024, not 2023.

• Citadel, Millennium, XTX, and Aspect: Citadel’s current leadership page identifies Andrew Janian as Interim CTO; current ML researcher and ML engineer roles describe LLM, pretraining, fine-tuning, distributed training, inference, and productionization as role requirements. Millennium’s AI Lab Q&A describes experimentation, AI-partner collaboration, and specialist hiring. XTX’s Canadian FX Committee materials record a presentation on deep learning in FX market-making; Aspect’s AIMA interview describes ML and practitioner use of ChatGPT across research and business functions. These sources describe roles, research, or practitioner statements; they do not provide a common deployment or performance measure.

• Caxton, Winton, Systematica, Marshall Wace, Brevan Howard, Voleon, and Squarepoint: The reviewed public surfaces expose combinations of LLM hiring intent, predictive ML, NLP, quantitative research, compliance tooling, careers material, or public personnel profiles. For Caxton’s public LLM-engineer listing, the evidence is hiring intent around RAG, proprietary knowledge, and document extraction. For Voleon, the evidence is a current personal/Scholar language-model personnel signal, not a firm-controlled program. For the other firms, the absence of a located public LLM/agent source is retained as a search boundary, not a claim about private capability.

• BlackRock: Its AI Labs page names Rachel Schutt and Stephen Boyd as co-heads and says the lab applies statistics, machine learning, optimization, stochastic control, and decision theory across retirement, trading, alternatives, and ETFs. A current AI Labs software-engineering role explicitly includes generative AI, production deployment, and alpha generation among the lab’s remit. BlackRock’s Augmented Investment Management material describes a systematic alpha-model system built from large signal libraries, while its official Asimov disclosure describes a proprietary equities-research tool. Its Systematic Investing page separately describes LLM use for macro narrative, analyst views, market consensus, and tradable signals. These are several public layers; they should not be collapsed into one model or one production claim.

• WorldQuant: A detailed Portfolio Manager, Agentic Systems role artifact describes live-book risk management, agentic systems with planning, tool use, memory, reflection, and collaboration, reinforcement-learning hyperparameter tuning, deep-learning model development, custom workflows, and human-in-the-loop checks. The direct URL no longer resolves as a live vacancy on the August 10 check. WorldQuant’s current official careers index still exposes AI Scientist, AI software, WQBrain AI research, deep-research, and LLM/AI-agent hiring signals. This is detailed public hiring intent; it does not establish a filled role, production architecture, model weights, or return attribution.

• QRT and Tower: QRT’s current careers board lists AI Platform Engineer and AI Platform Developer roles alongside technology, data, compute, and quantitative-research roles. The reviewed public board does not expose enough role detail to infer model use or deployment. Tower’s careers page explicitly describes machine-learning work, research tools, predictive signals, data management, compute, risk, and compliance; no current LLM or agentic investment workflow was located. QRT remains a platform-hiring signal, while Tower is a predictive-ML control case.

GMO: investment-thesis and research-method evidence; limited public platform detail

Strategy visible in public

GMO’s public AI posture is consistent with its long-standing valuation discipline. The firm describes a long-horizon, valuation-led, contrarian process, with quantitative, fundamental, and blended teams in its public approach. Its 2025 Systematic Equity letter shows where machine learning fits: a systematic architecture built around Value, Momentum, and Alerts, with advanced ML sentiment, network-aware momentum, corporate-behavior alerts, and plans for AI-based analytical tools and uncertainty estimates.

That is not a generic “AI will change everything” position. It is a capital-allocation framework. GMO’s 2025 Quality Strategy letter maps the AI economy as a four-layer stack—applications, large language models, compute, and data-center suppliers—then asks where capital spending and operating economics will accrue. The investment question is which layer captures durable economics after accounting for concentration, capital intensity, and competitive pressure.

Tom Hancock makes this framework explicit in GMO’s 2026 public video and event material. Hancock is Head of Focused Equity, a portfolio manager for Quality Strategies, a partner, and a computer-science PhD with prior software-engineering and research experience. His public AI work should be read as investment judgment around AI economics, not as evidence that he runs GMO’s internal AI platform.

An August 2026 GMO Systematic Equity research page adds current GenAI wording from GMO. Alongside lifecycle modelling, country–industry signals, transaction-cost modelling, and uncertainty optimization, GMO lists “Super Analyst” and describes it as deploying GenAI to add qualitative depth to quantitative breadth. The same page describes a glass-box research process in which positions trace to explicit research insights and says the Systematic Equity Team’s views run through August 2026. This is a firm-authored research-agenda disclosure, not an architecture or deployment audit: no provider, model identity, training or retrieval data, evaluation design, permissions, named owner, or AI-attributed performance is disclosed. See the capture note.

An additional Excess Returns interview and its full publisher transcript (audio published April 9 and transcript page dated April 10, 2026) gives the framework more structure. Hancock describes four layers—applications, hyperscalers, LLMs, and enabling infrastructure—and follows capital spending down the stack, while discussing proprietary data, customer embedding, growth versus maintenance capital expenditure, and uncertainty about LLM differentiation (03:51–07:33, 11:00–13:00, 19:00–22:00, 28:30–30:10). This is a clearer account of GMO’s public AI-economics and quality-investing lens; it remains separate from evidence about GMO’s internal GenAI platform, model inventory, data rights, or production authority. See the timestamped capture note.

GMO’s Form ADV adds a different signal. The firm says third-party or open-source AI tools such as ChatGPT may be restricted and identifies risks including confidential-data exposure, intellectual-property loss, hallucinations, and regulatory uncertainty. That is a governance disclosure, not a product announcement. It does show that AI-tool access and data handling have entered the firm’s formal risk vocabulary.

The same current GMO personnel surface adds Chris Heelan. GMO identifies Heelan as a researcher on its ESG Research Team responsible for ESG-related quantitative analysis across investment products and asset classes, and says he previously served as the firm’s Machine Learning Development Lead for Investment Data Solutions. The biography describes prior cloud-computing and machine-learning work on large-scale neuroscience data at Brown University, with undergraduate training at Vanderbilt and graduate training at Brown. This is useful role and academic-lineage evidence, but it does not identify Heelan as the owner of Super Analyst or a current GenAI leader, and it does not disclose a model, dataset, permission boundary, deployment stage, or performance result.

The public record is broader than investment commentary alone. GMO says it has built NLP infrastructure for differentiated alpha models. A live Data Platform Analyst role references AI tools for reference-data operations, validation, and production monitoring, while a Client Systems Manager role references GenAI integration in CRM and client-engagement workflows. These are operational and client-service signals, not evidence of an investment-facing agent platform.

GMO’s March 2026 quality-investing paper, March 2026 alternatives paper, and January 2026 AI valuation viewpoint extend the public investment framework. The papers discuss AI applications, LLMs, compute, suppliers, AI-related sentiment, uncertainty, and valuation. They are investment research and viewpoints; they do not establish a unified GMO operating strategy or current internal system ownership.

GMO’s 2026 Outlook page for Australia and New Zealand adds a current, first-party event route. It names Tom Hancock, Ben Inker, Catherine LeGraw, John Thorndike, and Tina Vandersteel and describes a panel of market predictions that includes AI-related views. The page is useful for dated speaker and investment-thesis coverage, but it is not a technology disclosure: it publishes no internal model, AI engineering owner, training data, evaluation, or production permission information.

A distinct April 2026 Moneycontrol interview with Arjun Divecha adds an individual-practitioner route that is easy to miss in a firm-only search. The video description identifies Divecha as a GMO senior adviser and says he uses AI in his own investment process. In the discussion, he describes model-and-agent-assisted coding, deep-research queries, personal apps, and a simple strategy-testing example (36:41–41:30). He also discusses model non-determinism, training cutoffs, and the need to check outputs (44:45–46:44). When asked about generating alpha, he explicitly presents the answer as personal experimentation rather than speaking for GMO and reports no success in those experiments (41:51–42:17). This is evidence of a named person’s public AI workflow and reliability concerns, not evidence of GMO-approved tooling, internal deployment, model ownership, investment authority, or performance. See the capture note.

Man’s current applicant privacy notice supplies a separate enterprise-governance signal. It says AI supports research, business-process streamlining, meeting records, compliance, and internal training; it also says AI-generated records may become business records subject to governance, audit, regulatory requests, and legal proceedings. This is an official data-governance disclosure, not evidence about AlphaGPT’s investment authority or model controls.

QRT’s official LinkedIn post adds a separate current academic-network route. It describes a recent Paris workshop of the firm’s Deep Finance and Statistics initiative with École Polytechnique, with topics including market simulation, trading algorithms, generative AI, and risk management. QRT says the partnership with École Polytechnique and its foundation has continued for six years and that it supports two Master’s programmes in Data Science and AI for Finance. The post names Charles-Albert Lehalle, Stefano De Marco, Marie Caillat, and Cristina Gastineau in connection with the workshop. This is firm-controlled partnership and topic evidence; it does not provide a recording, paper list, model, dataset, evaluation, production endpoint, agent permission, or performance result.

QRT’s current company feed, checked August 27, also exposes a wider public research and talent surface. One update says QRT teams attended ICML in Seoul, PyCon in Long Beach, and PyData in London, and opened applications for NeurIPS Sydney 2026 travel grants. Other updates describe a global coding challenge involving 110 interns across 18 teams and a MathQ/C3 programme that reached 800 students and brought 50 finalists from Oxford, Cambridge, Imperial College London, and UCL to mathematical, algorithmic, and agentic challenges. QRT’s commitments page lists support for the Python Software Foundation, NumFOCUS, Rust Foundation, IHES, ETH FinsureTech Hub, and the University of Zurich’s Portfolio Management Program. Together these are current ecosystem, conference, and talent-pipeline signals; they do not identify a QRT model, training corpus, evaluation fixture, production endpoint, agent permission, or investment result. See the capture note.

The self-authored CV of Orso Forghieri supplies a dated personnel and academic-lineage route. It lists a Qube Research & Technologies quantitative-research role in Paris from October 2025 through March 2026 and describes machine-learning and representation-learning work across datasets covering more than 3,000 equity assets, feature engineering, out-of-sample evaluation, statistical performance analysis, and Python/JAX infrastructure for learning-based decision systems under noisy data. The CV also lists a PhD in applied mathematics and reinforcement learning at École Polytechnique, supervision by Erwan Le Pennec, Hind Castel-Taleb, and Emmanuel Hyon, and publications in IFIP Networking 2025, EWRL 2024, and ICAPS 2024. These are self-reported dated-personnel claims, not a QRT paper or independent employment record; they do not establish QRT ownership of personal research artifacts or entry into a live strategy. The capture note preserves the source boundaries.

Key public people

Person	Publicly visible role	What can be inferred	What cannot be inferred
Tom Hancock	Head of Focused Equity; PM for Quality Strategies; partner	Public GMO investor associated with AI-economics analysis and quality/valuation positioning	Ownership of internal GenAI engineering or model governance
Warren Chiang	Head of Portfolio Management, Systematic Equity	Senior owner in the systematic research and portfolio-management chain	A specific LLM or agent program
George Sakoulis	Head of Investment Teams & Systematic Equity	Senior systematic-investment leadership	Current GenAI platform ownership
Ben Inker	Co-Head of Asset Allocation	Senior allocator whose public AI-bubble and valuation framing influences portfolio context	Engineering or AI-product leadership
Hylton Socher	Chief Technology Officer and partner; previously founded VaraQuest, focused on machine-learning portfolio construction and management	Public technology leadership with a machine-learning and AI academic background	The public biography does not establish current GenAI ownership, model selection, or production authority

Jeremy Grantham remains a highly visible GMO cofounder and long-run market voice, but his public AI commentary is a strategic viewpoint rather than evidence about GMO’s operating systems. That separation is important when reading high-profile investor commentary.

What GMO is not disclosing

The reviewed public record does not identify GMO’s internal foundation-model vendors, model-routing layer, retrieval architecture, agent framework, AI engineering head, user adoption, evaluation suite, or live portfolio authority. It also does not separate how much of the systematic ML language refers to established predictive models versus current generative AI.

Public-evidence boundary: GMO’s public AI evidence is visible primarily at the investment-thesis and research-method level. The visible strategy is “understand the economics, use machine learning where it improves signals, and keep valuation and uncertainty in charge.” The public evidence does not establish an agentic-investing platform.

Acadian: public investment-process and agentic hiring signals

Strategy visible in public

Acadian’s public materials show a systematic manager building from data and research infrastructure outward. Its systematic-edge page describes more than 100 investment professionals, more than 95 advanced degrees, 35-plus years of data, hundreds of millions of daily observations, tens of terabytes of data, broad market coverage, and tens of thousands of traded assets. These firm-reported scale figures provide context for Acadian’s research environment; the page does not itself establish that its AI strategy is modular.

A current Global Equity strategy page, retrieved August 27, 2026, gives a more direct but still high-level process statement: Acadian says curated AI and ML applications help extract information from complex datasets and support financial-data collection, analysis, and monitoring across a broader investment universe. The same page says the process uses more than 70 predictive factors, including ESG signals, and displays holdings marked as of July 2026. This is current first-party strategy positioning; it does not identify a model, training corpus, vendor, agent runtime, permissions, or AI-attributed return. The source note preserves that boundary.

A newly recovered Fear & Greed video route adds a more granular, speaker-reported process map. The interview describes separate AI modules for pattern discovery in large market datasets (01:00–01:18), analyst-bias correction and fundamental forecasting (01:36–02:00), and extracting signals from earnings-call language (03:35–04:10). Later, the interview describes those outputs feeding an automated investment process, expected-return and risk screening, portfolio optimization, and trading decisions (05:15–06:10). The English captions are automatic and privately archived; these timestamps support navigation and bounded paraphrase, not a speaker-verified transcript, model inventory, or independent evidence of AI-attributed performance.

The firm’s 2024 GenAI paper outlines a practical use-case map: missing-data imputation, synthetic-data generation, risk modeling, and language-model-based natural-language processing. It also keeps the difficult parts visible: low signal-to-noise, point-in-time data, specialized expertise, and the danger of treating generated output as investable evidence. Because it predates the current model generation, it is historical strategy context rather than proof of today’s tooling.

A current public signal comes from a 2025 named-practitioner interview on AI in investing. The speaker describes AI modules processing a broad equity universe, correcting for analyst bias, extracting information from earnings-call questions and answers, reading newsflow, and incorporating supplier/customer and peer-fundamental context alongside technical patterns. The modules feed expected-return forecasts, which then move through portfolio construction, risk and transaction-cost constraints, and order routing. Humans remain in the loop. The interview also gives a useful negative control: generic ChatGPT is not a portfolio-construction system without market data, financial models, source inputs, and a bridge from research output to expected returns.

Acadian’s public investment commentary also includes Owen A. Lamont’s January 2026 AI-bubble analysis and his February 2026 worst-case scenario analysis. These articles expose how a senior Acadian investor is framing AI economics, concentration, and downside risk for investment analysis. They are evidence about the firm’s public investment perspective, not evidence about internal GenAI infrastructure or implementation.

A current recruiting signal is the public VP, Investment AI Engineer role. The description places the role inside an investment team alongside portfolio managers and quantitative researchers. It calls for agentic solutions that accelerate repeatable investment research and portfolio tasks; reusable skills that encode institutional knowledge; sub-agents and controlled workflows; testing; human approval; scope, cost, and data guardrails; and metrics for adoption, quality, and time-to-value.

That is a concrete operating model: investment teams own the problems, AI engineers turn the problems into reusable workflows, and controls determine what can execute. The role still does not prove that the system is in production. It proves that Acadian is recruiting for that direction.

The title-blind media recovery adds two public routes. The CFA Institute podcast with John Chisholm is a 2021 executive conversation about machine learning, big data, ESG, value investing, quantitative culture, and data-science talent. Acadian’s May 19, 2026 investor-forum transcript adds management discussion of approximately 70 data and data-infrastructure staff, AI-enabled coding and research support, human judgment, generic-LLM limits, and out-of-sample robustness. The transcript also records a speaker saying an internal AI research agent was used to confirm a historical milestone: natural-language processing in production in 2008, with research reportedly going back to 2006. That is a dated management statement and agent-assisted historical recollection, not a current model lineage or deployment audit. These sources add temporal and executive context; they do not identify a model vendor, model weights, agent permissions, or AI-attributed returns.

The official 2021 Acadian ML podcast page adds Ryan Taliaferro and Vladimir Zdorovtsov as named speakers, while the Impact of AI/ML on Quant page frames the discussion around alternative data, quantitative-process applications, and implementation risks. A publisher mirror displays a different date for the same episode, so the date discrepancy is retained rather than normalized. Acadian’s current Credit’s Systematic Shift note further describes filings, earnings calls, news, issuer relationships, code generation, and agentic research machinery as relevant systematic-credit workflows. The full capture and boundary notes are in the new source note.

Acadian’s May 2026 investor-forum presentation makes the firm-level strategy more explicit: enterprise chatbots, agentic coding, AI-enabled services, and governance. The presentation also describes roughly 70 data employees across Investments and IT, spanning research/analytics, data quality and access, storage/compute, and AI. These are management disclosures about organizational design, not an independent headcount or deployment audit.

The recovered CIO episode transcript adds historical operating detail: Brendan Bradley describes portfolio-manager oversight as a catalyst for research, links research to portfolio management and implementation, and says the 2021 agenda included non-traditional alpha estimation under machine-learning-type methods. He also says NLP had been in Acadian’s investment process for a decade. The recording is useful for temporal process evidence; it does not establish current staffing, model versions, vendors, permissions, or AI-attributed performance.

The title-blind 2021 cybersecurity episode names Adam Connell as Acadian’s Director of Information Security and discusses user reporting, security-awareness testing, asset inventory, patching, and multifactor authentication. The timestamped capture adds adjacent control vocabulary for evaluating later AI data-access claims, but it does not establish an AI-agent policy or current security architecture.

The recovered co-founder episode transcript adds a second historical layer: Acadian’s first-party page exposes a public Podbean MP3, and the local chunked ASR supplies timestamped navigation. John Chisholm describes Acadian’s expanding data and processing capacity, says early ML work raised overfitting concerns (39:16–39:44), and attributes later incorporation of ML-based factors to team expertise and relevant academic backgrounds (39:44–40:16). A public CFA Institute transcript PDF provides a separate searchable transcript route without publisher timestamps. The episode was recorded before his June 2022 retirement, so it is temporal process evidence rather than a current personnel or system disclosure. The Acadian/QRT recovery note preserves the full title-blind route and its university-partnership boundary.

The completed title-blind feed pass also transcribed Episode 4 on ESG. That 2021 recording identifies Malcolm Baker as Acadian’s Director of Research and a Harvard Business School faculty member, Matthew Picconi as a Sydney-based ESG portfolio manager, and Ryan Taliaferro as Director of Equity Strategies. The discussion describes multi-factor analysis across a large security universe, messy ESG data, and the need to assemble information so it fits the investment process rather than plug an off-the-shelf data set directly into a model (15:25–17:29, 34:04–34:26). This is historical data/process and personnel evidence, not a GenAI deployment disclosure. The same pass found that Episode 2 is a macro/election discussion with no verified AI or model-use statement; it remains a completeness and historical-personnel record rather than AI evidence.

Acadian’s historical March 2019 ML paper adds a bounded research-method record. Its random-forest case study adapts Piotroski’s F-score, uses 2000–2012 for training and validation and 2013–2018 for out-of-sample comparison, restricts tree depth to three levels, and tests forests ranging from 300 to 3,000 trees. The paper reports 9.4 basis points per month for the ML F-score versus 6.2 for the baseline, but explicitly says the results are not investible and omit transaction costs, borrow costs, and other frictions. The source record preserves the controls and historical boundary; this is evidence about Acadian’s public ML research vocabulary, not current model performance or deployment.

The public investment-use record reaches beyond research commentary. Acadian’s 2023 Net Zero Alignment Model announcement describes an LLM-based model analyzing company reports and sustainability-target credibility. Its 2025 Responsible Investment Statement describes in-house AI, NLP, and LLM techniques for forward-looking signals, greenwashing detection, and supply-chain analysis. These public use cases are closer to text and ESG research than to autonomous portfolio decisions.

Acadian’s 2025 Client Conference adds a dated operating surface. The agenda names an LLM session, “A Day in the Life of Acadian Data,” and a trade-show description covering alpha forecasting, portfolio attribution, optimization scheduling, portfolio review, and ESG analytics that applies LLMs to unstructured data to generate signals. It also names Doug Eisenstein in investment data solutions and Jim Soper in technology, alongside research and systematic-credit leaders. The current Investment AI Engineer posting supplies a separate hiring signal for reusable investment skills, sub-agents, human review, testing, controlled execution, and adoption metrics. Together these sources clarify the public vocabulary around data, investment workflows, and intended controls; they do not establish a filled role, a live agent fleet, model vendors, or attribution.

A title-blind INSEAD Emerging Markets Podcast interview with Asha Mehta, published June 5, 2024, adds a dated Acadian alumni and regional-market route. The publisher identifies Mehta as Managing Partner at Global Delta Capital and describes her earlier Acadian roles as Lead Portfolio Manager and Director of Responsible Investing; Acadian’s historical profile separately identifies her as a Senior Vice President and Portfolio Manager. In the episode, Mehta describes using quantitative tools and data science to cover frontier and emerging markets at scale, then combining that breadth with fundamental and on-the-ground context (07:46–18:00). She also discusses systematic treatment of responsible-investing signals (32:00–35:00). This enriches the historical personnel and process map, but it is not evidence of current Acadian employment, current Global Delta GenAI use, transferred systems, model identity, permissions, or performance. See the capture note. Acadian’s public Systematic Methods archive, backed by a JSON results endpoint, returned 17 dated items in the August 20, 2026 capture. The archive adds a 2024 Generative AI in Systematic Investing paper that discusses foundation-model data requirements, NLP, domain tuning, and the limits imposed by insufficient training data. It also exposes first-party pages for Episode 3 on machine learning, co-founder John Chisholm, and CIO Brendan Bradley. A subsequent complete-feed reconciliation found eight RSS items and recovered all eight public audio enclosures, including the previously omitted executive and adjacent-function episodes. Episodes 1 through 7 now have timestamped local transcripts; Episode 2 is retained as non-AI macro context, Episode 4 as ESG data/process context, and Episode 5 as adjacent cybersecurity-control evidence. The trailer remains a completeness item, so no AI or model claim is made for it.

The personnel pass separates role-level signals from confirmed ownership. A self-described public profile associated with Acadian suggests a Vice President of Artificial Intelligence role, while a separate profile identifies an adjacent implementation position. Acadian’s official leadership page does not list a Chief AI Officer, Head of AI, or VP of AI. Neither profile establishes formal reporting lines, AI-platform ownership, language-model fine-tuning, or agent autonomy.

The current leadership page also identifies James Soper as CTO and James Crumlish as CIO. This corrects stale material that described Crumlish as CTO. Neither current biography assigns either executive ownership of Acadian’s AI platform or investment agents.

Key public people

Person	Publicly visible role/background	Signal	Boundary
Vice President of Artificial Intelligence	Public LinkedIn profile self-identifies this Acadian role; the source record is retained in the internal worker-layer ledger	Public signal of named AI leadership	The profile is self-described and does not establish formal firm confirmation, scope, or production authority
Senior VP / portfolio manager / research lead	Named Acadian practitioner in a 2025 podcast interview; name withheld from this public brief	Detailed public account of Acadian’s AI-assisted investment workflow	The public brief withholds the name and does not infer platform ownership
Vladimir Zdorovtsov	Director, Global Equity Research; PhD; former quantitative-research and ML-hedge-fund leader	Quantitative-research and ML background; current public role is research leadership, not a named GenAI executive role	The public biography does not establish current GenAI ownership
Javier Alcazar	Acadian’s 2023 investment-team announcement describes algorithmic trading, ML, deep/reinforcement learning, NLP, large/unstructured-data infrastructure, and generative modeling; an April 2026 public investment-committee packet identifies him as Senior Vice President, Portfolio Manager, Research, Acadian Asset Management (U.K.)	Current investment-research affiliation plus a direct ML/NLP background signal	Current role is corroborated outside Acadian’s own roster; background and title do not establish GenAI ownership or platform authority
Owen A. Lamont	Author of Acadian’s public AI-market commentaries	Public investment framing around AI concentration, economics, and downside scenarios	The articles do not establish a current AI-platform title or ownership of agentic workflows
VP, Investment AI Engineer	Current public role; name not disclosed in the listing	Direct signal of an emerging agentic investment-engineering function	The listing does not identify the person filling the role or prove production deployment

What Acadian is not disclosing

Acadian does not publicly specify which foundation models it uses, how investment skills are versioned, how agents are evaluated against research baselines, what data can leave the firm, how many employees use the system, or whether any agent can cross the boundary from recommendation to order submission. The public language points to controlled automation, not autonomous trading.

Public-evidence boundary: Acadian’s public evidence connects AI modules with investment research, forecast models, portfolio construction, execution, and human review. The public record does not establish how much of this workflow is in production, how it performs against a non-AI baseline, or whether the self-described AI leadership role owns the underlying systems.

Arrowstreet: visible GenAI platform architecture

Strategy visible in public

Arrowstreet’s investment and technology pages describe a quantitative manager organized around research, software, data, models, portfolio construction, automated portfolio management, and scalable systems. The public investment process is systematic and computational. It says little about generative AI as a source of alpha.

The investment page is also a useful negative-control source: it explicitly assigns model development and testing to quantitative researchers, and tools/platforms for quantitative research to quantitative developers with software and data-science expertise, without naming an LLM, agent, or AI leader. That absence should be tracked separately from the AI-platform hiring evidence; it is not evidence about private systems.

The firm’s Senior AI Platform Engineer listing changes the picture at the platform layer. Arrowstreet says it is building an AI Engineering function for firm-wide productivity and agentic capabilities. The role includes managed LLM inference through Amazon Bedrock, model access and routing across foundation models, Model Context Protocol servers and gateways, authorization, retrieval-augmented generation, autonomous-agent frameworks, and reusable patterns for model and vector-database access.

The controls are as important as the tools. The listing describes default-deny execution, filesystem/IAM/network boundaries, sandboxing, private endpoints, no open internet egress, pre-execution policy hooks, audit logs, and a rule that production modification or deletion requires human approval. It also calls for usage, cost, and adoption dashboards. This is an enterprise control plane, not a prompt library.

An adjacent Senior Software Engineer, Trading Systems listing places GenAI in the front-office systems surface. The current posting describes services used by traders, portfolio managers, and investment teams, integrations with order and execution systems, and workflows covering order lifecycle, execution, allocations, and post-trade processes. It asks the engineer to explore GenAI/LLM capabilities where they improve workflow, automation, or user experience, and treats experience with agentic frameworks as welcome. This connects model-assisted workflow design to a trading-systems team; it does not show that an LLM generates signals, routes orders, or has production authority.

The public role names a Head of AI Engineering as a reporting or organizational reference, but it does not publish that person’s identity. The listing alone does not justify assigning AI-platform ownership to a named executive. Public LinkedIn pages and recruiting posts add personnel and timing signals, but they do not establish investment use or production ownership.

Florent Monthel’s public GitHub account adds an individual engineering signal. The current public claude-code-pd-protection repository describes a Claude Code guardrail layer with production-target classification, deny-write/allow-read policy, human approval for unknown targets, AWS/host/database inventory, MCP interception, cache daemons, and OpenTelemetry/SIEM audit paths. Its public files include hooks for Bash and infrastructure/data MCPs, a read-only inventory MCP server, managed settings, tests, and telemetry configuration. The repository uses example inventory endpoints and contains no observed Arrowstreet credentials or firm-specific datasets in the reviewed material. It is evidence of Monthel’s publicly visible technical vocabulary and a recruiting/team signal, not evidence that the repository is an Arrowstreet system, that he is Head of AI Engineering, or that it is deployed there.

The adjacent Senior AI Security Engineer listing exposes the security-side design vocabulary: agent identities, delegated authorization, short-lived credentials, least privilege, signed requests, tamper-evident audit logs, tool isolation, prompt-injection and data-leakage threat models, and AWS/Azure controls. The direct Workday posting is currently unavailable, so this remains an indexed recruiting copy rather than a live official specification. The AI-platform listing also names Azure OpenAI alongside Bedrock, MCP registries/gateways, RAG, inference routing, observability, usage/cost dashboards, VDI tooling, and M365/Excel integrations. These details deepen the intended platform/security model without proving deployment.

The title-blind personnel sweep adds several distinct public surfaces. Ronald Lee’s personal site self-reports a current Senior Simulation Developer, Research Systems role at Arrowstreet, building tools for quantitative researchers and using AI to automate tasks. Forrest Bicker’s public resume self-reports a Quant Researcher role at Arrowstreet beginning in 2025, prior D. E. Shaw experience, and earlier machine-learning research under George Montañez at Harvey Mudd; it does not assign him AI-platform ownership. Xingyi Zhu’s public LinkedIn profile shows an Arrowstreet affiliation and a post about discussing data, technology, and AI in investment decision-making at an MFA conference, but the public page does not expose a current title. These are personnel and lineage signals, not proof of a named AI team or live model. The expanded Arrowstreet source note preserves the job, platform, Boston-recruiting, and conference routes separately.

Arrowstreet’s official Quant Finance Symposium announcement adds a conference-discovery route: the firm described a November 2025 event with keynote speakers, panels, workshops, and networking. No recording or AI-specific session is exposed in that post. The firm’s current homepage separately reports $344 billion-plus in AUM, 440-plus clients, and 500-plus employees as of June 30, 2026, alongside advanced data science and high-performance computing. These are current firm-reported scale and technology statements, not evidence of model inventory or performance.

A title-blind Trading Talk episode with Bill Tilford, published November 28, 2023, adds a former-personnel media route. The publisher describes Tilford’s Arrowstreet and CC&L experience and frames the 40-minute discussion around quantitative investing and building a successful quant team. A local timestamped ASR/diarization capture now recovers the discussion: Tilford describes historical CC&L/Arrowstreet collaboration and joint research, building proprietary risk models, using alternative data across different horizons, guarding against overfitting, and hiring across coding, mathematics/statistics/data science, and finance/economics. The episode is useful for historical practitioner and team context; it does not disclose a current Arrowstreet AI system, model inventory, current affiliation, platform permissions, or performance. The capture note records the timestamp locators and ASR boundary. The Apple and Substack pages establish episode identity/date; Spotify is an alternate distribution route.

The same publisher’s Piper Hoekstra episode, published January 13, 2023, adds historical CC&L Quantitative Equities practitioner evidence. Hoekstra describes a daily statistical model that constructs signals and risk/return forecasts, an optimization and portfolio-construction process, separate research/trading/process-management/systems responsibilities, ML use in data processing and signal construction, dataset checks for errors and survivorship bias, and a staged research-to-production model-update process. The capture note records timestamp locators from local ASR/diarization. This supports historical CC&L process and personnel context; it does not establish current CC&L employment, Arrowstreet systems, current model ownership, data rights, permissions, or performance.

Additional personnel traces include Ties de Kok’s public CV and research archive, which records Arrowstreet quant-research employment and separate public work involving GPT, fine-tuning, LLM textual analysis, NLP, and XBRL through 2026. The record supports public technical adjacency; it does not establish that the research was performed for Arrowstreet or that any resulting code reached production. A HackMIT-sponsored project adds an external GPT-4/LangChain prototype signal, not an internal platform disclosure.

Sophia Guan’s personal site adds a distinct MIT talent-pipeline route. She describes herself as an MIT undergraduate studying AI and mathematics, an MIT Sloan undergraduate researcher using multimodal models to study idea dissemination across short-form content, and an incoming Arrowstreet Quantitative Developer Intern on the Research Team. This connects a public multimodal-research profile to an intended Arrowstreet research-team internship, but the page is self-authored and future-tense: it does not establish completed employment, a current title, reporting line, transfer of the research, a firm-owned multimodal model, or production use. See the capture note.

The updated worker-layer audit finds a named recruiting signal around the AI-platform team, several current data/software/quantitative workers, and broader unverified organizational rosters. These signals show an operating layer around the platform build. They do not establish that any named person owns a live agent, a model-training program, or an AI-derived investment signal.

Key public people

Person / role	Public evidence	What it supports	Boundary
Florent Monthel	Public recruiting post asks his network for a Senior AI Platform Engineer for “our team”; his public GitHub repository describes Claude Code production-data guardrails	Named recruiting signal plus a public technical artifact covering production safety, inventory classification, MCP, human approval, and audit telemetry	Neither source establishes his formal title, reporting line, Arrowstreet deployment, or ownership of production AI
Ties de Kok	Public CV records Arrowstreet quant-research employment; his research archive contains separate GPT/LLM work	Current public quant-research affiliation and separate AI/ML research	The public record does not tie the GPT/LLM work to Arrowstreet
Sophia Guan	Personal site describes MIT AI/math study, MIT Sloan multimodal-model research, and an incoming Arrowstreet Quantitative Developer Intern role on the Research Team	MIT multimodal-research and Arrowstreet talent-pipeline signal	Self-authored, future-tense page; no completed-employment confirmation, internal model, data transfer, reporting line, or production evidence
Alex Herron	LinkedIn profile identifies an Arrowstreet software-engineering role; a NeurIPS record identifies external conversational-AI research	AI/ML-adjacent engineering signal	The public record does not connect the external research to Arrowstreet
Amy Wong	Public recruiting post describes hiring across AI Platform and AI Security	Recruiting and organizational-intent signal	Talent acquisition activity is not evidence of technical ownership

The Org adds an unverified candidate layer. Its Data and Engineering roster, Investment Services roster, and Quantitative Research roster are explicitly marked “Unverified,” and displayed counts changed across captures. The disqualification loop excludes the rostered names and counts from the public personnel record. They remain an internal validation queue, not confirmed employment, hierarchy, headcount, or AI ownership.

Current LinkedIn evidence, an official Arrowstreet biography, a paper, a conference appearance, a patent, or a first-party recruiting artifact is required before any individual is described as an AI practitioner or owner.

The worker-layer implication is organizational rather than competitive. Arrowstreet’s public record shows several interfaces that an AI platform would need to serve: research data, software and cloud engineering, investment systems, quantitative research, portfolio-management engineering, security, and talent acquisition. The public record does not show how those interfaces are governed, which teams are active users, or whether the platform has moved beyond hiring and design.

What Arrowstreet is not disclosing

The public record does not say how many models are in use, whether Bedrock is already serving production workloads, which investment teams have access, whether agents touch research code or portfolios, what evaluation set governs model changes, or whether any AI output contributes to live positions. The platform architecture is specific; the investment use case remains largely private.

Public-evidence boundary: Arrowstreet’s public GenAI engineering disclosure is architecture-focused. The public strategy described in the role is to standardize access, tools, permissions, telemetry, and safe execution before agents reach sensitive workflows. This is an architecture signal; it does not establish an AI-alpha disclosure.

CFM: model training, text research, and prediction-services infrastructure

Strategy visible in public

CFM’s public record also has a dated, title-blind history behind its current AI language. A 2020 Risk.net Quantcast page identifies Jean-Philippe Bouchaud and describes agent-based models as synthetic market systems that could serve as scenario generators for risk managers; its episode index separately marks machine-learning functions at CFM. The recovered SoundCloud recording adds timestamped detail: Bouchaud describes price, fundamental, weather, inventory, ML, and NLP inputs as possible signal sources (03:04–04:02); treats agent-based models as scenario generators for investigating unstable feedback loops (10:00–22:47); names portfolio construction, model allocation, execution, and signal generation as ML application areas (31:30–35:15); and describes cross-validation and adding nonlinear models alongside simpler working models (31:30–36:15). The term “agent-based model” is kept separate from modern LLM-agent terminology. A 2023 CFM-hosted Hedgeweek interview identifies Benjamin Roy as CTO and describes dataset-qualification protocols, alternative-data processing, ChatGPT-style tool testing, expanding data-science recruitment, and the gap between a proof of concept and daily production use. These dated routes add temporal context, but disclose no model inventory, corpus, benchmark, production permissions, or investment result. The capture note and audio recovery note preserve the source and boundary distinctions.

CFM’s current approach page places technology inside the full investment lifecycle: research models are built from statistical properties of financial instruments; technology implements strategies, trades across markets, and monitors risk and performance in real time. The page says CFM combines machine learning, AI, and cloud computing with more than 12 petabytes of financial and alternative data, and that strategies are tested and piloted before deployment. Its strategies page describes machine learning and AI as tools for using expanding data sources and uncovering new signals. These are firm-level operating claims, not a disclosure of an LLM-based trading strategy.

CFM’s current homepage provides a second first-party route for the same operating substrate: it reports more than USD 30 billion in AUM as of July 2026, 12 petabytes of stored data, and machine-learning techniques applied to global financial and alternative data for trading-signal research. The CFM research archive also lists dated 2026 research posts that can be mined as a continuing firm-media queue. A separate CFM Data Challenge forum provides a historical talent and methodology route: its public discussions cover supervised-learning choices, feature engineering, baselines, rankings, and a moderator’s warning that random row splits can leak same-date information across train and test sets. This is public firm/community methodology evidence, not a current model inventory or proof that challenge methods entered a live strategy. See the capture note.

The current CFM.com approach page makes the operating boundary more explicit: CFM says it integrates machine learning, AI, and cloud computing with more than 12 petabytes of financial and alternative data; tests and pilots strategies before deployment; monitors risk and performance in real time; and retains board authority to override algorithms during extreme volatility. A dated July 23, 2026 CFM research article also says its authors used LLMs to extract views and sentiment from a sample of mid-year broker reports. Its byline names Andre Breedt (Vice President, Research), Christian Dery (Head of Macro Strategy), Bengisu Kaynar (Associate, Macro Alpha), and Philip Seager (Head of Portfolio Strategy). This is direct first-party evidence of an LLM-assisted research/publication workflow and a named research team; the article does not identify the model, prompt/evaluation design, sample size, data rights, production endpoint, or portfolio authority. CFM’s Resilience Lab and academic-partnerships page add a separate environmental-research and academic pipeline: a June 2025 lab studying climate, biodiversity, pollution, and resource scarcity, plus CFM–ENS, École Polytechnique, Columbia, and foundation-funded research routes. These are research and partnership signals, not evidence that sponsored work is deployed in trading. The updated source note keeps the source classes and non-claims separate.

The title-blind media pass recovered a distinct historical CFM-controlled technical route: the firm’s 2018 Challenge Data mid-year solution video identifies Cyrille Delabre as the competition’s presenter and describes a volatility-prediction task built from stock and date time-series data. The walkthrough discusses recurrent/LSTM-style modeling, embeddings for stock/date categories, grouped statistical features, missing-value handling, dropout, cyclical learning-rate scheduling, fold-based validation, and averaging multiple model variants (00:30–02:26; 04:37–09:05). The public CFM-organized challenge specification confirms the financial-volatility objective and five-minute stock/date input structure. This is evidence of a dated public competition and participant methodology, not evidence that CFM adopted the solution, that it entered a live trading system, or that it produced investment returns. See the timestamped capture note.

CFM’s July 23, 2026 mid-year review adds a current research-use example: CFM says it used LLMs to extract key views and sentiment from a sample of mid-year broker reports and used the result as a “wisdom of crowds” anchor for discussing themes and risks through the balance of 2026. The page names Andre Breedt, Christian Dery, Bengisu Kaynar, and Philip Seager as authors or contributors. This is evidence of LLM-assisted document extraction and consensus analysis. It does not identify the model, disclose a trading signal, or establish that the extracted consensus enters portfolio construction or execution.

A scoped public weight-modification example is CFM’s Hugging Face case study. It describes LLM-assisted labeling with Hugging Face Inference Endpoints and Argilla, followed by fine-tuning compact models for financial named-entity recognition. The case study reports GLiNER F1 moving from 87.0% to 93.4% and solutions up to 80 times cheaper than using a large model alone. The task is financial text extraction. It does not establish investment reasoning, portfolio construction, autonomous research, or a general CFM finance language model.

CFM also exposes a research-to-production layer through a public Full-Stack Engineer listing. The listing describes predictive models used by automated trading systems, a Prediction Services platform, production deployment and monitoring, generative-AI agents for code generation, workflow automation and model transformation, AWS infrastructure, LangChain/LangGraph, vector databases, and observability. Because this is a hiring artifact, it establishes the architecture CFM is staffing for. It does not prove that every listed component is live or that the agents generate trading signals.

A separate CFM Prediction Engineer role is now marked filled, but its public description provides a more specific quality-control layer: a model registry, CI/CD, AI tooling, monitoring, independent validation, written quality assessments before deployment, drift detection, and retirement recommendations. It says the models power real trading decisions. Because the page is a filled-role artifact and does not identify the person hired, it supports a firm-designed validation and production boundary, not a current personnel claim or proof that the GenAI components in the separate listing are live.

CFM’s current Machine Learning Researcher role describes a newly formed quantitative-research team building AI-based alpha models and scaling foundation-model training and evaluation on multi-GPU infrastructure. Its ML Platform Engineer role covers data, training, evaluation, deployment, reproducibility, auditability, CI/CD, monitoring, incident readiness, and safe rollout. The current Data & Technology careers board separately lists a GenAI Security Engineer role involving sandboxes, guardrails, agent identity, authentication, RBAC/ABAC, RAG, vector databases, inference, MCP servers, and secure MLOps. Together, these are a detailed intended operating model; they do not prove the roles are filled or that the components are live.

The 2025 CFM profile hosted on the firm’s site reported that CFM hired mathematician Eric Vanden-Eijden to lead an internal AI effort intended to publish academic work and collaborate with researchers on new investment opportunities. A CFM ML Lab announcement names Vanden-Eijden as Head of the ML Lab and Anastasia Borovykh alongside him; Borovykh’s professional site describes current CFM work on ML-based alpha. Florentin Coeurdoux’s professional site describes current CFM research work building and deploying ML models. These are current or recent public lab/personnel signals, not evidence of a firmwide Chief AI Officer or a disclosed LLM production system. A 2025 Cumulus profile names LLMs as a research focus, places them against a large text-data and GPU/cloud substrate, and says CFM blends linear models, deep learning, NLP, and LLMs according to application. The same profile emphasizes that portfolio construction and execution remain separate parts of the process.

A later public CFM post identifies Eric Vanden-Eijnden as Head of Machine Learning and describes his New York seminar, “Generative Models for Quant Finance: From Mathematical Foundations to Forecasting.” CFM says the discussion covered using GenAI to develop signals and models and says the firm is already applying it. A separate CFM-ENS recruiting post advertises one or two two-year postdocs beginning in fall 2026 on diffusion models and generative-AI theory, supervised by Giulio Biroli and connected to the CFM ML Lab. These are current organizational, title, and academic-recruiting signals; no recording, model inventory, filled-hire confirmation, live permission map, or performance attribution is public. See the capture note.

The AIMA interview with Philip Seager is a useful control source. It describes scientific research, significance testing, decorrelation, alternative data, shared research/data platforms, and ML tools, while not disclosing a production LLM-agent system. Reading the interview together with the CFM case study separates the public extraction-model evidence from the broader systematic-investing platform evidence.

The recovered timestamped audio capture adds operating-scale and platform detail to that control case. Seager describes CFM as having more than 100 people in research and more than 200 dedicated investment staff, and discusses applying new approaches and datasets through shared research and data platforms, common tooling, cloud storage and compute, and additional machine-learning techniques (29:32–30:04, 34:52–36:28). These are speaker-reported figures and process descriptions; they do not establish a current organization chart, model inventory, deployment permissions, or investment performance.

A separate June 2, 2026 Traders Magazine interview with Philip Seager is more explicit about research workflow. Seager says CFM has built an agentic system that can take an academic or broker paper, test and code the proposed strategy, and output statistics for researchers to assess whether it could add to a portfolio. He also says a human validation procedure remains in place before implementation. This is a public CFM statement connecting agents to strategy research, but it is still an interview statement: it does not disclose the model, tools, benchmark design, approval permissions, production coverage, or realized investment results.

CFM’s January 2026 “Investing After the AI Honeymoon” interview adds a separate strategy layer. Christian Dery says CFM established a machine-learning lab and is investing in generative AI for sentiment and context extraction, classification, risk management, automated research, and coding productivity. The statement describes intended integration into the investment process, but does not identify a specific production model or agent owner.

A firm-hosted June 15, 2026 Pensions & Investments interview with Dery adds a systematic-global-macro boundary. Dery describes one unified research team spanning predictors, portfolio construction, and execution, and says AI/ML use cases include predictors, classification, network discovery, automated research, extraction from text, video, and audio, LLM-proposed network and classification structures, and agentic approaches that test and code ideas at scale. This is a practitioner strategy and process statement. It does not disclose a model, training corpus, evaluation threshold, live permission boundary, or return attribution.

A newly recovered Masters in Business interview transcript with CFM co-founder, Chairman, Head of Research, and Chief Scientist Jean-Philippe Bouchaud adds a dated leadership account of how the firm frames AI. At approximately 00:21:35–00:23:19, he describes AI as advanced data analysis, highlights text and high-volume data processing, and says CFM created a lab to transfer ML methods toward research while investigating their behavior and interpretability. At approximately 00:26:01–00:27:11, he distinguishes high-frequency data abundance from shorter low-frequency financial histories. At approximately 00:28:20–00:28:40, he describes exploring generative models for very long synthetic financial histories. Later, at approximately 00:35:02–00:39:28, he describes systematic risk forecasting and human intervention when an unexpected event falls outside a model’s knowledge. This is a timestamped practitioner account of CFM’s public research framing, synthetic-data direction, and human risk boundary; it does not identify a model, training corpus, validation artifact, production permission, or AI-attributed return. See the capture note.

Key public people

Person	Publicly visible role or source	What the public record supports	Boundary
Eric Vanden-Eijden	NYU mathematician; public profile says he leads an internal CFM AI effort	Named AI-research leadership signal connected to academic publication and investment research	The source does not publish his current system inventory, reporting line, or deployment metrics
Anastasia Borovykh	Named alongside Vanden-Eijden in CFM’s ML Lab announcement; current professional site identifies CFM ML/alpha work	Current ML/alpha practitioner and lab-builder signal	No public model inventory, reporting line, or deployment metrics
Florentin Coeurdoux	Current professional site identifies CFM research work building and deploying ML models for live strategies	Investment-side ML practitioner signal	The public record does not establish senior AI ownership or GenAI deployment
Christian Dery	CFM article identifies him as Head of Macro Strategy in a discussion of the ML Lab and AI in research	Investment-side sponsor/practitioner context	Not evidence that he leads the ML Lab or a GenAI platform
Andre Breedt; Bengisu Kaynar	Named with Christian Dery and Philip Seager on CFM’s July 2026 mid-year review	Current authorship signal for an LLM-assisted broker-report extraction and consensus-analysis exercise	Authorship does not establish model ownership, production deployment, or investment authority
Jean-Philippe Bouchaud	CFM Chief Scientist and public scientific leader	Scientific research culture, ML/AI framing, and academic bridge	Public commentary does not establish ownership of the LLM extraction or agent stack
Philip Seager	Head of Portfolio Strategy; named AIMA and Cumulus commentator	Public connection between portfolio strategy, data/ML/LLM research, and portfolio construction	Interview and profile do not provide model-level performance attribution
Benjamin Roy	Appointed CFM Chief Technology Officer in the 2022 official release; a June 2025 profile still used that title	Dated technology-leadership signal covering infrastructure, data, risk, and ML/cloud investment	2026 current status is not independently confirmed
Laurent Laloux	Current CFM board page lists Chief Product Officer and co-head of Technology	Current product and technology leadership around data and research/IT capabilities	The public record does not assign him current GenAI program ownership
Marc Potters	Current CFM board identifies him as CIO	Current investment-process, portfolio-construction, risk, and execution leadership adjacent to AI/ML decisions	No public source assigns him direct GenAI ownership

What CFM is not disclosing

CFM’s public record does not specify a complete model inventory, the training corpus or objective for its compact NER models, evaluation leakage controls, production user count for GenAI agents, live agent permissions, or any return attribution to LLMs. The evidence supports a layered record: narrow fine-tuning, public LLM research on text, predictive-model production services, and a named internal AI effort. It does not support collapsing those layers into one autonomous-investing claim.

CFM also maintains a public CFMTech GitHub organization, including a deep-reinforcement-learning portfolio-optimization repository, monitoring, pipeline, and quantitative-research artifacts. The public repositories are useful engineering and research evidence, but the reviewed organization does not expose CFM-owned LLM weights, proprietary prompts, an agent runtime, or a current production research platform. The deep-RL repository’s visible activity is historical and should not be presented as current GenAI deployment.

Balyasny: centralized Applied AI with team-specific research agents

Strategy visible in public

Balyasny’s OpenAI customer case study describes a centralized Applied AI team created in late 2022 with researchers, engineers, and domain experts. OpenAI reports approximately 20 people in that group and approximately 180 investment teams across asset classes and geographies. The firm describes an AI investment-research system designed to reason, retrieve, and act like an analyst, with a central platform and local team customization.

The disclosed control loop is described in operational terms, but remains firm/vendor-reported. Before production, Balyasny says it evaluates models across 12 or more dimensions including forecasting accuracy, numerical reasoning, scenario analysis, and robustness to noisy inputs, using internal benchmarks, tools, and proprietary financial data. The case study says GPT-5.4 is used as a reasoning engine alongside internal models selected by task. It also describes scoped tool and data access, traceable reasoning paths, compliance guardrails, structured user feedback, and an agent-orchestration layer.

An additional OpenAI for Business video post, cross-checked against OpenAI’s finance-solutions page, attributes a newer adoption and tooling account to Charlie Flanagan. The post says Balyasny’s internal AI platform was being used by 97% of the firm, spanning investment research, coding, and back-office automation; it also reports an economic-analysis workflow moving from about two days to about 30 minutes and describes an expansion from coding into earnings-report work. These are vendor-distributed, speaker-attributed claims: the post does not disclose the denominator, measurement window, platform model mix, evaluation fixture, error rate, permission boundary, or investment authority. It adds a current adoption and workflow signal, not independent proof of performance or autonomous capital allocation. See the capture note.

The public use cases are workflow-specific: a Central Bank Speech Analyst, a Merger Arbitrage Superforecaster that continuously updates deal probabilities, and deep-research agents that synthesize filings, research, earnings material, and other documents. OpenAI reports approximately 95% of investment teams using the platform and reports a central-bank scenario-analysis workflow moving from about two days to about 30 minutes. These are customer-case-study figures, not an independent productivity or performance study. The public roadmap names reinforcement fine-tuning, deeper financial-domain orchestration, multimodal inputs, and future model evaluation.

A later OpenAI GPT-5.6 release, published July 9, 2026 and rechecked August 27, 2026, identifies Alberto Da Costa as Principal Engineer, Applied AI at Balyasny and reports a Balyasny evaluation signal for complex financial research, token efficiency, headline categories, and multi-hop tasks. This is medium-high confidence for a vendor-hosted customer evaluation quote and named role; it does not show the benchmark fixture, sample, category definitions, model inventory, production permissions, independent audit, or P&L attribution.

Balyasny-affiliated authors also published a 2026 merger-arbitrage forecasting paper that exposes a more specific research artifact. It describes twelve specialized tool-using agents, timestamped financial-document retrieval, embeddings and reranking, GPT-4.1 or GPT-5 agents, a fine-tuned forecasting model, hindsight-derived reasoning traces separated from the time-constrained test set, and a final verification step against cited sources. The paper reports a held-out class-balanced Brier score of 0.151 across more than 400 large M&A deals. That is a paper-reported research evaluation—not live P&L, investment returns, or evidence that agents can allocate capital.

Other public surfaces add granularity. A Business Insider profile describes an Applied AI research experiment involving market-event forecasting and macro-research feedback. A separate Business Insider roundup names an internal chatbot, BAMChatGPT, reports an earlier staff-use figure, and identifies a data-science executive hire. These media sources are personnel and adoption signals; they do not independently verify the OpenAI case-study architecture or establish autonomous investment authority.

The current personnel pass adds Balyasny’s leadership roster, Peter Anderson’s public profile, and two MIT pages for the March 4, 2026 event: the CSAIL event record and the MIT Career Advising and Professional Development listing. Together they name Charlie Flanagan as Chief AI Officer, Dima Tymofieiev as Senior Research Engineer, and Peter Anderson as Head of Applied AI Research. The MIT CAPD listing advertises discussion of production AI systems at a multi-strategy manager, live demonstrations for portfolio managers, the merger-arbitrage paper as a multi-agent LLM project, and work on making LLMs efficient and scalable in high-stakes settings. This is organizer-published agenda language: it establishes the advertised subject matter and dated role associations, not proof that the session occurred, that a demonstration was delivered, or that the described systems have live investment authority. The arXiv record is separate evidence for the paper’s authors, task, evaluation, and ICML 2026 acceptance; it does not convert the event listing into a production disclosure.

A separate Quant Strats USA 2025 brochure provides a dated personnel and conference route that was not previously in this article’s Balyasny section. The March 11, 2025 New York programme lists Andrew Gelfand as Senior Quantitative Researcher on Balyasny’s Alpha Capture team and places him on a panel about machine learning versus fundamental analysis. The panel description asks whether ML is primarily useful for routine work or also for alpha generation, and how to make ML accessible across an organization. This establishes a brochure-listed role and session association as of that event; it does not establish Gelfand’s current title, attendance, what he said, a Balyasny model, deployment, permissions, or performance. The same brochure contains other firms’ speaker and vendor material, which remains separate evidence rather than a Balyasny system description.

Balyasny’s official How We Work and Investment pages describe Applied AI as a firmwide resource for investment teams. Its April 21, 2026 AI-hackathon report says employees across regions built agentic workflows and automated tasks. These are first-party organizational and experimentation signals; they do not prove that every prototype reached production or received investment authority.

A recovered Bloomberg Tech Disruptors interview, cross-checked against the public Tech Disruptors RSS enclosure, adds a dated practitioner view from April 2024. Flanagan describes a centralized Applied AI function, 48-hour hackathon prototypes, and a central-bank workflow that he says moved from about two days to about 30 minutes after decomposition and roughly four hours of coding (05:33–06:30 media clock). He describes a privately hosted GPT-4 setup via Microsoft in the firm’s own environment, internal retrieval over broker research and management transcripts, source-linked answers, and retrieval tuning for domain language rather than treating the foundation model as a database (09:28–16:10). He also says most resources were then aimed at process automation and research triage, while novel-question generation and proactive analyst/PM insights were being explored (11:56–14:28). These are self-reported 2024 workflow claims; they do not establish current architecture, model ownership, permissions, or investment performance.

A recheck of Balyasny’s official YouTube channel adds a dated, title-blind process signal. In How My Team Develops Investment Theses, the firm’s captions describe starting at the intersection of market-structure understanding and data-science or machine-learning methods, then testing hypotheses in an empirical system that can validate, refine, or refute them through repeated iterations toward operational deployment (00:00–00:50). A separate Gappy Paleologo clip describes quantitative researchers working with portfolio managers on day-to-day portfolio management, performance understanding, and new data (00:03–00:32). These are short, firm-controlled recruiting/practitioner statements; the captions do not identify a model, training corpus, benchmark, permission boundary, or AI-attributed result. The channel capture note records the additional clips and capture boundaries.

A separate March 2024 Hedgineer clip names Carson Boneck and describes a central-data-team onboarding step for incoming data analysts before they join an investment team. The publisher description and a local English ASR recovery emphasize learning the firm’s data landscape, coding ability, curiosity, and cross-functional collaboration (00:00–00:19). The speaker then describes a possible three-month pre-onboarding period for a data analyst joining ahead of a portfolio manager (00:19–00:47). This adds a personnel-to-data-platform interface, but the short third-party clip and automatic ASR do not establish an AI model, agent, deployment, permission boundary, or investment authority. See the capture note.

A title-blind Capital Allocators interview with Dmitry Balyasny, uploaded October 2, 2023, adds a dated operating baseline for Balyasny’s platform. Balyasny describes a staged process for entering new strategies that includes studying comparable businesses, defining an operating blueprint and guardrails, assembling specialist teams, and building research and technology stacks (15:54–18:45). He describes cross-strategy sharing around data monetization, technology, and portfolio construction (25:53–27:40; 39:04–40:20) and says large technology, data, and legal teams work across strategies (70:01–70:48). The recovered caption transcript contains no explicit AI, GenAI, language-model, or agent discussion; that is a negative-control result for this dated interview, not evidence about the firm’s other public records. See the capture note.

A title-blind Meb Faber interview with Vinesh Jha adds a former-PDT and alternative-data-provider route. The publisher identifies Jha as ExtractAlpha’s founder and says he was most recently an Executive Director at PDT; the interview describes ExtractAlpha’s entity mapping, point-in-time and survivorship-bias controls, hypothesis testing, and research on earnings-call NLP, Chinese news sentiment, patent/visa hiring data, and crowdsourced earnings estimates (12:50–26:41; 43:50–49:03; 52:14–64:45). ExtractAlpha’s current Solutions catalog lists US and Japanese earnings-call AI models, Japan news/POS data, an innovation model using patents and visas, digital-demand data, and Estimize. Its team page names Yunan Liu, Qayyum Rajan, and Alan Kwan and records their academic or prior-firm affiliations. These are dated practitioner and first-party provider signals, not evidence of current PDT usage, a specific hedge-fund customer, model weights, data rights, or investment performance. See the capture note.

Key public people

Person	Publicly visible role or source	What the public record supports	Boundary
Charlie Flanagan	Chief AI Officer; quoted in the OpenAI case study	Named central Applied AI leadership and public framing of AI-assisted investment research	The case study does not publish his reporting structure or model-ownership details
Kevin Byrne	Chief Operating Officer; quoted on centralized deployment and local customization	Executive operating-model signal for firm-wide rollout and controls	The case study does not provide independent adoption or control testing
Mike Grimaldi	Current CTO on Balyasny’s official leadership page	Technology leadership and implementation responsibility	No direct public evidence of AI-research ownership
Peter Anderson	Current personal site and MIT event identify him with Applied AI research leadership	Current investment-AI research leadership signal	No official Balyasny biography or model-ownership disclosure
Dima Tymofieiev	MIT event identifies him as a current Senior Research Engineer on Applied AI	Named current engineering practitioner signal	Event material does not disclose model permissions or performance
Carson Boneck	Current official leadership roster identifies him as Chief Data Officer	Data leadership adjacent to Applied AI	No separate public AI-practitioner remit
Damian Miraglia	Official Balyasny post identifies him as Principal Engineer, Data Science & AI and describes a BamGPT Operations Assistant use case	Current first-party practitioner and internal-tool signal	The post does not disclose system-wide ownership, permissions, or investment authority
Hinal Jajal; Chris Pulman	The 2026 merger-arbitrage paper names both; public sources place them in Balyasny-related Applied AI or macro-research contexts	Dated research-project and practitioner context	Paper authorship does not independently confirm current titles, reporting lines, deployment, or model performance

What Balyasny is not disclosing

Balyasny’s public record does not disclose the full model inventory, internal benchmark datasets or thresholds, per-team deployment stages, RFT status, model-training details beyond the customer-case-study description, production error rates, costs, or audited investment-return attribution. The OpenAI adoption and speed figures are vendor/customer-reported. The public record supports a centralized platform with local research workflows and human investment decision rights; it does not establish that the disclosed agents independently allocate capital.

Robeco: named Next Gen Quant leadership and agentic-investing experiments

Robeco’s public record became more specific after a Dutch-language title-blind pass. A June 2025 Investment Officer interview, also available through an English publisher mirror, identifies Mike Chen as Head of Next Generation Quant and describes Robeco experiments with agentic systems. The account discusses specialized agents, scenario simulation, portfolio adjustment, bounded trade execution, human control, a Dynamic Theme Machine, and an internal incubation programme. It states that the experiments were still in development. A public LinkedIn cross-post provides an additional social discovery route, but the substantive account remains publisher-reported.

Robeco’s current Dutch Agentic AI in asset management video-course page, modified August 21, 2026 according to its page metadata, names Mike Chen as Head of Next Gen Research and organizes the public discussion around alpha discovery, research, workflow design, market opportunities, governance, human assessment, and validation. The visible page text describes agents supporting idea screening, signal monitoring, hypothesis testing, portfolio evaluation, paper scanning, claim identification, and methodology summaries. The same page exposes official English, German, French, Spanish, and Japanese routes pointing to the same embedded course asset; these are retained as one canonical media record rather than counted as separate episodes. This is current first-party strategy and educational framing, not a model card or a disclosure of production permissions, investment authority, or performance. The capture note records the locale graph, embedded-video boundary, and follow-up route.

The recovered Robeco Asset Management video, published August 20, 2026, adds a timestamped process layer to that page summary. In English auto-captions, the speaker defines agentic AI as tool use and sequential action subject to human governance and permissions (00:10–00:26), then describes asset management as a chain from evidence gathering and hypothesis testing through investment and client experience (00:29–01:29). The recording explicitly separates model output, signal formation, and extracted alpha, with portfolio constraints, trading costs, risk modeling, and implementation shortfall between them (02:37–03:47). It sketches a paper-to-data-to-signal-to-alpha-model-to-simulated-trading workflow that presents evidence to a human investor, with signal inclusion requiring human authorization (04:00–04:46). Later sections identify data mapping, permissioning, model stability, and accountability as controls, and frame a verifiable investment workflow—with data checks and balances—as the relevant design object rather than a particular frontier model (09:38–10:08, 15:21–16:24, 17:10–17:33). These are automatic-caption paraphrases, not speaker-verified quotations; the recording does not identify a model provider, training corpus, internal benchmark, production telemetry, live portfolio permission, or AI-attributed return. The capture note records the recovery method and boundaries.

Robeco’s own material separates future architecture from current tool status. Its July 2025 AI-course document presents research, portfolio-management, trading, and orchestration agents as a future/scenario design with auditable reports and human governance. Its Q3 2026 Fundamental Equity Quarterly describes two proprietary in-house tools in trial and development: one for ideation, research, and analysis, and another for portfolio management, monitoring, exposure calibration, and administrative work. The same first-party publication says final fiduciary decision-making requires human expertise and judgment. These are first-party status disclosures, not an independent audit of production permissions, model choice, adoption, or investment results.

The official Mike Chen biography identifies his current role and prior PanAgora, BlackRock Systematic Active Equity, Google, and Morgan Stanley experience, along with University of Illinois degrees, MIT Sloan teaching, and public publications. That is useful personnel and academic-lineage evidence; prior roles and teaching do not establish that a particular method was reused at Robeco. A related Argan.ai practitioner article by CEO Reza Kahali discusses small task-oriented agents, context management, verification chains, domain transfer, and combination of AI with classical models. It is a practitioner account from a neighboring investment-technology firm, not evidence about Robeco’s systems or a performance study.

Argan’s current Agentic Strategy product page adds a product-level description: it names Reza Kahali as CEO/Fund Manager and Ali Sardeha as COO/Chief of Staff, and describes NINA as a self-built system combining classical quantitative models with agentic AI for company and macro analysis. The page’s portfolio-size and crisis-positioning claims remain attributed product-page statements; it does not disclose model versions, training data, agent permissions, execution logs, or an independent audit. See the capture note.

Robeco’s official Quant Street archive adds an 11-episode, dated media surface that was previously only tracked at archive level. Episode 5’s English page identifies Iman Honarvar and frames the discussion around AI, data, models, and practical quant use; Episode 6 covers next-generation signals; Episode 11 identifies Harald Lohre and discusses creativity and the human factor. Public Libsyn audio routes for Episodes 6 and 11 were downloaded and transcribed locally on August 26, 2026. In Episode 6, Honarvar describes next-generation signals as an evolution of traditional signals using machine learning, NLP, and alternative data such as text, audio, social posts, and credit-card transactions (00:41–01:11). He also describes combining those sources with analyst reports, management earnings calls, and news to seek information before it is reflected in stock prices (01:45–02:44). The same episode describes incorporating new signals into existing strategies where possible, while using a separate faster portfolio when signal horizon and portfolio speed differ (02:55–04:11). In Episode 11, Lohre places human judgment around hypothesis and investment-thesis design, model oversight, performance and trade monitoring, explanation, and checking model relevance as markets change (00:49–01:45). He also describes confirmation bias as a research risk and prediction/outcome feedback as a way to reassess models (04:17–05:36). These are timestamped local-ASR paraphrases of firm-hosted practitioner interviews; they do not disclose model identity, permissions, production stage, or AI-attributed returns. The capture note records the artifacts, hashes, proper-name corrections, and ASR boundaries. A German Episode 5 video separately provides an automatic-caption route on nonlinearities, infrastructure, domain expertise, and NLP/context-aware language models. The archive also has English, Dutch, and Traditional Chinese regional routes, which are tracked as localization surfaces rather than independent content.

A fresh title-blind pass recovered Robeco’s July 2, 2026 episode on emerging-market quant investing, which was not previously promoted as an episode-level record. The first-party transcript names Jan Sytze Mosselaar and Dijana Kostic and describes a broad emerging-market stock universe, proprietary signals, a research-project and monthly approval funnel, adaptation to changing earnings revisions and momentum, and human handling of data-quality, governance, currency, listing, trading, and implementation issues (01:46–03:07; 10:56–12:48; 15:14–18:41). Kostic and Mosselaar also connect new signals to alternative data, NLP, and ML, while describing model evolution and governance rather than a static black box (08:48–10:55). These are first-party practitioner statements; they do not identify model families, training data, feature definitions, permissions, production endpoints, or independently audited performance. See the capture note.

An adjacent professional-education route adds implementation vocabulary without adding a firm deployment claim. CFA UK’s 2025 three-part AI Agents 101, Automated Fundamental Analysis, and AI-Driven Portfolio Management pages name Carlos Salas, Hanane Dupouy, and Nicole Koenigstein and describe framework selection, multi-agent financial analysis, LlamaIndex AgentWorkflow, automated market tracking, risk controls, and rebalancing. These are event learning objectives and speaker biographies; no public recording, client identity, live authority, or performance evidence was recovered.

The CFA Netherlands event page also exposes a useful personnel graph around the same regional quant-AI conversation. It identifies Robeco’s Kristina Usaite, CFA, and Iman Honarvar, including their APG, Tilburg, Vilnius, Maastricht, and EPFL lineage; Northern Trust’s Guido Baltussen, including prior Robeco leadership and Erasmus/Tinbergen research lineage; and Northern Trust’s Michael van Baren, including prior APG and PGGM quantitative-research roles. These are dated event biographies and academic/employment routes. They do not establish that any person owns an AI system or that event participation maps to a disclosed production workflow.

The same graph produces stronger research evidence when followed into first-party and academic sources. Robeco’s April 2025 Quant Street episode with Iman Honarvar describes scalable AI infrastructure, practical use, and human creativity in quantitative investing; the page exposes metadata and a player, but no transcript was recovered. Robeco’s alternative-data and AI discussion, authored with Kristina Usaite, Laurens Swinkels, and Mike Chen, describes the four Vs of alternative data, machine-learning infrastructure, and the need to convert data into portfolio decisions. These are first-party strategy and workflow statements, not proof of a named production model.

An arXiv paper co-authored by Kristina Usaite and Robeco-affiliated collaborators describes a web-and-knowledge-graph pipeline for automated impact-investing scoring: text collection and filtering, knowledge-graph construction, classifiers, and a reported micro-average F1 of 0.89. It is a public research artifact, not proof that the same system is deployed in Robeco portfolios. Separately, a Robeco historical-factor research summary attributes a pre-1926 U.S. equity database to a team led by Guido Baltussen and describes random-forest and neural-network analysis. That establishes a dated ML research lane, not a current Northern Trust model inventory.

The title-blind video pass also recovered a distinct 2022 interview with Mike Chen while he was Head of Alternative Alpha Research at Robeco. Chen describes alternative data outside financial statements and market history, including satellite, location, web-search, social, speech, shipping, and test-flight data; he discusses occasional public-data collection through web scraping, natural-language processing and machine learning for noisy/high-dimensional data, and the use of a prior investment hypothesis to guide data selection (03:11–08:20; 12:15–16:51; 21:52–24:11). The publisher labels its transcript AI-generated. This is historical role-and-method evidence that complements, but does not replace, Robeco’s current agentic-AI disclosures; it does not establish current model ownership, data rights, production authority, or performance. See the capture note.

The Quant Street Episode 1 recording was recovered from its Libsyn player on August 29, 2026 and locally transcribed as a 393-second Dutch/English recording. In that dated interview, Harald Lohre describes a roughly 20-person quant-research team and a documented, transparent process; he discusses ML questions around nonlinearities, asymmetries, feature importance, and interactions (02:08–02:45; 03:20–04:08). These are local-ASR paraphrases with mixed-language recognition errors, not speaker-verified quotations, and they do not establish a named model, vendor, corpus, current headcount, production permission, agent, or AI-attributed return. See the capture note.

The same recovery pass resolved Episodes 2, 4, 7, and 9. The dated recordings add public descriptions of long- and short-horizon signal adaptation, multi-factor ranking, portfolio-construction algorithms, benchmark/risk constraints, customization, and human exceptions for events the model may not capture quickly. The capture note records timestamps, hashes, and the evidence boundary. None names a GenAI system, model provider, training corpus, agent permission, or AI-attributed performance.

Episode 8 is also now audio-recovered and locally transcribed. Vania Sulman describes an in-house portfolio-construction algorithm that combines expected alpha with risk, transaction costs, liquidity, turnover, and client sustainability constraints (02:19–03:06). She also describes an in-house risk model spanning benchmark-relative static exposures and dynamic exposures such as interest-rate and commodity risk (03:32–05:20). This is dated firm-hosted process evidence from automatic ASR, not a named GenAI, model-provider, training-corpus, permission, or performance disclosure.

Silvercrest: title-blind market-structure and AI-headline discussion

A separate Silvercrest Asset Management podcast, published August 21, 2026 and described as recorded July 15, names Senior Equity Analyst Chad Kusserow and Growth Equity Portfolio Manager Jeff Nevins. Silvercrest’s first-party LinkedIn announcement frames the episode around technology, social media, AI, volatility, and fundamental analysis; Silvercrest’s team index supplies the role cross-check. The automatic captions discuss algorithmic and programmatic trading, factor flips, rapid AI-related repricing, and a collaborative Microsoft Teams workflow for sharing and validating data points (01:31–06:40; 12:10–15:20). This expands the adjacent institutional-manager and title-blind media graph. It does not identify Silvercrest model weights, training data, vendors, permissions, execution rules, or AI-attributed performance, and the captions are not treated as verified quotations. The capture note records the evidence boundary.

For Michael van Baren, a PGGM first-party profile describes his historical work designing and implementing systematic-equity strategies in the Systematic Equities Strategies team. The CFA Netherlands biography separately places him at Northern Trust and titles his session “Machine Learning Applications in Quant Equity Investing.” The historical PGGM implementation description should not be generalized into a claim about Northern Trust’s current systems.

BlackRock: AI Labs, systematic alpha models, and an equities research assistant

BlackRock’s public record is layered rather than singular. Its AI Labs page describes a firmwide research and engineering group applying statistics, machine learning, optimization, stochastic control, and decision theory across retirement, trading, alternatives, and ETFs. The page names Rachel Schutt and Stephen Boyd as co-heads and identifies a wider academic advisory group. A current AI Labs software-engineering role adds generative AI, production deployment, alpha generation, operational efficiency, and responsible deployment to the stated remit.

A separate current Systematic Active Equity quantitative-research posting, retrieved August 27, 2026, asks for work across ML, AI, data science, economics, and large complex datasets. It describes alpha discovery and translation of research into live strategies, with exploration of emerging ML/AI techniques. This is a current hiring signal for the team’s research scope, not evidence that the role is filled, that a named model is live, or that AI-generated output has trading authority. See the source note.

The systematic-investing layer predates the current GenAI wave. BlackRock’s Augmented Investment Management material describes a system created in 2014 to train alpha-forecast models from a large library of financial statements, market indicators, news, analyst reports, and alternative data. Its current Systematic Investing page separately describes LLM use for macro narratives, analyst views, market consensus, and tradable signals.

BlackRock’s IDYN product brief, sourced by the firm to December 31, 2025, adds a dated scale and history claim for the Systematic platform. It says the platform began natural-language processing in 2007 and integrated AI and machine learning in 2013, and describes more than 1,000 investment signals, approximately 50 new signals per year, 40,000 earnings calls, more than 1.5 million analyst reports, and a 200-person team. These are firm-reported product-document claims; the brief does not disclose model versions, training splits, feature definitions, data permissions, live deployment boundaries, or AI-attributed returns. The underlying capture note keeps this Systematic product evidence separate from BlackRock AI Labs and Aladdin.

BlackRock’s March 2026 Systematic investing white paper, retrieved August 28, adds a more explicit description of the data and modeling surface. It says Systematic applies LLMs to broker reports for sentiment estimation, extending NLP that began with simple word counts, and gives examples of foreign-language sources, point-of-interest or infrastructure information, and biographic information about company managers as potential information categories. It also describes text analysis of news, broker reports, and central-bank communications for macro sentiment, plus proprietary models that extract context-aware topics and discussion points. A chart is labeled as more than 100,000 broker reports machine-read monthly, sourced to March 1, 2026. The white paper does not disclose model names, weights, training or evaluation splits, point-in-time controls, data licenses, or independently reproducible results; the reported scale and workflow remain firm-controlled disclosures. See the capture note.

The same paper describes LLM-assisted thematic-basket research and machine-learning applications in private equity and private real estate. In the private-market example, job postings are cited as an alternative dataset and the modeled target is an observable outcome such as an IPO or acquisition probability over a multi-year horizon, rather than a direct return forecast. The paper’s footnotes label the model-versus-universe selection-rate example as hypothetical rather than an actual success rate. It therefore adds a public modelized sourcing and prioritization route, not evidence of a live portfolio, investment authority, or realized performance. Its thematic workflow also references news, conference calls, social media, web searches, and other unstructured sources, with human-in-the-loop LLM synthesis before candidate long/short baskets are formed. These details should remain distinct from BlackRock AI Labs, Aladdin, and the Systematic podcast episodes.

A regional-language Cinco Días interview with Adam Riley, published June 18, 2026, adds a separate public personnel and workflow route. The article reports Riley’s description of a roughly 220-person systematic area, more than 300 data sources, and a mix of analyst reports, earnings-call transcripts, social media, news, product reviews, responsible-investment material, mobility, transactions, and sentiment. It says the team used NLP/ML on analyst reports as early as 2012 and now processes about 6,500 reports per day with context-sensitive models. Riley also describes continuous signal generation followed by cost-, risk-, and constraint-aware rebalancing and manager review, plus a public-web hiring-data example from the Covid period. These are interview-reported figures and examples; no model versions, training data, permissions, or AI-attributed performance are disclosed. The source describes BlackRock Systematic and should not be merged with AI Labs, Aladdin, or a particular hedge-fund mandate. See the capture note.

An April 14, 2026 InvestmentNews interview adds a separate named portfolio-manager route. The publisher identifies Jeffrey Rosenberg as a managing director and senior portfolio manager within BlackRock Systematic. In the embedded recording, Rosenberg distinguishes predictive AI from generative AI, describes fine-tuning and mapping unstructured data to companies, and links those methods to predictive-return models and factor-neutral long-short investing (00:18–00:50; 00:50–01:16; 01:50–02:16). The captions are automatic and the discussion does not identify a model family, training corpus, evaluation split, user permissions, production endpoint, or performance attribution. See the capture note.

BlackRock’s How AI Is Transforming Investing adds a separate systematic-equity disclosure: narrow fine-tuned LLMs, a “Thematic Robot” for equity baskets, human portfolio-manager input, and a model trained on more than 400,000 earnings-call transcripts covering more than 17,000 public companies plus historical market data. These are BlackRock’s own descriptions; the page does not publish the model, split policy, timestamp controls, benchmark artifact, or live-versus-backtest separation. The Macro Decoder separately describes semantic search over broker notes, sequential LLM scorers, forward-looking-statement extraction, long/short classification, and the MATT market-agreement framework. It also describes converting internal meeting summaries into macro-trading research. This is a workflow disclosure, not proof of autonomous trade execution.

The same Bloomberg Intelligence interview with Jeff Shen now has a recovered timestamped automatic transcript, adding detail beyond the publisher’s episode description. Shen describes a 1,500-signal/factor research platform and a job-posting signal he calls “long Python, short Excel” (17:25–19:18); roughly 30,000 datasets and an LLM-assisted process for generating new questions about a dataset (20:11–22:05); and machine-learning regime identification linked to factor performance (23:03–25:14). He also describes applying ML, neural networks, and deep learning to portfolio construction and treats AI as a broader umbrella that includes portfolio construction, risk management, and simulation (31:27–33:42). The discussion names Stanford professor Stephen Boyd in the firm’s academic-collaboration context (28:52–30:09). These are dated, speaker-attributed statements. They do not publish a complete signal or dataset inventory, model weights, vendor contracts, training/evaluation design, data rights, permissions, live deployment status, or AI-attributed performance. See the capture note.

A current BlackRock AI-investing page makes that public description more specific. It separates narrow, curated task tuning from general-purpose chatbots, describes earnings-call reaction forecasting and the “Thematic Robot” for long/short or long-only equity baskets, and names Raffaele Savi, Jeff Shen, Yaki Tsaig, and Taylor Dufour in the displayed Systematic-investing contributor list. The page says human-defined themes, portfolio-manager priors, transparent iteration, and correction remain part of the workflow. The related 2025 “Alpha re-imagined” paper separately lists open- and closed-source model training and firm-reported evaluation context. These are firm-authored disclosures, not a public model registry or independent evaluation: model weights, complete data provenance, leakage controls, current versions, permissions, order authority, and independently audited performance remain undisclosed. See the capture note.

The firm also publicly names an equities research tool: an official Active Investor page calls it Asimov and says it supports broader and deeper equities research and portfolio-level risk adjustment. This is a firm-owned product signal, but the public page does not identify its model family, retrieval sources, evaluation set, user population, permissions, or performance attribution. Asimov, AI Labs, systematic AIM, and LLM-based macro research should therefore remain separate evidence layers. The official Systematic Investing podcast archive and AI archive now reconcile to eight unique episode-level routes with public HTML transcripts; four are net-new beyond the earlier Systematic-only capture. Jeff Rosenberg’s April 2026 fixed-income episode discusses broadening sentiment analysis across central banks and companies, while Ronald Kahn discusses large signal libraries, alternative data, and human review; Jeff Shen’s thematic-investing episode describes NLP over broker reports, conference-call transcripts, and financial news, alongside LLM-assisted theme research. The four additional episodes add BlackRock editorial context on AI-linked alternatives, energy and infrastructure, AI’s investment history, and technology stock-picking, but do not disclose a new model, agent, or permission system. The earlier Rosenberg episode contains no public AI disclosure, which is a useful negative result. These are first-party interview statements and do not establish model identity, training data, live permissions, or AI-attributed returns. See the reconciliation note and machine-readable manifest.

A separate title-blind Flirting with Models interview with Jeffrey Rosenberg, published August 18, 2025, adds detail that is not in the BlackRock AI pages. Rosenberg describes replacing or augmenting earlier bag-of-words sentiment signals with LLM-based sentiment analysis through proposal, backtesting, validation, and signal-diversification or sunset decisions (46:48–48:50). He also describes fixed-income portfolio construction as an optimization of forecasts subject to risk and transaction-cost constraints, and stresses liquidity and market-impact estimates when testing historical ideas (49:10–51:18; 57:03–58:14). BlackRock’s current biography separately identifies Rosenberg as a Managing Director and senior Systematic Fixed Income portfolio manager. This is a dated executive account of a signal and implementation process; it does not identify the LLM, training corpus, evaluation split, permissions, production endpoint, or AI-attributed performance. See the capture note. A June 22, 2026 Bloomberg Intelligence Credit Crunch interview with Jeffrey Rosenberg adds a distinct, title-blind route. Rosenberg separates predictive AI or machine learning from generative AI (06:56–07:23), then describes LLMs as expanding context-aware extraction from broker research beyond ratings and price targets (17:24–19:58). He frames the workflow around sentiment signals, issuer breadth, temporal consistency, and systematic portfolio application; the podcast gives an illustrative scale of roughly 60,000 reports but does not name a model, provider, training corpus, data license, evaluator, or production endpoint. The recovered MP3 and local timestamped ASR are recorded in the capture note; this is a dated practitioner account, not an implementation or performance audit.

BlackRock’s 2024 response to the U.S. Treasury AI request for information supplies an important governance boundary: it describes AI Labs as established in 2018, a newer central hub for GenAI coordination, human review, audit, backtesting, monitoring, and risk governance, and states that BlackRock does not use AI to make autonomous, independent decisions. The 2026 proxy statement separately describes Asimov as an AI Research Platform used to scale equity research within the Portfolio Management Group. A September 1, 2026 S&P Global interview with Wei Li independently adds a named executive account that BlackRock uses both internal AI tools and external LLMs, identifies Asimov as an internal investment tool used by investors, and discusses agent experimentation and process redesign. Together, these sources support an AI research and decision-support layer with stated human controls; they do not disclose Asimov’s builders, architecture, or permission map.

BlackRock also maintains a public GitHub organization. The Blowfish repository is a firm-controlled semantic-search and RAG-evaluation artifact; HOLA exposes hyperparameter-optimization infrastructure; and AladdinSDK exposes public Aladdin APIs and investment-research tooling. Blowfish’s README names Thomas R. Barillot and Alex De Castro as BlackRock research authors. These artifacts establish public technical work and named authorship. They do not establish that any repository powers Asimov, AI Labs, Systematic, or a current investment decision.

An August 11, 2026 recheck adds a separate Aladdin layer. BlackRock’s Aladdin Copilot page describes a generative-AI product available to all Aladdin clients, with permission-dependent access, content filtering, hallucination-risk parameters, and a boundary that it does not provide investment advice or answer outside Aladdin platform limits. Current Aladdin roles add hiring and architecture signals: an AI, Data Science - Aladdin VP role names AI assistants, agentic workflows, RAG, evaluation, monitoring, and responsible AI; an Investment & Trading Compliance AI role names AI compliance assistants, agentic compliance workflows, LLM reasoning systems, knowledge graphs, evaluation, and governance; and an Aladdin Solutions Architect role describes Agentic SDLC delivery across investment and trading lifecycle workflows. These sources establish a client/platform and compliance layer. They do not disclose model inventory, prompt controls, adoption metrics, investment-recommendation authority, or performance attribution.

An August 13, 2026 conference-lane recheck adds an Aladdin architecture signal. The 2025 AI Engineer event schedule lists Brennan Rosales/BlackRock for “Agents in Investment Management: Aladdin Copilot” and describes an enterprise multi-agent platform for federated application development in a controlled and explainable environment. A related LangChain Interrupt recording, uploaded June 12, 2025, identifies Brennan Rosales and Pedro Vicente Valdez and describes production-ready Aladdin agents, LangGraph architecture, a plugin registry, a 50+ engineering-team contribution path, 100+ applications, and daily CI/CD evaluation. Because these are conference-host and recording-metadata surfaces rather than a retrieved transcript or product specification, they support BlackRock’s public platform vocabulary; they do not disclose model inventory, prompt controls, user permissions, adoption metrics, investment-advice authority, or performance attribution.

An August 15, 2026 conference-caption pass adds a separate investment-operations knowledge-app layer. The AI Engineer recording How BlackRock Builds Custom Knowledge Apps at Scale, mirrored locally from official YouTube captions captured July 2, 2026, identifies BlackRock engineering speakers Vaibhav Page and Infant Vasanth and describes document extraction, workflow automation, Q&A/chat, and agentic-system app categories. The presenters use a new-issue/security-setup workflow as an example, then describe a sandbox and cloud-native app-factory architecture for domain experts, with prompt and extraction-template versioning, evaluation datasets, LLM-strategy selection, access control, cluster and cost controls, validations, QC checks, and human-in-loop or four-eyes review. They report reducing complex app delivery from three to eight months to days. This is practitioner testimony from a conference recording, not an independent ROI, security, model, permission, or deployment audit.

An August 12, 2026 paper pass adds a research-publication layer. The ACM ICAIF 2025 accepted-papers page lists “Large Language Model Agents for Investment Management: Foundations, Benchmarks, and Research Frontiers” with five BlackRock-affiliated authors. This is public evidence that BlackRock researchers are publishing on investment-management LLM-agent foundations and benchmarks. The page is not a product disclosure, does not identify an internal platform, and does not show live deployment or investment authority.

A title-blind historical practitioner route adds useful baseline vocabulary. In a February 2021 Promotable recording, the host introduces Justin Sheets as a Vitruvian Capital co-founder/co-portfolio manager and describes earlier Barclays and BlackRock experience. Sheets discusses social and retail-flow signals, earnings-call reading, NLP over filings and social data, multiprocessing, and a stack spanning R, Python, SQL/MySQL, MongoDB, and kdb-style tick data (03:33–04:00; 08:12–10:39; 28:36–33:09). The recording is historical professional education and the employment details come from the host introduction and speaker narrative; it does not establish a current BlackRock title, current Vitruvian deployment, model inventory, data rights, permissions, or performance. See the capture note. This is evidence for the discovery method—non-AI job titles can surface research-stack and personnel routes—not evidence that the described workflow remains current.

Key public people

Person	Publicly visible role or source	What the public record supports	Boundary
Rachel Schutt	BlackRock Managing Director and co-head of AI Labs	Named current AI research and engineering leadership	The public record does not assign ownership of Asimov, AIM, or a specific LLM
Stephen Boyd	Co-head of AI Labs and Stanford professor	Named AI Labs leadership and academic research connection	The public record does not establish operational ownership of a production system
Raffaele Savi	Global Head of BlackRock Systematic	Current systematic-investing leadership adjacent to AI, macro, and signal research	Public pages do not assign him ownership of an LLM or Asimov
Jeff Shen	Co-head and Co-CIO of Systematic Active Equity	Current systematic-equity leadership adjacent to AI-enabled research	No model inventory or GenAI reporting line is public
Ronald Kahn	Global Head of Systematic Investment Research	Current research leadership in the systematic platform	The public record does not connect him to a specific GenAI product
Thomas R. Barillot; Alex De Castro	Named as BlackRock research authors in the Blowfish repository	Firm-controlled public semantic-search/RAG-evaluation authorship signal	Repository authorship does not establish current AI Labs employment scope, Asimov ownership, or production investment use
Mike Pyle	Deputy Head of the Portfolio Management Group; official Asimov transcript speaker	Current portfolio-management leadership and public Asimov disclosure source	He refers to “my colleagues”; the source does not identify him as builder or product owner
Nish Ajitsaria	Co-Head of Aladdin Product Engineering and executive sponsor for AI; official biography	Current enterprise AI sponsorship and platform-engineering signal	No public attribution to AI Labs, Asimov, or a specific investment system
Rob Goldstein	COO; official biography	Current operating/technology leadership associated with public Asimov disclosure	No evidence that he built or technically owns Asimov

A historical Quantopian route provides a useful non-GenAI comparison point. In a 2019-uploaded recording, the host introduces Jessica Stauth as a Quantopian research leader and describes selection of community algorithms for portfolio allocation. Stauth then describes deduplication of cloned or correlated strategies, common simulations with cost models, feature extraction from returns/positions/transactions, investability filters, risk-model checks, and leakage/overfitting controls (02:03–06:18; 07:58–12:15; 30:39–36:32). A dated QuantCon speaker page and related SSRN paper provide separate historical role and publication cross-checks. This is a Quantopian-era research workflow, not evidence of current Stauth employment, current Quantopian operations, or any current manager’s model architecture. See the capture note.

What BlackRock is not disclosing

BlackRock’s public material does not provide a unified AI organization chart, model inventory, Asimov architecture, prompt or retrieval controls, evaluation results, user counts, model-routing policy, or return attribution. The sources support firmwide AI research, systematic signal modeling, and a named equities-research tool; they do not support treating those as one autonomous investment engine.

WorldQuant: AI leadership plus a dated agentic-PM role artifact

WorldQuant’s 2026 Portfolio Manager, Agentic Systems role artifact was explicit about the intended intersection of portfolio management and agentic systems. The posting described a live trading book, risk and P&L responsibility, planning algorithms, tool use, memory, reflection, collaborative reasoning, reinforcement-learning hyperparameter tuning, deep-learning model development, custom agentic workflows, and human-in-the-loop checks against quantitative-trading acceptance criteria. The direct URL no longer resolves as a live vacancy on the August 10 check, so its evidentiary status is dated hiring intent rather than a current open-role signal.

WorldQuant’s first-party leadership page names Paul Griffin as Co-Chief Investment Officer and Chief Science Officer and says he leads the firm’s AI initiatives. That is a current firm-controlled leadership signal. It does not establish that Griffin owns the specific agentic-PM role, a named model, or a production deployment. WorldQuant’s May 20, 2026 perspective says AI is deployed across teams and that agentic systems may help generate ideas, while also stating that humans retain accountability. The article is firm thought leadership and expressly disclaims a relationship to a specific investment strategy or product.

The current WorldQuant careers index lists related AI software, WQBRAIN AI research, deep-research, and LLM/AI-agent roles. The current AI Software Developer role describes company-wide applications for information access, knowledge synthesis, strategic decision support, LLMs, vector databases, and finance-oriented AI products. The Senior AI Software Developer role uses similar London applied-AI language. A separate Quantitative Research Intern (LLMs & AI Agents) role places LLM/NLP/prompt-engineering experience alongside alpha research and signal analysis. The WQBRAIN AI Researcher role describes transformers, reinforcement learning, generative AI, fine-tuning, signal generation, and alpha enhancement on BRAIN. A current Lead Python Engineer, AI/ML Systems role describes AI/ML systems, LLMs, agents, data pipelines, and deployment across the business-facing technology organization. A distinct current AI Scientist posting places the role in the Artificial Intelligence team and explicitly names LLM-based agentic technology for developing and testing trading signals and algorithms, signal compression and combination, deep-learning architectures, model frameworks for investment professionals, and collaboration with portfolio managers and researchers. This is a more specific trading-research hiring signal than a general AI-software posting, but it remains hiring intent: the page does not establish a filled role, model/provider, evaluation design, permissions, live authority, or performance. See the capture note.

WorldQuant’s June 1, 2026 How AI Broadens Opportunities for All Quant Researchers adds a firm-owned research-culture statement. It describes AI agents helping individuals digest financial documents, generate hypotheses, run simulations, refine algorithms, and pursue many research threads; it also reports International Quant Championship registration growth from 80,000 students in 2025 to more than 156,000 in 2026. Gerry Beatty’s December 17, 2025 agentic-teammates article separately emphasizes agent-to-agent standards, oversight, training, and governance. Both pages carry WorldQuant’s standard disclaimer that they are not tied to a specific investment strategy or product. They support a public strategy and talent-development signal, not deployed investment authority.

A distinct March 23, 2026 leadership article by Igor Tulchinsky adds a dated account of how WorldQuant frames AI-enabled research participation. It reports that the 2025 International Quant Championship drew nearly 80,000 participants from 11,000 universities across 142 countries, and says individual participants used AI as a primary research partner to scan research, generate hypotheses, run simulations, and refine strategies. The article explicitly frames this as an IQC and education/research-culture example rather than a description of a particular WorldQuant investment product, and its disclaimer repeats that boundary. It therefore adds a leadership-authored talent and research-culture route, not evidence of a WorldQuant production model, model ownership, investment permissions, or returns.

WorldQuant’s May 8, 2026 IQC launch page provides a distinct dated platform and recruiting snapshot. It says the competition uses BRAIN to build and test quantitative-finance algorithms on historical market data, and reports more than 140,000 registrations from more than 150 countries, a $100,000 prize pool, more than 80 offers through the competition ecosystem, and a BRAIN footprint of 125,000 datasets, more than 500,000 users, and around 13,000 research consultants. The page quotes Igor Tulchinsky and Chief Strategy Officer Nitish Maini and mentions AI as a source of new research possibilities and talent pools. These are WorldQuant’s dated self-reported platform and talent figures; they are not an independent audit or evidence of a live investment model, permissions, or performance. The May snapshot is retained separately from the later June registration figure rather than reconciled into one total.

A title-blind personnel pass adds Zsombor Koman’s November 2025 profile, which identifies him as Vice President, Research, Emerging Technologies and describes a current focus on AI and large language models. This is a named research-surface and editorial-remit signal. It does not establish model ownership, a reporting line, a production endpoint, or investment authority.

The May 2026 Milken panel page and its public transcript add a separate executive disclosure. Andreas Kreuz is identified as WorldQuant’s Deputy Chief Investment Officer, and describes AI use across signal research, portfolio construction, and execution, plus ongoing work on multiple agents, human-defined objectives, accountability, infrastructure, and security. The transcript does not disclose model names, providers, evaluation thresholds, permissions, override rates, or AI-attributed P&L. It is a public executive statement, not an independent deployment audit.

These postings define a public capability-building and hiring thesis. They do not establish that the portfolio-manager role is filled, that the described agentic systems are in production, that reinforcement-learning updates affect live decisions, or that any returns are attributable to the systems. The public record names an AI-initiatives leader but does not provide a complete agent-platform roster or assign that leader ownership of the specific systems described in the vacancies.

Key public people

Person	Publicly visible role or authorship	What the public record supports	Boundary
Paul Griffin	Co-Chief Investment Officer and Chief Science Officer; official page says he leads WorldQuant’s AI initiatives	Current firm-controlled AI leadership signal spanning investment, research, and technology	No public assignment to the agentic-PM vacancy, model ownership, or live deployment
Andreas Kreuz	Deputy Chief Investment Officer; author of WorldQuant’s May 2026 AI perspective	Current investment-leadership authorship for public discussion of AI, unstructured data, agents, and human accountability	Thought-leadership authorship does not establish system ownership or trading authority
David Rukshin	Chief Technology Officer	Current technology leadership adjacent to AI/LLM infrastructure	No public AI-specific remit or model ownership
Gerry Beatty	Chief Technical Advisor	Current technology-strategy signal	No public AI-specific remit or deployment evidence
Zsombor Koman	Vice President, Research, Emerging Technologies; author of a WorldQuant profile discussing AI and large language models	Named emerging-technology research signal and public LLM focus	No public model ownership, reporting line, production endpoint, or investment authority

What WorldQuant is not disclosing

The public record does not disclose agent architecture, model providers or weights, training data, reward definitions, evaluation thresholds, live-book permissions, override rates, or model-specific P&L. The appropriate classification is detailed agentic hiring intent with a direct portfolio-management interface, not confirmed autonomous investing.

WorldQuant and WQU video surfaces — August 20, 2026

The public record also includes a substantial WorldQuant University (WQU) and WorldQuant Careers video surface that is easy to miss if discovery is limited to hedge-fund podcasts or titles containing “AI.” The official WQU homepage links an official YouTube channel and Facebook page. An inventory captured on August 20, 2026 identified 72 videos on the channel. The WQU news archive is another discovery route: its August 14, 2026 Boken Lin profile embeds Boken Lin: The Risk of Success, and YouTube metadata identifies WorldQuant University as the uploader. The channel mixes applied-AI course material, quantitative-finance education, alumni speakers, and institutional content; these routes should be treated as education, recruiting, and media evidence rather than as evidence of WorldQuant asset-management deployment.

Selected routes include The AI-Driven Leader, Applied AI: Interdisciplinary Approach, Applied AI: Containing a Quant Contagion, What Is a Financial Engineer?, Most Common AI Mistakes, When AI Fails: The Human Element, and The Accidental Hedge Fund Manager. The selected routes expose dated titles, speakers, descriptions, and English auto-captions; they add searchable vocabulary around applied AI, decision-making, risk, financial engineering, and quant careers, but do not disclose model weights, live permissions, trading performance, or investment authority.

WorldQuant’s Learn2Quant series is a separate firm-owned educational route hosted by Nitish Maini and covering alpha ideas, data categories, diversification, risk, and advanced research. WorldQuant Careers also has public Facebook video routes: the 2025 International Quant Championship announcement, the registration post, and the BRAIN recruiting/community explainer. These indexed Facebook routes describe competition mechanics, prizes, participation, and recruiting/community language. Direct Facebook archive inspection was login-gated during this pass, so they are page-level discovery records rather than transcript evidence. The firm channel also exposes WorldQuant BRAIN: Sign Up Today, an official video-level recruiting route that complements those Facebook pages. YouTube publisher metadata identifies the uploader as WorldQuant; the video is useful for tracing BRAIN and talent-funnel messaging, but it does not disclose a model, live-book permission, or investment result.

The separate WorldQuant firm YouTube channel, exposed by the firm’s official homepage and also reachable as @WorldQuantCareers, adds 94 indexed videos beyond the WQU archive. Its title-blind routes include Quantcepts: How Quants Can Partner with AI, How to Assess an Alpha, Sentiment Data, David Rukshin: WorldQuant is a Technology Company, and Nitish Maini’s Chess.com interview. The archive adds public terminology, a named technology-leadership route, BRAIN/IQC recruiting context, and a path for timestamped caption review. It does not establish model ownership, production deployment, live-book authority, or performance attribution. No separate corporate WorldQuant Facebook video archive was verified beyond the WQU and WorldQuant Careers pages.

The August 20 refresh added five caption-audited corporate routes that were not yet represented in the ledger: What is an Alpha, Price Volume Data, Carlos Nocito’s data-team profile, the Chess.com/IQC partnership, and Igor Tulchinsky’s Milken interview. Together they add public vocabulary around alpha construction, price/volume features, internal data generation, ML application in a data team, talent recruiting, and technology culture. They remain firm education, personnel, partnership, or dated leadership evidence—not proof of a named production model, live authority, or performance.

The full route map and capture boundaries remain in the repository source ledger. The key boundary is unchanged: WQU education, WorldQuant Careers recruiting, and firm-owned media can reveal public terminology, talent pathways, and stated research culture, but they do not prove that a named speaker owns a production system or that an educational or recruiting claim maps to a live portfolio.

Numerai: public AI-scientist workflow, predictive LLM, and MCP-enabled tournament automation

Numerai is not a like-for-like asset-manager peer, but it is directly relevant as a public control case for AI-mediated quantitative research. Its current homepage centers the workflow on architecting an AI scientist, installing an MCP server, and using coding agents against Numerai’s obfuscated stock-market dataset. The same page describes contributors submitting machine-learning predictions that are combined into a meta model for a quantitative global equity market-neutral hedge fund. This is a platform disclosure; it is not evidence that any outside agent receives independent authority over Numerai’s portfolio.

The NumerCon 2026 recap, accessed August 12, 2026, adds a language-model feature layer. Numerai says its Predictive LLM is an 8-billion-parameter model trained on more than 1 million articles to turn news into structured predictive signals across Numerai’s universe, and that it powers the Faith dataset. A Faith release forum post describes 186 new features in Dataset V5.1 and warns about early-era sparsity. These sources support a public model-generated feature release. They do not publish the model weights, article corpus, leakage controls, live portfolio attribution, or independent evaluation record.

The same NumerCon recap and Numerai’s public example-scripts repository describe open-source agent skills for experiment design, testing, evaluation, diagnosis, iteration, reports, and valid submissions. Numerai’s API and MCP documentation says the MCP uses the stateless 2026-07-28 protocol, supports Codex CLI, Cursor, and Claude Code, and exposes tools for tournament information, creating and uploading models, checking submissions, and GraphQL access. The docs also require scoped API keys and caution against overbroad permissions. This is public tooling and permission-surface evidence, not proof that agent outputs are accepted without user-configured credentials or human oversight.

A distinct February 23, 2026 Flirting with Models episode adds a founder-level account of how those surfaces are intended to operate. Richard Craib describes an MCP for agentic model management and an AI-scientist/AutoML loop that can spend roughly three days on a research task before returning a model (42:30, 44:50). He also describes a roughly 20-person team using agents to test whether trial data from a vendor is additive, and a Predictive LLM adapted from an open-source LLM to turn internet news into stock-prediction features (48:40, 49:54). The episode is a dated, speaker-reported workflow disclosure that complements Numerai’s first-party model and MCP documentation; it does not expose the research-agent code, model/version inventory, vendor datasets, evaluation fixture, permission configuration, production endpoint, portfolio authority, or independently audited performance. See the capture note.

The same recap also exposes five separate NumerCon video routes: Genesis, Data, Signal, Risk, and Singularity segments attributed to Richard Craib, Michael Phillips, Noah Harasz, and Michael Oliver. Caption review of all five recordings adds direct, time-linked public statements. Phillips says the Faith features were generated by an 8-billion-parameter Numerai Predictive LLM trained to turn news into stock-level predictive signals; Harasz says the Numerai MCP can let agents create and upload models, run validation diagnostics, and review performance; and Oliver introduces Numerai Risk alongside a claim about predictions from thousands of models on thousands of stocks. Craib’s meta-model route describes the stake-weighted combination of submitted signals. These are firm-published event statements, not independent audits: the recordings do not disclose weights, training corpus, leakage controls, evaluation design, accepted-submission rates, portfolio attribution, or autonomous capital authority. The full route map and caption boundaries are catalogued in the Numerai/WorldQuant/QRT source note.

The title-blind/person-specific queue also recovers a November 2020 Chai Time Data Science episode with Richard Craib, with an iVoox episode page that links Numerai Signals and the contemporary “Master Plan” material. The dated episode description centers the Numerai story, hedge fund, competition, and Signals launch; Numerai’s Signals announcement separately documents the historical any-model/any-dataset signal-submission design. This adds platform-evolution context before the current AI-scientist, Predictive LLM, and MCP surfaces, but does not establish their later architecture, a complete transcript, current deployment, permissions, or independently audited returns. See the capture note.

Two older video routes add historical research-process and practitioner context. In a Weights & Biases Deep Learning Salon uploaded October 28, 2020, Richard Craib describes Numerai’s public obfuscated-data interface, finance-specific temporal overfitting, feature neutralization, and combining less-correlated models (04:22–04:57; 05:10–06:15; 06:13–08:24; 09:12–11:18; 23:10–24:09). A separate first-party Numerai Office Hours S02E05, uploaded September 22, 2020, presents a participant identified as UKI/Yuki. Its description states that the participant operated a Japan-focused, market-neutral stock-selection approach using machine learning alongside traditional quantitative/statistical methods and links a public model page. The queue had incorrectly assigned this episode to Craib; the source note corrects that attribution. These are historical founder and practitioner accounts, not current model inventories, performance audits, or proof of Numerai’s internal deployment. The repository capture note preserves the caption and identity boundaries.

The recovered Numerai Office Hours S01E07 adds an anonymous participant-level route. “ZEN” describes software engineering, leading an unnamed company’s AI department, and hands-on work for clients (00:54–03:20); the same session discusses validation discipline and Meta Model Contribution as an incremental-information measure (05:49–07:38; 22:27–27:28). Because the guest’s surname and employer are not disclosed, this must not become a named personnel or firm AI claim. See the capture note.

The recovered Epicenter episode 191, published July 11, 2017, adds a historical founder account of Numerai’s architecture. Craib describes obfuscated financial data, prediction submission rather than model transfer, and learning a meta-model from long histories of contributor predictions (03:29–05:00; 22:03–24:29; 24:47–30:34). This helps establish the earlier design logic behind the later platform, Predictive LLM, and MCP disclosures, but it does not establish current model weights, data rights, agent permissions, portfolio authority, or performance. See the capture note.

The July 29, 2020 Outlier Ventures interview adds a second dated founder account. Craib discusses the machine-learning tournament, the contrast between Numerai’s internal model and crowd models, the value of uncorrelated models, and sourcing predictions without taking participants’ models (10:03–16:33; 18:02–20:52). It is historical founder evidence and does not establish the current Predictive LLM, MCP, agent permissions, portfolio authority, or performance. See the capture note.

The distinct Epicenter Episode 348, published July 14, 2020, adds a later founder account before the current AI-scientist and Predictive LLM material. Craib describes staking and meta-model contribution (04:52–11:29), an obfuscated dataset of approximately one million rows and 310 feature columns (19:13–21:19), and the rationale for limiting human interpretation of hidden feature groups (23:07–24:15). He also describes Numerai Signals as a route for external stock predictions and uses social-media sentiment as an example of a signal that might become more useful when combined with other information (42:56–47:51). This is historical founder evidence; the approximate figures and illustrative examples are not a current dataset or production-model disclosure. It does not establish current model weights, data rights, agent permissions, portfolio authority, or independently audited performance. See the capture note.

The Token Summit II — Numerai recording adds a short 2017 founder presentation on narrow AI, obfuscated data, staking, and a proposed stake-backed GPU/EC2 access mechanism (00:14–06:42). Craib describes approximately 530,000 rows and 50 features and reports an early staking observation (02:52–05:27); these are approximate speaker-reported historical statements, not a current dataset, launch record, benchmark, or performance audit. The recording does not establish current AI-scientist/Predictive LLM/MCP architecture, permissions, portfolio authority, or performance. See the capture note.

The distinct Token Summit I panel, held in the first Token Summit’s 2017 context, adds a separate founder account of Numerai’s external research network. Craib describes supplying stock-market data to data scientists who return machine-learning predictions to the fund, reports more than 16,000 connected data scientists, and explains staking as an incentive and anti-Sybil mechanism (05:35–07:34). These are speaker-reported historical claims; they do not establish current contributor counts, token economics, model architecture, permissions, portfolio authority, or performance. See the capture note.

The canonical Lex Fridman Podcast #159 page and direct recording, published February 7, 2021, add a dated founder interview spanning Numerai, hedge funds, AI in stock trading, and Numerai data. The publisher outline places the Numerai discussion at approximately 37:00 and the AI-in-stock-trading discussion later in the episode. Two shorter queue items are excerpts from this same interview and are retained as derivative provenance, not separate founder conversations. The route supplies historical context between the 2020 Signals account and later first-party NumerCon/AI-scientist disclosures; it does not establish current architecture, permissions, or independently audited returns. See the capture note.

The Other Life Episode 162 publisher page, Apple listing, and recording, published December 14, 2021, add a distinct founder interview focused on finance, truth, crypto regulation, artificial intelligence, and Numerai’s crowdsourced/tokenized hedge-fund model. Its show notes link Numerai’s contemporaneous performance announcement; those comparisons remain company-attributed, not independently audited evidence. This route helps time the public narrative between the 2021 performance report and later product surfaces, but it does not establish current architecture, permissions, or AI-attributed results. See the capture note.

Two additional historical captures refine the lineage without changing the current-state conclusion. In a 2020 Weights & Biases Deep Learning Salon, Richard Craib discusses obfuscated data, backtest-to-live deterioration, time-bucketed validation, feature neutralization, and combining less-correlated models (04:22–11:18, 23:10–24:09). A separate Numerai Office Hours S02E05 recording identifies a Japan-based participant presented as UKI/Yuki and describes a market-neutral, machine-learning and statistical-modeling context. The participant’s surname and current employer are not established, and the recording is not evidence of Numerai staff or internal strategy. Both are historical, caption-assisted routes; neither establishes current model weights, agent permissions, production authority, or independently audited performance. See the capture note.

Two more first-party Numerai recordings add named personnel and infrastructure context. In the Q4 2020 Fireside Chat, the introduction identifies Anson Chu as CTO and Richard Craib as founder and CEO; Chu’s engineering update discusses data-pipeline reliability, dashboards and diagnostics, API evolution, validation-data cadence, and model-slot infrastructure (00:21–00:34; 14:15–17:28; 26:29–28:26). In Office Hours S02E06, Numerai’s description identifies Michael Phillips as a data scientist, links prior leadership of a DecisionIQ data-science team that put thousands of ML models and data pipelines into production, and links his Meta Model Contribution article and public code. The recording itself covers walk-forward validation, multiple holdouts, metric choice, genetic search, PCA, and regime-overfitting concerns (06:35–09:54; 14:03–15:40). These sources sharpen historical personnel and research-process lineage, but do not establish current employment, current system ownership, model weights, production continuity, or independently audited performance. The repository capture note keeps description-based biography separate from the recording.

The Numerai Keynote from the 2019 conference, uploaded October 24, 2019, provides an earlier first-party bridge between the platform and named technical leadership. The publisher’s segment map identifies Richard Craib on Numerai’s fundamentals and historical backtests, Anson Chu presenting Numerai Compute and Staking 2.0, and Professor Marcos López de Prado discussing financial machine learning (00:03–28:51). The captions add historical references to open APIs, containerized model deployment, blind holdout data, and the separation between participant models and the meta-model. This is dated conference and product evidence, not a current system inventory or independent performance record.

Numerai’s first-party Office Hours S01E13 recording, uploaded September 16, 2020, is a separate community-participant route. Keno Leon describes a public tournament workflow in which model stability matters, training can take from roughly one week to a few months, and about 70% of the work is setup and tooling rather than modeling and training (10:27–14:02). The account also describes repeated statistical-model experimentation and a public dashboard/data-download project (12:32–14:02, 20:04–20:59). This is participant experience and platform-community evidence, not proof of Numerai employment, internal workflow, portfolio deployment, or performance.

Numerai’s Office Hours S01E03 recording, uploaded September 4, 2020, adds a distinct personnel and technical-background route for Michael Oliver. Numerai’s description identifies him as a long-time participant and states that he was hired full-time as a Data Scientist a few months after the recording; Oliver’s first-party May 2020 forum introduction says his planned role was to interface between tournament predictions, tournament rules, incentives, and the hedge fund. The recording documents his prior neural-network work, Python/PyTorch and GPU use, time-series validation concerns, and requests for better model visualization (04:38–07:23, 10:57–12:01, 14:50–15:19). This is dated historical personnel and community evidence; it does not establish Oliver’s current role, Numerai’s current internal implementation, portfolio authority, or performance.

A separate Numerai Office Hours S01E04 recording, uploaded September 4, 2020, captures Richard Craib discussing stability across periods, validation-to-live generalization, separation of signal generation from portfolio optimization, feature-neutral targets, and the role of the meta model (14:18–15:41; 17:01–20:17; 31:50–33:31). This is a founder/community account of the platform’s historical research incentives, distinct from the participant episodes above. It predates the current Predictive LLM, AI-scientist, and MCP surfaces and does not establish current model weights, agent permissions, production authority, or independently audited performance. See the capture note.

Numerai’s official site links careers to Wellfound. The current AI Scientist role on that careers host, accessed August 12, 2026, describes a new team building AI for continuous quantitative research, including hypothesis generation, experiment execution, text and filing extraction, researcher tools, and evaluation infrastructure for systems that can influence portfolio construction. Because the role is hosted outside Numerai’s own domain, the evidence is classified below firm-site disclosure. It is still material hiring intent because Numerai’s own site routes candidates to that host. It does not prove a filled role, deployed system, live trading permission, or AI-attributed return.

QRT: AI-platform hiring, quantitative generative modeling, and a public research fork

Qube Research & Technologies (QRT) has several distinct public layers. Its current careers board lists AI Platform Engineer, AI Platform Developer, and Production Support Engineer — AI & LLM roles alongside quantitative research, data science, compute, and infrastructure positions. These titles establish current hiring intent. The accessible postings do not disclose the filled team, model architectures, production endpoints, or a direct link to live trading.

The current board additionally shows AI Platform Engineer openings in Wrocław and Mumbai, an Academic Partnerships & Research Lead, and Data Scientist roles across London, Paris, Hong Kong, fixed income, options, and pricing. QRT’s company page describes the surrounding platform as supporting structured and unstructured data from tick-level to multi-year horizons, with research teams spanning engineering, computer science, physics, mathematics, data science, and fundamental analysis. These current title and platform surfaces sharpen the organizational map; they do not identify a named AI-platform owner or prove that the listed vacancies are filled.

The current title and media routes are catalogued in the Numerai/WorldQuant/QRT source note.

The accessible role pages add specific platform and control vocabulary: AI Platform Engineer — Mumbai describes production RAG, embeddings, vector databases, retrieval, model serving, evaluation, observability, and agentic workflows; AI Platform Engineer — Wrocław describes shared AI services; Senior AI Delivery Engineer describes coding agents, CI/CD, code review, guardrails, and human oversight; and Senior Product Security Engineer — AI & LLM describes prompt-injection, data-leakage, permissions, monitoring, and safe-execution concerns. The Production Support Engineer — AI & LLM role, checked through Greenhouse’s public job API on August 13, 2026, adds front-office trading flows and LLM-based systems, monitoring, release management, algorithmic-trading platforms, risk systems, and LLM hosting environments. These are role specifications, not proof each control or service is deployed.

QRT’s QRT Labs initiative, launched with Imperial, Cambridge, and Oxford, describes research across AI, decision-making, complex systems, mathematics, and computing, including foundation models and agentic systems. It also supports a large PhD and postdoctoral research community. This is an academic and talent-development signal; it does not establish that QRT trains a proprietary foundation model or deploys agents in its investment books.

QRT also maintains a title-blind Quant Signals External Contributors page inviting outside contributors to provide predictive signals. That is a distinct public intake route from QRT’s employee AI-platform vacancies and academic partnerships. QRT’s anti-impersonation notice identifies its official LinkedIn company page and WeChat account alongside the website, which gives the regional-language discovery process a firm-declared source-of-truth list. Neither page discloses contributor terms, data rights, evaluation, accepted signals, model identity, production integration, or portfolio authority. See the capture note.

An indexed public profile for Peter Buckland-Merrett labels him a Quantitative Technology Director at QRT. A profile-activity card also surfaced the QMUL podcast Just Sayin’ IT, but the canonical Spotify episode and host post identify Check Point Software guests, not Buckland-Merrett or QRT. The route is retained as an identity-resolution disqualification, not as QRT media or AI evidence. See the capture note.

The institutional announcements add concrete scale and personnel context. Oxford says the coordinated programme will directly fund more than 70 early-career researchers across mathematics, statistics, computer science, and engineering. Imperial names foundation AI models, agentic systems, high-performance computing, cybersecurity, hardware design, and advanced mathematical modelling among the programme’s areas. The Oxford Statistics bulletin adds a QRT travel-grant route for ICML 2026, while Yorgos Deligiannidis’s public profile identifies him as a QRT Quantitative Research Director with prior Oxford work in computational statistics, machine learning, generative modelling, diffusion models, and MCMC. These are partnership and academic-lineage signals, not evidence that a QRT-sponsored research topic is deployed in a live portfolio.

QRT’s public Technology Summit post adds a separate firm-media and partner route. QRT says the London summit brought together 130+ colleagues from its global technology community and frames the discussion around research, engineering, infrastructure, systems design, models, optimization, and hardware. It names AWS, NVIDIA, and Snowflake as external partners at the event. This is a company-post signal about technology-community activity; it does not identify speakers, workloads, procurement terms, model architecture, production status, or investment permissions.

A fresh title-blind pass adds an organizational and infrastructure layer. QRT’s May 2026 ESG report reports growth from 1,400 to more than 2,000 employees in 2025, 13 global offices, a QRT Academy, a researcher induction week, a Tech Camp for technology hires, 20 university partnerships, and 135+ student scholarships or bursaries for 2025/26. These are company-reported talent and education figures, not AI staffing or capability measures. The report does not identify a proprietary AI lab, model inventory, or deployment record.

The official Data Engineer, Infrastructure / Cloud Engineer, Quantitative Research / Trading, and Storage Operations Engineer postings, updated August 26–28, 2026, expose a current hiring surface around datasets for research and trading workflows, diverse-data predictive signals, research-to-live-deployment workflows, PyTorch/TensorFlow/JAX familiarity, and Hong Kong/London on-premise HPC storage supporting research and trading teams. The postings are role specifications, not evidence of filled positions, installed frameworks, or a named LLM/agent. QRT’s public GitHub organization also lists infrastructure repositories and forks including LiteLLM, KServe, Restate, Apache Arrow, Apache Doris, Airflow, FastAPI, SQLModel, and OpenTelemetry. The LiteLLM fork exposes a public litellm_internal_staging default branch and AI-gateway metadata; that is a firm-associated code surface, not proof of internal customization, procurement, production use, or trading linkage. See the capture note.

The same follow-up adds four public personnel and academic-lineage routes. Sidharrth Nagappan’s site identifies him as a QRT Quantitative Research Analyst from January 2026 after a 2025 internship and records Cambridge work supervised by Mateja Jamnik and Nikola Simidjievski; Zhang Naifu’s CV identifies a QRT Quantitative Researcher in Singapore since September 2022 and describes continuous-time-series embeddings plus earlier vision-speech-action robotics research; Richard Bergna’s CV identifies a QRT Quantitative Researcher since April 2024 and describes ML and mathematical models for future-return prediction, with Cambridge PhD supervision by Jose Miguel Hernandez-Lobato and Pietro Liò; and Antoine Bodin’s site records a QRT quantitative-research role in Zürich alongside EPFL/Harvard ML-theory lineage, but was last updated in November 2023. These are self-authored personnel and academic records, not proof that QRT uses their academic methods, voice/vision work, embeddings, or models. See the capture note.

QRT’s public GitHub organization contains a public quant-mind repository. The repository is a fork of LLMQuant/quant-mind and describes financial-knowledge extraction, parsers, embeddings, RAG/Data MCP, and workflow or agent components. Its README presents some agent-memory and cross-document-reasoning capabilities as future or aspirational functionality. The correct classification is a firm-associated public research/experimentation artifact with uncertain provenance—not a QRT-originated model or a live investment system.

The JUXT XT25 conference recap and agenda add a named conference route that was not in the earlier QRT capture. The recap identifies Steve Pegg as QRT’s Head of Risk and P&L Technology and moderator of a panel with QRT Quantitative Technical Director Zohar Melamed; the embedded recording also identifies Melamed’s QRT role (role introduction). The publisher frames the panel around AI adoption in regulated environments, legacy-system modernization, and resilient risk and trading platforms. This supports named personnel and conference-topic evidence. It does not establish a QRT model, vendor deployment, dataset, permission boundary, or live investment use; automatic captions are retained only as locators, not as a clean transcript.

QRT also has a public predictive and generative-model research record. A 2026 JMLR paper lists QRT affiliations and studies nonparametric generative modeling for financial time series. Earlier work covers recurrent neural networks for collateral posting and policy-gradient methods for financial decision problems. These sources concern statistical prediction, stochastic control, and synthetic financial data; they should not be relabeled as LLM or agentic investment deployment.

The Singapore academic and conference search adds three separate public routes. The ACM ICAIF’25 Industry Day programme identifies Georgios Papaioannou as QRT Research Director and moderator of a panel on cutting-edge AI-in-finance research; it also identifies Chao Zhou as an Industry Day co-chair. The NUS Institute for Mathematical Sciences RIPS 2025 poster lists QRT among past sponsors of a programme applying mathematics, statistics, data science, and machine learning to industry projects. Separately, NUS PhD candidate Qian Wang’s public research page records a 2025 invited talk at QRT’s Singapore office on using LLMs to make fair and unbiased judgments about factors, alongside Wang’s public work on financial benchmarks, LLM trading agents, and trustworthy LLM judgment. These routes extend the visible research and talent network around QRT; they do not establish a QRT-owned model, adoption of Wang’s methods, training corpus, production endpoint, agent permission, or investment result. The source hierarchy and boundaries are recorded in the QRT Singapore academic-media capture note.

A separate YellowDog customer case study adds a named compute-infrastructure relationship. YellowDog says QRT uses RayDog to orchestrate Ray-based ML workloads and its broader platform to coordinate AWS workloads, including large parallel simulations against historical and real-time market data. The PDF names Jon Fautley as QRT’s Head of Cloud Infrastructure and describes workload submission, monitoring, isolation, multi-region scaling, and clusters reaching 10,000+ nodes. It also reports customer-story outcomes such as moving some simulations from hours to minutes and enabling multiple model iterations per day. These are vendor-reported infrastructure and workflow claims; they do not identify QRT model architectures, permissions, trading authority, or AI-attributed returns.

An additional AWS Summit London 2025 recording makes the QRT platform story more concrete. In the session, the QRT presenter describes AWS as supporting financial research, strategy development, backtesting, and a large data platform; the captioned walkthrough covers the evolution from S3 and a PostgreSQL/EC2 setup to EKS-backed services, researcher data access, workload submission, and backtesting controls (data-platform discussion, backtesting discussion). Public event promotion identifies Jon Fautley, Jon Hammant, Jonathan Kerr, and Adam Temple. This recording is infrastructure and research-process evidence, not a GenAI disclosure: it does not name an LLM, agent, training corpus, model evaluation, permission map, or AI-attributed return. The caption recovery and exact locators are recorded in the QRT AWS Summit source note.

QRT’s NVIDIA GTC San Jose 2026 session adds a distinct, named infrastructure lane. NVIDIA identifies Jerome Vienne, QRT’s HPC Performance Engineer, and Chris Turpin, QRT’s HPC Network Engineer, in a session on QRT’s B200 HGX/Spectrum-X research cluster. Their transcript describes tens of thousands to potentially hundreds of thousands of daily research-cluster jobs, QRT’s move from heavy public-cloud use toward an on-premise Iceland data center, layer-by-layer GPU/NVLink/network validation, Slurm scheduling by scalable unit, and real-training validation rather than microbenchmarks alone (cluster rationale, validation and scheduling). The page associates the session with BlueField, HGX, NCCL, NVLink/NVSwitch, Spectrum-X, and Blackwell. This is high-confidence evidence for a public QRT infrastructure presentation and medium-confidence evidence for automatic-transcript wording; it does not disclose a model, training corpus, agent, permission map, live investment linkage, or AI-attributed performance. The full route and boundaries are recorded in the QRT NVIDIA/French interview source note.

Two French-language interviews with Laurent Laizet add dated leadership and regional-language evidence. The November 2021 interview describes a data-centric QRT approach, satellite imagery, machine-learning transformation of unstructured data, and recruiting for technical theory, new data sources, and new markets. A January 2024 interview says QRT develops mathematical models and machine learning for asset-movement prediction and describes collaboration among data scientists, quantitative researchers, and quantitative traders working with satellite images and financial news (2021 interview; 2024 interview). These are publisher-transcribed executive statements, paraphrased here from French; historical asset/headcount figures are not independently audited and are not treated as current. Neither interview discloses an LLM, named model, training corpus, agent permission, production endpoint, or performance result.

Key public people

Person	Publicly visible role or evidence	What the public record supports	Boundary
Rémi Leluc	Personal site says he has been a QRT quantitative researcher since June 2024; LinkedIn corroborates affiliation	Current quantitative-research signal involving stochastic optimization, reinforcement learning, federated learning, Monte Carlo, and ML	No public GenAI deployment or platform ownership
Pierre Henry-Labordère	QRT-affiliated 2026 JMLR paper and public profile	Current research-affiliated generative modeling and stochastic-control signal	Generative financial time-series work is not LLM/agent evidence or proof of production use
Mohamed Hamdouche	QRT affiliation in the same JMLR paper; public profile	Research-affiliated generative time-series and quantitative-finance signal	Current job title and production remit are not publicly specified
Chao Zhou	University of Queensland seminar identifies him as QRT Quantitative Research Director	Current/qualified quantitative-research leadership and ML discussion	No public GenAI production claim
Laurent Laizet	QRT co-founder/CIO and sponsor named on QRT Labs	Current investment leadership and research-sponsorship signal	Public source does not establish technical AI/ML ownership
Daniel Giamouridis	Oxford QRT Labs announcement identifies him as a QRT Quantitative Research Director	Current research-leadership signal within the QRT Labs partnership	No public AI-platform ownership or production deployment claim

What QRT is not disclosing

QRT’s public sources do not disclose a named head of AI or AI platform, a complete current AI roster, a proprietary foundation-model training program, model cards, training data, internal benchmarks, agent permissions, deployment dates, or AI-attributed returns. QRT’s public research and hiring surfaces support multiple AI/ML and platform signals, but they do not provide a complete chain from model to investment decision.

Tower Research Capital: predictive-ML infrastructure plus public agentic-research discussion

Tower’s careers and engineering pages describe research tools, historical-data simulation, machine learning, hardware acceleration, low-latency systems, model infrastructure, risk, compliance, and large-scale data. A current Machine Learning Engineer listing, also mirrored by Built In, describes distributed GPU training, large-scale experimentation, model/version/artifact management, data lineage, feature engineering, backtesting, resource scheduling, and ML telemetry. This is a role-description signal for a research platform; it does not prove that any specific model is live or that the role was filled.

Tower’s public AI material also discusses GenAI and agentic research workflows. A January 6, 2026 article describes AI agents, MCP, multi-agent systems, and LLM-powered workflows being explored across trading, research, and engineering. The associated Columbia practitioner seminar identifies the talk as “AI Agents in High-Frequency Trading Environments” and says it covered current implementations at Tower. Tower’s May 26, 2026 AI-in-capital-markets article adds knowledge graphs, agent harnesses, governance, observability, security, cost controls, and auto-research-style systems that browse, synthesize, analyze, and report.

The same quiet-firm pass recovered Tower Research Ventures’ 2024 AI x Wall Street event report. Tower’s venture arm says the event included Edge AI Research Lab, OpenBB, and Cohere, with demonstrations involving OpenBB, MindsDB, Cohere, Genesis Computing, TrendTrophets, Markets EQ, and WaverlyAI. The page discusses investment-data analysis, compliance review, sentiment analysis, note-taking, reporting, synthesis, RAG knowledge bases, embeddings, and enterprise guardrails. Tower explicitly says the information was not independently verified; this is an ecosystem and event record, not proof that Tower adopted any named vendor or connected a demo to a trading system.

These sources establish company-reported agentic research and implementation activity. They do not identify a specific model, training corpus, evaluation score, autonomous trading permission, execution path, or investment performance attribution. Tower’s August 6, 2026 ICML report adds a research and recruiting signal around overfit models and data leakage, not a disclosed production model.

Tower’s Synthetic Software conference recap separately describes code-generation models used in internal production and AI applied to risk surveillance and monitoring. This supports internal software and control-workflow use. It does not establish that the same systems generate investment signals, approve positions, or execute orders.

The current Tower site adds several previously untracked routes. Its August 11, 2026 time-series seminar note says NYU’s Andrew Wilson visited Tower’s New York office to discuss time-series foundation models, data generation, zero-shot forecasting, uncertainty representation, and synthetic data with trading-team members familiar with deep learning and LLMs. Tower’s August 13, 2026 Citi-panel recap attributes discussion to CIO John Cogman and mentions HFT/MFT convergence, new datasets, factor models, risk and portfolio monitoring, continuous markets, and AI/compute capacity. A current employee profile describes a unified web framework used across trading, engineering, and support and explicitly frames the work as tools for people rather than virtual agents. These are first-party research, executive, and engineering signals; they do not establish a deployed model, training corpus, agent permissions, or live investment authority.

Two older Tower research-community routes also expand the public record. Tower’s NeurIPS 2024 retrospective discusses time-series benchmarks, low-compute transformers, irregular-data features, and AI-driven development tools. Its ICML 2024 retrospective discusses code agents, repository testing and refactoring, feature transfer, quantization, time-series foundation models, and secure ML. Separately, Tower’s IIT Bombay ML for Social Good announcement describes a 2024 CSR education lab, names Professor Kameswari Chebrolu as principal investigator, and mentions AI/ML models, BodhiTree, evalPro, cLabs, and H100 support. Tower Research Ventures’ Princeton NLP event recap records a July 9, 2024 discussion of SWE-agent, SWE-bench, code-agent interfaces, and benchmark construction, while disclaiming independent verification of the reproduced historical rates. These routes are technical, academic, social-good, and venture ecosystem evidence; they remain separate from proof of Tower trading-system adoption. The capture note records the source boundaries.

The latest first-party media pass recovers two standalone Tower routes. Tower’s August 4 NLP for Finance article describes classification, extraction, sentiment, semantic search, summarization, question answering, communications surveillance, and research support, then names embeddings, retrieval-augmented generation, domain adaptation, evaluation, permissions, freshness, and monitoring as production concerns. Tower’s August 6 ICML 2026 recap describes a Seoul sponsorship and an inaugural Production Trading System Challenge in which nearly 800 participants were asked to identify an overfit model and detect a data leak. These are firm-authored system vocabulary, research-community, and recruiting signals; they do not disclose a Tower model, training corpus, benchmark rubric, production endpoint, trading permission, or investment result. See the updated capture note.

Key public people

Person	Publicly visible role or evidence	What the public record supports	Boundary
Ramit Sawhney	Tower identifies him as Global Head of Core AI & ML in a career spotlight, 2026 AI article, and AI-agents article	Current directly AI/ML-focused leadership signal covering time series, GNNs, NLP, transformers, LLMs, and agentic research discussion	Public sources do not disclose model ownership, deployment scope, or investment attribution
John Cogman	Tower’s Columbia conference report identifies him as CIO	Current senior investment leadership with public AI/ML strategy exposure	Not a technical AI personnel or model-ownership signal
Saksham Sharma	Current Tower employee spotlight	Current quantitative-development signal supporting research technology for a global trading team	No explicit GenAI remit or system ownership

Tower’s ventures page and its Atomic Canyon investment announcement document external AI-company investments. Those are venture exposures and must remain separate from evidence about Tower’s own research or trading systems. Tower’s GitHub organization has no public repositories or public members in the reviewed surface, and no assigned AI/ML patent was located.

What Tower is not disclosing

Tower does not publicly disclose model names or weights, training data, evaluation suites, production deployment dates, agent permission boundaries, kill-switch design, override rates, or AI-specific returns. The public record supports a predictive-ML engineering substrate and company-described agentic research exploration; it does not establish autonomous capital allocation.

Public code artifacts: ownership is evidence, not deployment

Public repositories reveal different layers of a firm’s technical surface. A firm-controlled repository provides direct evidence of public technical work, while a personal repository or job mirror provides a different kind of signal. Repository ownership alone does not establish that the code powers a live investment system.

Firm	Public artifact	What it supports	Boundary
BlackRock	Blowfish, HOLA, AladdinSDK	Semantic-search/RAG evaluation, hyperparameter optimization, and public Aladdin APIs; Blowfish names Thomas R. Barillot and Alex De Castro as research authors	No public link to Asimov, AI Labs, Systematic, or live investment decisions
QRT	quant-mind, pi, and QRT organization	Firm-associated public RAG, embeddings, knowledge graphs, MCP, and agent-tool artifacts	quant-mind is a fork with substantial upstream provenance; QRT authorship and deployment are not established
Man Group / AHL	ArcticDB and LiteLLM fork	Public time-series/data infrastructure and an organization-owned LLM gateway fork with routing, logging, cost, and guardrail vocabulary	No Man-specific model customization or investment deployment is disclosed by the repositories
CFM	CFMTech, Deep RL for Portfolio Optimization, and CapitalFundManagement on Hugging Face	Firm-controlled research organization, historical portfolio-optimization code, and a public model-hosting organization; separate firm case study documents Llama 3.1-based NER annotation and fine-tuning	Direct public ownership and a historical deployed annotation workflow; no current GenAI trading deployment or named live model is disclosed
Point72 / Cubist	CSP, ccflow, csp-gateway, and csp-bot	Firm-owned streaming, simulation, workflow, orchestration, and chatbot infrastructure; CSP identifies Point72 development	Infrastructure evidence; the chatbot is not proof of an LLM investment system
Two Sigma	functional_semantic_types and its NAACL paper	Public semantic-type ontology generation using foundational models	Repository and paper do not establish current author employment or investment deployment
Jane Street	torch and GTC 2026	OCaml/PyTorch GPU bindings and training-performance/profiling work	No public code path connects these artifacts to a GenAI trading workflow
Jane Street	Hugging Face organization and neural-network puzzle	Public model artifacts released for mechanistic-interpretability puzzles, including DeepSeek/Qwen/PyTorch-related artifacts	Public research artifact; no model cards, inference-provider deployment, or production trading-model link
G-Research	GR-OSS organization, ParquetSharp, FastTrackML, and Armada	Firm-linked data libraries, ML experiment tracking, and large-scale scheduling infrastructure	Public engineering evidence; no direct public GenAI-to-investment link
G-Research	Bobbit	Firm-owned browser-controlled multi-agent coding runtime with human steering, workflow state, recovery, and sandbox/worktree controls	Public agent-runtime evidence; no investment or trading deployment
Marshall Wace	Recursive Language Model workshop	Organization-owned educational artifact about recursive language models and coding agents	One workshop repository does not establish a production RLM program or investment use
D. E. Shaw	Public organization and versioned-hdf5	Public engineering and data infrastructure	No direct investment-business GenAI artifact
XTX Markets	ternfs	Large-scale distributed data/filesystem infrastructure	No public LLM or agent code
Balyasny	BalyasnyAI on Hugging Face	Firm-associated public embedding model and disclosed finance-retrieval research	Public model is not proof of the current production platform or investment performance

The negative results are also bounded: no relevant firm-controlled AI repositories were located for GMO Asset Management, Acadian, Arrowstreet, WorldQuant, Citadel, Millennium, Schonfeld Investments, Systematica, Brevan Howard, Caxton, Squarepoint, or Tower Research Capital in the reviewed organizations. CFM, Balyasny, QRT, BlackRock, Man Group, Two Sigma, and Marshall Wace have separate public artifacts described above. These are public-artifact search boundaries, not evidence that private code or systems do not exist.

August 11 verification delta: new artifacts, current-status changes, and disqualifications

This pass did not locate a new public disclosure of autonomous order authority or an audited AI-attributed return stream. It did locate several changes to the public evidence map.

Current firm disclosures and operating signals

Man Group’s July 28, 2026 half-year results now explicitly describe embedding cross-functional agentic workflows across the firm. This is a firm-level operating statement, not proof that a named AHL or Numeric system is live. The current Man Numeric page identifies Daniel Taylor as CIO, Ed Fang as Director of Research, and Tom Taylor as Co-Head of Front-Office Engineering; none is publicly assigned ownership of GenAI.

Balyasny’s public evidence has a newly surfaced scale claim. A National podcast page attributes approximately 2,000 AI agents and 5,000 daily tasks to Dmitry Balyasny. The episode description does not provide a transcript, architecture, permission map, or measurement method. Balyasny’s official leadership page identifies Dmitry as Managing Partner and CIO, Charlie Flanagan as Chief AI Officer, Mike Grimaldi as CTO, and Kevin Byrne as COO with oversight touching Data Science & Applied AI and Technology. The podcast figures and OpenAI’s smaller Applied AI-team description measure different things and are not reconciled by either source.

Acadian’s current 10-Q contains no matches for “artificial intelligence,” “AI,” or “machine learning,” while its earnings-call transcript discusses enterprise AI productivity, coding, selected AI-enabled research services, and shared guardrails. This is a disclosure-surface difference, not evidence that Acadian does or does not use AI internally. The VP Investment AI Engineer posting also contains location and AUM-copy inconsistencies; those are treated as recruiting-platform drift, not as evidence of multiple roles or a filled position.

Arrowstreet’s two previously reviewed AI postings now show conflicting status signals: LinkedIn marks the Senior AI Platform Engineer and Senior AI Security Engineer roles as no longer accepting applications, while the Workday platform page still renders an application flow. No successful hires are named. Jollson Varghese’s public experience page self-reports a Director of Software Engineering role beginning February 2026 and discusses responsible AI and automation, but this is personal self-reporting and does not establish firm-wide AI ownership.

WorldQuant’s official career index still presents the LLMs & AI Agents internship under the WorldQuant-hosted route, even though older Greenhouse detail can remain date-conflicted. The same current firm-hosted surface also lists WQBRAIN AI Researcher, AI Software Developer, Senior AI Software Developer, and Lead Python Engineer, AI/ML Systems roles. The correction is current hiring-surface evidence only; it does not establish filled roles, model inventory, deployment, live-book authority, or return attribution.

QRT’s current board lists AI Platform Engineer roles, but the direct Wrocław posting and Mumbai posting currently render application forms without the previously observed role detail. Earlier RAG, serving, evaluation, and observability language is therefore dated hiring evidence rather than freshly reverified architecture. QRT’s company feed now provides a more specific current training signal: its Quant Research Induction Week included backtesting, “agentic quant research,” and the path from idea to market for new joiners across nine offices. This is a firm-social training statement, not proof of a named tool, model, production deployment, or trading permission.

New public technical artifacts

Bridgewater published Noisy Data Breaks RLVR on August 3, 2026, with an associated paper and code hosted by the external uiuc-kang-lab organization. The work is Bridgewater-linked research on reward-label noise and data curation. The external repository should not be described as Bridgewater-owned code, and the paper does not establish live investment deployment.

Jane Street’s verified Hugging Face organization exposes several public model artifacts, including large DeepSeek-V3 variants, a Qwen2 warm-up model, and a smaller PyTorch model. Jane Street’s engineering article frames the releases as mechanistic-interpretability puzzles; the model pages do not provide model cards or inference-provider deployment. These are firm-owned public research artifacts, not disclosed production trading models. A community “backdoor” interpretation was later corrected in the Hugging Face discussion and is excluded.

Jane Street’s current Machine Learning Researcher role adds hiring-language specificity: it names LLMs and RL agents as part of the relevant ML landscape while saying the role trains models for the next generation of deep-learning-based trading strategies. The Visiting Researcher page separately says Jane Street uses third-party models and custom models trained on its own infrastructure. These sources support current model-training and hiring intent; they do not show that any LLM itself drives trading decisions or disclose weights, data, evaluation thresholds, or live metrics.

G-Research’s firm-owned Bobbit repository adds a public agent-runtime artifact. Its README describes a browser-controlled multi-agent coding command center with leads, coders, reviewers, testers, workflow state, human steering, session recovery, provider consistency, authentication handling, and sandbox/worktree protections; public releases extend through v0.16.3. This is evidence of an agentic software-engineering runtime and its controls, not evidence of investment or trading deployment.

Personnel and entity corrections

XTX’s Companies House history records Joshua Leahy’s directorship termination on March 30, 2026, and contemporary reporting reports his departure from the CTO role and XTX. An older MIFIDPRU disclosure is stale for current-personnel purposes. No verified successor was located. XTX’s current website reports machine-learning forecasts across more than 53,000 instruments, approximately 25,000 GPUs, and roughly one exabyte of storage; these are firm-reported infrastructure figures and do not disclose LLM use or agent authority.

Voleon’s current management page identifies Jeremy Rosenblatt as Managing Director of Strategic Projects and Head of Software and Systems. This adds a current technology leadership signal around a predictive-ML organization; it does not establish a GenAI program. PDT’s February 2026 Generating Alpha podcast provides current founder-level context on model-building and risk but no LLM disclosure; a separate episode with Vinesh Jha identifies him as a former PDT director and is excluded from current personnel evidence.

Winton’s current MIFIDPRU disclosure records Carsten Schmitz’s retirement as Co-CIO and Simon Judes becoming sole CIO. A separate public LinkedIn signal associates Judes with a TracFox founding-team entry, creating a current-status conflict that the public record does not resolve. The article retains Judes as CIO based on Winton’s official leadership page and does not infer departure or exclusive affiliation.

D. E. Shaw’s current public hiring surface now includes an Applied AI Engineer, Fundamental Equities AI Product Analyst, and AI Enablement Strategist. These postings describe agents, skills, retrieval, production deployment, idea sourcing, document synthesis, data licensing, information barriers, and compliance. An August 11 recheck also located a Strategic Intelligence: Software & AI Research Analyst role in Financial Research, focused on software and AI markets, AI adoption, AI infrastructure, product roadmaps, and investment-relevant implications; a Product Manager - AI Vendor Tools role covering Claude, ChatGPT, Copilot, and Gemini; and an AI Engineer - Human Capital role mentioning RAG, agentic architectures, Cursor, and Claude Code. These remain hiring-intent and operating-surface evidence. Market research, vendor-tool ownership, human-capital use, and the separate Cove private-equity AI role are not treated as DESIM or liquid-strategy deployment proof.

August 12 verification delta: adjacent quantitative-manager controls

This pass added a compact adjacent-firm control cluster. These sources are useful because they show how prop-trading and market-making firms describe AI differently from asset managers, but they should not be merged into the GMO, Acadian, or Arrowstreet evidence layers. Confidence is high for the existence of the cited first-party pages and current role language; confidence remains low for any inference about autonomous capital allocation, model ownership, permissions, or AI-attributed P&L.

Jump Trading’s current AI/ML page, accessed August 12, 2026, describes machine learning across trading, research, and core infrastructure; frontier work in deep learning, reinforcement learning, LLMs, and generative modeling; LLM agents and assistants served through API and HPC; NLP signal generation; custom foundation models; under-24-hour model-deployment-to-trader feedback; 50+ AI tools used daily; and 75%+ firmwide weekly LLM usage. This is firm-owned operating evidence. It does not disclose model weights, permission maps, trading authority, return attribution, or the methodology behind the usage figures.

An August 31, 2026 recheck of the same first-party AI/ML page adds two displayed leadership titles: Lucas Baker, Head of LLM R&D, and Loren Puchalla Fiore, Head of ML Engineering. The page also describes a split between AI/ML specialists embedded with trading teams and central groups working on custom foundation models, LLM agents, scalable systems, and infrastructure. These are current page labels and firm-reported operating descriptions; they do not establish start dates, reporting lines, ownership of every system, model weights, training data, permissions, or investment outcomes. See the dated capture note.

Hudson River Trading’s current Machine Learning & AI page, accessed August 12, 2026, identifies HRT AI Labs, describes deep learning as core to trading, names low signal-to-noise, distribution shift, adversarial adaptation, ultra-low latency, and real-time event processing as operating constraints, and says AI researchers work on features, architectures, training dynamics, and trading-decision logic. A current AI Researcher, LLMs role page confirms the role title, Strategy Development function, locations, and job ID, while the public page text available in this pass does not expose the full role body. These sources support a current AI/ML research surface; they do not show a public LLM trading model, strategy parameters, or reproducible performance. The page-level capture is preserved in the HRT source note.

The title-blind queue also resolves a dated HRT optimization route. Miles Lubin’s JuliaCon recording identifies him as JuMP’s BDFL and an HRT employee speaking personally (00:07–00:18); JuMP governance and Lubin’s public research page independently document his project role, HRT affiliation, MIT Operations Research lineage, and optimization publications. The material supports a public personnel and open-source software relationship. It does not show that JuMP is used in HRT trading, disclose an HRT model, or establish production permissions or performance. A separate JuliaCon HiGHS session was checked and reassigned to Julian Hall rather than Iain Dunning; the capture note records that negative attribution result.

The personnel pass adds three distinct public lineage routes. Sean Mann’s research page identifies him as an HRT AI Researcher and HAIL member building deep-learning systems for trading; it records MIT SB and MEng training, Devavrat Shah/LIDS advising, and public work on time-series forecasting, matrix completion, and differentiable simulation. Kevin Luo’s page identifies him as an HRT algorithm developer from August 2024 and records Harvard statistics/mathematics training plus public work on high-dimensional cross-validation, out-of-sample risk, and neural-network pruning. Claire Donnat’s CV records a 2019 HAIL AI-team fellowship focused on deep learning for time series and market-structure analysis; her University of Chicago profile records her current statistics faculty role and Stanford lineage under Susan Holmes and Jure Leskovec. These pages expose personnel, education, and academic research themes—not a complete HAIL roster, HRT-owned papers, model inventory, data rights, production permissions, or AI-attributed performance. The capture note keeps current roles, historical fellowship, and academic artifacts separate.

A newly recovered JuliaCon 2014 JuliaOpt presentation adds a separate historical lineage signal for Dunning: the official archive identifies him and Joey Huchette as MIT Operations Research Center doctoral students presenting open-source optimization packages. The recording is useful for understanding optimization and academic background, but it contains no Hudson River Trading reference and is not evidence of an HRT system or contemporaneous employment. The capture note keeps this date-scoped academic artifact separate from current HRT AI Labs evidence.

An additional HRT-controlled LinkedIn walkthrough features Marc Khoury of HRT AI Labs with Dwarkesh Patel discussing how a researcher might construct and evaluate a predictive signal from historical order-book data. The embedded transcript describes top-five bid/ask snapshots, imbalance and momentum features, incremental predictive value, changing-market parameter drift, and the many research decisions beneath a simple signal question. It also contains a “34% improvement” remark about combining features, but the surrounding exchange says that adding a feature to a linear model cannot lower in-sample R-squared when zero weighting is available; the figure is therefore an illustrative in-sample modeling comment, not a return or out-of-sample result. A June 10, 2026 Blockworks account adds secondary context on a speed-versus-predictive-power tradeoff, including a claim that one-minute forecasts may be difficult to monetize when inference is too slow. This is first-party educational media plus secondary commentary; it does not expose the model family, label horizon, split, costs, permissions, latency budget, production endpoint, or realized performance. The capture note records that browser review recovered the publisher transcript, while no episode-specific video ID or downloadable recording was exposed.

A separate HRT AI Labs video, published July 7, 2025, gives a more explicit organizational description. HRT says HAIL is a researcher-and-engineer team that collaborates with trading teams rather than owning trading, uses supervised, unsupervised, and reinforcement-learning techniques, and has models used across asset classes and time horizons. The speaker also describes petabyte-scale tick-data holdings, news and alternative data, a large compute cluster, and latency-aware design for high-rate market events. The official English caption track was recovered on August 28, 2026, adding a durable provenance hash to the existing timestamped transcript. These are narrated first-party claims, including the stated reach of HAIL-trained models; the video does not disclose model versions, labels, permissions, architecture, independent telemetry, or AI-attributed returns. See the timestamped capture note.

HRT’s March 13, 2026 HPC video adds a physical-infrastructure lane. The firm describes its Lefdal Mine Datacenter in Norway, cold-seawater cooling, renewable power, dense racks, direct-liquid-cooled NVIDIA GPUs, high-speed interconnects, and storage designed to feed large data volumes to accelerated compute. It also describes plans for larger models and more data, alongside narrated rack-density and investment figures. These are first-party facility claims, not an independent capacity or capex audit; the video does not disclose GPU counts, training runs, model families, data rights, utilization, or model performance. The capture note records the boundaries.

The June 5, 2026 Odd Lots follow-up, also distributed through YouTube and Apple Podcasts, adds a separate generative-AI and infrastructure layer. Iain Dunning describes coding, experiment ideation, experiment monitoring, and attempts to compare agent-generated signals with human quant research (09:04–10:39); he also describes objective task-level evaluation, power and site constraints, multi-year GPU-capacity planning, and an estimated $100–$200 per employee per day for token use on his team (23:31–24:17). The recording distinguishes research assistance from HRT’s predictive market models and does not identify a public LLM trading model, a named custom-chip contract, firmwide spend, model permissions, or AI-attributed performance. See the timestamped capture note.

A title-blind archive pass also recovered HRT’s 2023 algorithm-developer profile of Francisco Rivera. Rivera describes algorithm developers as owning strategy creation, maintenance, and improvement on a medium-frequency team, with Python notebook research feeding C++ production-system work and reusable analysis shared through a codebase (00:22–02:32). The recording identifies Harvard applied mathematics/economics and additional computer-science study, an HRT internship, and an algo-developer-lead promotion at the time of recording; HRT’s team surface and data-quality article provide separate public identity and authorship cross-checks. This is a dated role-taxonomy and research-to-production signal, not an AI or GenAI disclosure; it does not establish a current title, model inventory, permissions, or performance. The capture note records the evidence boundaries.

An older HRT automation roundtable, published November 7, 2023, describes automation as spanning research, data, production trading, and maintenance. Participants discuss standardized repeated tests, common abstractions across markets, automated monitoring and computer response in live trading, and retaining more manual work where market structure is difficult to abstract. They also discuss language models as a future aid for code, design documents, code review, and contextual interpretation. This is historical first-party operating-model evidence, not proof of a current LLM deployment or autonomous trading authority; see the capture note.

HRT’s first-party ICML 2026 conference post adds a current conference and personnel route. It identifies Iain Dunning as Head of HRT AI Labs, advertises his session “Scaling Deep Learning in Financial Markets” on July 6, 2026, and describes an HRT research-and-engineering presence at Booth B201. The post and company feed establish the event and public topic scope, but no recording, transcript, slide deck, model, dataset, benchmark, or production-system detail was recovered. This is conference-participation evidence, not evidence of a particular deployed model, trading permission, or performance result. See the capture note.

A separate NVIDIA GTC 2026 on-demand session adds an HRT infrastructure route that the registry had not yet connected to the firm. The session, published March 17, 2026, is titled “The Blueprint for a Modern Resource-Responsible AI Factory” and names Gerard Bernabeu Altayo as HRT Compute Systems Lead. NVIDIA’s description frames the session around HRT’s AI-factory trajectory, data-center design, accelerated workloads, resource efficiency, and the end-to-end path from AI-driven model development toward deployment. This is vendor-hosted session metadata and advertised scope; the directly retrievable page does not provide a stable full transcript in the reviewed capture. It therefore supports the named role and infrastructure topic, not exact GPU capacity, model-training runs, data rights, production permissions, trading authority, or performance. See the capture note.

A title-blind HRT employee profile adds a Singapore operating and deployment-control route. The speaker identifies herself as Agnes, a Python Engineer on the Operations Developer Team, and describes infrastructure for monitoring and regulating trading machines, connecting brokers and exchanges under market restrictions, and maintaining deployment processes for trading strategies (00:23–01:06). She also describes choosing between batch, streaming, and asynchronous processing as data grows (01:29–01:45). This is useful evidence about the platform, compliance, and strategy-release layer around trading; it is not evidence of GenAI, an AI model, an agent, or a trading decision right. The recording supplies no surname or exact role date. See the capture note.

A second title-blind HRT employee profile identifies a speaker named Bati as an algorithm developer in HRT’s London office. The 2023 recording describes researching and implementing signals, modifying existing signals, and improving implementation efficiency; it places the speaker on a medium-frequency team and names statistics, machine learning, Python, communication, and skepticism about research errors as important skills (00:14–02:29). Bati also describes a probability-theory PhD and postdoctoral experience before moving into finance (02:41–03:20). This adds role-taxonomy and research-to-production context, not GenAI evidence. The recording supplies no surname, university, paper, current title, model inventory, permissions, or performance result. See the capture note.

A title-blind Code of Conduct episode adds a historical HRT personnel and engineering-practice route that the Jane Street query had misclassified. The video title identifies Jessie Newman as an Algo Engineer at Hudson River Trading. A Society of Women Engineers profile says she joined HRT in 2019, worked on testing infrastructure, and helped lead the firm’s Women’s Employee Resource Group; PyLadiesCon’s speaker page separately records three years of high-frequency-trading work at HRT, and a public LinkedIn post records HRT algorithm-engineering recruiting activity. Recovered English automatic captions add a dated first-person description of creating tools used to create trading algorithms launched across different stock markets (01:14–01:25). The captions are not speaker-verified and support only historical personnel, hiring, testing, and engineering-context claims—not AI/GenAI use, current employment, model ownership, permissions, or performance. See the capture note.

Jane Street’s first-party software-engineering mock interview and recording add a separate dated recruiting route. The company identifies the presenters only as software engineers named Grace and Nolen and provides timestamps for the interview exercise and guidance on communication, clear and correct code, and practice. This is engineering-culture evidence, not AI or trading-system evidence; no surnames, current team assignments, model ownership, or production claims are inferred. See the capture note.

Jane Street’s State Machine Replication episode with Doug Patti adds dated control context around Concord and Aria. The publisher identifies Patti as a Client-Facing Tech developer and describes Concord’s availability and audit-trail requirements; the recording discusses deterministic testability, replay, access to replicated topics, and automating parts of message/order handling (03:03–03:34; 23:38–28:09; 48:32–50:35; 65:09–68:37). This is first-party engineering evidence, not AI/GenAI or autonomous-capital evidence. See the capture note.

The title-blind queue also recovered the November 9, 2015 Software Engineering Daily interview with Yaron (Ron) Minsky and its recording. Minsky corrects the host’s introduction at 00:51–01:02, saying he is Head of Technology and no longer Head of Quantitative Research, a useful date-scoped organizational distinction. He describes real-time multi-source market data, accurate timestamps, bulk historical research, and rapid iteration as trading-technology requirements (01:07–02:17); later he distinguishes risk tiers across order-sending, build, and monitoring infrastructure (44:00–44:26) and describes redundant ways to stop trading as a deliberately engineered safe state (49:04–50:58). This historical external interview helps explain the control vocabulary around later Jane Street infrastructure and developer-tool disclosures. It predates current GenAI and does not establish a current model inventory, agent permissions, data rights, trading authority, or performance. See the capture note.

The recovered Strange Loop 2018 session “Data Driven UIs, Incrementally” adds a separate conference route and a more concrete view of Jane Street’s data-intensive internal interfaces. In the recording, Minsky frames the firm’s proprietary trading work around complex, changing data (00:11–00:56), demonstrates a table with roughly 100,000 rows and dynamic filtering/sorting (01:17–02:21), and explains Incremental, persistent data structures, and diff/patch as ways to limit recomputation (08:41–10:14; 20:49–26:18). He closes by describing UI design as an optimization problem and diff/patch as a bridge across functional data, incremental computation, network protocols, and browser APIs (33:39–35:46). This is historical engineering evidence, distinct from the 2012 Caml Trading and YOW! routes, not ML/GenAI deployment evidence. It does not establish current model weights, data rights, agent permissions, trading authority, or performance. See the capture note.

Optiver’s current research page, accessed August 12, 2026, describes statistical modeling, machine learning, high-performance compute, petabytes of historical and live market data, simulation, peer review, production constraints, and an AI Lab exploring deep learning, reinforcement learning, LLM-based analysis, signal discovery, forecasting, execution quality, and risk decisions. This is first-party research-process evidence. It does not identify model vendors, language-model weights, an agent runtime, or AI-attributed investment results.

DRW’s current AI Engineer and Data Developer roles, accessed August 12, 2026, describe an AI/ML role inside a collaborative multi-asset trading team, end-to-end AI infrastructure, continuous training, validation, monitoring, feature/model stores, production ML/AI model maintenance, LLMs with RAG and fine-tuning, and a Data Developer role in the AI and Multi Asset Systematic Strategies team building RAG pipelines, embeddings, vector storage, retrieval-quality evaluation, and inference data workflows. These are official hiring signals; they do not prove roles are filled, systems are live, or any model has trading authority.

A separate live DRW Quantitative Researcher — Commodities listing, checked September 2, 2026, adds a title-specific research-to-production route. It requires experience using AI and LLMs and describes systematic commodity strategies, alpha-generating signals, large and diverse datasets, data pipelines, predictive-model validation, deployment, monitoring, updating, and post-trade diagnostics. This is stated hiring scope, not evidence that the role is filled or that the listed methods are deployed. The page does not disclose model families, vendors, training data, evaluation design, agent runtime, capital permissions, or AI-attributed returns. See the capture note.

DRW’s first-party AI page identifies Wachi Bandara as Director of AI and says the firm is expanding across functions touched by AI. A separate DRW interview with Bandara identifies him as Head of Artificial Intelligence and describes a remit spanning machine learning for unstructured or difficult datasets, firm-specific information, model testing, responsible use, hallucination controls, human validation, and model-input and bias considerations. Bandara’s background description includes image processing, computer vision, yield-curve arbitrage, equity risk, and founding machine-learning companies before joining DRW in 2023. This is unusually explicit firm-published strategy and leadership evidence, but it still does not identify a model family, provider, training corpus, benchmark, data contract, agent permissions, production endpoint, capital authority, or AI-attributed result. The unverified The Org directory lists two “Quantitative Researcher - AI” entries, but those names and the displayed roster remain discovery leads rather than confirmed DRW personnel. See the capture note.

Capstone Investment Advisors’ current AI Engineer, AI Infrastructure Engineer, and open positions board, accessed August 12, 2026, add a first-party adjacent asset-manager control. The AI Engineer role sits in Tech - Data and describes AI/ML/LLM-driven models, tools, workflows, agents, chat agents, copilots, automation pipelines, RAG, MCP servers and clients, function calling, and internal integrations. The AI Infrastructure Engineer role sits in Tech - Infrastructure and describes secure execution environments, distributed agent development, Kubernetes, observability, compliance, security controls, and support for firmwide AI-augmented development workflows. This is current hiring-intent and architecture vocabulary; it does not prove the roles are filled, that the systems are live, that models have investment authority, or that any return is attributable to AI.

The official Goldman Sachs Exchanges interview with Capstone founder and CEO Paul Britton, recorded February 20 and published February 28, 2025, adds an executive operating-model statement. In the recovered audio, Britton says AI is discussed daily at Capstone, describes a staff of roughly 330–340, sets an ambitious goal of saving 50,000 hours, and tells employees to examine repeatable daily workflows for automation so that people can spend more time on higher-value work. The source is a dated speaker account verified against the official publisher page and direct audio; it does not disclose a model or provider, completed savings, implementation inventory, data rights, permissions, investment authority, or AI-attributed performance. The timestamped capture note records the local transcription and boundaries.

IMC Trading’s current Software Engineer - AI Powered Engineering role, accessed August 12, 2026, adds a prop-trading developer-workflow control. IMC describes an Agentic AI engineering team building agents, MCP servers, retrieval systems, compile/test/evaluation pipelines, static analysis, safety checks, performance gates, code review, telemetry, and production-critical controls for AI-generated code. This is first-party evidence of agentic software-engineering infrastructure. It does not disclose a model inventory, trading-system agent authority, investment workflow, or AI-attributed P&L.

Numerai’s current homepage, NumerCon 2026 recap, MCP documentation, example-scripts repository, and AI Scientist role on its official careers host, accessed August 12, 2026, add a crowdsourced hedge-fund control case. The sources describe an AI-scientist workflow, a Predictive LLM used to generate Faith dataset features, open agent skills, MCP tools for research and submissions, scoped API keys, and hiring intent around autonomous quantitative research and evaluation. This evidence is public and current, but it does not disclose model weights, data-rights controls for the article corpus, independent evaluation, internal portfolio authority, or AI-attributed returns.

A September 1, 2026 title-blind refresh adds two current routes. Numerai’s May Engineering update reports company-described agent-tooling activity: tens of thousands of MCP sessions for model building, prediction uploads, diagnostics, and performance retrieval, plus thousands of clones of the Numerai Skills examples. The official Engineering archive shows this is part of a continuing monthly firm-media stream. Separately, the Numerai homepage routes Careers to Wellfound, whose checked page shows three San Francisco engineering openings: AI Scientist, Senior Quantitative Researcher, and Lead Software Engineer. The AI Scientist posting names Autoresearch, LLM training/fine-tuning, RL, agents, evaluation, text/filing extraction, and researcher-workflow automation; the quantitative and software roles add risk, optimization, data ingestion, ML/CI/CD/ETL, research-platform, and trading-pipeline responsibilities. These are first-party or routed hiring signals, with the Wellfound descriptions hosted outside Numerai. They do not establish named incumbents, model weights, training corpora, data rights, production permissions, portfolio authority, or independently audited performance. See the capture note. The official NumerCon archive exposes two YouTube playlists containing 18 replays. A caption audit found additional title-blind sessions with Victor Haghani of Elm Wealth, Richard Dewey of Allometry Labs, Howard L. Morgan in a historical Numerai conversation, Anson Chu, CrowdCent/NumerBlox, Jo-Fai Chow, and an unresolved guest identified only as “Nick.” The replays add public timestamps for MCP-based model workflows, Faith features, risk forecasts, staking, synthetic data, and platform scoring. They are platform, tournament, guest, and historical evidence; they do not establish autonomous authority over Numerai fund capital.

The archive refresh also promoted three direct 2022 full-session routes that had been described in the capture note but were not previously distinct ledger records: Intelligence, Data, and Money. The captioned Intelligence session describes True Contribution as a feedback loop from staked model signals through a convex portfolio optimizer to portfolio returns and stake updates; the Data session covers benchmark models, data release, and synthetic-data research; and the Money session names guest affiliations including Jonathan Larkin, Joey Krug, and Howard L. Morgan. These are historical Numerai presentations and guest-context routes, not evidence of current implementation, current guest-employer practice, or autonomous capital authority.

August 13 verification delta: QRT, regional controls, and entity separation

This pass added one official QRT support-role detail and several regional first-party pages. QRT’s public Greenhouse API body for Production Support Engineer — AI & LLM ties the role to front-office trading-flow issues, LLM-based systems, monitoring, release management, algorithmic-trading platforms, risk systems, and LLM hosting environments. This is support-surface and hiring evidence. It does not establish a filled role, production endpoint, permission map, trading authority, model identity, or AI-attributed return.

The regional pass added official Chinese-language pages for High-Flyer, Lingjun, Mingshi, WizardQuant, DeepWin / 蝶威量化, and Metabit Trading / Qianxiang / 乾象, plus QTS Canada-linked methodology and legal-entity records, a named Dynamic Funds AI/quant role, and a Macquarie QIS Australia-adjacent control. High-Flyer’s homepage separates AI compute, an AI-used quant hedge fund, and AI basic research, while naming Fire-Fly II, hfai.nn, hfreduce, and 3FS. Lingjun’s pages describe an implemented research agent, ML/deep-learning signal development, AI coding tools for risk-model development, large-model processing of unstructured and multimodal data, Hong Kong/offshore products, and Beijing/Inner Mongolia supercomputing centers. Mingshi’s homepage describes AI-selected factors, nonlinear ML weighting, generated signals, and a factor/AI/optimization/risk-control/trading research pipeline. WizardQuant’s careers pages add first-party hiring and infrastructure language for AI labs, ML platforms, AI clusters, training/inference, model lifecycle work, and trading-scenario loading. DeepWin’s official page identifies Shanghai DeepWin Private Fund Management Co. and describes AI quantitative-investment, reinforcement-learning, and Agent process modules, while Chinese Securities Journal reports describe a firm-hosted DeepWin Agent research workflow. Metabit’s firm page reports private-fund-manager registration P1071383, AI-driven quantitative-trading, RL/POMDP, HPC, DL/ML platform, data-engine, and risk-control language; a JuiceFS customer case adds vendor-published cloud research-infrastructure context. QTS adds a firm article on feature selection in generative-AI-era financial ML, while public filing and LEI records separate the Delaware legal entity from the British Columbia headquarters and principal-place-of-business signal. Dynamic Funds adds a named investment-research support role; Macquarie QIS adds a bank/QIS methodology comparator rather than a hedge-fund-manager peer. These sources are material public signals, but they do not show model inventories, live permission maps, independent deployment audits, or AI-attributed performance.

August 14 verification delta: current hiring corrections

This pass corrected the WorldQuant current-status boundary and added one APAC Point72/Cubist role. WorldQuant’s firm-hosted career index currently lists the LLMs & AI Agents internship, WQBRAIN AI Researcher, London AI Software Developer and Senior AI Software Developer roles, and a Yerevan AI/ML systems role. That updates the older stale-mirror reading of the internship job ID: the WorldQuant-hosted listing is current, while date drift in older Greenhouse detail remains a recruiting-platform caveat. The WQBRAIN role describes LLM fine-tuning and signal/alpha generation on BRAIN, but the public source remains a hiring artifact rather than filled-role, deployment, or investment-authority evidence.

Point72/Cubist’s current AI Data Scientist role adds a Hong Kong/Singapore regional signal. The role sits in Focus Systematic Data and describes AI-powered data products, in-house AI/ML model training, fine-tuning, evaluation, deployment, proprietary and external datasets, LLM use cases, and production governance for systematic data workflows. It is high-confidence evidence for a current official hiring surface; it does not disclose the hire, model versions, live endpoints, trade permissions, or investment attribution.

August 14 source-expansion pass: new channels and findings

The second source-family pass widened the search beyond wires and ordinary firm news pages. The new raw source ledger records the retrieval date, source family, direct links, and disqualification boundary. The result is a wider evidence map, not a cross-firm score.

Official firm pages and investor-relations archives

• Acadian: its current news index shows July 2026 hiring activity in Southeast Asia, portfolio management, trading, and North American credit distribution. The May 2026 investor-forum release describes systematic-credit and global-wealth expansion and a tax-aware long/short launch. The primary investor-forum transcript adds management’s account of a roughly 70-person data/data-infrastructure group, an NLP production history dating to 2008, internal AI-research-agent use, and a human-reviewed boundary around investment judgement. These are direct management disclosures about operating context; they do not establish current model versions, permissions, or return attribution.
• Man Group: the Anthropic partnership names Claude, Claude Skills, Claude Code, and agentic workflows across investment and corporate functions. Its AlphaTrend research note describes pre-specified research stages, parallel signal proposals, and a comparison of Claude 4 Sonnet and GPT-5 that found different proposal correlations and creativity profiles. The 2025 annual report reports more than 100 internal AI plugins and investment-team reorganisation. These are useful disclosures about workflow design and organisational activity; they do not prove a single production model or provide an independent performance comparison.
• Millennium: the June 2026 AI Lab Q&A describes a separate environment for early-stage experimentation, AI-company collaboration and potential co-development, and specialist hiring. It adds a current organisational layer to Millennium’s technology and AI-advisory material without disclosing partner names, model inventory, or investment authority.
• CFM: current official material describes a machine-learning lab and generative-AI investment, with text/video/audio extraction for sentiment and nuance, classification and network discovery, automated research, and agentic test-and-code workflows. Investing After the AI Honeymoon and Why Systematic Global Macro Is Having a Moment support those claims. The older financial-news corpus case study describes retraining generic neural networks on a specific corpus, while the Hugging Face case study is a narrower financial-NER fine-tune. This separates historical model adaptation from current multimodal and agentic research language.
• Arrowstreet: the current official homepage reports $344B+, 440+ clients, and 500+ employees as of June 30, 2026, with data science, high-performance computing, and next-generation-platform language. This supersedes the older March 31, 2026 observation. It is a dated scale disclosure, not evidence about AI capability or investment results.

Vendor, partner, and customer surfaces

Anthropic’s financial-services agents page adds a partner map that the wire and careers sweeps cannot provide: Dun & Bradstreet, Fiscal AI, Financial Modeling Prep, Guidepoint, IBISWorld, Intralinks, Third Bridge, Verisk, Moody’s, and FactSet are named in data/MCP or workflow contexts. The same page quotes Citadel’s Head of Core Engineering on Claude for Excel and Walleye’s CEO/CIO on Claude Code adoption. Anthropic’s Claude for Financial Services page adds Box, Daloopa, Databricks, FactSet, Morningstar, Palantir, PitchBook, S&P Global, Snowflake, and implementation partners including Accenture, Deloitte, KPMG, PwC, Slalom, TribeAI, and Turing; it also quotes Bridgewater’s AIA Labs CTO on an Investment Analyst Assistant. These pages establish a public vendor ecosystem and attributed customer statements. They do not establish that every manager uses every partner, nor that a named data vendor is connected to a particular live strategy.

Balyasny’s OpenAI case study adds model-evaluation detail: a central Applied AI group, internal benchmarks, tests across forecasting accuracy, numerical reasoning, scenario analysis, and robustness to noisy inputs, and task-by-task selection between GPT-5.4 and internal models. Jane Street’s CoreWeave agreement adds an approximately $6B AI-cloud commitment and $1B equity investment, while CoreWeave’s GTC 2025 recap describes a Jane Street/CoreWeave session about training and fine-tuning quantitative-trading AI models. These are vendor/company disclosures. They do not publish Jane Street’s complete model workload, permission map, production endpoint, or AI-attributed return record.

Walleye is a newly discovered candidate for the tracked universe. Anthropic’s page and Walleye’s official site describe a roughly 400-person multi-strategy manager and broad Claude Code adoption; Walleye’s company LinkedIn page also reports early testing of GPT-5.5 and Opus 4.7, plus recent partner promotions. Those claims are retained as company/vendor statements. They are not upgraded to independent model tests or evidence of capital authority.

The title-blind Every AI & I transcript adds an attributed account from Will England, Walleye’s CEO, CIO, and managing partner. In the May 2025 interview, England says roughly 75% of employees used ChatGPT weekly or almost daily, roughly one-third used AI coding tools such as Windsurf, every fundamental-investing team used the internal Current product, and the firm required baseline AI proficiency across departments. He also says the firm recorded nearly all internal Zoom and risk calls and used LLMs over their transcripts for institutional memory, recurring-risk context, and exploratory prediction under the name “Borg.” The transcript does not disclose Current’s model or data architecture, recording permissions, evaluation design, portfolio authority, or AI-attributed performance. Every identifies itself as helping Walleye with AI training and implementation, so these remain self-reported statements in a commercially involved publisher context. See the capture note.

Every’s broader finance-consulting archive adds a workflow and tooling route without naming most client firms. Its investment-analyst playbook describes a progression from ChatGPT projects over SEC filings and earnings releases, through Claude Skills and Daloopa via MCP, local proprietary models and notes in Claude Cowork, to a Claude Code command that maintains a recurring investment dashboard. The companion consulting-practice account describes custom plugins, shared skills, defined-job agents, and an unnamed fast-growing hedge fund whose initial automation opportunity was in compliance, operations, recruiting, and administration rather than investment decisions. These are vendor-described workflows and anonymized client claims; they do not establish a named fund’s production stack, permissions, or performance. See the capture note.

A second Every AI & I transcript, with head of consulting Natalia Quintero, adds implementation detail from an unnamed private-equity client. The client mapped research, diligence, market mapping, portfolio management, and other investor tasks by team; Every reports that tailored prompts over a decade of SharePoint investment-thesis material reduced a draft investment memo from a stated two-to-three weeks to roughly 30 minutes. Quintero also describes Every’s own Claudie project-management agent, built in Claude Code with skills, commands, subagents, and repeated discarded iterations. These are vendor and anonymized-client statements, not evidence about a named hedge fund, and the transcript does not provide a controlled time study, error rate, model configuration, permissions, or performance attribution. The source is retained in the Every workflow capture note.

Every’s 2026 predictions archive also says Natalia Quintero used Claude Code and ten subagents to help an unnamed hedge fund build a market map of AI tools, including CEO profiles and company differentiation. The first version is described as taking about two hours. The fund, source universe, prompts, review process, and downstream use are not disclosed, so this is a vendor-described research-agent route rather than evidence about a named manager.

An older but distinct vendor-media route is the FinTech Files interview with Exabel’s Mark Fleming-Williams, published June 23, 2021. The captions describe an interface that combines alternative and traditional data, applies quantitative machine-learning methods to historical relationships, and lets an investor form forecasts as new data arrives (17:12–20:17). The discussion also separates sourcing, cleaning, entity mapping, enrichment, and analysis, and describes larger hedge funds doing more of the cleaning themselves as part of their process (28:04–31:00). This is a vendor and general practitioner account, not evidence that a named manager used Exabel, nor evidence of a customer contract, model inventory, data rights, production permission, or investment result. See the capture note.

The title-blind vendor pass also resolved a deeper LinqAlpha route. Its AI Lab, introduced August 18, 2026, frames the program around understanding, measuring, and building “Alpha Intelligence.” LinqAlpha’s Third Square Capital customer story describes a Devil’s Advocate Agent that challenges theses against filings, expert-call transcripts, and industry commentary. An AWS partner case study adds a four-stage flow—thesis definition, document ingestion, assumption decomposition, and source-linked counterarguments—and names a multi-agent implementation using EC2/Python orchestration, Textract, S3, RDS, OpenSearch, Claude Sonnet 3.7 for document-structure enrichment, and Claude Sonnet 4 for retrieval and rebuttal generation through Bedrock. The post names LinqAlpha personnel including Suyeol Yun, Jaeseon Ha, Subeen Pang, and Jacob (Chanyeol) Choi, and links the FinAgentBench paper, whose 26,000 expert-annotated examples separate document selection from key-passage retrieval. Choi is also a named author of FinDER, whose public dataset record contains 5,703 expert-generated finance query, evidence, and answer triplets. The separate DeepSeek-R1 function-calling repository exposes iterative tool calls, structured-output validation, reasoning traces, and recovery from tool or parse failures. These are inspectable research and engineering artifacts; they do not establish that a named client uses the dataset, repository, or model in production. LinqAlpha and AWS also make customer-count, agent-count, and speed claims; those remain partner/customer marketing claims without an independent audit, model registry, data-rights map, permission map, or return attribution. AWS calls the company Boston-based, while LinkedIn lists New York headquarters and a Cambridge location, so the geography is recorded as unresolved rather than treated as a single headquarters fact. See the capture note.

A Korean-language media pass adds a material but separately bounded layer. Three reports spell the company LinkAlpha rather than LinqAlpha; the leadership names and institutional-investment product are consistent with the existing LinqAlpha record, so this is retained as a likely naming variant pending a corporate legal-name cross-check. The August 20 AI Lab report describes an in-house lab with five full-time researchers in New York and Seoul, UNIST Professor Yongjae Lee as research lead, and University of Florida Professor Alejandro Lopez-Lira as academic adviser. It reports research on model interpretation of financial information, reliability of AI judgments, alpha/risk systems, more than 13 papers at ICML, ACL, and ACM ICAIF, and the FinDER and FinAgentBench datasets. It also reports joint research with J.P. Morgan, BlackRock, Blackstone, State Street, and MIT. These are publisher and company-reported statements; the public report does not enumerate the full lab, provide collaboration agreements, or show that the datasets or papers are deployed in a client portfolio.

The Series A coverage and platform report describe a 34 billion won financing round, a platform combining market data, disclosures, research, internal documents, and meeting records, and the ability to route different tasks to OpenAI, Anthropic, and Gemini. They name Fidelity, BNP Paribas, and Schonfeld as institutional users or clients, while the Seoul report repeats a company claim about a first-place MTEB search/embedding result in May 2024. Those are vendor/media claims, including AI-translated reporting; they do not establish a specific customer’s model routing, score construction, data permissions, or investment authority. MoneyToday’s board report adds John Chang as a named board member, with prior Barclays and Deutsche Bank leadership roles, and a global expansion/hiring route. The capture note preserves the spelling, source-language, personnel, funding, and evidence boundaries.

The same first-party surface now advertises an invitation-only LinqAlpha AI for Finance Summit in London on September 16, 2026, hosted at Deutsche Bank. Its stated theme is the convergence of systematic and discretionary investing under agentic AI. The page names Ed Ralph (Executive Director, Data Science Strategy), Alistair Jackson (Senior Business Analyst), and Caio Natividade (Managing Director, Global Head of QIS Research), and lists a panel titled “Signals at Scale: Agentic AI in Systematic Investing.” It also says the session is closed-door, with no recording or press under the Chatham House Rule. This adds an upcoming conference and personnel route, but the public page does not expose attendance, private discussion content, model/provider choice, data rights, permissions, or deployment. See the updated capture note.

The same first-party route exposes a separate AI for Finance Summit 2026 Asia held in Seoul on July 10, 2026. Its public roster links Atlas Wang of XTX Markets, Srijan Sood of J.P. Morgan, Joo Lee of Arrowpoint Investment Partners, Chang Hwan Sung of Zentific Investment Management, LSEG, AWS, LG AI Research, and multiple Korean securities and asset-management organizations. It also lists Lukasz Szpruch of the Alan Turing Institute and academic hosts from MIT and UNIST. LinqAlpha’s post-event recap reports approximately 100 attendees and frames the discussion around verification bottlenecks, runtime governance, domain-specific models, and Asian-language fragmentation. The event was under the Chatham House Rule and no public recording or transcript was identified. The roster and recap are evidence of public programming and LinqAlpha’s interpretation of the event, not evidence of any named firm’s deployed model, partnership scope, data rights, agent permissions, or investment performance. Joo Lee’s personal post is retained as a self-described Arrowpoint perspective rather than an Arrowpoint-wide policy.

The LinqAlpha route also exposes a public engineering artifact. Its DeepSeek-R1 agent repository names Suyeol Yun, Subeen Pang, Yongjin Kim, and Chanyeol Jacob Choi and documents iterative function calling through Fireworks, structured JSON validation, database/search tools, bounded iterations, and recovery from tool or parsing failures. A public team announcement describes proprietary database and agentic workflows for investment research; those claims are kept as personnel statements rather than independent production evidence. LinqAlpha’s current AI Lab index additionally lists research on LLM investment bias, financial embeddings, agentic retrieval, disclosure signals, mention-market forecasting, and semantic filtering for prediction-market lead-lag signals. An OpenBB partner post describes an API route for company screening, earnings-transcript analysis, and tailored dashboards, with a Singapore event demonstration. These artifacts show a public research, engineering, and partner surface; they do not disclose customer data permissions, model fine-tuning, live portfolio authority, or AI-attributed returns.

An additional Asia-linked vendor route is MaxQuant AI Platform. An AWS case study describes Bedrock AgentCore, Lambda, EventBridge, DynamoDB, and API Gateway supporting a finance platform whose reported use cases include data parsing, strategy backtesting, customer profiling, multi-strategy portfolio optimization, feature extraction, model invocation, and strategy evaluation. A Chinese-language Future Securities partnership report adds the product names News Intelligence Engine, Predictive Analytics Core, and Cross-Asset Intelligence across equities, commodities, foreign exchange, and crypto. These are vendor/customer and secondary-media claims, not proof of a hedge-fund deployment, regulated-manager status, order authority, or performance. The AWS case names Zeng Jingfeng as CEO, while the partnership report quotes Jeffery Zeng as founder; the reviewed sources do not establish that those names refer to the same person. See the capture note.

The directly retrieved MaxQuant site adds company- reported modules for multi-agent sentiment and narrative analysis, volatility/momentum/trend-reversal monitoring, dynamic position sizing, stress testing, stop-loss optimization, and strategy-attribution views. It also claims “10,000 AI strategy evolutions daily” and labels the product early access. These are first-party marketing claims; the site provides no model card, evaluation protocol, customer deployment record, permission map, or independent performance verification.

Personnel and title corrections

Two Sigma’s current AI outlook, Part I identifies Matt Greenwood as Chief AI Innovation Officer, Jeff Wecker as CTO, Jin Choi as Head of Technique Forecasting, Mike Schuster with AI Core, and Ben Wellington as Head of Complex Feature Engines. Part II adds cautions around model scaling, interpretability, multimodality, knowledge-cutoff leakage, and backtest overfitting. This corrects the earlier blanket statement that the reviewed record had no named GenAI owner; the title is now directly evidenced, while ownership of a particular model remains undisclosed.

Millennium’s official Gideon Mann profile continues to identify him as Global Head of Artificial Intelligence, while the AI Lab Q&A identifies Vlad Torgovnik as CIO and explains the lab’s remit. Acadian’s official news archive is now a required temporal source for 2026 appointments. Walleye’s official people page lists Will England as Managing Partner, CEO, and CIO; Robert McGehee as CIO of Quant; and Luke Anderson as CTO. These are title and organisational signals, not proof that a named executive owns a particular model or agent.

What remains missing

The newly found pages still do not provide a full model registry, training-data rights, weights, reproducible benchmark results, current adoption by business unit, agent permission maps, production-stage labels, or audited return attribution. “Fine-tuning,” retrieval, prompting, post-training, early access, and pretraining remain separate evidence labels. A partnership is not treated as data use unless the source explains that use. A job title is not treated as a filled role. Vendor pages and LinkedIn feeds are preserved with retrieval dates because their text and partner lists can change.

August 14 lab and academic-lineage pass

The new AI-lab and academic-lineage ledger adds university-industry chairs, author pages, paper graphs, conference bios, parent-company and regulatory archives, technical artifacts, and hiring APIs as separate discovery channels. The central distinction is whether a firm names a lab, exposes an internal AI group without using that label, or only shows AI-adjacent hiring and research.

Short list of named labs and lab-like internal groups

Firm	Public classification	Named people located	Evidence boundary
Bridgewater	AIA Labs, a named AI research and investment lab	Greg Jensen; Blake Cecil, Deputy CIO, Alpha Engine and AIA Labs; Nina Lozinski, Head of AIA; Aaron Linsky, Head of Engineering, Alpha Engine; Oliver Simon, Head of AI & ML Investment Strategy; Jasjeet Sekhon (historical Bridgewater role; now Google DeepMind); Rohan Alur and Daniel Kang (research coauthors)	Firm pages and papers expose the lab’s remit, current partnership titles, and artifacts; they do not expose a complete current roster, model registry, permissions, or independent return attribution
CFM	ML Lab, a named ML research lab	Eric Vanden-Eijnden, Anastasia Borovykh, Giulio Biroli through the CFM-ENS chair	Named remit, academic bridge, and detailed AI-native hiring are public; current team membership and deployed model ownership are not
Millennium	AI Lab / dedicated early-stage experimentation environment	Vlad Torgovnik, Gideon Mann; Vaibhava Goel remains a secondary-source lead	The firm describes experimentation, partner collaboration, and hiring; it does not publish a lab roster or partner/model inventory
Schonfeld	FE AI Lab backed by SchonAI	No named FE AI Lab researchers located in the first-party article	The program, model partnerships, and pilot controls are public; lab staffing and academic lineage remain open
BlackRock	BlackRock AI Labs; adjacent traditional asset-manager comparison	Rachel Schutt, Stephen Boyd, Trevor Hastie, Rob Tibshirani, Emmanuel Candes	Public lab leadership, publications, and finance applications are visible; product-level deployment and attribution are not
Balyasny	Central Applied AI group, not consistently called a lab	Charlie Flanagan; additional team names require corroboration	Centralisation, evaluation, BAM ChatGPT, and agent workflows are public; the complete researcher roster is not
Citadel / Citadel Securities	Internal AI/ML research and platform groups, not a publicly named lab	William Hamilton, Jianbo Chen, Honglin Yuan, Wei Tan, Yifeng Tao	Academic profiles and AI/ML hiring are public; the link from individuals to the disclosed assistants or GQS systems is not
Man Group / Man AHL	Central AI platform plus Oxford-Man joint research laboratory	Matthew Hertz, Tushara Fernando, Anthony Ledford, Martin Luk	AI strategy and academic-commercial links are public; a separate internal AI-lab roster and model ownership are not
Two Sigma	AI Core / AI Innovation and research platform, not a formally named lab	Matt Greenwood, Mike Schuster, Jeff Wecker, Lei Chen, Gene Li, Kishor Jothimurugan, Qiwen Cui, Lingxue Zhu, Myung Jin Choi	Titles, research topics, and hiring are public; the formal lab boundary, model ownership, and trading permissions are not

This is a classification table, not an assessment of capability. “Named lab” means the firm itself uses lab language or describes a dedicated lab. It does not mean that every activity is production or that the lab controls capital.

MIT / Harvard / Stanford lineage → public employment map

This is a dated employment map, not a university leaderboard. “Current” means the cited firm, university, or personal page presents the affiliation as current at retrieval; “former,” “incoming,” “fellow,” and “academic partner” are kept separate. The map covers the firms in this article and the adjacent comparison points where the public record is unusually specific.

Person	Public education or research link	Public employment / affiliation located	Status and boundary
Rohan Alur	MIT EECS PhD candidate; advisors Manish Raghavan and Devavrat Shah	Bridgewater Principal Research Scientist; AIA Forecaster	Current on his public profile; Bridgewater role and MIT research are direct, but a portfolio assignment is not disclosed
Madalina Persu	MIT PhD and MEng in CS; Harvard mathematics BA	Two Sigma Quantitative Researcher, VP	Yale’s 2025 event page says she has been at Two Sigma since 2018 and works on AI for predictive mid-frequency equity signals; current title should still be rechecked before publication
Abe Ejilemele	MIT MEng and BS in CS	Two Sigma Quantitative Researcher	Current on Two Sigma’s profile; the source describes his path from internship to full-time researcher, not a specific model
Sarath Pattathil	MIT EECS PhD under Asu Ozdaglar	Two Sigma Quantitative Researcher	His public page says “as of July 2023”; retain as date-stamped employment, not an unqualified 2026 claim
Matthew J. Staib	MIT EECS PhD; Stanford MS EE and BS Mathematics	Two Sigma Quantitative Researcher, according to public CV	CV says 2020–present, but no current firm bio was found in this pass; date-sensitive lead
Cenk Baykal	MIT CS PhD	Former Two Sigma quantitative researcher; public profile currently identifies Google Research	Former Two Sigma, current Google Research on the public profile; do not count him as current Two Sigma
Michael R. Douglas	Harvard BA; Caltech PhD; current Harvard CMSA research scientist	Former Renaissance Technologies researcher, 2012–2020	Harvard’s official profile supports the dated former affiliation and current machine-learning research; it does not identify Renaissance project ownership or a current hedge-fund role
Jonas Metzger	Stanford PhD Economics; Stanford Data Science postdoc	Citadel Quantitative Researcher, Alpha Research	Current on his personal profile; Stanford training and Citadel employment are direct, but the public page does not assign a named AI system
Honglin Yuan	Stanford ICME PhD under Tengyu Ma	Citadel Securities Quantitative Researcher	Current on his public profile; research covers deep-learning theory, optimization, and federated learning, not a disclosed firm model
Benoit Zhou	Stanford graduate; teaching and research activity across Stanford AI/ML courses	Citadel Securities Quantitative Researcher, options	Current on his public profile; the page does not establish a Stanford degree type or a Citadel AI assignment
Jason Cheuk Nam Liang	MIT Operations Research PhD	Citadel Quantitative Researcher, listed as first position by an MIT faculty alumni page	Faculty alumni page placement evidence; current title and AI scope require a firm-controlled refresh
Wenyu Chen	MIT Operations Research PhD	Citadel Quantitative Researcher, listed on an MIT advisor CV	Academic CV placement evidence; current title and AI scope require a firm-controlled refresh
Pratik Rathore	Stanford PhD in Electrical Engineering	Citadel Securities Quantitative Researcher, starting fall 2026	Incoming appointment, not current employment as of retrieval; research is optimization and ML training related
Charlie Flanagan	Harvard Master of Software Engineering; Stanford ML/Python instructor	Balyasny Chief AI Officer	Current on Balyasny’s leadership page; Stanford is teaching history, not a Stanford degree claim
Rachel Schutt	Stanford MS in Engineering-Economic Systems/OR; Columbia Statistics PhD	BlackRock AI Labs co-head	Current on BlackRock’s official biography; BlackRock is an adjacent asset-manager comparison, not a hedge fund
Stephen Boyd	Stanford professor	BlackRock AI Labs co-head / research leader; Stanford faculty	BlackRock names him in AI Labs leadership; Stanford faculty status is separate from any firm employment contract
Jeremy Andre	Stanford MSc Financial Mathematics	Man Group / Man AHL Senior Quant, Portfolio Monetisation	Current on Man’s official profile; prior Balyasny employment is also disclosed, but the page does not make him an AI-lab employee
Serginio Sylvain	MIT mathematics and economics BA	GMO Systematic Equity quantitative analyst and partner	Current on GMO’s official research page; GMO publicly connects his team to NLP infrastructure, but not to a named GenAI model
Chris Heelan	Vanderbilt electrical/biomedical engineering; Brown graduate training in electrical engineering and innovation management	GMO ESG Research Team researcher; previously GMO Machine Learning Development Lead for Investment Data Solutions	Current role and former ML-development title are on GMO’s official page; no public link to Super Analyst, current GenAI ownership, or production deployment
Jasjeet Sekhon	Former Harvard professor; Yale and UC Berkeley professor; Cornell PhD	Former Bridgewater Chief Scientist / Head of AI; current public profile also lists DeepMind	Treat Bridgewater as former or role-date-sensitive, not current without a refreshed firm bio; AIA coauthorship is direct, model ownership is not
Suproteem Sarkar	Harvard Economics doctoral student	Two Sigma Fellow	Fellowship and academic affiliation, not proof of employment by Two Sigma
Jun Liu	Harvard professor	Two Sigma academic partner	Academic partnership, not employment; Two Sigma’s university page names the relationship

The negative result is also informative. This pass did not locate a current MIT/Harvard/Stanford-trained head of AI on a firm-controlled Acadian, Arrowstreet, CFM, or Millennium biography. Acadian has public ML/AI strategy material and hiring signals; CFM and Millennium have named AI groups; and Arrowstreet has a historical MIT CSAIL industry relationship. None of those facts should be converted into a named-person employment claim without a person-specific, date-stamped source. A search result for Arrow Electronics and MIT CSAIL was excluded because it is not Arrowstreet Capital.

Personnel-to-research linkages

CFM. CFM’s ML Lab announcement names Eric Vanden-Eijnden and Anastasia Borovykh and connects the lab to Giulio Biroli through the CFM-ENS Data Science Chair. Borovykh’s public research page identifies her as an Executive Director at CFM working on machine learning for alpha, after Imperial ML faculty work; her public papers include deep-network dynamics and federated-learning leakage. Vanden-Eijnden’s NYU publication page includes neural-network trainability, density estimation, and active importance sampling for rare data. Biroli’s ENS profile lists high-dimensional statistics, machine learning, random matrices, and generative-AI theory. These topics are technically compatible with CFM’s public language around weak/non-stationary financial data, predictive models, and multimodal extraction. That compatibility is an inference; no source ties a named paper to a live CFM alpha model.

CFM’s current AI-native quant-researcher role is a particularly valuable source: it names target definitions, training recipes, distributed foundation-model training, MoEs, long-context transformers, vision-language models, RLHF/RLAIF/DPO, evaluation, monitoring, and GPU optimisation. It is detailed hiring intent, not proof that each method is deployed or that the role is filled.

Bridgewater. The public AIA Labs page lists current research artifacts on PAT, expert-judgment replication, AIA Forecaster, and RLVR generalization. Rohan Alur’s research page identifies MIT EECS training with Manish Raghavan and Devavrat Shah and links his small-data, human-judgment, and LLM work to Bridgewater. Jasjeet Sekhon’s Yale profile records a Cornell PhD, University of British Columbia undergraduate training, and research in causal inference, interpretable ML, and experimental design. Daniel Kang’s publication page includes agent evaluation, reinforcement-fine-tuning transfer, and safety research; his coauthorship on AIA publications is direct evidence of collaboration, not a public internal title.

The research-to-workflow linkage is unusually legible here: causal inference and interpretable ML connect to credible judgment; small-data forecasting connects to sparse event prediction; RLVR work connects to post-training and verification; and the AIA papers explicitly discuss financial information filtering and forecasting. The source still does not provide an independently reproduced map from those papers to live portfolio positions.

Citadel and Citadel Securities. Public researcher pages reveal a broader academic talent graph than the firm’s AI product pages. William Hamilton’s profile covers graph representation learning, network science, and NLP. Jianbo Chen’s profile records a UC Berkeley Statistics PhD advised by Michael Jordan and Martin Wainwright, with ML, statistics, and optimisation interests. Honglin Yuan’s profile records Stanford ICME training under Tengyu Ma and work in deep-learning theory, optimisation, and federated learning. Wei Tan’s profile records a Tsinghua PhD, IBM Watson research, GPU/distributed ML, and deep learning/NLP in quantitative finance. Yifeng Tao’s profile records Carnegie Mellon ML/CS training and computational-healthcare ML.

These backgrounds suggest several research lanes—structured/graph data, statistical learning, distributed training, and unstructured data—but they do not establish that these people built Citadel’s public equity assistant, Claude-for-Excel workflow, or GQS models. The firm/entity distinction between Citadel LLC and Citadel Securities is retained.

The official Citadel Securities YouTube archive surfaced 85 uploads on August 28, 2026, including several previously untracked, title-specific routes. A fresh subtitle retrieval recovered public English VTT tracks for all seven selected records, upgrading the longer items from metadata-only to timestamped automatic-caption evidence. Peng Zhao’s CEO-on-AI short frames new AI applications around the firm’s quantitative-research, compute, machine-learning, and AI operating description. A Jim Esposito item from the Milken Institute Global Conference divides the AI discussion between operating-cost reduction and client experience, and links data ingestion, curation, risk-model revision, pricing, and speed to market (00:00–01:34). The longer Peng Zhao–Alfred Lin discussion names Jeff Maurone as COO of Technology and Low Latency and lists AI, automation, and client experience as topics at the October 2025 Future of Global Markets event. Its caption-backed discussion separates machine-learning models from LLMs and agents (04:46–05:17), describes unique data, data-use insight, algorithm-capable people, and workflow as four ingredients in an AI capability (13:52–14:28), and discusses coding agents, search, summarization, compliance forms, and the choice between automating a workflow and waiting for tools to improve (32:45–35:10). The discussion also places conceptual problem selection and asking the right question among remaining human contributions (38:40–39:07). Two related conference records and welcome-address metadata expose event scope around AI, machine learning, digital assets, event contracts, and regulation. Finally, title-blind recruiting shorts for Yiming and Sriram describe statistically tested signal research and systematic-equity-options quoting work (00:00–01:21, 00:00–01:21). These are firm-controlled descriptions, captions, and recruiting narratives; captions are not speaker-adjudicated, and no selected item supplies a model inventory, dataset, evaluation, permission map, autonomous trading authority, or performance record. See the archive capture note.

Two Sigma. The current AI outlook names Matt Greenwood as Chief AI Innovation Officer, Jeff Wecker as CTO, Jin Choi as Head of Technique Forecasting, Mike Schuster with AI Core, and Ben Wellington as Head of Complex Feature Engines. The surrounding talent graph contains researchers with public pages: Gene Li’s RL and decision-making profile, Lei Chen’s transformer and graph-learning papers, Kishor Jothimurugan’s formal-methods and RL profile, Qiwen Cui’s multi-agent-RL profile, Lingxue Zhu’s high-dimensional and network research, and Myung Jin Choi’s MIT statistical signal-processing profile.

The potential linkages are method-level: reinforcement learning and formal methods relate to agent control; graph and network work relates to structured alternative data; signal processing relates to noisy measurements; and transformer research relates to representation learning. Two Sigma’s own outlook also explicitly discusses leakage, interpretability, multimodality, and backtest overfitting. These are useful public research clues, not proof of which researcher owns a production model.

Millennium and Balyasny. Gideon Mann’s Millennium biography records Brown and Johns Hopkins training, Google Research NY, and Bloomberg ML leadership. His public NLP record includes BloombergGPT, but that paper is not evidence of a Millennium model. Matthew Hertz’s Man profile is a platform signal for Man’s central AI strategy rather than an academic-paper signal. Balyasny’s official Charlie Flanagan profile confirms the Chief AI Officer role, Columbia MBA, Harvard software-engineering master’s, and Google/QuantRes background. A public applied-AI panel describes the central team as six engineers and six researchers; its roster is not independently confirmed here. Balyasny’s public model-evaluation disclosure is kept separate from its public academic-lineage disclosure.

Other source families to keep mining

The next high-yield layers are university chair seminar archives, author-level coauthor graphs, conference speaker pages, firm and parent annual reports, regulatory exhibits, public patents, model cards, GitHub organisations, hiring APIs, podcasts, and technical-event recordings. For every new lead, preserve the source date and classify it as firm-controlled, vendor/customer, academic, recruiting, social, or third-party. The article will not treat paper-topic overlap as evidence of alpha or deployment without an explicit firm-controlled connection.

August 14 follow-up: leadership, conferences, partners, and role history

Four additional passes were run after the academic-lineage map: identity resolution, conference/video mining, provider and data-partner extraction, and dated role-history reconstruction. The four raw ledgers are leadership resolution, conference/video mining, provider and data partners, and role history.

Newly resolved personnel signals

The first-party and event-host sources materially improve the dated personnel map without proving model ownership:

• Tower’s official seminar article identifies Ramit Sawhney as Global Head of Core AI and ML and describes work on agents, tool use, structured data, and human feedback. His public research history includes NLP, time series, multimodal finance, and agentic systems; that is a research-lineage signal, not a disclosed production model.
• Jump’s March 2026 infrastructure announcement identifies Joe Stam as Head of Research Technology and Alex Davies as CTO. It describes broad model experimentation, accelerated infrastructure, and a long-running NVIDIA relationship, but does not disclose a model registry or strategy assignment.
• The Georgia Tech AI and Future of Finance program provides event-date title evidence for Charlie Flanagan at Balyasny, Vaibhava Goel at Millennium, Mike Moreau at Schonfeld, Ramit Sawhney at Tower, Vladimir Zdorovtsov at Acadian, and Gerardo Rodriguez at BlackRock. These are dated speaker-bio records; they should not be silently treated as August 2026 firm biographies.
• Point72’s official leadership page confirms its current CTO, Cubist, and algorithmic-trading leadership lanes. A separate “Head of AI ML Foundations” identity remains unresolved.

Third-party directory candidates for Citadel Securities and WorldQuant remain discovery leads only. Arrowstreet’s official technology page confirms the technology function and platform remit, but not the person holding the unresolved Head of AI Engineering title.

Conference and video mining

The Georgia Tech conference is now a priority source because its March 2026 program places multiple tracked firms in one public implementation panel. The CEPR–Imperial–Goldman Sachs hedge-fund conference adds an academic/practitioner paper trail. CQF’s official AI and Machine Learning in Quant Finance conference page lists a September 16, 2026 online event with Petter Kolm, AllianceBernstein Director of AI Research Yuyu Fan, and quantitative portfolio manager Tony Guida. The page describes Fan’s news-derived company-network and graph-neural-network covariance work, and Guida’s public LLM/knowledge-graph investment-process profile. These are dated speaker and research-topic signals; the event is not yet a recording or deployment record. Finteda’s London events add a practitioner and vendor surface, while Jump’s public ICML 2026 material is retained as a lead until the official program or recording is archived.

The retrieval rule is strict: capture the official agenda, speaker title on the recording date, slides or transcript, named tools and datasets, evaluation language, and permission boundaries. Event attendance or sponsorship is not treated as a partnership or deployment claim.

The title-blind replay search also recovered a Harrington Starr FinTech Focus TV episode recorded at Future Alpha: Alan Parry, CTO at YellowDog, interviewed by Oliver Knight, published June 29, 2026. The Amazon Music listing and other directory copies identify a roughly five-minute episode about compute, cloud infrastructure, simulations, machine-learning workloads, AI, and research productivity. This adds a useful vendor and event-media route, and research productivity. The embedded Harrington Starr TV recording exposes an English automatic-caption track. At approximately 00:39–01:06, Alan Parry describes compute orchestration across cloud and on-premises environments for simulation, machine-learning training, inference, HPC, and batch workloads; at approximately 01:16–01:21, he mentions spot compute and lower cost; and at approximately 02:06–02:11 and 02:40–02:56, he frames the platform around quantitative-trading institutions and compute-intensive research. These are speaker-reported vendor statements from automatic captions; the temporary VTT SHA-256 is dd4895408bc8e9e92e33e5f84e727df4936ee772608f38677ee7f5eca5b97312. The recovered recording does not name a hedge-fund customer or disclose a model, dataset, permissions, customer cost figure, trading authority, or performance result. It is therefore not assigned to Bridgewater, QRT, or any other fund that has a separate YellowDog customer story. The episode and its cross-platform copies are catalogued in the conference source note.

A systematic sweep of the same publisher’s feed recovered additional title-blind episodes. Capital Markets 2026 names LSEG, Tech Advocates, and Asim Farooq of an unnamed proprietary trading firm while discussing Asian market access, Hong Kong/Mainland China, extended trading, zero-DTE options, infrastructure, and AI as an augmentation layer with human oversight. AI in Capital Markets Starts With the Foundations names five vendor leaders and describes data readiness, governance, front-office chat automation, iterative coding-agent loops, and human review. These pages expand regional, infrastructure, and control-plane coverage, but do not identify the proprietary firm, a fund’s model, or a live deployment.

The Evolution of AI in Capital Markets adds a dated Rimes AI-product leadership signal: Theo Bell is identified as Head of AI Product, with prior Palantir and Goldman Sachs experience. The publisher describes an AI lab, later embedding engineers in product teams, small task-specific agents, support-ticket triage, and a data-environment digital twin. This is a vendor operating account, not evidence of a hedge-fund customer or an investment model. The Structural Shift in Market Data adds a market-data-provider discussion of alternative data, prediction-market information, and agents as direct data consumers, without naming a fund client.

Finally, America’s Trading Conference 2025 with the FIX Trading Community previews a buy-side AI panel with four unnamed firms and separately names Jane Street in a 24/5-trading panel. The page does not establish that Jane Street was part of the AI panel, so it remains an agenda-discovery record rather than Jane Street AI evidence. The official final agenda does resolve that AI-panel roster: Iro Tasitsiomi (T. Rowe Price), Jean-Marc Nosbusch (Crescent Bay Capital Management), Rashmi Rao (L&G Asset Management America), and Travis Ni (GIC), moderated by Kevin Houstoun. FIX’s speaker biographies describe Iro as Head of AI & Data Science at T. Rowe Price and Travis as Head of Equities Trading Americas at GIC with an AI-development remit. This is an official agenda and dated biography surface; it does not prove attendance, recording content, or production deployment. The five YouTube routes and their caption boundaries are recorded in the conference source note.

An earlier title-blind Harrington Starr episode with Usman Khan, Founder of APEX:E3, adds a vendor-side AI and data-platform surface. The publisher describes models processing SEC filings and global market research into actionable insights. The linked YouTube recording adds timestamped discussion of an asset-class-agnostic data architecture, agent/tool orchestration for selecting among quantitative models, SEC filings and long documents as automation targets, and a multilingual research interface with a claimed connection to more than three million research papers. A 2024 Deus X Capital announcement claims that APEX:E3’s A.L.I.C.E. uses models from Google Cloud, OpenAI, Meta, Anthropic, and its own cognitive architecture, alongside an Oxford collaboration. These remain company, publisher, and investor claims—not independent evidence of a named tracked fund’s use, model weights, permissions, evaluation, or investment results.

Provider and data-partner map

The new registry separates firm-specific relationships from ecosystem-only availability. Directly attributable rows now include Balyasny–OpenAI, Man–Anthropic, Jane Street–CoreWeave, Walleye–Anthropic, Citadel Securities– Anthropic workflow evidence, Bridgewater–Anthropic assistant evidence, CFM–Columbia/PER and financial-NER research, and BlackRock–Snowflake data platform evidence. Anthropic’s broader finance ecosystem names FactSet, Moody’s, S&P Global, Morningstar, PitchBook, Daloopa, Databricks, Snowflake, Palantir, Box, and other partners, but those are not assigned to every fund.

The vendor sweep adds a second ecosystem trail through Rogo’s official updates archive: Rogo publicly lists partnerships or integrations with Daloopa, Dow Jones Newswires, MT Newswires, PitchBook, Microsoft, Snowflake, and SS&C Intralinks, and says it acquired the HF0-backed financial-agent company Offset. Rogo says its Daloopa integration brings source-linked fundamentals for more than 5,500 public companies into workflows used by 35,000 finance professionals. These are vendor-layer signals, not evidence that a particular tracked hedge fund uses a given connector or grants it investment authority.

The Hedgineer Episode 6 recording with Daloopa co-founder Thomas Li adds a more specific data-quality architecture. Li describes independent ML/AI models for different extraction tasks, human fundamental analysts verifying machine output and feeding corrections back into the system, a preference for preserving company-specific disclosure detail rather than forcing a universal taxonomy, and Excel/API delivery for investment teams (12:58–14:15, 09:20–11:58). This is founder/vendor evidence about the data and validation layer, not evidence of a named hedge fund’s deployment, predictive model, training corpus, permissions, or investment authority. See the timestamped capture note.

The Hedgineer Episode 4 recording with Sachin Kullkarni is now backed by the recovered public audio and local ASR. Apple’s description identifies him as Head of Data Science at an unnamed multibillion-dollar credit hedge fund with prior Point72 experience. The recording describes Jupyter, AWS/S3, Snowflake, Tableau/QuickSight, Airflow, and Materialize, and places LLMs in coding assistance, summarization, unstructured-to-structured mapping, and reference-data glue (00:31–00:48; 17:19–22:34). The employer remains unresolved by design; this is speaker-reported workflow evidence, not attribution of the stack to a named fund. See the capture note.

The missing fields remain model/version, data entitlement, residency, RAG versus fine-tuning, tool permissions, evaluation set, human approval, and pilot/shadow/assisted/live status. Public partnership language rarely fills those fields.

Role history and temporal corrections

The historical pass recovers dated transitions rather than flattening all titles into a current roster: Balyasny’s Applied AI group is dated to late 2022 in OpenAI’s retrospective; Point72/Cubist’s quant leadership changed in 2025; Man’s AI strategy gained a dated Anthropic partnership in February 2026; Jump’s research-technology leadership is documented in March 2026; and the Georgia Tech speaker program gives a March 2026 snapshot across several firms.

The history is still incomplete. Closed job pages, archived LinkedIn headers, exact start/end dates, internal reporting lines, and partnership renewal or termination dates remain unverified. Current, former, event-date, and third-party statuses therefore remain separate throughout this article.

August 14 quiet-firm artifact pass

The quiet-firm artifact ledger reclassifies several firms that had thin public AI records:

• Voleon now has firm-controlled evidence linking its leadership and research culture to machine learning, including Stanford, Harvard, Berkeley, and Princeton academic lineages. The public record still does not identify a complete AI roster, model inventory, or AI-attributed performance.
• XTX Markets publicly connects machine-learning forecasts to trading and exposes technical infrastructure through its research-cluster, XTY Labs, and TernFS pages. TernFS is a notable engineering artifact: the firm says it was designed for growing ML workloads and moved into production in 2023. The public record does not connect a named model or dataset to a specific strategy. The first-party TernFS technical post adds the engineering boundary: XTX describes a distributed filesystem for raw market data and GPU-job communication, and says storage activity is largely associated with machine-learning work. This is infrastructure evidence, not a model, data-rights, permission, or performance disclosure; see the capture note. A follow-up official careers/API check adds current XTY Labs foundation-model and agent-prototype hiring plus ML acceleration infrastructure language; that evidence remains hiring and infrastructure evidence, not deployment or return attribution.
• PDT Partners publicly lists Applied ML Scientist, Quantitative Researcher, and Research Engineer roles. Its process page says researched models are peer reviewed, empirically validated, and then deployed to automated trading systems. This is a process disclosure and hiring signal; it is not evidence that a named role is filled or that a GenAI system is live.
• Systematica publicly describes a Data R&D team and quantitative-research roles across its global offices. This adds a data-research operating signal, but the reviewed page does not disclose a GenAI system, model family, or named AI owner.
• Squarepoint publicly connects its investment team to machine learning, large-scale data analysis, quantitative models, and automated trading systems, while its careers pages recruit ML and PhD talent. The reviewed pages do not identify a GenAI platform or current AI leader.
• G-Research now has a direct firm-controlled AI/ML surface: its overview describes ML, AI, infrastructure, and technology innovation; its Technology Innovation Group exposes open-source engineering; and its news surface identifies a NeurIPS 2026 presence and an ELLIS academic partnership. These are public research and engineering signals, not proof of a finance foundation model or live investment agent.
• A current August 26, 2026 recheck of G-Research’s quantitative research and machine-learning page adds a named Research Lab, a Machine Learning College, and current role families spanning data science, ML research, NLP, quantitative research, and fundamental equity research. The page also describes a researcher profile built around mathematics, physics, ML, computer science, engineering, and postgraduate research. Its current news module links Cambridge and Warwick scholarship routes, while the news index lists the August Imperial College relationship and the July ELLIS partnership. The current Mar scholar profile describes multimodal graph learning, information theory, probabilistic modelling, and uncertainty in large language models. The Warwick announcement describes postgraduate scholarships and mentoring from G-Research researchers; the Imperial announcement describes six mathematics and machine-learning PhD scholarships. This strengthens the public talent and academic-network evidence; it does not establish employee status for scholars, exclusivity of their work, model ownership, or investment deployment.
• G-Research’s ICML 2026 paper-review page, dated August 27, 2026, adds a current firm-hosted research-media route. It identifies Johann as a Quant Research Manager and Fabian as a Senior Quantitative Researcher; Fabian describes predicting future returns of stocks and other assets and highlights agentic workflows as a research topic while leaving their usefulness for research ideation open. This is a dated firm commentary and personnel signal, not evidence of a deployed research agent, model inventory, permissions, or performance. The page does not publish surnames for either employee.
• G-Research’s current machine-learning roles guide adds a distinct recruiting and operating-surface record. It describes the ML team as working across noisy financial datasets, research, infrastructure, and production, and lists role families in data science, NLP, ML research, and quantitative research. The page also names employee profiles using first names only, including an ML engineer describing a project that continued in use and an open-source software manager. This is useful evidence about the firm’s talent funnel and public engineering narrative, but it does not establish the identity of those employees, a specific investment model, model ownership, or live trading authority.
• Renaissance Technologies is not blank in public sources, but its official about page and home page expose research and infrastructure scale rather than GenAI specifics: about 300 employees, 90 PhDs, a database receiving more than 40 TB per day, more than 50,000 cores, and a network above 200 Gbps. These are firm-reported operating figures, not a model-quality or AI-capability assessment. No public GenAI, language-model, named AI-lab, or AI-leader artifact was located in this pass.
• Renaissance Technologies adds a secondary media route: the Investology episode with Greg Zuckerman links a timestamped publisher page and YouTube recording. It discusses research culture, data, systematization, and whether AI is a new phase or variation of quantitative investing. This is historical/secondary context, not a current Renaissance employee statement or firm-controlled disclosure.
• A separate Goldman Sachs Exchanges interview with Renaissance CEO Peter Brown, published September 11, 2023 and recorded July 27, 2023, is a first-party publisher route. Goldman says the episode covers Brown’s career, the building of Renaissance, and the historical role of computer models and algorithms in the firm’s growth. The official YouTube recording was recovered and transcribed with timestamped local ASR; the official transcript locator remains a viewer/challenge route. In the recording, Brown discusses historical IBM work on speech recognition, machine translation, early generative language-model concepts, bilingual/legal text, and computer-accessible data. The full transcript is retained privately; ASR-derived details remain medium confidence pending audio spot-checking. This strengthens the historical personnel and research-lineage record but does not disclose Renaissance’s current model inventory, GenAI program, permissions, production authority, or performance.
• Marshall Wace has an official quant application guide that prohibits LLM chatbots and search engines during assessments. That is an assessment-control signal only; it cannot be extended into a claim about the firm’s internal research or production use of language models.

QRT and Caxton remain in the direct-artifact queue. QRT searches also collided with an unrelated Indian QRT Capital; that result is excluded. The next loop will search technical blogs, academic seminars, patents, repositories, closed-job mirrors, and dated leadership pages.

Regional coverage: United Kingdom, China/Hong Kong, Canada, and Australia

The Hebrew title-blind pass also recovered an official Israeli prospectus for Harel Alfi Benedek (5E) US Equity AI QUANT Strategy, a mutual hedged-fund wrapper whose stated offering window ran from December 21, 2023 through December 13, 2024. Harel’s current mutual-hedged-fund page still names the strategy and describes the AI-based stock-selection relationship, while a September 29, 2024 annual-report notice identifies fund number 1201508, Alfi Benedek as external adviser, and Erez Shapira as principal investment manager as of June 30, 2024. The filing says Alfi Benedek developed an AI-based stock-selection algorithm with Bintek Advanced Models, which it describes as researching models using statistical, machine-learning, and AI methods. It describes quantitative models learning stock trends and behavioural patterns from trading data, volume, and selected macroeconomic data, with long, market-neutral, short, and currency-hedging modes. The filing says the manager may act on model recommendations at discretion. This is direct prospectus evidence of an AI-labelled fund design and a human control boundary, but not proof of operation after the 2024 reporting period, model architecture, training data, validation, execution logs, or independently verified performance. See the Hebrew prospectus note.

The China Fund News short forum video with Lingjun CIO Ma Zhiyu, published 2026-05-28, is now backed by a recovered public MP4 and local automatic Chinese ASR. At 00:45–01:11, Ma frames AI as part of continued iteration of investment technology and capability; the earlier 00:00–00:45 window covers macro conditions, Chinese-equity allocation demand, and systematic-investing consistency. This is a brief speaker-context artifact. It does not establish a new model, dataset, vendor, permission boundary, production endpoint, or AI-attributed result. See the capture note.

The article already covers several UK-based managers, including Man/AHL, Marshall Wace, Brevan Howard, Winton, and Systematica. A regional expansion adds smaller or differently disclosed public control cases. These records are not a geographic scorecard: they show what each source makes observable, and what remains undisclosed.

A fresh title-blind search adds two manager-side regional routes that were not previously in the ledger. Deep Edge Fund describes an Israeli medium-frequency ML fund trading CME futures across equity indices, commodities, currencies, and Treasuries, names Amit Antebi and David Khotoveli, and says dozens of automated strategies run concurrently. An Israeli company-record search result lists Deep Edge Fund Management Ltd. as an active private company, but does not establish a securities licence or independently verified fund offering. AI Financial Capital is described by MANAFA as a subsidiary and multi-asset systematic manager using algorithms and AI-driven strategies across foreign exchange, equities, indices, and commodities; its corporate profile adds technical-analysis, behavioural-finance, price-pattern, signal, and execution language. These pages establish public positioning and stated scope, not model identity, data rights, regulator authorization, live permissions, AUM, or AI-attributed performance. The capture note preserves the promotional-language and entity-boundary caveats.

The thin-language Hebrew pass adds I Know First, a fintech whose current pages describe self-learning forecasts across roughly 13,500 assets and name co-founder/CEO Yaron Golgher, R&D lead/CTO Dr. Lipa Roitman, and Tel Aviv University mathematician Prof. Yakov Yakubov as a consultant. A June 29, 2025 Calcalist podcast describes Gabi Diamant, Dor Aligula, Or Aligula, and Mor Hazan forming a hedge fund that used AI to analyze markets and then turning an internal fund tool into a global financial-analytics platform. A 2016 Globes report gives historical detail on six forecast horizons, quantitative-only inputs, and a proposed fund with an unnamed financial partner; its proposed launch and $150–200 million target were not verified. The current material describes predictive decision support rather than a disclosed generative-AI system, and does not establish a current fund, named institutional customers, model inventory, data rights, permissions, or independently reproducible performance. See the Hebrew capture note.

The Hebrew title-blind pass also recovered an official Israeli prospectus for Harel Alfi Benedek (5E) US Equity AI QUANT Strategy, a mutual hedged-fund wrapper whose stated offering window ran from December 21, 2023 through December 13, 2024. The filing says Alfi Benedek developed an AI-based stock-selection algorithm with Bintek Advanced Models, which it describes as researching models using statistical, machine-learning, and AI methods. It describes quantitative models learning stock trends and behavioural patterns from trading data, volume, and selected macroeconomic data, with long, market-neutral, short, and currency-hedging modes. The filing names Kobi Alfi, Yona Benedek, and fund manager Erez Shapira; it says the manager may act on model recommendations at discretion. This is direct prospectus evidence of an AI-labelled fund design and a human control boundary, but not proof of current operation, model architecture, training data, validation, execution logs, or independently verified performance. See the Hebrew prospectus note.

The Russian-language pass adds Alfa Capital Quant, an exchange-traded mutual fund page that says machine-learning technologies and AI elements, with professional-manager participation, generate signals for Russian equities over day-to-week horizons. The page names Nikita Elenberger, displays dated sector and issuer holdings, and includes Alfa Capital’s management-company licence references. It is a product-level and regulatory-manager surface rather than a hedge-fund disclosure; the model, features, training data, validation design, and AI-attributed performance remain undisclosed. An Indonesian route adds Bareksa’s announcement with STAR Asset Management and QUANTIT, plus a BINUS Finance event summary naming Ariyanto Dipo Sucahyo. The newly recovered Bareksa product article identifies STAR Fixed Income Neo AI Dollar, launched January 7, 2026, as predominantly USD fixed income with up to 15% in global equities; it says Quantit AI supports analysis and stock selection while STAR retains investment authority. A QUANTIT company post names the equity-sleeve system as the FINTER AI agent. An Infobanknews launch report names STAR director Erwin Faizal, QUANTIT CIO Seong Min Park, QUANTIT CEO Han DuckHee, and Bareksa co-founder/CEO Karaniya Dharmasaputra, and attributes a 10-year 5.7–8.5% annual backtest claim to the launch coverage. A Korean DBR profile adds QUANTIT’s claims about Vietnam robo-adviser work, an NYU educational-tool or hackathon route, and discussions with Singapore institutions and hedge funds; those statements do not establish converted deployments. The BINUS summary describes AI-based factor modeling, a three-tier alpha/portfolio/fund architecture, knowledge-graph risk controls, social-sentiment monitoring, and network-based contagion analysis. The FINTER model, training data, reproducible evaluation, execution logs, and independent performance remain undisclosed, and the referenced Bareksa product video remains unrecovered. See the capture note.

The China-language pass adds GokuTech / 念空科技, whose official pages name Wang Xiao as founder partner and CIO and describe a research/IT-heavy organization, a factor-to-model-to-backtest workflow, and separate roles for deep-learning research, factor-pool alpha prediction, execution-cost monitoring, strategy deployment, options, and HFT systems. The company’s strategy page describes factor generation, submodel combination, AI-model training, and historical backtesting. A July 8, 2025 Shanghai Securities News interview reports Wang’s account of AI use since 2017, LLM and multimodal-data exploration, and a separately described AllMind technology entity. A May 23, 2025 36Kr interview and same-day Sina Technology interview further attribute to Wang a three-person AI team formed in 2017, a claim that 90% of live models had shifted to neural-network/Transformer methods by 2019, and a vertical model fine-tuned from Qwen3. The related SASR paper lists Shanghai Goku Technologies Limited, AllMind, and Shanghai Jiao Tong University in its affiliation block and proposes adaptive SFT/GRPO post-training for reasoning tasks. This is a useful bridge between a public quant firm’s hiring and research language and an affiliated LLM research artifact, but the paper is not a financial experiment and does not establish that SASR or AllMind research is used in GokuTech portfolios. Company infrastructure, staffing, AUM, compute, live-model share, and Qwen3 fine-tuning claims remain self-reported or interview-reported; no model inventory, data-rights record, live permission map, independent performance audit, or AI-attributed return was retrieved. See the capture note.

In the UK, GSA Capital describes systematic strategies across asset classes, proprietary systems, and research supported by high-performance infrastructure. Its public careers page describes systematic-strategy research and infrastructure hiring, but the reviewed pages do not disclose an LLM, agent runtime, or GenAI investment workflow. Quantellence publicly describes an AI-native quantitative-investment operation guided by autonomous AI-agent teams. Because the page does not establish a regulated fund vehicle, current assets, filled roles, or a production permission map, it is retained as a UK public positioning and formation signal rather than treated as equivalent to an established manager’s operating disclosure.

A University of Cambridge Careers Service Q&A with Jan Stańczuk adds a named GSA personnel and workflow lane. Published in November 2025, it identifies Stańczuk as a GSA quant trader and researcher with a PhD in applied mathematics and machine learning focused on GANs and diffusion models. He describes exposure to idea generation, data analysis, hypothesis testing, model building, portfolio construction, tooling, backtesting, paper trading, and real-time trading, and recounts moving code into production within weeks and a strategy into deployment after about a year. These are employee recollections in a university careers interview, not an independent deployment audit. His ICML 2024 paper is academic evidence about diffusion-model geometry; it does not establish GSA use of that method, an LLM or agent program, investment authority, or performance. The full personnel and evidence-boundary note is here.

An additional GSA personnel and academic-lineage route comes from Sebastian East’s self-authored professional page, which says he is a quantitative researcher at GSA Capital and describes work at the intersection of optimization, control theory, and machine learning. The page lists an Oxford DPhil in optimization and control, prior lecturing in Bristol, and an AI research-scientist internship at Nnaisense. His imitation-learning paper studies stable imitation learning for nonlinear control; a University of Oxford researcher-to-finance booklet independently presents East as a GSA quantitative researcher and repeats the Oxford/Bristol path. The UCL London Hopper Colloquium page separately identifies Neeraja Bhamidipati as a GSA quantitative researcher and gives a systematic-trading talk abstract centered on trading costs, market impact, and execution. These sources expand the visible personnel and research lineage; the academic paper is not evidence of GSA use, and none of the sources establishes a GSA AI/GenAI model, agent, dataset, production deployment, permission boundary, or return. See the capture note.

In mainland China and Hong Kong, High-Flyer / 幻方 publicly says it has used machine learning since 2008, applies neural networks to large data, develops natural-language understanding, and explores fully automated quantitative trading. Its homepage separately presents AI compute, an AI-used hedge fund, and AI basic research, while naming Fire-Fly II, hfai.nn, hfreduce, and 3FS. The fund page identifies mainland and Hong Kong legal entities, which matters because the AI-company and fund surfaces should not be merged without entity-level evidence. Aspect adds a UK-to-mainland-China legal-entity signal rather than a new AI signal: its December 17, 2024 press release says Aspect Capital (China) Limited completed AMAC private-fund-manager registration on December 13, 2024, after QFI status in 2022 and WFOE registration in the Shanghai Pilot Free Trade Zone in March 2024; that is regional market-access evidence, not a GenAI or model-deployment disclosure. Lingjun / 灵均 describes an implemented research agent, ML/deep-learning signal development, AI coding tools for risk-model development, and large-model processing of unstructured and multimodal data; a separate Hong Kong/offshore page describes Lingjun Capital Hong Kong, offshore products, and Beijing/Inner Mongolia supercomputing centers. Mingshi / 鸣石 describes AI-selected factors, nonlinear ML weighting, generated signals, and a research pipeline split across factor, AI, optimization, risk-control, and trading stages. WizardQuant / 宽德 adds a separate hiring, infrastructure, and lab-positioning signal: its careers pages link the WILL AI-lab hiring surface, say strategy research uses statistical and machine-learning tools, and describe AI for quant scenarios, ML platforms, AI cluster customization, training/inference, model lifecycle/repository work, safe trading-scenario loading, deep-learning deployment, simulation/backtesting, portfolio optimization, and low-latency systems. A June 22, 2025 official article describes AI Agent as a perceived next stage, AI-driven Smart Beta framing, and WILL as an industrial AI R&D lab created by WizardQuant; it does not establish a regulated-product boundary for WILL, model inventory, live deployment, or AI-attributed returns. DeepWin / 蝶威量化 adds another mainland-China control case: its official page identifies Shanghai DeepWin Private Fund Management Co., lists registration P1070101, and describes AI quantitative-investment, reinforcement-learning, and Agent modules, including strategy agents using reward signals, execution agents updating from fill feedback, and risk constraints entering the reward function. Chinese Securities Journal’s March 20 event report and March 30 industry survey describe a firm-hosted DeepWin Agent workflow with five roles across analysis, research, code, evaluation, and fund-manager review; the described workflow spans broker-report factor extraction, factor optimization, code/backtest execution, performance judging, and factor intake. This is a firm/media public signal, not independent proof of model inventory, live permissions, product or fund-vehicle boundaries, deployment audit, or AI-attributed performance. Longqi Scientific Investment / 龙旗 adds a separate mainland/Hong Kong regional signal: its current official page identifies Hangzhou Longqi Scientific Investment, Longqi Scientific (Hong Kong) Limited, AMAC private-fund-manager licensing language, SFC Type 9 asset-management language, index-enhancement, market-neutral, and market-timing fund families, and AI/ML techniques described as trading since 2018. The same page reports two different SFC CE numbers, BPL583 in the legal section and BNP998 in the footer, so the CE number is retained as a discrepancy requiring direct SFC-register verification rather than resolved by inference. JoinQuant / 聚宽 describes a registered private-fund manager combining market data, quantitative research, and AI-driven investment research, including a model layer and factor/feature work. 中量财富 publicly lists machine-learning strategies, sentiment analysis, order-book models, and high-frequency research. Separately, IQuestLab, its IQuest-Coder model page, and the IQuest-Coder repository publish code-LLM/model artifacts; one public researcher page links IQuest Research, UbiQuant to coding-LLM and agent work. Because no official Ubiquant fund disclosure or regulator record was retrieved for this lane, it is retained as technology-affiliation evidence, not a quant-fund deployment claim. These are firm-controlled descriptions, hiring surfaces, or public model/code artifacts, not independent evidence of model quality, current deployment, portfolio permissions, or AI-attributed returns; Chinese-language claims also require preserving the original wording and translation boundary. In mainland China and Hong Kong, High-Flyer / 幻方 publicly says it has used machine learning since 2008, applies neural networks to large data, develops natural-language understanding, and explores fully automated quantitative trading. Its homepage separately presents AI compute, an AI-used hedge fund, and AI basic research, while naming Fire-Fly II, hfai.nn, hfreduce, and 3FS. The fund page identifies mainland and Hong Kong legal entities, which matters because the AI-company and fund surfaces should not be merged without entity-level evidence. Aspect adds a UK-to-mainland-China legal-entity signal rather than a new AI signal: its December 17, 2024 press release says Aspect Capital (China) Limited completed AMAC private-fund-manager registration on December 13, 2024, after QFI status in 2022 and WFOE registration in the Shanghai Pilot Free Trade Zone in March 2024; that is regional market-access evidence, not a GenAI or model-deployment disclosure. Lingjun / 灵均 describes an implemented research agent, ML/deep-learning signal development, AI coding tools for risk-model development, and large-model processing of unstructured and multimodal data; a separate Hong Kong/offshore page describes Lingjun Capital Hong Kong, offshore products, and Beijing/Inner Mongolia supercomputing centers. Mingshi / 鸣石 describes AI-selected factors, nonlinear ML weighting, generated signals, and a research pipeline split across factor, AI, optimization, risk-control, and trading stages. WizardQuant / 宽德 adds a separate hiring, infrastructure, and lab-positioning signal: its careers pages link the WILL AI-lab hiring surface, say strategy research uses statistical and machine-learning tools, and describe AI for quant scenarios, ML platforms, AI cluster customization, training/inference, model lifecycle/repository work, safe trading-scenario loading, deep-learning deployment, simulation/backtesting, portfolio optimization, and low-latency systems. A June 22, 2025 official article describes AI Agent as a perceived next stage, AI-driven Smart Beta framing, and WILL as an industrial AI R&D lab created by WizardQuant; it does not establish a regulated-product boundary for WILL, model inventory, live deployment, or AI-attributed returns. DeepWin / 蝶威量化 adds another mainland-China control case: its official page identifies Shanghai DeepWin Private Fund Management Co., lists registration P1070101, and describes AI quantitative-investment, reinforcement-learning, and Agent modules, including strategy agents using reward signals, execution agents updating from fill feedback, and risk constraints entering the reward function. Chinese Securities Journal’s March 20 event report and March 30 industry survey describe a firm-hosted DeepWin Agent workflow with five roles across analysis, research, code, evaluation, and fund-manager review; the described workflow spans broker-report factor extraction, factor optimization, code/backtest execution, performance judging, and factor intake. This is a firm/media public signal, not independent proof of model inventory, live permissions, product or fund-vehicle boundaries, deployment audit, or AI-attributed performance. Longqi Scientific Investment / 龙旗 adds a separate mainland/Hong Kong regional signal: its current official page identifies Hangzhou Longqi Scientific Investment, Longqi Scientific (Hong Kong) Limited, AMAC private-fund-manager licensing language, SFC Type 9 asset-management language, index-enhancement, market-neutral, and market-timing fund families, and AI/ML techniques described as trading since 2018. The same page reports two different SFC CE numbers, BPL583 in the legal section and BNP998 in the footer, so the CE number is retained as a discrepancy requiring direct SFC-register verification rather than resolved by inference. JoinQuant / 聚宽 describes a registered private-fund manager combining market data, quantitative research, and AI-driven investment research, including a model layer and factor/feature work. 中量财富 publicly lists machine-learning strategies, sentiment analysis, order-book models, and high-frequency research. Separately, IQuestLab, its IQuest-Coder model page, and the IQuest-Coder repository publish code-LLM/model artifacts; one public researcher page links IQuest Research, UbiQuant to coding-LLM and agent work. Because no official Ubiquant fund disclosure or regulator record was retrieved for this lane, it is retained as technology-affiliation evidence, not a quant-fund deployment claim. These are firm-controlled descriptions, hiring surfaces, or public model/code artifacts, not independent evidence of model quality, current deployment, portfolio permissions, or AI-attributed returns; Chinese-language claims also require preserving the original wording and translation boundary.

Going International Asset Management / 高盈量化 adds a Hong Kong manager-side surface. The site identifies 高盈国际资产管理有限公司 / Going International Asset Management Limited, describes AI-driven quantitative trading across Hong Kong, the United States, Japan, Singapore, and Europe, and displays SFC central number BGA871 with Type 9 language. The SFC register is the authoritative lookup surface, but this pass did not capture a direct current firm-specific extract; the identifier is therefore retained as a public-site claim plus cross-check, not as a fully reproduced current registration result. Dated Chinese Fund News coverage and a Hong Kong finance-media interview page attribute statements to chairman Wu Chao and describe AI-agent task decomposition across financial, price/volume, news, and alternative data, large-model and reinforcement-learning integration, microsecond-scale execution language, and claimed research cooperation involving Tsinghua University and Hong Kong Polytechnic University. The official 高盈科技 January 2026 Tsinghua Hong Kong Forum page adds a dated event and personnel route: it names Wu Chao as a keynote speaker, records his description of end-to-end AI with reinforcement learning and specialized-chip integration, and states that human validation remains part of the workflow. The company’s history page lists a 2026 strategic-cooperation agreement with Zhongke WenGe, while its high-frequency solution page describes deep-learning price prediction, order-book microstructure, intraday timing, and reduced reliance on traders. These are company-published positioning and event statements; the separate Hong Kong Economic Journal report repeats additional partnership and data-scale claims that remain secondary/promotional until independently documented. These are dated public statements and firm descriptions; they do not publish a model inventory, training data, agent permissions, or independent performance audit. GDA / GROW is a separate public-brand lead that describes agentic research, testing, and execution with human approvals. Its legal identity, regulator record, named team, independent methodology, and live authority remain unresolved; its numerical backtest and return statements are retained as unverified first-party marketing claims and are not used in comparison tables. A separate Chinese-language AI Odyssey episode, published April 8, 2026, features a guest presented as a small hedge-fund founder under the alias Wang Huageng / 大厨. The publisher transcript discusses price/volume regime analysis, cross-questioning GPT and Claude to challenge model flattery, and the possibility of AI-agent interfaces changing financial-data platforms. Because the guest’s legal identity, fund name, employer, model, and track record are unresolved, this is a regional discovery and practitioner-vocabulary route only; it is not assigned to Gao Ying, High-Flyer, BedRock, or another named manager. See the capture note. Finally, CITIC Securities’ “信谛听” 1.5 coverage is an adjacent securities-platform case, not a hedge fund: public reports describe natural-language strategy generation and backtesting, adaptive code-error correction, source-oriented retrieval, and DeepPaw multi-agent research integrating reports, news, market data, and filings. This supports a regional financial-data and agent architecture record, not a claim about private-fund deployment, point-in-time backtests, live authority, or returns. The full evidence note is here.

Metabit Trading / Qianxiang / 乾象 adds another mainland-China public signal. Its firm-controlled Chinese page reports private-fund-manager registration P1071383 for Beijing Qianxiang Private Fund Management Co., Ltd., Beijing and Shanghai offices, RMB 10 billion management scale, full-process automated alpha mining, an end-to-end reinforcement-learning framework, a POMDP framing for incomplete market information, hundreds of thousands of CPU cores, a 20 EFLOPS-class GPU cluster, a self-developed DL/ML training and inference management platform, a self-developed data engine for 10PB-level global market data, and full-process monitoring/risk-control language. A separate JuiceFS Chinese customer case describes Metabit as an AI-centered quantitative-investment company and presents a cloud research-infrastructure workflow spanning feature extraction from market and unstructured data, AI model training, signals, portfolio optimization, simulated orders, backtesting, elastic compute, Kubernetes, POSIX storage, and IP isolation. This is firm-controlled operating language plus vendor-published infrastructure evidence, not a direct AMAC extract, fund document, model inventory, training-data license, production permission map, autonomous capital-allocation disclosure, independent performance audit, or AI-attributed-return record; the vendor case’s older scale statement is retained as dated vendor context rather than overriding the current firm page.

In Canada, QTS Capital Management describes supervised-ML models and a data-intelligence approach within a multi-strategy trading firm; its feature-selection article discusses transformers, variational autoencoders, pretraining, fine-tuning, and sample-specific feature selection for financial ML. QTS is Canada-linked rather than purely Canadian: the 2026 SEC Form D/A for QTS Partners, L.P. identifies a Delaware issuer with a British Columbia principal place of business and QTS Capital Management LLC as general partner, while the QTS Capital Management LLC LEI record lists Delaware legal jurisdiction and a British Columbia headquarters address. Evovest describes a systematic equity process combining fundamental analysis and machine learning; its public ETF disclosure states that portfolio selection relies on proprietary ML models. The Maritime Fund describes a machine-learning group developing, testing, and deploying quantitative strategies and AI tools for risk metrics, screening, and idea generation. Dynamic Funds adds a named Canadian role signal: its official profile says Ahmad Golaraei joined in 2025 as an AI & Quant Investment Analyst applying emerging technologies to support investment research, with prior ML, NLP, and generative-AI work at CI Global Asset Management. Connor, Clark & Lunn Investment Management adds a larger Canada-based quant-equity operating signal: its Quantitative Equity Data Science role places a data scientist in a Vancouver Quantitative Equity Team embedded in a quantitative equity fund managing more than $111 billion, with machine learning, AI, NER, knowledge graphs, integrated data-model evaluation, data-quality testing, alpha-research support, and production deployments. The same page separates CC&L Investment Management from the broader CC&L Financial Group, which it says manages more than $222 billion across affiliates. Mackenzie Investments adds another Canada-based quant-equity signal: its 2026 market outlook says the Global Quantitative Equity Team uses ML, NLP, LLMs, and cloud computing with a human overlay, applies LLMs to earnings-call sentiment, and uses NLP on alternative data. A separate 2025 practitioner PDF says the team uses non-linear ML for revenue and earnings estimates and native-language NLP to read financial statements. These sources support Canadian ML, automation, role, and operating-surface signals, but the reviewed public record does not establish a Canadian finance-specific agent authority, production permission map, live trading authority, model inventory, independent performance audit, or AI-attributed contribution.

In Australia, Avangard Investments describes an AI-driven systematic Australian-equity fund. Its current performance page names A.L.F.R.E.D. as the proprietary AI system behind the strategy and reports 5.5 billion data points, 28,000 daily price inputs, up to 100 years of cleaned historical financial data, and an ASX-listed equities and ETF universe of 2,500+ securities. Current FAQ, governance, and people pages identify Avangard Investments Pty Ltd as a corporate authorised representative of FB Corp Limited under AFSL 557810, describe A.L.F.R.E.D. as an in-house machine-learning and portfolio-optimization system, identify Alfred “Alf” Eggo as CEO, portfolio manager, and architect of A.L.F.R.E.D., and distinguish Avangard from similarly named entities. The performance page separates historical Avangard Alpha Fund performance from the new fund structure effective July 1, 2026. This is current entity, fund-structure, named-system, and firm-reported data-scale evidence, not proof of model design, live permissions, an independent ASIC extract, live GenAI agents, independent performance, or AI-attributed return. Antipodes has a June 2026 AI & Alternative Data role spanning AI, quantitative research, and investing. The role is unusually explicit for an Australian manager: it names PM/analyst agent, skill, and MCP catalogue ownership; systematic research and portfolio integration; alpha-signal validation and implementation; Claude Code, Codex, and Copilot-style workflows; AgentOps, evals, observability, audit trails; and Anthropic/OpenAI SDKs. This is a high-confidence official hiring/workflow signal, not evidence that the role was filled, the catalogue exists in production, agents have portfolio permissions, or returns are attributable to AI. RQI Investors adds an Australian active-quant equities signal: its June 16, 2025 release says the RQI Australian Diversified Alpha Long Short Fund launched with A$500 million from UniSuper, uses ASX 300 long-short exposure, and identifies AI and machine learning as part of the systematic process for finding alpha opportunities. RQI’s March 2024 AI paper page separately describes institutional uses including automated reporting, topic modelling, NLP on news and earnings-call transcripts, adaptive trading algorithms, and portfolio optimisation. Its First Sentier Curious archive provides a second first-party route to the June 2026 “Quant edge” episode with RQI’s Joanna Nash and David Walsh; it corroborates the episode identity, roles, and recording date but is not counted as a second episode. These are official manager communications; they do not provide model identity, production-stage maps, independent signal attribution, or proof that every described AI use is active in the long-short fund. Boronia Capital describes a Sydney-based, research- and technology-focused systematic manager, but the reviewed page does not add a GenAI or LLM disclosure. Macquarie QIS is an Australia-headquartered bank/QIS comparator rather than a hedge-fund manager peer: its current QIS page says its technology platform incorporates machine learning and intraday trading, while an older official Protean case describes reinforcement-learning signals for an FX-volatility QIS index. These are Australian manager, fund-structure, hiring, and adjacent bank/QIS signals; they do not establish filled roles, model weights, live permissions, current GenAI agents, independent performance, or return attribution.

The NAB Morning Call route adds a separate Australian media capture for Minotaur Capital. In the February 6, 2026 interview, Armina Rosenberg describes the internal system as ingesting roughly 30,000–35,000 articles per week from 173 sources, including Indonesian, Japanese, and Spanish media (04:18–04:43). She describes AI-generated company snapshots, deeper initiations, thesis-validation reports, and continuously running stock-specific agents (05:37–06:24; 06:31–06:54). She says a human remains at every decision gate and that portfolio construction and sizing remain with the two founders (08:56–09:22; 10:57–11:43). In discussing operating economics, she says the firm has no investment analysts, uses roughly 20 large language models, and makes thousands of API calls per day (26:24–27:10). She also describes moving from one-off memos toward a continuously monitored pipeline that frames debates and stress-tests arguments (29:00–30:19). This is a publisher-hosted founder account, not an independent audit; Minotaur’s first-party material reports 174 rather than 173 sources, and neither route discloses the model roster, versions, data rights, error rates, permissions, or AI-attributed performance. See the timestamped capture note.

A separate May 25, 2026 Experts in the Loop episode adds a publisher transcript and a Minotaur first-party cross-link for the same firm’s public AI workflow. The episode identifies Armina Rosenberg as co-founder and portfolio manager and Thomas Rice as co-founder; Rosenberg describes Rice as a fund manager and software developer who built Taurient (06:38–08:27). The later account describes Taurient processing about 30,000 articles weekly across 174 sources for global-equities idea generation and structural-change triage (15:03–18:36), with a human at every decision gate (18:36–21:28). It also describes source-document checks, treating models as junior analysts, routing work across more than 20 models to manage cost, and using coding agents and Claude Skills as part of the workflow (23:41–28:38; 43:18–46:46). The Minotaur page independently confirms the “Taurient” spelling and 174-source description. This is a distinct episode from NAB, not a second confirmation of live permissions or returns: the publisher transcript and firm page do not disclose the model roster, versions, prompts, data rights, evaluation denominators, API spend, agent permissions, production endpoints, or independently audited performance. See the timestamped capture note.

The title-blind Invest with AI episode with Doug Clinton adds a separate named Intelligent Alpha route. The publisher identifies Clinton as Intelligent Alpha’s founder and describes the firm as using frontier models for investment analysis and portfolio-management work; the transcript records his origin experiment after ChatGPT’s release, a stated hybrid quant/fundamental process, model-mediated knowledge graphs and EBITDA ontologies, model routing, and an all-AI ten-stock portfolio experiment (02:39–06:38; 16:55–20:26; 30:23–32:36; 45:19–48:28). The episode also explicitly grades the system as a “B+ analyst” rather than treating it as a complete substitute for investment judgment, and discusses leadership ownership and uncertainty (19:21–20:26; 43:13–49:09). These are a named founder’s account and a publisher transcript, not independent evidence of unrestricted capital authority: the route does not disclose model versions or weights, training and retrieval data, data rights, evaluation fixtures, routing logs, portfolio permissions, live order authority, or independently audited performance. The related source note keeps Intelligent Alpha separate from Clinton’s named Deepwater context and from unrelated vendor claims.

Three additional adjacent systematic-manager controls were verified outside the required regional expansion. Pictet Asset Management’s Quest AI-Driven franchise adds a Switzerland/Singapore/Canada source-family control: official 2026 and 2025 pages describe proprietary AI stock selection across thousands of companies, “crystal box” forecast attribution, weekly rebalancing with human oversight, quarterly retraining, rigorous data governance, a one-month forecast horizon, and boosted-tree use of IBES data. The February 2026 Quest AI article, April 2026 governance article, and October 2025 practitioner page establish a public AI-strategy and governance disclosure, not model weights, a full feature list, live portfolio permissions, independent performance, or AI-attributed return evidence. RAM Active Investments publishes firm-controlled research on LLM-based newsflow representations and fine-tuning LLMs for stock-return prediction, while its company page describes an in-house deep-learning infrastructure with more than 500 alpha inputs for single-stock return prediction. The July 2024 paper also appears in the EMNLP 2024 Industry Track, which verifies the public research artifact but not its use in live portfolios. T. Rowe Price Integrated Equity published a July 2026 systematic-research note describing LLM-assisted business-quality scoring, analyst/LLM prompt refinement, internal-database integration, hallucination-reduction controls, weekly batch scoring, software-resilience analysis, reported correlation/classification observations, and explicit look-ahead-bias and overfitting caveats. These adjacent controls add public research-workflow and methodology evidence; they do not disclose model vendors or weights, portfolio permissions, production adoption, audited signal contribution, investment recommendation authority, or AI-attributed returns.

A title-blind media and hiring expansion adds several distinct evidence classes. Pictet’s David Wright gives a timestamped practitioner account of one-month relative-return forecasting, point-in-time features, decision trees, feature interactions, retraining, and human-defined feature universes. RQI’s David Walsh discusses nonlinear signals, idiosyncratic alpha, machine learning, transaction costs, and portfolio optimization in a full publisher transcript. Quantedge CEO Suhaimi Zainul-Abidin describes generative AI as a research-productivity, data-cleaning, and execution aid while saying it is kept out of production models; the episode’s scale and return figures remain publisher/guest claims. A second Strand interview with Suhaimi provides a separate dated account: the guest describes a final human check before trading instructions go to execution (07:42–08:21), prioritizes live portfolio management before research automation (09:03–10:17), and says current AI use is around research summaries, data cleaning, and simple backtesting while production models remain deterministic in his account (15:57–17:06). The local transcript is automatic-caption evidence without diarization; neither interview establishes Quantedge’s model inventory, provider contracts, data rights, or AI-attributed performance. These remain practitioner statements, not independent audits.

The July 28, 2026 Sophron episode with Marc Nunes adds a Singapore-based crowdsourced-signal route that was absent from the earlier title-blind queue. The publisher identifies Nunes as AlphaNova’s co-founder and CEO. In the timestamped local transcript, he describes code-level leakage checks, a proprietary overfit screen, submission limits, and a live observation period (04:13–06:13); a roughly 50%-correlation ceiling and compressed existing-signal context (24:35–29:18); manual Claude-assisted review with an intended contest-to-trading pipeline (28:12–30:22); and a neural-net prediction layer separate from portfolio engineering (30:22–35:16). He also describes launching research-agent swarms for idea generation and using agents for coding and code review while retaining human judgment (41:18–47:10). The capture note records the ASR limitations and source boundaries. These are named speaker and company-methodology claims; they do not establish current model weights, training data, agent permissions, a live signal roster, or independently audited performance.

AlphaNova’s homepage generally describes selected signals entering a live trading system. Its July 28 competition-close discussion says individual May-competition signals were not viable as standalone strategies after realistic transaction costs and describes future one-month live simulations, while a separate profit-sharing discussion states that no signals had yet been onboarded to the live environment at that stage. The article preserves these as dated, scope-specific first-party statements rather than inferring a current deployment status.

Moreton Capital Partners exposes an AI Lab and public language around agentic AI, LLMs, systematic commodities, prediction markets, academic partnerships, interpretable construction, and human-plus-machine workflows. Arrowpoint’s quantitative/machine-learning role names deep learning for financial time series and cross-asset signals, regime analysis, AWS research/trading systems, and research-to-production code. Its founder, Jonathan Xiong, separately describes AI agents approaching analyst-like research capability, but identifies access to proprietary, broker, expert-network, and transcript data as an unresolved bottleneck in the 2025 Macro Hive interview (timestamped capture note). DRW’s Montréal AI/ML Research Intern posting lists possible forecasting or LLM projects alongside deep learning, reinforcement learning, graph neural networks, and telemetry. These are first-party positioning, hiring, and speaker-account signals; none establishes a filled role, a complete model inventory, capital permissions, or return attribution.

The current Moreton AI Lab page adds specific lab design language: high-performance computing and ML tools for prototyping, an intended development/test/deployment remit, university and AI-platform partnerships, and stated principles of commercial relevance, research excellence, and data sovereignty. Moreton’s leadership page names Les Finemore as Co-Founder/CIO and displays Carlos Martinez and Bruno Fragoso in ML-platform and model-training roles, Emiliano Lopez in data-access infrastructure, Joel Miguel Maya in quantitative trading and risk infrastructure, and Jesus Guzman in weather-data and AI-data systems. A closed prediction-market quantitative-analyst listing described fair-value estimation, calibration, alternative-data pipelines, backtesting, and live-performance evaluation across event categories. The Alpha Signal Program describes an external model-intake path in which submitted signals may be tested through execution and risk infrastructure before an allocation decision. A Spanish-language El Cronista México report separately attributes to Moreton plans for Mexico City and Abu Dhabi, an AI lab with UNAM and Tecnológico de Monterrey links, possible FEMSA and Coppel relationships, and commodity models using satellite, climate, logistics, and language data. That report is secondary and time-bounded; the claims are retained as verification leads, not established partnerships or capital events. These are current first-party, historical recruiting/program, and secondary regional descriptions; they do not establish named university partners, filled-role dates, model providers or weights, training data, agent permissions, capital allocated, or audited performance. See the expanded capture note.

The Alternative Data Podcast route for Vlad Johnson is a corrected Isla Capital route, not an Eisler Capital record: the episode introduction identifies Johnson with Isla Capital’s Systematic Futures work. In local ASR from the public Acast enclosure, Johnson describes ML for quantitative-finance alpha research, unstructured alternative data including news, earnings calls, Commitments of Traders reports, and weather, and LLM-era topic-conditioned sentiment (15:50–17:37). He describes GenAI as useful for research automation—decoding events, mapping them to instruments, and testing candidate transformations—and distinguishes that automation from direct alpha claims, retaining human ideation and judgment (22:06–24:01). These are personal, automatic-ASR-mediated practitioner statements; they do not establish Isla’s current model inventory, data licenses, vendors, permissions, production endpoint, or performance. See the audio-recovery note. A related Nico Smuts episode identifies him as a former Citadel senior data scientist and discusses a public-data/Telegram research route, crypto-market sentiment timing, messy-data pipelines, and AI-era hiring. In the recovered public audio, Smuts describes an artisanal scrape of government data with an earnings read roughly three weeks before a print (08:19–09:33), and reports that positive Telegram sentiment lagged crypto prices while negative sentiment was coincident (22:38–25:28). He also says AI may make complex-pipeline and messy-data work more accessible, while over-reliance can narrow junior development (18:53–21:41; 36:08–41:33). These are personal, automatic-ASR-mediated historical statements; they do not establish Citadel’s current model inventory, data licenses, vendors, permissions, production endpoint, or performance. See the audio-recovery note. Both pages carry personal-view disclaimers. Finally, the Simons Institute Industry Day program lists Bridgewater’s Rohan Alur as Principal Research Scientist/AIA Forecaster and Voleon’s Sahand Negahban as Research Scientist. The program verifies conference metadata and affiliation, not talk content or firm-system deployment.

The recovered Evercore Flow of Funds episode with Acadian adds a direct, timestamped systematic-credit route. Kelly Young and Scott Richardson describe Acadian’s process as quantitative and technology-enabled, combining financial intuition with data science (01:03–02:47). Richardson’s concrete AI/ML example is credit-risk forecasting: ingesting hundreds of variables to detect nonlinear interactions relevant to default or rating migration, with the stated goal of improving out-of-sample forecast accuracy and credit excess-return forecasts (18:35–19:43). He also describes shared raw and derived data across equity and credit research and says he is not disclosing specific IP (19:43–21:03). The audio and timestamped ASR sidecars were durably retained privately on September 2, 2026; the public article contains no transcript body. This is a speaker-reported ML use case, not a model inventory, provider disclosure, data-license map, production-permission record, or return attribution. See the audio-recovery note.

Fidelity Investments Canada’s QRI supplies a Canadian institutional-manager comparator from a first-party May 2026 transcript. Karishma Kaul describes a quantitative-research organization led by Neil Constable, with roughly 200 quants supported by more than 250 technologists, central data-quality and validation controls, alternative-data work, model/backtest review, transaction-cost modelling, and systematic fixed-income layers spanning allocation, credit selection, macro risk timing, style signals, and flow or toxicity measures. Kaul also reports that recently adopted AI tools made the team two to three times more productive in research validation, new-data ingestion, and model review; that is a participant-reported productivity claim, not an independent time study or evidence of AI-attributed returns. The public record does not identify model providers or weights, training corpus, licenses, production permissions, or portfolio authority. The Spotify route is retained as a cross-platform identity check.

The Alpha Exchange interview with Jessica Stauth, identified by the publisher as CIO for Systematic Equities at Fidelity Investments, adds a separate December 2025 practitioner account. The public audio was recovered and processed with local WhisperX-MLX, producing a private timestamp layer. Stauth discusses earnings-call transcript mining, normalizing unstructured text into ticker/date-level fields, testing contextual LLM sentiment against simpler pattern-recognition approaches, broad data coverage, historical consistency, pre-specified risk guardrails, and slower recalibration of an underlying alpha model (33:35–36:17; 40:49–44:59; 46:47–47:14). This supports a dated research-process description, not a Fidelity model inventory, provider, training corpus, data-rights map, production-permission record, autonomous trading claim, or performance attribution. The capture note records the audio and ASR provenance, and the timestamped extract separates supported findings from unknowns.

J.P. Morgan Asset Management’s Hong Kong-based On Investors’ Minds APAC archive adds a dated, title-blind workflow route. In Episode 163, Fiona Harris describes analysts writing predetermined company game plans and using AI analysis to look for thesis drift across emails and meetings, with possible peer and competitor comparison (09:29–12:04; local ASR locator). The same episode says AI is used more broadly in day-to-day work at J.P. Morgan, but names no model, vendor, permissions, data-retention policy, evaluation set, or investment authority. Episode 160’s AI correction briefing adds public discussion of CFO scrutiny of token or volume costs and routing expensive models toward higher-value tasks (02:15–03:02; capture note). These are dated first-party practitioner statements and generic enterprise-finance observations, not evidence of a complete J.P. Morgan or hedge-fund AI system.

Bernstein’s Rupal Agarwal episode adds a Singapore/Asia research and event route. Bernstein’s personnel page identifies Agarwal as Asia Quantitative Strategist, while the April 2026 episode describes a move from summarization toward agentic workflows, unstructured-data research, governance, role-specific training, and human escalation. It names examples such as thesis and sentiment analysis, voice and text pattern recognition, scenario generation, and a vendor series involving Anthropic, Fiscal AI, Boosted AI, Delupa, AlphaSense, Rogo, Hebbia, LinqAlpha, Finster, and QuantConnect (01:58–06:41; 11:58–19:23; 24:15–27:32; capture note). Bernstein’s symposium disclosure separately names Agarwal’s Blackbook and participants from BlackRock, Tenucia Partners, LinqAlpha, and Model ML. These sources establish dated research and event disclosures; they do not establish Bernstein procurement, vendor partnerships, internal model ownership, or any firm’s production permissions or performance.

The same podcast’s related-episode graph adds title-blind personnel context: Natalya Dmitriyeva is described as former senior data-strategy executive at Two Sigma and Schonfeld; Evan Reich as having held data roles at Millennium, SAC Capital, Quandl, and Verition; Charles-Albert Lehalle as a professor formerly affiliated with ADIA and CFM; and Christina Qi as Databento’s CEO and former Domeyard founder. The recovered Lehalle audio adds a bounded historical CFM/ADIA workflow account: data-analytics teams, satellite imagery, and the conversion of high-dimensional alternative data into lower-dimensional inputs for standard alpha teams (15:54–18:27), plus discussion of data lineage and pipeline instrumentation for diagnosing performance changes (27:28–31:54). It also covers causal/world-model and embedding concepts (12:09–13:48; 45:13–47:35). These are speaker-reported historical/conceptual claims and do not establish current CFM/ADIA systems, vendors, permissions, or performance; see the ASR recovery note. The recovered Christina Qi episode adds a vendor-side AI demand signal: Qi says unnamed AI firms bought market data for direct AI use cases and may train on it (32:42–33:07), while distinguishing conventional market-data ML from LLM-assisted analysis (35:50–37:50). No customer is named or confirmed, so this is not a partnership or hedge-fund deployment claim; see the ASR recovery note. Vendor-side episodes with Brickroad, YipitData, and Consumer Edge add data-sourcing, consumer-data, transaction-data, and pricing routes. These publisher pages are discovery and personnel evidence; they do not establish current firm systems, customer relationships, or portfolio use.

The recovered full Evan Reich episode adds a temporal correction to the personnel route. The episode discusses data sourcing as an interface between quant, fundamental, research, legal, and compliance teams (05:39–05:55), says AI-enabled data-loading tools could reduce some loading work while leaving human interaction and creativity important (15:27–15:55), and discusses data-licensing limits on loading vendor data into an LLM (56:06–57:05). Most importantly, the host says Reich has left Verition and is now Chief Product Officer and Head of AI at BWG Global; Reich then describes BWG as bringing together BWG and Off the Record research operations and exploring AI-mediated research consumption (53:20–55:43). This current-role statement is host/guest-reported and requires first-party BWG confirmation. It should not be back-projected onto Verition, Millennium, SAC, or Quandl. See the ASR recovery note.

The regional search also surfaced two different Asia evidence classes. Citadel Securities’ current Machine Learning Researcher — PhD Graduate (Asia) posting covers Hong Kong and Singapore and names deep learning, NLP, unconventional and unstructured data, implementation, backtesting, and research documentation. It is hiring-intent evidence, not proof of a filled role or a particular model. Dnalyaw’s official site and founder Wayland Zhang’s public site describe a Hong Kong AI-native quant lab, reinforcement-learning trade-idea generation, language-model risk extraction, and a four-gate order veto; Zhang’s site also provides a University of Toronto computer-science lineage. These are self-authored architecture and biography claims. The reviewed pages do not independently verify regulatory status, live capital, model details, evaluation, or performance, so the displayed operational metrics are not treated as validated facts.

Regional evidence remains uneven. The Mackenzie/Dimensional media follow-up adds a Canadian quant-leader podcast, a publisher transcript, and a third-party Dimensional AI-investing video. The current Chinese-language media follow-up adds Lingjun, DeepWin, Mingshi, and WizardQuant video, roadshow, and recruiting routes while preserving original-language boundaries. A new Singapore/Hong Kong pass adds the AIMA Mark Wong / Dymon Asia transcript and an August 2026 Wu Chao / Going International Asset Management interview report. The AIMA catalog pass also adds Kate Smaje, Freddie Parker, Peter Kim, Darren Bowdern, and Kher Sheng Lee as regional and industry-context records. These dated public executive and media statements are not evidence of current deployment, model ownership, permissions, or performance. The next passes should prioritize Chinese-language official pages and registration records, Hong Kong SFC entity separation, Canadian securities filings and fund facts, and Australian ASIC/wholesale-fund disclosures before adding personnel or deployment claims.

Two Australia/Hong Kong source upgrades now separate first-party product and legal-entity surfaces from podcast-only evidence. Macquarie Asset Management’s March 2025 systematic-investing article describes a global Systematic Investments team, active quantitative equity strategies, more than 1,000 proprietary signals, cloud and Python research infrastructure, factor-calibration models, and AI/ML use in strategies and risk management, while its disclosures keep Macquarie Asset Management advisers and regional distributors distinct. RQI’s Hong Kong launch release identifies the RQI Global Value Fund as a Hong Kong vehicle, names an AI-enabled Alpha Signal Overlay and four investment leaders, and states that First Sentier Investors (Hong Kong) Limited issued the material without SFC review. These are product, strategy, personnel, and distribution disclosures; they do not disclose model inventories, data rights, production permissions, independent performance, or AI-attributed returns.

A recent title-blind Odds on Open interview with Suhaimi Zainul-Abidin adds a Singapore-based systematic-manager route. The episode describes Quantedge’s research as systematic and quantitative, then draws a firm-reported boundary between production investment models and supporting AI: Suhaimi says the production models are rules-based mathematical algorithms and that generative AI is not included in them in his account, while AI can support research productivity, data management, data cleaning, and execution (28:30–31:08). He also emphasizes understanding why models produce outputs and how risk and capital are allocated (18:44–20:24). Quantedge’s current team page identifies Suhaimi as CEO and Li Zixi as Head of Technology; it does not assign Zixi an AI-specific remit. This is recent speaker-reported evidence with a first-party title cross-check, not a model or deployment audit. It does not establish Quantedge’s model inventory, training data, vendor, evaluation, permissions, or AI-attributed performance. See the capture note. The same episode’s Spotify publisher page and recovered full transcript add coverage of the surrounding operating model: the discussion describes simple, economically motivated rules, a preference for developing early-career researchers, and a separation between support tooling and the production strategy (23:22–35:31). This is a capture-quality upgrade to the existing episode record, not a second interview or an independent audit. The publisher/guest’s scale and return figures remain separately attributed claims.

Quantedge’s July 7, 2026 first-party AI article adds a separate, more explicit control-plane account and names Xu Wei Chen as Vice-President, Quantitative Research alongside CEO Suhaimi Zainul-Abidin. The article describes AI as a supervised workflow actor for daily portfolio-change synthesis, documentation maintained alongside code, first-pass code review, anomaly flagging, and test-suite construction. It states that every AI-assisted output requires human sign-off, that the tools are sandboxed away from the core portfolio-management pipeline and deterministic strategy rules, and that proprietary data and code are processed under a protocol intended to prevent training of external models. This is a first-party strategy and control statement, not an implementation audit: the public article does not disclose model providers or versions, training data, evaluation fixtures, adoption coverage, permissions, order authority, or AI-attributed performance. See the source note.

A title-blind India pass also upgraded the existing JioBlackRock lead from podcast metadata to a timestamped transcript. In Rishi Kohli’s interview, Kohli describes combining BlackRock’s data-driven technology with Jio’s digital distribution (08:06–09:28), and describes Aladdin as integrating traditional and alternative data, alpha/risk/transaction-cost models, portfolio construction, execution, and pre-/post-trade controls (10:50–12:44). He describes a local/global investment-team mix (13:10–13:57) and an initial objective of fully systematic and quantitative products (15:35–16:56). JioBlackRock’s SAE research page and ProFolio page independently expose machine-learning, alternative-data, Aladdin, and human-oversight language; a SEBI addendum records Kohli’s CIO appointment effective August 4, 2025. These sources establish a public India-facing systematic-investing and platform record, not a complete model inventory, data-rights map, production permission map, or independently audited outcome. See the expanded source ledger.

A separate title-blind FINANZ26 route adds a Swiss panel with Georg von Wyss of BWM Value Investing, Reda Jürg Messikh of Pictet Asset Management, and Dennis Hagander of WealthArc. The publisher recording records von Wyss describing LLM-assisted exploratory research, conference-call summaries, and programming macros with human verification (11:30–13:38); Messikh describing a shift from preconceived factor ideas toward algorithms learning from examples (09:39–10:35), a COVID-era internal experiment on cross-data interactions (14:02–14:55), specialist-integrated data/model infrastructure (19:47–21:50), four-eyes checks and no automatic trade handoff in his account (33:34–34:55), and a private AI model placed into infrastructure for quant experimentation (39:04–40:26). Hagander describes lineage, governance, ownership, and data quality in a multi-custodian foundation (15:15–17:04). The recording and event metadata are useful for named personnel and workflow boundaries; they do not disclose model weights, training corpora, data licenses, evaluation fixtures, live permissions, or independently audited returns. See the FINANZ capture note.

Two additional French-language Swiss routes broaden the allocation layer. Fundo’s Cortex page describes a machine-learning allocation tool that combines large financial datasets, investment constraints, and multiple algorithms for investment-committee decision support, and names Bruno Maumené as CIO. Fundo’s home page places the firm’s origins at EPFL, while a linked Le Temps article provides older media context on Cortex. Kepler Unigestion describes a partnership created in 2025 between Kepler Cheuvreux and Unigestion, combining advanced AI, human expertise, quantitative methods, market data, and fundamental analysis. These are public positioning and partnership disclosures; they do not establish current model versions, training data, client deployment, portfolio permissions, or AI-attributed performance. See the capture note.

A fresh role-title search also recovered Macro Hive’s July 3, 2026 interview with Stefan Jansen, founder and CEO of Applied AI. The episode places LLMs in hypothesis and feature-variation work, describes agents as constrained controllers that call deterministic tools and emit standardized, auditable research artifacts (13:52–19:11), and discusses RAG, provenance, knowledge graphs, execution-focused reinforcement learning, and human goal-setting (19:11–33:30). Jansen’s public book and seminar materials add a six-library workflow spanning data, feature engineering, diagnostics, models, backtesting, and live operations; they do not identify his unnamed former investment employer or any client deployment. This is a methods and search-vocabulary comparator, not evidence about a tracked firm’s internal system, permissions, or returns. See the capture note.

The AIMA episode with Kate Smaje is now backed by a recovered public Acast enclosure and temporary local ASR. The McKinsey Global Head of Technology and AI describes research as an early agentic entry point, keeps human review near capital allocation, and discusses access controls, LLM gateways, and off-the-shelf versus proprietary-data/model use (15:22–17:20; 19:47–21:50; 37:38–39:23). This is industry context, not evidence of a named hedge fund’s model or deployment. See the capture note.

Asia-based Acadian-scale and quantitative-manager AI pass — August 13, 2026

The AUM comparison included several broad Asian asset managers that are not direct systematic peers. This pass therefore separates product-level AI, quantitative-investment models, GenAI operating infrastructure, and AI-themed investment activity. The full source inventory is in sources/06-industry-verticals/asia-acadian-scale-ai-public-signals-2026-raw.md.

Mirae Asset Global Investments

Mirae’s 2025 investment brochure describes a domain-driven AI asset-management program that has been developed since 2016 by quantitative investment professionals and engineering teams. The stated process covers data selection and preprocessing, learning objectives, model evaluation, and model selection across equities, bonds, commodities, quant-active, absolute-return, covered-call, and asset-allocation strategies. The same document connects the program to robo-advisory, digital advice, personalized portfolios, and a planned expansion through Stockspot and Wealthspot.

Mirae also exposes product-level model detail. Its Korean AI Global Momentum fund page says an AI algorithm uses returns, volatility, rate spreads, and cross-asset correlations to generate asset-level momentum scores and portfolio weights through a deep-learning neural network. That is evidence of a named investment product using deep learning. It is not evidence of an LLM, an agentic research system, or a common model across Mirae’s global AUM. No named current AI executive or public foundation-model program was located in the reviewed sources.

SCB Asset Management — Thailand

SCB’s Machine Learning Thai Equity prospectus states that the fund uses quantitative analysis and machine-learning techniques in a manager-developed system for factor-based Thai-equity selection. The document lists นายพูนศักดิ์ โล่ห์สุนทร as Executive Director of the Machine Learning investment group and records his University of Southern California electrical-engineering, mathematical-finance, and mathematics degrees. It lists นายสถิตย์พงษ์ จันทรจิรวงศ์ as Director of Machine Learning Equities, นายกฤช จันทร์หนัก as a quantitative-equity manager with prior QIS Capital, J.P. Morgan Securities Hong Kong, and WorldQuant Research Thailand roles, and นายณัฐวุฒิ เดชบดินทร์ as a quantitative-equity manager with Chulalongkorn financial-engineering and nano-engineering degrees. These are prospectus-listed roles and biographies for the stated period, not proof of current employment beyond the document, individual model ownership, or AI-attributed performance. The English product page and SCB Securities’ AI allocation product page provide the related product-policy context. See the regional source note.

Nikko / Amova Asset Management

The reviewed Amova pages primarily expose broad active/passive investment capabilities, fundamental research, and AI as a market and investment theme. The post-2025 Amova identity is clear, but the public sources reviewed here do not identify a firm-controlled GenAI platform, an internal AI/ML investment team, a named model, or an agent runtime. That is a negative-search boundary rather than a conclusion about private activity. The reviewed evidence supports keeping Amova/Nikko in the partial-coverage queue until its careers, Japanese-language research, and internal technology disclosures are searched more fully.

Samsung Asset Management

Samsung’s official profile books identify an AI & Quant EMP Team that develops portfolio solutions using quantitative and AI/ML algorithms. Its multi-asset materials describe AI and quantitative algorithms in asset-allocation strategies spanning ETFs and funds. This is a direct team-level and product-method signal, but the official sources do not name the algorithms, model providers, LLMs, agents, or a current AI executive.

A dated Maeil Business interview identifies Lim Byung-hyo as head of the AI Quant Management Team and discusses asset-allocation products and crisis management. That personnel item is retained as dated secondary evidence, not as a current roster confirmation. A third-party people-data page names another purported AI Innovation Team head, but it is not used as public confirmation without a firm-controlled source.

Asset Management One

Asset Management One’s 2017 Financial Innovation Team announcement is unusually specific for a traditional manager: it describes Big Data infrastructure, AI technology, integration into both quantitative and judgmental investment analysis, and collaboration with Mizuho, universities, and venture companies. Its current Deep AI fund page says the firm uses a proprietary deep-learning model to select relatively attractive non-Japan global equities; the page reported approximately JPY 60.8B in net assets on July 17, 2026. The current organization chart retains Financial Innovation and Quantitative & Index Strategies as separate investment functions.

Additional Japanese-language reporting published June 3, 2026 names Deep AI fund managers Junichiro Tobita and Tatsuya Oka and gives more process detail: AI scores approximately 4,000 developed- and emerging-market stocks using price, financial, earnings-forecast, and news-text inputs; software then constructs the portfolio with a separate risk model; human managers make adjustments for concentration, large-cap exposure, and turnover. This is a manager interview, not an independent audit, and it does not identify a model version, training-data rights, production permissions, or a GenAI or agent system. See the capture note and original Japanese interview.

The public evidence therefore establishes a historical AI organization and a current deep-learning product. It does not establish a current GenAI program, model inventory, named AI owner, agent workflow, or model-to-portfolio permission map.

Nomura Asset Management and Nomura Group

Nomura exposes several separate AI layers. Its Innovation page says the Asset Management Innovation Lab researches management methods using machine learning, statistics, and information science. The current Nomura Asset Management career page describes quantitative strategy research, model implementation, and backtesting, with AI/ML, NLP, financial text, news, and disclosure analysis listed as relevant skills. A separate DX Strategy AI/Data Team role describes developing applications with group-approved LLMs, prompt testing, workflow prototyping, and AI governance.

At group level, Nomura announced a November 2025 collaboration with OpenAI involving Deep Research, proprietary and external data, investment advice, market analysis, and data solutions, alongside security and governance requirements. Current Asia-Pacific listings include AI Data Scientist and AI Engineer roles. The public profile reviewed for a Nomura Innovation Lab researcher identifies only “Kei” by first name; it is not used as a fully identified personnel record. These sources establish investment-research and approved-LLM development intent across Nomura’s group, but they do not establish one AI owner, a complete model inventory, or production portfolio authority inside Nomura Asset Management.

ChinaAMC

ChinaAMC’s public AI record is more visible at the product and stewardship layers than at the internal investment-technology layer. Its AI ETF page and factsheet identify Li Jun as portfolio manager and report the fund’s net assets, but the product tracks an AI-sector index; that does not prove AI is used to select the portfolio. ChinaAMC also launched a 2026 report on ESG in the AI era, covering AI-related opportunities, risks, and responsible-AI practices.

The reviewed sources did not disclose a ChinaAMC foundation model, internal GenAI platform, named AI engineering leader, agentic research workflow, or AI-attributed investment record. ChinaAMC’s very large total AUM is therefore not treated as evidence of comparable AI depth.

Dymon Asia Capital

Dymon’s public AI disclosure is practitioner-level rather than architectural. In an AIMA interview, Co-CEO, COO, and CRO Mark Wong says the firm is examining GenAI for productivity, operational-process streamlining, and processing large data volumes. He frames it as augmentation of human decision-making and names privacy and bias as concerns. Dymon identifies Sian Goh as Partner and Head of Research and Strategy; a J.P. Morgan podcast discusses the firm’s combination of local market knowledge, micro analysis, and macro developments.

The reviewed record does not identify a Dymon AI team, model, LLM provider, agent runtime, evaluation system, or live investment permission. Dymon remains a useful Asia-based hedge-fund control case, but its public evidence is not comparable in type to a job-specified AI platform or a disclosed investment product.

Eastspring, SPARX, Simplex, and Macquarie QIS

Eastspring’s current quantitative page describes a proprietary alpha model and Asian-focused quantitative research. A 2021 paper by Ben Dunn, Head of Quantitative Strategies, discusses ML/AI as a complement to quantitative investing and emphasizes interpretability and causal relationships. A current Singapore posting for Senior Manager / Assistant Director — AI Researcher, Quantitative Strategies adds a more specific hiring signal: it describes novel-signal discovery, structured and alternative data including text/news, time-series and cross-sectional modelling, regime-shift controls, walk-forward evaluation, leakage controls, model-risk documentation, explainability, monitoring, and controlled retirement of decaying signals. This is a detailed role specification, not confirmation that the position was filled or that a named model is in production. No current GenAI platform or agent deployment was located.

SPARX’s AI signals concern venture investing: the PKSHA SPARX Algorithm Fund and Mirai Creation investments support AI, advanced software, and world-model companies. Those are investment themes and portfolio-company exposures, not evidence that SPARX uses AI to manage its listed-equity strategies.

Simplex describes proprietary models for multi-QIS strategies, alternative risk premia, overlays, portfolio optimization, and independent risk monitoring. Its reviewed current QIS page does not name AI, ML, GenAI, or agents.

Macquarie QIS provides a separate Australia-headquartered bank/QIS comparator. Its public materials describe a machine-learning platform, reinforcement-learning signals for systematic volatility with Protean Capital, ML-based economic-regime classification and signal calibration, and a library of more than 1,000 signals. A 2026 Macquarie Group/CGM briefing reports approximately 20 AI use cases implemented, around 30 in the pipeline, about 70% of CGM engineers using AI to accelerate delivery, advanced trading signals and AI-driven insights, and agentic solutions in the operating roadmap. The QIS-specific sources and the group-level CGM metrics are kept separate; the latter is not treated as proof of an LLM connected to QIS portfolios.

Coverage boundary after this pass

The Asia comparison now has substantive public AI evidence for Mirae, Samsung, Asset Management One, Nomura, ChinaAMC, Dymon, Eastspring, SPARX, Simplex, and Macquarie QIS, with Amova/Nikko remaining a partial-coverage case. “Substantive” describes the amount and specificity of public evidence reviewed, not a judgment about any firm’s capability or results. Across all of these firms, the public record still does not provide a complete current personnel roster, model inventory, agent permissions, production endpoints, internal benchmarks, or independently audited AI-attributed performance.

Newly verified regional quant and technology routes

The August 30 Chinese-language title-blind pass adds a separate recruiting layer. Inno Asset / 因诺 announced a “large model × quant” campus program on April 7, 2026; the detailed role description, updated March 30, describes LLM use in strategy research, factor discovery, signal construction, unstructured announcements/reports/news, and evaluation for predictive ability, robustness, interpretability, and backtest results. The university announcement and job mirror establish recruiting intent; they do not name a model/provider, filled employee, data licence, deployment endpoint, or permission boundary.

Qilin Investment / 启林 exposes a December 22, 2025 university-hosted posting for AI algorithm research, portfolio optimization, and machine-learning systems. It names feature extraction, return prediction, deep-learning networks, reinforcement learning, PyTorch/TensorFlow, multi-period and multi-signal optimization, and optional TensorRT, mixed-precision, and distributed-inference experience. A separate March 2, 2026 Zhejiang University posting adds high-frequency Level-2/tick-data research, fundamental and alternative-data research, portfolio-capacity work, and Python machine-learning-systems engineering; it also describes a team with graduates of Tsinghua, Peking, Fudan, Shanghai Jiao Tong, USTC, and Stanford, plus IT staff from Microsoft and major Chinese technology companies. These are employer-provided recruiting and organization-description claims, not proof of filled roles, individual academic lineage, installed frameworks, model ownership, live strategies, or performance. Public profiles for Shihang Song and Jianxiong Cui remain personnel leads, not confirmed AI-leadership evidence.

Micro Trading / 微观博易 adds a 2026 internship route with a large-model application-development track. The listing describes agent evaluation and deployment, autonomous planning and tool use, Tool/Model/Memory orchestration, and Hugging Face, OpenAI API, LangChain, AutoGen, MCP, and A2A. A separate Chinese-university posting describes the manager’s low-latency automated trading and machine-learning research surface. These are recruiting records, with the 2026 internship closed May 11; they do not establish internal agent ownership, production scope, or live trading authority. A public profile for Qi Zhao references a FinText financial embedding trained on Dow Jones Newswires, but the public profile does not establish Micro Trading ownership or use.

The Wukong Investment AI-quant engineer posting, published March 3, 2026, names deep learning, time-series models, LLMs, NLP, computer vision, and reinforcement learning for factor, signal, and trading research. It asks the hire to work with business teams and fund managers to identify LLM application scenarios and iterate predictive models. The page is a university-hosted hiring artifact: it does not name a model/provider, filled employee, dataset, permission map, or performance result.

ChengQi Funds currently lists machine-learning research among formal roles. Its detailed machine-learning researcher posting, dated April 11, 2022, specifies reinforcement learning, deep learning, and machine learning for quantitative and portfolio strategies, including neural networks, decision trees/GBDT, and reinforcement learning. The detailed role is historical; the current careers page confirms a role family but not a current program, model, or owner.

Two adjacent Singapore controls clarify the entity boundary. Capital Park claims an 80/20 AI-agent-and-human operating model and says agents watch markets, run research, place hedges, and log activity, but the same page states that it is a private investment office managing only proprietary capital rather than outside funds. GIC’s AI Alpha Group posting is a sovereign-wealth-fund route, not a hedge-fund record; the current page says the position is filled, while the earlier indexed description referenced a GenAI-enabled fundamental-investment process with RAG, fine-tuning, multi-agent systems, and contextual engineering. Neither control establishes current model inventory or deployment. See the capture note.

Temasek’s Investment Data Science posting adds a separate Singapore institutional-investor comparator. The “Investment Researcher” role bridges sector and market investment teams with Investment Data Science, and names web traffic, mobile-app usage, and natural-language data as example alternative datasets. It describes AI-enabled workflows for sourcing information, company analysis, hypothesis testing, and synthesis, plus external-supplier evaluation and dataset-quality assessment. The title itself does not contain AI, which makes it a useful title-blind discovery pattern. This is a current official job specification, not proof of a filled role, model/provider, training corpus, production deployment, investment authority, or performance. See the capture note.

JQ Investments / 佳期投资 provides a mainland-China quant-fund route that was not present in the earlier regional set. Its current public company profile describes a Shanghai-based quantitative hedge fund focused on equities, futures, and derivatives, lists machine learning and deep learning as specialties, and describes a team with backgrounds at universities including Harvard, Princeton, Stanford, Tsinghua, and Peking University. JQ’s positions page exposes separate quantitative-research, deep-learning, and technology-development categories, while a 2021 Tsinghua recruiting notice describes statistical and machine-learning methods and says deep-learning researchers form part of JQ’s core competencies. The university list and company profile are self-descriptions, and the recruiting notice is historical: these sources do not establish a complete current roster, individual academic lineage, model inventory, agent use, permissions, or performance attribution.

A dated recruitment follow-up adds current university-board routes without converting them into personnel evidence. JQ’s careers page displays 2026 campus recruiting across quantitative research, deep learning, technology development, and operations. A Cornell posting opened August 19, 2026 and describes statistical and machine-learning methods for quantitative research; an MIT posting opened the same day and describes the research/software bridge, Python, advanced C++, probability, statistics, algorithms, and performance engineering. A Boston University posting opened August 7, 2025 and expired August 6, 2026, describing PhD-level deep-learning research and publication expectations. These are university-hosted vacancy records, not filled-role confirmations, academic lineages, model disclosures, or production evidence. See the JQ recruiting capture note.

Jasper Capital adds a distinct China-A-share manager route that was incorrectly attached to Arrowstreet in the discovery queue. The 2021 Opalesque video page and related manager article identify Dr. Bo Huang as a senior portfolio manager and head of Jasper’s quantitative investment team, and describe quantitative long-only, market-neutral, and directional strategies focused on China A-shares. The article attributes the opportunity set to retail-driven mispricing and market microstructure and describes the manager as mid-frequency; these are dated publisher and manager-attributed statements, not an independent strategy assessment. The linked YouTube recording now supplies an English automatic-caption track with navigation around 02:09–02:21 for technical, fundamental, alternative-data, and event-driven signals/models; 07:15–07:22 for forecasts, risk, and transaction/hedging-cost assumptions; and 08:06–08:14 for a Barra risk-model reference. The captions are not manually verified. Jasper Capital Hong Kong says the Hong Kong entity was founded in December 2017 and is supported by Jasper Capital International in mainland China. A Blackstone sub-adviser profile separately describes alpha-factor research, factor integration, portfolio construction, and algorithmic execution, and reports $464.0 million AUM as of March 31, 2023; Blackstone says the information came from or was derived from the sub-adviser. This route supports dated strategy, entity, and personnel context, but it does not establish current AI/GenAI use, a model inventory, training data, data rights, production permissions, current AUM, or AI-attributed performance. The capture note records the caption recovery and queue correction.

Dynamic Technology Lab adds a Singapore-based manager route. DTL’s official page describes a MAS-regulated licensed fund manager, proprietary OMS and risk-monitoring systems, global multi-asset trading, and machine learning in investment decision-making. A machine-learning quantitative-researcher internship posting hosted by Sungkyunkwan University describes data cleaning, model research, machine/deep-learning trading signals, backtesting, simulation, statistical analysis, and Chinese NLP as a desired skill. This is evidence of firm positioning and a dated hiring artifact, not a named current researcher, live model, portfolio permission, or AI-attributed result.

Roger McIntosh’s ExpertGate interview adds an Australian institutional-quantitative route. The AICRO profile identifies him as Investment Director at IFM Investors and records La Trobe mathematical-statistics/applied-mathematics training plus a University of Melbourne quantitative-finance master’s degree; a Financial Standard appointment report independently dates his December 2024 move to IFM’s quantitative-equities team. The interview describes factor-based and sustainable portfolio construction, data-provider workflows, and academic engagement (00:59–05:50; 07:04–10:37). Its publisher description says McIntosh is interested in ML/AI for factor ranking and dynamic asset allocation, but that is a research interest, not evidence that IFM adopted a named model or agent. See the capture note.

Tensor Investment Corporation is retained as an adjacent technology-provider route rather than a hedge-fund-manager entry. Its official site describes AI-systems research across commodities, fixed income, and digital assets, an experimental agentic-AI division peripheral to trading, and services for CTAs, market makers, hedge funds, and digital-asset institutions. Its listed services include deep-learning execution, compute scaling, model parameterization, strategy creation, and agentic systems for alpha discovery, feature generation, and strategy evaluation. These are Tensor’s own product and customer-category claims; no customer identity or named-fund deployment was located, so the claims are not transferred to any tracked manager.

Reference points: different public operating models

CloudQuant / AnacondaCON: historical research-throughput design

A 2017 AnacondaCON presentation by Morgan Slade, identified in the publisher description with CloudQuant, describes a Python-accessible cloud simulation platform, alternative-data access, and a model for bringing more data scientists into strategy research (01:40–02:19; 12:10–12:59). Slade frames machine learning as a way to shorten predictor search while leaving research questions and judgment with people. The presentation also gives an illustrative case involving a non-programming analyst learning Python and changing reported strategy diagnostics after applying machine learning (13:11–17:58). Those figures are vendor-presented examples with no public benchmark protocol or independent replication, so they are not used as manager-performance evidence. The slide deck and capture note preserve the historical and attribution boundaries.

Man AHL: public strategy-design and agentic-workflow disclosure

Man AHL publicly describes more than three decades of systematic research and an Oxford-Man Institute connection. Its 2025 “Big Innovation Imperative” article describes a suite of AI agents for research productivity and front-to-back trend-signal generation, while retaining human researchers in the process. Its 2025 Alpha Assistant article describes an agent with access to proprietary data, internal libraries, analytical tools, and an approval step before execution. Its 2026 AlphaTrend article describes a specialized, predefined workflow for generating, implementing, and researching trend signals.

The media layer is broader than those firm-authored AI articles. Two title-blind publisher appearances with Russell Korgaonkar and Man AHL’s 2021 Trend Following Radio interview add dated CIO, Oxford physics, research-responsibility, model-aging, execution, position-sizing, and risk context. The 2023 Top Traders Unplugged page also exposes a full timestamped transcript; the 2023 capture note records the private audio provenance, while a 2021 local-ASR note records the recovered recording and navigation-only transcript layer. They are historical practitioner-media records, not evidence about current AlphaTrend or AlphaGPT implementation, permissions, or performance.

A separate title-blind Bloomberg Television interview with Luke Ellis, published March 18, 2021, adds a dated Man Group executive account of trading Bitcoin futures and physical markets. Ellis describes liquidity and capacity as variables that can change over time, and uses Chinese egg futures as an example of a market whose temporary speculative liquidity later receded; he says such temporary liquidity can make a contract useful for models without making the opportunity durable (00:07–01:31). This is market-selection and model-capacity evidence, not an AI disclosure: the clip does not identify a model, data source, allocation rule, current title, permission set, or performance result. The capture note records the metadata-versus-transcript attribution boundary.

Man’s public AlphaGPT disclosure gives the workflow more definition: an Idea Person proposes hypotheses, an Implementer converts them into executable research code against proprietary tools and databases, and an Evaluator applies statistical, risk, and economic checks before a signal can be considered for deployment. Man says the workflow remains under human oversight and strategic direction. Its AlphaTrend article names Claude 4.0 Sonnet and GPT-5 in a broad idea-generation experiment, which is a model-use disclosure rather than a complete model inventory. Man’s February 2026 Anthropic partnership announcement and 2025 annual report add enterprise context around Claude, ManGPT, AlphaGPT, training, governance, and an extensible internal platform. These sources still do not disclose organization-wide user counts, model-evaluation records, or performance attribution.

An additional August 5, 2026 interview with Greg Bond on The Bridge by iCapital gives a current executive account of the same general operating layer from a different publisher. Bond describes “digital researchers,” institutional skills files or harnesses that encode research context, and broad adoption through both user-facing interfaces and direct coding tools (21:29–26:08). He also discusses using varied model-generated views to look for relatively uncorrelated research ideas, multi-reviewer checking, and simulated market scenarios for stress testing (23:29–24:11; 26:03–29:42). The transcript is speaker-reported and does not disclose model versions, training data, data rights, permission boundaries, rollout metrics, capital authority, or AI-attributed returns; see the source note.

A previously unindexed WatersTechnology report, published September 24, 2025, adds a dated publisher account of Gary Collier’s public statement at Bloomberg’s Investment Management Summit in London. The accessible opening says Man agents had generated independent alpha ideas, while Collier also said the ideas were still vetted by human committees and were not in a straight-through process at that time. This is a source-reported statement about the idea-generation and authorization boundary, not evidence of realized returns, deployed capital, a named agent, or a particular Man fund. The article is subscription-limited after the opening; the capture note preserves that access boundary.

An additional first-party Man route fills a historical operating gap. In the March 2024 AI Diary of a Quant, Russell Korgaonkar describes Man AHL’s use of AI across data analysis, strategy development, and execution; identifies ManGPT as an employee tool using Microsoft’s ChatGPT API; and describes GPT-assisted processing of filings, social material, data-quality checks, and data-aggregator reports. The article also describes human-selected parameter searches, random-noise comparisons, information-decay checks, out-of-sample testing, and reinforcement-learning-based adaptive intelligent routing. These are dated, firm-reported operating statements, not a current model registry or independent performance evidence.

The same first-party personnel graph adds Chao Xia’s current Man Numeric profile, which lists her as Deputy Director of Research, Investment Committee member, co-portfolio manager, and research-team manager with published focus areas including machine learning, artificial intelligence, alternative data, event strategies, and macro timing. Slavi Marinov’s current profile lists him as Head of Equities Alpha, Systematic, and records a former Head of Machine Learning role at Man AHL spanning hardware, research frameworks, and live-trading ML alphas. Current titles and former responsibilities are kept separate; neither page establishes ownership of AlphaGPT, AlphaTrend, or a current investment permission set. The capture note preserves the dates and boundaries.

A dated Man AHL first-party account, published November 17, 2016, supplies a separate historical ML lineage. Anthony Ledford is identified as Chief Scientist, and the article says AHL had researched ML for roughly five years and had ML-based systems trading in a multi-strategy client portfolio since early 2014. It describes the Oxford-Man collaboration and a method-transfer example in which Bayesian classification work from Galaxy Zoo supernova research was applied to extracting predictive signals from broker recommendations. This is useful evidence of a historical research pathway and a dated firm statement, not evidence of current AlphaGPT or AlphaTrend scope, current personnel, model weights, data rights, permissions, or performance. See the capture note.

A title-blind YouTube recovery adds four dated Man AHL media routes that should remain in separate evidence lanes. Anthony Ledford’s FundForum interview (2016) describes heterogeneous data, financial-market forecasting, and research → company-funded test trading → client-portfolio eligibility (00:25–03:08). The official AHL Optimisation explainer (2016) teaches cross-sectional portfolio optimisation and convex risk constraints (00:03–03:25). A MongoDB case study (2016) reports an approximately 25x tick-throughput increase for a Man AHL tick-storage/distribution system (00:32–01:26); that remains a vendor-published, speaker-attributed claim. A Man AHL-hosted Simon Knowles talk (2018) supplies model/compute vocabulary but identifies Knowles as a Graphcore co-founder and does not establish Man AHL adoption. None of these recordings discloses a current LLM, agent, training corpus, permission map, autonomous order authority, or performance result. See the follow-up capture note.

An Anthropic Code with Claude session adds a separate vendor-conference claim: AI-researched, backtested, and proposed signals were described as running real capital at Man Group, alongside a governed skills framework, a core data layer, roughly 750 developers and quants, and more than 100 skills. This does not identify the production system as AlphaGPT or AlphaTrend, and it provides no live P&L, capital, risk limits, approval rate, or signal names. The July 2026 Odd Lots discussion adds token-spend and enablement context, but its detailed transcript remains a secondary source.

Two current first-party Man surfaces add operating context around those named workflows. The Technology at Man page says technology supports research, decision-making, execution, and reporting; reports more than 600 technologists and quants, more than 200 datasets added in 2025, and $7.5 trillion in notional trading volume as of December 31, 2025; and names Gary Collier as Chief Technology Officer. It also identifies ArcticDB as an open-source project developed into an enterprise time-series data solution. These are firm-reported platform and scale statements, not evidence that all of that infrastructure supports GenAI or that a named AI system controls a portfolio. Man’s August 11, 2026 AI boundary article, authored by Gregory Bond, discusses how AI could alter the balance between internal coordination, outsourcing, and data integration. That is organizational strategy commentary, not evidence of Man’s procurement choices or a production deployment. The source note preserves the boundaries.

A current UK conference route adds agenda-level corroboration for the same personnel and operating vocabulary. IA Engine’s EmTech Global 2026 agenda lists Gary Collier, CTO of Man Group, on a “Scaling Agentic AI in Investment Management” session about autonomous agents and GenAI across research, distribution, client service, quant research, and operations. His speaker biography says his remit includes Man Group technology and data science, front-office technology underpinning investment decision making, ROSA, and other enterprise systems. This is conference metadata, not a recording or transcript; it does not disclose Man’s model inventory, data rights, permission boundary, deployment scope, investment authority, or performance.

The historical infrastructure trail is more specific. MongoDB’s November 10, 2015 customer announcement says Man AHL released Arctic, a MongoDB-powered tick store, as open source; the announcement calls it the primary market-data store for quantitative researchers and reports a 40x cost saving, a 25x processing improvement, storage at 40% of the prior solution, and model fitting 25x faster. A companion MongoDB customer video describes researchers using large datasets, simulations, and historical backtests, and links the data layer to tick distribution and a scalable compute cluster (00:48–01:04, 01:04–01:26). These figures are dated vendor/customer claims: the public material does not provide a benchmark protocol, independent replication, current architecture, or evidence that Arctic powered AlphaGPT, AlphaTrend, or any current GenAI system. The capture note keeps the infrastructure route separate from current model and agent claims.

The title-blind media pass adds a separate distribution and research layer. Man’s official YouTube channel exposes seven public videos, including caption-backed routes on systematic credit, emerging-market alternative data, Python training, data engineering, risk engineering, and market-data infrastructure. The 2025 emerging-market episode credits Ori Ben-Akiva and Jason Moore in its publisher description and discusses alternative data, geolocation data, geopolitical events, model-use mismatch, and human oversight (00:00–00:14, 01:14–01:20, 02:45–03:20). It does not name an AI model or claim a GenAI workflow; the capture note preserves the automatic-caption and no-deployment boundaries. The official Exceptional Data episode and Algorithmic Advantage episode provide edited first-party transcripts on data ingestion, Python, quantitative credit, risk modelling, portfolio construction, and electronic execution; the Trend Setters episode adds dated Man AHL trend-following context. The Oxford-Man Institute archive adds 2026 machine-learning and quantitative-finance conference recordings and an AI-in-capital-markets talk. These routes expand the public vocabulary and personnel/media map, but they do not establish that a named speaker owns AlphaGPT, AlphaTrend, or a live trading system. The full capture preserves caption, academic, and deployment boundaries in the research ledger.

Key public people

Person	Publicly visible role or authorship	What the public record supports	Boundary
Russell Korgaonkar	Head of Systematic, Man Group; CIO, Man AHL	Public executive framing for AI, research, markets, and risk-management work at Man AHL	The public articles do not assign him day-to-day ownership of a specific agent implementation
Martin Luk	Senior Quantitative Researcher, Man AHL; participant/co-author in public LLM research	Named authorship for Alpha Assistant and AlphaTrend; public connection between LLM work and quantitative research	Participation and authorship do not establish sole engineering ownership or production authority
Tarek Abou Zeid	Partner and Head of Client Portfolio Management, Man AHL	Co-authorship of the AlphaTrend research-workflow disclosure; client-portfolio-management perspective	The public record does not establish model-development ownership
Otto van Hemert	Senior Advisor, Man Group	Co-author of the AlphaTrend disclosure and a named senior research contributor	The source does not specify operational responsibility for the system
Giuliana Bordigoni	Portfolio Manager and Director of Alpha Research, Man AHL	Publicly listed Alpha Research leadership adjacent to the disclosed research workflow	The public record does not tie the role to a named GenAI system
Tushara Fernando	Publicly identified with data and AI strategy; speaker in the Anthropic session	Public connection between firmwide AI enablement, governed skills, and the reported production-signal claim	The vendor session does not identify production systems, capital, or performance
Matthew Hertz	Publicly identified with the central AI platform	Central-platform personnel signal adjacent to ManGPT and firmwide enablement	A personnel page does not establish model ownership or deployment scope
Gary Collier	Current Man Group CTO; formerly CTO of Man AHL in dated material	Current technology leadership and a historical Man AHL technology connection	The public record does not establish current ownership of AlphaGPT, AlphaTrend, or a specific model
Slavi Marinov; James Blackburn	Public Man profiles identify equities-alpha and platform-engineering leadership	Investment and platform interfaces around the AI program	Public profiles do not tie either person to AlphaGPT or AlphaTrend ownership

The important distinction is workflow location. Man AHL’s public material places GenAI inside quantitative research and strategy design rather than presenting a general chatbot as a replacement for portfolio construction. The disclosures identify tools, internal context, workflow stages, human approval, and evaluation boundaries; they do not establish implementation scope across the broader organization or investment performance.

AQR: the skeptical control case

AQR’s public ML material presents machine learning inside a disciplined, theory-led systematic process. Its broader “Can Machines Learn Finance?” work is useful because it emphasizes the low signal-to-noise environment, small economic effects, data mining risk, and the continuing value of economic reasoning and simpler models. The evidence is not a current GenAI platform disclosure. It is a control against overstating what more generated features or more backtests mean.

A current AQR UCITS page, Can Machines Learn Finance?, states that ML is used as a tool within the systematic investment process to examine relationships among valuation, momentum, quality, macro conditions, and other financial data. It explicitly describes economic grounding, pattern selection, and real-world validation intended to reduce overfitting. This is a current first-party method statement, not evidence of a GenAI platform or an AI-attributed return stream. The previously indexed AQR video-introduction locator, which search results still describe as Bryan Kelly explaining ML in finance, currently resolves to AQR’s 404 page; it remains a locator failure rather than a recovered video source.

AQR’s A New Paradigm in Active Equity, published February 5, 2025 and updated February 19, 2025, adds a title-blind research-framework route. The paper discusses large language models, machine learning, and alternative data in active equity and presents systematic investing as a way to engage with those changes. It expressly says the discussion is not specific to an AQR strategy or product. This is firm-published analytical vocabulary, not evidence of a named AQR model, agent, provider, dataset, production workflow, portfolio authority, or AI-attributed performance. See the capture note.

The recovery pass found a regional mirror of that video: AQR Australia’s Can Machines Learn Finance? page embeds a public Brightcove player and describes Bryan Kelly as AQR’s Head of Machine Learning. The four-minute-15-second player asset was downloaded and locally transcribed on August 28, 2026; its SHA-256 is recorded in the AQR/GMO source ledger. The timestamped audio-backed pass describes finance as a low-signal-to-noise setting, treats interpretability as relevant to fiduciary risk, and names return forecasting, risk–return modeling, risk, transaction costs, and NLP as AQR application areas (00:00:28–00:04:03). The player’s internal title is a generic “Blank Template - Apr 20, 2026,” so the page text controls identity and the transcript remains an automatic ASR layer rather than a verbatim human transcript. A separate 2026 AQR/J.P. Morgan interview with Jordan Brooks says new data and AI tools have contributed to NLP-derived signals and ML-based signal weighting within a fundamentally driven multi-asset process. AQR’s leadership page identifies Brooks as Principal and Head of the Macro Strategies Group. The interview is useful current method evidence; it does not identify model vendors, weights, permissions, or AI-attributed performance.

AlphaSimplex’s Kathryn Kaminski is a newly recovered current practitioner-media route. In the October 31, 2025 Meb Faber episode, Kaminski is identified as Chief Research Strategist and co-portfolio manager for AlphaSimplex’s Managed Futures and Global Alternatives strategies; the episode’s 35:02 chapter is explicitly titled “Implementing AI.” In the captured 34:51–35:34 segment, she describes long-standing machine-learning use as a tool, including summarization and other tasks, and discusses possible analyst-role support, code writing, and process facilitation. This establishes a named public statement about current ML use and contemplated applications, not a model, vendor, agent-permission map, or AI-attributed performance. The local audio/transcript note records the MP3 hash, WhisperX-MLX capture, and source boundaries.

A historical title-blind route adds a different kind of AlphaSimplex evidence. In Top Traders Unplugged Episode 99 and Episode 100, published in April 2018, Robert Sinnott discusses financial-ML non-stationarity, signal decay, economic constraints, and transparency. The Episode 100 transcript attributes approximately 40% of portfolio risk to trend signals whose weights were adjusted by interpretable tools such as decision trees and kernel regression; it also describes a path from research or academic ideas through implementation, testing, out-of-sample evaluation, and deployment. This is a dated practitioner account, not a current AlphaSimplex model inventory or independent performance audit. It does not establish GenAI, current personnel, data rights, permissions, or AI-attributed returns. See the capture note.

A second newly recovered CFA Institute conversation adds a personnel-lineage route relevant to ADIA and AQR. Published March 18, 2025, the recording identifies Marcos López de Prado as ADIA’s Global Head of Quantitative Research and Development, Cornell professor of practice in financial machine learning, former AQR Head of Machine Learning, and former Guggenheim quantitative-investment-strategies leader (00:20–01:23). Its framing connects financial ML to black-box risk, causal inference, factor limitations, and causal-discovery models (01:56–02:22); the speaker also discusses responsible use and misuse of AI in investing (07:50–08:27). The recording establishes role history and research vocabulary, not ADIA’s deployed models, permissions, collaborators, or performance. See the capture note.

An official-channel follow-up recovered three additional AQR video routes that were absent from the stable-ID ledger: Why AQR?, AQR in 60 Seconds: Ph.D. Program, and AQR Research. The 2015–2020 clips use firm-controlled recruiting and research language: applied academic-style work, quantitative research applied to data, empirical testing, systematic exposures, transaction costs, liquidity, and the need to explain portfolio construction. Their English captions are automatic and contain entity errors, so the capture note records only short timestamped paraphrases. These videos add historical first-party channel coverage and research-culture context; they do not establish a current GenAI platform, a named speaker from the captions, model ownership, or deployment.

Registered-fund prospectuses expose additional strategy labels

The SEC filing sweep adds a different disclosure surface. These documents are not substitutes for a private manager’s internal model documentation, but they often state the investable target, feature family, model class, and human override more precisely than a general firm page.

• AQR: A 2026 funds prospectus says the adviser uses machine learning and natural-language processing for certain strategies and names consumer transactions and behavior, social-media sentiment, internet-search and traffic data, and model/data testing as part of the risk discussion. A separate preliminary prospectus for AQR Delphi describes a long/short framework built around beta, quality, value, and proprietary quantitative models. The Delphi filing is preliminary and the fund had no operating history; it is strategy-design evidence, not a performance or production claim. (AQR Funds Prospectus; AQR Delphi preliminary prospectus)
• Acadian: A Voya prospectus says Acadian pays hard dollars for third-party data and research that support quantitative processes and describes automated broker selection. An older Acadian prospectus biography identifies Constantine Papageorgiou’s research focus as including machine learning and pattern recognition. Together these filings support a data-intensive quantitative infrastructure and a dated personnel signal; they do not identify a current GenAI model, vendor, or agent. (Voya prospectus; Acadian prospectus biography)
• Arrowstreet: The reviewed prospectus sweep did not find a comparable Arrowstreet-specific AI model disclosure. Its public GenAI evidence remains the current platform/security hiring surface described above. This is an important negative result: filing silence does not negate the platform signal, and the platform signal does not establish an investment model.
• Adjacent controls: Counterpoint’s filing describes gradient-boosted trees and neural networks ranking U.S. companies and ADRs across value, reversal, momentum, profitability, sentiment, and stability variables. QRAFT describes deep learning and Bayesian neural networks for relative four-week price appreciation. Goldman Sachs Asset Management says NLP and ML may extract information from textual or audio datasets for quantitative portfolios. These are registered-product disclosures from adjacent managers, not evidence about GMO, Acadian, or Arrowstreet. (Counterpoint prospectus; QRAFT prospectus; GSAM prospectus)

The full filing map, including Sparkline, Draco, Bluerock, FINQ, Sterling, FS, and Blackstone, is in the prospectus audit. The article keeps registered-product evidence separate from private hedge-fund evidence because a prospectus can describe a fund’s stated process without revealing how a related adviser operates across other vehicles.

G-Research: public engineering and runtime signals

The prior deep dive identified public roles at G-Research spanning AI engineering, ML/HPC architecture, principal AI/ML engineering, GenAI software, and quant software. The firm’s Core AI Engineer description covers on-premise open-model serving, centralized MCP, governed data/tool access, and secure sandboxes for autonomous agents. These are infrastructure and hiring signals. They are not proof that G-Research has deployed an investment agent or trained a finance language model.

Bridgewater: AIA Labs, PAT, and the Interrupt 2026 keynote

Bridgewater’s AIA Labs publicly describes a dedicated AI research and investment lab connected to Pure Alpha and an artificial investment associate. The firm’s PAT disclosure describes combining codified investment knowledge, large language models, agentic workflows, and software architecture to support exploratory investment research. Its public AI pages also describe AI tools in portfolio management, trading, portfolio risk management, and asset allocation, while retaining risk disclosures around error, security, and oversight.

A separate Bloomberg Television interview with Greg Jensen, published July 8, 2024, adds dated vehicle-level context. A timed transcript recovered on August 30 supplies a missing navigation layer: Jensen describes 25 people focused on using machine learning for investing (02:05–02:18), says off-the-shelf language models need to be combined with data models (02:26–02:31), and describes a goal of having machine learning generate investment ideas and algorithms from codified human intuition (02:35–03:08). Bloomberg’s public video description links Jensen to a roughly $2 billion vehicle using machine learning as the primary basis of decision-making. A Fortune account carrying Bloomberg reporting further reports an initial capital estimate, a smaller Pure Alpha test sleeve, possible inclusion of models from OpenAI, Anthropic, and Perplexity, and a potential shift toward data-science hiring. Those details are secondary reporting, and the vehicle’s legal identity and continuity with today’s AIA/PAT surfaces are not public in the reviewed sources. The recording therefore belongs in the dated launch and organizational-context layer, not as proof of a current model inventory, provider deployment, or strategy-level performance. See the timestamped capture note.

The AIA Labs page also publishes Learning to Replicate Expert Judgment in Financial Tasks, which Bridgewater describes as a training process for tuning models to investor information-filtering tasks. The same page links that work alongside PAT and other AIA research. This adds a public model-customization and task-evaluation signal to the lab record; it does not independently establish general investment autonomy or return attribution.

The current AIA Labs page also states that systems have been tested with real capital, manage billions of dollars, and generate alpha, and that AI tools may inform portfolio management, trading, risk management, and asset allocation after oversight. Those are Bridgewater’s own claims. The page does not provide audited returns, strategy-level P&L, benchmark, drawdown, fee treatment, or model-by-model attribution. The page’s AI research index adds current publications on forecasting, RLVR, and financial-task model tuning, including links to public research code.

Bridgewater’s July 28, 2026 AIA Labs video, featuring Greg Jensen, adds a separate evaluation and model-adaptation signal. Jensen describes evaluating model utility by task completion cost and quality, names an internal Investment Systemization benchmark with separate task categories such as forecasting and planning, and says the lab is adapting external and open models with reinforcement learning for Bridgewater’s purposes. He also gives a three-to-five-year expectation for material firm transformation. These are executive statements in a firm-hosted recording; the public route does not disclose the benchmark fixture, model versions, training data, reward design, permissions, production stage, or performance attribution. The caption capture preserves the timestamped evidence and its automatic-caption caveat.

An earlier title-blind Bloomberg Live interview adds organizational context rather than another system disclosure. The public caption capture connects Bridgewater’s discussion of AI to talent strategy, differentiated thinking, and the firm’s organizational future. It does not identify a model, agent, benchmark, data source, permission boundary, or investment authority, so it is retained as strategic context only. The capture note records the source and caption boundary.

Publicly named Bridgewater personnel include Greg Jensen, who is identified with AIA Labs and Alpha Engine/Pure Alpha leadership. Bridgewater’s current partnership page lists Blake Cecil as Deputy Chief Investment Officer, Alpha Engine and AIA Labs; Aaron Linsky as Head of Engineering, Alpha Engine; Nina Lozinski as Head of AIA; and Oliver Simon as Head of AI & ML Investment Strategy. Nina’s separate July 2026 profile still uses a Co-Head of AI & ML Investment Strategy title, so titles are kept source-specific rather than merged. Jasjeet Sekhon’s current profile places him at Google DeepMind after his former Bridgewater role. The reviewed material is firm-reported and temporally mixed. It does not independently establish the current team size, reporting lines, performance attribution, or production authority of any specific AIA system.

The title-blind personnel pass also recovered a 2019 MIT IDSS seminar whose description identifies Sekhon as Head of Causal Inference at Bridgewater at the time. The talk covers causal-effect estimation, heterogeneous treatment effects, high-dimensional covariates, and the distinction between predictive ML and causal inference (09:20–09:35; 20:42–21:01; 28:03–28:14). This is date-scoped academic-lineage and methods evidence, not evidence of a Bridgewater trading model, current employment, model ownership, portfolio use, permissions, or performance. The capture note preserves the recording and caption boundaries.

An earlier 2016 CODE plenary recording, titled for Sekhon and Johan Ugander, adds academic discussion of heterogeneous treatment effects, outcome-versus-effect modeling, adaptive experimentation, and validation (01:51–11:31; 35:57–40:33; 47:05–48:42). The recovered recording is an academic session and does not support assigning the methods to Bridgewater or mapping every later segment to a specific speaker. It therefore remains methodology and date-scoped lineage context—not firm deployment, a current role, model ownership, or performance evidence. See the capture note.

An adjacent 2018 MIT CODE plenary recording, titled for Eva Ascarza and Jas Sekhon, adds applied causal-ML material on treatment-effect estimation, causal random forests, neural networks, transfer learning, validation, and interpretability (02:10–04:48; 23:21–29:14; 37:35–40:33; 67:01–68:37). The recovered audio does not support assigning the methods to Bridgewater or reliably mapping every segment to one speaker, so this remains academic methodology and possible date-scoped lineage context—not firm deployment, a current role, model ownership, or performance evidence. See the capture note.

A separate Sackler Big Data Colloquium recording adds a 2016 academic route on large-scale randomized experiments, heterogeneous treatment effects, researcher degrees of freedom, blocking, adaptive experimentation, and validation (00:00–14:40; 39:03–48:42). The recording does not establish Bridgewater employment at the time or use of the methods by the firm; it remains methodology and date-scoped lineage context only. See the capture note.

The Interrupt 2026 recording library is a newly added conference source surface. It contains 23 on-demand sessions from LangChain’s May 2026 San Francisco conference, including Bridgewater | Building Pat, the AI Pocket Analyst Tool. Bridgewater’s official session page confirms that the presentation was recorded on May 19, 2026 and identifies Brendan McManus, Michael Ran, and Santi Weight as the presenters. The recording is materially more specific than the short PAT description: McManus identifies himself as applied-AI team lead, Ran as investor lead, and Weight as technical lead. A closing slide also names Jon Schory, Paul Batterman, Nilesh Patel, Zain Khalid, Aqib Dar, James Tyler, and Hayley Williamson in a special-thanks list; the slide does not assign those people roles or ownership, so they are not promoted to current AI personnel.

The recording describes PAT as an investigation-oriented research agent, explicitly separate from how Bridgewater trades. The presenters say it was deployed internally several months before the talk and was being used daily by hundreds of investors. They describe two connected layers: a chat agent that gathers context, searches data, asks clarifying questions, and creates a plan; and a coding agent that turns the plan into Python/Pandas analysis. The chat layer is implemented in LangGraph for persistent state, cancellation, and continuation. Its tools include unstructured search and time-series search. The coding layer treats the plan as a natural-language Python project: tasks map to data-frame-producing functions, code is generated in parallel by sub-agents, a dependency graph controls execution, static analysis and validation agents run before results are accepted, and cached Python execution avoids reloading or recomputing intermediates. These are public implementation statements from the presenters, not an independent code audit.

The disclosed data and control boundary is unusually concrete. Each user’s PAT is described as having a distinct context and tool set because Bridgewater investors have different information entitlements. The system searches a near-real-time database of millions of documents, including broker research, earnings transcripts, and internal emails, with thousands of new items arriving daily; it also searches a time-series database containing tens of millions of internally modeled series. The presenters describe conventional retrieval and reranking combined with human-like inspection of frequency, currency, and whether values fit prior expectations. They report an improvement from roughly 50% to 90% accuracy for that search task, but do not define the benchmark, denominator, test set, or independent evaluator. The system then asks clarifying questions, builds a detailed data-frame/schema plan, executes Python under oversight, checks computed values and visualizations, and returns an interactive report using Bridgewater’s internal charting and text-deck tools.

A YellowDog case study dated March 29, 2026 adds a separate infrastructure layer to Bridgewater’s public record. YellowDog says Bridgewater’s quant teams use multi-region AWS Spot capacity and workload orchestration to run large ML workloads, with production runs regularly reaching more than 35,000 nodes. The case study reports a vendor-estimated research-velocity increase of more than 7× and estimated Spot-cost savings versus on-demand pricing. These figures are customer-story claims, not an independent audit; the page does not identify model weights, training data, PAT permissions, portfolio decisions, or strategy-level performance. The safe inference is a named vendor relationship and a disclosed high-scale compute workflow, not a unified account of every Bridgewater AI system.

The same vendor trail exposes a separate Fulcrum Asset Management case study. YellowDog identifies Fulcrum as an independent investment manager and names Jago Westmacott as Chief Technical Officer. The page describes on-premise and multi-cloud workload execution, private-cloud separation, data-sovereignty controls, automatic recovery from pre-empted instances, and compute-intensive financial analyses; it reports a vendor/customer claim that workload runtime was halved. This is an asset-manager infrastructure disclosure, not evidence of a hedge-fund AI or GenAI program, model ownership, or investment results.

YellowDog also publishes an anonymized high-throughput scheduling paper describing a global hedge-fund quant workload with millions of tasks and tens of thousands of nodes. Because the firm is not named and the PDF is product marketing, it is retained as an infrastructure pattern rather than assigned to any manager in this map.

The learning loop is also public. Agents review completed conversations for behavioral mistakes, context gaps, and user steering; a user can explicitly trigger a “Teach” flow; the system creates a benchmark expected to fail, changes context repositories or the harness until it passes, checks that the wider suite does not regress, and sends a pull request for review. The technical presenter reports that, on the team’s test suite, two agents produced exactly the same code about 95% of the time and characterizes the result as a deterministic coding agent. That percentage and the reported 4x code-generation speed comparison with Claude Code are firm-presented test claims, not independently reproduced results; the talk does not disclose the suite, prompts, model versions, sample sizes, or production error rates.

This keynote changes the public evidence category for Bridgewater from “LLMs and agentic workflows are described” to a more detailed firm-reported operating account: entitlement-aware retrieval, specialized agents, deterministic code generation, executable validation, caching, interactive research outputs, and benchmark-driven context updates. It still does not establish PAT as a trading model, identify the full model/provider inventory, disclose prompts or permissions in machine-readable form, or provide an audited link from PAT to returns, positions, execution, or capital allocation. The broader AIA Labs mission should therefore remain separate from the narrower PAT deployment claim.

The personnel lineage has a separate academic route. The 2018 CODE Plenary recording presents Jasjeet Sekhon on transfer learning for estimating causal effects with neural networks, covering data efficiency, low signal-to-noise, shared and experiment-specific architecture, and out-of-sample transfer across experiments (23:21–29:26; 37:37–43:47; 44:47–47:38). Bridgewater’s current public profile identifies Sekhon as a former Chief Scientist and Head of AI, but this academic recording is not a Bridgewater event and does not establish that the method was used at the firm. It is useful for academic-lineage mapping only; it does not disclose Bridgewater data, models, permissions, deployment, or performance. See the capture note.

An adjacent THOR Behind the Ticker episode with Bob Elliott adds a separate former-Bridgewater practitioner route. The publisher describes Elliott as a former Bridgewater practitioner and co-founder of Unlimited Funds, and describes Unlimited’s replication product as using a proprietary Bayesian machine-learning model that treats inferred manager positioning as path-dependent rather than relying only on long rolling return regressions. The same page frames the products as replication technology and labels performance information as hypothetical or back-tested. This is a first-party product/media description, not evidence about Bridgewater’s current systems, live permissions, model weights, training data, or realized performance.

An additional title-blind Excess Returns interview with Bob Elliott, dated June 9, 2023 in the Apple Podcasts listing, supplies a separate practitioner description of the same general research problem. Elliott describes combining experience building proprietary strategies with statistical-learning methods (21:21–22:58), inferring likely positions from returns and market outcomes (24:15–25:04), building components by sub-strategy before combining them (25:49–26:19), and expressing views across roughly 60 liquid markets (27:09–28:08). This makes the architecture more specific than the title alone, but it remains dated, speaker-reported, and separate from Bridgewater’s current systems. It does not establish model weights, training data, live permissions, or independent performance. See the capture note.

The same title-blind route recovered Pitch The PM EP035, published June 10, 2026, with Eric Moster. The episode identifies Moster as CEO of Portrait Analytics/Portrait Research and describes his prior Citadel, Millennium, and Surveyor Capital roles. Portrait’s current product site describes connected idea discovery, cited deep research, and thesis monitoring. Its Intelligence page describes codifying investor frameworks, testing them against legacy ideas, evaluating a company universe, and producing framework-specific research outputs; its Easel guide describes a conversational agent that plans and carries out multi-step investigations across documents, the web, market data, and user materials.

The recording makes the operating design more explicit. Moster describes a data-grounded chat agent, source-linked answers, long-form research, and monitoring that filters for relevance, incrementality, and historical context (05:34–17:07). He says the team built domain-oriented datasets and evaluation sets to refine the system as models change. Portrait Intelligence is described as converting an investor’s framework into criteria, applying it across a broad universe, using agents for fundamental work, and returning framework-fit research candidates. A separate Game Tape workflow uses historical trades or positions with contemporaneous context to produce postmortems and refine future framework-based idea generation. These are speaker and vendor descriptions, not an independent benchmark or proof of investment authority; the capture note preserves the episode timestamps and ASR boundary.

The episode also gives a useful negative boundary: Moster says the team discontinued an open-ended experimental agent that did not solve a concrete problem and dropped a structured-financial-data/model-building initiative to focus on fundamental research. That is a dated company product-scope decision, not a general conclusion about agents or financial-model automation. Third Bridge’s July 2025 collaboration announcement separately describes integrating its expert-interview library into Portrait’s LLM-supported research workflow. The 2025 announcement names David Plon as CEO and co-founder, while the June 2026 episode identifies Moster as CEO; the sources establish a dated title discrepancy but not the transition date or reporting structure. No reviewed source names Portrait’s foundation models, training corpus, customer identities, permissions, or AI-attributed returns. See the capture note.

The title-blind route also recovered an Odds on Open interview with Doug Garber, with Spotify and Pitch The PM distribution pages. The 2025 recording identifies Garber as a former Citadel analyst and Millennium portfolio manager, and describes domain research, platform systems, factor and risk-model overlays, portfolio constraints, analyst training, and PM–analyst feedback loops (00:25–03:16; 09:55–10:30; 19:13–30:25). This is historical, speaker-reported operating context that helps identify what the title-blind route can recover; it does not establish current firm policy, AI/LLM use, model ownership, data rights, permissions, or performance attribution. See the capture note.

Jane Street: machine-learning trading infrastructure and internal GenAI tools

Jane Street’s machine-learning page describes deep-learning models that drive trading strategies, infrastructure for training and inference, researchers and traders studying models in production, and work spanning LLMs, reinforcement learning, computer vision, and classical ML. The same current page displays firm-reported scale markers of tens of thousands of high-end GPUs, more than one exabyte of storage, and $400 billion in daily filled dollars; it does not define the measurement date, the portion allocated to ML, or an independently audited denominator. Its quantitative-research page describes researchers analyzing large datasets, building and testing models, creating trading strategies, and writing the implementation code. The page-level capture is preserved in the Jane Street source note.

Jane Street’s current ML Research Engineer role adds a distinct platform layer: training and inference infrastructure, neural networks, tree ensembles, production-transition experience, workflow APIs, and reproducible research code. Its Visiting Researcher program separately states that Jane Street uses both third-party models and custom models trained on its own infrastructure and invites academic collaboration across ML, programmable hardware, systems, and applied mathematics. These pages clarify the public organization and model categories, but do not identify providers, model weights, data rights, evaluation splits, production endpoints, per-model ownership, permissions, order authority, or performance. See the capture note.

A title-blind recovery adds a dated engineering baseline: Yaron Minsky’s 2012 Caml Trading talk describes OCaml as the first tool used across monitoring, trading, research, and other systems, alongside correctness concerns, critical-code review before production, a shared repository, integrated compilation, and unit tests (16:38–17:25; 17:44–19:28; 25:38–26:15; 54:54–55:44). The speaker also links the platform transition to trading-related research and strategy development (67:00–67:50). This is historical software-factory and control evidence, not a current AI or GenAI deployment record; it does not establish model permissions, agent-written production code, trading authority, or performance. See the capture note.

The regional check adds a Hong Kong-specific public surface. Jane Street’s current Hong Kong quantitative-research role describes model building, time-series analysis, feature engineering, distributed-training debugging, studying model behavior in production, petabyte-scale data, and large CPU/GPU clusters. The current Hong Kong quantitative-trader and trader-internship pages place trading, research, and machine learning in the same regional department and keep the ML language tied to statistical modeling, deep learning, simulated strategies, and trading education. A separate current Hong Kong Machine Learning Performance Engineer role adds training and inference performance, CUDA/Triton/CUTLASS/cuDNN/cuBLAS, GPU networking, and distributed-training algorithms to the regional infrastructure signal. The footer identifies Jane Street Hong Kong Limited as regulated by the Hong Kong SFC with CE No. BAL548. This strengthens the Hong Kong legal-entity and hiring-intent lane; it does not identify a GenAI model, data license, agent permission map, production endpoint, autonomous order authority, or AI-attributed performance.

The GenAI-specific public signal is separate from the trading-model disclosure. Jane Street’s AI Assistants team article describes internal generative-AI systems, RAG applications, and plugins built at an AI hackathon. The public pages do not identify a firm-wide GenAI executive, foundation-model inventory, or a direct link between those internal assistants and live trading decisions. The reviewed Jane Street sources are therefore evidence of a machine-learning trading organization plus internal GenAI tooling, with the boundary between the two left undisclosed.

The current From Code to Woodcraft article adds custom models and tools for employee LLM use. Jane Street’s AI Engineering at Jane Street talk describes code-generation models, editor integration, workspace context, and build/typecheck/test evaluation. Public engineering artifacts include ocaml-torch and the GTC 2026 repository, which expose ML runtime and performance-engineering work. These artifacts support internal ML/GenAI engineering activity; they do not reveal production trading-model weights, features, live metrics, or a link from AI Assistants to trading decisions.

The first-party Signals & Threads episode with In Young Cho, published March 12, 2025, adds a direct research-process account. Host Ron Minsky introduces Cho as a long-serving Jane Street researcher and trader who recently moved into the research group and was leading work in machine learning (00:12–00:30). The discussion then walks through data collection and correction, predictors/responders, moving from simple regressions to more expressive methods as interactions grow, and the productionization boundary between exploratory notebooks or spreadsheets and production systems written in OCaml (14:41–20:36). This is first-party practitioner evidence about research workflow and data quality, not a model inventory, current title registry, training-corpus disclosure, agent-permission map, autonomous order authority, or performance record. The local transcript is automatic and retained for timestamp navigation and paraphrase; the repository source note records its provenance.

Jane Street’s first-party “Get to Know Us” video featuring JP, uploaded July 18, 2024, adds a Hong Kong equities-trader route whose description says JP has been full-time since 2017. JP describes small modeling improvements, desk-level research conversations, checking model behavior when data is out of distribution, expanding CPU/GPU access, and internal libraries for sharing modeling techniques across regions and desks (00:09–02:02). The profile is useful because it ties the ML infrastructure story to a regional trading-and-research role without naming a model or provider. JP’s surname is not established; the video does not prove firmwide practice, agent permissions, production model ownership, or performance.

The same RSS archive also contains Why ML Needs a New Programming Language with Chris Lattner, published September 3, 2025. Minsky introduces Lattner as an external guest; Lattner discusses MLIR, heterogeneous accelerators, model-serving infrastructure, and Modular’s Mojo/Max platform (00:12–00:20; 12:46–13:37; 19:30–21:01). This is useful title-blind AI-infrastructure context and a discovery route for portability and compiler questions, but it does not establish Jane Street usage, procurement, partnership, or deployment. The repository source note keeps that boundary explicit.

The title-blind pass also recovered a 2017 SF Scala interview with Yaron Minsky. The recording links Jane Street through its description and Minsky describes work on the firm’s core development platform, the quantitative-research group, compiler tooling, production infrastructure, and trading systems (00:46–01:39; 06:08–07:15). Later sections discuss shared libraries, incremental computation, and compiler development (17:00–18:10; 24:05–24:20; 31:54–32:18). This is historical engineering and personnel evidence, not an AI disclosure or a performance claim; it does not establish current titles, model inventory, data rights, permissions, autonomous order authority, or returns. See the capture note.

A separate first-party Jane Street profile for Alok links the unnamed speaker to New York’s Trading, Research, and Machine Learning department, identifies him as a 2018 intern who joined full-time in 2020, and describes research work spanning dataset construction, pattern search, model training, strategy development, reusable firm tools, and theoretical problem solving. The profile also says his prior interests included quantum information theory and machine learning for clinical applications. This is a recruiting/culture artifact and intentionally does not infer a surname, current ownership, proprietary model, GenAI system, data rights, permissions, or performance. See the capture note.

A February 2026 Jane Street engineering-blog post by Edwin Morris provides a separate employee-level workflow signal. Morris identifies himself as a designer on the options desk and describes using Claude to prototype changes in Jane Street’s internal JSQL interface, iterate against a running codebase, and obtain user feedback. He says the prototype is treated as a disposable living proposal before an engineer takes ownership of production code. This is a firm-hosted personal account of product-development practice, not a firmwide policy or evidence that Claude receives sensitive market data; it does not establish model permissions, deployment scope, trading integration, or performance. It is useful precisely because it separates rapid AI-assisted prototyping from production ownership.

Jane Street’s Tech Talk with Arjun Guha adds a title-blind technical-media route with a full HTML transcript and a linked recording. Guha is identified as a Northeastern computer-science professor, and the talk concerns LLM behavior on programming languages, low-resource languages such as OCaml, developer adaptation, and evaluation design. One useful control is separating model limitations from missing libraries, environment, or tool configuration when evaluating code-model output. This is external academic content hosted by Jane Street; it does not establish Jane Street authorship, a trading-model use case, model permissions, or performance. See the capture note.

The title-blind pass also recovered Jane Street’s 3Blue1Brown roundtable, published August 21, 2026, and its first-party companion article. The recording describes a real-time auction for compute-cluster time (03:37–04:15), references petabytes of market data (11:58–12:06), and has a participant describe deep learning as the “single largest driver” of the systematic trading the firm has (09:53–10:03). That last statement is a firm-controlled self-description, not a comparative assessment or independently measured attribution. The 3Blue1Brown partner page adds partner-side descriptions of ML models, GPU resources, historical events, and automated trading; those are marketing claims and are kept separate from the first-party recording. See the capture note.

A separate title-blind Predicting Alpha interview with Craig Newbold, published in 2021, identifies him as a former Jane Street practitioner. Newbold says he tried modern deep-learning methods directly on trading problems (05:20–05:39) and discusses how domain expertise might transfer into trading research (05:48–06:22). This is historical, third-party former-personnel evidence; it does not establish a current Jane Street system, production deployment, permissions, or performance. See the capture note.

The Bug Bash 2026 talk by Ron Minsky is a separate first-party engineering capture. Antithesis dates the page May 13, 2026, its agenda identifies Minsky as Jane Street’s Co-head of Technology, and the page embeds the recording. In the recording, Minsky describes an AI-assistant team established for software engineers and says agent uptake later increased (03:12–04:45). He then draws a separate boundary around ML architectures and models whose outputs guide automated trading systems (04:47–05:22). The talk places type systems, tests, code review, deterministic simulation, and formal verification around agent-assisted software development. This supports a public engineering-control and organizational account; it does not identify model providers, training data, permissions, trading integration, or performance. The transcript is an automatic caption layer and the adoption/scale statements are speaker-reported. See the timestamped capture note.

Point72 / Cubist: training-language signals alongside a systematic research organization

Point72 and Cubist expose several different layers. Cubist’s official firm page describes more than 600 team members as of January 1, 2026, systematic research across liquid asset classes, machine-learning-based return prediction, data-science work, and quantitative development. The Market Intelligence page connects alternative-data sourcing to investment professionals and Compliance, then describes AI/ML models and tools applied to large data collections. These sources establish a systematic research and data-product environment; they do not establish GenAI-agent deployment.

The current Cubist Machine Learning Researcher role names deep learning and natural-language processing as relevant experience and describes the full research loop: data ingestion, analysis, methodology selection, implementation, testing, and performance evaluation. Additional Point72 roles use AI-system language: an NLP/AI Engineer role names LLMs, agents, RAG, open-source fine-tuning, post-training, and benchmarking; a Fundamental Equities AI Engineer role describes AI products moving from prototype to production within compliance-approved boundaries. August 11 rechecks add a GenAI Security Engineer role covering security controls for agentic and human-in-loop GenAI systems, MCP/tool-calling protection, prompt-injection and data-exfiltration threat models, and audit evidence; an AI Solutions Architect role in Market Intelligence that maps investment-team research workflows into AI use cases; and a Cubist Machine Learning Engineer role mentioning synthetic data, MCP agents, and a production-support AI agent. An August 14 recheck adds a regional AI Data Scientist role in Hong Kong/Singapore, covering AI-powered data products, in-house AI/ML model training, fine-tuning, evaluation, deployment, and production governance for systematic data. These are current role requirements. They do not prove that a named model was fine-tuned, that a role was filled, or that an agent can approve investment decisions.

Point72’s leadership page identifies Geoffrey Lauprete as Head of Cubist Systematic Strategies and Ilya Gaysinskiy as Chief Technology Officer. Point72’s Market Intelligence team also places data scientists and engineers alongside investment professionals and Compliance. The public record does not identify a firm-wide GenAI owner, model inventory, evaluation suite, or production user count. A separate media trail around the AI-focused Turion strategy is evidence about investment exposure and fund structure, not evidence about internal AI systems.

Point72’s July 27, 2026 San Francisco office announcement adds a current first-party strategy and regional-media route. It says the San Francisco-based Turion team attended the opening and describes Turion as an AI-focused public-equities fund. The announcement also names Harry Schwefel and Michael Sullivan among the participants. This establishes Point72’s public description of a named strategy and office presence; it does not establish Turion’s legal structure, assets, portfolio, model stack, datasets, AI staffing, deployment, permissions, or performance, and it should not be conflated with Point72’s internal AI systems. See the capture note.

The current Cubist Internal Alpha Capture quantitative-research role is a separate signal from the data-product role. It describes AI-driven equity-trading signals built from proprietary data and a research loop spanning ideation, method selection, implementation, evaluation, and eventual application; preferred methods include sequence models, graph neural networks, reinforcement learning, and LLMs. The current Hong Kong/Singapore AI Data Scientist role instead sits in Cubist Data Services and names in-house model training, fine-tuning, evaluation, deployment, LLM-based extraction/search/automation, and production governance. This separation suggests distinct public hiring lanes—signal research and regional data-product operations—but does not establish shared models, filled roles, live endpoints, or decision authority. See the capture note.

The title-blind Bloomberg Odd Lots episode with Joe Peta, published September 5, 2024, adds a personnel and operating-process route. Its description identifies Peta as a former Head of Performance Analytics at Point72, and the automatic captions record a discussion of portfolio-manager evaluation, skill attributes, and the noise in realized results (05:18–05:39, 06:58–07:19, 09:12–09:33). This is third-party former-personnel evidence about an evaluation concept, not a Point72 policy or disclosure of AI, GenAI, model ownership, deployment, portfolio authority, or performance. See the capture note.

A title-blind former-personnel route surfaced Thomas Li’s August 2026 Invest with AI episode. Podscan identifies Li as a former Point72 TMT analyst and provides timestamps on data factories, MCP reliability, knowledge graphs, post-training, finance evaluation, and the boundary between quants and fundamental analysts. It is current Daloopa/vendor and guest commentary, not evidence of Point72’s current internal systems.

Point72’s first-party Perspectives archive expands the media graph beyond job pages. It links a June 2026 Cubist AI-chess hackathon describing model training, autonomous workflows, and agentic evaluation in a recruiting/engineering exercise; a March 2025 global technology hackathon; and a 2025 CTO profile describing AI exploration and technology expansion. A April 2026 Sajid Ahmed profile identifies him as Head of India and Head of APAC Technology and mentions AI-related teams in Bengaluru. These are firm-controlled technology, recruiting, and organization signals; they do not disclose model weights, portfolio permissions, or investment performance.

Schonfeld: investment-team enablement with explicit pilot controls

Schonfeld’s May 2026 FE AI Lab disclosure is a direct first-party account of an investment-workflow program. It describes portfolio managers and analysts learning to automate earnings preparation, idea generation, document analysis, inbox triage, and spreadsheet workflows, with access to Schonfeld’s internal systems. The stated workflow runs from research intake toward portfolio construction, but the source does not say that an AI system determines positions or sends orders.

The same page names SchonAI as a proprietary platform and describes model partnerships with Anthropic and OpenAI. It says new tools pass through a structured pilot process and are evaluated for downside before rollout. A March 2026 Schonfeld Q&A says SchonAI is used by a large majority of employees, with portfolio managers described as the largest user group. A current Senior Software Engineer role describes background workflows over market news, security data, SEC filings, email, and other sources, with MCP gateway work and an agent-evaluation framework. These are firm and recruiting disclosures, not an independent adoption or performance audit.

Two current Fundamental Equity COO roles sharpen the operating-layer signal. The Software Engineer - Fundamental Equities and Quantitative Developer - Fundamental Equities postings place LLM APIs, MCP servers, agentic workflows, multi-model architecture, dashboards, APIs, data pipelines, and a shared AI plugin and skills library inside a business-embedded team serving Fundamental Equity investment professionals. These are current hiring-intent signals. They do not show filled roles, model routing, production authority, adoption metrics, or investment performance.

Public materials do not disclose the internal model-routing design, tool permissions, evaluation datasets, failure rates, or whether SchonGPT remains a separate product name or a component of SchonAI. A vendor interview reports an earlier OpenAI-based prototype, later Kubernetes deployment, approximately 150 tools, and approximately 30 bot configurations; those figures remain vendor-reported and are not treated as independently verified.

Public personnel evidence previously reviewed in the worker-layer ledger connects a Generative AI Engineer to LLMs, RAG, Bedrock, Azure OpenAI, Claude, LangChain, and internal SchonGPT work. Schonfeld’s public company page also identifies Mansi Kapadia as Head of AI, while its official leadership page lists Steve Harmon as Interim CTO overseeing infrastructure, data, cybersecurity, and AI initiatives. The current public record does not publish the reporting relationship between those roles or a complete platform-team roster.

The expanded Schonfeld capture note consolidates these firm, personnel, podcast, and hiring routes with their evidence boundaries here.

G-Research: a governed runtime and production LLM controls

G-Research’s Core AI Engineer role describes a centralized platform for on-premise open-model inference, model serving, developer tooling, centralized MCP servers, and secure sandboxes for autonomous agents. The role also names Kubernetes operations, observability, capacity planning, self-service APIs, and controlled environments. The role description provides detailed infrastructure evidence, but it remains a role description and does not identify a live investment agent.

The current G-Research role family makes that public surface more granular. The Applied AI Engineer role describes an AI layer serving quantitative research, engineering, risk, and operations, with production LLM systems, agents, RAG, MCP, evaluation, observability, LoRA, and DPO in scope. The Core AI Engineer role describes centralized on-prem open-model inference, model serving, developer tooling, MCP servers, and secure agent sandboxes. Separate ML/HPC engineering and ML research roles cover distributed training, inference optimization, accelerator evaluation, and quantitative research into deep learning, reinforcement learning, NLP, Bayesian methods, and approximate inference for market prediction. These are current first-party role descriptions; they do not identify the hires, a complete model/provider inventory, data rights, live users, investment-agent permissions, or performance attribution. See the capture note.

G-Research’s production code-review article adds implementation detail. The LLM is treated as untrusted; structured output is validated against an authoritative rules index; deterministic severity fields are computed outside the model; truncation and malformed responses receive bounded recovery; and the system records provider, model, cost, and quality metrics. The tool runs in CI/CD and posts non-blocking pull-request comments. These controls show how a quant firm can operationalize an LLM without granting it final authority. They do not establish use in portfolio construction, signal generation, or trading.

The firm-controlled code trail now includes Robocop, which G-Research’s official open-source page lists as an automated LLM code-review wrapper. The repository README says it sends diffs to GPT-5 for code review and is intended as the basis of a CI pipeline. This is public code-ownership evidence for software-engineering review automation; it is not evidence of investment research, trading deployment, or model authority.

G-Research’s current vacancies page lists Core AI, Applied AI, machine-learning, natural-language-processing, workflow, observability, and security roles. Its current NLP Researcher role names domain-adaptive pretraining, instruction fine-tuning, quantization, and LoRA for large text corpora and possible market-predictive features. An Applied AI role names production LLM systems, RAG, multi-agent orchestration, evaluation, observability, MCP, and fine-tuning. These are material hiring signals, but they do not identify a deployed finance LLM or live trading loop.

The title-blind video pass adds three firm-uploaded routes. In G-Research at NeurIPS 2022, the firm connects market prediction, conference-based ML learning, quantitative-research and engineering development, and recruiting; the accompanying NeurIPS review archive names researchers and the papers they selected. The Alex Davies mathematics symposium recording and firm interview add an academic/talent route through machine-learning-assisted mathematical discovery and a Cambridge PhD under Zoubin Ghahramani. The Michael I. Jordan lecture is a dated G-Research Lecture Series surface covering uncertainty, data-sharing constraints, and finance’s large-data setting. These recordings strengthen the firm’s public conference, academic-network, and recruiting map; they do not establish current collaboration, model ownership, investment use, or performance.

The firm-controlled GR-OSS OUT Podcast is a further title-blind media route. Its JSON feed and RSS feed list 26 episodes, public MP3 enclosures, and linked YouTube videos. Four recovered videos add bounded engineering evidence: the Databricks episode discusses Unity Catalog AI’s agentic tool-calling framework, hosted foundation models, and ML/GenAI training cycles (01:08–01:16; 06:48–06:53); Kubernetes and AI/ML Workloads covers batch scheduling, Armada, and Kubernetes work for AI workloads (00:00–00:16; 01:12–01:32); and two newer episodes discuss responsible AI-assisted coding and automation/security boundaries (episode 25, episode 26). The captioned routes are automatic-transcript navigation aids, not human-verified quotations. This is firm-controlled engineering and open-source evidence; it does not establish a finance model, investment deployment, portfolio permission, or performance. The capture note records the episode-level evidence and boundaries.

An additional title-blind Data Flowcast episode names Christos Bisias as an Open Source Software Engineer, Apache Airflow, at G-Research. The publisher describes Airflow use for large-scale data transformations and upstream contributions to scheduler throughput and OpenTelemetry support. Its key-takeaway map connects G-Research’s machine-learning and big-data market-prediction context to tracing, custom spans, a CI-enforced YAML metrics registry, and earlier concurrency limits in scheduler queries (01:20; 04:30–12:10). The Airflow Summit speaker page independently corroborates Bisias’s G-Research Open Source role and Airflow contribution. This adds a named data-platform and observability contact; it does not identify a finance model, proprietary corpus, agent permissions, investment authority, or AI-attributed return. See the timestamped source note.

The public record does not disclose a complete AI leadership roster, a named finance-language model, model weights, proprietary training data, or the boundary between the central platform and investment research teams.

Title-blind executive media recovered from Generating Alpha

The podcast practitioner article and its linked source ledger add four public routes that were not present in the media ledger because the episode titles do not lead with AI. Pete Muller of PDT discusses overfitting, model trust or modification, and the place of large models beside established quantitative practice; the relevant excerpts are from public automatic captions and do not identify a PDT GenAI system. A Dmitry Balyasny episode adds leadership context and discussion of collaboration across research domains, while separate Balyasny/OpenAI sources remain the basis for specific Applied AI and agent claims. Ryan Tolkin of Schonfeld discusses AI efforts and the importance of identifying material data, and Jeff Yass of Susquehanna provides adjacent context on prediction markets, manipulation risk, and the limits of quantification. None of these four episodes discloses a named model, training corpus, permission map, live trading authority, or AI-attributed performance.

A new publisher-hub pass also resolved the iframe links in the Systematic Traders weekly roundup, published August 2, 2026. Standpoint, Quantedge, and Nick Baltas reconcile to existing records; the Jane Street item is third-party interview preparation rather than a firm disclosure. The net-new item is Susquehanna’s official “Finding Signal in a World of Noise”, published July 27, 2026 and produced in partnership with the International Congress of Mathematicians. First-name-only contributors describe markets as a low-signal, high-noise setting, a hypothesis-to-test-to-validation loop, AI-assisted pattern exploration, verifiable-reward environments, and firm computing resources for scaling-law study. Susquehanna’s separate NeurIPS page provides independent role context for some contributors, but surnames are not inferred for the video’s first-name-only participants. The captions are automatic and the claims remain firm-produced research-culture and technical-positioning evidence; no model, corpus, benchmark, permission system, deployment, or performance is disclosed. See the capture note.

AQR: public method and validation constraints, limited current GenAI disclosure

AQR’s Can Machines “Learn” Finance? paper remains a public statement of the firm’s methodological concerns: finance has low signal-to-noise, data-mining risk matters, and economic theory and human expertise remain part of the process. AQR’s 2024 Can Machines Time Markets? adds a contrast between large-data prediction settings and market-timing settings with fewer independent observations. These sources are research and control evidence, not current GenAI deployment documentation.

The public record identifies Bryan Kelly with machine-learning leadership and has affiliated point-in-time language-model research in the local source ledger. It does not establish that the affiliated training work is deployed inside AQR portfolios. The reviewed AQR surfaces also do not identify a current GenAI executive, internal agent platform, model vendor inventory, or production user base.

A title-blind Hoover Institution interview with Cliff Asness, published October 23, 2025, adds a founder-level AQR route. The episode introduces Asness as an AQR co-founder and describes his University of Chicago finance PhD under Eugene Fama; AQR’s current leadership page provides a first-party title and biography cross-check. Asness says machine learning had moved into AQR’s existing funds, names Bryan Kelly, Andrea Frazzini, and Laura Serban in that discussion, and identifies natural-language processing and text classification as an area of application (52:32–56:29). A public Podbean enclosure was also recovered and rechecked locally on August 30, 2026; it corroborates the timestamped discussion but remains an automatic-ASR navigation layer. He also connects the firm’s method to data quality, model complexity, and overfitting controls. This is a dated executive account and academic-lineage record, not a current model inventory or deployment audit; it does not establish model ownership, training data, vendors, evaluation results, permissions, or AI-attributed performance. See the capture note.

HSBC Asset Management: explicit boundary between AI-assisted research and portfolio decisions

HSBC Asset Management’s first-party article “AI won’t pick your stocks”, published August 18, 2026 according to the indexed publisher record, identifies Daniela Hamoui as global head of the quant product specialist team and describes the Active Quant Equity team’s current boundary. The article says AI supports internal research and productivity, back-tests, strategy simulations, and signal refinement, while the models making buy and sell decisions are kept outside direct AI decisioning. It also describes experimenting with business-news text to create topic-intensity and tone measures, including a recession-narrative index, with signals checked against existing indicators and constrained or dropped when unstable.

This is a useful asset-manager control case because it makes the decision boundary explicit: text-derived features and simulation can enter the research process without the publisher claiming that an LLM selects stocks. The source is a first-party strategy statement, not an independent audit. It does not disclose model identity, vendors, training data or rights, evaluation design, production scale, portfolio authority, or performance. The source ledger records the evidence boundary and title-blind query extensions. HSBC should remain separate from the hedge-fund cases in this article.

A recovered HSBC Markets and Securities Services podcast adds named execution-side context that is distinct from the asset-management article. Its introduction names Ed Dugen for next-generation equity-execution algorithms, Chris Ul as global head of equity-execution quants, and Paris Panesi as head of systematic trading strategies for spot FX and commodities, and describes Panesi’s artificial-intelligence doctorate. In the discussion, participants describe AI/ML in market making and equity execution, including pricing, order handling, nonlinear interactions, and execution improvement (12:31–15:30); they also discuss NLP research and safeguards, domain expertise, constraints, explainability, and monitoring (15:55–16:18, 21:59–25:49). This is a 2025 bank-hosted discussion with automatic captions, not evidence about HSBC Asset Management’s portfolio models or a disclosed bank model inventory. The capture note preserves those entity and evidence boundaries.

The title-blind Research Uncut debut, uploaded August 20, 2026, adds a separate HSBC Global Investment Research route. Alex Andronov, identified in the video description as Global Head of Business Development, discusses research becoming machine-readable data, LLM discoverability, research evaluation, transcript processing, and the line between scalable document handling and human analyst/corporate access (00:58–02:54; 05:04–10:30; 12:49–18:59). Substantive Research’s June 2026 event page independently confirms Andronov’s role and exposes a recurring event/archive route with buy-side and vendor personnel leads, including Carrie Anton, Director of Market Data at Jain Global. These are sell-side and event-network signals; they do not establish HSBC Asset Management or Jain Global model inventory, deployment, permissions, or performance. The capture note keeps the entity and agenda/remarks boundaries explicit.

A title-blind Rational Reminder interview with Andrew Chen adds a reproducibility and model-risk control. Chen’s Open Source Asset Pricing project publishes data and code for standardized predictor construction, while the OpenSourceAP repository exposes separate signal, portfolio, and shipping code. His public research page now links July 2026 papers on high-throughput/data-mined asset pricing and “What Useful Alphas?”, alongside 2025 work on peer-reviewed theory. The interview covers out-of-sample decay, multiple testing, point-in-time data, transaction costs, and the distinction between a statistical predictor and an investable strategy (07:08–09:16; 27:10–34:09; 43:54–50:11). This is academic/control evidence, not a hedge-fund deployment claim; the capture note preserves the paper, code, sample, and inference boundaries.

Two Sigma: LLMs as feature and research-system inputs

Two Sigma’s 2026 AI in Investment Management outlook frames AI as an operating layer for quantitative research, with LLMs widening idea generation and shifting attention toward evaluating more hypotheses. Its Part II adds company-aware tools, timestamp and knowledge-cutoff leakage concerns, governance, and overfitting risk. The firm presents these as research and operating principles, not as a claim that LLMs make trading decisions autonomously. The second article also describes frontier models and tools being integrated across research, production environments, incident management, and business processes. That is an organization-level operating statement, not a system specification.

The public LLMs in Action presentation, dated October 29, 2024, connects Matt Greenwood, Chief Innovation Officer and Head of Investment Management Engineering, and Ben Wellington to LLM use in the investing process. A June 2025 interview used an earlier feature-forecasting title for Wellington; a current 2026 Two Sigma article identifies him as Head of Complex Feature Engines. A current official Generative AI quantitative-software role describes a one-petabyte unstructured-data corpus, an interactive research platform, automatic signal mining, LLM/NLP services in research and production pipelines, and millions of daily production-system events. Two Sigma’s 2026 feature-research article describes LLM outputs becoming queryable text data and accelerating exploration of previously expensive feature ideas. Its AI Core Team article describes use for productivity and feature extraction while retaining human judgment.

Two Sigma’s August 11, 2026 follow-up to Ben Wellington’s Flirting with Models episode adds a first-party summary to the recovered audio and timestamped ASR. Wellington describes features as economically meaningful facts, a shared platform through which useful features can be reused, and three possible locations of alpha: data access, feature generation, and forecasting method. He also discusses generated text as a possible research surface, the risk that broad automation reduces researcher diversity, and the value of preserving originality and orthogonality. A CEO-video/micro-expression example is presented as an exploratory illustration, not as evidence of a traded signal. The capture note preserves the episode date, timestamps, and ASR boundary.

An August 11 recheck adds three more official hiring signals. The Techniques Engineering role describes LLM and agentic engineering, agent frameworks, agent-evaluation frameworks, synthetic data, and production-ready research libraries. The Modeling Data Scientist role describes LLM-based featurization and agentic workflows for intraday datasets, with evaluation frameworks for generated features. The Post-Training Research Scientist role describes autonomous experiment iteration, RLHF, DPO, reward modeling, and post-training for financial time series and quantitative reasoning. These are current hiring and research-system signals. They do not show filled roles, model weights, agent permissions, live trading authority, or AI-attributed returns.

The public boundary remains material. Two Sigma does not disclose a complete model inventory, filled-role status, agent permissions, live adoption metrics, or a direct path from its internal GenAI tools to a trading order. The evidence connects LLMs to feature research, research productivity, and post-training hiring intent; it does not prove a language model is itself the trading model.

A newly indexed January 2021 SuperDataScience interview with Claudia Perlich adds a dated personnel and operating-model layer. The publisher identifies Perlich as a senior data scientist at Two Sigma and describes a Strategic Data Science team with a broad R&D mandate, work on less-direct datasets and affiliates, public/open-data activity, and a small nonprofit-oriented data effort (13:40–17:10). Her account decomposes the research path into economic hypotheses, entity characterization, signal validation, portfolio optimization, and execution, and describes embeddings or distance learning for an entity-matching problem, alongside Python, cloud compute, and Ibis in the then-current research environment (22:08–32:48). The same discussion emphasizes data skepticism, expected-versus-observed checks, and hiring for complementary hypotheses and unusual backgrounds (36:50–46:35). This is a 2021 speaker account, not a current Two Sigma organization chart, model inventory, filled-role record, permission map, or performance result; the source note preserves the timestamped transcript privately.

The title-blind pass also found The Data Standard interview with Jatin Dewanwala. The publisher identifies him as Head of Research at Hildene Capital Management and describes prior structured-credit work at Metacapital, the founding of Krivi around large-scale data analysis and machine learning for structured-products investing, and fixed-income systems work at Bay Crest Partners. The same page lists Columbia financial-engineering and Delhi College of Engineering training. This adds a structured-credit personnel and research-lineage node, but the page does not expose a transcript or a named Hildene AI system; the source note keeps the career chronology, episode framing, and evidence limits separate. No current Hildene model, data source, vendor, production permission, or performance is inferred.

The same publisher’s October 2021 interview with Drew Conway adds a separate private-markets route. The publisher describes Conway as a Two Sigma Senior Vice President working on private-investment decisions and records a data-science/investment-team “buddy system” with one-to-one pairing, regular shared workflow, and measurement of the data-generating process before modeling (28:27–45:10). Conway distinguishes a low-error or attractive backtest from a result that changes underwriting, and says hiring conversations should test how candidates reason through ambiguous measurement problems as well as basic technical skills (37:20–46:08). This is a dated private-markets operating account, not evidence about current Two Sigma public-market systems, current role status, model ownership, or performance; the source note records the private transcript retention and boundaries.

A title-blind Beryl EDU discussion, published on the archive April 14, 2024, adds a dated alternative-data bridge. The captions identify Tony Berkman discussing his Majestic Research and Two Sigma-era experience and compare deep, single-name data work in concentrated portfolios with cross-sectional data use across many names. He describes simulations, candidate models, marginal-data-value analysis, and the role of discretionary experience in feature design (00:13–03:24). Daniel Sandberg, identified through a separate Beryl profile, describes broader demand for niche datasets, PySpark, AWS ML stacks, and multi-terabyte cloud workflows (05:53–08:13). This is third-party educational and vendor-side evidence, not a current Two Sigma system disclosure, named dataset purchase, permission map, or performance record. The Beryl archive itself is now a discovery surface for adjacent AI/LLM, causality, and alternative-data episodes. See the capture note.

A title-blind August 2025 Odds on Open interview with Bill Mann identifies him as a former AQR and Two Sigma fundamental researcher and founder of Harmonic Insights. The timestamped capture adds practitioner detail on point-in-time data conventions, firm-specific data construction, and a view that reconciliation and comparison tasks are candidates for agentic workflow automation (03:42–04:41, 12:52–15:20). Mann also describes using LLMs to help fundamental researchers process company documents after starting Harmonic Insights (24:03–25:28). This is an alumni/founder account, not a current AQR or Two Sigma deployment disclosure; it does not establish models, permissions, production status, or investment authority.

Two Sigma’s public technical artifacts add a separate evidence layer. The firm-owned functional_semantic_types repository describes using foundational models to generate a functional semantic-type ontology, with embeddings, prompt utilities, model-generated code extraction, and graph analysis. The related NAACL industry paper describes downstream uses such as data validation, mapping, and joins. This is direct firm-owned GenAI research evidence. It does not establish a live trading agent, model-serving runtime, or portfolio-management deployment.

The current personnel record also needs qualification. Two Sigma’s 2026 outlook identifies Matt Greenwood as “Chief AI Innovation Officer,” while a 2025 firm article uses “Chief Innovation Officer.” The discrepancy is retained rather than resolved by inference. Jeff Wecker is identified as CTO and Mike Schuster with the AI Core team; Jin Choi discusses supervision, leakage, overfitting, and monitoring; and Ben Wellington is identified in current material as Head of Complex Feature Engines. These are public strategy and practitioner signals, not proof that any named person owns a particular model or agent runtime.

A title-blind follow-up adds a dated platform and software-culture layer. In a June 2019 TWIML interview with Matt Adereth and Scott Clark, Adereth describes a modeling platform spanning data ingestion and cleaning, reusable transformations, point-in-time simulations and backtests, researcher-facing tooling, model deployment, and in-house compute (04:27–05:58; 09:13–14:48; 18:30–21:25). He reports hundreds of researchers and thousands or tens of thousands of simulations in some workloads, and describes standardizing common infrastructure without prescribing every modeling tool (09:13–10:19). Clark describes SigOpt’s experimentation and optimization layer and an academic-conference relationship with Two Sigma (06:11–06:45; 10:20–11:22). These are dated speaker accounts, not a current vendor inventory, model registry, agent permission map, or performance claim. The capture note preserves the boundaries and private transcript hashes.

The same pass recovered Julia Meinwald’s 2017 PyData talk on Two Sigma’s open-source program. She describes an open-source coordinator role, an internal committee and approval process, IP and license concerns, and contributions involving BeakerX, Flint, Jupyter, pandas, Spark, and scientific Python (01:18–03:18; 05:01–05:09; 07:22–08:24; 14:57–19:42). Two Sigma’s current open-source page corroborates the BeakerX/Flint projects, major upstream contributions, meetups, and NumFOCUS support. This is historical engineering-culture and dependency-governance evidence, not disclosure of proprietary model code or current GenAI deployment. A 2021 Michigan MIDAS seminar separately identifies Ben Wellington’s NLP and public-data work; it is retained as personnel context and not treated as an investment-model disclosure.

The current Two Sigma events archive also exposes a media and conference trail that title-only searches miss. It links an Anyscale/Ray on the Road locator for Mike Schuster, Head of AI Core, and an official 2022 webinar by Justin Sirignano covering deep learning for price-move prediction and reinforcement learning for order strategies. Two Sigma’s ICML 2025 review adds the firm’s conference-sponsorship and research-translation surface, covering LLMs, diffusion, alignment, active learning, and sequential inference. These are media, academic-engagement, and research-process signals; the Anyscale recording remains a locator until independently captured.

A person-name expansion adds two missing temporal routes for Mike Schuster: a 2021 Robot Brains episode, a 2024 AI & the Future of Work publisher episode, and a 2025 Two Sigma article summarizing practical LLM use, feature extraction from earnings calls and Fed speeches, team design, data quality, and human review. The 2024 Buzzsprout page exposes Mike Schuster metadata and an audio element, but its visible show-notes body is mismatched to a Russ Fradin/Larridin episode and the MP3 returns HTTP 403 to local requests. The same episode is mirrored on YouTube, where English automatic captions were recovered on August 28, 2026. The caption layer identifies an episode-era AI Core team of roughly 25, describes a mix of engineering and modeling/research roles, and discusses finance-versus-technology-company compute constraints and cloud/GPU tooling (08:20–13:16). It also covers the need for cross-functional teams on large systems and finance-specific risk limits (13:46–17:16). These are dated practitioner statements, not a current headcount or technical audit; see the caption recovery note. The dates and role descriptions form a public personnel/media trail; they do not disclose a complete AI Core roster, model inventory, or trading authority.

The Boston Quantara episode with Yiannis Antoniou supplies an adjacent, transcript-backed governance route rather than a hedge-fund deployment record. Antoniou distinguishes chat-style LLMs from agents that collect information, run functions, synthesize outputs, and initiate actions (03:45–07:48); he calls for recording, replaying, and explaining agent action paths, documented decision boundaries, human-approval thresholds, and a chain of trust (14:32–18:40). The discussion places ultimate liability with the financial institution deploying an agent and describes human touchpoints, escalation, supervision, and provenance as design requirements (20:54–23:48). This is Lydatum/financial-services governance commentary, not evidence of a Lydatum customer, hedge-fund system, model choice, data rights, investment authority, or performance. See the transcript-recovery note.

The same name-expansion pass recovered Jump Trading infrastructure surfaces that generic AI searches omitted. Joe Stam’s NVIDIA GTC session exposes a current research-technology and CUDA route. Alex Davies’ VAST webcast and Redpanda customer case add vendor-described HPC storage, streaming telemetry, market-data, and AI-data infrastructure; the PDSW archive and Cordial Q&A add dated infrastructure and business-boundary context. The official VAST MP4 was recovered and locally transcribed on September 2, 2026; the four-minute recording describes machine-learning data volumes, removal of spinning disks from research pipelines, read-heavy archive performance, and storage supporting researcher/model-training workloads (approximately 00:04–03:52). The original media and timestamped sidecars are retained privately; see the capture note and private archive checkpoint. A newer VAST FWD 2026 customer story names Lucas Wojcik, an HPC Systems Engineer, and describes a shared multiprotocol data layer, API-driven operations, workload-stress testing, and vendor-reported throughput changes. Its embedded recording was found but did not yield a usable caption or audio in this pass; the capture note keeps the written vendor account separate from transcript evidence. These are infrastructure and personnel records, not evidence of a named alpha model or autonomous capital allocation.

The official Jump Trading YouTube channel adds a 2025 Custom HPC Stack video. Jump’s firm-authored description and chapter metadata describe CPU/GPU clusters, multi-site interconnection, direct liquid cooling, Nvidia H100 racks, high-bandwidth interconnects, custom software, and infrastructure intended to support research and trading systems. The checked route did not expose usable caption text, so the description is treated as the source of those statements rather than as a transcript. This is a concrete public infrastructure route; it does not disclose model weights, training data, permissions, a named alpha system, or performance.

Citadel: separate fundamental-research assistants from systematic ML

Citadel’s fundamental-equity chatbot report describes an internal assistant for locating details in public filings, summarizing sell-side research, and tracking executive keywords. The report says the tool was intended to accelerate research while leaving investment judgment with people. It quotes Umesh Subramanian in his then-CTO role; Citadel’s current leadership page identifies Andrew Janian as Interim CTO. A Citadel Reuters NEXT post corroborates the research and risk-assessment purpose. These are media and company-social sources, not an official architecture document. Citadel’s current official Data Strategies Group page describes AI/ML and alternative-data work across fundamental and quantitative investors, but not the chatbot’s implementation.

A Reuters account adds more specific first-party detail: Citadel’s AI Assistant was described as using licensed transcripts, regulatory filings, brokerage research, and Citadel’s own investment strategies; it reportedly highlights risks and generates portfolio-specific research and reading lists. The report says the tool had been rolled out over the prior year and was used regularly by nearly all equities investors. This supports a firm-reported internal research-assistant deployment, while leaving model, permission, retention, evaluation, and performance details undisclosed. Reuters report via Investing.com

A July 2026 Goldman Sachs interview records Ken Griffin describing an internal agentic workflow that reads a finance paper, reproduces its results, verifies the findings, and runs out-of-sample tests. That is public evidence of research acceleration as described by Citadel’s CEO; it does not establish a live portfolio loop, autonomous order authority, model identity, or AI-attributed P&L. Li Deng is excluded from any current Citadel personnel list: public material places his current affiliation at Vatic Investments, not Citadel.

Citadel’s official Global Quantitative Strategies page describes a fully automated systematic strategy and an integrated research platform spanning alpha design, portfolio construction, and execution. The current page also says the platform is powered by modern agentic AI frameworks so researchers and engineers can focus on human-led work. Its 2026 EQR materials describe the data-to-forecast-to-portfolio-to-execution loop and simulation before deployment. These are predictive-ML, systematic-investing, and platform signals. They should not be merged with the fundamental-equity chatbot into a single GenAI claim, and the GQS page does not disclose model identity, permissions, evaluation records, or AI-attributed P&L.

An independent eFinancialCareers profile, linked from Citadel’s company feed, adds a personnel and academic-lineage route without naming a model. The subject is identified only as Navid and is described as an EQR alpha researcher in New York who joined after a Columbia Statistics PhD focused on machine-learning theory. The profile describes overnight simulation runs, written research notes, manager challenge, quantitative-developer collaboration, and repeated signal testing. Because the surname is not published, this person is not merged with other Citadel personnel; the account supports a date-scoped role and workflow description, not current AI practice, a GenAI system, investment authority, or performance. See the capture note.

An independent eFinancialCareers profile, linked from Citadel’s company feed, adds a personnel and academic-lineage route without naming a model. The subject is identified only as Navid and is described as an EQR alpha researcher in New York who joined after a Columbia Statistics PhD focused on machine-learning theory. The profile describes overnight simulation runs, written research notes, manager challenge, quantitative-developer collaboration, and repeated signal testing. Because the surname is not published, this person is not merged with other Citadel personnel; the account supports a date-scoped role and workflow description, not current AI practice, a GenAI system, investment authority, or performance. See the capture note.

The academic-recruiting surface adds a Boston/MIT route. Citadel’s GQS PhD Colloquium describes a multi-day exchange in which machine-learning and deep-learning PhD students and postdocs share research with GQS researchers. MIT CSAIL lists an upcoming September 9, 2026 Citadel and Citadel Securities technical event with participants from GQS, EQR, and Data Strategies and broad topics in mathematics, statistics, and machine learning. A self-authored researcher page identifies Jianbo Chen as a Citadel Securities quantitative researcher working across machine learning, statistics, and optimization. These are academic-network and personnel signals; they do not assign a paper or method to a Citadel system, establish a filled recruiting outcome, or disclose model ownership, data rights, permissions, deployment, or performance. See the capture note.

An August 27 first-party page pass adds a more explicit Data Strategies Group description. Citadel’s Quantitative Research page describes DSG as a central quantitative-research team working across investment strategies on alternative data, AI/ML research, and quantitative modeling. It says researchers build models from noisy and complex sources to infer how the real world evolves, then work with investment teams on investment problems. The page also describes access to large proprietary datasets, research and development tooling, and computing and simulation resources. This is a firm-controlled organizational and hiring surface; it does not identify a model, training corpus, filled role, permissions, or performance.

Citadel’s Commodities page adds a strategy-specific infrastructure datapoint under its Citadel Energy section: approximately 100 dedicated engineers and more than 17 TB of average daily data ingested and processed, both stated as of July 2026. The page names Python data tools, Google Cloud services, Kubernetes, React, Postgres, GitHub, and AI tooling. The scope is the commodities engineering platform described on that page, not a firmwide measurement; it does not establish a particular AI model, data license, production decision right, or return attribution.

Citadel’s current privacy policy supplies a separate non-investment governance boundary. It says third-party AI tools may support business operations such as meeting transcription, summarization, recording, and recruiting workflows; it describes human review for decisions about individuals and opt-in biometric speaker identification where applicable. This is evidence about firmwide operational controls and data handling, not evidence of an investment model or trading workflow. See the capture note.

The public record also contains current AI and ML recruiting and research signals. A current GQS Machine Learning Researcher role names deep learning, NLP, LLMs, pre-training, fine-tuning, and reinforcement learning; a Machine Learning Engineer role covers distributed training, inference optimization, and productionization. These roles do not resolve whether “trained on” language in secondary reporting means retrieval, fine-tuning, or another form of adaptation. The reviewed sources do not publish model weights, evaluation results, live-error rates, model-specific ownership, or a direct link between GenAI assistants and trading decisions.

A title-blind Odd Lots episode with Daniel Morillo, published October 7, 2024, adds a former-Citadel quantitative-research route. Bloomberg identifies Morillo as a former Citadel partner and Head of Equity Quantitative Research; his current Freestone Grove biography lists him as Head of Quantitative Strategies. In the episode, he describes AI as another step in the evolution of data and analytics, says his current firm has invested in AI work, and distinguishes summarization or theme extraction from producing a differentiated investment view (49:15–51:59). He also describes domain-specific questions as important when applying models to market research. This is a dated, third-party speaker account and current Freestone Grove personnel evidence; it does not establish Citadel’s current AI program, a named model or dataset, permissions, deployment stage, or performance attribution. See the capture note.

Another title-blind route recovers Michael Watson’s December 2025 Odds on Open interview, with Spotify, Apple Podcasts, and an iVoox publisher mirror. The episode host presents Watson as a former Citadel Managing Director who progressed from software engineering to equities-engineering leadership. A separate 2019 Data Engineering Podcast interview records Watson and Robert Krzyzanowski discussing Citadel data-engineering teams, data evaluation, cataloging, infrastructure, and Jupyter integration. These sources add dated personnel and engineering context; they do not establish Watson’s current Citadel affiliation or connect him to Citadel’s later AI assistant or systematic platform.

Watson’s current Hedgineer account is a separate vendor-side signal. He describes client-deployed MCP tool groups for operations, research, portfolio analytics, and fund operations; a Claude Code SDK wrapper with pre-deployment evaluation; and knowledge-graph construction across databases, research models, third-party systems, and SharePoint (25:49–29:18). Hedgineer’s company-authored Claude Skills note separately describes model-invoked skills, an internal marketplace, Usage Analytics MCP and OpenTelemetry, plugin hooks and subagents, and group-level access governance. These are Hedgineer product and engineering claims, not evidence that Citadel uses Hedgineer. The public sources do not identify a client, model weights, training corpus, data rights, production error rates, investment permissions, order authority, or AI-attributed performance. See the capture note.

An additional title-blind route is the May 1, 2024 Norges Bank Investment Management interview with Citadel founder and CEO Ken Griffin. In the captioned discussion, Griffin describes internal AI uses for drafting emails, summarizing research, introducing memos and documentation, tagging data, and helping software engineers with productivity (12:54–13:32). He also says Citadel had used machine learning for roughly eight or nine years and describes it as important for asset pricing, with a smaller role in risk management (13:32–14:02). This is a dated executive account from an NBIM production, not a Citadel architecture disclosure; it does not identify model families, datasets, vendors, permissions, evaluation results, or AI-attributed performance. The timestamped capture and boundary notes are in the source note.

The title-blind queue also recovered an Odds on Open interview with Tom, described by the publisher as a former Tudor Investment Corp and Moore Capital quantitative PM; a secondary episode index identifies him as Tom Costello. The episode does not name his current fund. He says that unnamed fund uses AI for research-assistant work such as data cleaning, data-structure preparation, and associating related datasets (25:37–26:28). He also describes a historical 2003 system that parsed news-wire text, required roughly two years of stored data for backtesting, traded on an approximately three-to-twelve-day horizon, and later required hedge changes as the strategy decayed (44:26–48:10). The source is a dated speaker account with automatic-caption limitations: it does not identify the current fund, model, corpus, costs, permissions, live performance, or independent replication. Those current-fund statements are not assigned to Tudor or Moore. See the capture note.

NBIM’s March 24, 2026 AI Summit recording is a useful institutional-manager control case. Presenters describe a cloud/data-warehouse foundation, an AI team, organization-wide training, Anthropic-supported ambassadors, and a responsible-AI operating model with human involvement in investment-related decisions. The use cases include a multi-agent block-trade research workflow, AI media monitoring, multi-agent preparation for company meetings, an in-house forensic-accounting model described as in production, and a two-stage responsible-investment screen that escalates flagged companies to human review. NBIM’s official Strategy 28 and Responsible Investment 2025 report provide separate written support for AI-assisted investment processes and LLM-based risk screening. The video is partly automatic speech recognition with entity errors, so the article uses the recording for timestamped workflow evidence only after cross-checking first-party pages. It does not establish a complete AI roster, model inventory, permissions, or independent performance attribution; see the capture note.

The title-blind pass also recovered a distinct Citadel Securities Future of Global Markets 2025 recording with Jensen Huang, published October 14, 2025 after the October 6 event. Its official description says the firm’s engineers, traders, and researchers use quantitative research, compute, machine learning, and AI to support analytics and other firm challenges. In the captioned discussion, an external-speaker exchange describes quantitative trading moving from human-engineered features toward AI and refers to Citadel Securities as a customer (28:42–29:02); a later segment describes a progression through machine learning, deep learning, embeddings, and multimodality in quantitative trading (36:46–37:15). This is a firm-controlled event and external-speaker account, not a Citadel Securities model or deployment disclosure. It does not identify customer scope, model family, training data, permissions, or performance, and Citadel Securities remains a separate entity from Citadel LLC. See the timestamped capture note.

Millennium: firmwide AI advisory, agents, and end-user customization

Millennium’s technology page describes a dedicated environment for early-stage enterprise products and collaboration with AI companies. It names commercial and proprietary AI tools, agentic infrastructure, and investment-team customization. The page says the technology organization includes AI experts, software engineers, quant modelers, and data scientists, but it does not publish team size by AI function or a model inventory. A June 2026 AI Lab Q&A gives that environment a more explicit charter: experiment with emerging products before enterprise adoption, collaborate with AI partners, and attract specialist talent. A contemporaneous Bloomberg report describes the same remit from an internal memo, including early access and product assessment. These are dated organizational and partner signals, not evidence of results.

The firm’s Gideon Mann interview identifies him as Global Head of AI and gives more detail on the operating model: AI work is split between centralized and federated teams; the core team builds foundational technology while partnering with business users on targeted applications; and an internally built search platform lets employees interrogate proprietary content and developers add search to their own applications. Mann also describes larger coding tasks and iterative research as practical uses of AI. A June 2026 AI Lab Q&A describes experimentation and collaboration with AI companies. The AI hackathon article describes agentic AI, RAG, and MCP workflow prototypes across a multi-region technology organization. Millennium’s privacy notice adds an enterprise-control signal: AI systems are subject to approval processes, employee-use policies, human oversight, and spot checking. These first-party descriptions establish an operating model and search-platform claim; they do not disclose model versions, retrieval permissions, evaluation results, or investment authority.

Millennium’s August 2026 Anthropic announcement adds a distinct risk-management partnership route. It says Anthropic forward-deployed engineers are working with Millennium technology and risk teams to build, pilot, and optimize a supervised digital risk analyst. The stated functions include explaining daily risk changes, interrogating data, retaining information across interactions, and surfacing risk insights across asset classes. Millennium also says it will test Anthropic’s frontier models against sophisticated firm work. This is first-party evidence of co-development and model-evaluation intent; it does not disclose model versions, retrieval or tool architecture, evaluation scores, risk thresholds, production rollout, autonomous action, or AI-attributed performance.

The regional trail adds implementation context without resolving deployment. Millennium’s 2025 hackathon recap names agentic AI, RAG, and MCP workflow automation among project areas across New York, Miami, London, Dublin, and Tel Aviv. An ADAPT / ML Dublin report records a Millennium-sponsored event where Paolo Aloe discussed technology integration and Pat Lenihan presented natural-language SQL, multi-step research agents, and cross-model LLM standardization. The March 2026 Dublin event page is a recruiting surface that combines AI, technology, quantitative strategies, and execution services. These sources expose public topics and talent routes; they do not identify live systems, teams, model versions, or authority boundaries.

Current personnel pages add implementation detail. Daniel Tymecki is described as building agentic tools for investment professionals; Amuthan Kannan describes chatbots and multiple agents for testing and edge-case discovery; and Scott Rofey is identified in a June 2026 first-party interview as Global Head of FIC and Cross Asset Risk. Rofey describes parsing market content into tailored insights, interrogating market and risk data, combining skills through agents, and faster code-based model testing. A current Applied Cloud and AI Engineer role adds architecture language around production-ready LLM applications, multi-step agents, retrieval, Agent Harness infrastructure, Agent Flywheel trace capture, evaluation gates, hallucination-rate measurement, cost/latency telemetry, MCP servers, and named agent frameworks. These are first-party personnel and hiring accounts. They do not disclose the model versions, evaluation results, deployment map, or investment authority of the described tools.

Three additional first-party pages widen the personnel and talent map. Anish Mandalika is identified as an AI engineer on the Equity AI team who uses large language models to improve fundamental-equity portfolio-manager processes and emphasizes data and process modeling. Nitesh Ranveer is identified as the new Head of Research in India; his June 2026 Q&A describes a Bengaluru research team integrating financial modeling, earnings coverage, sector analysis, internal tools, and AI-enabled workflows while retaining judgment and accountability with analysts. Millennium’s January 2026 campus-recruiting article says its global internship program had grown to nearly 200 interns and added a specific AI-intern track for designing, developing, and implementing AI solutions. These are firm-authored role and talent statements, not proof of filled-system ownership, model inventory, production permissions, or investment performance.

Millennium’s official account of its participation in MLDS 2026 names Vaibhav Jain as Team Lead, AI Engineering and says he presented in Bengaluru on building reliable AI agents in production. The organizer’s speaker page describes his prior BlackRock and J.P. Morgan AI/ML-oriented roles and frames the talk around tool chaining, reflection, human-in-the-loop design, selective multi-agent systems, observability, and governance across research, monitoring, operations, and reporting. This is a useful personnel and public-topic signal for Millennium’s India technology presence; it does not establish a live internal system, model inventory, partner, evaluation result, investment use, or autonomous authority. Jain is a separate person from Vaibhava Goel, whom the Georgia Tech programme lists as Millennium’s Head of Machine Learning Research. See the capture note.

The title-blind MIT recruiting route adds a current quantitative-research control. Millennium’s 2027 London Quantitative Researcher Intern listing describes market and alternative-data analysis, statistical and machine-learning methods, backtesting, data normalization, portfolio-optimization tooling, and support for live strategies. It also asks candidates to apply AI tools in quantitative workflows and explain methodology, rationale, and output validation. This is a dated hiring specification, not a filled-role record: it does not identify an AI leader, model, vendor, training corpus, permission boundary, production system, or investment result. See the capture note.

A separate former-personnel and open-artifact route comes from Karan Vora’s public account. Vora says a Lead Quantitative Developer role for a planned Millennium macro pod was cancelled before his scheduled start after the pod was dissolved; this should not be treated as evidence that he joined Millennium or that Millennium deployed his proposed design. His post describes point-in-time market and macro data, event-driven pipelines, replayable agent workflows, provenance, and deterministic quantitative tools. His public Standard-Tools repository documents typed tools for LLM agents, leakage-purged walk-forward validation, model-to-backtest bridging, and audit records. The repository is maintained by Vora, not by Millennium; the capture note keeps the employment and deployment boundaries explicit.

The public materials therefore expose a platform-and-advisory model with local customization. They do not disclose which investment workflows are live, which models are fine-tuned, what agents can execute, or whether any investment outcome is attributed to an AI system.

D. E. Shaw: a visible applied-AI hiring layer with business-boundary issues

The D. E. Shaw Group’s current careers page lists Applied AI Engineer, Senior Product Manager—Applied AI, Product Manager—AI Vendor Tools, and machine-learning researcher roles under Quantitative Strategies and Technology. The descriptions reference AI agents, agentic systems, generative-AI technology, and internal vendor-tool products. A current DESIM Portfolio Strategist role says the investment-management team uses advanced research and technology and seeks comfort applying generative-AI tools.

The Applied AI Engineer listing adds implementation vocabulary that was not visible in a title-only count: bespoke agents and user applications, production ownership, coding agents, reusable skills, agent frameworks, large-scale retrieval, shared infrastructure, and firmwide adoption. The Machine Learning Researcher listing separately connects ML engineering to large-scale knowledge discovery in financial data and systematic research. These are current role descriptions, not evidence that the roles are filled or that the described systems have trading authority. The source note is here.

The current careers index also renders a Senior LLM Researcher, Lead Software Developer — GAI Infrastructure Services, AI-enablement, and AI-engineering roles. The Machine Learning Researcher description connects ML/software engineering to large-scale knowledge discovery in financial data and systematic research; the Product Manager — AI Vendor Tools description names Claude, ChatGPT, Copilot, and Gemini as internal vendor-tool surfaces. These are current role-family and workflow signals, not filled-role or production evidence.

A current Fundamental Equities - AI Product Analyst role is more directly investment-workflow specific. It sits in Financial Research, describes integrating agentic AI tools into core analyst workflows, and asks the hire to map processes with analysts and portfolio managers; redesign idea sourcing, document synthesis, screening, and model maintenance; prototype with Claude Code; iterate from investment-team feedback; package successful workflows; and comply with data licensing, information barriers, and compliance standards. This is a current hiring signal for fundamental-equity workflow redesign. It does not establish production deployment, model inventory, agent permissions, autonomous investment authority, or performance attribution.

The current careers index exposes a wider set of adjacent surfaces. A Strategic Intelligence: Software & AI Research Analyst role covers software, generative AI, cloud, semiconductors, cybersecurity, data, automation, and enterprise-technology adoption as inputs to long-cycle investment analysis. A Product Manager — AI Vendor Tools role names Claude, ChatGPT, Copilot, and Gemini as internal vendor-tool surfaces and assigns ownership of integrations, evaluation, roadmaps, and adoption. A separate AI Engineer — Human Capital role names LLMs, RAG, Cursor, Claude Code, and agentic architectures for talent workflows, while an AI Enablement Strategist role describes firmwide AI fluency and workflow enablement.

The same role family also sharpens the boundary between applied infrastructure and investment use. The Machine Learning Researcher listing describes high-performance knowledge discovery in financial data, proof-of-concept implementations, efficient deep-learning training workflows, GPU/low-level optimisation, and deployment into systems affecting decision-making and trading. The Applied AI Engineer listing describes bespoke agents, reusable skills, large-scale retrieval, shared agent infrastructure, coding agents, and concept-to-production ownership. Together these pages establish firm-controlled hiring intent across research, infrastructure, vendor management, and enablement; they do not prove that the roles are filled, identify an AI leader, reveal model or data inventories, or establish a live agent’s permissions or investment performance. See the capture note.

The current D. E. Shaw firm overview adds dated operating context: it reports more than $100 billion in investment and committed capital as of June 1, 2026, more than 750 developers and engineers, systematic and discretionary capabilities, and a stated willingness to build internal platforms when external ones do not meet firm standards. These are firm-reported scale and technology statements. They do not identify the AI staffing share, compute allocation, model inventory, or an AI-attributed investment result.

These are separate signals. Quantitative Strategies hiring supports an investment-technology and applied-AI layer; DESIM’s role supports employee use of GenAI in portfolio-analysis work; the broader D. E. Shaw Group also has unrelated AI and machine-learning activity in human capital, private equity, and venture-studio contexts. The public sources do not establish a single investment-business AI program, a current AI executive, or a direct path to trading decisions.

The firm’s public LinkedIn company feed adds a current, title-blind conference route: at its ICML 2026 booth in Seoul, D. E. Shaw said it was asking researchers and academics how they use AI to have better research days. That supports a firm-controlled recruiting and research-conversation signal around AI-assisted work, but it does not name an owner, model, dataset, deployment endpoint, trading permission, or performance result.

The personnel disqualification pass found no named current investment-side AI/ML leader corroborated by a first-party biography. DESRES research, Arcesium, venture-studio, private-equity, and vendor-tool personnel are separate entities or functions and are not counted as D. E. Shaw investment AI personnel.

A title-blind historical-media check recovered a 2019 Data Driven NYC interview with Pedro Domingos, identified there as D. E. Shaw’s Head of Machine Learning, with a recoverable video route. Domingos’s academic CV separately records a 2018–2019 Head of Machine Learning Research role at D. E. Shaw and his academic/ML background. This strengthens the dated personnel-to-academic route, but it remains historical and does not establish a current D. E. Shaw role or production system. See the capture note.

A separate historical practitioner route now fills a thin Crabel lane. The South Park Commons AI Speaker Series article, dated February 26, 2018, identifies Justin Nelson as a Crabel Capital Management commodity-futures trader and attributes to him a description of experimenting with neural networks for systematic futures strategies. The article discusses risk management, strategy life cycles, non-stationary markets, data labelling and cleaning, model interpretability, bootstrap aggregation, clustering, autoencoders, and language-based sentiment signals. These are dated practitioner observations, not proof of Crabel’s current models, data, deployment, permissions, or results; the article’s performance assertions and heuristic estimate about idea failure are excluded. See the capture note.

The entity boundary is especially important for D. E. Shaw Research’s public research page, which describes machine-learning applications in computational chemistry alongside supercomputing and drug discovery, and its official ICML 2026 post. A NeurIPS 2025 talk page adds a dated scientific-ML artifact: Peter Skopp’s abstract describes physics-based synthetic data from Anton molecular-dynamics simulations and methods for incorporating that data into multimodal large language models for drug discovery. These are DESRES scientific-research signals; they are not evidence about the investment-management business, financial datasets, trading models, or portfolio authority.

The official D. E. Shaw YouTube channel also exposes a title-blind Systematic Strategies video, uploaded in 2022. Its automatic captions describe statistical forecasts for financial instruments, in-house quantitative tools, a unified system processing exchange data near real time, and a machine-learning clustering project (00:01–01:39). The later discussion frames open-ended research around practical impact rather than publication (01:42–02:09). The video does not identify the speaker in its metadata. The capture note treats this as historical, firm-produced workflow evidence—not a current model inventory, GenAI disclosure, permission map, or investment-authority claim.

A separate title-blind Wall Street Oasis episode provides a publisher transcript for an unnamed practitioner’s historical path through a national research laboratory, D. E. Shaw, Jump Trading, and Citadel. The account discusses specialty recruiting, supercomputing and Python experience, risk-management and Monte Carlo engineering, shared research infrastructure, and a Citadel research-cluster build at dated transcript points (11:09–18:24; 26:11–30:42; 42:24–43:31). The capture note keeps this as historical, self-reported personnel and engineering context; it does not establish current employment, AI/GenAI use, a named system, or firm-wide practice.

PDT: applied ML and automated trading, without public GenAI linkage

PDT’s work page describes a scientific process in which quantitative models are researched, peer reviewed, empirically validated, and then deployed to automated trading systems. Its careers page currently lists an Applied ML Scientist, quantitative researchers, research engineers, and platform roles. The Applied ML listing describes devising, implementing, evaluating, and iterating statistical methods for trading strategies.

The current Research Engineer role, accessed August 12, 2026, adds infrastructure detail: the research-engineering team partners with quantitative researchers on infrastructure for alpha, signal, and portfolio construction, optimizes models for real-time trading-system inference, and asks for experience building infrastructure for training or fine-tuning large ML models. This is official hiring evidence for large-model research infrastructure. It does not identify LLMs, GenAI agents, a filled role, production status, or AI-attributed returns.

This is direct evidence of predictive-ML research and automated trading infrastructure. The reviewed public sources do not identify LLMs, GenAI agents, a model-adaptation program, or a research assistant used by portfolio managers. “No public GenAI evidence found in the reviewed surfaces” is the correct boundary; it is not evidence that PDT lacks such systems privately.

Public personnel evidence identifies Rushi Nadimpally with PDT’s research-engineering function, with a MLSys listing using the title Head of Research Engineering. This is a current research-engineering signal with public corroboration, not evidence of GenAI ownership or a finance-language-model program.

A separate academic-recruiting route resolves one otherwise anonymized PDT panelist. Columbia’s public event page and a Princeton event page describe “John” as a Princeton economics PhD and Tufts mathematics/economics graduate whose research covered India’s employment-guarantee program and microfinance contracts, followed by Highbridge’s global-macro group and PDT. A public LinkedIn profile for John Papp lists PDT and Princeton; his coauthored American Economic Review paper and public India employment-guarantee paper match the event biography, while The Org’s unverified profile supplies the Highbridge-to-PDT timeline. This is high-confidence identity matching and academic-lineage evidence, not a first-party PDT biography; it does not establish current AI ownership, a GenAI system, transfer of academic methods into trading, or performance. The event’s “Max” remains unresolved and is intentionally not assigned a full name. See the source ledger.

A Stanford-hosted profile for Kurt Tadayuki Miller says he is a quantitative researcher at PDT Partners and records Berkeley PhD/MA training with Michael I. Jordan as advisor, Stanford degrees, and publications on Bayesian nonparametric latent-feature models, link prediction, dynamic graphs, and variational inference. A University of Washington event page from October 2020 and a 2021 Women in Machine Learning workshop programme independently corroborate a historical PDT researcher named Kurt/Miller in a careers or quantitative-research context. The current status remains unverified. The academic topics are not relabeled as PDT trading methods: no public source links them to a live PDT model, LLM, GenAI system, agent, dataset, or investment result. See the capture note.

Two additional public personnel routes add Boston/MIT and quantitative-finance career context without adding a public AI claim. An NSF Institute for Artificial Intelligence and Fundamental Interactions event page lists Matthew Rispoli of PDT Partners for a December 2025 industry lunch and describes his Harvard Physics PhD, Harvard–MIT Center for Ultracold Atoms affiliation, and quantum-gas-microscopy research. Harvard’s Physics PhD record names Markus Greiner as advisor for Rispoli’s 2019 thesis, and the Greiner Lab profile records his 2013–2019 graduate period and publications, including work using artificial neural networks to analyze non-equilibrium quantum states. A public Harvard-affiliated paper lists him as a 2016 coauthor. The IAIFI page is current-dated role evidence; the Harvard pages and paper are academic-lineage evidence. Separately, the 2021 Women in Machine Learning programme lists Winnie Yang in a PDT Partners careers session, and a University of Michigan event page identifies her in 2021 as Executive Director, Trading Model Implementation while describing her Waterloo computer-engineering degree and prior Morgan Stanley, HBK, and KCG/Virtu path. A public professional profile and unverified organizational profile provide additional PDT affiliation and later-title leads. The later title and current status are not independently verified. Neither route identifies an AI remit, LLM, agent, dataset, production system, portfolio authority, or performance. See the capture note.

XTX Markets: predictive ML, XTY Labs hiring, and compute

XTX’s careers material describes machine-learning techniques, extensive computation, and price forecasts across financial assets. The page presents a research and engineering environment with a fast feedback loop around quantitative models. It does not identify LLM agents, generative-AI workflow tools, or a finance-language-model training program.

The public record is therefore useful as a predictive-ML control case. It shows why “AI in trading” cannot be treated as synonymous with GenAI. Atlas Wang’s University of Texas profile, Stony Brook seminar, and the XTY Labs announcement support a current research-director/AI-Lab signal tied to financial time series and market-data research.

The March 9, 2026 Stony Brook AI Innovation Institute AI3 seminar supplies a specific university-hosted description in this route. It identifies Atlas Wang as Research Director at XTX Markets and says the firm frames algorithmic trading as a deep-learning and foundation-model problem, with time-series modelling, large-scale optimisation, representation learning, and foundation models in the research agenda. The abstract claims forecasts for tens of thousands of instruments, more than $300 billion in global trading volume, and fully automated execution without discretionary human intervention; these are speaker/company claims reproduced by the event page, not an independent audit. The biography says Wang founded and leads XTX’s New York AI Lab for foundation models on financial time series and market data, while on leave from the University of Texas at Austin. The page does not provide a recording, model weights, training corpus, evaluation protocol, permissions, or independent performance evidence.

A current AI Research Internship - XTY Labs listing and the corresponding official careers API record add a more specific public signal. The role sits in Quantitative Research, names XTY Labs as a 2024 XTX division led by Atlas Wang, and describes work across machine learning, optimization, generative AI, foundation models, distributed training, NLP, time-series analysis, control/optimization, reinforcement learning, models, agents, and software prototypes. It also lists LLM and multimodal-framework experience as desirable. A separate Machine Learning Performance Engineer listing in Tradingdev ETD Tech describes an ML Performance and AI Acceleration function focused on training and inference platform performance, optimizing compilers for accelerated computing, irregular bitwidth numerics, novel hardware architecture design, and mapping AI models from JAX graphs to transistors.

These XTX sources are material hiring and infrastructure evidence, not a deployment record. They do not disclose filled-role status, model weights, a complete model inventory, production permission maps, autonomous trading authority, or AI-attributed returns. XTX Ventures is a separate venture arm and is not merged into the market-making personnel record.

XTX’s first-party Ventures page adds a separate external-AI ecosystem signal. It says XTX Ventures partners with founders from seed through Series B, focuses on technical founders with AI and machine-learning expertise, and offers technical validation, infrastructure-scaling support, architecture guidance, code review, customer introductions, and access to an engineering and developer network. This may reveal where the broader XTX group is willing to place technical attention and capital, but the page does not establish that any portfolio company supplies XTX Markets, that a venture investment became an internal dependency, or that a named technology is used in trading. The venture-arm page must therefore remain separate from XTX’s market-making, XTY Labs, and trading-infrastructure evidence.

A separate first-party research-talent route is XTX’s December 2025 pure-mathematics funding announcement. XTX says it committed £26.37 million across seven UK universities for PhD and postdoctoral positions entering in 2026–2028, with more than 100 early-career research positions planned. The announcement explicitly connects advanced mathematics with AI, cryptography, defence, and finance. This is evidence of a university research-funding network and stated talent-pipeline intent. It does not identify awardees, XTX hires, principal-investigator relationships, a model, dataset, production system, or transfer of university work into trading. The capture note keeps that institutional route separate from XTY Labs, TernFS, and market-making disclosures.

Capula: recovered Asia-Pacific conference audio, with no firm AI disclosure

The title-blind media pass recovered the Bloomberg Asia Centric excerpt and linked FICC Focus full-forum episode from Bloomberg’s June 3, 2026 Volatility Forum in Singapore. Publisher metadata identifies Oliver Chan as a portfolio manager at Capula Investment Management, alongside portfolio manager Ivan Nurminsky of Dymon Asia and derivatives personnel from Optiver. The recovered publisher transcript adds timestamped navigation for discussions of single-stock ETFs, retail options, dispersion, dealer gamma, hedging, and regional volatility (01:23–02:00, 03:35–05:20, 12:01–17:35).

This closes a capture gap and improves the Asia-Pacific personnel and media trail. It does not add evidence about Capula’s AI or GenAI strategy: the AI references are general market commentary, and no model, dataset, vendor, agent permission, production endpoint, portfolio authority, or performance claim is established. Speaker labels are generic and the transcript is automatic, so it is a navigation artifact rather than a verbatim quotation source. See the capture note.

A separate personnel and academic route materially expands the Capula AI search without converting academic work into a firm-system claim. Marta Grzeskiewicz’s public research page identifies her as a Portfolio Research Scientist at Capula and describes work combining economics with machine learning, reinforcement learning, multi-agent systems, agent-based modeling, and deep learning for market and macroeconomic decision problems. The page lists a University College London PhD in Economics and Machine Learning and Cambridge affiliations, and links to work on physics-informed neural networks, inverse reinforcement learning, causal inference, and computational economics. Her public LinkedIn announcement separately records joining Capula in that title while retaining Cambridge teaching and research. This establishes a named, self-reported personnel and research-lineage route; it does not disclose Capula’s model inventory, data, deployment, permissions, or performance, and the academic projects are not relabeled as Capula trading methods. See the capture note.

A second Capula personnel route comes from Jason Ho’s public profile, which describes him as an AI and full-stack developer on Capula’s Euro RV trading desk. The Org’s Capula office directory separately lists him as a Full Stack Developer, but does not confirm the AI or desk remit. UCL records link Ho to a 2021–2022 MotionInput V3 project using machine-learning computer-vision libraries and to a 2020–2024 MEng Computer Science degree; his profile also describes a GitHub Copilot experiment in a test-to-code dissertation. This adds a current, self-reported investment-technology personnel lead and a dated UCL technical lineage. It does not establish Capula’s model inventory, data, vendor, evaluation, deployment, permissions, investment authority, or performance, and the university work is not treated as Capula research. See the capture note.

The title-blind hiring pass adds a separate, dated platform signal. An archived Capula Quantitative Developer — Data Platform & Risk Analytics listing specified production Python in Docker, SQL/DuckDB/Parquet data services, low-latency C# components, AWS, CI/CD, infrastructure-as-code, and P&L/VaR/scenario/exposure analytics over tick and end-of-day data. The listing is now expired, so it is not evidence of a current vacancy or filled role; it is useful as a dated description of platform and risk-engineering requirements. Capula’s current LinkedIn jobs page separately displays Technology Graduate Analyst and Quantitative Strategist (PhD) vacancies. Neither route identifies a GenAI system, model vendor, agent, data corpus, deployment state, or investment authority. See the capture note.

The same note records a historical academic-lineage lead rather than a current personnel claim: an MIT-hosted CV lists Jaume Vives i Bastida as a 2015 Capula quantitative-research intern and separately describes MIT work in econometrics, statistical learning, machine learning, causal inference, regularization, neural networks, and synthetic controls. It also lists a later Two Sigma PhD Symposium presentation. The CV does not establish post-internship Capula employment, method transfer, or Capula use of those models; those academic topics remain separate from Capula system evidence.

A new cross-firm conference route comes from AIMA’s Technology & Innovation Day 2026 overview, agenda, and speaker archive. The London event took place on June 24, 2026 and listed Amira Amin, COO of Liquid Strategies at Magnetar, and Jon Freedman, CTO of Capricorn Fund Managers, on an operational-excellence panel. Its agenda describes an ipushpull technology snapshot for real-time trader-chat capture, standardisation, RAG, human-in-the-loop methods, and pre-trade analytics; it also describes Coremont Clarion Copilot as an AI assistant for natural-language portfolio insights. An AI-governance panel covers validation, data controls, auditability, bias testing, recordkeeping, vendor oversight, and resilience. These pages add conference-role and vendor-workflow metadata, not evidence that Magnetar, Capricorn, Marshall Wace, or any other named manager uses those products or operates a particular model. No replay, transcript, model card, customer identity, data-rights record, permission map, or performance evidence was recovered. See the capture note.

Aspect Capital: AI/ML as a research method with explicit caution around deployment

Aspect’s 2025 J.P. Morgan podcast page identifies Martin Lueck discussing machine learning, AI, and LLMs in the investing process. Its 2023 podcast summary records caution about using machine learning, while the firm’s quantitative-ML challenge emphasizes repeatable evaluation, feature selection, rapid iteration, and risk analysis. Aspect’s team profile for Bas Monsewije discusses ML for feature selection and the role of technology in research and portfolio construction.

The expanded media pass adds several independent routes. The AIMA interview with Anthony Todd says Aspect has researched machine-learning models for years, is exploring and using ChatGPT in selected areas, and treats intellectual-property risk as a serious boundary. The J.P. Morgan transcript with Martin Lueck describes constrained datasets, interpretability, and ML across signal generation, volatility forecasting, portfolio construction, and execution. Top Traders Unplugged provides an older transcript with the same model-first and anti-overfitting boundary, while the Macro Hive route adds a 2023 title-blind podcast surface. A current The Derivative episode with CIO Christopher Reeve adds an AI-related academic biography and model-evolution discussion, but its substantive focus is systematic macro and portfolio construction. A public timestamped transcript is now retained privately for navigation; it remains a third-party, non-audio-verified text source. The expanded Aspect source note records the capture and its boundaries.

The same J.P. Morgan interviews now have stable YouTube and caption records. In the 2025-recorded Martin Lueck video, the captioned discussion places an underlying premise and interpretability around financial ML, then describes constrained-data testing and caution about convincing but false LLM outputs (16:51–19:55). In the 2023-recorded Anthony Todd video, the captions attribute to Todd a statement that client capital was invested across Aspect ML models and that the firm had developed 23 ML-based models exploiting different effects (00:58–01:19); the discussion later characterizes the approach as constrained and supervised (13:54–15:04). The 23-model statement is a dated speaker claim with no model list, data license, validation split, or return attribution. The captions are automatic, and the new capture note preserves that boundary. These routes add historical and current-public discussion; they do not establish a current GenAI system or autonomous investment authority.

The previously unresolved Podbean enclosure is now recovered and locally transcribed. The 2023-recorded audio attributes to Todd a four-part model-evolution frame—data, models, processing technology, and market coverage—and a description of broader data inputs, AWS cloud work, and more than 190 markets in the Aspect Diversify programme ([02:50–05:57], local audio). It also describes price, economic, option-derived, and alternative data—including NLP, sentiment, fund flows, weather, and shipping—and frames ML as a hypothesis-led tool for diffuse multi-input effects ([11:44–16:45], local audio). These are dated practitioner statements that add operating detail to the video caption record; they do not establish current coverage, model identity, training data, cloud scope, production permissions, or AI-attributed returns. See the Aspect title-blind source note.

The public material supports a hypothesis-driven systematic research process that uses ML selectively and tests models against financial constraints. It also exposes technology and talent interfaces: Mike Kwan’s profile describes quant-development infrastructure, interactive backtesting, trade scheduling, Arrow Flight, and cloud-native architecture; the OSGD ML challenge exposes a public research-and-recruiting interface around overnight futures forecasting; and Aspect’s technology interview names Gemma Hagen and discusses cloud migration, data, software, production, and execution. These sources do not disclose a GenAI assistant, model-training pipeline, LLM vendor, or autonomous investment authority. The evidence is practitioner, technology, and research-process disclosure, not a current GenAI deployment record.

The current Martin Lueck leadership profile and 2025 podcast support a current investment-research leader discussing ML, AI, and LLMs. Bas Monsewije’s profile supports a current ML-aware research role. Neither source establishes a dedicated GenAI team, LLM deployment, or named AI-platform owner. A historical ML-challenge participant without current corroboration is excluded from the current personnel record.

A recovered 2021 ReSolve episode, identified by Aspect’s first-party announcement, adds historical practitioner detail from Aspect’s Director of Investment Solutions Razvan Remsing. ReSolve’s publisher transcript PDF supplies speaker labels and timecodes, and a public recording was downloaded for local recovery. The conversation discusses flow data as alternative data, sentiment and forward-looking information, conditional and trend models, risk forecasting, and differences between financial and commodity data (approximately 10:58–11:15, 18:00–19:10, and 23:55–26:05). This is a dated 2021 practitioner-media source; it does not establish a current model inventory, GenAI deployment, data rights, live permissions, or performance. See the capture note.

A title-blind review of Aspect’s current careers surface adds an operational view: research is organized around signal generation, portfolio construction, and market execution, with quantitative researchers, portfolio managers, and quant developers using automated execution and backtesting infrastructure. Bas Monsewije describes ML for feature selection; Mike Kwan describes quantitative-development work spanning backtesting, trade scheduling, and cloud-native data infrastructure; and Sunny Ratilal adds a business-automation route across execution, settlement, and fund accounting. These are useful title and workflow signals, but they do not establish LLM, agent, or GenAI deployment. The careers page also exposes three Vimeo video routes; they are retained as media-discovery routes, not treated as investment evidence. See the Aspect source note.

PGIM Quantitative Solutions: language models described as research inputs

The archive-reconciliation pass recovered the canonical recording for Top Traders Unplugged ALO35, featuring George Patterson, whom the publisher identifies as Managing Director and CIO of PGIM Quantitative Solutions. In the recovered YouTube recording, uploaded June 6, 2026, Patterson describes language models as a source of market data and research signal for quantitative-equity and multi-asset work, and traces a progression from bag-of-words methods through BERT and FinBERT to LLM-based theme identification (00:32:29–00:34:35). He also describes tracing a recommended position back to raw underlying data as a transparency objective (00:33:13–00:33:32). The Podscan episode page provides an alternate episode-level locator.

This is a dated, named-CIO practitioner account with a recoverable recording and publisher transcript route. It is evidence of Patterson’s public description of PGIM’s research process; it does not establish model weights, training data, data rights, live deployment coverage, portfolio permissions, autonomous trading authority, or AI-attributed performance. Captions are automatic and are retained for navigation rather than treated as human-verified verbatim text. See the capture note.

The separately captured Flirting with Models episode with Stacie Mintz, published August 3, 2026, adds a second PGIM Quantitative Solutions evidence lane. PGIM identifies Mintz as Managing Director and Head of Quantitative Equity. In the episode, she describes a sequence for LLM research that starts with the concept or dataset, selects the simplest suitable tool, and increases complexity only as needed; she gives non-financial quality measures such as board composition and innovation as examples of text-derived inputs (38:33–40:12, automatic local ASR). She then describes extracting comparable information from company text and using board-member linkages as an example of turning qualitative information into a systematic signal (40:32–42:48). The discussion names hallucination, memorization and look-ahead bias, model selection, out-of-sample time, repeated-prompt stability, and linkage back to company sales or earnings as validation concerns (43:13–45:54). Earlier, Mintz describes PGIM’s in-house risk model as a diversification engine and a source of customization (06:34–10:36). See the timestamped capture note.

This is a dated practitioner account, corroborated by PGIM’s first-party role and quantitative-equity pages. It adds research-design and validation language, not a public model inventory or production record. The sources do not establish the model provider, training corpus, data rights, evaluation fixtures, adoption coverage, permissions, live portfolio authority, or AI-attributed performance; the local transcript is automatic ASR and is retained for navigation and paraphrase rather than verbatim quotation.

Voleon: named predictive-ML leadership, without a public GenAI layer

Voleon’s official overview describes investment management through machine learning and says its models serve financial prediction. The firm’s management page supplies a more specific current personnel signal: Michael Kharitonov is CEO; Jon McAuliffe and Vasco Chatalbashev are Co-Chief Investment Officers; Prem Gopalan is CTO; and Dave Tolliver is Chief of Technical Staff. The same page connects Chatalbashev to building Voleon’s predictive models and trading systems, Gopalan to portfolio optimization and market-impact estimation, and Tolliver to research leadership for regional equity strategies.

This is useful evidence of named ownership around predictive modeling, portfolio construction, optimization, and research systems. It is not evidence of a public LLM, GenAI assistant, agentic workflow, or autonomous investment authority. Voleon’s public careers surface adds current hiring context, but the reviewed pages do not disclose a language-model inventory, agent permissions, evaluation suite, or GenAI-to-trading path. A current public Jiafan He homepage and Google Scholar profile identify him as a Voleon Member of Research Staff, with machine-learning, reinforcement-learning, large-language-model, and optimization research interests. This is a personal/Scholar personnel signal with a verified Voleon email, not a firm-controlled LLM program disclosure or deployment claim.

The Stanford Advanced Financial Technologies Laboratory page for Mike Ryerson adds a dated, title-blind research-talk route. Stanford records a May 3, 2018 talk, “The Benefits of Machine Learning for Quantitative Investing,” and identifies Ryerson as a Voleon senior member of research staff. The abstract says the talk would address why machine learning may suit quantitative investing and Voleon’s general approach to implementing those techniques. This is a university-hosted abstract, not a recording or transcript; it does not establish the current Voleon model stack, deployment status, permissions, or performance. See the Voleon capture note.

A title-blind Bloomberg Masters in Business interview with Jon McAuliffe adds historical practitioner evidence. The publisher identifies McAuliffe as Voleon co-founder and CIO and describes a systematic investment process built around machine learning, data, and a proprietary predictive engine; the interview’s closing discussion emphasizes that predictive quality still requires risk management. The YouTube capture is dated September 13, 2023 and now supplies an English automatic-caption timestamp layer. The captions place the process/risk/execution discussion around 00:41:21–00:43:04 and the extension of machine learning to additional assets around 00:47:15–00:47:22. This strengthens the historical predictive-ML record, but it does not establish Voleon’s current LLM or agent architecture, model inventory, permissions, or performance attribution; the captions are navigation evidence, not a manually verified transcript.

Two newer public routes sharpen the current personnel and strategy record. A Commonfund Forum 2026 summary, dated April 29, 2026, identifies Mark Refermat as leading Voleon’s machine-learning portfolio-strategy team and summarizes a market-neutral approach focused on dispersion among individual companies. It attributes to Refermat an explanation of machine learning searching complex, nonlinear relationships across datasets, including satellite parking-lot imagery, and records risk limits, daily risk-committee oversight, and automated monitoring as controls discussed by the panel. A Canada Alpha Generation Symposium 2025 biography gives the title Managing Director, Machine Learning Portfolio Strategy, and says Refermat works with the research team on strategy analytics and serves on the risk committee; it also records prior Man AHL and GAM Systematic roles and an academic background in cross-asset liquidity and quantitative investment strategies. These are allocator/event descriptions, not a complete current roster or independent performance assessment. See the capture note.

A July 2026 WebsEdgeScience video about Voleon adds a recent public media route that a current job-title search would miss. The video describes Voleon’s origin in machine-learning research for quantitative finance and systematic investing, and names Michael Katzanelis as a partner and co-founder in the recording (00:33–01:20). It describes work on stock-return forecasting, optimal trading, new machine-learning models, feature ideas, and strategies built from scratch, alongside an academic-style research environment with clear metrics (01:32–03:15; 03:25–04:21). The closing discussion describes candidate signals such as mathematical facility, computational thinking, and curiosity. The video is hosted by a third party, does not clearly identify every speaker, and supplies no model inventory, data rights, live headcount, permissions, or performance attribution. See the capture note.

Coatue: a named sector head describes an AI-native research redesign

A publicly accessible Coatue investor presentation adds a much more concrete current-firm record. The Q2 2026 CTEK presentation says Coatue spent approximately $53 million in 2025 on data, data infrastructure, data science, AI, software, and the associated personnel, and had more than 500 datasets in production. It describes Mosaic as the internal home for data-science insights and says Coatue is building an AI “Brain” that combines internet and external data with internal data-science, Mosaic, and firm-domain context. Those figures and labels are Coatue’s own, and the presentation cautions that the tools may not benefit every fund.

The same presentation names four analyst-facing agents: a Meeting Agent for preparation, an Earnings Agent for extracting updates and implications from public-company results, a Brainstorm Agent for developing stock or thematic theses, and an Expert Agent for identifying proprietary research angles. It places them above a separate orchestration layer that schedules and coordinates agents against the data layer. Importantly, Coatue explicitly says the Brain and agents assist analysts, do not make investment decisions, and are not built into the investment-decision process itself. The document does not disclose models, providers, prompts, user counts, evaluation results, permission semantics, or AI-attributed returns. It is also labeled for accredited-investor and qualified-client use and “not intended for public use or distribution,” although Coatue hosts the file at a public URL; the article therefore treats it as a dated, firm-controlled disclosure rather than as independently audited evidence.

Two earlier primary sources establish chronology and one current model provider. Coatue’s November 2024 SEC filing says the firm began substantial investment in its proprietary data-science platform in 2015, integrates Mosaic with trading, order-management and treasury systems, and launched the Coatue Brain in 2023. A February 2026 firm-authored Anthropic post reports 35 data scientists and engineers; active Claude use for research-report compilation and long-idea generation; Claude Code agents running data-analysis scripts for hours; and Claude Skills under construction for fetching earnings transcripts. It does not say Claude is the Brain’s underlying model or publish task evaluations, permissions, incident data, adoption counts or investment effects. Coatue also says it co-led Anthropic’s financing and may benefit from the position, creating a disclosed vendor/customer/investor conflict that should accompany product-use claims.

This chronology also limits what can be attributed to Frank Long. A January 2026 Coatue announcement names him Head of AI & Partner on the Hedge Fund team. Because the Brain predates his appointment by roughly three years, the title does not establish that he created Mosaic, launched the Brain, owns the four named agents, or manages the pre-existing data-science organization. His prior public technical lineage includes a Goldman-assigned multimodal time-series foundation-model patent that combines time series with time-stamped exogenous data such as news and supports task-specific decoder heads. The patent establishes disclosed design and inventorship, not deployment at Goldman, migration to Coatue or current investment use.

A historical 2019 Rev 2 panel supplies a more explicit organizational baseline. The recording description identifies Thomas Laffont as Coatue co-founder and senior managing director, Alex Izydorczyk as Head of Data Science at Coatue, and Matthew Granade as Chief Market Intelligence Officer at Point72. The panel describes a Coatue data-science team spanning engineering, science, and analysis, and a platform intended to support public- and private-market work (02:20–05:16). The speaker describes the team as roughly forty people at that time; this is a dated speaker account, not a current headcount.

The panel also discusses alternative-data sourcing, data-product construction, legal/compliance review of data provenance and permissioning, and the distinction between data-driven, model-driven, and model-informed investment processes (11:17–16:29; 19:14–21:27). It names research-infrastructure tools including Domino and Databricks in a discussion of managing data-science work (36:52–38:05). This is historical panel evidence and does not establish current personnel, model ownership, dataset licenses, production deployment, agent permissions, autonomous order authority, or performance. See the capture note.

A separate October 2025 Sourcery interview identifies Michael Barton as a Coatue sector head and supplies a useful operating-model route that a firm-name-only search could miss. Barton describes a Coatue data-science platform built during the firm’s cloud transition, with company data brought together for analysis (33:49–34:33). He then describes a proposed AI-native workflow for research triage, model construction, combining disparate datasets, and automating steps in the investment process; the discussion includes hiring analysts to help redesign workflows and a future in which sector coverage is supported by multiple agents (34:35–36:00).

The same interview describes weekly KPI tracking across covered companies using credit-card data, email traffic, and other real-time inputs, including companies outside the portfolio (57:56–59:17). It also describes a human decision gate in which detailed analysis and models are reduced to a concise investment pitch for review (43:54–46:06). This is a named practitioner’s account and forward-looking operating vision, not proof of a firmwide agent deployment, a specific agent count, model ownership, data rights, autonomous order authority, or performance. The publisher’s scale figures are not treated as audited AUM. See the capture note.

The Berkeley CLIMB Voleon Seminars archive also currently lists 2025–2026 talks on online allocation, decision-focused value of data, multiple-testing control, human-AI systems, and privacy, with speakers affiliated with MIT, Stanford, Harvard, and Carnegie Mellon. This confirms a current Voleon-associated academic event surface and research-topic mix. It does not show which employees attended, what Voleon sponsored internally, or whether any method entered a live system. The public record still does not disclose a Voleon LLM/GenAI program, model inventory, training corpus, permissions, or AI-attributed returns.

A current Voleon Senior Machine Learning Engineer posting, posted May 13, 2026 and checked August 27, makes the research-to-production boundary more explicit. The first-party role says it partners with research staff to productize machine-learning models that drive quantitative trading strategies; it also names ingestion, feature engineering, validation, data-quality monitoring, experiment and model-evaluation pipelines, model serving, feature stores, distributed training, and deployment coordination. The posting describes Voleon as a technology company applying AI/ML to finance and reports a company-stated base range of $290,000–$395,000 plus bonus. This is high-confidence evidence of hiring intent and the technical interfaces the firm describes, not proof that the role is filled, that a named model or LLM is live, or that any return is attributable to it. See the capture note.

Euclidean Technologies is an adjacent, value-oriented quantitative manager with a more specific historical ML description. In a 2022 Excess Returns interview, co-founder John Alberg describes forecasting future operating income, using uncertainty estimates to inform a margin-of-safety process, and combining deep learning with simpler models for value-trap prediction (29:15–32:58, 33:45–36:20). The conversation also describes neural networks processing financial statements and deriving relationships beyond hand-specified ratios (40:36–40:57). This is a third-party, dated practitioner account: it does not establish current Euclidean continuity, an LLM or agent, model weights, data rights, live permissions, independent evaluation, or performance. See the capture note.

Winton, Systematica, Marshall Wace, Brevan Howard, Caxton, and Squarepoint

Winton’s working-at page describes centralized data provision, execution infrastructure, quantitative research, and technology roles. Its current Simon Judes profile, 2026 podcast page, and recovered Bloomberg recording support a current investment-leadership discussion of AI, machine learning, alternative data, cloud technology, and large-language-model developments (11:09–12:00). The recording does not identify a Winton model, AI/ML department, training record, or production AI claim. No reviewed Winton page identifies an LLM or agentic investment workflow. See the capture note.

Winton’s official news index also preserves a June 23, 2026 company LinkedIn update about the Goldman Sachs European Hedge Fund Symposium in Cannes. The update says CIO Simon Judes discussed AI’s role in developing investment infrastructure, the search for durable alpha sources, and talent in investment management. It is event-summary metadata only: no recording, transcript, slide deck, model, vendor, dataset, evaluation, or production detail is published on the indexed route. The source note keeps it separate from the recovered podcast evidence. A current title-blind check of Winton’s opportunities page adds a dated staffing route: the page lists Data Engineer and Data Analyst roles for the Quantitative Platform, Cloud Engineer, post-trade software roles, Quantitative Developer in Commodities, and MENA Equities Quantitative Researcher, with displayed locations in London and Abu Dhabi. The page does not use an AI or LLM title. This is evidence of current data, cloud, post-trade, and quantitative-research vacancy metadata—not evidence that roles are filled, that Winton lacks private AI work, or that any model, dataset, agent, or production system is involved. See the capture note. An earlier Top Traders Unplugged transcript adds timestamped research-process and machine-learning constraints: multiple-hypothesis control, avoidance of selective backtest presentation, and limited independent observations at slower trading horizons. It is historical third-party executive evidence, not a current model or deployment disclosure. Cambridge’s official technical-talk archive now shows a multi-year GSA technology and recruiting route rather than a single isolated event: its 2021 page names Joris Peeters and describes trading systems, data pipelines, risk calculations, historical backtesting, new data streams, and AWS simulation; the 2024 page names Mihai Enache and describes technology teams, stacks, execution services, and the idea-to-execution order lifecycle; and the 2025 page names Junhui Yang and describes data pipelines, backtesting, monitoring, exchange-data compression, approximately tenfold lower replay time, and rapid intraday-signal generation. The archive also records a 2022 GSA talk by Julian Roth. These are university event descriptions and personnel profiles; no recording was recovered, and they do not establish an LLM, GenAI workflow, model inventory, permissions, or performance. See the GSA capture note.

The current Winton homepage also exposes an official Vimeo profile and employee-video routes for Smitha in Internal Audit, Stephen in Investment management, Alex in Client operations, and Rae in Human Capital. These add firm-controlled role and culture media, not AI-system evidence. A separate direct check also verified Winton’s official YouTube channel and its 2023 Institutional Investor fireside video; the captions discuss allocator portfolio fit, trend following, risk parity, and strategy comparison, but not AI (00:33–00:57, 01:26–01:36). A different older directory ID still fails direct existence checks and remains a stale route. The corrected route details and caption boundary are in the Winton media source note.

Systematica’s Leda Braga profile and a current podcast support current investment-leadership commentary on AI’s role in research and signal development. A dated Stanford Women in Data Science keynote, published April 9, 2018, gives a more concrete historical description: Braga separates signal generation from portfolio construction, discusses exotic data and execution signals, and identifies sparse financial data and overfitting as central constraints. Additional title-blind routes include a Women in Data Science interview, a Money Maze episode, and SS&C event coverage. The recovered audio for the Portuguese-language Outliers episode adds timestamped, automatically transcribed executive commentary: Braga describes coding a scalable quantitative decision process (06:11–06:20) and requiring mathematical, team-based justification for risk cuts by examining failing model hypotheses (15:01–15:44). These are translated ASR paraphrases, not exact quotations. A CNBC Events transcript attributes to Braga a separate data-research team that treats and selects features from imagery, text, sound, unformatted text, and consumer data before alpha-making. Its culture page identifies a Data R&D team, and a current software-engineering role mentions AI process modernization. Systematica’s August 2025 Responsible Investment Policy adds a more specific data-to-signal record: ESG-provider scorecards and information from Refinitiv, corporate disclosures, and research papers are combined with a proprietary thematic score, alpha factors, restrictions, and exclusions before the policy says the cost-controlled signal enters the algorithmic trading process (source extraction). This is a dated policy disclosure, not evidence of an LLM, agent, or current AI owner. No reviewed source names a current GenAI owner, language model, model card, or production AI investment workflow. A historical QCon speaker is excluded because the current public affiliation is elsewhere.

The current Systematica homepage adds a separate scale and scope record: it displays “$13+ Billion AUM,” labels the figure “Firmwide Assets Under Management,” and lists trend following, macro non-trend, multi-strategy, equity market neutral, and customised-solutions strategy families alongside offices in Europe, Asia, and the United States. The page does not state the AUM measurement date or provide vehicle-level composition. This updates the scale map only; it is not an AI disclosure and does not establish a model, agent, deployment, permission, or performance claim. See the capture note.

A title-blind personnel and modality route adds Ruofan Zhou’s personal research site, which reports “2020-now” as Quantitative Analyst at Systematica Investments and lists a 2015–2020 EPFL Computer Science PhD plus earlier Tsinghua degrees. A public LinkedIn search result also associates Zhou with Systematica in Geneva and describes computer-vision and deep-learning experience. The personal site footer says it was last updated in September 2020, so the employment line is historical and date-sensitive rather than current-role confirmation. EPFL’s thesis-defense record and Süsstrunk profile identify Sabine Süsstrunk as thesis director and document the academic lineage. Zhou’s public papers cover realistic image super-resolution, GAN-based blur-kernel modelling, stochastic frequency masking, joint demosaicing/super-resolution, and microscopy-image mapping (KMSR project, frequency masking, demosaicing/super-resolution, W2S). This adds a computer-vision/scientific-ML lineage, not a financial-vision, LLM, or GenAI deployment claim. The reviewed sources do not connect these methods to Systematica code, data, portfolio construction, permissions, or performance. See the capture note.

The next title-blind pass adds a named data/AI leadership and governance lane. A Bloomberg Professional Services recap of a November 2025 London summit identifies Grégoire Dooms as Systematica’s Head of Data Research & Development and attributes to him a discussion of AI lowering the cost of unstructured-text processing and enabling sector- or asset-specific feature-extraction pipelines. A Neudata conference route identifies Richard Waters as Head of Data Engineering and describes legacy-stack migration toward AWS and Snowflake with common governance/access for internal and third-party data. These are named practitioner and event/vendor statements, not a model inventory or production proof. Public sources give Dooms differing titles, including Product Manager and Portfolio Manager, so the discrepancy is retained rather than reconciled by assumption.

The Carey Olsen risk-governance interview identifies Ben Dixon as Systematica’s General Counsel and Chief Compliance Officer. Dixon refers to work by “our head of data” involving AI models for alpha generation and data extraction, and describes enterprise-chat uses such as policy review, board-minute preparation, regulatory updates, and horizon scanning alongside guardrails, human review, and records/transparency concerns. The recording does not name the Head of Data or disclose model, vendor, training, evaluation, or production details. Systematica’s culture page and current platform-engineering postings add a first-party Data R&D and AWS/on-premise automation/platform surface without mentioning LLMs, GenAI, agents, or model training. See the capture note.

A separate title-blind SBAI episode published January 6, 2026 features Braga and Systematica General Counsel/CCO Ben Dixon discussing trade errors, near misses, control processes, disclosure, and a no-blame reporting culture. The recovered public audio has a private timestamped ASR capture; the discussion is relevant to the control environment around systematic automation, not evidence of an AI system, model, data source, or performance result. See the capture note.

Marshall Wace’s technology internship programme describes AI-powered automation, coding-agent projects, and workflow development. Its public GitHub organization includes RAG and recursive-language-model workshops, while its privacy policy describes AI/ML in transcription, research, and due-diligence workflows. Its Quant Associate Programme guide exposes a virtual-machine Data Lab, permitted Python/R/Matlab/Excel analysis, trading-signal case work, and a stated prohibition on LLM chatbots during applicant tests. A Bloomberg IBVAL disclosure describes a vendor ML product used for pricing insights. No named current AI/ML leader or Marshall Wace-trained model passed the corroboration loop.

The current Marshall Wace technology page adds a distinct firm-media surface: it describes ongoing investment in systems, data, infrastructure, and people, plus staff expectations around automating and streamlining complex processes. Its teams page separates data engineering, platform engineering, production engineering, and software engineering responsibilities around acquisition, cleansing, enrichment, automation, resilience, and end-user tools. These pages describe organizational capability and workflow scope, not a model inventory or an AI-attributed result. A historical academic route is also now documented: the Stanford-hosted 2002 Kernel-Based Reinforcement Learning in Average-Cost Problems lists Dirk Ormoneit with Marshall Wace Asset Management in its author-affiliation block and studies stable kernel/local-averaging approximations for reinforcement learning. That establishes a dated author-affiliation and research-method signal, not firm adoption, current employment, or continuity into present AI/GenAI systems. See the capture note.

An upcoming UK event adds a current speaker-title route without adding a model claim. The official AFI Innovation Summit agenda dates the London event to October 8, 2026 and lists Stefan Delmarco as global head of quant implementation at Marshall Wace and James Hylands as head of AI at Fasanara Capital for an AI-use session. It also lists Flora Hudson-Evans, director of HR at Aspect Capital, and Simon Bowie-Britton, CTO at Trium Capital, in separate operating sessions. This is organizer-published upcoming agenda metadata; no recording, transcript, model, data-rights evidence, permission map, or production disclosure is available. See the capture note.

The current Marshall Wace AI Placement 2027 listing adds a firm-controlled hiring artifact. It says AI Development teams support Finance, Operations, Risk, Trading, and Portfolio Management; build services, models, and data assets that bring generative AI into investment and operations workflows; and work on an AI platform, backend services, LLM pipelines, and a RAG engine. It also says candidates may design and train models for firmwide use cases and that the stated aim is to let every engineer use AI/ML without being a specialist. The source extraction classifies this as current hiring intent. It does not establish a filled role, model or vendor identity, data permissions, evaluation results, investment authority, or performance attribution.

An additional title-blind route comes from IC Hack 26, held at Imperial College London from 31 January to 1 February 2026. Imperial’s event report identifies Marshall Wace as title sponsor, while the public event schedule lists two Marshall Wace workshops. A participant-authored project timeline records ERLA, a second-place project in the Marshall Wace “Emergent Behaviour” track, and its public GitHub repository describes a recursive literature-research agent: parallel Scouts search Semantic Scholar, summarise papers with LLMs, validate summaries at token level, and generate hypotheses; a Master Agent manages branches and splits research by field, topic, or time. The repository describes a pure-Python Sentinel/Detector/Explainer validation pipeline using LettuceDetect and ModernBERT-NLI, and lists Claude via OpenRouter, PyMuPDF, FastAPI, Convex, and PyTorch. A public LinkedIn profile result reports Marshall Wace experience and associates its author with the project, while a separate participant post gives a related project account. The public profiles expose inconsistent teammate attribution, so no complete team roster is asserted. This is a dated student project and sponsor/event route, not evidence that ERLA was built inside Marshall Wace or integrated into its investment workflow. The repository’s groundedness figure is a project claim without a published held-out benchmark, denominator, or independent audit. See the capture note.

The new August 26, 2026 NBIM interview with Marshall Wace co-founder Paul Marshall is a title-blind founder route that materially deepens the public account. Marshall describes TOPS as evolving from a sell-side idea-measurement and virtual-portfolio system into a strategy using optimization, algorithmic trading, and active machine learning to mine signals and expand data inputs (audio 19:25–20:03). The same episode is also available as an episode-specific NBIM YouTube recording; its English captions are retained as a second timestamp locator, not a separate source event. He distinguishes earlier sentiment and balance-sheet processing from AI systems that can consider broader context, and says fundamental teams have built agentic portfolio workflows by encoding their skills into portfolio systems (audio 20:07–22:08). He also describes a future in which agents could recursively analyze data and develop signal implementations, while the same exchange discusses the firm’s technology staffing. Later, Marshall describes overnight testing of human-created skills and prompts and a possible future self-correcting loop (audio 32:15–32:56). These are current founder statements captured from automatic ASR; they do not establish a named head of technology, current agent count, model/provider inventory, data rights, evaluation results, permissions, live order authority, or AI-attributed performance. See the timestamped capture note.

Brevan Howard’s current Quantitative Strategist listing describes an employer-reported macro-research platform with LLM-assisted ingestion, RAG, embeddings, local/open-weight models, guardrails, evaluation, and tracing, with outputs delivered to portfolio managers and traders. A separate Senior Quant Developer artifact, mirrored through LinkedIn and a third-party job index and apparently closed on August 8, 2026, names Claude, Cursor, GitHub Copilot, and Codex for AI-assisted development; it also describes internal benchmarks, agentic code-review and regression-test pipelines, and LLM use cases for anomaly detection, report narration, and reconciliation triage. Because the first-party posting was not recovered, this is historical secondary hiring evidence, not current firm policy or deployment proof. Tim Mace’s public profile and a Financial News report support current affiliation and a Head of AI report, but the exact title is not shown on Brevan’s official team page. Treat the workflow as employer-described and the title as qualified, not confirmed by first-party leadership material. A Flirting with Models episode and a later At the Forefront interview add historical Bin Ren/Systematic Investment Group and SigTech routes; those describe a former-employee and vendor lineage, not current Brevan deployment. A 2019 AI/Data Science in Trading brochure separately names Sebastien Guglietta as Brevan’s then Co-Head of Computational Intelligence Systematic Strategies, a dated conference affiliation. AIMA’s Technology & Innovation Day speaker page also lists Mike Sanders as Chief Technology Officer at Brevan Howard in the prior-year speaker line-up. This is a dated conference-organizer biography, not a Brevan-controlled leadership page; it qualifies a public technology-title signal but does not establish a current title, AI remit, reporting line, model ownership, or deployment.

The personnel and academic-lineage pass also found Daniel Libman’s public profile, which displays a Brevan Howard association and lists prior Bar-Ilan University work on Volume Prediction with Neural Networks, Forecasting Quoted Depth With the Limit Order Book, and Mutual Information between Order Book Layers. The Frontiers paper uses deep feed-forward networks to study per-minute quoted-depth forecasting, while the Entropy paper record studies information dependence across order-book layers; the earlier volume-prediction paper describes a hybrid LSTM, SVR, and autoregressive approach. This adds a publicly visible microstructure/ML lineage route, not a Brevan-owned research disclosure. The profile does not show a title or remit, and the papers do not establish that Brevan uses these methods, that Libman works on them there, or that any related model is deployed, authorized for portfolio decisions, or associated with performance. See the capture note.

The current Brevan Howard surface also exposes a distinct historical conference route. The Quant Conference’s 2019 London programme lists Sebastien Guglieta as Co-Head of Brevan Howard’s AI Strategies Group for a keynote on AI-based macro strategy, human–machine collaboration, and causality. The page says access to the speaker-allowed recordings is restricted to ticket holders, so only the title and displayed role are promoted. A separate Imperial College people page identifies Sesh Karri as a Research Fellow in Machine Learning at the Brevan Howard Centre and describes work on ML for finance and tick-by-tick trading rules. Karri’s self-authored research page records an academic lineage through Francis Bach, Marc Deisenroth, Adrian Weller, and Vladimir Kolmogorov. The centre and conference routes are kept separate from Brevan fund operations: they do not establish a current AI organization, model inventory, training corpus, deployment, permissions, or performance. See the capture note.

Caxton is disqualified under the current strict screen. Its official site and careers page do not disclose current AI/ML/LLM personnel or deployment. The relevant Python LLM Engineer listing is stale or third-party hiring intent and does not survive the current-role corroboration loop. A separate FCA MIFIDPRU disclosure states that Caxton Associates LLP acts as a delegated sub-investment manager for Caxton Associates LP, reports more than US$10 billion of group client assets and approximately 160 staff as at December 31, 2024, and identifies currency, financial, commodities, and securities markets. This is a legal-entity and scale correction, not evidence of AI deployment or a current AI team. See the capture note.

A recovered Alternative Data Podcast episode with Vadim Khidekel adds historical context without changing that current-status conclusion. The publisher describes Khidekel as a semi-retired buy-side quant with Art Advisors and Teza Technologies experience; in the indexed automatic transcript, he describes Art Advisors’ historical roots in a Caxton quant group and a progression from price/volume models to analyst revisions, financial statements, and later alternative-data experiments (00:07:34–00:10:45, 00:12:32–00:16:34, 00:21:15–00:23:59). The discussion also describes coverage, crowding, timing, and horizon constraints for credit-card, satellite, employment, social, and patent data (00:26:30–00:27:37, 00:51:22–00:52:58). This is historical practitioner evidence about a former Caxton-related quant lineage, not evidence of current Caxton AI/GenAI, model deployment, permissions, or performance. See the timestamped capture note.

A title-blind Apple Podcasts pass also recovered Bin Ren on AI at Work: Finance and a historical Bruce Kovner/Caxton episode. The former adds founder/vendor and historical Brevan context; the latter is historical founder media. Neither establishes current Brevan or Caxton AI deployment, permissions, or performance, and no performance figures from the Kovner episode are used.

Squarepoint’s experienced-professional page and early-careers page support systematic quantitative ML and automated strategies. The official STAC New York agenda names Vinit Adya of Squarepoint on a trading-infrastructure panel and places it beside an AI track on LLM evaluation, production RAG, agentic-era MLOps, benchmarks, validation, and neurosymbolic AI with quantitative finance. Adya’s public post is personal commentary on agentic inference consumption, while his NYU profile records electronic-trading architecture teaching. These sources add personnel and conference routes, not proof of Squarepoint adoption or a current firm-controlled LLM workflow.

A separate title-blind personnel route adds a different modality. Peetak Mitra’s self-authored homepage states that he currently works as a Quantitative Researcher at Squarepoint and specializes in AI-driven weather and energy forecasting to inform trading strategies. It describes prior real-time power-market forecasting at Equilibrium Energy and weather-intelligence work at Excarta. His public LinkedIn profile supports a Squarepoint affiliation but does not expose a complete current title or reporting line. Mitra’s 2021 paper on compressing learned fluid-simulation models and 2023 paper on local machine-learned CFD correction document prior scientific-ML work on efficient inference and model-plus-solver workflows; they are not Squarepoint research disclosures. See the capture note. The route adds a named, self-reported finance role and weather/energy modality, but it does not establish a firm-wide GenAI program, model ownership, deployment status, permissions, or performance.

A second title-blind personnel route adds Daqian Shao. A public LinkedIn profile result reports a Systematic Quant Researcher role at Squarepoint and describes an Oxford PhD focused on reinforcement learning and causal inference. Oxford’s doctoral profile names Marta Kwiatkowska as supervisor, records a 1 February 2026 leaving date for the doctoral-student entry, and describes work spanning deep learning, reinforcement learning, causal inference, hidden confounders, robustness, and decision policies. The Alan Turing Institute profile links him to quantitative robustness evaluation of AI-driven decision pipelines. His public papers cover instrumental-variable and double-machine-learning methods for decision policies (2024), causal imitation learning under confounding (2025), temporal-logic-constrained reinforcement learning (2023), and sample-efficient decision policies (2026). These papers describe academic methods, not Squarepoint models or alpha research: the reviewed material does not connect them to Squarepoint code, data, production deployment, portfolio authority, or returns. Because the LinkedIn and Oxford time anchors differ, the employment association remains profile-reported rather than independently dated. See the capture note.

These firms remain underexplored in different ways: some disclose investment leadership and predictive ML but not GenAI; one discloses an employer-described LLM-assisted macro workflow with qualified personnel evidence; and one has only stale LLM hiring signals. “No public evidence found” remains a search boundary, not a claim about private capability.

An archived Squarepoint Quantitative Researcher — Equity listing, marked removed April 21, 2026, adds a dated implementation vocabulary that is absent from the generic firm pages. The mirrored description names Python and Shell automation, KDB/Q and BigQuery, large unstructured datasets, time-series forecasting, simulation and application validation, live-trading-automation monitoring, statistical-arbitrage research, and cross-region model projects. Squarepoint’s current experienced-professionals page corroborates the broad connection between research, ML, quantitative models, automated strategies, and high-performance trading systems. The archived listing is secondary and does not prove that a role was filled or that every named tool is current; it does not identify an LLM, GenAI system, agent, model inventory, data rights, permissions, deployment endpoint, or performance. See the capture note.

A separate London School of Economics recording supplies historical, title-blind context from a December 2020 conversation. Local ASR locates Marshall’s account of Alpha Capture as a sell-side virtual-portfolio and idea-database process (02:36–03:02), his discussion of financial-statement processing, human intervention, and alternative data (14:03–17:23), and his description of high-volume idea processing, execution-cost work, and a former Citadel algorithmic-trading hire (44:18–45:22). This is a distinct public university recording, separate from both the premium Capital Allocators item and the current NBIM interview. It is historical speaker-reported evidence with automatic-ASR proper-noun limitations; it does not establish Marshall Wace’s current architecture, model inventory, permissions, agent authority, or performance. See the capture note.

The broader public-signal tape

The local public-signal tape and research signal landscape place these firms beside Two Sigma, Bridgewater, Balyasny, Schonfeld, D. E. Shaw, Tower, XTX, CFM, and G-Research. Those peers reveal different layers: research agents, artificial-investor experiments, platform governance, ML infrastructure, or controlled human-machine workflows. They should not be collapsed into one score because several are hedge funds, research firms, or technology analogues rather than direct asset-manager peers.

Two new title-blind control routes sharpen the cross-firm vocabulary. The CFA Institute conversation with Francesco Fabozzi, published July 29, 2026, separates encoder-style text representations from generative language models and discusses sentiment versus return prediction, textual factors, time-ordered evaluation, synthetic data, multi-agent model risk, difficult gold sets, and human-controlled portfolio decisions (15:43–24:57, 34:56–50:57). This is academic/practitioner context from Johns Hopkins and Yale-affiliated speakers, not evidence about a named fund. The MSCI Perspectives transcript, dated March 12, 2026, names Hitendra Varsani and Andrew DeMond and describes agentic workflows across index construction, portfolio exposure, risk, governance, and MCP-connected datasets, with human validation retained. It is vendor-side operating-model evidence, not proof of adoption by any tracked manager. The source notes and MSCI capture record preserve the boundaries.

The title-blind pass also recovered a useful vendor/exchange route that is easy to miss because it is framed as market-data infrastructure rather than hedge-fund AI. In a December 12, 2025 Beryl interview, viaNexus CEO Tim Baker describes a platform built from an IEX acquisition, normalization across partner, exchange, and market data, and MCP access for agent-based systems (01:47–03:35). He describes constraining an LLM to prescribed platform sources rather than open-web retrieval (04:04–05:07) and a SIX partnership for white-label exchange-data delivery (06:29–07:05). SIX’s October 29, 2025 announcement independently confirms the co-developed distribution platform and planned testing of performance, entitlement models, and operational resilience. This is a dated vendor/exchange disclosure, not evidence that GMO, Acadian, Arrowstreet, or another tracked manager uses the platform; it does not establish model identity, customer permissions, completed production rollout, investment authority, or performance. See the capture note.

Another Beryl title-blind recording, “What are some Big Alternative Data Sourcing & Scaling Mechanism?”, published April 26, 2024, adds an operating constraint to the vendor map. The panel describes alternative-data teams working in silos, independently defining fields, losing a “golden copy,” and accumulating data sprawl as cloud pipelines and dataset counts expand (00:43–01:35). It then describes a convergence of data-engineering and analytical skills, including SQL-based tooling for data analysts (01:49–02:05). The captions are automatic and the panelists are not reliably identified; these are dated vendor-side workflow observations, not evidence of a named manager’s architecture, staffing, model, permissions, or performance. See the capture note.

Public-signal dimensions

The evidence below is organized by signal type. It is descriptive, not a ranking or a cross-firm score.

Dimension	What the public record shows	Evidence	Confidence boundary
AI inside investment research	Acadian’s public materials connect modular AI signals to forecasts, portfolio construction, execution, and human review	Named-practitioner interview and current investment-AI engineering role	Workflow direction is public; production status and performance are not
GenAI platform engineering	Arrowstreet’s public role describes LLM routing, Bedrock, MCP, RAG, agents, telemetry, and default-deny controls	Current AI-platform job listing	Architecture intent is visible; live adoption and investment use are not
AI investment judgment	GMO’s public material frames AI through valuation, quality, capital intensity, systematic signals, and uncertainty	GMO letters, video, and Form ADV	Investment thesis is public; internal GenAI machinery is not
GenAI strategy design	Man AHL publicly describes GenAI-assisted trend-following strategy design	2025 firm article	Specific implementation and performance are not public
Narrow financial-model fine-tuning	CFM publicly describes compact-model fine-tuning for financial named-entity recognition	CFM/Hugging Face case study	Task-specific extraction evidence; not investment reasoning or autonomous research
Centralized Applied AI deployment	Balyasny publicly describes centralized model evaluation, scoped tools, team-specific agents, and investment-research workflows	OpenAI customer case study	Vendor/customer-reported workflow; independent performance and control audit are not public
Skeptical control	AQR’s public ML material emphasizes theory, signal-to-noise, validation, and data-mining risk	AQR learning material and related research	Historical ML material is not current GenAI deployment evidence
AI engineering and runtime staffing	G-Research public roles cover AI engineering, ML/HPC, GenAI software, on-prem serving, MCP, and sandboxes	Current role descriptions and prior deep dive	Capability and hiring intent are public; investment deployment is not
Artificial investment associate and AI research lab	Bridgewater publicly describes AIA Labs, an artificial investment associate, PAT, LLMs, agentic workflows, and integration with investment research	AIA Labs and PAT disclosures	Firm-reported architecture and use cases; independent performance attribution is not public
Machine-learning trading organization	Jane Street publicly describes deep-learning models for trading, training/inference infrastructure, and research in production	ML, quantitative-research, and machine-learning hiring pages	ML trading role is public; GenAI-to-trading linkage is not
Adjacent prop-trading AI controls	Jump, HRT, Optiver, and DRW publicly describe AI/ML infrastructure, low-latency trading constraints, LLM/RAG hiring, AI Labs, or LLM-agent workflow surfaces	First-party AI/ML, research, and careers pages checked August 12, 2026	Useful control cases; not comparable to asset-manager public records and not evidence of autonomous capital allocation or AI-attributed returns
Crowdsourced hedge-fund agent control	Numerai publicly describes an AI-scientist workflow, Predictive LLM-generated features, open agent skills, MCP access, and current AI Scientist hiring	Numerai homepage, NumerCon recap, MCP docs, GitHub example scripts, and official careers host checked August 12, 2026	Platform and tournament evidence; not proof of autonomous firm-capital authority or independently audited model contribution

Personnel coverage and title-signal sweep — August 13, 2026

The roster has now been audited separately from the firm capability summaries. This matters because a title such as CTO, Head of Research, or Quantitative Developer can sit close to AI work without proving AI ownership. The new personnel title-signal audit classifies each signal as a named AI owner, named practitioner, role-led AI function, AI-adjacent leader, ML-only signal, or disqualified lead.

Direct-title records in the current sample include Man Group’s Head of Data and AI, Millennium’s Global Head of Artificial Intelligence, Technology, Tower’s Global Head of Core AI & ML, Bridgewater’s AI & ML Investment Strategy leadership, Balyasny’s Chief AI Officer, WorldQuant’s AI-initiative leadership, and H2O’s CTO/Head of Innovation in Generative AI. Title-free or title-adjacent records are CFM’s ML Lab and agentic research personnel, Arrowstreet’s platform/security hiring and public recruiting artifacts, Point72/Cubist’s GenAI/MCP/security roles, Schonfeld’s investment AI lab, and G-Research’s Core AI and Applied AI roles.

The main gaps are also title-specific. GMO has public ML/technology and investment leadership but no current named GenAI owner in the reviewed record. Jane Street, AQR, PDT, XTX, Aspect, Voleon, Winton, Systematica, Marshall Wace, and Squarepoint expose meaningful quantitative or ML signals without a current firm-controlled GenAI owner. Brevan Howard has a useful employer-described LLM workflow and a qualified Head-of-AI lead, but the title needs firm-controlled confirmation. Caxton’s LLM role does not survive the currentness and source quality loop.

This is a coverage map, not a ranking of people or firms. See the full title taxonomy, evidence grades, source ledger, and next-sweep protocol in the linked audit.

The second sweep expands the vocabulary further. State Street Investment Management’s current role uses “Insight Generation and AI Adoption” rather than Chief AI Officer while covering GenAI, agentic AI, modeling, and enterprise platforms. Marathon’s Head of Data Management & AI role connects data architecture and governance directly to investment-research synthesis and portfolio-manager partnerships. These are adjacent investment manager control cases, not hedge-fund personnel rankings.

The sweep also found a historical ExodusPoint Head of AI announcement, a named but self-published Systematic Strategies quantitative-trading practitioner writing about agentic alpha research, and an uncorroborated third-party claim about a Magnetar “Head of AI Quant.” The first is not current-roster confirmed, the second is practitioner evidence rather than a firm platform audit, and the third remains excluded pending firm-controlled confirmation.

ExodusPoint’s current Global Privacy Notice adds an operational-AI signal that is distinct from its historical personnel announcement. The notice says the firm may record and transcribe calls, video conferences, and meetings using AI-powered technology; it names audio/video, AI-generated transcripts, participant and meeting metadata, business-record and compliance purposes, role-based access controls, periodic access reviews, cross-border transfers, and retention. This establishes a firm-stated voice-data and governance surface. It does not identify a vendor or model, establish model training, reveal meeting coverage or adoption, connect the workflow to investment research, or grant an agent any trading or portfolio authority. See the capture note.

A separate historical personnel loop fills in the data-sourcing layer around ExodusPoint’s early years. The original Business Insider report from October 2019 identifies Chris Petrescu as the then-head of data strategy and describes his remit as finding alternative datasets for portfolio-management teams. The report attributes a 15-person data team to Anil Chandroth, then head of data science and Petrescu’s former boss. CP Capital’s first-party biography independently describes Petrescu’s former ExodusPoint remit across quantitative and discretionary teams and his earlier WorldQuant data-strategy role. A July 2021 Business Insider report later reported Chandroth’s departure and placed the two leaders within a contemporaneous 158-person technology organization. These are dated historical disclosures, with some details attributed to sources close to the firm; they do not establish current ExodusPoint personnel, current architecture, GenAI use, dataset licensing, or investment authority. See the capture note.

The later personnel timeline is more specific but still bounded. A firm announcement for Stephen Luterman names him Chief Technology Information Officer and says he would lead technology, analytics, and data functions. A separate firm announcement for Peter Cotton names him Chief Data Scientist and cites quantitative analytics and data-science innovation in global markets. A third firm announcement for Shen Xu names him Head of Artificial Intelligence, cites AI, machine learning, and NLP expertise, and says he would enhance the Data Science Platform and build AI infrastructure. ExodusPoint’s current About Us page reports 718 total headcount and $14.9 billion AUM as of July 1, 2026, while the displayed senior roster lists a Head of Infrastructure but not a Chief Data Scientist or Head of AI. The announcements establish dated appointments; they do not prove current incumbency, reporting lines, model use, or GenAI deployment. The overall headcount is not a technology-team count. See the capture note.

Two newly recovered first-party recruiting routes expose additional architecture vocabulary. BlueCrest’s AI Cloud Platform Engineer posting places AI-platform delivery between Front Office AI Technology and Technology Infrastructure, with Azure and on-premise foundations plus AWS/GCP Gemini/Vertex AI enablement, enterprise AI SaaS integration, identity controls, DLP, monitoring, governance, and cost management. A second BlueCrest AI DevOps Engineer posting adds RAG pipelines, MLOps frameworks, GPU orchestration, Azure Foundry, Bedrock, agents, and MCP integration, and describes support for live systems critical to hedge-fund operations. RBF Capital’s Lead AI Platform Engineer posting describes a new data lake over market data, SEC filings, and alternative data, with ML analytics, natural-language interfaces, RAG, LLM fine-tuning, and multi-agent frameworks, plus a proposed three-to-five-person specialist team. RBF’s homepage separately says it transitioned from a hedge fund into a private family office in 2012, so it is not counted here as a current hedge fund. These are hiring-intent and entity-status records. They do not establish filled roles, provider adoption, live deployment, model performance, data rights, or investment authority. See the capture note.

The third sweep found that the hidden ownership layer often sits in product and control titles. Point72’s current public careers surface includes an AI Product Analyst, Market Intelligence role that owns AI products through approval, testing, deployment, support, adoption, and model governance. Its AI Validation Engineer, Macro Technology role connects AI-assisted development to regression packs, release gates, trading workflow validation, and human review. The same careers surface lists GenAI infrastructure, NLP/AI, and ML-related roles. These are current firm-controlled hiring signals; they do not identify the people hired or prove production permissions.

Citadel’s current GQS ML Researcher role explicitly names large language models, pre-training, fine-tuning, reinforcement learning, and deep learning in systematic-investing research. The GQS ML Engineer role adds distributed training, inference optimization, internal ML libraries, and research/production tooling. This moves the public evidence beyond generic “machine learning” hiring language, but it still does not disclose a named GenAI leader, model inventory, strategy-level use, or return attribution.

The geographic sweep adds three distinct records. High-Flyer’s official English page describes AI-focused quantitative trading, neural-network and NLP research, a deep-learning trading timeline, and the internally built Fire-Flyer platform. Ubiquant’s official site describes a technology-focused quantitative research firm using AI and advanced data analytics, but does not name an AI owner or disclose a GenAI system. M37 Management LP’s AI Engineer posting describes agents, Claude Code, MCP, multi-agent collaboration, evaluation, and investment workflow integration. SEC and LinkedIn records corroborate M37’s identity, but the job posting is university-hosted rather than firm-controlled, so it remains an emerging-manager role-led signal. M37 is distinct from Move37 Capital.

Finally, Trexquant’s current careers page supports an ML-for-alpha-discovery and portfolio-construction signal, not a current public GenAI or agent-ownership claim. This distinction matters: title-blind searching finds more evidence, but it also creates more opportunities to mistake research methodology, hiring intent, and control-plane ownership for a disclosed live system.

The regional sweep also verified a distinct China-based AI-quant route for Baiont Quant. In a May 20, 2025 Financial Times interview, founder and CEO Feng Ji describes a roughly 30-person team, an internal foundation model, short-horizon predictions from trading data, dynamic trade combination, and an organizational approach that treats factor discovery, signal generation, modeling, and strategy work as one machine-learning problem. The interview also records claims about a large compute-per-researcher ratio and 13 programming-contest gold medalists. Baiont’s first-party company site corroborates the AI-driven quantitative-hedge-fund identity and computer-science/AI talent positioning. These are unusually specific public statements, but they remain interview and company claims: no independent model card, training corpus, AUM audit, performance replication, data-rights record, or portfolio-permission map was retrieved. The public record does not establish whether Baiont’s foundation model is generative or whether it is used across every investment stage. See the capture note.

August 14 extended loop: emerging AI-native formation signals

The extended loop found a separate public layer below established-manager disclosures: newly formed AI-native investment platforms and vendors founded by former hedge-fund practitioners. These records are useful for tracking personnel migration, workflow vocabulary, and the types of systems being proposed to funds. They are not merged into the established-manager universe.

KelAI. Its Y Combinator profile and official site describe an autonomous research loop covering idea generation, data analysis, code, backtesting, validation, live monitoring, and PM feedback. The founder is publicly described as a former WorldQuant systematic-equity portfolio manager and former Millennium ML lead. The profile also claims institutional deployment and signals running since October 2025. KelAI’s SEC Form D identifies Jeremie Cohen as CEO and reports a $4.975 million offering sold to 14 investors as of July 16, 2026. That filing supports a corporate exempt offering, not a regulated hedge-fund vehicle or a performance record. The client, signal history, model inventory, and deployment permissions remain undisclosed.

WithAI. Its Y Combinator profile describes Multiplier as a governed command center connecting structured and unstructured data, tools, files, secure inference, changelogs, ontology maintenance, and firm knowledge. The same profile identifies a co-founder as a former Bridgewater investor and builder of LLM investment systems. WithAI’s own site says it is backed by angels including Bridgewater co-CIOs Greg Jensen and Karen Karniol-Tambour. These are company-published personnel and sponsorship claims, not independent confirmation of Bridgewater endorsement, customer identity, or client-level deployment.

Kimpton AI. Its Y Combinator profile describes a platform built inside a quantitative systematic fund that ingests a mandate, positions, and transaction data, generates structured trade proposals, and supports adversarial review while leaving decisions with portfolio managers. The founders are described as former Goldman Sachs and Vistra engineers who co-founded Level III Capital. The claimed institutional client and asset exposure are not independently verified; Kimpton is retained as a platform signal rather than classified automatically as a hedge fund.

Fifth Era AI Access Fund I. This is a different formation pattern: a pooled investment vehicle rather than an AI trading platform. Its March 2026 SEC Form D reports $4.85 million sold to 27 investors and names Fifth Era entities as managerial parties. Fifth Era’s public thesis focuses on private companies at the convergence of Internet, AI/Agentic, and Blockchain technologies, while its team page identifies a Partner for AI Investments. The public record supports an AI-focused venture/access vehicle, not a systematic hedge-fund peer or evidence of model-driven trading.

Ai Funds High Conviction US Equity AI-Managed ETF. SEC materials provide an adjacent registered-product control case. The prospectus describes a BAILA (Bayesian AI Learning Algorithm) engine for high-conviction US-equity selection, macro-regime forecasting, and risk-aware portfolio construction, and identifies Milliman as sub-adviser for selected Ai Funds products. This is not hedge-fund evidence; the public filings do not provide model weights, training data, independent evaluation, or AI-attributed performance.

Standard Signal. Its Y Combinator profile positions a new AI-native hedge-fund formation effort around end-to-end research and execution, with a founder who previously built Phind and trained language models. A dated Form D record for Standard Signal, LP reports $0 sold and 0 investors as of the June 22, 2026 filing. That is a dated financing snapshot, not evidence that no later capital was raised. The current company site adds explicit claims about reasoning-model training with reinforcement learning, autonomous execution inside infrastructure-enforced risk limits, logged decisions, a market-neutral strategy, and ML Research Engineer and Quantitative Researcher openings. The site also claims a live Sharpe ratio above 3, but publishes no underlying return series, model artifacts, broker/order evidence, data rights, or independent audit. Those claims remain company positioning; the capture note keeps the formation, recruiting, financing, and performance boundaries separate.

The full evidence ledger is hedge-fund-ai-emerging-formation-signals-2026-08-14-raw.md.

The next title family is Head of Automation. The exact title is more common in trading venues and market infrastructure than in hedge-fund personnel pages, so the object of automation matters. Man Group’s current AHL Risk Automation Lead role connects automation and AI adoption to investment-risk workflows, Python tools, and LLM integration. Point72’s Security Automation Lead connects automation to auditable security controls. State Street’s Nick Delikaris appointment connects the title family to algorithmic trading, business-process engineering, analytics, and platform services. These records suggest that automation leaders can be the bridge between investment or risk teams and AI infrastructure, but they do not by themselves establish GenAI ownership or autonomous investment authority.

Each source addresses a different operating question. GMO’s public research frames the economics of AI and systematic signals. Acadian describes how AI modules connect to the investment workflow. Arrowstreet specifies a permissioned platform for models, tools, and agents. Man AHL describes agentic strategy-design workflows. AQR documents validation constraints relevant to these claims.

Key Data Points

The figures below are source-specific indicators, not comparable measures. They describe different populations, dates, definitions, and levels of verification; the article does not convert them into a score or ranking.

Date	Data point	Source and credibility
2025-12-31	GMO Form ADV reports approximately $76.8B in discretionary AUM and discusses restrictions and risks around third-party AI tools	GMO Form ADV; primary regulatory disclosure
2025	GMO systematic research describes Value, Momentum, Alerts, ML sentiment, network-aware momentum, and a 2026 agenda including AI-based analytical tools and uncertainty estimates	GMO Systematic Equity Year-End Letter; primary firm research
2026-03-19/20	QuantVision 2026’s public agenda brings together named speakers associated with Citadel, Man Group, Two Sigma, Millennium, DRW, and BlackRock around ML architectures, data, multimodal signals, agents, and investment-management AI	Fordham publication, Fordham recap, and published agenda; event and roster metadata, not recording-backed firm disclosure
2025-07-28	A named Acadian practitioner describes modular AI signals across earnings, analyst bias, Q&A, newsflow, fundamentals, technical patterns, expected returns, portfolio construction, and order routing	Fear & Greed interview on AI in investing; named-practitioner interview; medium-high
2026	Acadian’s systematic-edge page reports 100+ investment professionals, 95+ advanced degrees, 35+ years of data, 620M+ daily observations, and 65K+ traded assets	Acadian systematic edge; primary firm page; self-reported scale
2026	Acadian’s public Investment AI Engineer role calls for agentic investment workflows, reusable skills, human-in-loop controls, testing, and adoption/quality/time-to-value metrics	Acadian LinkedIn job listing; primary hiring signal
2026	Arrowstreet’s public AI-platform role calls for Bedrock, model routing, MCP, RAG, agents, usage/cost dashboards, and default-deny execution controls	Arrowstreet Workday listing; primary hiring signal
2026	Arrowstreet’s LinkedIn company page displays 539 employees and concurrent openings for AI platform, AI security, data platform, quantitative research, and quantitative development roles	Arrowstreet LinkedIn jobs page; current public hiring surface, not a headcount audit
2026	A named public recruiting post seeks a Senior AI Platform Engineer for “our team”; another employee post describes hiring across AI Platform and AI Security	Florent Monthel post and Amy Wong post; employee recruiting signals, not proof of ownership or production
2026	Florent Monthel’s public GitHub repository describes production-target classification for Claude Code, deny-write/allow-read policy, human approval, MCP hooks, inventory APIs, caching, and OpenTelemetry/SIEM audit paths	claude-code-pd-protection; public personal technical artifact; no observed Arrowstreet credentials or firm-specific data
2025	Man AHL publicly describes GenAI as a way to accelerate trend-following strategy design	Man AHL; primary firm article
2026	CFM’s public approach page reports AI/ML/cloud integration with 12+ petabytes of data and pre-deployment testing/piloting; its public job listing describes Prediction Services and GenAI agents for code/model workflows; a firm-hosted systematic-macro interview adds text/video/audio extraction, network/classification structures, and agentic test/code language	CFM approach, Prediction Services listing, and systematic global macro interview; firm page, hiring signal, and practitioner interview
2024-12-03	CFM reports LLM-assisted labeling and compact-model fine-tuning for financial NER, with F1 moving from 87.0% to 93.4%	CFM case study; company-reported, task-specific, not investment reasoning
2026	Balyasny’s OpenAI case study reports centralized Applied AI, 12+ model-evaluation dimensions, GPT-5.4 plus internal models, scoped agents/tools, and approximately 95% investment-team use	OpenAI case study; vendor/customer-reported, no independent performance audit
2026	BlackRock’s current AI Labs role describes generative AI, production deployment, alpha generation, operational efficiency, and responsible deployment; the firm’s official page names Rachel Schutt and Stephen Boyd as co-heads	AI Labs role and AI Labs; firm-owned hiring and organizational evidence
2026	BlackRock’s official Asimov disclosure describes a proprietary equities-research tool; its Systematic Investing page describes LLM use for macro narrative, analyst views, consensus, and tradable signals	Asimov disclosure and Systematic Investing; firm-reported product and workflow claims
2026-07-02	BlackRock engineering speakers at AI Engineer describe investment-operations knowledge apps, a sandbox and app-factory architecture, extraction templates, validation/QC checks, access and cost controls, and human-in-loop review	How BlackRock Builds Custom Knowledge Apps at Scale and local caption source record; practitioner conference testimony, not an independent ROI, model, security, permission, or deployment audit
2026	WorldQuant’s dated Portfolio Manager, Agentic Systems role artifact described live-book management, planning, tools, memory, reflection, collaborative reasoning, RL tuning, custom agent workflows, and human-in-the-loop checks; the direct URL no longer resolved as a live vacancy on 2026-08-10	WorldQuant role artifact; archived/dated hiring evidence, not proof of a filled role or deployment
2026	WorldQuant’s official leadership page identifies Paul Griffin as Co-CIO and Chief Science Officer and says he leads the firm’s AI initiatives; a May 2026 article says AI is deployed across teams while humans retain accountability	WorldQuant leadership and AI perspective; current firm-controlled leadership and firm-claim evidence, not model or live-book attribution
2026	WorldQuant’s current careers surface lists AI Scientist, AI software, WQBRAIN AI research, deep-research, and LLM/AI-agent roles; QRT lists AI Platform Engineer/Developer roles; Tower describes ML and predictive-signal engineering	WorldQuant career index, QRT jobs, and Tower careers; hiring and firm-posture signals with no common deployment measure
2026	Voleon’s official management page names current leadership tied to predictive models, trading systems, portfolio optimization, market impact, and technical staff	Voleon management and overview; first-party predictive-ML and personnel evidence, not GenAI deployment evidence
2026-08-12	Adjacent control pages add first-party AI evidence: LLM agents/API-HPC serving and usage figures at Jump; HRT AI Labs and deep-learning trading constraints; Optiver AI Lab LLM-based analysis language; DRW AI/RAG/fine-tuning hiring; Capstone AI/LLM/RAG/MCP hiring; and IMC agentic developer-workflow engineering	Jump AI/ML, HRT Machine Learning & AI, Optiver research, DRW AI Engineer, DRW Data Developer, Capstone AI Engineer, Capstone AI Infrastructure Engineer, and IMC AI Powered Engineering; first-party operating or hiring evidence, not model inventory, trading authority, or return attribution
2026-08-12	Numerai’s public surfaces describe an AI-scientist workflow, an 8-billion-parameter Predictive LLM trained on more than 1 million articles for Faith dataset features, open agent skills, MCP research/submission tools, scoped API keys, and AI Scientist hiring	Numerai homepage, NumerCon 2026 recap, Faith release, Numerai MCP docs, example scripts, and AI Scientist role; platform and hiring evidence, not autonomous firm-capital authority or return attribution
2026-08-13	QRT’s official Greenhouse API adds production-support language for front-office trading flows and LLM-based systems; regional sources add High-Flyer AI compute/fund separation, Lingjun research-agent and offshore context, Mingshi AI-factor pipeline, QTS methodology plus Delaware/British Columbia entity separation, and an IQuest/Ubiquant technology-affiliation model-code signal	QRT role, QRT Greenhouse API, High-Flyer, Lingjun AI article, Lingjun Hong Kong article, Mingshi, QTS article, QTS SEC Form D/A, QTS LEI, IQuestLab, and IQuest-Coder; first-party/regulator/public-code signals, not deployment audits, permission maps, fund-boundary proof, or AI-attributed performance
2026-08-13	Mainland China regional sources add DeepWin / 蝶威量化 entity and research-agent workflow evidence: official P1070101 identity, AI quantitative-investment/RL/Agent process language, five-role DeepWin Agent research workflow, broker-report factor extraction, code/backtest execution, performance judging, and fund-manager review vocabulary	DeepWin official page, Chinese Securities Journal event report, and Chinese Securities Journal industry survey; firm page plus media reports of a firm-hosted event, not independent deployment, permission, product/fund-boundary, or performance proof
2026-08-14	WorldQuant’s firm-hosted career index currently lists LLM/agent, WQBRAIN, AI software, and AI/ML systems roles; Point72/Cubist adds a Hong Kong/Singapore AI Data Scientist role for systematic data workflows	WorldQuant career index, WorldQuant LLMs & AI Agents internship, WQBRAIN AI Researcher, WorldQuant AI Software Developer, WorldQuant Senior AI Software Developer, WorldQuant Lead Python Engineer, AI/ML Systems, and Point72/Cubist AI Data Scientist; current official hiring surfaces, not filled-role, deployment, permission, live-book, or performance evidence
2026-08-14	Antipodes adds a current Australia-specific agentic-investment hiring signal; WizardQuant adds a firm-controlled Chinese-language AI-lab strategy article beyond its careers pages	Antipodes careers and WizardQuant AI-era indexing article; official pages with high confidence for hiring/lab-positioning language, not proof of filled roles, production catalogues, model inventory, portfolio permissions, regulated-product boundaries, independent evaluation, or AI-attributed returns
2026-08-14	XTX’s official careers API adds XTY Labs foundation-model/agent-prototype hiring and a separate ML Performance and AI Acceleration role; Aspect adds a mainland-China legal-entity and AMAC-registration signal	XTX AI Research Internship - XTY Labs, XTX Machine Learning Performance Engineer, XTX official careers API at https://api.xtxcareers.com/jobs.json, and Aspect AMAC registration press release; official hiring and firm press-release evidence, not filled-role, model inventory, deployment permission, autonomous trading, AI-attributed return, or GenAI-deployment proof
2026-08-15	Connor, Clark & Lunn Investment Management adds a Canada-based quant-equity operating signal: an official Quantitative Equity Data Science role names ML/AI, NER, knowledge graphs, integrated data-model evaluation, data-quality testing, alpha-research support, and production deployments inside a Vancouver Quantitative Equity Team	CC&L Quantitative Equity Data Science role; official hiring and operating-surface evidence, not filled-role, model-inventory, LLM/agent-runtime, permission-map, live-trading-authority, or AI-attributed-performance proof
2026-08-15	Longqi Scientific Investment adds a current mainland/Hong Kong regional signal: official pages identify Hangzhou Longqi Scientific Investment, Longqi Scientific (Hong Kong) Limited, AI/ML techniques described as trading since 2018, fund-family categories, and SFC Type 9 language while preserving an unresolved CE-number discrepancy	Longqi official page; firm-controlled evidence, not direct SFC-register extraction, model inventory, deployment audit, permission-map, fund-level product documentation, independent performance, or AI-attributed-return proof
2026-08-15	RQI Investors adds an Australia-based active-quant equities signal: official pages describe a June 2025 long-short fund launch with UniSuper seed capital, AI/ML in the systematic process, and an AI paper covering NLP, topic modelling, adaptive trading algorithms, and portfolio optimisation	RQI long-short fund launch and RQI AI paper page; official manager communications, not model identity, production map, permission, audited signal contribution, investment authority, or return-attribution proof
2026-08-15	Avangard Investments’ Australia pages add a named-system and entity signal: A.L.F.R.E.D. is described as the proprietary in-house AI/ML and portfolio-optimization system for the Avangard Systematic Australian Equity Fund, with firm-reported data scale and a July 1, 2026 fund-structure transition	Avangard performance, FAQ, governance, and people; firm-controlled evidence, not independent ASIC extraction, model design, live GenAI-agent, permission, performance, or return-attribution proof
2026-08-15	RAM Active Investments adds a Switzerland-based adjacent systematic-manager control: official research pages describe LLM-based newsflow representations, fine-tuned LLMs for stock-return prediction, and deep-learning infrastructure with 500+ alpha inputs; ACL Anthology verifies the 2024 EMNLP Industry paper record	RAM research, RAM company page, and EMNLP Industry paper; firm-controlled and public paper evidence, not live-portfolio, model-weight, production-stage, independent-performance, or return-attribution proof
2026-08-15	T. Rowe Price Integrated Equity adds an adjacent systematic-research control: its July 2026 note describes LLM-assisted business-quality scoring, analyst/LLM prompt refinement, internal-database integration, hallucination-reduction controls, weekly batch scoring, software-resilience analysis, and look-ahead-bias/overfitting caveats	T. Rowe Price systematic AI research note; official firm research evidence, not model-vendor, model-weight, portfolio-permission, production-adoption, audited-signal-contribution, investment-authority, or return-attribution proof
2026-08-16	Mackenzie Investments adds a Canada-based quant-equity signal: official materials say the Global Quantitative Equity Team uses ML, NLP, LLMs, and cloud computing with a human overlay; applies LLMs to earnings-call sentiment; uses NLP on alternative data; and uses non-linear ML plus native-language NLP for emerging-market company estimates and financial-statement reading	Mackenzie 2026 Market Outlook and How Arup Datta uses AI to help track volatile emerging markets; official manager materials, not model inventory, permission-map, autonomous-trading, independent-performance, or AI-attributed-return proof
2026-08-16	Pictet Asset Management’s Quest AI-Driven franchise adds an adjacent systematic-manager control: official pages describe proprietary AI stock selection, forecast attribution, weekly rebalancing with human oversight, quarterly retraining, rigorous data governance, a one-month forecast horizon, boosted trees, and IBES data use	Pictet Quest AI article, Pictet governance article, and Pictet practitioner page; official manager pages, not model weights, full feature list, live permission map, independent performance audit, or AI-attributed-return proof
2026-05-19	Anthropic describes AI-researched, backtested, and proposed Man signals as running real capital, with governed skills and a shared data layer	Code with Claude session; vendor-conference claim, no public P&L or linkage to a named Man system
2026	CFM reports a narrow financial named-entity recognition fine-tune improving F1 from 87.0% to 93.4%	CFM/Hugging Face case study; company-reported, not independently replicated
2026	G-Research publicly describes on-prem open-model serving, centralized MCP, governed tools/data, and secure autonomous-agent sandboxes	G-Research Core AI Engineer; primary hiring signal
2026	Point72 roles name LLMs, agents, RAG, open-source fine-tuning, post-training, benchmarking, and compliance-approved AI products; Cubist roles name deep learning, NLP, and research-to-evaluation workflows	Point72 NLP/AI Engineer, Fundamental Equities AI Engineer, and Cubist ML Researcher; current hiring signals
2026-05-12	Schonfeld describes FE AI Lab workflows, SchonAI, Anthropic/OpenAI partnerships, structured pilots, and downside review; current roles add MCP, agent evaluation, business-embedded Fundamental Equity AI integration, and a shared AI plugin and skills library	FE AI Lab, Senior Software Engineer, Software Engineer - Fundamental Equities, and Quantitative Developer - Fundamental Equities; firm article plus hiring signals
2026	G-Research describes production LLM code review with source-of-truth validation, bounded recovery, provider abstraction, cost telemetry, and non-blocking CI comments; its firm-owned Robocop repository adds a public code artifact for automated LLM code review; separate roles name domain-adaptive pretraining, LoRA, RAG, and multi-agent orchestration	Code-review article, Robocop, NLP Researcher, and Applied AI Engineer; firm-owned sources
2026-08-11	Two Sigma official roles describe LLM/NLP production pipelines, automatic signal mining, agentic research libraries, LLM-based featurization, and post-training research for financial time series	Generative AI role, Techniques Engineering, Modeling Data Scientist, and Post-Training Research Scientist; official hiring signals
2026-08-11	Millennium publicly names a Global Head of AI, investment-team agentic-tool personnel, agentic infrastructure, internal search, end-user-built applications, and approval/spot-checking controls	Technology page, Gideon Mann interview, Daniel Tymecki profile, Applied Cloud and AI Engineer role, and privacy policy; firm-owned sources
2026	Citadel’s public record combines an internal equity-research assistant, an academic-finance research-validation agent account, current LLM/fine-tuning hiring, and explicit human-judgment boundaries	Reuters NEXT post, GQS ML Researcher, and investment-decision article; mixed firm/media evidence
2026	D. E. Shaw’s current careers surface lists Applied AI, AI vendor-tools, and ML research roles; DESIM asks portfolio-strategy applicants for comfort applying GenAI tools; a Fundamental Equities AI Product Analyst role adds analyst/PM process mapping and agentic workflow redesign	D. E. Shaw careers, DESIM role, and Fundamental Equities AI Product Analyst; primary hiring signals
2026	PDT lists an Applied ML Scientist role and describes peer review, empirical validation, and deployment to automated trading systems; no public GenAI linkage was found	PDT work and Applied ML Scientist; primary firm/careers evidence
2025-07-15	Aspect’s cofounder discusses AI, ML, and LLMs in systematic investing; firm material describes feature selection, repeatable evaluation, and financial-context controls	J.P. Morgan podcast page and ML challenge; practitioner and firm sources

These figures are not comparable measures of AI capability. The figures are generally firm-reported operating scale or role requirements. None is an independent measure of model quality, alpha, or production reliability.

August 20 official YouTube channel expansion

The platform-specific pass recovered official YouTube channels that were not represented in the earlier podcast-led queue. CFM’s channel exposes 14 videos, including a 2026 Philip Seager Milken recording, a Columbia alternative-data initiative video, and older first-party Machine Learning material. A caption check found no public English captions for the 2026 Milken recording. The 2019 Columbia video’s automatic captions place the CFM–Columbia alternative-data initiative, seminar/workshop, and satellite-derived inventory example at approximately 00:28–01:56; this is historical research-culture evidence, not a current dataset, model, collaboration, or investment-result disclosure. The capture note records the caption status and boundaries.

Balyasny’s official channel contains 70 indexed videos. Titles and descriptions add several title-blind routes: investment-thesis development, the quant mindset, technology teams for complex markets, technology at BAM, and quantitative-researcher/portfolio-manager collaboration. These are first-party recruiting and practitioner-media signals. They expand the personnel and workflow search surface but do not establish systemwide model ownership or live capital authority.

Bridgewater’s official channel returned 100 entries from a capped inventory. Relevant routes include A Critical Time for AI, Global Outlook: CIOs on AI, AI Today and Tomorrow, and Inside Bridgewater’s Research Process. The recent AI video exposes public English automatic captions and an authorized alternate player route allowed a local ASR audit, but the transcript contains proper-noun uncertainty and is not used here for verbatim claims. The video route is therefore recorded as a high-confidence caption-backed media route with an automatic-transcription caveat, not as independently validated system evidence.

Two Sigma’s official channel contains 27 indexed videos, including the historical Machine Learning Models of Financial Data, Deep Learning for Sequences in Quantitative Finance, and the third-party Anyscale talk by Two Sigma’s Head of AI Core. The two first-party recordings expose English automatic captions and were caption-audited on August 20, 2026. The 2022 presentation introduces Justin Sirignano and discusses recurrent neural networks for price moves and reinforcement learning for optimal order strategies (price-move route, order-strategy route); the 2021 David Kriegman presentation frames sequence models around stock prices, alpha modeling, and buy/sell order decisions (alpha-modeling route, order-decision route). These are dated firm-owned educational and research-context disclosures, not current model cards, production permissions, live-book evidence, or performance attribution. They should remain time-scoped and separate from current hiring and AI-outlook pages.

GMO LLC’s official channel adds a 16-video strategy archive covering climate, ESG, emerging-market, event-driven, quality, resources, alternative-allocation, and value-investing topics. Selected routes include GMO Quality Strategy, GMO Alternative Allocation, GMO Resources Strategy, and Shades of 2000. This is a firm-controlled personnel and strategy-media surface; it does not disclose a GenAI model inventory or connect these explainers to a production AI system.

Acadian adds two title-blind public video routes. The Fear & Greed discussion, dated July 28, 2025, is a video companion to the already tracked AI-and-investing podcast route. The Alpha Exchange interview with Owen Lamont, dated March 11, 2025, omits AI and hedge-fund terms from its title but exposes an Acadian portfolio-manager role and MIT economics PhD/academic-finance lineage in its public metadata. Caption review adds a bounded market-view signal: Lamont discusses AI-linked corporate capex and valuation interpretation (18:10, 24:54, 25:25). This remains third-party interview evidence, not a disclosure of Acadian’s GenAI architecture, budget, or deployment.

The search also produced a same-name Arrowstreet YouTube channel whose description identifies an architecture firm. It was disqualified from Arrowstreet Capital coverage after checking the channel identity and linked subject matter. The full channel map, capture status, and false-positive control are in the official quant-firm YouTube expansion source note.

The GMO, Acadian, and Arrowstreet core-firm checks are recorded in the core-firm YouTube expansion note. One identity-sensitive Acadian search result was withheld under the public-redaction rule and is not used as evidence here.

September 4, 2026 — title-blind hiring routes

Recent hiring searches add implementation detail that does not appear in an “AI lab” label. Point72’s NLP Engineer posting is an employer-controlled early-career role in Systematic Investing. The posting describes an NLP engineer joining a quant-research team, using LLMs, agents, and RAG; combining internal and external textual datasets; formulating research hypotheses “to derive alpha”; and building GenAI solutions with internal models and external APIs. This is a public hiring specification. It does not identify the person hired, a model, data rights, production permissions, or investment results.

M37 Management LP’s AI Engineer posting, hosted by Yale’s Office of Career Strategy, describes a Menlo Park long/short equity hedge fund focused on AI-related companies. The advertised role covers autonomous finance agents for investment research, financial-data and news analysis, and workflow orchestration; named tools and capabilities include Claude Code, OpenClaw, document parsing, web research, data extraction, code generation, multi-agent collaboration, benchmarking, real-time monitoring, and web crawling/scraping. The posting also names security, compliance, and ethical-AI requirements. Because this is a university-hosted posting and not a firm-controlled technical disclosure, it remains a role-led signal. It does not establish that the system was built, deployed, or granted investment authority, and M37 is distinct from Move37 Capital.

An adjacent asset-management route is Société Générale Haussmann Management Japan’s Structured Portfolio Manager posting, with a LinkedIn mirror. The indexed official text, published May 7, 2026, combines derivative- and index-linked portfolio management with Python automation, triggers linked to calendar events, portfolio exposures, and flow events, modular strategy logic, execution checks across email/Bloomberg/FIX, dashboards, and model-governance controls. It asks candidates to use Copilot/AI tools for scripting, documentation, workflow generation, recurring code tasks, and LLM-assisted code verification. The official page returned 410 when rechecked, while the LinkedIn copy preserved the same text. This supports a dated hiring-intent and workflow-design record, not a claim about a filled role, selected provider, or live system.

These routes reinforce a search rule for the wider universe: inspect systematic investing, NLP, market-intelligence, validation, platform, automation, portfolio, and product-control titles even when the title does not contain AI. Hiring language can reveal intended workflows and control surfaces, but it cannot by itself establish model ownership, investment authority, firm-wide adoption, or performance. The capture note records the source classes and access caveats.

August 20 conference-replay and event-route expansion

The official GMO 2025 Conference page adds a first-party replay route that does not appear in the YouTube channel inventory. Its “Quality Investing” session spotlight names Tom Hancock, Head of the Focused Equity team, and describes a discussion of GMO’s fundamental approach to the opportunities and risks of investing in AI. The replay is contact-gated, so the public page supports session metadata and personnel linkage only—not a transcript, model inventory, or production-system claim.

A title-blind CFA Society Houston event page records a January 21, 2026 Zoom webinar with Ben Inker, Co-Head of GMO Asset Allocation, on AI as an investment-bubble thesis and on other investable areas. The page also records his quantitative-equity and asset-allocation history at GMO. It is a dated event and personnel route, not a recording transcript or evidence about GMO’s internal AI systems.

The route map and capture boundaries are recorded in the GMO conference and event-video expansion note. These additions expand discovery coverage; they are descriptive records, not a cross-firm capability ordering.

August 20 title-blind regional media follow-up

The South African title-blind pass adds Ghost Stories #109, a 30 July 2026 publisher transcript and video conversation with Reza Fakie, identified as portfolio co-manager of Old Mutual Investment Group’s Global Managed Alpha Fund. Fakie describes a multi-factor global-equity process, benchmark-relative portfolio controls, continuous testing/backtesting, and a preference for understandable models. The AI discussion is more concrete about workflow than about autonomous investment: he describes using Claude and Claude Code for programming, database queries, analytics, and commentary review while retaining subject-matter verification of outputs. The recovered YouTube captions provide timestamped navigation, including the AI workflow discussion around 30:14–32:49. This is a regional publisher and speaker-reported account; it does not disclose Old Mutual’s model inventory, training data, permissions, production endpoint, or independently verified AI-attributed performance. See the capture note.

The AIMA China Quant Spotlight, held in Hong Kong on May 6, 2026, adds an official regional conference and personnel map. Its agenda says AIMA introduced a China Quantitative Investment, AI and Big Data Task Force and names William Ma as Global CIO of GROW Investment Group, Ming Xu as Quantitative Research Director at Qube Research & Technologies Hong Kong, Jimmy Fan as Winton’s Research Director for Greater China, Shen Yi as Chairman and CIO of Shenyi Investment, and managers and academics from Frontier Asset Management, OP Investment Management, and the University of Edinburgh. The agenda also includes a panel titled “Systematic Methodologies in Quant Investing: From China to Global Markets — AI, and the New Rules of Quant Investing.” The page exposes no replay, transcript, speaker remarks, model inventory, or performance evidence, so this is an event and discovery-network record rather than a substantive firm-system disclosure.

The title-blind regional sweep also surfaced several new personnel and media routes. In Korea and Greater China, an S&P Global Korea Quantitative Investment Conference agenda names Tianyan Capital founder/CIO Chris Xie and a session connecting systematic finance with neural-network chess systems; a Tech42 Korea interview adds SQR founder/CEO KyuYul Lee; an ETMarkets interview and AlphaGrep profile describe Bhautik Ambani and public AI-driven systematic-strategy positioning; and a PMS Bazaar transcript identifies Estee Advisors VP and Head of Investments Vivek Sharma while describing AI-assisted interview analysis, data preparation, risk rules, and algorithmic portfolio-management workflows. These are agenda, interview, and firm-positioning sources, not independent evidence of model inventory or AI-attributed performance.

The Chinese-language pass adds Luoshu Investment interview coverage, a Peking University lecture announcement for Jiukun’s Xi Li covering ML/AI applications in factor mining, signal combination, and portfolio construction, and a Peking University/The Paper reference to an intentionally unnamed Blackwing AI basic-research lab head. Mengxi coverage describes founder/general manager Li Xiang, an AI Lab established in early 2025, and AI-related work across factor research, portfolio optimization, execution, and risk; Douyin and Bilibili add a person-level Mingshi/Cai Xian route. A separate 2020 Mingshi interview now has recovered Chinese audio and local segment-timestamped ASR: the speaker describes scaling from roughly 20 to nearly 1,000 factors, AI-assisted factor combination, and a dated comparison between AI-added and non-AI factor logic (00:38–05:50). These are translated, speaker-reported claims from a third-party historical video, not an independent evaluation or current deployment record. The unnamed role remains unnamed and no current deployment conclusion is drawn.

The Block Central Chinese-language episode, mirrored on Bilibili, now has a recovered DASH audio stream and local Chinese ASR with timestamps. Charles Shang identifies himself as BCY Labs CTO and describes a “Transparent and Verifiable AI” framing, factor decay, text/sentiment factorization, an agent/harness coordination layer, and a proposed multi-agent sandbox for simulating interacting actors (01:45–02:19; 08:03–08:39; 26:23–27:17; 37:46–39:17). He characterizes present industry use as more often AI-assisted data analysis and factor extraction than autonomous capital allocation (21:43–22:45). This is a regional practitioner discussion and BCY Labs affiliation record, not evidence of a covered hedge fund’s production system, model inventory, data rights, permissions, or performance. See the recovery note.

Canada, Australia, and New Zealand produced additional title-blind routes outside conventional hedge-fund wording. Scotia Global Asset Management identifies Dan Yungblut and describes InvestIQ as an internal research-assistance tool; a Thinking Ahead Institute fireside chat names RBC Global Asset Management’s Jaco van der Walt and discusses AI, agents, governance, and quantitative research. A Desjardins post describes dynamic-factor, portfolio-construction, and financial-plus-sentiment prediction projects with Concordia and HEC Montréal researchers, while a Manulife AI research-analyst hiring post describes co-development of AI research tools with investors. These sources establish public roles, collaboration, or hiring intent—not filled headcount or production permissions.

The same pass found Fidelity Australia’s AI-investing webinar, Australian Retirement Trust’s webcast archive, and an Avangard person-level LinkedIn route. In New Zealand, Milford’s podcast/transcript discusses AI agents and commercial returns, while the Guardians of New Zealand Superannuation FY2022–23 annual report records a historical AI-powered New Zealand-equities portfolio-manager proof of concept that used value, momentum, quality, and volatility inputs and went live with small initial capital during that reporting period. The latter is explicitly historical and does not establish current status.

An additional title-blind HFR Podcast episode, published July 16, 2026, identifies Stephan Kessler as Morgan Stanley’s Managing Director and Global Head of Quantitative Investment Strategies Research. Kessler describes using AI as a research collaborator, with multiple agents running parallel questions inside a constrained workflow: a pre-coded backtester, point-in-time and properly lagged data, transaction-cost assumptions, and a pre-coded portfolio-construction method. He also describes extracting sentiment and changes from annual reports for systematic tests, while retaining human verification. This is a firm-affiliated practitioner account and a useful control vocabulary for research-agent design; it does not establish a named production model, live trading authority, model inventory, or performance attribution at Morgan Stanley or any hedge fund.

An ICLR 2025 expo-panel abstract adds a distinct ADIA quantitative-research route. Under the title “Bridging Specialized ML Research and Systematic Investing,” the page names Mathieu Tolle and Gautier Marti and describes a quantitative R&D team using LLMs, retrieval-augmented generation, agent-based systems, variational autoencoders, graph neural networks, and multimodal signal processing for dataset curation, feature extraction, trading-signal construction, and investment-hypothesis generation. The abstract also proposes interacting ML-agent ecosystems as a research direction. This is an event abstract and team self-description, not a recording or a deployment inventory: it does not establish current speaker titles, model ownership, datasets, permissions, evaluation, order authority, or performance. The capture note keeps those boundaries separate from the existing ADIA personnel interview.

The title-blind sweep also recovered a January 20, 2026 Blushing Quants interview with Oren Tapiero. The publisher calls him a quantitative researcher at “Tidal,” while a public profile, role listing, and Tydal’s first-party site point to the Israeli digital-asset algorithmic-trading company tydal; the publisher spelling is retained as historical metadata. In the local automatic ASR, Tapiero describes a digital-asset systematic-trading technology company, a fully automated momentum strategy he says has been live since mid-2021 and manages millions of dollars (01:05–01:42), and a research process organized around target selection before feature/model choice (14:48–17:38), parsimonious features (30:42–32:21), causal reasoning (41:31–47:08), and walk-forward, regime-aware testing (55:07–59:58; 01:03:26–01:03:50). Tydal’s site separately describes a proprietary momentum algorithm, adaptive models, and multi-layered risk management across digital assets. These are speaker-reported and company-reported statements, including automatic rather than speaker-verified transcript material; they are not an independent performance audit, model inventory, permission map, or proof that every described method is deployed. See the capture note.

The UK and continental-European search added an upcoming CFA UK Borealis Global Analytics webinar, an Investment Association agenda entry for 3AI, a French-language AFG/Institut Louis Bachelier GenAI event, Smart Wealth’s AI-process and performance surfaces, and a Quoniam event post covering practical AI, automation, and research agents. A media-integrity recheck found that the Investment Association page’s embedded YouTube player resolves to a 130-second video titled “INVESTMENT ASSOCIATION ANNUAL CONFERENCE 2025”, uploaded July 18, 2025; it is not evidence of a 2026 session recording. These remain event, vendor, or company self-description surfaces; they should be read as source-specific records rather than a cross-firm ordering. See the video-integrity note.

The same European watchlist now includes CFA Society Netherlands’ Artificial Intelligence in Risk Management event, scheduled for October 1, 2026. Its programme names Marcel Prins, former COO of Robeco and APG-AM; Peter Strikwerda, former global head of digitalization and innovation at APG-AM; and representatives from Zanders, Bunq Bank, DNB, and AFM. The stated agenda covers current AI, generative AI, and agentic AI in risk management, including governance, reliability, scaling, controls, and decision-making. This is an upcoming event and personnel-discovery surface: no recording, transcript, model, deployment, or investment result was public on the check date. See the capture note.

The official GAIIM 2026 programme adds another upcoming, title-blind event route. The hybrid conference is advertised for September 29–30, 2026 at Columbia University’s Faculty House in New York. Its preliminary agenda includes agentic buy-side workflows, Claude Code and Cursor, agentic equity analysts, MCP and connected research stacks, risk/compliance/operations/governance, and an Anthropic session on agents in financial services. The public speaker inventory lists Pete Petersen as CTO of Causeway Capital Management, Michael Soss as CIO of Millburn, and George Ho as Managing Member of NIV Asset Management, alongside finance-data and research vendors including Daloopa, Hudson Labs, Canary Data, and AlphaSense. This page establishes an advertised event, agenda, and listed roles; it does not establish attendance, session content, a firm’s internal model or agent deployment, permissions, or investment results. The capture note records the post-event recovery plan.

The official GAIIM 2026 programme adds another upcoming, title-blind event route. The hybrid conference is advertised for September 29–30, 2026 at Columbia University’s Faculty House in New York. Its preliminary agenda includes agentic buy-side workflows, Claude Code and Cursor, agentic equity analysts, MCP and connected research stacks, risk/compliance/operations/governance, and an Anthropic session on agents in financial services. The public speaker inventory lists Pete Petersen as CTO of Causeway Capital Management, Michael Soss as CIO of Millburn, and George Ho as Managing Member of NIV Asset Management, alongside finance-data and research vendors including Daloopa, Hudson Labs, Canary Data, and AlphaSense. This page establishes an advertised event, agenda, and listed roles; it does not establish attendance, session content, a firm’s internal model or agent deployment, permissions, or investment results. The capture note records the post-event recovery plan.

CFA Society Hong Kong’s official AI x FinTech Symposium 2026 page adds a separate Asia-facing route for September 17, 2026. Its programme covers AI in investment workflows, agent deployment, investment infrastructure, and institutional digital assets. The public role inventory lists Kevin Kwan as Head of Data Science at Bloomberg Enterprise Data; Douglas Chan as eBroker founder and chairman; Bryan Chik as a BlackRock Director of Data Science; Alice Wong as a J.P. Morgan Asset Management portfolio manager; and Brooksley Kang as Head of Digital Asset at Aletheia Capital. The page also exposes an HKU quantitative-finance research route around language-model agents and an open-source investment-desk system, but the person-level attribution is intentionally omitted here under the public redaction rule. This is event metadata and organizer biography evidence, not a transcript or proof of any firm’s model inventory, deployment, permissions, or investment results. The capture note records the names, roles, and post-event recovery plan.

The Investment Association’s IA Talks AI archive adds a dedicated UK investment-management media route that a search for fund names or quantitative titles can miss. The archive says the series began in May 2023 and covers implementation, investment-specific use cases, policy, regulation, and practical considerations; its embedded official YouTube playlist contained 35 entries on the August 27, 2026 inventory check. Newly listed 2026 episodes name Next Gate Tech CEO and co-founder Davide Martucci on clean data, deterministic workflows, LLM use, agentic systems, and governance; Adam Grainger of Agentic Risks on agents, accountability, and the human–agent organisation; and Deloitte’s Dimitri Tsopanakos on scaling AI and measuring its performance. An alternate public-player route recovered automatic English captions for all six 2026 entries. Those captions add bounded signals: Next Gate Tech frames data ingestion and harmonisation across formats as a prerequisite to automation (02:24–03:29) and describes LLM use for unstructured files (06:57–07:50); Regulex describes a planned platform joining macro data, news, holdings, NAVs, analytics, and CRM into a morning brief, with later agent plans, while saying the product remained under development (03:25–06:21); and Agentic Risks defines agents in terms of planning, permitted tool calls, cross-system action, and accountability (02:40–04:18). A Deloitte episode contains an unattributed speaker claim about an unnamed hedge fund operating thousands of autonomous agents; it is retained only as an uncorroborated caption-level statement and is not assigned to any manager. The IOSCO episode adds a supervisory vocabulary around risk-based oversight, hallucination, opacity, disclosure, operational resilience, outsourcing, governance, and model oversight in financial services (01:01–07:46). Captions are automatic and not audio-verified, so these are institutional, vendor, regulatory, and event-media signals—not evidence of a covered manager’s model inventory, training data, permissions, production deployment, or investment performance. The archive capture note records the episode IDs, six caption hashes, and recovery boundary.

The recovered Evolution Exchange Singapore episode is another title-blind route: its page title does not identify a particular fund, but the January 5, 2026 discussion brings together Jiri Pik, Ernest Chan, and Jared Broad around practical portfolio workflows. The locally recovered audio describes decomposing portfolio work into data processing, idea screening, scenario analysis, monitoring, optimization, coding, and strategy testing (04:50–05:32; 22:54–23:04). It gives two particularly concrete control patterns: an MCP-connected portfolio-correlation agent (10:11–10:38) and a monitoring agent that alerts on portfolio deviations (11:22–11:47). Chan describes “corrective AI” as a way to identify or correct human errors rather than decide whether to trade (15:53–16:30), while the closing discussion frames the role as augmentation with human judgment retained (30:42–31:17).

This is a practitioner/provider conversation, not a named hedge-fund disclosure. It supports a task-level control taxonomy, but does not establish a particular fund’s production deployment, model performance, data rights, permissions, or autonomous order authority. Proper nouns and quoted wording remain subject to audio spot-checking; the capture note preserves the local transcript and hashes.

The Podscan recovery pass converted four previously metadata-only or missed routes into timestamped transcript records. Goldman Sachs’ Robyn Grew conversation describes ManGPT, Alpha Assistant, and AlphaGPT, with the CEO describing curated data, human/quantitative review before commissioning, and model-change mapping through the firm’s quant processes. Dmitry Balyasny’s Bloomberg Masters in Business episode adds a title-blind account of BAM’s shift toward internally built technology as it added more quantitative and macro strategies; the transcript includes speaker-reported counts of more than 500 technology staff and more than 100 data/AI staff. Acadian’s Scott Richardson Credit Edge episode adds timestamped discussion of machine-learning default models, LLM-assisted ingestion and reporting, agentic workflow acceleration with human review, and systematic-credit coverage. These are first-party or practitioner accounts captured through a secondary transcript service; they are not independent audits, model cards, permission maps, or performance attribution.

A new Odds on Open interview with former Two Sigma quant Omer Seider provides a separate historical personnel and workflow route. Seider describes a Two Sigma program that normalized expert opinions through structured web forms across thousands of companies and multiple horizons, then discusses generative AI and digital analysts as tools for data validation and context curation. Because this is former-employee testimony, it is kept separate from current Two Sigma first-party strategy disclosures and does not establish current reporting lines, model ownership, or deployment.

The transcript recovery also repaired the evidence lane for AXA IM Core / BNP Paribas Asset Management’s quantitative-portfolio discussion. Ram Rasaratnam and Chris Iggo describe a neural-network model that they say entered production roughly nine years earlier, machine reading of earnings-call transcripts for sentiment, precision, and language changes, a patent-derived innovation signal, and an internally hosted code-repository assistant. The speakers also describe model-importance checks, testing, and portfolio-manager or team sign-off. Their account of identifying regional-bank risks before the 2023 failures is retained as speaker-reported and is not independently validated here. The originally resolved Ausha MP3 was an unrelated inflation-bonds episode; the Podscan timestamped transcript recovery is therefore cited as a separate capture route.

A new Australian route comes from Magellan’s In The Know episode with Vinva Research Manager Rob Franklin. Franklin describes proprietary data, broad systematic equity coverage, text processing, and an AI coding workflow that can plan, write, test, and evaluate research code. He says a task that formerly took days can sometimes be attempted in an hour or two, and describes a preference for transparent intermediate targets that humans can validate rather than direct return prediction. The conversation also puts talent, then data, then infrastructure and compute at the centre of systematic-investing requirements, while retaining human judgment, diversification, risk controls, and fiduciary accountability. These are dated practitioner statements; no model weights, vendor contract, autonomous capital authority, or independent performance attribution is disclosed.

The official Jane Street Tech Talks archive also yielded two transcript-backed ML infrastructure routes that a hedge-fund or AI keyword search could miss. Making GPUs Actually Fast is presented by Sylvain Gugger and Corwin de Zahr and discusses Jane Street’s CPU context for low-latency trading, PyTorch, data loaders, GPU training, kernel fusion, and custom-kernel tradeoffs. Building Machine Learning Systems for a Trillion Trillion Floating Point Operations is a Jane Street-hosted talk by Meta PyTorch compiler researcher Horace He on compiler optimization, large-scale GPU training, and distributed ML systems. Both pages expose full HTML transcripts and YouTube recordings/captions. They are technical-media evidence about systems and research vocabulary; they do not establish a Jane Street portfolio model, training corpus, live permission, or performance result. The timestamped capture note preserves the capture boundaries.

The current regional pass recovered a dated Caixin interview with Lu / 陆政哲, identified on the original page as High-Flyer Quant’s CEO, plus an English mirror and translation. The 2020 interview attributes to High-Flyer an AI-centered workflow spanning market, fundamental, and structured alternative data, deep-neural-network training, portfolio generation, and programmatic execution, and describes an AI Lab and internal compute investment. It is historical executive media, not a current model registry: no weights, current personnel allocation, evaluation design, permissions, or independent performance are disclosed. The source ledger preserves the 2020 date and keeps High-Flyer’s finance systems separate from later DeepSeek research.

The same title-blind pass added two publisher routes for Ernest Chan and QTS Capital Management: Interactive Brokers’ January 2025 Machine Learning in Finance episode, which includes a publisher transcript and career lineage, and the Mutiny Fund episode 17 page, which adds a separate QTS/Tail Reaper route. These are practitioner and publisher records. They do not establish a current QTS production model, agent permission, live portfolio authority, or performance result.

August 20 emerging-manager and WorldQuant media follow-up

The emerging-manager sweep found several public AI descriptions that were not visible through a hedge-fund-plus-AI title search. Castle Ridge’s W.A.L.L.A.C.E. page describes a proprietary system that creates, maintains, and evolves portfolios, detects behavioral patterns, learns and adapts, and formally interacts with portfolio managers. Its team pages identify a Chief Scientific Officer, VP of R&D, and machine-learning scientists, including public academic and publication lineage. Castle Ridge’s news archive also functions as a route map to older conference and media appearances; the Canaccord release is a dated partnership record. These are firm-controlled descriptions and archive metadata, not an independent audit of models, permissions, data rights, or outcomes.

Bayswater Technologies uses “AI-native,” specialized agents, and a centralized world-model description for a long/short-equity process, while explicitly retaining a human portfolio-manager decision point in its public account. The page also provides a founder route through Marble Bar, B2C2, Imperial College London, Queen’s University Belfast, Oxford, and UCL. Machine Capital publicly describes machine-learning and automation, and its AI Equity Fund page specifies neural-network return forecasts, portfolio construction under a predefined volatility threshold, and cash reduction when full exposure would breach that threshold. These are first-party product descriptions; they do not disclose the network version, training corpus, data rights, live permissions, or independently measured performance. TradeWell Capital and Machina Capital publicly describe related machine-learning or AI research platforms without disclosing LLMs, agent permissions, training corpora, or model-level evaluation. Seldon Capital adds a company-social identity and archived university recruiting language combining fundamental research with machine learning; the job pages are historical/archived and are not treated as current vacancies. TC43’s SEC Form D establishes an entity route, while the description of its ML framework comes from a secondary adviser profile, not a direct firm technology page.

An additional temporal link now comes from Acadian’s July 30, 2026 Q2 earnings discussion. The published earnings transcript attributes to President and CEO Kelly Young the statement that Acadian welcomed Jonathan Briggs and other members of the former TC43 team, describing their research, data-engineering, and modeling capabilities as complementary to Acadian’s systematic platform and research agenda. Quartr’s event summary surfaces the same personnel transition, while Acadian’s investor-relations archive confirms the Q2 reporting event and date. This is a dated management statement reproduced by secondary transcript services, not a first-party personnel roster. It does not establish Briggs’s current Acadian title, transfer of TC43 code or data, a particular model, an AI/GenAI system, agent permissions, or performance. The transition capture note preserves the entity boundary and disqualification loop.

An adjacent historical media route adds context on Briggs without changing that boundary. In the December 24, 2021 Investor’s Podcast interview, Briggs is identified as Delphia’s CIO and discusses a modeling choice that separates sparse, nonstationary return observations from more structured fundamental and cash-flow variables (00:37:59–00:44:00). He describes using machine learning with business KPIs and surprise relative to expectations, and separately describes data, compute, talent, and framework costs plus probabilistic portfolio evolution (00:55:47–00:58:48). Delphia’s first-party retrospective records his June 2020 CIO appointment, the April 2021 market-neutral quant-strategy launch, and the later open-sourcing of InvestOS. This is historical named-practitioner and company-history evidence, not proof of a current Briggs title at Acadian, transfer of TC43 or Delphia code/data, a current model, agent permissions, production deployment, or performance. See the capture note.

The WorldQuant gap was a platform-discovery failure. Its current careers page links to both the official YouTube channel and the WorldQuant Careers Facebook page; the Facebook video page is therefore recorded as a separate first-party video surface. The technology archive exposes several AI/BRAIN titles, and the leadership page names Paul Griffin as Co-CIO and Chief Science Officer, explicitly describing him as leading WorldQuant’s AI initiatives. It also identifies technology and research leadership, including Gerry Beatty, David Rukshin, and Rohit Agarwal, plus an employee-video reference to a machine-learning, deep-learning, and data-science center of excellence. Three recent essays add explicit but still high-level strategy language: Nitish Maini on AI agents in quant research, Andreas Kreuz on AI in quant investing, and WorldQuant’s July 2026 AI/BRAIN perspective. These are public role and positioning signals; no model weights, data licenses, evaluation results, or fund-specific deployment claim is made.

The archive audit adds useful detail to that bounded reading. Maini’s June essay describes agents reviewing financial documents, generating hypotheses, running simulations, and refining algorithms in the International Quant Championship context. Kreuz’s May essay says AI is used across teams and may help structure unstructured data and generate ideas, while retaining human accountability. A July WorldQuant perspective discusses feature selection, overfitting, and the simulation-to-live gap through an IQC and sports analogy. These statements are first-party positioning and education, not a disclosed production architecture. The older RavenPack World of Alphas page adds a 2018 conference route for Nitish Maini covering Alpha Factory, human/machine complementarity, research, portfolio management, risk control, and execution. A 51-second RavenPack highlight video had no captions, but its public audio was recovered and transcribed locally. The narration describes RavenPack news datasets and headlines, NLP learning words or phrases associated with favorable or unfavorable news for a stock, historical checks of predictive quality, and neural-network techniques (00:05–00:47). This is a historical presentation highlight, not a full-session transcript or current WorldQuant system disclosure; it does not establish current vendor terms, model inventory, data rights, deployment, or performance. See the ASR recovery note.

The same archive audit surfaces a separate WQU talent and methodology lane. A WQU alumni profile describes Adrian Dunkley’s 2025 election experiment using multiple named general-purpose models, independent forecasts, adversarial debate, and confidence-weighted aggregation. A WQU–Sazience partnership announcement describes applied-data-science coursework, mentoring, consultancy projects, and an Africa-focused talent pipeline. Both are relevant to public AI-method and talent discovery, but neither establishes a WorldQuant asset-management system, model ownership, or investment use.

The title-blind pass also recovered Nitish Maini’s April 2025 Market Maker interview, which has searchable chapters and a publisher transcript covering BRAIN, Learn2Quant, the International Quant Championship, and Maini’s career route. A RavenPack 2018 Symposium page identifies Maini’s historical “World of Alphas” presentation and exposes a direct YouTube recording route. A recent J.P. Morgan forum post by Paul Griffin names Cubist’s Pusheng Zhang and describes proprietary-data, experimentation, and human-constraint themes. David Rukshin’s WorldQuant technology post connects AI-era developer roles to the firm’s Global Technology Conference in Riga. The podcast and social pages add public personnel and research-culture evidence; they do not establish current model inventory, permissioning, deployment, or performance.

The later RavenPack/WorldQuant Data Creation Challenge announcement adds a distinct public data-partnership and talent-network route. RavenPack says WorldQuant BRAIN consultants used the Bigdata.com Search API to turn unstructured financial content into research datasets and then build alphas in a six-week educational challenge. RavenPack’s results post reports participation and submission totals and includes publisher-provided transcript text. These are challenge-level and publisher-reported facts; they do not establish an exclusive data license, a live investment-system workflow, dataset quality, model ownership, or strategy performance.

A caption audit of the official Quantcepts: How Quants Can Partner with AI adds more specific public vocabulary than the title alone suggests. The video says statistical, machine-learning, and deep-learning models can be applied to BRAIN data; it describes predicting missing historical observations, turning quarterly fundamental metrics into weekly or daily observations, predicting returns one or five days ahead, and compressing multiple data fields into fewer predictive features (00:30–01:21). The video also says this can improve alpha turnover and performance, but that is WorldQuant’s educational claim, not an independently reproducible result. No architecture, training corpus, evaluation split, portfolio permission, or production attribution is disclosed. The timestamped capture and boundaries are in the WorldQuant video surface note.

The source-level evidence, dates, identity controls, and disqualification boundaries are in the emerging-manager and WorldQuant follow-up note and the regional quant/ML media expansion note.

The wider title-blind conference sweep adds useful cross-firm public routes without supporting a ranking. BattleFin Discovery Day Miami lists WorldQuant CTO David Rukshin, Bridgewater research/data personnel, Balyasny data sourcing, PDT data strategy, and RTW data strategy among its speakers, while a Data Score agenda review frames sessions around data-vendor integration, compliance, and enterprise GenAI. The Neudata Summer Data Summit agenda adds explicit session descriptions for Morgan Stanley QIS “skeptic” agents, T. Rowe Price coding agents, and data-strategy sessions involving WorldQuant, Jain Global, Balyasny, PDT, Jump, BlackRock, SIG, Quantbot, and Columbia Threadneedle. These are event metadata and agenda claims; they do not establish any firm’s deployed architecture, permissions, cost, or performance. The Eagle Alpha fall New York event is an additional September 10, 2026 watch route with public AI-infrastructure, agentic-workflow, prediction-market, and web-scraping topics.

The speaker graph also adds a Neudata London recap with WorldQuant’s Eugene Miculet and data-strategy practitioners from Man Group, CFM, Final, and Agami. Its public account describes data-evaluation pipelines, automated sourcing, out-of-sample concerns for AI-generated datasets, and vendor documentation; it is a secondary event recap, not a WorldQuant system disclosure. The BlackRock comparison gains a separate research route through AlphaAgents, a 2025 paper coauthored by Dhagash Mehta that describes role-based fundamental, sentiment, and valuation agents, debate, logging, human review, and retrieval evaluation. The paper, KAIST seminar listing, and NVIDIA GTC program support distinct publication, personnel, and conference routes; they do not establish BlackRock deployment or performance.

The QuantVision 2026 agenda adds a separate conference-roster route that was not represented as its own record in the media ledger. Fordham’s Spring 2026 publication and official recap confirm the March 19–20 event and its focus on quantitative finance, AI in asset management, alternative-data research, and human judgment. The public agenda lists a machine-learning panel with Samson Qian and Arkin Gupta of Citadel and Matt Rowe of Man Group, a data/AI fireside involving Claudia Perlich of Two Sigma, multimodal-alpha discussion with participants from Millennium and DRW, and a BlackRock keynote by Dhagash Mehta. It also lists a keynote on alternative-data research agents by Charlie Marin of Quanted. A separate March 10 Rebellion panel page attached to the event labels Dorothy Ruderman as Data Strategy at Verition Fund Management and Mark Fleming-Williams as Head of Data Sourcing at Capital Fund Management; it also lists former Two Sigma, Schonfeld, and Point72 routes. The panel description covers geolocation, satellite, supply-chain, and unstructured-web data, vendor transparency, model interpretability, compliance, and AI-driven data synthesis. These are publisher agenda and speaker-label signals; no recording or transcript was recovered, and they do not establish current employment after the event, system ownership, deployment, permissions, or performance. See the conference capture note. The Verition personnel graph also now contains a provisional technology-architecture route. An unverified The Org team page labels Vishnu Mavuram “Head Of Technology And AI Architecture” and Srinivasa Chinnam CTO, while public LinkedIn profiles and aggregator/LinkedIn cross-checks connect both people to Verition. Mavuram’s public posts discuss unified data planes, MCP, AI-friendly engineering assets, and agentic development with compliance guardrails. This is a useful personnel-discovery and architecture-vocabulary route, but the title source is explicitly unverified and the posts are personal; it does not establish a Verition platform, model, deployment, permissions, or investment use. See the capture note.

Quantbot adds a distinct first-party ML and data-infrastructure route. Its current home page describes machine learning and cloud computing, data-team collaboration, in-sample/out-of-sample analysis, an in-house backtesting platform, paper trading, operational checks, and benchmark-gated live trading. Its About Us page names Paul White as CEO, Ashar Mahboob as CIO, Thomas Crimi as CTO, Benjamin Cilia as Chief Data Officer, and Rohit Thakare as Chief Trading Officer; the founder Q&A says the firm hired specialist talent for ML alpha generation, sources new datasets, and uses computing-on-demand. A dated academic case by Sodhi and Tayur describes a 2021 Quantbot Alpha Optimal Combination problem: combining noisy, correlated ML-generated alpha candidates under risk constraints, with data scientists reporting to the CIO and a technologist to the CTO, using Toshiba’s cloud Simulated Bifurcation Machine with Carnegie Mellon collaboration. The paper reports a retrospective 252-trading-day comparison of twenty near-optimal solutions against a prior heuristic. That is a specific, paper-reported historical portfolio-combination experiment, not evidence of current GenAI or LLM use, continuing partnership, production continuity, autonomous authority, or independently audited performance. See the capture note.

RTW Investments adds a different data modality rather than an AI claim. Its first-party team page identifies Alex Ewing as Principal, Data Strategy, says he joined in August 2025, and describes the integration of complex healthcare and claims-based datasets into investment research and corporate development. RTW’s key-facts page describes a science-led life-sciences process, while a June 2026 SEC prospectus describes internally developed genetics-based analytical tools and research across scientific, clinical, regulatory, and commercial milestones. The public record does not say that RTW uses AI, GenAI, LLMs, or trial-outcome models; it does identify a concrete healthcare-data and personnel route for the clinical-research queue. The role description, prior-employer history, and regulatory filing should not be merged into a claim about current model architecture, data licensing, or trade authority. See the capture note.

Engineers Gate adds a personnel and academic-lineage route, not a confirmed AI-system disclosure. A University of Texas profile and dated CV identify Sujay Sanghavi as a former Engineers Gate senior quant and founding member of an algorithmic-trading team; his current research concerns the architecture and training of large language and representation models, with IIT Bombay, UIUC, and MIT LIDS lineage. A public Oliver Orejola profile identifies him as a Machine Learning Research Scientist at Engineers Gate, with a Tulane mathematics PhD and publications on wavelet random matrices and Hurst-exponent analysis. A UIUC lab roster labels Xiaoxiao Shi “Head of SPM @ Engineers Gate,” while a separate corporate-history page places that role in 2014–2019 and lists a later Occudo position. The public records do not establish an Engineers Gate AI lab, current firm ownership of these research topics, model inventory, data, deployment, or investment authority; the Shi route remains temporally conflicted. See the capture note.

Tudor has a separate, dated practitioner-media route. The Meb Faber Show transcript identifies Ulrike Hoffmann-Burchardi as a Tudor portfolio manager focused on digital, data, and disruptive innovation, and records her discussion of purpose-built models, human control, data/AI/domain knowledge, transparency, ethics, and guardrails. She describes a possible architecture in which domain-specific AI analysts feed investor and risk-management AI layers. This is a public conceptual framework from an August 2023 recording, not evidence that Tudor built or deployed those layers. The source predates current model generations and does not identify a provider, training corpus, evaluation, permissions, production endpoint, or AI-attributed performance. See the capture note.

Caxton adds a recruiting signal around workflow automation and quantitative-data infrastructure. A public Bengaluru Broker Relations and Vendor Analyst posting includes AI in workflow-automation and process-efficiency initiatives and lists GPT, SQL, and Tableau as desirable skills. A separate Quantitative Developer posting describes a Quantitative Development & Data group supporting portfolio-manager alpha generation, strategy deployment, and risk management through Python libraries, web services, dashboards, databases, and ETLs for market and quantitative data; a public mirror adds alternative-data ETLs and a Bengaluru-office context. These are job-description signals, not evidence of a filled role, selected provider, model deployment, training corpus, permissions, or investment impact. An unverified The Org team page is retained only as a personnel-discovery lead. See the capture note.

Point Three Group adds a market-level recruiting route without a firm assignment. Its public AI/ML Research Scientist opening describes work spanning machine learning, NLP, optimization, predictive models, and capital-markets problems for hedge funds and trading firms, and asks for advanced quantitative training plus first-author publications or conference presentations at venues such as ICML, ICLR, or NeurIPS. Because Point Three is a recruiting intermediary and names no hiring client, this is evidence about demand-side role vocabulary and a new search route—not evidence about any tracked fund’s team, model, deployment, or performance. See the capture note.

August 20 second-pass manager and media expansion

The next candidate pass added a regulator-and-media layer around several smaller or less-covered managers. Meridian & Saturn’s official site describes China A-share and global systematic strategies, ML signals, a research-to-production history, and named quantitative and technology leadership. The Monetary Authority of Singapore directory independently confirms the Singapore fund-management licensee and lists Zhang Feiyun as executive director and CEO. The firm’s scale and compute figures remain self-reported.

PharVision provides a public machine-plus-human description for a systematic, market-neutral, factor-neutral US-equity process, names its founders, and separates research, data engineering, and technology functions. Its SEC Form ADV corroborates the adviser identity. Signum Investments describes a quantitative and machine-learning layer over fundamental long/short equity, with Michael Scafati as managing partner and CIO and an MIT Sloan lineage. Move37 Capital describes AI-driven global strategies and names Kunal Gautam and Dr. Mushtaq Shah; its relationship to Laven Advisors is supported by the firm’s own disclosure and the FCA/Laven route, not by an independent model audit.

A Bulgarian-language LinkedIn partnership post adds a regional PharVision–SoftUni education and talent route, reposted by Blago Baychev. Its underlying LinkedIn MP4 was recovered and processed with local Bulgarian ASR. The timestamped, unverified transcript describes a systematic US-equity fund using machine learning and AI (00:16–00:56), a technical mix including machine learning, software engineering, linear algebra, and optimization (02:08–02:25), an eight-person team with machine-learning/AI, physics, mathematics, economics, and software-engineering backgrounds (02:35–03:14), and the SoftUni relationship as education and talent development (06:29–07:23). These are bounded paraphrases from unverified ASR, not quotations or a translation-certified transcript; they remain partnership and recruiting context rather than model or deployment evidence. See the capture note.

QuantumStreet AI adds a media and partner trail that was not in the earlier manager set. The title-blind Channel Insider interview links an embedded YouTube episode and Simplecast audio featuring Art Amador. Its recovered English automatic captions add timestamped speaker-reported detail: QuantumStreet is described as an IBM watsonx fintech partner (00:25–00:45); the platform combines fundamental, macro, technical, news, and social-media data (03:02–03:19); IBM is described as supporting multilingual processing (05:42–06:05; 13:08–13:17); and generative summaries and security/market forecasts are described as outputs (10:57–11:27). Amador also describes a BNP Paribas collaboration on a multi-asset AI-powered index and a structured-product route involving XP in Brazil (14:21–16:00). QuantumStreet’s March 2026 release describes knowledge graphs, earnings-call and regulatory-filing inputs, and SHAP explanations; its quoted performance is explicitly backtested and hypothetical. The IBM watsonx material provides the foundation-model partner route. These interview statements are automatic-caption navigation evidence and do not independently establish current client deployments, model inventory, data rights, or results. See the capture note.

The same guest is listed on a separate Capital & Power episode distributed through Amazon Music, dated April 27, 2026. An AInvest article links a matching AInvest Official YouTube recording published April 21, 2026. Its public English automatic captions add timestamped, speaker-reported detail: named market, macro, SEC-filing, and news inputs (03:48–04:48); an adaptive approach as market conditions change (04:48–05:05); SHAP-based signal attribution across categories and individual inputs such as EPS, RSI, and technology-sector sentiment (06:17–07:11); and named institutional relationships plus an index-provider/access-route description (00:40–00:46; 01:13–01:23; 03:03–03:21). The source note records the caption hash and boundaries. These claims are not an independent audit of relationships, data rights, processing volume, model behavior, or performance, and promotional performance language is excluded from the evidence base.

The first-party AInvest Studios / Capital & Power catalog adds a title-blind manager-media surface that was not present in the earlier registry. Public YouTube caption recovery resolved separate appearances by Stephen McClurg of Canary Capital, Hal Lambert of Point Bridge Capital, Seth Cogswell of Running Oak Capital, Dan Rasmussen of Verdad Advisors, Bob Elliott in a former-Bridgewater context, Michael Green of Simplify, and David Bahnsen of The Bahnsen Group. The reviewed captions add market and portfolio commentary; only McClurg’s AI-data-center discussion and Lambert’s AI-theme discussion are clearly AI-related, and neither discloses an internal model or deployment. The remaining routes are useful negative controls against treating generic manager appearances as AI evidence. See the catalog capture note.

Runtime Fund publishes a stated five-stage lifecycle from research through development, training, controlled live validation, and human oversight, with Raed Malhas and Deniz Erkan named in leadership. Voleon adds a current research-analytics engineering role focused on foundational datasets, metadata, lineage, governance, cloud, and on-premise research infrastructure. Marduci and XYZ Capital’s hiring PDF remain watchlist entries because their public AI or ML language lacks sufficient independent identity and operating evidence. Axiome is tracked as an adjacent AI-native investment-intelligence platform, not yet as a verified hedge-fund vehicle.

The full source, media, personnel, and disqualification ledger is in the TC43-adjacent manager and media follow-up note. These additions do not support a cross-firm ordering conclusion.

A regional/title-blind media pass adds three distinct evidence patterns. A Quantedge interview with CEO Suhaimi Zainul-Abidin describes generative AI as useful for research productivity, data cleaning, and execution while keeping it outside production models; this remains publisher/index evidence until the recording is recovered. Arrowpoint’s official Singapore role describes deep learning for financial time series and cross-asset signals, regime/stress testing, and production-quality research code in an Asia-focused pod structure. DRW’s Montréal AI/ML role names forecasting, LLMs, deep learning, reinforcement learning, graph neural networks, and time-series work, while DRW’s Canada page identifies a Mila partnership. These are distinct podcast, hiring, and regional-partnership signals; none establishes model ownership, filled roles, trading permissions, or AI-attributed returns.

August 24 personnel, regional, and title-blind refresh

The personnel pass resolves several public identities without converting personal research into firm deployment evidence. Guy Davidson’s public CV lists him as a Machine Learning Researcher at Jane Street from January 2026 and identifies an NYU Center for Data Science PhD supervised by Brenden Lake and Todd Gureckis. His public task-representation paper, SAGE-Eval paper, and research page are personal or academic artifacts; the reviewed sources do not show Jane Street authorship, firm-owned code, or trading deployment. The CV also lists a Meta FAIR role with an unresolved date overlap, so no additional current-employer inference is drawn.

A follow-up recovery pass tested three more same-name recordings—CppCast 328, ACCU 2021, and Game Industry Conference 2024—and disqualified them as Jane Street evidence. Their introductions identify the separate Creative Assembly/Six Impossible Things C++ and game-engineering speaker. The collision ledger records the hashes and negative attribution boundary; name similarity is not treated as an employment link.

A title-blind Capital Thesis interview with Agustin Lebron identifies him as a former Jane Street trader and describes Echol Technologies as a research firm using reinforcement learning for algorithmic-trading systems. The timestamped captions discuss simulation-to-live mismatch, historical-data cleaning, survivorship and dividend issues, production-data drift, and feeding live-trading observations back into model development. This is former-personnel and third-party media evidence. It does not establish a current Jane Street system, Echol production permissions, model quality, or performance.

The same pass adds current public personnel records around adjacent comparison firms. Martin Ferianc is listed as a Machine Learning Engineer at G-Research from March 2026, with public descriptions of financial-data ML systems, distributed-training libraries, and training/inference optimization; his linked papers and repositories are personal artifacts. Michael Weihao Song is listed as a Quantitative Researcher at Point72 from 2026, with prior JPMorgan Asset Management experience and public research spanning financial ML, NLP, multimodal models, and graph learning. Andreas P. Mentzelopoulos is listed as a Quantitative Researcher at Cubist Systematic Strategies from June 2026; his public MIT-linked deep-learning artifact concerns underwater-image diffusion and is not evidence of Cubist ownership or deployment. G-Research’s Oxford scholarship announcement identifies Daniel Marks as a statistical-machine-learning DPhil candidate in a scholarship relationship, not as G-Research staff.

The regional sweep adds manager-controlled signals with explicit boundaries. Mackenzie GQE describes a roughly ten-person team using ML and NLP alongside human judgment and managing approximately C$10 billion, according to the manager’s own account. RQI Investors describes ML and AI/NLP as research tools for nonlinear patterns, management communications, and qualitative information. Vinva provides public personnel context for data-infrastructure and quantitative-systems work, including Reshma Joseph’s 2026 appointment and Rob Franklin’s ML, computer-vision, and NLP background. Lingjun’s 2026 update uses “technology AI-ization” and describes coordination across research, risk, markets, and operations. These are firm-reported positioning and personnel signals; they do not identify model weights, portfolio permissions, or AI-attributed returns.

The completed vendor and conference lane also recovered Captide’s LangChain customer story, LSEG’s AgentSmyth collaboration, LSEG Deep Research, and Snowflake’s financial-services agent architecture. These sources describe multi-agent research, structured financial data, source-linked outputs, and governed workflows, but none links the named products to GMO, Acadian, Arrowstreet, or another covered manager. They are retained as ecosystem context rather than evidence of adoption by a target firm.

The title-blind queue also surfaced a second public interview with Pico’s Ian McIntyre, dated September 23, 2025. The distributor listing and chapter markers cover direct connectivity, colocation, fixed-income latency, Corval network observability, cloud and AI in market data, resilience, and DORA. It is vendor-side infrastructure context and a distinct episode from the existing April 2026 recording; it does not identify a customer, hedge-fund purchase, model, agent, or investment decision. The capture note records the episode-level boundary.

The queue also recovered a title-blind Odds on Open interview with recruiting director Jesse Skaff, with a matching Apple Podcasts page. The January 2026 episode discusses hiring across Citadel, Point72, Balyasny, and Millennium, including the recruiting market’s demand for data-science and AI talent, risk ownership, and how recruiters assess technical and behavioral fit. This is recruiter testimony about labor-market signals, not evidence of any named firm’s internal AI system or investment permission structure. A separate At the Forefront episode with Pico’s global head of market data supplies a vendor-side route into real-time market-data volume, cross-asset infrastructure, and execution systems; it is retained as stack context, not as a target-manager disclosure.

Adjacent allocator, operations, and systematic-media recovery

The title-blind media lane recovered several useful boundary cases outside direct manager disclosures. A Capital Allocators episode with GHR Foundation’s Andrew Dorle describes Tetrix as the foundation’s first “truly AI-native” system for private-equity transparency, exposure, and analytics, alongside Canoe AI capabilities and Claude Enterprise. The episode also describes deliberate adoption: test distinct utility, preserve existing systems while they improve, and do not assume AI replaces the operations function. This is allocator-operations evidence, not hedge-fund deployment evidence.

A Capital Allocators interview with Paloma Partners COO Michael DeAddio says Paloma’s quants have used machine learning for years and describes his personal use of Claude for coding and spreadsheet work. The relevant captions discuss specifying the task and quality-checking the output, and describe shortening the path from an idea to a usable, debugged process. Personal use is not evidence of firmwide deployment, production agents, trading permissions, or alpha.

A J.P. Morgan systematic-commodities episode describes AI as useful for algorithm development, new-market exploration, and faster prototyping, while retaining manual oversight and safety checks before production. It is a bank commentary on systematic trading rather than a named live-trading model disclosure. A SmarterMarkets episode with iPushPull CEO Matthew Cheung adds a vendor-side governance route: agent identity, permissions, connectivity, and control layers are presented as prerequisites for trading and post-trade workflows. The episode is forward-looking and does not establish customer deployment.

Finally, a Capital Allocators episode with Colby College investment-operations director Alex Gentilli describes choosing an AI overlay around FinPilot rather than replacing the existing risk-management stack, while flagging token-based cost uncertainty as a system-design consideration. The publisher transcript also records Gentilli’s former Arrowstreet investment-operations role in 2014 and his later participation in organizing that function under one partner. This is former personnel and operating-process evidence, not an AI-platform disclosure. The episode does not establish current production status, model configuration, spend, investment-decision impact, or performance.

The Capital Allocators episode with RCP Advisors’ Adam Ciborowski adds a title-blind investment-operations route. RCP’s first-party announcement identifies Ciborowski as Principal and Head of Research, Portfolio Monitoring, and Data Management, and describes centralized data infrastructure, document-workflow automation, a decision to build proprietary AI solutions rather than rely only on third-party platforms, and data quality as a foundation for AI initiatives. The public episode description describes an inbox workflow that identifies incoming documents, extracts defined attributes, routes them for analyst review, and uploads them to a CRM. Local ASR of the recovered audio adds the speaker-described use of Microsoft Copilot, Databricks, Azure, OCR/schema work, an Intel inbox, roughly 50 target attributes, and analyst spot checks (12:18–20:30); these timestamps are machine-generated and not publisher chapter markers. This is a private-equity investment-manager operating-model record, not a hedge-fund deployment disclosure; the source does not establish a model family, training corpus, evaluation set, production permissions, investment authority, or AI-attributed performance. The capture note records the audio, hashes, transcript, and recognition boundary.

The firm-specific search recovered a deeper CFM personnel and research route. A Mixtape interview with Jérémy L’Hour identifies him as a quantitative researcher in CFM’s volatility-arbitrage team and an affiliated CREST researcher. The captions connect his academic work to synthetic control, causal inference, variable selection, machine learning for econometrics, natural-language processing, alternative data, and skepticism about noisy backtests. This links a named CFM researcher to public academic methods and research topics; it does not show that any paper or method is used in a live CFM strategy.

An Evercore discussion with Acadian CEO Kelly Young and Scott Richardson adds an executive route into systematic credit. Richardson describes using machine-learning tools over hundreds of variables to detect nonlinear associations in default and credit-risk forecasting, while the firm connects equity-derived data across a broader credit universe. The speakers also describe transparency through signal-level attribution and retain human interpretation around the systematic process. These are speaker-reported workflow and organizational claims, not model cards, independent evaluations, or AI-attributed performance.

QRT’s title-blind FPGA Usage in Trading Systems presentation identifies Brás Patta as part of the firm’s London FPGA team and discusses market-data ingestion, order entry, deterministic processing, latency/resource trade-offs, and FPGA acceleration for parallel data-processing and ML/AI tasks. The FPGA Horizons speaker page corroborates the speaker’s QRT role and low-latency trading remit. This is infrastructure evidence, not a disclosure of a QRT model, agent, or investment permission.

A Spotify episode with former Point72 proprietary-research head Kirk McKeown links his current Carbon Arc work to a prior Point72 research-platform role. The interview discusses structuring alternative data, domain context, decision trees, and guardrails around AI tools, including a caution that research processes and human learning remain important. It is former-personnel testimony and current-vendor positioning, not a Point72 disclosure or evidence of current Point72 systems.

The title-blind QuantSpeak archive adds a dated Neo Ivy route that is relevant to the article’s personnel and operating-model map. In the March 16, 2023 interview with Renee Yao, the CQF Institute recording identifies Yao as Neo Ivy’s founder and describes her public “third-generation” framing: training an AI system to generate new research ideas rather than relying only on human-led historical-pattern research (05:26–06:41). She describes nonlinear interactions as an explainability constraint in portfolio management (16:36–19:26) and says Neo Ivy built a system from scratch when legacy quantitative infrastructure did not support the approach (21:25–22:21). At 29:31–31:58, she discusses AI as a possible way to reduce some human idea-generation and operating requirements. These are dated, speaker-reported statements; they do not establish Neo Ivy’s current model inventory, production permissions, staffing, performance, or autonomous capital authority. The source note records the automated-caption limitations and independent first-party corroboration boundaries.

QuantSpeak also adds adjacent manager and control evidence. Tony Guida’s RAM AI interview identifies him as Co-Head of Systematic Macro and discusses price-based features, alternative data, reproducibility, p-hacking, and out-of-sample discipline (YouTube recording). Sonam Srivastava’s Wright Research interview describes reinforcement-learning feedback for stock/ETF allocation and a return/risk/diversification/churn objective (YouTube recording). Samit Ahlawat’s J.P. Morgan interview supplies explicit leakage, survivorship, hyperparameter, train/test, and deployment-pipeline controls (YouTube recording). These are dated speaker accounts and comparator methodology, not evidence of current production systems or performance. The combined capture note keeps the evidence tiers separate.

The same title-blind QuantSpeak archive adds a vendor/data-modality route through Grant Fuller’s 2023 interview and its CQF Institute recording. Fuller, identified as co-founder and CEO of Irithmics, defines vicarious risk as risk other market participants see but a portfolio manager does not (08:01–09:16). He describes institutional allocations and information effects as the underlying data problem (11:31–12:52), states a 177,000-listed-company coverage figure, and explicitly places the product on the risk-information side rather than presenting it as an alpha generator (15:08–15:41). Irithmics’ first-party product page separately names deep neural networks, reinforcement learning, exposure-change forecasts, and web/API delivery. This is vendor/CEO testimony and product positioning; it does not establish a hedge-fund customer, proprietary dataset, data rights, model performance, or live investment authority. The source note records the university and evidence boundaries.

The earlier QuantSpeak AI/ML archive adds four non-duplicative evidence lanes. NVIDIA guests Tim Wood and John Ashley discuss GPU compute, data, derivatives pricing, risk, and possible text/news/regulatory-filing workflows (episode; 40:58–42:00). This is vendor framing, not evidence of a customer implementation. Arthur Böök discusses neural-network interpolation of Monte Carlo prices and deep-learning forecasts of implied-volatility surfaces using S&P 500 options (episode; 02:53–06:26). Jörg Kienitz’s Acadia episode discusses model-agnostic hedging, local regression, Gaussian-mixture methods, open-source software, and C++ (episode; 03:01–10:58). Paul Wilmott discusses AI-assisted coding and quant work (episode; 05:26–06:04). These are vendor, research, professional, and practitioner sources—not evidence of hedge-fund deployment, performance, or capital authority. The archive note records each evidence boundary.

The same archive adds four named manager and infrastructure routes. Jan Rosenzweig’s Pine Tree interview discusses Sharpe-ratio allocation, equal-risk contribution, higher moments, and tail-risk-aware portfolios (00:55–04:48). Hari P. Krishnan’s SCT Capital interview discusses positioning, carry, and agent-based risk (01:54–06:19). Misha Fomytskyi’s Vola Dynamics interview describes derivatives infrastructure, hedge-fund portfolio-management experience, high-frequency market making, and quant software formation (03:00–06:20). Elie Ayache’s ITO 33 interview identifies a financial-software route focused on convertible bonds, equity-to-credit modeling, and volatility-surface calibration (00:40–01:30). These are dated practitioner and software-company sources. They do not establish current AI deployment, model performance, customer use, or investment authority. The manager/risk archive note keeps the evidence classes separate.

The archive-completion pass adds three further routes relevant to the public-signal map. Natalie Packham’s London Whale episode discusses correlation stress testing, stock/credit portfolios, and risk-management authority (00:48–01:18; 09:18–10:30). Araceli Venegas-Gomez’s Qureca episode discusses quantum-finance proof-of-concept work and workforce training (02:19–05:30). Carol Alexander’s episode links market-risk modeling, Algorithmics, early coded models, and a public GenAI-in-finance editorial route (00:54–01:55; 11:49–12:54). These are model-governance, emerging-compute, and research-community signals—not evidence of a current hedge-fund AI system. The archive-completion note records all 14 recovered episodes and the nine reviewed without article promotion.

The title-blind queue also recovered a March 15, 2021 Excess Returns interview with Kevin Zatloukal. The episode introduction presents him as a former Google and Microsoft programmer with an MIT computer-science PhD, a University of Washington computer-science teacher, and a collaborator in OSAM’s external Research Partner program (approximately 00:36–01:18). OSAM’s first-party announcement corroborates that dated background and says the collaboration explored value traps, factor relationships, and random forests. OSAM’s decision-stump paper, clustering paper, and 2019 quarterly letter add firm-published evidence of simple nonlinear methods, clustering, and software for automating machine-learning tasks within the historical Research Partner program. This is an academic-to-investment and external-collaboration route, not evidence of current OSAM employment, a current model inventory, live permissions, or AI-attributed performance. The timestamped capture note records caption limitations and dates separately.

A second, distinct OSAM media route is a February 2023 Flirting with Models conversation with Chris Meredith, then the firm’s Co-Chief Investment Officer. Meredith describes a full-stack research platform as three linked layers—data, analytical tools, and the people interpreting results—and gives historical examples of structuring scanned Moody’s manuals, classifying financial line items, vectorizing 10-K and MD&A text, mapping entities in shipping and news data, and building in-house analytics rather than accepting a vendor’s finished score (02:10–02:48; 05:16–07:18; 08:08–09:35). He also describes a dated research triage template that estimates possible return, transaction-cost reduction, or risk reduction against affected assets and data cost, plus a process for retaining failed ideas for later out-of-sample tracking (10:29–12:20; 12:52–13:21). This is historical platform and research-governance evidence, not current GenAI deployment or performance evidence; the capture note records the automatic-caption boundary.

An adjacent, title-blind route comes from Momentum’s interview with DiligenceVault founder Monel Amin. The publisher transcript describes the allocator/manager diligence workflow and chapters on AI in due diligence (14:11–23:54). DiligenceVault’s first-party AI hackathon account reports four short-cycle agent prototypes for pitch-deck screening, Form ADV peer analysis, distressed-debt sourcing, and early portfolio monitoring. Separate internal-adoption and BuildWithAI posts describe internal RFP, contract, support, email, API, and insight workflows. A July 2026 DV Assist 2.0 release names AI Product Manager Ryan Wilding and describes a governed Prompt Library, Review Agent, enterprise connectors, source traceability, approvals, and client-data isolation from foundation-model training; the AI platform page adds Form ADV change analysis, source-linked memos, and human approval. This is vendor-side evidence about the operating layer around asset managers and allocators, not evidence of a named hedge fund’s use, underlying model stack, or performance. The capture note records the source boundaries.

A separate title-blind route connects Sequentum CEO Sarah McKenna’s April 1, 2026 interview to a historical WorldQuant seed-round account. The interview discusses web and alternative-data acquisition, deterministic validation, compliance and quality controls, versioned/replayable workflows, and human-approved agents (01:46–02:31; 12:23–18:57; 30:53–end). A 2024 Sequentum profile corroborates McKenna’s CEO title and describes scripting, AI enrichment, validation, auditability, and data delivery. This adds a data-factory and acquisition-governance route around WorldQuant, not evidence of WorldQuant’s current model inventory, customer permissions, or performance. The capture note records transcript and promotional-claim boundaries.

Blushing Quants #33 adds a methodological control rather than a firm disclosure. Antonio Marrazzo discusses factor construction, regime-aware allocation, point-in-time data, release-date alignment, survivorship and look-ahead bias, realistic labels, overlapping samples, purged/embargoed validation, and transaction costs (03:39–08:46; 13:00–16:30; 38:28–45:00). His Rice/Fulbright profile identifies a graduate data-science and capital-markets background, but no current hedge-fund employer. This is useful for testing public AI/factor claims; it is not evidence of deployment by any tracked firm. The capture note keeps the personal-research boundary explicit.

August 26 title-blind vendor/data-layer route

A title-blind search also recovered a May 18, 2026 Beryl Elites interview with Peter Hafez, identified by the publisher as RavenPack’s Chief Data Scientist, with a matching publisher transcript. Hafez describes a vendor-side evolution from NLP and processed analytics toward knowledge graphs, entity resolution, grounded retrieval, and use-case-specific “micro data sets.” His example is a tariff-exposure workflow: map transition channels, retrieve relevant content, use an LLM for verification layers, and produce structured exposed/not-exposed labels (04:18–07:19). He also discusses repeated classification, majority voting, LLM-as-judge, and repeated mind-map generation as ways to seek convergence when quant users require reproducibility (07:50–11:28). Bigdata.com separately describes its product as RavenPack’s AI grounding layer with search, citations, workflows, and financial content. This route is useful architecture and vendor-strategy evidence, not a disclosure by GMO, Acadian, Arrowstreet, or a named customer: it does not establish customer adoption, model weights, training data, permissions, live trading authority, or performance. The capture note preserves the transcript and evidence boundaries.

A separate Beryl EDU panel, published March 21, 2024, is a title-blind route into the same data layer. Its publisher description names Jessica Stauth, then identified with Fidelity Investments, alongside John Farrall of Vertical Knowledge, Ken Zockoll of Spatial Risk Systems, Justyna Kosianka of Ursa Space Systems, and MarketWatch’s William Watts. The panel discusses customers applying alternative data to models, a shift from tabular inputs toward long review corpora, and the cost/benefit question for proprietary-data sandboxes (00:17–02:49). Zockoll’s contribution describes a knowledge graph connecting billions of structured data points and CNN/computer-vision feature extraction (03:18–04:20); the panel then describes data fusion, asset/location/time retrieval, and focused narratives from web-scraped collections (04:27–05:27). Fidelity’s FMR brochure supplement independently dates Stauth’s Fidelity role and prior Quantopian leadership, while Spatial Risk’s current company page identifies Zockoll with data governance and a spatial knowledge graph, and Ursa Space’s current platform page describes GeoAI and multimodal data fusion. These are separate dated and current context sources; they do not show that Fidelity or any tracked hedge fund used these vendors, nor do they disclose model weights, customer identity, data rights, production permissions, or performance. See the capture note.

Another Beryl EDU recording, titled What are LLMs in Finance and What are Benefits & Risks of Generative AI?, supplies a partial but useful source-factory route. The recovered captions distinguish generative output from language-model probability scoring and describe a novelty workflow that compares a model trained on an earlier period with later earnings-call, news, and regulatory-document language (03:15–05:30). The broader Beryl description lists Michael Oliver Weinberg (Columbia), Ruben Falk (AWS), Daniel J. Sandberg (S&P Global), Martin O. Ouko (TIAA), and Sateesh Kumar Challa (Société Générale); Columbia independently identifies Weinberg, and AWS independently identifies Falk with capital-markets data, ML, and AI work. This recording also reinforces why the discovery program searches job postings, workforce text, reviews, embeddings, taxonomies, and company identifiers—not only episodes with “hedge fund” in the title. The caption route ends at 05:30 and does not establish a current TIAA, Société Générale, AWS, S&P Global, or hedge-fund model, customer, permission, or performance claim. See the capture note.

A further Beryl recording, How do LLM, ML and AI influence Investment Management - its Strategy & Efficiency?, makes the boundary between automation layers more explicit. The panel describes algorithmic execution around ETFs and less-liquid fixed income, then separates bulk webinar transcription and summarization from the harder task of ranking which sources deserve human attention (00:00–06:31). It also raises data provenance, copyright, nonpublic information, on-site versus off-site deployment, client-data isolation, and the need to return to the original source when a summary may have changed meaning (06:33–10:35). The recording’s publisher description names participants from Columbia, AWS, S&P Global, TIAA, and Société Générale; the recovered captions are unlabeled and partial. This is a dated workflow and governance route—not evidence of any named employer’s current model, customer, permissions, or performance. See the capture note.

The Beryl archive also contains a title-blind NLP/ML and alternative-data discussion, published April 14, 2024. Its recovered captions put the emphasis on the data layer: cleaning and distilling large datasets, structuring LinkedIn profiles, job postings, and Glassdoor reviews, mapping millions of job titles into occupations and seniority levels, and using unsupervised learning, taxonomies, embeddings, and company IDs before the result enters a conventional investment workflow (00:37–03:31). The panel then stresses data-centric MLOps, purpose-built datasets, provenance, and millisecond order-book data as a high-volume ML use case (03:33–05:37). This is methodological evidence that supports title-blind personnel and data-source searches; it does not identify a hedge-fund customer, current deployment, model, permissions, or performance. See the capture note.

Another title-blind Beryl recording, How does AI Democratize Quant Tools?, published March 19, 2024, adds an implementation pattern. The publisher lists participants from ActiveViam, IBM, Microsoft, and RockCreek. The recovered discussion links cloud-backed environments and digital assistants to investment/technology collaboration, then proposes starting with a narrow filing task—flagging changes in risk language—and adding adjacent workflow blocks over time (00:00–05:05). It also distinguishes foundation-model and non-generative AI uses from generative AI and connects nontraditional data to question design (05:09–07:12). The captions are unlabeled and include speaker-reported adoption examples that are not independently verified. This is a dated implementation route, not evidence of any named employer’s current model, customer, permissions, deployment, investment authority, or performance. See the capture note.

A third title-blind Beryl recording, Markets are Dynamic Ever Changing Beasts. How do LLM Adapt to Rapid Market Changes?, adds a distinct multimodal boundary. In a short, partial exchange, the panel questions whether LLMs alone solve changing-market adaptation, then discusses combining language with computer vision and recommender systems around distinctive data such as satellite observations (00:00–02:00). It separately describes “digital labor” or virtual analysts as a workflow-automation use case that does not require a predictive-market claim (02:03–02:22). Only approximately 2:22 minutes were recoverable; the captions are unlabeled. This is a dated architecture and workflow discussion, not evidence of a named manager’s model, customer, permissions, deployment, investment authority, or performance. See the capture note.

The archive also contains a title-blind alternative-data evaluation discussion, published April 14, 2024. The panel describes the evaluation “catch-22” as a capability and cost problem: cloud access and specialist vendors can reduce the initial burden, while internal data-science, data-strategy, and engineering work can be spread across many datasets (00:46–02:14). It then discusses trials before purchase, open-source tools, hypothesis-led experiments, guided samples, and vendor white papers as inputs to diligence (02:16–04:26). The captions are unlabeled and partial, and no vendor, customer, dataset, model, permission, deployment, or performance result is identified. See the capture note.

A further title-blind Beryl panel, What Sets Current Generative AI and LLMs apart, and What are their Use Cases?, published April 5, 2024, adds a more specific grounding and finance-use-case route. The recovered discussion covers advisor-conversation transcription and retrieval, investment-research ideation, internal question routing, document generation, and a shift from fine-tuning toward grounding long filings such as 10-Qs and 10-Ks (01:46–06:19). It then describes extracting features from risk and MD&A language changes across 10-Ks, with contextual language models proposed as an extension to a rules-based measure (06:23–07:59). The captions are unlabeled and partial; the organization, team, universe, model, backtest, permission, and result are not identified. See the capture note.

The Beryl archive also contains a title-blind Key Trends affecting Alternative Data recording, published April 14, 2024. The panel discusses customizable and private data, non-coder tooling, and the operational burden of cleaning, validating, testing, and legally sourcing many datasets. One panelist distinguishes document summarization from predictive time-series work and raises time-series language models as a forward-looking possibility (00:33–03:56). These are dated panel opinions, not a current manager strategy, named model, dataset, deployment, or performance claim. See the capture note.

A second title-blind Beryl panel, How to Streamline Alternative Data in the Most Cost Efficient Way?, describes an agentic data-ingestion pattern. An unnamed participant discusses point-and-click agents that pull from webpages, documents, databases, and APIs, apply process automation, sentiment or translation enrichment, and deliver formatted outputs, alongside permissions, audit logs, transparency, and governance (00:15–01:19). The panel also connects qualitative-text/NLP ingestion and web or employment data with changing macro conditions (02:11–04:36). The captions are unlabeled and do not identify the platform, customer, fund, model, data rights, production deployment, or performance. See the capture note.

A third title-blind Beryl recording, Where does the Alternative Data Vendor’s Responsibility end, and the Fund’s work begin?, makes the diligence boundary explicit. The panel assigns value testing to the data consumer while assigning the vendor responsibility for accuracy, timeliness, clarity, and collection methodology; it then separates raw feeds from usable interfaces/APIs and discusses how processing depth affects audience, pricing, and interpretation (00:00–02:50). It also raises vendor continuity, synthetic-data risk, contracts, and third-party diligence (02:53–03:25). The captions are unlabeled and identify no provider, customer, model, permission, deployment, or performance result. See the capture note.

A correctly matched Beryl recording, How Accurate are LLMs given the Challenges of Obtaining Proprietary Information?, adds a licensing and data-access boundary. One participant says a Bloomberg Events News Alert System product was evaluated for analytics and possible model training, but that the training plan was stopped after legal review identified contractual limits (02:06–02:50). The panel also discusses time-varying data value and a multi-source operating layer around human judgment (04:52–06:18). This is a speaker account from a dated educational panel, not Bloomberg contract evidence or a named fund’s policy, model, deployment, or performance. See the capture note.

What are the Key Elements to Evaluating Alternative Data with Multiple Datasets? adds a cross-sectional data-quality route. The panel discusses breadth across comparable companies, historical depth, point-in-time treatment, resilience to missing or erroneous observations, entity resolution, parent-company mapping, satellite-data interpolation, and data-use restrictions (00:15–05:37). The captions are automatic and unlabeled; this is methodology evidence, not a named manager’s data architecture, model, permission set, or performance. See the capture note.

What Modeling Techniques are employed in Alternative Data Analysis? adds an operational reliability route. The panel discusses orchestration across many streams, missing/stale/noisy data, interpolation and replacement proxies, ticker and entity mapping, cloud delivery, copilot-assisted domain work, and the continued role of human monitoring when data breaks (00:37–05:53). The discussion does not identify a current fund, model, vendor, deployment, or result. See the capture note.

With Abundance of Data, how do we efficiently deliver it to End Users? adds a delivery-layer route. The panel discusses cloud access, natural-language interfaces for domain experts, composable services, infrastructure-as-code, security review, cost estimation, and limiting model infrastructure to required operating hours (00:00–07:20). These are dated platform-architecture discussions, not evidence of a named manager’s cloud deployment, budget, permissions, or performance. See the capture note.

August 26 title-blind alternatives-panel route

The corrected Beryl archive mapping also surfaces a title-blind discussion of risks introduced by patchy alternative data, published April 14, 2024 (publisher page). A participant contrasts short histories and regional gaps in alternative data with the long histories available for traditional factors, then recommends starting with an economic hypothesis and testing each intermediate link rather than relying on price correlation or a backtest alone (00:00–02:03). The panel also discusses data shelf life, the possible value of combining independent sources, and shorter decay horizons for some timely datasets (02:48–04:34; 06:16–06:57). The captions are automatic and unlabeled; this is dated methodology evidence, not a current manager’s validation policy, named dataset, or performance result. See the capture note.

The same archive contains a title-blind panel on anticipated LLM and AI developments, published April 14, 2024 (publisher page). Its 2024 forecasts emphasize use-case-specific models, data quality, proprietary data, quantitative inputs, and Transformers applied to particular finance problems rather than one universal model (00:18–02:37; 06:03–06:39). A participant says time-series foundation-model work was being followed and tested at that time (03:35–04:07). These are panel forecasts, not evidence of a tracked firm’s model ownership, training corpus, deployment, or results. See the capture note.

The archive’s alternative-data investing-process discussion, published April 14, 2024 (publisher page), adds a participant’s spoken account of work at “Chimera.” The account separates nowcasting from forward projection, describes combining independent datasets, and gives cohort and cross-shopping analysis—including customer migration in a pet-commerce example—as ways to study business activity (00:00–03:18). A second participant defines alternative data broadly to include text, credit-card, satellite, and biographical sources for questions about intangibles, sentiment, and forward-looking indicators (03:40–04:57). “Chimera” remains a spoken, dated account rather than an independently verified current affiliation or deployment; the automatic captions do not establish a named fund, dataset, model, permissions, or performance. See the capture note.

A fourth title-blind Beryl panel on demand for alternative data, also published April 14, 2024 (publisher page), records a historical data-buyer account from an unnamed pension-fund innovation lab. The speaker describes testing roughly 20 datasets with commodity and fixed-income teams; other panelists mention historical crypto order books, retail order flow, port and shipping logistics, Chinese trucking observations, geolocation, credit-card receipts, and earnings-call language as examples of data demand (00:00–04:18). The “deception” wording in the discussion is retained only as a discovery lead, not as evidence of a verified detector or trading result. Captions are automatic and unlabeled, and no current institution, vendor, data-rights arrangement, model, deployment, or performance is established. See the capture note.

The Beryl archive also contains a title-blind Google and buy-side data-infrastructure session, listed April 4, 2024 (publisher page). A Google representative describes a data-exchange environment, BigQuery and Analytics Hub access, and an AI/machine-learning layer over retail, geospatial, store-buying, exchange, and traditional market data (00:13–02:59). The recording says some buy-side users were already combining these sources, but it names no customer, model, contract, permission set, production system, or result. The speaker’s surname is not reliably recoverable, so this remains a vendor-side platform account. See the capture note.

A separate title-blind Beryl discussion of alternative data across bull and bear markets, published April 14, 2024 (publisher page), adds a validation and governance route. The panel discusses stress-period testing, the difficulty of identifying a regime in real time, and the risk that apparent regime-switching benefits may cluster in isolated events (00:10–02:00). It also links enterprise data organization to building AI models from internal and external sources and describes physical-asset, geospatial, satellite, and regulatory data for supply-chain and climate-risk questions (02:01–05:58). These are dated panel and vendor claims, not a tracked firm’s regime model, governance implementation, investment authority, or performance disclosure. See the capture note.

Three older Beryl title-blind routes add named practitioner and operating-context evidence. In a January 7, 2020 interview with Adrian Sisser, the Seven Eight Capital founding partner describes a quant-equity/statistical-arbitrage workflow and says the firm had used substantial ML/AI for a long time (01:29–02:14). The recording and description do not identify models, features, data rights, live permissions, or a reproducible result; a separate approximate-return statement in the description is promotional and lacks period, benchmark, counterfactual, and audit details. A second January 7, 2020 interview with Robert Ciemniak describes research and data-collection automation and neural-network classification of short text, including Twitter feeds and news headlines (00:36–01:23). That is vendor-side implementation context, not evidence of a hedge-fund customer, current deployment, or benchmarked accuracy. Finally, a December 22, 2019 interview with Olga Kane frames alternative-data diligence around written sourcing policies, internal versus outsourced collection, quality verification, vendor concentration, and replacement planning (01:32–02:54). These dated recordings extend the discovery map but do not establish current practice at a tracked manager. See the capture note.

The queue also contained a missing episode-level record for the December 9, 2024 Beryl EDU discussion of AlphaSense and unstructured-content search. The publisher description names Evan Reich as Verition Fund Management’s Global Head of Data Strategy and Sourcing, alongside leaders from RavenPack, AlphaSense, Bain, and TD Cowen. The recording describes semantic expansion across filings, transcripts, broker research, and expert-network material; search over internal analyst notes and theses; and a generative-search mode that returns source citations (01:04–04:02). AlphaSense’s time-compression and client-use statements are vendor claims, not independent time studies or proof of Verition adoption. The public material does not disclose Verition’s tools, permissions, model provider, data-isolation design, or production workflow. See the capture note.

The title-blind sweep also recovered a 2019 Learn Data Science meetup recording whose description links to Two Sigma’s official Using News to Predict Stock Movements Kaggle competition. The official competition page documents a historical public research artifact: Two Sigma as host, a news-to-stock-movement task, market and news fields, and a ten-day market-residualized forward-return scoring target. This is evidence of a public competition schema and community research surface, not proof that any notebook or winning method entered Two Sigma’s production stack. The recording had no usable caption track in the capture, so the source note relies on publisher metadata and the official Kaggle pages.

The title-blind queue also surfaced Deloitte’s November 14, 2024 IMpact interview with Niall Hurley, identified by the publisher at recording time as Eagle Alpha’s CEO. Deloitte’s publisher transcript describes Eagle Alpha’s alternative-data discovery, profiling, delivery, collection-rights and PII/MNPI checks, lagged testing, backtesting, and ongoing compliance monitoring (09:01–15:00). This adds a vendor-governance control-plane route and a named alternative-data executive; it does not establish Eagle Alpha’s current status, a client roster, a model, or any fund’s internal permissions. See the capture note.

A separate Hedgineer Podcast episode with Niall Hurley, published September 2, 2024, adds a more technical alternative-data supply and preparation route. The episode describes card-transaction data for retail nowcasting, mobile panels, IoT, regulatory filings, web scraping, and corporate CRM/ERP systems as data sources (02:01–02:28; 08:00–08:28). It also describes using embeddings and cosine similarity to map companies, drugs, events, and other entities across datasets without a shared identifier, followed by thresholds and review prompts (45:15–48:48). The speakers say data preparation can consume roughly 80% of work on complex datasets and that some mappings can be reduced from weeks or months to seconds or about a minute; those are practitioner estimates, not independent benchmarks (43:43–45:10; 47:05–48:48). This remains vendor-side evidence: it does not identify a customer, data right, model revision, production deployment, investment permission, or result. See the capture note.

The sweep also recovered the July 11, 2022 Flirting with Models interview with Ralph Smith, identified by the publisher as BlueCove’s Head of Research. The publisher describes BlueCove’s corporate-credit and interest-rate mandates and a discussion of research organization, liquidity and bond-availability assumptions, and appropriate backtest use. This is a useful fixed-income quantitative-process control, not an AI disclosure: neither the episode metadata nor the recovered recording establishes an AI/GenAI program, model inventory, production deployment, or performance result. See the capture note.

A Global Alts Miami 2026 panel, published by iConnections, adds three distinct public lanes that do not appear in a hedge-fund job-title search. Ryan Teal, Head of Operational Due Diligence at Albourne Partners, describes breaking diligence into tasks and subtasks and reviewing them through ML, AI, and agent lenses; he names document review, offering documents, questionnaire flags, manager benchmarking, and background checks, and says meeting transcription was still not productionized in his account (19:51–21:49). Chris Ackerson, SVP Product at AlphaSense, describes an AI interviewer that researches a topic, interviews experts asynchronously, and creates channel-check data for sector demand and price signals (08:02–10:22). Amadeo Alentorn, Head of Systematic Equities at Jupiter, separates long-standing statistical learning from GenAI and says GenAI has accelerated the path from research ideas and academic papers through testing and implementation; he gives a speaker-reported estimate that the number of improvements reaching the live model roughly doubled (05:58–07:31). The panel also emphasizes human responsibility, source citations, vendor due diligence, data-sharing controls, AI policies, and incident response. These are named speaker and vendor claims, not independent evaluations or disclosures by GMO, Acadian, Arrowstreet, or another tracked manager; they do not establish model weights, customer identity, permissions, or AI-attributed performance. The capture note records the survey and promotional-claim boundaries.

August 26 title-blind founder/platform route

A Harrington Starr TV interview with Zuber Seth, introduced as Orchid’s co-founder, adds a founder-side platform route that a search limited to hedge-fund names or AI job titles could miss. Seth describes Orchid as an AI platform for asset management and says the project began as a hedge-fund concept before being repackaged as software after an FCA approval problem related to the founders’ age. He says the initial workflow covered equity and macro research, model work, and parts of analyst activity; he gives an early speaker-reported estimate of roughly 20–30 minutes of analyst time saved per day (00:16–06:26). Those are founder and host claims, not a regulator finding, a verified KIA relationship, a customer case study, or a measured productivity benchmark.

The product description is more specific than a generic “AI assistant” label. Seth describes a set of task-specific agents with routing, and a planned modular version that would construct a workflow for each prompt rather than depend on a fixed agent roster (08:20–10:08). He also identifies compliance and back-office document work as augmentation targets and says Orchid could make people twice as fast in some workflows (19:30–21:37). The recording does not disclose the model family, tool permissions, data rights, customer identities, evaluation design, error rates, production telemetry, or investment authority. It therefore supplies an architecture and diligence lead, not evidence of live trading or relative capability.

The public status is time-sensitive. Orchid’s current asset-management page positions a private LLM for asset managers with research, risk-attribution, and portfolio-construction capabilities. Companies House records ORCHID AI LIMITED as active and incorporated on January 22, 2025, while its officer record records Seth’s directorship as resigned on November 10, 2025 and its PSC record records his control as ceased on November 11, 2025. That corrects the temporal interpretation of the interview: “co-founder” is a dated description, not automatically a current officer title. The capture note keeps the speaker claims, current product positioning, and corporate records separate.

August 26 finance-risk conference route

August 26 title-blind alternative-data practitioner route

The March 5, 2026 Momentum in B2B Tech with AI episode, also indexed by Amazon Music and Listen Notes, adds a data-strategy route that a fund-name or AI-title search can miss. The publisher identifies Jack Killea as Head of Data Strategy and Sourcing at Maiden Century. In the public English caption track, Killea discusses sourcing and testing data, the difficulty of converting data volume into usable context, and a workflow architecture connecting data, models, and role-specific tasks (01:30–01:46; 14:42–15:10; 19:34–20:35). He also says his team is applying agentic workflows to internal processes and suggests starting with manually performed work across research, data science, engineering, sales, and support (31:00–31:28). This is a speaker-reported practitioner and vendor-adjacent account. It does not identify a named hedge-fund customer, model family, training corpus, data-rights arrangement, production permission, trading deployment, or performance result. See the capture note.

The NYU Stern Volatility and Risk Institute’s 2026 recording adds a finance-risk and governance route that is easy to miss when discovery is limited to manager names. The event page identifies Vasant Dhar as moderator, Melissa Koide of FinRegLab, and Andrea Bonime-Blanc of GEC Risk Advisory for the “AI Opportunities and Risks” panel. The discussion asks whether AI increases or decreases systemic risk and focuses on model type, training data, deployment context, nondeterminism, explainability, identity, privacy, stakeholder impact, third-party evaluation, and cross-disciplinary review (00:04–16:31). This is governance and financial-system context, not evidence of a named hedge-fund deployment.

The conference agenda also exposes a useful personnel and research graph: Petter Kolm’s deep-learning work on limit-order books; Marcos López de Prado’s ADIA quantitative-research leadership and causal-factor-investing topic; Bryan Kelly’s AQR machine-learning leadership and news-pricing research; and research routes at NYU, Harvard, and Stanford on asset embeddings, inflation uncertainty, and financial stability. Those agenda entries establish public conference participation and research topics. They do not establish that any paper, method, lab, or model is owned or used by a live investment strategy. The timestamped capture note preserves the separation between the recording, the official agenda, and the evidence gaps.

August 26 MIT finance-decision research route

The MIT IDE 2026 presentation by Eric So adds an academic route for testing financial research agents. MIT identifies So as lead of its AI in Financial Markets and Decision-Making group, whose research examines how AI, human behavior, and market incentives interact. The presentation covers three streams: a purpose-designed financial-advisor intervention versus an off-the-shelf LLM, an experiment varying whether a reporting assistant is told to maximize profits, and the effect of AI reliance on later human recall (00:36–06:31).

The most transferable diligence idea is to treat the objective in the prompt as an experimental variable. So says that adding profit-oriented language to otherwise similar reporting tasks changed risk escalation and urgency judgments in his study (06:34–11:18); he also presents a speaker-reported 83% recall result from a student study (13:19–15:05). These are academic presentation claims, not evidence about a hedge fund’s system. The recording does not disclose a commercial model, a fund deployment, full study instruments, or a performance result. The capture note records the methodology gaps and the follow-up paper routes.

August 26 India/CFA investment-research route

An Indian Association of Investment Professionals presentation with Vikram Srinivasan adds a regional, title-blind route into investment-research infrastructure. The recording introduces Srinivasan as Needl.ai co-founder and identifies Kuntal Shah as his partner; Oaklane Capital separately lists Srinivasan as an advisor and describes his technology background. The talk describes a data layer spanning public web, social posts, regulatory and exchange sources, enterprise drives, email, chat, and notes; a RAG/search and custom-agent layer; source-linked reports; and on-premise or customer-cloud deployment for data-localization requirements (04:37–10:26). Needl.ai’s current product pages provide a separate vendor positioning route.

The recording also demonstrates a portfolio/company-monitoring workflow using earnings calls, analyst reports, filings, and investor presentations, and describes a vendor-reported APAC renewable-energy trading case with Australian and New Zealand exchange alerts. The speaker claims the traders made up to $250,000 per day after three weeks; the customer, baseline, P&L attribution, and counterfactual are not disclosed (15:01–22:22). This is therefore a useful source-factory and verification lead, not a verified customer outcome or hedge-fund deployment. The capture note records the regional route and evidence boundaries.

August 26 title-blind systematic-manager operating-model route

The title-blind Flirting with Models interview with Kevin Cole, published August 29, 2022, identifies Cole as CEO and CIO of Campbell & Company. He describes a systematic multi-strategy research process in which ideas are evaluated against the existing model collection, tested on held-out out-of-sample data, peer reviewed, and continued, reduced, or retired (18:54–24:22). He also describes research groups organized around investment styles and risk, with data engineering, data science, and core infrastructure functions (35:11–36:57), and a recurring project-governance process called “Pulse” (37:15–40:35).

This is a useful dated non-GenAI baseline for understanding a systematic research factory: it supplies operating-process and personnel-function evidence, not evidence of Campbell’s current AI, LLM, or agent stack. The episode does not disclose model providers, training data, permissions, production telemetry, live investment authority, or performance attribution. The capture note preserves the source and temporal boundaries.

Third Point: founder-level AI adoption and a retained human interface — August 2026

Third Point’s official company LinkedIn post points to Daniel Loeb’s Invest Like the Best interview, also distributed through Apple Podcasts and YouTube. In the public caption track, Loeb describes bringing in computer-science-native AI experts for specific projects, having them coach the broader team, using a system integrator, and encouraging employees to find applications. He says some staff run agents overnight while others use AI mainly for queries, and refers to Claude in the context of individual self-improvement (58:08–62:56). He also describes a human interface remaining in capital allocation and frames technology fluency plus industry understanding as part of the modern analyst role (18:00–18:43; 60:45–61:20). These are founder statements and automatic-caption evidence about adoption and organizational boundaries. They do not establish a Third Point AI lab, model inventory, permission map, production endpoint, order authority, or AI-attributed performance. See the timestamped capture note.

Voya Machine Intelligence: a dated AI-equity and human-machine disclosure

A Beryl EDU recording published in the archive on April 14, 2024 adds a named asset-manager route. Gareth Shepherd, cross-checked against Voya’s 2020 announcement, describes Voya Machine Intelligence as running dedicated AI equity strategies, separating workflow productivity from alpha generation, and combining AI signals with discretionary processes (00:28–02:33; 04:17–04:48). He names LLMs for reporting, stock narratives, news-trigger identification, and ESG research, and describes reinforcement learning and Bayesian networks in fixed income (02:50–04:02).

A separate Nasdaq TradeTalks recording, published May 17, 2023, identifies Shepherd as Co-Head of Voya Machine Intelligence and Portfolio Manager. He describes AI applications across portfolio construction, basic research, and client reporting, characterizes investing as a non-linear problem, and emphasizes human curation of the data the system can access (02:41–03:02; 03:15–03:40). The recording is a dated practitioner account: Nasdaq provides no written transcript, and it does not identify a model family, training corpus, evaluation design, permissions, production authority, or independently measured performance. It should not be merged with later Voya strategy materials without reconciling dates, vehicles, and definitions. See the timestamped capture note.

Nasdaq also hosts a distinct May 20, 2024 TradeTalks x SALT recording, identifying Shepherd as Managing Director and Portfolio Manager at Voya Financial. In that later recording, he describes early use for idea generation and portfolio red-flag review before a human portfolio-manager discussion (03:07–04:02). This is a separate dated source, not a correction to the 2023 page; Nasdaq provides no written transcript, and the recording does not identify a model, data source, evaluation, permissions, deployment authority, or independent performance. See the 2024 capture note.

A second Beryl recording from the same 2024 archive adds a distinct causality-and-forecasting route. Its publisher description names Shepherd with Voya, Peng Cheng with J.P. Morgan, and participants from Robotic Online Intelligence, Scotiabank, and NVIDIA. The panel describes causal graphs for intraday idiosyncratic risk, causal discovery for separating confounders, mediators, and colliders, and counterfactual controls for estimating the remaining-day impact of a timed price-target change (00:29–04:59). A Voya participant then discusses transformer models for selected return and volatility forecasts (05:20–06:49); later discussion describes incremental productionization, human-observable signals, and an LLM as a possible bridge from metadata to an initial causal graph before numerical time-series modelling (10:02–12:22). Voya’s 2020 announcement and Beryl’s Peng Cheng profile provide separate historical role cross-checks; J.P. Morgan’s AI research page and market-making reinforcement-learning paper provide institutional context. The recording is a dated panel with unlabeled automatic captions: it does not establish a current Voya or J.P. Morgan model, customer, open-source dependency, live permission, performance result, or current reporting line. See the capture note.

The public record requires temporal and scope discipline. The recording reports more than a billion dollars and a live track record, but those are speaker-reported figures without an independent audit or a clear aggregate/vehicle definition. Voya’s current Machine Intelligence strategy page and 2026 strategy brief describe proprietary non-linear ML for fundamental analysis and state that the models described in those strategy materials do not use generative-AI algorithms. Voya’s June 2026 leadership announcement separately places Christine Cappabianca over the combined Machine Intelligence and Quantitative Equity suite. These sources do not demonstrate that the products, dates, figures, models, or reporting lines are identical. See the capture note.

Heptagon Capital: thematic AI investment framework, not an internal technology disclosure

Nasdaq’s 2026 TradeTalks Q&A identifies Alex Gunz as a Fund Manager at Heptagon Capital and records the firm’s public Future Trends framing. Gunz places AI alongside cybersecurity, robotics, data-center power, quantum, and space within a long-term, multi-thematic process. He says the process seeks themes that can grow in importance across economic conditions and where public policy, if involved, is a tailwind. His 2025 Q&A uses similar criteria and discusses AI’s possible long-run productivity effects. A related 2024 Nasdaq video places Gunz alongside Harmonic Security and Bailard in a cybersecurity discussion. Although Nasdaq supplies no written transcript, the public JW Player recording was recovered and processed locally: it provides navigation for data-privacy and employee-use concerns (01:37–02:20), Harmonic’s speaker-reported observation of roughly 5,000 AI applications across protected enterprises (02:10–02:20), data-governance and zero-trust framing (09:14–09:28), and log-based anomaly/pattern detection (15:03–15:27). Those claims belong to the cybersecurity speakers, not to Heptagon. The recording adds investment-context evidence about cybersecurity and enterprise-AI risk; it does not disclose Heptagon’s internal AI systems, model inventory, training data, permissions, deployment, or performance. See the updated capture note.

Late-queue verification: Dimensional, Jump, and personnel movement

The expanded search recovered a current Dimensional founder interview that was not in the article’s normalized media registry. The Meb Faber Show episode index lists “David Booth: 45 Years to $1 Trillion at Dimensional | #646” on August 21, 2026. The publisher describes Booth as Dimensional’s founder and says the episode discusses the firm’s history, its University of Chicago and Eugene Fama lineage, reported scale, and AI investing as a market theme. This is current founder/media and investment-framework evidence. It does not disclose a Dimensional AI lab, internal model inventory, training data, agent permissions, production deployment, or AI-attributed performance.

The same pass recovered a distinct academic-discovery surface for Jump Trading. The UCL ELLIS Computational Statistics and Machine Learning seminar page says the series is sponsored by Jump Trading, connects the series to UCL’s AI Centre, Gatsby, computer-science, and statistics communities, and says most recordings are uploaded to YouTube. Its 2025–2026 archive lists talks on reinforcement learning, neural-network training, AI research agents, contextual bandits, and Bayesian online learning. This verifies sponsorship and a route for episode-level recording recovery. It does not identify Jump employees among the speakers, establish Jump ownership of the academic work, or disclose a Jump trading model, permissions, or performance.

A third route strengthens temporal personnel tracking without being folded into current Balyasny evidence. Citi Research’s E46 podcast, dated August 28, 2025, identifies Heath Terry as Citi’s Global Head of Technology and Communications Research and states that he previously led centralized TMT research at Balyasny before joining Citi in May 2025. The episode discusses AI use cases, enterprise technology, and private-company research. It supports a dated former-employer affiliation and current Citi role; it does not establish current Balyasny employment, Balyasny AI ownership, or any model, permission, deployment, or performance claim. The capture note preserves the source boundaries.

FinLLM 2026: conference and academic-network route

The first-party FinLLM 2026 workshop page records an August 15, 2026 International Symposium on Large Language Models for Financial Services in Bremen, organized with E Fund Management, Tsinghua University, The Hong Kong Polytechnic University, the University of Manchester, the University of Kent, and Wuhan University. Its organizing committee identifies Qiang Yang and Liyuan Chen as general chairs and Shuoling Liu as executive chair; the page describes Liu as E Fund’s Chief Information Officer, a FinTech Executive Committee member, and head of its Innovation Lab.

The agenda is a useful title-blind map for further research. It names FinSAgent for SEC-filing analysis, open-weight financial-text evaluation, multi-agent sentiment analysis, document-grounded DCF valuation, financial-statement fraud detection, unstructured-data Treasury regime detection, trustworthy AI and accountability, market-feedback adaptive RAG, LLM-agent benchmarking for algorithmic trading, financial-news extraction, autonomous financial-signal discovery, quantized-LLM memory editing, and multi-agent trading with an internal contest mechanism. The workshop call also lists multimodal and cross-lingual finance, token economics, model collaboration, permission controls, hallucination mitigation, and regulatory alignment.

The programme establishes a conference, institutional, and research-topic route. It does not establish that E Fund or any listed institution built, funded, owns, or deploys a named system, nor does it provide model weights, dataset rights, evaluation splits, recordings, permissions, investment authority, or independently measured performance. Treat author affiliations and agenda titles as leads for separate paper, code, personnel, and recording verification. See the first-party workshop page.

The agenda led to the primary FinSAgent preprint, revised July 21, 2026. It describes role-specialized agents anchored to 10-K structure, corpus-conditioned query decomposition, and multi-path retrieval with a feature-gated reranker that separates evidential validity from semantic similarity. The paper documents five financial QA benchmark routes, ablations, human review, noise-sensitivity testing, latency, and error analysis. Its author list does not, on the abstract page, establish that E Fund, SimpleWay.AI, McGill University, or another workshop-affiliated institution owns or deploys the system. This is academic methodology and author-reported evaluation, not evidence of a hedge fund’s production system, trading use, data rights, agent permissions, investment authority, or independently replicated performance. See the primary paper and workshop programme.

The indexed first-party agenda also exposes a speaker and institution graph that was not visible in the earlier topic-only capture. It lists MIT’s Zijie Zhao on market-feedback adaptive RAG and event-driven market-impact signals; Manuela Zaidan of Intesa Sanpaolo on generative and agentic AI in creditworthiness; Chun Chet Ng of AI Lens on financial retrieval; Irem Demirtas of Prometeia on system-level validation; Mengting Chen of StepFun on a multi-agent trading system; and university researchers working on fraud detection, Treasury regime shifts, financial-news extraction, quantized-model memory, and algorithmic-trading benchmarks. Four talks are marked pre-recorded, but the public page does not expose their video, audio, slides, or transcripts. These are event-listed affiliations and research topics, not evidence of investment-firm employment, deployment, model ownership, or performance. The source is the first-party FinLLM programme.

The workshop call also frames cloud/small-model collaboration, local-model privacy and latency, token economics, permission controls, hallucination mitigation, systemic stability, and cross-border regulatory alignment as central questions. It uses “OpenClaw” and a “One-Person Company” paradigm as discussion examples for autonomous financial workflows. This is organizer framing, not evidence that an investment firm uses that framework or has granted an agent portfolio or order authority.

The personnel route is now more specific. Jiangpeng Yan’s self-authored profile identifies him as a research director in E Fund’s Intelligent Solution Research Center, with a stated focus on financial foundation models and agentic engineering for investment research, risk management, and decision-making. The profile records a Tsinghua Ph.D. under Prof. Xiu Li, prior Tencent AI Lab and Huawei work, and selected finance-AI publications. The publisher page for When DeepSeek-R1 meets financial applications independently lists Yan as an author and E Fund Management as an author affiliation. Separately, the E Fund-organized NLPCC 2026 investment-advisor-agent task repository lists Yan as a task contact and describes daily macro and sector ETF allocation agents using financial news and historical prices in an academic backtest. These sources establish a public personnel, research-lineage, and academic-task connection. They do not establish E Fund production deployment, model ownership, proprietary data, agent permissions, investment authority, or AI-attributed performance.

Dimensional: named 2026 ML research route

Xing Hong’s February 2026 paper identifies Dimensional Fund Advisors as the author’s affiliation and studies machine-learning volatility forecasts using intraday S&P 500 data from 1998–2023. The abstract compares linear ML, gradient-boosted trees, and neural networks with heterogeneous autoregressive benchmarks, and evaluates target-volatility portfolio control. It reports similar out-of-sample performance between linear ML and HAR, a predictive increment for the tested gradient-boosted-tree models among nonlinear methods, mixed neural-network results, lower results from more granular predictor sets, and modest incremental target-volatility gains from ML. FINRA’s public report independently records Hong as a Dimensional researcher in its employment history. This is a dated research and personnel disclosure, not evidence of a firm-wide AI lab, GenAI program, production model, model ownership, training-data rights, agent permissions, live investment authority, or AI-attributed performance. The earlier 2024 volatility paper with Wei Dai, Hong, Robert C. Merton, and Mathieu Pellerin is retained as a separate research route. See the capture note. The title-blind Financial Coconut episode with Wei Dai was missing an accessible text track, so the public Acast enclosure was recovered and processed with local English ASR. Dai describes a global research agenda focused on systematic return drivers and their translation into real-world portfolios (02:52–04:09); she later describes academic participation in the Investment Research Committee’s scrutiny of portfolio additions (20:31–20:50). These are automatic-ASR paraphrases, not exact quotations, and add research-process context rather than a Dimensional AI or GenAI disclosure. See the audio-recovery note.

The pre-recorded-title follow-up recovered two additional public implementation routes. PRISM and the accompanying AI Lens repository expose a financial-document retrieval stack with prompt engineering, in-context examples, optional multi-agent coordination, four workflow variants, role-specific agents, concurrency controls, checkpoints, and Azure OpenAI configuration. The authors report evaluation on FinAgentBench, FiQA-2018, and FinanceBench, including latency, token, cost, and ablation analysis; the reported NDCG@5 is 0.71818 for the best stated FinAgentBench configuration. A second route, Backtrader-Bench and its public repository, publishes executable question verification, a 160-question pool, a balanced 30-question set, generator-solver screening, run-level artifacts, and a possible reinforcement-learning corpus route for quantitative-trading workflows. The papers and code are academic or fintech artifacts, not evidence of hedge-fund deployment, proprietary data, investment authority, or live returns. Reported benchmark figures remain author claims bounded by their stated test designs.

The late-queue pass also found an upcoming firm-hosted talent and research surface. Man Group’s Sofia AI Hackathon is scheduled for October 24, 2026 and is described as a one-day exercise using curated financial datasets in a market-simulation environment, with Man technologists and portfolio-manager guidance. The participant profile includes data scientists, machine-learning engineers, and software developers without requiring prior finance experience. This is an event and recruiting signal. It does not establish that the simulated portfolios, prototypes, or participant work reach production or receive trading permissions.

A separate D. E. Shaw first-party article, Machine Teaching: What I Learned From My Optimizer, adds a dated investment-workflow route. The page describes optimizers used in systematic and discretionary contexts and discusses iterative interaction between portfolio managers and computational tools, including sensitivity to inputs and constraints, transaction costs, position sizing, correlations, crash scenarios, and common-investor risk. It presents the optimizer as a way to surface assumptions and behavioral biases. This is a firm-authored optimizer and human-machine workflow description, not a GenAI or LLM disclosure; it does not identify a model family, training corpus, named AI personnel, agent permissions, production endpoint, or independently measured performance. The page carries a 2023 copyright notice; the capture note preserves that date boundary.

The next firm-media gap pass recovered four distinct routes. Renaissance Technologies’ official careers page lists research, data, financial-infrastructure, real-time-trading, research-infrastructure, systems, security, and data-center roles in East Setauket and New York. It does not use AI or ML labels, identify filled personnel, or disclose models, permissions, deployment, or performance; it is also separate from the unrelated Renaissance Learning renaissance.com site. XTX Markets’ official home page says machine-learning price forecasts support trading and liquidity provision across multiple asset classes, and describes a growing research cluster and Finland data-centre project. Those are current company statements; the page does not identify model architectures, training data, evaluations, individual ownership, permissions, or independently audited results.

The personnel and academic-lineage lane adds one carefully bounded Renaissance route. Joao Carreira’s self-authored profile says he currently works at Renaissance Technologies and previously completed a UC Berkeley Ph.D. in the RiseLab on large-scale serverless systems for machine-learning workloads, advised by Randy Katz and Pedro Fonseca. It lists public systems/ML work including Cirrus, Pronghorn, and A Case for Serverless Machine Learning; Fonseca’s Purdue profile independently lists the co-authored papers and describes his reliable and secure systems lab. This establishes a self-reported current-affiliation and academic-lineage route, not a Renaissance-owned publication, named AI responsibility, financial model, GenAI deployment, or investment authority. See the capture note.

Two Dimensional routes add temporal and title-blind context. Its October 10, 2025 “Will AI Take Your Job?” episode features Mark Gochnour and University of Chicago economist Kevin M. Murphy and discusses AI’s substitution/complementarity effects on work and Dimensional’s academic-network framing. Dimensional’s December 3, 2025 leadership announcement records Wei Dai becoming Global Head of Research and Peter Dillard becoming Global Head of Investment Engineering on January 1, 2026, while Savina Rizova remains Co-CIO and chairs the Investment Research Committee. The announcement describes academic collaboration, Dai’s empirical research, Dillard’s former Chief Data Officer role, and a team of more than 70 investment-system experts. Neither source establishes a Dimensional GenAI owner, AI lab, model inventory, training corpus, agent permissions, production deployment, or AI-attributed performance. See the gap-pass capture note.

An additional Australia-facing Australian Financial Review / Chanticleer interview, dated October 27, 2024, identifies Gerard O’Reilly as Dimensional’s co-CEO and co-CIO and former research head. In its AI section, O’Reilly describes AI as assisted implementation: likely useful for coding, analysis, and trading-workflow efficiency, but not necessarily a major way to pick stocks or out-guess market prices. The interview also points to proprietary historical pricing and accounting data and gives Australia-specific operating context. This is a dated executive position distributed through a firm-controlled document route, not proof of a current firm-wide policy, AI lab, model inventory, training corpus, deployment, agent permissions, or AI-attributed performance. It also does not merge DFA Australia with other Dimensional legal entities or with any fund vehicle. See the capture note.

The title-blind media and personnel pass adds public routes around the same research organization. Rob Harvey’s April 27, 2025 Behind the Ticker episode identifies him as a Dimensional vice president and links episode-level YouTube, Buzzsprout, and Spotify routes. Its show notes describe evidence-based ETF construction, daily portfolio management, and a high research bar for portfolio changes. Orion’s Savina Rizova episode identifies Rizova as Co-CIO and Global Head of Research, records her Dartmouth/University of Chicago lineage, and places big data, AI, and machine learning in the public outline at approximately 28:20–31:37. Orion’s Gerard O’Reilly episode identifies O’Reilly as Co-CEO, Co-CIO, and former Head of Research, with AI and innovation placed at approximately 32:10. These routes add personnel, academic-lineage, and media-surface context; they do not establish a Dimensional AI lab, GenAI owner, model inventory, training corpus, agent permissions, production deployment, or performance attribution. The event catalog is retained as a first-party enumeration route for recorded webcasts and future event artifacts. A public researcher job listing adds hiring-intent evidence for research spanning idea generation through publication or strategy implementation, large datasets, statistical methods, and programming; it does not prove that the role was filled or identify an AI owner. See the title-blind capture note.

The same gap pass recovered two additional first-party surfaces. Systematica’s current home page lists trend-following, macro non-trend, multi-strategy, equity market-neutral, and customized-solutions strategy families and global offices in Europe, Asia, and the United States. It is a firm and strategy surface, not an AI disclosure. Numerai’s September 4, 2025 engineering update announces Numerai-Tools v0.5.0 for submission cleaning/validation and churn/turnover calculations, changes to Signals diagnostics and staking, a fireside-chat route, and a hiring link. This is dated product and tournament tooling evidence; neither item establishes model inventory, training data, agent permissions, live fund architecture, or independently audited performance. See the capture note.

An official-media pass also recovered two Balyasny and two Citadel routes. Balyasny’s News & Insights index links firm-produced coverage of its annual AI hackathon and Applied AI team, while its Technology page describes applications, infrastructure, datasets, and AI tools supporting investment, risk, and business-infrastructure functions and identifies Mike Grimaldi as CTO. These are first-party media and operating-surface signals; they do not disclose model providers, training data, evaluation, permissions, production endpoints, or performance. Citadel’s EQR page identifies Perry Vais as Head of Equity Quantitative Research and describes a workflow combining data, quantitative research, technology, portfolio construction, and systematic strategies. Its February 20, 2026 EQR article presents an observation–modeling–action cycle linking data trails to research, models, liquidity, portfolio construction, execution, and systematic-strategy deployment. Neither Citadel page identifies an AI model, GenAI program, training corpus, agent permissions, or independently measured performance. See the official-media capture note.

Resolving the linked pages adds more specific signals. Balyasny’s April 21, 2026 annual AI Hackathon article says its Applied AI team hosted a global employee event in which participants built agentic workflows to automate complex tasks and improve daily processes. Its March 6, 2026 Applied AI article describes researchers, engineers, and domain experts building workflow-embedded tools with evaluation, traceable reasoning, and feedback loops, and identifies Charlie Flanagan as Chief AI Officer. These remain firm-authored/vendor-linked descriptions without a complete model inventory, training corpus, permission map, or independent outcome audit. Citadel’s August 17, 2026 EQR alpha-research article identifies Max Bane as a Quantitative Research Lead and places generative AI and agent systems at the research-process frontier alongside hypothesis generation, out-of-sample testing, cost/risk checks, simulation, and live-market feedback. Citadel’s March 17, 2026 Data Strategies Group article profiles Solène Chabanier and describes satellite imagery, prices, text, statistical/AI/ML processing, and distinct research-to-production responsibilities. Neither article identifies a specific GenAI model, training corpus, data license, agent permission, or independently measured performance. See the page-level capture note.

The next regional and platform gap pass added five distinct routes. Lingjun’s January 27, 2026 Chinese-language summit article identifies CIO Ma Zhiyu and describes continued prediction-model iteration, expanded short- and medium-horizon signal work, market-sensitive risk-parameter adjustment, and ML/AI as research tools. The English summary is a translation/paraphrase of the original page, not a substitute for it. The FIAM Montréal Summit 2026 program schedules sessions on AI in asset management, governance, deployment, and agentic finance, with named participants from BlackRock Systematic, Northern Trust Asset Management, Fidelity, Apollo, DCM Systematic, and McGill. Its biographies also expose cross-firm personnel and academic links, including a description of Finaix as an agentic-AI portfolio manager; that organizer-published description is retained as an unverified company claim. The program is a speaker and network surface, not a recording or transcript. See the regional gap-pass note.

A title-blind LinkedIn and careers pass adds a dated Jump Trading conference route. A public LinkedIn announcement says Jump sponsored ICML 2026 in Seoul and promoted an expo talk titled “Multi-Agent System Design and Evaluation for Quantitative Finance,” naming Lucas Baker and Loren Puchalla Fiore as speakers and describing work on building and evaluating multi-agent architectures at scale. It also mentions an AI4Physics workshop and discussions of AI research and alpha modeling. This is social-media announcement evidence, not a recovered talk or technical paper: the benchmark, model inventory, training data, permissions, and deployment are not public in the post. Jump’s official careers surface supplies a separate AI/ML recruiting route, but the checked role shell did not expose detailed job text; hiring intent is not filled-role or production evidence. See the capture note.

First-party follow-up pages sharpen two of those routes without expanding the claims beyond what is public. Finaix’s own site presents the company as an AI-driven asset manager and describes AI/ML-based portfolio-management models intended to adapt to market and behavioral changes. That page corroborates Finaix’s public AI/ML positioning alongside the FIAM program, but its registration statement is recorded as company self-description and the pages do not establish an autonomous product, live portfolio decisions, model family, or performance. Borealis Global Analytics separately names a “CIO Co-pilot,” Factor View, Macro View, and Sailesh S. Radha; its use-case page describes country-selection and rebalancing workflows, an AI analyst, and a database of indicators and returns spanning more than 30 years and 80 countries. The CFA UK webinar listing supplies the dated event route around AI-enabled country allocation. These are company- and event-controlled product surfaces, not evidence of model weights, data contracts, customer deployment, execution permissions, or independently audited results. See the first-party route note.

Optiver’s current research page now has a normalized registry record alongside its existing article citation: it describes an AI Lab spanning deep learning, reinforcement learning, LLM-based analysis, signal discovery, forecasting, execution quality, and risk decisions within a research lifecycle constrained by simulation and live implementation. A distinct July 22, 2026 Amsterdam event page describes a keynote on AI in research and engineering workflows, evaluation, and improvement of live systems. These are first-party research and event statements; they do not disclose model weights, vendors, training data, agent runtime, permissions, deployment telemetry, or AI-attributed trading results. Finally, a May 18, 2026 CafeMutual interview identifies Bhautik Ambani as CEO of AlphaGrep Mutual Fund and attributes to him AI use in data analysis, signal generation, risk monitoring, and anomaly detection within regulatory boundaries. That interview is a mutual-fund vehicle and media claim, not a technical disclosure or independent audit. See the regional gap-pass note.

A subsequent verification pass adds a new Citadel personnel route and a corrected XTX discovery path. Citadel’s April 28, 2026 EQR profile identifies Nick Butler as Head of Research and describes a data-to-prediction-to-trade pipeline spanning dataset identification, forecasts, market-impact modeling, portfolio construction, optimization, simulation, and market feedback. The page describes his advanced training in physics, applied mathematics, and economics and a PhD focused on game theory, but does not name the institution in the reviewed text. This is a named-person and research-workflow record, not a GenAI disclosure. A Chinese-language LinkedIn post published August 25, 2026 links to the already captured Atlas Wang Information Bottleneck episode and supplies a current regional-language/social route. The public YouTube pointer had no captions on recheck, so it is not counted as a second transcript or episode. The capture note preserves both boundaries.

Two additional conference routes extend the personnel and regional graph. Loyola University Chicago’s AI in Financial Services 2026 agenda places Carson Boneck, Balyasny’s Chief Data Officer, and Shirley Zhang, Balyasny’s Senior AI Engineer, on a May 14 session with LSV portfolio manager Guy Lakonishok about massively parallel architectures and world models in quantitative finance. The organizer’s description names high-volume data processing, synthetic data, counterfactual market-impact analysis, backtesting, and reinforcement-learning-agent training as agenda topics; a separate session covers autonomous agents, simulation, and stress testing with Northern Trust and BMO data leaders. This adds titles and a dated session surface, but does not identify which institution supplied which example or prove that any described system was deployed. The Monash–Q Group Finance Colloquium adds an Australia-based route dated April 17, 2026, with J.P. Morgan’s Head of Applied AI and Data Science for Global Research, a Monash paper on how AI agents can mislead investors, and a session on AI’s impact on research production. Its biographies add UNSW, Monash, and quantitative-equity research context. Both are organizer-controlled event and biography surfaces, not model, data, permission, production, or performance disclosures. See the capture note.

A separate 2025 conference route adds more implementation detail around Balyasny’s public research-platform description. The Enterprise Search & Discovery program describes a platform said to process more than 10 million documents daily, using custom embedding models, document chunking, diversity-driven reranking, AI-agent integration, and Excel compatibility; it also reports 200% year-over-year growth in data sources and user adoption. The related speaker biography identifies Shirley Zhang as Senior AI Engineer, says she develops and maintains the financial research platform and contributes to a firm-wide AI platform, and records her prior data-engineering role. A related public announcement uses “Data Engineer” for the later Loyola session, so the title difference is preserved as a dated discrepancy. The architecture and scale figures are organizer-published claims; no recording, model inventory, retrieval evaluation, provider, data-rights map, permission boundary, or independent audit was recovered. See the capture note.

A related Australian conference page adds a second dated title surface for J.P. Morgan’s Berowne Hlavaty. The Sydney AI in Financial Services Summit page dates the event May 7, 2026 and lists Hlavaty as “Global Head of Applied AI & Data Science.” It places him in a workforce-and-automation think tank with AMP’s Mike Way and Alceon’s Sze Ding, identified as Director of Artificial Intelligence and Principal of Level26 Labs. The agenda discusses automation, retraining, human oversight, governed analytics, model inventories, bias testing, monitoring, and agentic-system controls in Australian financial services. The Monash page uses “Head of Applied AI and Data Science for Global Research,” so both public labels are retained as dated surfaces. This is event metadata and a speaker-network route, not evidence of J.P. Morgan model implementation or a relationship between J.P. Morgan and Level26 Labs. See the capture note.

The next first-party verification pass adds three distinct Jump Trading routes. Jump’s current AI/ML page describes machine learning and AI across trading, research, and core infrastructure, including classical models, deep learning, reinforcement learning, LLMs, generative modeling, custom foundation models, LLM agents and assistants through API and HPC, NLP signal generation, and firm-reported compute, simulation, feedback, tool-use, and weekly-LLM-usage figures. The page does not publish model weights, training corpora, evaluation protocols, permissions, or independently audited performance. A March 17, 2026 NVIDIA Vera Rubin announcement names Joe Stam as Head of Research Technology and Alex Davies as Chief Technology Officer and describes a deep-learning financial-research data-center expansion; it also says LLM inference is one part of a broader model-architecture workload. This is a dated infrastructure and personnel statement, not proof that every described workload is live or tied to a named strategy. Finally, Jump’s fellowship page lists an AI/ML category and 2026–27 fellows at Harvard, MIT, and Stanford whose published topics include diffusion and generative models, robust probabilistic ML, and LLM-integrated statistical learning and reliable agentic systems. The fellowship page is an academic-network route, not evidence of employment, supervision, collaboration, model transfer, investment use, or deployment. See the capture note.

An August 26 Bridgewater recheck adds a current roster snapshot and two separately indexed research artifacts. The firm’s Partnership page currently lists Blake Cecil as Deputy Chief Investment Officer, Alpha Engine and AIA Labs; Nina Lozinski as Head of AIA; Theo Saarinen as Chief Research Officer, AIA Labs; Oliver Simon as Head of AI & ML Investment Strategy; Aaron Linsky as Head of Engineering, Alpha Engine; and Oliver Radwan as Head of Technology. The general People archive still contains a December 2023 description of Jasjeet Sekhon as Chief Scientist of AIA Labs and Head of Machine Learning, while his separate profile places him at Google DeepMind after a former Bridgewater role. These pages are temporally inconsistent, so the article treats the archive entry as stale/conflicting rather than resolving it by inference. A July 20, 2026 AIA paper names Rohan Alur, Daniel Kang, and Yuxuan Zhu and reports a PAC-Bayes/compression analysis using Qwen3.5-4B across math, programming, general knowledge, and Text-to-SQL; it publishes prompt, reward, sampling, and quantization details and links to an arXiv paper and code. Bridgewater’s August 3, 2026 noisy-data paper separately reports a re-verification pipeline, Qwen2.5-Math-7B experiments, and an external uiuc-kang-lab repository. These are firm-published model-training and data-quality studies, not investment benchmarks or proof of a production trading model. The capture note preserves the title conflict and evidence boundaries.

The August 27 six-part X thread by Daniel Kang adds a distinct AIA Labs research route in collaboration with UIUC and Thinking Machines. The linked technical report describes ReViSQL-K2.6: Kimi-K2.6 fine-tuned through Tinker with RLVR on BIRD-Platinum, an expert-reviewed text-to-SQL training set. The disclosed method combines data re-verification with reward checks for semantic SQL equivalence and for using supplied external knowledge. The report publishes benchmark-specific figures, including 91.37% under its single-sample configuration and 92.97% with 16-sample self-consistency on Arcwise-Plat-SQL, alongside reported per-task costs. Those are results on corrected academic text-to-SQL tasks, not evidence of a Bridgewater production investment model, financial-data training, portfolio authority, or return attribution. The capture note preserves the six thread URLs, model/data recipe, and ownership and deployment boundaries.

Bridgewater’s June 10, 2026 official video adds a separate recruiting and personnel route. Its written description identifies Will Barnes as Head of Policy Research and Nick Hamilton as a full-time Investment Associate on the macro and policy team after winning the firm’s forecasting competition. The recording presents forecasting, trade policy, tariffs, and macro/policy work as connected parts of that route (00:14–00:47, 01:15–01:57). This supports current firm-controlled personnel and recruiting evidence; it does not disclose the competition’s scoring data, hiring rubric, AI/ML system, model ownership, training data, production deployment, portfolio authority, or performance attribution. The automatic captions render Barnes’s surname inconsistently at the introduction, so the written description is treated as authoritative. See the capture note.

A separate October 2020 top1000funds recording supplies a title-blind historical founder route: its title and written description identify Ray Dalio as Bridgewater’s founder, co-chair, and co-chief investment officer at publication. Public automatic captions are now available for timestamp navigation, but review of the recovered track adds no substantive AI or AI-system disclosure. The recording remains dated founder and investment-philosophy context, not AI evidence. The same query pass disqualified five apparent Linsky/Dalio-related hits as unrelated people, generic aggregator material, or an identity-unresolved upload; those records are excluded from Bridgewater evidence counts. See the audit note.

The latest official-surface recheck adds detail without converting generic quantitative infrastructure into an AI claim. Renaissance Technologies’ About page reports approximately 300 employees, 90 PhDs, a research database growing by more than 40 terabytes per day, more than 50,000 computer cores, and more than 200 gigabits per second of connectivity. Its current Research Scientist listing describes data science, statistics, and applied mathematics for trading algorithms and financial-market models; the Research Engineer listing adds research/data-processing software, technical prediction and trading models, CPU/GPU knowledge, and distributed-computing skills; and the Research Infrastructure Programmer listing describes C++ infrastructure used by roughly 150 programmers and scientists for systematic trading and complex statistical models. These are first-party operating and hiring signals, not evidence of GenAI, a named AI owner, a particular model, or AI-attributed performance. Squarepoint’s Experienced Professionals and Early Career Opportunities pages separately connect machine learning, quantitative models, automated strategies, research platforms, and academic recruiting pathways. They provide firm-authored positioning and talent evidence, not a model inventory or deployment audit. The capture note records the source boundaries.

High-Flyer’s public Chinese-language technical surface can now be read at higher resolution. Its AI research page describes an AI research lab, an AI-compute sponsorship program using idle FireFlyer II capacity, and full-time scientist recruitment. The linked HFAiLab GitHub organization publishes hai-platform, OpenCastKit, ffrecord, and hfai-models, whose public examples span biology, vision, NLP, time series, weather, multimodality, and graph learning. The hfai model gallery and first-party technical posts document model reproduction/optimization, task scheduling, haiscale, haiprof, hfreduce, and the 3FS storage path using Direct I/O, io_uring, RDMA reads, and ffrecord. This adds a public AI-lab and engineering layer to the existing fund-level AI record; it does not establish that the public models are trading models, disclose proprietary weights or data, or prove live portfolio permissions or AI-attributed performance. The capture note preserves the Chinese-source and entity boundaries.

Ubiquant’s current English and Chinese pages describe a technology-led quantitative organization, report approximately RMB 80 billion in AUM as of Q4 2025, and place its AI/big-data strategy in 2019 and internal plus external joint laboratories in 2021. The Chinese 800亿人民币 wording is the same RMB 80 billion figure in Chinese financial notation. Its UbiquantAI GitHub organization exposes five public repositories, including iDO, a local AI desktop assistant with activity capture, searchable knowledge, task generation, and grounded chat. The same public surface includes One-shot Entropy Minimization, URM, Fleming-R1, and Fleming-VL, covering post-training, recurrent reasoning, medical reasoning, and multimodal medical inputs. A public GitHub profile says “Researcher @ Ubiquant” and links to two of those repositories. These artifacts establish public research and productivity-tooling routes; they do not establish financial-data training, investment deployment, portfolio permissions, or AI-attributed performance. The capture note preserves authorship, translation, and entity boundaries.

Ubiquant’s current careers page describes a campus-recruiting funnel and says 90% of core investment-research and technology team members graduated from named Chinese and international universities. Its official LinkedIn feed adds current public routes around an AI-focused campus tour, AI algorithm researcher and data-science hiring, an ICML AI networking event, and sponsorship of the ICML 2026 Deep Learning for Code Workshop in Seoul. Separately, the public IQuest-Coder repository and March 17, 2026 technical report describe 7B, 14B, and 40B code models, repository-evolution training, reasoning-driven RL, 128K context, loop variants, tool use, and CLI-agent examples for Claude Code and OpenCode. A secondary July 2026 job listing attributes an IQuest long-horizon-agent role to Ubiquant and mentions planning, multi-step tool calls, agent harnesses, and long-chain evaluation; that attribution remains secondary because the IQuest repository is a separate public organization. These routes expand the public recruiting, conference, and code-agent evidence without establishing filled personnel, IQuest legal ownership, financial-data training, investment deployment, portfolio permissions, or AI-attributed performance. The capture note preserves the affiliation boundary. The separate IQuestLab Hugging Face organization currently lists 14 team members, 23 models, 4 datasets, and recent UBio-MolFM-V1.5, HOTE-8B, and UniReason-Med routes. This is a dynamic IQuestLab catalog, not evidence of Ubiquant ownership, financial deployment, or production use; the capture note records the date-scoped boundary.

An independent title-blind media pass adds two allocator and manager routes that are useful comparison points without changing the article’s no-ranking posture. The Excess Returns interview with Elena Khoziaeva, identified by the publisher transcript route, describes Bridgeway’s factor research, replication, regime checks, statistical defenses against data mining, and team criticism; its 55:02 segment is explicitly titled “How Bridgeway is using AI.” The transcript presents AI as an aid to data, text analysis, and research while retaining human judgment. It does not identify Bridgeway’s model family, training corpus, data license, agent permissions, production status, or performance.

The J.P. Morgan Making Sense episode on Blackstone’s hedge-fund investing platform identifies Joe Dowling, global head of Blackstone’s Multi-Asset Investing business, and Riad Abrahams, head of Strategy, Risk and Quant Analytics in BXMA. The publisher page and timestamped transcript route describe manager selection, managed accounts, seeding, data, and AI tools for research, portfolio-company monitoring, and due diligence. The speakers’ description of Blackstone’s proprietary data and infrastructure is firm-reported and not an independently measured inventory. This is allocator/platform commentary, not proof of a Blackstone manager’s model architecture, data contract, permission map, live authority, or returns. The source note records both routes and the evidence hierarchy.

August 27 title-blind Risk.net routes: execution automation and causal research

An archive search by workflow terms rather than by hedge-fund, AI, or job-title keywords recovered a dated AllianceBernstein FX execution account. Risk.net’s public opening says AB’s multi-asset trading team deployed an FX algorithm wheel within the preceding year and describes the article as a hands-free foreign-exchange execution case. The body is access-limited after the opening, so the public record supports only the dated deployment statement. AB’s first-party equities page separately describes STAR systematic order routing, and its September 2025 policy describes ALFA automated liquidity filtering and analytics for bond-market information aggregated from more than 70 sources. Those are adjacent automation records, not proof that the tools share an architecture or that the FX wheel uses ML or GenAI. No public source reviewed here identifies the wheel’s model, vendor, routing rules, human-exception policy, permissions, evaluation, or execution-quality result. See the capture note.

A separate title-blind Risk.net account identifies Alik Sokolov as managing director of machine learning at RiskLab, University of Toronto, and co-founder and CEO of Sibli. The readable article describes LLM-assisted construction of causal graphs: one cited project used expert views gathered from a large daily news stream, while a later project used GPT-4 to organize 153 factors into clusters and causal charts. Risk.net reports that the GPT-4 groupings matched correlation-based versions on monthly-return prediction, were less cross-correlated and easier to interpret, and aligned with statistical causality tests for two-thirds of the proposed relationships. It also describes prompt chains, human review, and a possible strategy-search loop. These are reported research findings and speaker projections, not an independent replication or evidence of a named hedge fund’s live system. The route expands the modality map beyond sentiment scoring: an LLM can be used to organize hypotheses and causal structure before numerical modelling. It does not disclose Sibli customers, model weights, prompts, data rights, benchmark splits, leakage controls, portfolio permissions, or investment authority. The capture note keeps the method and evidence boundaries separate.

The same title-blind archive yielded a recoverable Risk.net Quantcast episode with Blanka Horvath and Gordon Lee. The 36:37 public recording is now retained with timestamped sidecars. Horvath describes variational-autoencoder market generators, synthetic financial paths, anonymisation, outlier detection, and overfitting controls; Lee describes the limited contemporary use of GANs and restricted Boltzmann machines in pricing and capital-generation practice. The speakers discuss an experiment using roughly 250 samples, signature representations that accommodate irregularly observed paths, and generating additional paths for data-hungry applications such as reinforcement-learning deep hedging (08:40–14:30, 19:12–25:10). This is a dated quantitative-finance research route, not evidence of UBS production adoption or hedge-fund deployment. It does not disclose model weights, data rights, validation results, portfolio authority, or performance. The capture note records the audio hash, ASR limits, and entity boundaries.

The same archive also recovers a deeper 2025 Risk.net Quantcast episode with Dario Villani and Kharen Musaelian. Risk.net dates it June 24, 2025 and describes a QCML method developed with BlackRock collaborators for finding similar high-yield bonds when liquidity constrains a purchase; the Qognitive-hosted paper provides the formal similarity-learning route. The recording places the bond use case at 18:21–22:23, describes feature-rich sparse/outlier settings, distinguishes quantum mathematics from running the method on classical computers (09:14–14:13; 25:13–27:04), and separately discusses equity, healthcare, insurance-claims, and robotics applications (34:20–50:02). Qognitive’s current team page lists Villani as co-founder and CEO and Musaelian as co-founder, president, and CSO. The episode also contains a speaker-reported IBM quantum-hardware collaboration (01:01:16–01:04:56). This is public company, paper, partner, and dated practitioner evidence; it does not establish current Duality ownership, investment-book deployment, model weights, data rights, agent permissions, or independently audited financial performance. See the capture note.

The same pass screened BlackRock, BGI and the big quant pivot. Its public preview establishes a dated account of BlackRock’s systematic-equity history, but does not expose enough AI or current operating detail to meet the promotion threshold. It remains a tracked lead rather than an article finding.

The expanded title-blind search also recovered two previously absent records in Fundamental Edge’s Invest with AI archive. The Implied episode with Ying Hua identifies Hua as founder and CEO of Implied and describes prior portfolio-management roles at Citadel and Balyasny. Its publisher YouTube recording now supplies a public caption layer: Hua describes in-house live earnings-call transcription with a stated sub-one-second delay and preservation of raw audio timing and disfluencies (11:53–14:14), a public-data scraping example (15:42–17:10), human-rubric flagging (32:32–35:00), and scheduled, persistent process-learning workflows (45:32–49:46). The Canary episode with Joe O’Donnell identifies him as Canary’s founder and CEO and describes prior Tiger Global portfolio management. Its public YouTube mirror adds automatic-caption timestamps for the layered data/intelligence/tool architecture and an AI forensic-accounting workflow covering roughly 10,000 companies (06:30–09:34), fine-tuned task models (11:32–12:20), a headless API/MCP route (13:35–14:10), idea-generation agents producing long investment reports (17:26–18:48), analyst-labeled examples for task-specific tuning (31:39–34:41), and hybrid customer use through API/MCP (47:42–50:16). These remain O’Donnell/Canary statements; the direct Buzzsprout MP3 returned 403, so the timestamps are caption navigation rather than audio-verified quotations. They do not establish customer identities, data provenance, model weights, fine-tuning corpus, agent permissions, portfolio authority, or performance. Prior-employer history is not treated as evidence that the current products were used at those firms. See the capture note.

The same archive’s inaugural episode, dated June 8, 2026, supplies an additional title-blind operating-model route. Brett Caughran and Khe Hy’s public YouTube recording adds an automatic-caption layer: Hy describes prior hedge-fund-manager research work at BlackRock, Caughran describes prior fundamental long/short investing, and the speakers frame the shift from chatbots to agents (00:39–04:31). This is useful for query expansion across firm, vendor, and conference sources whose titles omit hedge-fund or AI terms, but it does not establish a named manager’s deployment, customer identity, model inventory, data provenance, permissions, investment authority, or performance. The caption track is navigation evidence, not a verbatim transcript. The capture note preserves the boundary.

August 27 AI-thesis manager and SEC structure route

The title-blind queue also led to a first-party and regulator-record route for Situational Awareness. The author-hosted June 2024 essay frames advanced AI through deep-learning trendlines, compute, power, and industrial infrastructure. It is thesis provenance, not a disclosure of the fund’s model or portfolio process. A June 2026 SEC 13F identifies Situational Awareness LP as institutional investment manager for the quarter ended June 30, 2026. A June 29, 2026 Schedule 13G names the adviser, general-partner, LLC, and fund entities and identifies Leopold Aschenbrenner as managing partner/control person and Carl Shulman as co-portfolio manager. These filings establish legal and dated personnel relationships plus reportable-security records; they do not establish total AUM, leverage, private assets, model architecture, training data, AI deployment, portfolio permissions, or performance. The capture note keeps this primary evidence separate from the secondary podcast commentary.

The title-blind media expansion also recovered Innovantage Podcast #49 with Bas Kooijman, whom the publisher identifies as CEO and co-founder of DHF Capital. A public Podbean enclosure was downloaded and processed locally with WhisperX-MLX, producing 922 timestamped automatic segments. DHF’s first-party Nova Fund announcement separately describes a Luxembourg AI-driven multi-asset fund and proprietary AI systems spanning news, market microstructure, strategy optimization, risk, and capital allocation. Kooijman reports a “100% AI” / “zero human touch” fund, 187 agents, risk and news roles, an optimizer reviewed after 100 trades, a 2025 dry run, and a 32% efficiency figure (47:44–52:53). Those are speaker- or firm-reported claims. The recording also describes alerts, reports, risk blocks, manual oversight, and approval/control loops, creating an unresolved tension with “zero human touch.” The public record does not establish agent count, model weights, training data, data rights, audited performance, or portfolio authority. See the capture note.

August 27 Australia/systematic-manager route: Vinva

The title-blind pass recovered an Inside the Rope interview with Morry Waked, identified as Vinva Investment Management’s Managing Director and Head of Investments. The public SoundCloud route and Podscan transcript place his account in a global-equity systematic process: data and technology scale analysis across thousands of companies, while people contribute investment ideas, model architecture, and judgment (22:45–25:08; 36:22–37:15). The interview reports approximately $43bn and roughly half Australian and half international equities at recording time (26:35–27:02). Vinva’s current first-party home page reports a different, later figure—$65bn+ AUM as at August 4, 2026, alongside 45+ countries and 15,000+ stocks—so those figures are retained as date-scoped company or speaker claims rather than reconciled by inference.

The interview gives unusually specific public AI boundaries. Waked says Vinva has its own GPUs, runs AI models internally rather than placing them on the cloud at that point, and uses proprietary models over broad company and market relationships (30:16–33:32). Asked about conviction in earnings-call voices, he says the firm had used conference-call information as far back as 2007–2008, while noting that transcripts, sentiment tools, and management coaching had made the signal more competitive (33:32–34:27). This is dated practitioner evidence about call information and signal crowding, not evidence of a current acoustic model, deception classifier, label set, lead time, or return attribution.

Vinva’s first-party team page supplies the personnel layer. It names Juliane Krug as Research Manager for equity security-selection models, with a PhD in Finance from UNSW; Danny Lo as Research Director for security selection, risk, transaction-cost modelling, and portfolio construction; Therdsak Tangkuampien (“TK”) as Research Director working on optimization and with a PhD in statistical machine learning from Monash; Robert Franklin as Research Manager with machine-learning, computer-vision, and NLP experience; and Reshma Joseph as a 2026 Investment Data Analyst working on data infrastructure and quantitative systems. It also identifies Ethan Hansen in portfolio and trading systems after IMC execution-technology work, and Jenny Chen, Duncan Forrest, and Gangeyan Ganesalingam in investment-technology and data-feed responsibilities. These biographies establish public roles and training histories; they do not map any person to a specific model, voice pipeline, permission boundary, or production decision.

The distributor’s April 2026 quarterly update separately says AI and ML support research productivity, data sourcing, and investment-signal development. Its page describes adaptive signals, global linkages, short-horizon data, diversification, and risk controls, but does not name a model or dataset. The capture note records the podcast transcript, current AUM discrepancy, personnel routes, and disclosure limits.

An Australian allocator/research-provider route adds a useful cross-manager observation. Zenith Investment Partners’ Investment Researcher episode, hosted by Growth Team analysts Ethan Spiegel and Bradley Antman, is now backed by the official Podbean feed, recovered audio, and a local timestamped ASR transcript. The speakers say their manager conversations and prior global-equities survey show six recurring use-case buckets—operational efficiency, information sourcing, thesis challenge, industry research, data processing, and quant coding (04:09–05:07). They describe quant-side examples involving AI-assisted cleansing of anonymized card and foot-traffic data and faster coding (12:48–14:17), while describing IP leakage and cost as constraints and internally housed models as a protection measure (15:17–16:05). In Zenith’s account, agentic workflows were present but still a minority in manager meetings, and stock-specific internal modelling was not the common use case (20:52–22:16). Zenith also describes internally housed tools for peer analysis, performance attribution, regime/outlier identification, and follow-up diligence (22:49–24:06). These are attributed Australian allocator observations, not evidence about any named manager’s model, data rights, deployment, permissions, investment authority, or performance; the 86% survey figure is retained as Zenith-reported rather than independently audited. See the timestamped capture note.

August 27 institutional-capital-markets platform route: Goldman Sachs Marquee AI

The same title-blind pass recovered Goldman Sachs Exchanges’ “Building AI Systems for Capital Markets”, dated August 24, 2026. Goldman identifies Chris Churchman as head of Marquee, its digital platform for institutional and corporate clients, and co-chair of the Global Banking & Markets AI Working Group. Churchman says Marquee AI was internally available at the time of recording and describes a workflow that decomposes a user question into research topics, retrieves firm research, trading-floor commentary, Market View analytics, and pre-trade tools, then can write and run Python for calculations (02:00–05:14).

The episode’s public implementation detail centers on provenance and permissions. Churchman describes sentence-level grounding to a firm source or auditable calculation, and discusses the difficulty of separating facts from interpolations (04:15–05:14; 05:28–09:40). He then describes agentic, environment, and “mandate” engineering: secure environments, entitled access to institutional tools and knowledge, and explicit answers to what an agent may do under whose authority (09:50–11:22). He also distinguishes automating an existing process from reconceiving the job to be done, using an agent that understands a firm-specific visual-structuring tool as an example (18:15–20:31).

This is an asset-manager and capital-markets platform comparator, not evidence about a private hedge-fund system. The source does not identify the model provider or weights, training corpus, benchmark fixture, error denominator, routing policy, client rollout, production permissions, trade authority, or AI-attributed performance. The capture note preserves the first-party transcript and those boundaries.

August 27 title-blind Odds on Open expansion: Q Capital and Variational

The August 21 Odds on Open interview with Lucas Schuermann is a newer title-blind route into quantitative-fund formation and market infrastructure. Publisher pages identify Schuermann as founder and CEO of Variational and describe his earlier Q Capital and Genesis Trading roles. The public transcript describes historical fully algorithmic, market-neutral FX and crypto strategies, statistical-arbitrage-style relationships, inter-exchange arbitrage, and basis (00:00–02:18); it then describes a research-group path spanning differential geometry, computational neuroscience, machine learning, robotics, and Bayesian machine learning (05:32–06:27). The account identifies Schuermann as Genesis’s VP of Engineering and describes building a core electronic market-making system (08:10 onward). The episode’s publisher description adds Variational’s on-chain perpetual-futures and aggregated-liquidity context, growth-curve analysis, and AI’s effect on technical skills. These are speaker and publisher claims: the reviewed material does not disclose a current LLM, training corpus, agent, model weights, data rights, production permissions, or AI-attributed performance. It is included because the title omits hedge-fund and AI terms while exposing relevant quant, data, infrastructure, and academic-lineage routes. See the capture note.

August 27 title-blind System2 practitioner route

The Odds on Open discussion with System2 CEO and co-founder Matei Zatreanu adds a distinct practitioner and vendor-side route. The public transcript discusses the data-science layer in fundamental hedge funds, including talent constraints, the risk that a polished model output is unsupported, organization-specific context, expert-network automation, qualitative-data synthesis, causal mapping, and the difference between bespoke investment judgment and reusable workflows (00:00–08:28; 10:18–20:33; 31:43–44:03). A recovered public audio enclosure and local sample transcription independently align with the graph-and-entity discussion at approximately 20:18–21:13; the sample is automatic and not speaker-adjudicated. Zatreanu’s System2 profile supplies the first-party founder role and career/education context, while the publisher’s YouTube page supplies an alternate media route. See the capture note.

The August 27, 2026 Odds on Open interview with Noah Kann adds a title-blind emerging-manager and personnel route. Venari’s official site identifies Kann and Ethan Kann as co-founders and co-CEOs and describes research, portfolio-decision, construction, risk-tracking, and repeatable-framework responsibilities. The publisher description and transcript describe a discretionary macro process across equities, ETFs, options, and occasionally commodities/futures, with attention to defense, energy, data-center infrastructure, drawdown limits, and hiring for continuous improvement (11:53–17:00; 46:43–54:08). AI appears as an investment theme and in discussion of technology/infrastructure, not as a disclosed Venari model or workflow. The route therefore supports personnel, process, and AI-infrastructure-theme tracking only; it does not establish Venari GenAI use, a model/provider, data rights, evaluation, production deployment, trading permission, or performance. See the capture note.

The most useful control point is methodological rather than a claimed system feature: the episode describes a novel result as the point at which validation burden should increase, because an apparently differentiated output may reflect missing data, unsupported inference, or thesis-conforming context. That is a speaker’s operating view, not an independently tested investment result. The episode does not identify a named customer, model weights, provider, training corpus, data license, evaluation fixture, permissions, production endpoint, trade authority, or AI-attributed performance. Podscan dates the episode March 27, 2026, while YouTube metadata dates it March 26; both dates are preserved in the capture note, along with the related-episode follow-up queue.

August 27 AI-native manager formation route: Aethon Fund

The title-blind expansion also closed a stale Aethon pre-launch lead. A July 15, 2026 company release says Aethon was launching with $50 million, including a fund-of-funds separately managed account anchor and other commitments. The release describes proprietary market signals, machine-enforced rules, AI handling execution, and a strategy set spanning long, short, momentum, mean reversion, and stealth accumulation. It lists George Kailas as founder/leader, Dave Lauer as CTO, Ezi Ozoani as Head of AI, Joe Bernstein as Head of Trading, Erik Smolinski as Head of Risk Management, Dave West as Head of Infrastructure, and Matt Carter as Head of Communications. These are company-release claims, not independent employment or fund-document verification.

The same-day TBPN interview transcript records Kailas describing the transfer of Prospero.ai signal libraries into Aethon, LLM-generated strategy candidates, cross-market strategy combination, and opportunistic allocation among strategy groups. The interview says an August/September launch was expected, while the release is framed as a launch announcement; that timing discrepancy is retained rather than resolved by inference. A later public founder announcement describes a $50 million debut, a Prospero feedback loop involving more than 200,000 retail investors, Kailas moving to Chairman of Prospero, and Adam Plante becoming Prospero CEO. These sources expose a public formation and operating-model narrative, but not model weights, training data, signal definitions, provider contracts, evaluation splits, execution logs, customer identities, permissions, or AI-attributed performance. The capture note preserves the source and entity boundaries.

August 27 Rokos Capital Management: Head of AI and public GenAI routes

Rokos Capital Management’s company post identifying Luke de Oliveira as Head of AI says he led a firmwide knowledge-sharing session on AGI, how such systems operate, open versus closed models, industry effects, and work advancing the firm’s use of AGI. A later RCM company post describes AI as helping power the firm’s next chapter and expanding its opportunity set beyond macro. These are first-party public strategy and responsibility statements; they do not identify a model, provider, dataset, evaluation, or production endpoint.

RCM’s Investor Day post describes presentations and panels across Macro, Equities, Credit, Emerging Markets, and Systematic teams, including a panel involving Quant, Tech/AI, and Portfolio Strategy leadership. It names Jason Sippel as Deputy CIO and Global Head of Markets but does not name the Tech/AI panel participants or provide the panel materials. This is an organizational and follow-up signal, not a technical disclosure.

The public Luke de Oliveira profile and personal site identify him as RCM’s Head of Artificial Intelligence and describe prior work at Meta on Llama, machine-learning leadership at Twilio, the founding of Vai Technologies, and Yale/Stanford and laboratory research affiliations. The CaloGAN paper records his co-authorship on generative-model research for particle-physics simulation; linked papers cover transformer summarization and fine-tuned language representations. These sources establish personnel background and academic lineage, not an RCM investment implementation or research objective.

The GenAI Zürich schedule archive lists de Oliveira’s “Reasoning in the Open” session, while his public announcement supplies the personnel link. The archive presents a date/template inconsistency, so the event date is not normalized by inference. The capture note preserves the source hierarchy and unresolved details.

The reviewed material does not establish RCM’s model or provider inventory, weights, training corpus, data rights, evaluation records, agent permissions, production endpoint, order or portfolio authority, or AI-attributed performance. The company posts support public positioning and a named AI responsibility; the personal profile and papers support background only. The Tech/AI panel’s other participants remain unidentified in the reviewed public material.

August 27 AI-native manager surface: Rykos Capital

The title-blind search also surfaced Rykos Capital’s first-party site and its platform page. The pages describe an agent-based investment workflow that ingests SEC filings, earnings-call transcripts, market and alternative data, and FRED indicators; performs document analysis, anomaly and management-tone analysis, historical simulation, Monte Carlo analysis, position sizing, and compliance checks; and produces trade ideas and draft memos for portfolio-manager review before execution. The platform page organizes the claims into a research dashboard, filing-intelligence layer, and risk/compliance console. This is a useful public description of a proposed or claimed AI-native operating model, but it is firm-controlled marketing evidence rather than an implementation audit.

The team page names Anish Rudra as Co-Founder & CIO and Raj Karnati as Co-Founder & CPO. Rudra’s self-authored experience page describes other current and dated roles and does not mention Rykos, so the current-role record remains unresolved. The Rykos site itself also uses “fully autonomous” language while describing PM review before execution; the contradiction is preserved rather than converted into an autonomous-trading claim. A targeted search did not locate a corroborating regulatory record, public code, model card, paper, benchmark artifact, or independent track record. The capture note records the claims, identity conflict, and follow-up requirements. Nothing here establishes model weights, training data, data rights, named vendors, production deployment, portfolio authority, or performance.

August 27 title-blind thematic-manager route: Kurv and the AI-memory thesis

The August 23 Behind the Ticker interview with Howard Chan adds a distinct asset-manager route. The publisher identifies Chan as founder and CEO of Kurv Investment Management and frames the Kurv Memory Select ETF (KMEM) around a manager-reported thesis that memory is a binding constraint in AI infrastructure. Kurv’s July 1 launch release and current KMEM page provide first-party product context; the prospectus is retained as the formal product route. The interview describes concentrated exposure and monitoring of hyperscaler spending, order backlogs, and new entrants as thesis-break indicators. This is external investment-thesis evidence about AI infrastructure, not evidence of Kurv building or deploying AI internally: no LLM, training corpus, agent, data-rights map, portfolio permission system, or AI-attributed performance is established. See the capture note.

August 27 title-blind Man Group portfolio-manager route: Jack Barrat

The August 12, 2026 Algy’s Investment Podcast episode with Jack Barrat adds a current title-blind Man Group personnel and workflow route. Man Group’s biography identifies Barrat as a Portfolio Manager focused on UK equities and records his Matterley and UBS history and Cambridge training. In the public transcript, Barrat describes Man’s ability to invest in team growth and new technologies, a five-person team seeking a sixth member, and an approximate 18-month ramp for a new hire to become a net contributor (01:44–03:46). The episode’s AI segment (12:09–17:43) frames the combination of data, investment work, and developing technology as a potentially large productivity improvement and presents AI as a change to fund-management work rather than a disclosed autonomous trading system. These are attributable statements from one Man GLG investment team; they do not establish a firmwide staffing model, LLM or vendor, training corpus, agent, permissions, production endpoint, or AI-attributed performance. The capture note preserves the transcript and boundary.

September 10–11 IEEE CIFEr 2026: newly tracked agentic-finance conference route

The official IEEE Computational Intelligence in Financial Engineering and Economics (CIFEr) 2026 program adds a newly tracked conference route for September 10–11 in Tokyo. Its keynotes cover analysts and agentic AI in investment management and generative AI in financial markets. The detailed program lists sessions on point-in-time backtest auditing with LLMs, institutional portfolio construction, automated strategy improvement, policy-driven agentic frameworks, LLM-trading-agent herding stress tests, reinforcement-learning market makers, market-impact modelling, multimodal text/time-series forecasting, disclosure analysis, synthetic financial data, and quality-diversity market-making. This is academic and practitioner-adjacent research-program evidence, not evidence of a covered fund’s participation or deployment. The capture note records the source and follow-up boundary.

August 27 title-blind Citadel personnel route: Jerome Busca

The Blushing Quants interview with Jerome Busca is a useful historical personnel and infrastructure route that a search limited to current firm titles would miss. The publisher describes Busca as a quantitative trader with more than 25 years of experience and identifies prior work in Citadel’s hedge-fund business, including a mortgage desk and a systematic futures operation. The public episode enclosure was recovered and locally checksum- and duration-checked on August 28, 2026; the transcript remains an automatic Podscan artifact rather than a verbatim transcript. In that timestamped transcript, he discusses the evolution from pre-Python research tooling to modern data infrastructure and AI (11:16–13:00), cautions that easier AI-assisted backtesting can increase overfitting and false confidence (52:43–55:41), and describes individual AI tooling alongside the continuing need for professional database and alternative-data infrastructure (57:43–59:45). He also describes AI as reducing the time to obtain data and test ideas, while pointing to unusual and alternative datasets (01:03:16–01:04:55).

This is historical former-personnel testimony and a methodological discussion. It does not establish Busca’s current employer, a current Citadel reporting line, a Citadel model or dataset, production permissions, trading authority, or AI-attributed performance. The capture note preserves the recovered enclosure metadata, transcript source, and automatic-transcription boundary.

August 27 researcher and allocator/IR routes: López de Prado and AIMA

The first-party quantresearch.org video archive adds a researcher-led route that does not depend on a fund name in the title. It lists eight 2026 videos under Marcos López de Prado’s name or co-authorship, covering causal discovery in finance, the ADIA Lab Causal Discovery Challenge, machine learning for portfolio construction, causal factor investing, Sharpe-ratio inference, false-discovery control, and the strategy-assembly-line framing. The archive names Vincent Zoonekynd in the co-authored “Causality in Factor Investing” entry and describes him as being of ADIA. These are public research and video-index records; they do not establish a current fund employer’s production model, proprietary data, investment permissions, or return attribution. The capture note records the eight video IDs and the affiliation boundary.

The AIMA “Agentic AI for IR and Allocators” event page adds a named allocator-facing route. Its June 12, 2026 speaker list identifies Gardner LaMotte as Head of Client Solutions at Anchorage Capital, alongside CENTRL’s Eric Hoerdemann and investment-management adviser James Barber; Michelle Noyes, AIMA’s Head of Americas, moderated. The agenda covers asset-owner due diligence, manager investor-relations workflows, agent and information permissioning, privacy, and buy-versus-build decisions. A replay link is public but returned HTTP 401 during the check, so no claims are made about the discussion itself. This is evidence of a public role and event scope, not evidence of Anchorage’s model inventory, deployment, data rights, or investment authority. The capture note preserves the access boundary.

Robeco’s current Agentic AI in asset management series adds a first-party research-lead and governance route. The page identifies Mike Chen as Head of Next Gen Research and organizes an 18-minute conversation around alpha, research, workflows, the AI economy, and governance, with follow-on theme pages scheduled from August 25 through September 22, 2026. Robeco’s agentic-alpha page describes screening ideas, monitoring signals, testing hypotheses, and portfolio review as possible workflow support, while its process article names signal admission, capital commitment, and changes to workflow boundaries as decisions that should remain non-delegable. This is firm-authored operating-model language; it does not disclose a model family, training corpus, production permissions, or AI-attributed investment result. See the capture note.

August 27 Canadian academic and financial-innovation route: Rotman’s John Hull Forum

The University of Toronto Rotman’s John Hull Financial Innovation Forum was held in person on May 4, 2026. Its program framed AI in capital markets through delegation of decision-making, autonomy, risk, governance, and institutional accountability, and announced the John Hull Financial Innovation Fund for research and teaching in derivatives, risk management, and emerging AI/ML applications in finance. The keynote, “Who Makes Investment Decisions?: AI Agents, Autonomy and Accountability,” was presented by Andrew Ang, President and Co-founder of Tau Balance and a former BlackRock Head of Factors. The panel named leaders from CIBC Capital Markets, National Bank of Canada, BMO Global Asset Management, TD Bank Group, Montréal Exchange, and RBC; Rotman identifies Ing-Haw Cheng as an AI/ML-to-finance researcher and the forum’s opening speaker. Rotman’s June recap confirms that the forum took place. This is a dated Canadian academic and financial-network route, not evidence of a shared system or a manager-specific implementation: no replay, transcript, slides, model family, training data, data rights, agent permissions, investment authority, or performance attribution was located. See the source note.

The publisher-media pass also recovered Risk.net’s coverage of Andrew Ang’s AI-agent paper, with a WatersTechnology mirror. Risk.net dated its article April 24, 2026 and described Ang’s public strategic-asset-allocation system as using roughly 50 agents across macro analysis, risk assessment, and return-correlation estimation; WatersTechnology published the mirrored account April 27. This is secondary publisher reporting that makes an already-tracked paper discoverable through the finance-media route. The publisher pages do not add model versions, prompts, evaluation fixtures, capital permissions, production telemetry, or performance evidence. The paper remains the primary architecture source, and the capture note keeps the article, mirror, paper, and Canadian forum as separate evidence layers.

Two additional conference routes broaden the research and personnel map. Singapore Management University’s UMD/SMU/UBS Quant Investment Forum, held June 15–16, 2026, provides a post-event program with papers and slides on LLM-guided hypothesis discovery, machine disagreement and AI error, knowledge-informed deep learning, semantic alignment of fund and firm information, active ML trading, text-managed portfolios, foreign bias in AI financial predictions, and 30 years of daily hedge-fund trades. SMU’s post-event account says more than 150 academics and practitioners attended and names researchers including Hong Zhang, Shihao Yu, and Jianfeng Hu. This is a research-artifact and network route hosted by UBS, not evidence of UBS’s internal systems or of any paper author’s fund deployment. The linked LLM-hypothesis paper describes a human-designed symbolic search environment and iterative proposal–test–revision loop; its reported results remain paper claims pending independent code/data verification. See the capture note.

The official Concordia AI in Finance 2026 page schedules an October 13–15, 2026 Montréal conference, including a public livestreamed keynote and a PhD/postdoc workshop. It names Mihail Velikov of Penn State as keynote speaker and lists a program committee including Russ Goyenko, Michael Weber, Guofu Zhou, and Mamatha Adinarayana Swamy. This is an upcoming academic capture target; the page does not yet provide a complete accepted-paper program or a fund-participant list, so it is not treated as evidence of any manager’s AI strategy. See the conference-route note.

An adjacent code audit recovered the public AAPM repository for the paper Empirical Asset Pricing with Large Language Model Agents. The repository’s code combines news embeddings with learned asset embeddings in a feed-forward return-prediction model; its configuration names BGE-large embeddings and a dated GPT-3.5 model for news analysis, while the README requires external WSJ, CRSP-derived factor, and daily-return data. The code exposes dated train, development, and test partitions and experiment logging, making this a concrete research implementation rather than a generic LLM claim. The paper’s reported performance improvements remain author-reported results: the repository does not establish a fund connection, production deployment, data redistribution rights, agent permissions, or independently audited returns. See the code-and-paper capture note.

Mercer’s 2026 AI in Asset Management Survey provides a dated calibration sample rather than a firm comparison. Mercer reports responses from 131 asset managers: 55% reported AI integrated into at least one investment process, 27% reported pilot/proof-of-concept use, and 18% reported no integration. It also reports 73% using AI for operational efficiency, 68% using it as an investment-process insight partner, and 5% granting autonomous or semi-autonomous authority for investment recommendations or trades. Mercer’s allocator analysis reports 69% citing data constraints as a significant barrier, 28% reporting internally developed AI models, and 57% reporting one to five dedicated AI specialists. The survey was conducted online through Mercer’s extended network, so the figures are not a probability sample and should not be used to infer the state of any named manager. They are useful for locating questions about data, staffing, governance, workflow position, and decision authority. A title-blind Planet MicroCap episode with Sam Namiri now adds a dated Ridgewood workflow route: the publisher identifies him as Partner and Portfolio Manager, and local ASR captures AI layered onto valuation screening to prioritize management-meeting verification (13:29–15:11) plus a speaker-reported Gemini/web-traffic hallucination anecdote (09:18–10:01). The capture note preserves the automatic-ASR and no-deployment boundary.

August 27 India and Japan title-blind quant-media routes

An India-focused title-blind pass expanded the existing IIQF lead into a four-playlist discovery surface. The IIQF podcast hub links series on AI/ML applications in finance, practitioner conversations, BFSI machine-learning skills, and disruptive AI in BFSI. Direct playlist enumeration recovered visible titles on generative-AI applications in finance, AI applications in trading, front-office valuation and pricing, risk management, climate risk, finance ML tool requirements, ML technology stacks, and BFSI use cases. The hub labels the series as 12, 1, 2, and 6 episodes, while the visible playlists returned 14, 7, 2, and 6 items including specials; that catalog discrepancy is retained. A recovery refresh then obtained automatic English captions for the generative-AI episode and AI-in-trading episode. The first discusses synthetic data under data scarcity and LLM aggregation of news/social sentiment as an augmentation layer (01:46–03:08; 06:32–08:05); the second discusses execution analytics and speculative event-driven, price-discovery, and cross-market research ideas (10:23–12:10). These are speaker-reported educational discussions. The automatic captions are not a reviewed transcript, and neither episode establishes a named fund’s model, data rights, production permissions, trading authority, or performance. See the capture note.

The same regional pass recovered a distinct CFA Society India Pune program dated March 14, 2026. Its official page names Shailendra Abhyankar, Founder and CEO of FinSoftAi, for social-media and news narrative analysis, and Chetan Mehra, Head of quantitative trading and machine-learning methodologies at Bajaj Alternates, for quant and AI in investment decision-making. The page attributes to Abhyankar a FinSoftAi description involving more than 80,000 news and investor-community sources, including Reddit, Stocktwits, and Bluesky, with sentiment, buzz, and data-quality signals combined with price momentum in live trading strategies. That is speaker/company evidence hosted by the event organizer, not an independent audit or a named hedge-fund deployment; no recording or transcript was located.

Japan adds a talent-pipeline route through the Matsuo Institute AI Quant Trading course. The Tokyo University-linked institute’s page describes a free ten-session student course covering multifactor models, time-series models, machine-learning alpha search, LLM applications, and AI in finance, with J-Quants Pro data supplied by JPX Research Institute. It names the institute’s financial-team personnel and says high-performing students may apply for a financial-team internship. The course rules frame the competition as educational/research activity and disclaim investment use of submissions, so this is a regional talent and curriculum signal rather than evidence of a live fund system. See the regional route note.

A separate Matsuo–Iwasawa Lab financial-machine-learning course page adds a university-lab route that was not in the earlier Matsuo Institute record. The page describes a nine-session spring 2026 course spanning market prediction, dataset creation, labeling, backtesting, CTA operations, stochastic optimization, quantitative strategy, and post-LLM strategy investment. It identifies Kim Kangsoo as a Matsuo–Iwasawa Lab staff member and lists instructors including Satoshi Sekioka, CEO of AI Nest, alongside researchers, university-affiliated instructors, and a trader. Most sessions may be viewed from archives, although the page notes that some external-guest sessions may not be archived. This is a curriculum and talent-network signal, not evidence of a live fund system, model ownership, production permissions, or performance; the page does not provide recordings or implementation artifacts. See the regional route note.

The Korean-language Korea Investment Week 2026 program adds a current regional conference surface. It names Jongwon Roh as CIO of Infinity Global Asset Management for a session on Korean hedge-fund management through AI and quantitative strategies, Sunghoon Kim as Head of Hedge Fund Division at Kiwoom Asset Management for a low-volatility multi-strategy session, and Jupyo Hong as Head of Hedge Fund Division and Executive Managing Director at GVA Asset Management for a structural-alpha session. It also lists Rob Huisman of Robeco for a session on AI and next-generation research. The event ran May 12–15, 2026; the linked materials and replay pages require participant verification, and direct retrieval did not expose recordings or slides. These are dated program and personnel signals, not evidence of any named manager’s AI implementation, model inventory, or performance. See the regional source note.

The POSTECH-hosted Asian Quantitative Finance Conference 2026 call-for-papers page adds a separate Korean academic conference route. It dates the 10th AQFC to July 16–18, 2026 in Pohang and lists AI and machine learning in finance, fintech, derivatives and risk, asset pricing, financial econometrics, and financial markets among its topics. The page says the call for papers is closed but does not expose a speaker roster, accepted-paper program, recording, or firm affiliation. It is therefore a dated discovery and conference-taxonomy signal, not evidence of any manager’s AI implementation, model inventory, or performance. See the regional source note.

The ADIA Lab homepage and relationship page add a distinct Abu Dhabi research-lab route. ADIA Lab describes itself as an independent laboratory for data and computational sciences and says it is a separate legal entity from the Abu Dhabi Investment Authority, with its own governance, budget, and boards. The ADIA Lab Symposium 2026 page schedules the event for October 26–28, 2026 at the Four Seasons Abu Dhabi, says speakers and further details will be published as confirmed, and links to the Lab’s YouTube archive; the homepage says many events are streamed for registered participants. The Research Series lists August 10, 2026 finance publications on causal factor investing, spectral risk parity, false-discovery control, signal-to-noise inference, portfolio risk, and machine-learning-enhanced Markowitz selection. The team page identifies Horst Simon as Director and lists a research network including Marcos López de Prado, Alex Lipton, Robert Engle, Alex Pentland, Miguel Hernán, Guido Imbens, and Jack Dongarra. This is lab, research-lineage, and conference-capture evidence; it must remain separate from claims about the investment institution’s systems, models, permissions, production deployment, trading authority, or performance. See the ADIA/allocator route note.

The official ADIA Lab YouTube archive adds a recording-level lane that was not visible in the static event page. Local playlist enumeration recovered 30 playlists, including 20 individually titled Day 3 recordings from the 2025 symposium’s Digital Economy program. Selected videos include Sandy Pentland on AI agents, Michael Wolf on the Hedged Random Forest, Alexei Kondratyev on quantum-computer data anonymisation, Alex Lipton on monetary circuits, and the Best Paper Award in Financial Data Science. Recovered automatic captions connect Wolf’s talk to inflation forecasting and a named Hedged Random Forest (02:26–02:29; 14:50–14:53), Kondratyev’s to ML for risk-factor analysis and portfolio optimization and quantum ML for sensitive data (00:25–00:29; 02:00–02:10), and Pentland’s to AI agents across business processes (02:02–02:04). This is an official research-video archive and caption layer; it does not establish an ADIA investment-business model, production deployment, permissions, trading authority, or performance. See the recording capture note.

The complete ADIA Lab 2025 symposium agenda adds the speaker and topic map behind those recordings. Its Day 3 program names Guido Imbens on AI and productivity, Alex “Sandy” Pentland on AI agents and the digital economy, and parallel sessions on digital money and the economy, robust AI techniques, and infrastructure for the digital economy and finance. The agenda lists Alexander Lipton, Shafi Goldwasser, Michael Wolf, Thorsten Neumann, Davor Svetinovic, Juan Corchado, Jonathan Ledgard, Alexei Kondratyev, and Thomas Hardjono across those sessions and says recorded sessions were available online. These are organizer-published event and recording facts, not evidence of common employment, a shared model, ADIA investment-business deployment, permissions, or performance.

What This Means for Your Organization

For an allocator or investment-committee member, the relevant diligence question is not “Does the manager use AI?” Ask where AI sits in the chain from data to feature to forecast to portfolio to order, and which boundary still requires human approval. Acadian’s public material illustrates one investment-workflow disclosure. Arrowstreet’s job listing provides one platform and permission-control disclosure. GMO’s material illustrates one way to examine AI exposure through cash flows, competition, and valuation. Man AHL, Bridgewater, and Jane Street provide additional public evidence at research-workflow, investment-lab, and trading-infrastructure layers.

Request eight concrete artifacts from any manager making an AI claim: a current use-case inventory; named accountable owners; a model and vendor inventory; point-in-time and leakage controls; evaluation results against a non-AI baseline; a human-approval map; usage, cost, and failure telemetry; and a production-stage description separating pilot, shadow mode, assisted workflow, and live execution. If the manager cannot provide those artifacts, treat “AI strategy” as positioning rather than operating evidence.

August 27 Risk.net Quantcast: ADIA/HRP portfolio-construction route

The previously unresolved Risk.net Quantcast episode with Alexandre Antonov adds a recording-level route around portfolio construction and ADIA-linked quantitative research. Risk.net’s February 7, 2025 page identifies Antonov as a quantitative research and development lead at Abu Dhabi Investment Authority and links the discussion to the paper Overcoming Markowitz’s Instability with the Help of the Hierarchical Risk Parity (HRP): Theoretical Evidence. The public SoundCloud track was recovered and transcribed locally on August 27, 2026. The related SSRN record lists Antonov, Alex Lipton, and Marcos López de Prado as authors and describes an analytical comparison of noise in allocation weights.

Antonov frames the buy-side problem differently from sell-side calibration: stochastic calculus and probability transfer across settings, but drift and alpha are less directly observable, so the volatility component should not be made unnecessarily complex when the drift is uncertain (00:34–03:47, local recording). He then describes the Markowitz baseline as an optimization whose covariance-matrix inversion can amplify estimation noise (03:53–09:34). HRP is described as a two-level construction: cluster assets, apply an allocation inside each cluster, then allocate across the resulting cluster portfolios (07:40–09:00). The episode’s stated objective is to calculate and compare weight noise for HRP and the Markowitz baseline (09:15–10:35). These are explanations of the authors’ paper, not an independent replication.

The clustering route is relevant to the project’s ML map because Antonov says clusters can be chosen logically, such as by country, sector, or asset group, or generated statistically using covariance-derived distances and machine-learning or numerical methods (10:50–11:31). He describes possible uses as estimating confidence in optimization weights, selecting cluster counts, and combining outputs from multiple allocators using estimated weight noise (approximately 21:00–26:00). He also describes analytic validity conditions—moderate component counts, an untreated and invertible covariance matrix, and approximately Gaussian returns—and possible extensions to higher-order uncertainty, eigenvalues, eigenvectors, other optimizers, and causal theory (approximately 17:00–30:00). The precise applicability boundaries remain those of the paper and recording; they should not be generalized to arbitrary portfolios.

This route establishes a named ADIA-linked research topic, a finance-specific portfolio-construction framework, and a public research connection among Antonov, Lipton, and López de Prado. It does not establish ADIA implementation, a specific fund vehicle, a production optimizer, proprietary data, model-provider relationships, agent permissions, autonomous allocation, or attributed performance. The timestamped capture note preserves the audio hashes, ASR sidecars, and evidence boundaries.

August 27 Risk.net Quantcast: hedge-fund counterparty stress without an AI label

The title-blind Risk.net episode with Fabrizio Anfuso adds a risk-management route that a search limited to AI, machine learning, or hedge-fund titles would miss. Risk.net’s May 23, 2025 page identifies Anfuso as a senior technical specialist at the Bank of England and describes a top-down counterparty-risk framework using Gaussian copulae and mixtures of Gaussian distributions. The public SoundCloud recording was recovered and transcribed locally on August 27, 2026. The related Risk.net paper, published April 29, 2025, lists Anfuso and Dimitrios Karyampas and describes the same wrong-way-risk and leveraged-counterparty framework.

Anfuso distinguishes a bottom-up model built from detailed portfolio and balance-sheet information from a top-down model that represents extreme-tail dynamics with a smaller parameter set (00:00–05:52, local recording). He describes combining a Gaussian copula with a mixture of Gaussian distributions: the copula conditions exposure scenarios on default, while the mixture switches between ordinary and stressed calibrations to generate heavier tails without replacing an institution’s existing risk-factor engine (05:52–10:03). He presents this as a production-conscious modelling choice constrained by implementation cost, regulation, and pricing reconciliation. These are speaker-reported methodological statements, not an independent production audit.

The episode explicitly discusses hedge-fund counterparties. Anfuso identifies potential future exposure, expected exposure, and the full distribution of counterparty exposures as outputs, then says the distributional framing can support analysis across a portfolio of hedge-fund counterparties (17:31–19:34). For hedge-fund portfolios, he says banks may already have stress calibrations; the harder parameter is the copula level, which may need to reflect net asset value, leverage, liquidity, and unencumbered cash (24:31–27:45). He also says limited disclosure is itself a risk input and discusses possible third-party or blind-sided aggregation models (27:45–31:46). None of this identifies a specific fund’s positions, a vendor relationship, or a bank’s implementation.

This route matters for the AI map because it is a negative control: the source contains a detailed finance-specific computational framework but no LLM, generative-AI system, autonomous agent, or model-training programme. It also shows why title-blind discovery matters: material quantitative infrastructure can appear under counterparty-risk, stress-testing, or Archegos terminology rather than AI terminology. The public publisher LinkedIn post is retained as a distribution and discovery surface, not as independent technical evidence. The timestamped capture note preserves the audio hashes, sidecars, and Bank of England policy disclaimer.

The route establishes a named researcher, a paper-backed counterparty-risk framework, and a direct hedge-fund stress-testing use case. It does not establish any named hedge fund’s model inventory, data rights, production deployment, agent permissions, trading authority, or investment performance. It also does not establish that the Bank of England endorses the views expressed.

August 27 Risk.net Quantcast: autoencoding yield curves across CompatibL, QRM, and Bloomberg

The title-blind Risk.net episode with Alexander Sokol, Andrei Lyashenko, and Fabio Mercurio adds a high-specificity finance-ML route. Risk.net’s March 28, 2025 page identifies Sokol as founder and head of quantitative research at CompatibL, Lyashenko as head of market pricing and risk models at Quantitative Risk Management, and Mercurio as global head of quantitative analytics at Bloomberg. The public SoundCloud recording was recovered and transcribed locally on August 27, 2026. The related Risk.net paper and SSRN record provide the paper and affiliation layers.

The speakers describe autoencoders as a way to learn a small set of latent variables from historical yield-curve data, then use those variables inside a risk-neutral interest-rate framework (approximately 06:45–21:48, local recording). They explain the encoder/decoder and the autoencoder manifold as a low-dimensional representation of the full curve, while stressing that training alone does not guarantee no-arbitrage (approximately 21:48–42:12). The paper’s framework therefore combines neural-network representation learning with explicit mathematical conditions for pricing and hedging; it is not presented as an unconstrained prediction model.

The application discussion covers interest-rate derivative pricing and hedging, portfolio valuation, simulation, XVA, scenario analysis, asset-liability management, and tail-dependent exposure measures (approximately 42:12–50:00). The discussion also offers a valuable deployment boundary. When asked whether the framework was already applied at the participants’ firms, the recording contains “not yet” responses at approximately 50:27–51:15. Later, one participant discusses expected inclusion in the respective companies’ software, but the multi-speaker ASR does not reliably identify who is speaking. Risk.net’s publisher summary separately describes CompatibL implementation as an expectation rather than a completed release. This supports dated implementation intent, not confirmed production use or client adoption.

The historical route also links the project to a 2023 presentation by Jesper Andreasen, then head of quantitative analytics at Verition Fund Management, which helped bring the researchers together. That is a project-origin and conference connection, not evidence that Verition uses the framework. CompatibL’s first-party machine-learning note separately describes a bounded use of autoencoders for curve-shape representation and says other model components remain classical, with validation and explainability as design objectives. It is company positioning, not an independent audit of the joint paper.

This source establishes a named finance-native ML architecture, a cross-firm research collaboration, explicit no-arbitrage and validation constraints, application areas, and a dated “not yet” production boundary. It does not establish a hedge fund’s use, a released software version, proprietary training data, client deployment, agent permissions, trading authority, cost reduction, or investment performance. See the timestamped capture note.

August 27 Risk.net Quantcast: Bank of America volatility-model route

The title-blind Risk.net Quantcast episode with Lyudmil Zyapkov adds a public Bank of America quant-research route that does not use AI or GenAI language. Risk.net identifies Zyapkov as a senior quantitative analyst at Bank of America and describes a fractional Gamma Clock model built from two correlated Gamma processes with a fractional Brownian-motion component. The public SoundCloud recording was recovered and transcribed locally on August 27, 2026. The related Risk.net article, publisher PDF, and SSRN record provide the public research layer.

The episode describes a model in which correlated Gamma processes represent random timing associated with underlying and volatility jumps, while fractional Brownian motion supplies a path-dependent volatility-persistence component (approximately 01:43–15:24, local recording). Zyapkov describes calibration to spot volatility, forward volatility, and, where available, forward-volatility-skew slope (approximately 05:45–07:34). Risk.net’s publisher summary reports the Gamma Clock’s time-dilation interpretation and daily-calibration framing; these are publisher and speaker descriptions, not an independent validation.

The practical route is specific. Zyapkov says the work was initially examined for FX options and forward-volatility agreements, had been moved to an equity index, and could potentially extend to equities and commodities; he describes interest-rate adaptation as more difficult (approximately 17:15–18:18). He identifies a full Monte Carlo simulation as a missing component (approximately 18:66–19:13) and says the work was not yet at internal model-validation stage, while describing possible benchmark and volatility-pattern signal-detection uses (approximately 21:47–22:27). Later he describes extended testing and validation as prerequisites to productionization (approximately 22:49–23:69). These statements establish a public research and implementation boundary, not a live Bank of America or hedge-fund deployment.

This is a useful negative control for the AI map. The source contains a finance-native quantitative model and a possible algorithmic-trading signal-detection use, but no machine learning, generative AI, large-language-model, agent, or fine-tuning disclosure. The public route therefore adds model-design and personnel evidence without supporting a claim about Bank of America’s AI strategy. See the timestamped capture note.

August 27 Risk.net Quantcast: HSBC market-making and information-risk route

The title-blind Risk.net Quantcast episode with Alexander Barzykin adds a current public HSBC market-making research route. Risk.net’s author page identifies Barzykin as a director in HSBC’s foreign-exchange, rates, and commodities team in London, specialising in algorithmic execution and market making. The public SoundCloud recording was recovered and transcribed locally on August 27, 2026. The related information-risk paper and internal-liquidity paper provide the named research layer.

The episode describes two market-making problems: adverse selection from informed or fast-moving client flow, and “price reading,” where a dealer’s risk-management quotes can reveal inventory or supply-demand information (approximately 03:20–07:13, local recording). It also discusses clients with long-term signals or additional data as a source of information asymmetry (approximately 06:09–06:32). These are model motivations and speaker explanations; the recording does not provide a client-level dataset, a production decision trace, or a measured HSBC result.

The methods are finance-native: dynamic programming, stochastic optimal control, perturbative or series-expansion techniques, and an optimal-stopping problem for deciding when to take internal liquidity (approximately 10:08–11:32 and 25:05–27:26). The internal-liquidity discussion concerns client algorithms, information barriers, internalisation, and market footprint (approximately 21:00–22:01). Barzykin links the work to a CFM–Imperial market-microstructure workshop and says Robert Boyce worked on the paper during an HSBC internship (approximately 23:20–23:50). That establishes a public project and personnel connection, not a general programme, shared proprietary data, or a firm-wide implementation.

This is another negative control for the AI map. The recording contains detailed algorithmic-execution, order-flow, stochastic-control, and market-microstructure material, but no machine learning, generative AI, large-language-model, agent, or fine-tuning disclosure. It demonstrates why discovery cannot require AI terminology in the title. The route does not establish HSBC production use, a hedge-fund implementation, a named client, model inventory, data rights, vendor partnership, trading authority, or performance. See the timestamped capture note.

August 27 Risk.net Quantcast: BNY quant operating model and AI-judgment route

The title-blind Risk.net Quantcast episode with Gordon Lee adds a current BNY personnel and operating-model route. Risk.net identifies Lee as head of markets quants at BNY and its author page describes responsibility for XVA and pricing models, alongside a prior UBS quantitative-analytics route. The public SoundCloud recording was recovered and transcribed locally on August 27, 2026.

The material AI disclosure is about work allocation, not a BNY technology inventory. Lee says that AI can reduce the cost and time of coding and model-writing execution; he expects junior quants to be valued more for judgment about whether a model is worth pursuing, which trade-offs are acceptable, and which risks are worth taking (approximately 18:56–20:08, local recording). He extends the point to repetitive knowledge work, describing it as increasingly AI-enabled or AI-assisted and saying that workers reach judgment-oriented responsibilities earlier (approximately 20:08–21:00). These are speaker-reported views. The recording names no model, provider, internal tool, agent, training corpus, evaluation set, or production endpoint.

The broader operating model is also relevant to automation research. Lee describes visibility, a clear account of why work matters, intermediate milestones, alignment, stakeholder trade-offs, and explicit boundaries as filters through which technical work becomes actionable in a large organization (approximately 03:47–18:56). This offers a public hypothesis about where automation leaves management work—judgment, prioritization, accountability, and coordination—but it is not a measured BNY productivity result or a disclosed policy. Lee’s separate 2023 LLM commentary is public author commentary and is kept separate from firm deployment evidence. See the timestamped capture note.

August 28 regional-language recovery: BNY GenAI platform and Sarthak Pattanaik

The title-blind regional-language pass recovered Risk.net’s English article on BNY’s risk-centric GenAI approach and its Japanese localized counterpart. The English page, published February 23, 2026, identifies Sarthak Pattanaik as BNY’s chief data and AI officer and frames the public discussion around a centralized platform, governance, risk management, and talent. Pattanaik’s public framing treats AI as a foundational layer for future technology products; it does not name an internal architecture or budget.

The Japanese publisher page exposes a longer excerpt that reports 99% of BNY employees have access to GenAI tools, lists coding assistance, research support, modelling, and risk-management applications, and describes approximately 20,000 of an approximately 50,000-person workforce as actively developing agents. These are Risk.net-reported figures from a localized publisher surface, not independent measurements. “Access,” “use,” and “active agent development” are different quantities, and the article does not provide the underlying denominator, timestamp, adoption logs, agent registry, model/provider inventory, evaluation controls, or permission map. The capture note preserves the regional-language discovery route and the evidence boundary.

This route adds a named AI executive and a public operating-model signal, not proof of a specific BNY deployment. It should remain separate from Gordon Lee’s practitioner discussion: Pattanaik’s article concerns BNY’s enterprise GenAI framing, while Lee’s recording concerns quant work allocation and judgment. Neither source establishes trading use, autonomous action, financial performance, or a complete internal reporting line.

August 28 first-party partner route: BNY’s Eliza platform and OpenAI

OpenAI’s December 12, 2025 customer story is a separate, vendor-authored source that makes BNY’s public GenAI strategy more concrete. It identifies ChatGPT and API as the products, names BNY’s centralized AI Hub and internal Eliza platform, and says Eliza combines governance with leading models including OpenAI frontier models. The story names Sarthak Pattanaik as chief data and AI officer, Watt Wanapha as deputy general counsel and chief technology counsel, Michelle O’Reilly as global head of talent, and Ed Fandrey as head of sales and relationship management. These are named participants and role descriptions in the case study, not a complete ownership map.

The operating model described by OpenAI has three visible layers: a data-use review board covering intellectual property, cybersecurity, engineering, data, privacy, and third-party relationships; an AI release board for pre-production review; and an Enterprise AI Council for senior oversight and policy alignment. The case study says Eliza embeds model selection, agent development, sharing, permissions, security, approval flows, tagging, telemetry, and oversight in a governed environment. It also describes mandatory training, a daily seven-minute enablement series, private agent builds that later became shareable within selected teams, and a selected group experimenting with ChatGPT Enterprise and deep research. This is a public description of control design, not an audit of implementation or approval outcomes.

OpenAI reports more than 125 live use cases and 20,000 employees actively building agents. It gives examples including a Contract Review Assistant, reported to reduce review time from four hours to one across more than 3,000 annual vendor agreements; a People Business Partner Agent; a Lead Recommendation Engine; a permission-aware Metrics Agent; and a Risk Insights Agent that surfaces portfolio signals for analysts. It also describes controlled “digital employees” for payment-instruction validation and code-security work, and says deep research is being explored for risk modelling, scenario planning, strategic decisions, legal questions, and client preparation. The 75% legal-review result, 99% training/access figure, and 46% agent-growth figure are vendor/customer-story claims with no disclosed control group or independent audit. The source does not establish model versions, fine-tuning, data rights, API volume, vendor exclusivity, autonomous trading, portfolio authority, or investment performance. See the capture note.

Risk.net title-blind route: Mark Higgins and AI-assisted derivatives research

The title-blind Risk.net pass recovered New LLMs are proving to be surprisingly good quants, published March 18, 2026. The article identifies Mark Higgins as a retired Beacon Platforms co-founder and former co-head of quantitative research at JPMorgan, and reports that he used two LLMs on an options-pricing problem. The publisher page is a discovery and identity source; it does not expose a prompt log, model versions, benchmark, or production deployment.

Higgins’s public LinkedIn post adds the workflow detail: five candidate modeling ideas, ChatGPT for mathematical generation, Aristotle for mathematical validation, a coding AI for implementation, manual testing, and a CUDA Monte Carlo implementation on a desktop GPU. The account is a practitioner report, not a controlled productivity study or proof of institutional adoption.

The associated arXiv paper is a separate technical layer. It presents an affine Double Heston parameterization with stochastic spot/volatility correlation and studies implications for FX barrier options and volatility swaps. Its reported price effects are model-analysis results under stated calibration assumptions, not live trading performance. The earlier 2014 paper shows that the research problem predates the 2026 AI workflow, and a public Beacon-era deep-hedging presentation provides historical derivatives/ML context. None of these sources establishes hedge-fund adoption, confidential data access, autonomous research authority, or a production trading endpoint. See the capture note.

August 28 Cambridge University Algorithmic Trading Society conference route

The first-party CUATS Quant Conference 2026 page records a March 6, 2026 Cambridge event with six presentations spanning order flow, option risk, statistical arbitrage, FX information risk, generative deep hedging, and replication controls. Its “Learning to Trade: From Market Simulation to Hedging” abstract describes market simulators and reinforcement-learning environments for derivatives trading and risk management; the speaker biography identifies Hans Buehler as an Oxford visiting professor, former Deutsche Bank and J.P. Morgan quantitative-research leader, and former XTX Markets deputy CEO and co-CEO. The page also identifies Alexander Barzykin of HSBC, Pavel Ioselevich of CFM, and Marco Dion of QRT. Ioselevich’s abstract describes PCA and a data-driven factor model for option-portfolio risk, while Dion’s abstract focuses on multiple testing, data snooping, non-stationarity, regularization, ensembles, and out-of-sample controls. These are agenda abstracts and speaker biographies, not recordings or proof that any named firm adopted the presented methods. The page provides no model inventory, data rights, production permissions, autonomous authority, or investment-performance evidence. See the capture note.

August 28 BattleFin Montauk investment-data and AI network route

BattleFin’s first-party Montauk: Data & AI on the Edge page dates an application-only summit to August 10–11, 2026. It frames the event around agentic AI in investment research, modern research and data stacks, alternative data, consensus and variant perception, and the analyst workflow. Its public Luminary Committee names Sameer Gupta, Chief Data and Analytics Officer at Apollo Global Management; Debbie Lawrence, Group Head of Data Strategy & Management at LSEG; Tony Berkman, identified as former Two Sigma Data Science; Stewart Stimson, identified as a former Jump Trading data thought leader; Martin Vulliez, President of Cadian Capital; and Dan Entrup, Co-Founder of AggKnowledge. The page says roundtables were closed-door and limited to 12–15 participants. This adds a personnel and discovery-network route, not session evidence: no recording, transcript, model/provider inventory, dataset rights, permissions, production endpoint, or investment-performance result is public on the reviewed page. See the capture note.

August 28 upcoming open-source and allocator/data conference routes

The official Open Source Quantitative Finance (osQF) page schedules its annual conference for October 23–24, 2026, at the University of Illinois Chicago, with a PyData Chicago event on October 22. The organizer describes topics spanning AI, advanced risk tools, decentralized finance, econometrics, high-performance computing, market microstructure, portfolio management, and time-series analysis, all around open-source software for financial model development and trading. Its public committee links provide additional researcher and practitioner surfaces, but no covered hedge-fund affiliation is inferred from those links. The official Battle of the Quants London 2026 page dates its event to October 22, 2026, at the Royal Automobile Club and describes allocator and data breakfasts focused on quantitative-manager strategies, datasets, and signals. It advertises AI, machine learning, LLMs, agentic AI, and alternative data, but the accessible page does not expose named speakers or session recordings. These are useful future recovery targets and title-blind discovery routes, not evidence of attendance, firm implementation, dataset purchase, model ownership, permissions, or performance. See the capture note.

August 28 regional investment-technology routes: Sydney and Singapore

The STAC Summit Sydney page dates an upcoming September 10, 2026 event at the Hilton Sydney and explicitly targets technical leaders in trading and investment, including machine-learning, deep-learning, data-engineering, and operational-intelligence roles. Its displayed speakers include representatives of ASX, Commonwealth Bank of Australia, Macquarie, Platinum Asset Management, Goldman Sachs, and market-infrastructure and technology firms such as Avelacom, AWS, and WEKA. The accessible page exposes names and organizations but not agenda abstracts, recordings, model names, datasets, or firm implementation. The MSCI Singapore Investment & Risk Summit page dates an invitation-only September 2, 2026 event and lists Chao Jen Chen, Head of Augmented Intelligence at Fullerton Fund Management; Hemang Mandalia, quant-equities portfolio manager at AlphaGrep; and Joo Won Lee, co-founder and CTO of Arrowpoint Investment Partners, alongside leaders from GIC, Bank of Singapore, and Eastspring. Its agenda includes AI adoption in risk and portfolio management, scaling AI intelligence across investment and risk, and factor-aware construction of AI exposures using MSCI datasets and optimization tools. These are event and organizer/vendor agenda signals, not evidence that the named firms deploy a particular model, agent, dataset, or investment process. See the capture note.

August 28 Fullerton first-party Augmented Intelligence disclosure

Fullerton’s current Augmented Intelligence page says data-science techniques are integrated into portfolio-manager analysis and describes tools for sentiment shifts, market data, signals, and observable trends. It says the firm built an Augmented Intelligence team with machine-learning, data-science, and quantitative-finance skills, dating that team development to 2019. Fullerton’s November 7, 2023 Insights page links a four-page supplement that dates investment-process incorporation to 2020 and a dedicated team to April 2021; the two dates are retained as separate first-party claims. The supplement describes a multi-step machine-learning clustering model for growth, inflation, liquidity, and risk appetite; an in-house NLP sentiment dashboard for earnings calls, investor/shareholder meetings, and conference presentations; and an illustrative tree-based classifier for top-quintile stock prediction using technical, fundamental, estimate, and valuation inputs, filtered by market capitalization, liquidity, and forward EPS CAGR, refreshed monthly, and discussed in team meetings. This is unusually specific public workflow evidence, but it does not disclose code, model versions, training windows, data rights, evaluation design, turnover, costs, live permissions, or realized performance. The separate MSCI Singapore event page lists Chao Jen Chen as Fullerton’s Managing Director and Head of Augmented Intelligence; that is a personnel lead, not proof that Chen owns every workflow in the older supplement. See the capture note.

August 28 China title-blind manager route: MY Capital

MY Capital’s What Do We Do page describes a Chinese hedge fund founded in 2013 around algorithmic trading and quantitative investment. It says data management, model development and validation, strategy implementation, and trading take place within an in-house end-to-end research and investment platform, and lists a Hong Kong office opened in 2023. Its Process page describes high-performance computing, distributed data storage, predictive signals, risk-constrained portfolio construction, and proprietary execution algorithms that consider market impact and market microstructure. The Management page identifies CTO Xu Xinxin as overseeing modular toolsets and centralized AI expertise; it also provides academic lineage for founder and CIO Raymond Guo Xuewen (Tsinghua and University of East Anglia), equity Co-CIO Wu Ganghui (USTC and Fudan), and CTA/macro Co-CIO Yang Min (Peking University and Caltech, with prior Goldman Sachs and Winton roles). This is a useful China/Hong Kong platform and personnel route, but it does not identify model weights, training data, external data rights, evaluation results, current AI-tool contents, or strategy-level performance. See the capture note.

August 28 CenterBook title-blind alpha-capture route

The Podscan episode, also available through an Apple Podcasts listing, identifies David Stemerman as CenterBook Partners’ CEO, CIO, and co-founder. He describes an alpha-capture and active-extension model that receives independent managers’ positions, price targets, scenarios, qualitative inputs, and portfolio rules, then applies systematic risk modeling, portfolio construction, and execution analysis. At approximately 05:57–08:49, 14:18–16:20, and 58:59–01:03:13, the account frames research judgment and systematic portfolio construction as complementary layers and uses a “centaur” analogy. CenterBook’s public position-paper page says it has more than 10 years of proprietary data across more than 200 long/short fundamental-equity managers and more than three years of global trading; its contributors page describes daily post-market position files through an administrator or OMS, contributor reporting, and trading safeguards. These are first-party and practitioner claims, with an automatic transcript used for navigation. The route is relevant to the human/machine and data-governance dimensions of quant-fund operating models, but it does not disclose GenAI, LLMs, autonomous agents, model providers, training data, customer permissions, or independently audited performance. See the capture note.

August 28 Citco / Alternative Fund Insight title-blind middle-office AI route

The Alternative Fund Insight episode, published July 2, 2026 and recoverable through the show’s Acast RSS feed, identifies Ryan Fitzgerald as Citco’s Head of Middle Office Solutions and Declan Quilligan as Head of Hedge Fund Services. Fitzgerald describes three operational AI use cases: treasury fraud protection, automation of manual collateral-management work, and automatic trade matching with attention directed to exceptions (14:38–15:07). He also describes an LLM interface over operational products and data (15:07–16:03) and a possible future in which an agent manages processes while people manage the output (14:09–14:38). Citco’s July 14 first-party interview independently confirms Fitzgerald’s title and describes the underlying service surface as treasury management, trade operations, collateral management, data consolidation, and selective outsourcing. This is a useful operating-layer route for understanding where automation may sit around a hedge fund; it is vendor/provider evidence, not proof of any named manager’s production deployment, model choice, permissions, or investment use. See the capture note.

August 28 Curious Quant title-blind FX and alternative-data route

The archive recovery found a title-blind Curious Quant episode whose RSS title spells the guest “Saeed Amed.” The transcript and public professional profile identify him as Saeed Amen, a Turnleaf Analytics co-founder and CTO with earlier systematic-FX work at Lehman Brothers and Nomura and a Cuemacro route. The spelling discrepancy is recorded rather than silently normalized. This is a practitioner and data-methodology route, not evidence of a tracked hedge fund’s deployment.

The episode places machine learning mainly in data structuring—extracting sentiment from natural language and potentially combining trend, carry, growth surprises, PMI, and other features—rather than claiming a direct FX price model (34:43–36:19). It distinguishes fast event-based news trading from slower aggregation, describing day-to-weeks sentiment aggregation, shorter-horizon momentum, and possible longer-horizon fading (38:36–39:52). For emerging markets, Amen presents local-language news as an untested hypothesis rather than a measured result (25:31–26:18). The discussion also describes anonymized CLS flow data grouped into broad categories such as corporate and buy-side flow (40:08–40:32).

This adds three research lanes to the cross-firm map: multilingual macro-news coverage, the separation of feature construction from price-signal generation, and anonymized flow data as a possible contextual input. The speaker explicitly leaves production use and statistical validation open; the recording does not identify a fund customer, model provider, training corpus, data license, evaluation split, permission boundary, or performance result. See the timestamped capture note.

August 28 Curious Quant title-blind Boston HFT and AI boundary route

The same archive recovery found a second title-blind Curious Quant episode with Christina Qi. The automated transcript misrenders her name in places; her public professional site and LinkedIn profile identify her as Christina Qi, former Domeyard founder and current Databento CEO. The episode describes a Boston HFT operating context and a dated practitioner view of how AI fit into that environment.

Qi says the firm used limited AI for strategy selection and selected obscure strategies, while latency sensitivity constrained direct use in much of the trading stack; she places more plausible use in middle- and back-office work (17:37–18:33). She describes open-source tooling as a way to let interns work with data without exposing core technology, and says research hiring and internal process improvement could be a better use of marginal spend than paying to be the very fastest (18:48–22:15). She also describes strategy durability as highly variable, from weeks to years, with revisions over time (30:53–31:42).

This is a useful control against equating HFT with broad AI deployment: the account separates latency-critical trading from research selection, process automation, and later technology transfer to discretionary teams. It does not establish current Domeyard or Databento systems, a tracked firm’s model inventory, data rights, agent permissions, production authority, or independently measured performance. See the timestamped capture note.

The unfiltered Risk.net pass also recovered Patrick Hagan’s 2021 Quantcast episode, whose title does not contain “AI” or “hedge fund.” The SoundCloud recording was downloaded and transcribed locally. Risk.net introduces Hagan as managing director of Gorilla Science and associates him with the SABR model; its dated personnel report records a former J.P. Morgan CIO quant role and a Deutsche Bank consulting route. These are historical quant and model-risk records, not current hedge-fund employment evidence.

Hagan discusses convexity-adjustment methods, rough volatility, Lévy-flight models, calibration, and hedging residuals (01:27–18:34). His London Whale discussion emphasizes the limits of historical VaR and the need for scenario risk when regimes change (18:34–26:56). He also describes potential hedge-fund counterparty-risk opacity (27:39–29:57). The episode contains no AI, GenAI, LLM, agent, or provider disclosure; it is retained as a negative control for the title-blind process rather than as AI evidence. See the capture note.

August 28 Risk.net/Danske Bank differential-ML route

The title-blind pass recovered Antoine Savine and Brian Huge’s September 2021 Quantcast episode, whose title does not itself expose “AI” or “machine learning.” The SoundCloud recording was downloaded and transcribed locally. Risk.net identifies Huge as a senior specialist quant at Saxo Bank and Savine as chief quantitative analyst with Superfly Analytics at Danske Bank. Risk.net’s related differential-machine-learning article and the pair’s public GitHub organization provide independent publication and code surfaces.

The speakers describe combining automatic adjoint differentiation with neural networks and derivative-aware training. Their stated use cases include pricing and risk approximations for trading books, XVA, CCR, SIMMVA, backtesting, risk heat maps, and dynamic reports; Huge says the target is real-time computation on trader workstations (02:35–09:39). They describe differential PCA as using pathwise sensitivities to identify relevant risk factors and reduce dimensionality before training (11:16–15:12). The public repository contains demonstrations and production-oriented implementation notes, but not proprietary trading books or production code.

Savine says Danske Bank had deployed AAD in 2015 and says differential ML was being deployed for XVA at Danske Bank at the time of the September 2021 recording (07:40–09:39; 16:33–18:34). This is dated speaker-reported bank-workflow evidence. It concerns derivative pricing and risk analytics rather than alpha generation, and it does not establish current deployment, model providers, proprietary data rights, autonomous investment authority, or performance. See the capture note.

August 28 Risk.net/Pine Tree Market Neutral personnel and tail-risk route

The unfiltered Risk.net archive recovered Jan Rosenzweig’s May 2023 Quantcast episode, whose title does not contain “AI” or “hedge fund.” The SoundCloud recording was downloaded and transcribed locally. Risk.net identifies Rosenzweig as a portfolio manager at Pine Tree Market Neutral and a visiting senior research fellow at King’s College London; in the recording he describes Pine Tree as a small equity market-neutral quantitative fund and says he is one of two portfolio managers.

Rosenzweig describes a family of fat-tail and extremal risk measures, including a decomposition of liability hedging and risk-adjusted return, with analytical interpretability as a design objective (01:26–08:14). He discusses moving from variance-oriented optimization toward tail-focused and minimax behavior, independent components, non-linear factor interactions, and focused hedging (08:25–15:47). In discussing the LDI crisis, he distinguishes the funding gap from a portfolio-optimization failure and does not claim that his method would have prevented it (16:27–20:13).

The episode contains no AI, GenAI, LLM, agent, or provider disclosure. It is retained as a title-blind tracked-manager personnel and methodology route, not as AI evidence or proof of Pine Tree’s current systems, data, permissions, production status, or performance. See the capture note.

August 28 Risk.net title-blind MIT adaptive-markets and ML-governance route

The unfiltered Risk.net archive also recovered an Andrew Lo Quantcast episode whose title contains neither “hedge fund” nor “AI.” The direct SoundCloud recording and the MIT Sloan faculty profile resolve an ASR spelling of “Andrew Lowe” to Andrew W. Lo, Harris Professor of Finance and director of MIT’s Laboratory for Financial Engineering.

The episode distinguishes algorithmic execution for transaction-cost reduction from delegating investment decisions to algorithms (19:43–21:04). Lo connects the August 2007 statistical-arbitrage unwind to similar algorithmic exposures and argues that broader observations of market participants and their choices could support dislocation analysis (21:50–26:37). He describes a research discipline of stating a market worldview before testing, correcting for multiple searches, and rejecting unsupported theses (29:35–37:44).

The most relevant AI signal is methodological: Lo describes using machine learning not only to verify predictions but eventually to generate hypotheses, with perturbation-based interpretation and inductive-logic approaches to distill parsimonious explanations (35:09–45:37). He also describes combining behavioral science, computer science, AI, machine learning, data analysis, neuroscience, and psychology to model decision-making and support portfolio/risk tools (46:11–47:15). This establishes a dated academic-practitioner research route and a strong title-blind discovery pattern; it does not establish a current tracked-hedge-fund deployment, model provider, training data, production permission, or performance result. See the timestamped capture note.

August 28 Risk.net JPMorgan/XTX deep-hedging and simulator route

The unfiltered Risk.net archive recovered a Hans Bühler Quantcast episode, with the SoundCloud recording and a related drift-removal paper. Risk.net identifies Bühler as JPMorgan’s global head of equities analytics, automation and optimisation at the time. A later Risk.net personnel report records his move to XTX Markets as deputy CEO, while an XTX regulatory disclosure lists him as co-CEO in the reviewed document. These are date-scoped personnel and research links, not evidence that the JPMorgan work transferred to XTX.

At 03:31–05:21, Bühler describes statistical hedging for index vanilla products and expanding work around S&P and single-stock options, while positioning deep hedging around complex products such as cliquet options. At 06:42–08:04, he explains drift removal as a way to stop a neural network from turning a hedging mandate into an unintended directional strategy. He separates that risk-control objective from alpha generation and says the engine operates within established classic risk, capital, and limit frameworks (10:39–11:25).

The data-engineering route is equally important: Bühler describes GANs, variational autoencoders, and signature-based market dynamics (09:22–09:46), and says sparse observations motivate market and client-behaviour simulators on which higher-order models can be trained (12:46–13:46). His academic lineage through Hans Föllmer and Alexander Schied is stated in the recording (11:52–12:31). The route establishes concrete, dated bank research and a later XTX personnel connection; it does not establish current XTX implementation, any tracked hedge fund’s use, model weights, data rights, autonomous permissions, or performance. See the timestamped capture note.

August 28 Risk.net/HedgeFacts nonlinear-risk and quant-research route

The unfiltered Risk.net archive recovered a Matthew Dixon Quantcast episode with a direct SoundCloud recording. Dixon’s public research site describes work spanning machine learning, stochastic control, quantitative finance, and financial model risk. The episode is not a tracked-hedge-fund disclosure, but it exposes a named platform route: at 29:34–30:30, Dixon says the nonlinear risk-decomposition method was developed for HedgeFacts, which he describes as a cloud risk-management solution for hedge funds.

The method decomposes nonlinear derivative portfolio risk across factors, managers, instruments, and subportfolios, while preserving cross-factor correlations (00:18–01:19; 11:11–14:15). Dixon says the implementation is designed for rapid what-if restructuring without repeated Monte Carlo simulation and that source code was provided (30:39–32:25). The episode also records planned expected-shortfall work as future research (32:36–34:06).

This creates a useful vendor and research-partner lane: a quant researcher’s method was described as being built for a hedge-fund operations/risk platform, with a public software and methodology surface. The source does not identify client firms, contract terms, GenAI or LLM use, model providers, data rights, production permissions, autonomous investment authority, or performance. See the timestamped capture note.

August 28 Risk.net/DZ Bank algorithmic-differentiation and legacy-code route

The unfiltered Risk.net archive recovered a Christian Fries Quantcast episode with a direct SoundCloud recording. Risk.net identifies Fries as head of model development and methodology at DZ Bank and a professor at LMU; the LMU profile independently confirms that route. Public finmath documentation exposes a related backward-automatic-differentiation implementation.

The episode describes AAD for Bermudan-option and XVA sensitivities, including the connection between backward differentiation and neural-network backpropagation (04:22–09:18). Fries emphasizes runtime order of magnitude, parallelism, clean code, understandable logic, and legacy-system integration rather than an abstract speed contest (14:43–17:00). He says DZ Bank was reimplementing the university Java implementation for industry use (23:05–23:38), with follow-up work on forward sensitivities and approximations for margin valuation adjustment (24:24–25:52).

This is a model-risk and software-engineering route relevant to how quantitative institutions move research into production: academic code, bank reimplementation, clean-code requirements, and fallbacks for legacy systems. It does not establish any tracked hedge fund’s use, GenAI or LLM deployment, model providers, training data, permissions, autonomous investment authority, or performance. See the timestamped capture note.

August 28 Risk.net title-blind Imperial/CFM machine-learning and quant-risk route

The unfiltered Risk.net Quantcast archive recovered Damiano Brigo’s January 2018 episode, whose title contains neither “hedge fund” nor “AI” in its primary episode slug but whose article explicitly discusses both machine learning and quantitative finance. The SoundCloud recording was downloaded and transcribed locally. Risk.net identifies Brigo as Chair of Mathematical Finance at Imperial College London; Imperial’s current faculty profile independently records that role and his CFM-Imperial Institute of Quantitative Finance affiliation.

Brigo’s account gives four useful research boundaries. First, he calls for mathematical work on black-box ML outputs, convergence conditions, and interpretability (01:40–02:53). Second, he describes information-theoretic metrics for locating useful information in investment firms’ large customer, market, and information datasets, explicitly connecting that problem to hedge funds and buy-side firms (03:09–04:24). Third, he frames optimization as a way to move risk work from measurement toward active risk and capital management (04:24–05:48). Fourth, he distinguishes then-current statistical/ML systems from human-level AI and highlights regulatory, treasury, collateral, systemic-risk, and model-calibration problems that remain quantitative research targets (08:33–10:14; 14:29–22:57).

This is an academic/industry feeder and a title-blind discovery result. It does not establish a tracked hedge fund’s current AI or GenAI deployment, model inventory, provider, training data, permissions, autonomous investment authority, or performance. See the timestamped capture note.

August 28 Artur Sepp/LGT title-blind quant and applied-GenAI route

The unfiltered Risk.net archive recovered Artur Sepp’s August 2023 Quantcast episode, whose title does not identify AI but does expose a named quant practitioner, a prior systematic-hedge-fund route, and buy-side model-development context. The SoundCloud recording was downloaded and transcribed locally. Risk.net’s profile page identifies Sepp as head quant at LGT Bank and records his prior systematic-hedge-fund and family-office experience; his current first-party site describes him as Global Head of Quantitative Analytics at LGT Private Banking, leading a quant team of 10+ across portfolio construction, factor analytics, systematic macro, and applied GenAI.

The current public AI signal is hiring and workflow language, not a model reveal. LGT’s GenAI/ML Engineer listing names a portfolio-analysis system and asks for ML, LLM, RAG, Python, and financial-data analysis. A public Sepp repost frames the role as joining his quant team and describes automated portfolio analysis using generative AI. Those sources establish a recruiting and intended-system surface; they do not establish a filled role, model provider, training corpus, data rights, evaluation, permissions, or production status.

The public GitHub profile adds a separate code surface: ten public Python packages spanning portfolio analytics, factor models, portfolio construction, stochastic-volatility and option modelling, trend-following, private-asset analysis, and Bloomberg data access. The repositories are useful for understanding the published quant workflow and reproducibility posture, but they do not establish that LGT uses the code in production. The RCM Alternatives transcript separately identifies Sepp’s earlier Quantica Capital systematic trend-following hedge-fund experience.

The recovered audio describes robust volatility modelling, risk decomposition, portfolio construction, and DeFi interaction-risk controls (00:45–06:27; 15:37–17:08; 38:27–45:09). It is a dated research and personnel route, not evidence that the current LGT applied-GenAI description uses the open-source packages or the historical Quantica methods. See the capture note.

August 28 Risk.net/Fidelity inverse-reinforcement-learning route

The unfiltered Risk.net archive also recovered Igor Halperin’s December 2022 Quantcast episode, whose episode title does not contain “hedge fund” or “AI,” even though Risk.net’s page describes Fidelity quants working on machine-learning techniques for investment strategy. The SoundCloud recording was downloaded and transcribed locally. Risk.net identifies Halperin at the time as a senior quant analyst at Fidelity’s AI Center for Asset Management; its author page records a Fidelity quantitative-research role. These are dated publisher records, not a current employment or deployment claim.

The episode describes a two-stage research design. Inverse reinforcement learning infers a reward function from observed manager positions and trades, after which reinforcement learning optimizes against that inferred objective (01:33–05:43). In the paper’s example, the model operates on 11 sector allocations; a separate momentum rule maps sector changes to individual stocks (06:32–08:10). Halperin says the reported improvement is backtest-only and comes from a group of managers with shared objectives and benchmarks (09:21–10:14).

The design boundary is unusually clear. Halperin says the construction does not use neural networks and is intended to preserve interpretability (13:04–15:40). He describes the work as continuing research and a possible future helper for fund managers, not as a disclosed live Fidelity product. Later discussion identifies multi-agent RL and tensor networks as research directions (32:52–38:30). This adds a dated Fidelity-linked research route and a useful distinction between interpretable RL research, backtesting, and production evidence. It does not disclose current Fidelity systems, model providers, training data, permissions, autonomous investment authority, or live performance. See the timestamped capture note.

Sources

The source ledger at sources/06-industry-verticals/gmo-acadian-arrowstreet-ai-public-signals-2026-raw.md records retrieval dates, source tier, signal type, and evidence gaps. The earlier State of AI deep dive and worker-layer ledger are also included in the provenance chain. Selected primary sources are linked below.

New source note: Third Point founder media route — Daniel Loeb on AI and investment work

The August 14 expansion ledger is hedge-fund-ai-source-expansion-2026-08-14-raw.md. It records the new official archives, vendor/customer pages, partner ecosystem, current titles, conference surfaces, and negative-search boundaries used in the added section above.

The lab and lineage ledger is hedge-fund-ai-labs-academic-lineage-2026-08-14-raw.md. It records named labs, lab-adjacent groups, public personnel pages, academic training, paper links, cautious research-topic linkages, and disqualification rules.

The August 14 follow-up ledgers cover leadership resolution, conference/video mining, provider and data-partner evidence, and dated role history. They preserve current, former, event-date, and third-party statuses separately.

The extended loop also adds emerging formation signals, separating AI-native vendors, new-manager formation claims, founder personnel migration, and former-firm affiliations from established-manager evidence.

• GMO, Our Approach, current firm page; investment philosophy.
• GMO, Systematic Equity Year-End Letter 2025, 2025/2026; systematic research direction.
• GMO, Part 2 Form ADV, as of 2025-12-31; regulatory AI-risk and AUM disclosure.
• GMO, Quality Strategy 2025 Year-End Letter, 2026; AI ecosystem and quality-investing analysis.
• GMO, Top Market Predictions for 2026, 2026-01-07; Tom Hancock role/background.
• GMO, We Asked a $30 Billion Manager Where AI Profits Will Actually Go, 2026-04-09; public video with Tom Hancock.
• GMO, Technology Innovation to Generate Alpha, current research page; NLP infrastructure signal.
• Fordham Gabelli School of Business and Rebellion Research, QuantVision 2026 publication and agenda, Fordham recap, and detailed public programme, March 19–20 2026; conference speaker and topic-discovery route without a recovered recording.
• GMO, Hype vs High Conviction, March 2026; current AI investment framework.
• GMO, Data Platform Analyst and Client Systems Manager, current public roles; operational AI and GenAI hiring signals.
• Acadian, Our Systematic Edge, current firm page; data and team scale.
• Acadian, Generative AI in Systematic Investing: The Sizzle and the Steak, 2024-04; Tier 3 historical GenAI thesis.
• Acadian, Credits: The Systematic Shift, 2026-06; current AI/NLP credit workflow positioning.
• Acadian, VP, Investment AI Engineer, retrieved 2026-08-05; primary hiring signal hosted on LinkedIn.
• Acadian, interview on AI in investing, 2025-07-28; named-practitioner podcast.
• Acadian, Vladimir Zdorovtsov, current leadership page; quantitative-research background.
• Acadian, Acadian Adds to Investment Team With Key Hires, 2023; Javier Alcazar background.
• Acadian, 2026 Investor Forum presentation, 2026-05-19; AI strategy pillars and data-organization disclosure.
• Acadian, Net Zero Alignment Model, 2023; LLM-based ESG/text use case.
• Acadian, 2025 Responsible Investment Statement, 2025; in-house AI/NLP/LLM use cases.
• Arrowstreet, Investment, current firm page; systematic process.
• Arrowstreet, Technology, current firm page; technology platform.
• Arrowstreet, Senior AI Platform Engineer, retrieved 2026-08-05; primary platform-hiring signal.
• Arrowstreet, Senior AI Security Engineer, retrieved 2026-08-05; indexed security-hiring signal.
• Arrowstreet, official homepage, retrieved 2026-08-14; current $292B+ and 500+ employee scale observation as of 2026-03-31, plus data-science and high-performance-computing language.
• Arrowstreet, Ties de Kok CV, retrieved 2026-08-05; public technical adjacency, not deployment proof.
• Arrowstreet, claude-code-pd-protection, retrieved 2026-08-05; personal guardrail artifact, not firm-owned code.
• Man AHL, The Big Innovation Imperative, 2025; GenAI and strategy design.
• Man AHL, AI, Agents and Trend, 2025; Alpha Assistant and proprietary research workflow.
• Man AHL, AlphaTrend and Agentic Research Workflows, 2026; specialized signal-research workflow.
• Man AHL, Code with Claude session, 2026-05-19; vendor-reported production-signal claim with no public attribution.
• Man AHL, LLMs are hallucinating, 2025; RAG, human-review, uncertainty, and query-decomposition controls.
• Man, Tushara Fernando and Matthew Hertz, current people pages; data/AI strategy and central-platform personnel signals.
• Man, Anthropic partnership, AlphaTrend, and 2025 annual report, retrieved 2026-08-14; Claude workflows, model-behavior comparison, 100+ internal AI plugins, and reorganisation language.
• CFM, ML Lab announcement, AI-native quant researcher, Anastasia Borovykh, Eric Vanden-Eijnden publications, CFM-ENS chair, and Giulio Biroli, retrieved 2026-08-14; named lab remit, personnel, academic bridge, training vocabulary, and research-topic evidence.
• Bridgewater, AIA Labs, current firm page; AI research lab, artificial investment associate, and investment-process disclosure.
• Bridgewater, How AIA Labs Built PAT, 2026; LLMs, agentic workflows, and investment research tooling.
• Bridgewater, Artificial Intelligence, current research hub; named AIA Labs leadership and AI-risk context.
• Bridgewater, Nina Lozinski, 2026; AIA Labs staffing and role disclosure.
• Bridgewater, Blake Cecil, current people page; AIA Labs and investment-talent/technology role.
• Bridgewater, Oliver Simon, 2026; AI & ML Investment Strategy role.
• Bridgewater, Jasjeet Sekhon, current profile; current DeepMind role and former Bridgewater AI role.
• Bridgewater, AI research index, current; research and code-linked publication index.
• Jane Street, Machine Learning, current firm page; ML trading models and training/inference infrastructure.
• Jane Street, Quantitative Research, current firm page; research, model, strategy, and implementation roles.
• Jane Street, Iterating on intelligence, current firm article; internal GenAI systems, RAG applications, and plugins.
• Jane Street, From Code to Woodcraft, current firm article; custom LLM-use tooling.
• Jane Street, AI Engineering at Jane Street, 2025 talk; code-generation models and build/test evaluation.
• Jane Street, ocaml-torch and GTC 2026; public ML runtime/performance artifacts.
• Jane Street, Machine Learning Researcher and Visiting Researcher pages, accessed 2026-08-12; current LLM/RL-agent hiring vocabulary and custom/third-party model-training statement; no direct LLM-to-trading proof.
• CFM, Our approach, current; AI/ML/cloud/data and risk-control substrate.
• CFM, Hugging Face fine-tuning case study, 2024; narrow financial-NER weight modification.
• CFM, CFM AI effort profile, 2025; Eric Vanden-Eijden and internal AI effort.
• CFM, Prediction Services listing, current public job listing; GenAI agents and production model services.
• CFM, Why Systematic Global Macro Is Having a Moment, 2026 firm-hosted interview; AI/ML use cases, multimodal extraction language, and agentic test/code workflow.
• CFM, Investing After the AI Honeymoon, alternative-data initiative with Columbia, and The Next Level of Alt Data, retrieved 2026-08-14; ML lab, GenAI, multimodal text/video/audio research, alternative-data partnership, and historical financial-news retraining.
• Balyasny, OpenAI AI research engine case study, 2026; centralized Applied AI, model evaluation, agents, and scoped deployment.
• Anthropic, Financial-services agents and Claude for Financial Services, retrieved 2026-08-14; named data/MCP partners, Citadel Claude-for-Excel disclosure, Walleye adoption statement, and Bridgewater assistant statement.
• Walleye, official site, people page, and company LinkedIn, retrieved 2026-08-14; strategy groups, titles, Claude Code adoption, model-testing claims, and promotion signals.
• Jane Street/CoreWeave, AI-cloud agreement and GTC 2025 recap, retrieved 2026-08-14 and 2026-08-27; compute-capacity, equity-investment, and vendor-conference model-workload disclosure without Jane Street-authored model inventory, production permission, or return proof.
• Balyasny, Applied AI market-event profile, 2025; personnel/workflow signal.
• Balyasny, BAMChatGPT roundup, 2025; internal-assistant and adoption signal.
• BlackRock-affiliated paper, ACM ICAIF 2025 accepted-papers page, accessed 2026-08-12; LLM agents for investment-management foundations and benchmarks; research publication, not product or deployment proof.
• BlackRock / AI Engineer, How BlackRock Builds Custom Knowledge Apps at Scale, official YouTube captions captured 2026-07-02 and inspected locally on 2026-08-15 via source record; practitioner conference testimony for knowledge-app architecture, not independent ROI, security, model, permission, or deployment proof.
• WorldQuant, career index, LLMs and AI Agents intern, WQBRAIN AI Researcher, AI Software Developer, Senior AI Software Developer, and Lead Python Engineer, AI/ML Systems, accessed 2026-08-14; current firm-hosted LLM/AI hiring surfaces; no filled-role, deployment, live-book authority, or performance proof.
• Two Sigma, AI outlook Part I, Part II, and ACM CAIS 2026 preview, retrieved 2026-08-14; current AI titles, agentic-system controls, multimodal and leakage cautions, and conference-source discovery.
• Bridgewater, AIA Labs, Nina Lozinski, Rohan Alur, Jasjeet Sekhon, Daniel Kang publications, and AIA research index, retrieved 2026-08-14; lab remit, named personnel, academic lineage, and current research artifacts.
• Bridgewater, Nick Hamilton forecasting-competition video and capture note, June 10, 2026; official personnel and recruiting evidence naming Hamilton and Will Barnes, with forecasting and macro/policy context but no AI-system or performance disclosure.
• Bridgewater, Ray Dalio founder recording and title-blind/name-collision audit, October 12, 2020; third-party historical founder/investment-philosophy metadata, recovered automatic captions, and five disqualified or unresolved query matches; caption review adds no AI-system disclosure.
• Citadel/Citadel Securities, William Hamilton, Jianbo Chen, Honglin Yuan, Wei Tan, and Yifeng Tao, retrieved 2026-08-14; public researcher identities, academic training, and paper-topic signals; no individual-to-system assignment inferred.
• Two Sigma, Gene Li, Lei Chen, Kishor Jothimurugan, Qiwen Cui, Lingxue Zhu, and Myung Jin Choi, retrieved 2026-08-14; public academic lineage and ML/RL/graph/signal-processing research.
• BlackRock, AI Labs, Rachel Schutt, and Stephen Boyd papers, retrieved 2026-08-14; named lab, academic advisors, public finance/optimization research, and asset-manager comparator evidence.
• Millennium, AI Lab Q&A, Gideon Mann, and BloombergGPT, retrieved 2026-08-14; lab remit, personnel lineage, and prior finance-LLM publication.
• Balyasny, Charlie Flanagan, OpenAI case study, and Applied AI panel, retrieved 2026-08-14; executive remit, team-size statement, model evaluation, and applied-AI workflow.
• Millennium, AI Lab Q&A and Gideon Mann, retrieved 2026-08-14; AI Lab remit, CIO disclosure, and Global Head of AI title.
• Point72/Cubist, AI-chess hackathon, retrieved 2026-08-14; recruiting/engineering and agent-evaluation signal outside financial deployment.
• AQR, Machine Learning, current page; systematic ML posture.
• CFM, Fine-tuned financial NER case study, current public case study; narrow task-specific weight modification.
• Jane Street, Quantitative Researcher, Hong Kong, Quantitative Trader, Hong Kong, and Quantitative Trader Internship, Hong Kong, retrieved 2026-08-28; regional ML/trading hiring and SFC-regulated Hong Kong entity footer evidence, not GenAI deployment or investment-authority proof.
• AQR-affiliated research, Point-in-time language models, 2026; research evidence, not public AQR deployment evidence.
• G-Research, Core AI Engineer, current role; infrastructure and agent-runtime intent.
• Point72, AI/ML — Investment Services, ML Infrastructure Engineer, GenAI, and AI Engineer, L/S Equity, current job listings; training/serving and investment-team integration signals.
• Bridgewater, AIA Labs, current firm page; investment-process AI and human-machine collaboration.
• Schonfeld, FE AI Lab, current firm article; provider-backed workflow integration.
• State of AI local deep dive, Hedge-Fund AI Capability Audit, July 2026; capability taxonomy and peer context.
• State of AI local ledger, Hedge-Fund AI Worker Layer, July 2026; LinkedIn role evidence and temporal boundaries.
• State of AI local research, Hedge Fund Research Signal Landscape 2026, broader peer context.
• State of AI local findings, Hedge Fund Research Public-Signal Tape, broader public-signal discovery tape.
• Point72 / Cubist, Cubist Systematic, NLP/AI Engineer, Fundamental Equities AI Engineer, Cubist ML Researcher, and AI Data Scientist, current 2026; systematic ML, LLM, agent, fine-tuning, regional AI-data-product, and research-to-evaluation signals.
• Schonfeld, FE AI Lab, Hannah Jiang Q&A, Senior Software Engineer, and leadership page, 2026; SchonAI, adoption, research workflows, MCP, evaluation, and AI leadership.
• Schonfeld, Software Engineer - Fundamental Equities and Quantitative Developer - Fundamental Equities, current 2026; business-embedded Fundamental Equity AI-integration hiring signals.
• G-Research, Core AI Engineer, NLP Researcher, Applied AI Engineer, LLM code-review article, and Robocop, 2026; infrastructure, model-training intent, production LLM controls, and firm-owned code-review artifact.
• AQR, Systematic Equity, Wholesale Managed Futures PDS, and Scaling Point-in-Time Language Models, current/2026; ML/NLP disclosure and affiliated LM research boundaries.
• Two Sigma, AI in Investment Management Outlook I, Anything Can Be Language Now, Generative AI quantitative-software role, Techniques Engineering, Modeling Data Scientist, and Post-Training Research Scientist, 2026; feature research, LLM/NLP infrastructure, agentic workflows, and post-training hiring signals.
• Citadel, GQS Machine Learning Researcher, Global Quantitative Strategies, Reuters NEXT post, and investment-decision article, 2025–2026; assistant, validation-agent, ML, and human-judgment boundaries.
• Millennium, Technology and AI, Gideon Mann, AI Lab Q&A, Daniel Tymecki, Applied Cloud and AI Engineer, and privacy policy, 2026; AI leadership, agentic infrastructure, investment-team agent tools, internal search, and controls.
• D. E. Shaw, current careers and DESIM Portfolio Strategist, current 2026; investment-business and group-boundary checks.
• D. E. Shaw, Fundamental Equities - AI Product Analyst, current 2026; analyst-workflow and compliance-boundary hiring signal.
• PDT, work, careers, and Applied ML Scientist, current; predictive ML and automated-trading evidence without public GenAI linkage.
• PDT, Research Engineer, accessed 2026-08-12; alpha/signal/portfolio-construction research infrastructure, large-ML-model training/fine-tuning infrastructure, and real-time inference hiring signal; no LLM, filled-role, or return-attribution proof.
• XTX Markets, clients and careers, current; price-forecasting and market-making ML evidence.
• XTX Markets, AI Research Internship - XTY Labs, Machine Learning Performance Engineer, and official careers API at https://api.xtxcareers.com/jobs.json, accessed 2026-08-14; foundation-model/agent-prototype hiring and ML acceleration infrastructure signals.
• Aspect, J.P. Morgan podcast, ML challenge, and Bas Monsewije profile, 2022–2025/current; ML, evaluation, and AI/LLM practitioner signals.
• Aspect, AMAC registration press release, 2024-12-17, accessed 2026-08-14; Aspect Capital (China) Limited legal-entity and mainland-China private-fund-manager registration signal, not an AI deployment disclosure.
• Winton, working at Winton and leadership, current; systematic infrastructure and personnel.
• Systematica, strategies, culture, and management, current; systematic and Data R&D signals.
• Marshall Wace, about and teams, current; quantitative/data infrastructure signals.
• Brevan Howard, home and careers, current; macro and quantitative personnel surfaces; no promoted GenAI workflow source located.
• Caxton Associates, official site and careers, current; global-macro and personnel surfaces; no promoted GenAI workflow source located.
• Voleon, official site and management, current; ML-investment identity and management; no public LLM/agent architecture located.
• Voleon, Jiafan He homepage and Google Scholar profile, accessed 2026-08-12; personal/Scholar current language-model personnel signal with verified Voleon email; no firm-controlled LLM program or deployment proof.
• Squarepoint, experienced professionals, early careers, and 2026 MIFIDPRU disclosure, current; systematic ML and entity/governance evidence.
• Jump Trading, AI/ML, accessed 2026-08-12; LLM agents/API-HPC serving, custom foundation models, usage figures, simulations, and trader-feedback loop; no model inventory or AI-attributed returns.
• Hudson River Trading, Machine Learning & AI and AI Researcher, LLMs, accessed 2026-08-12; HRT AI Labs, deep-learning trading constraints, and current role-title evidence; no public LLM trading model or reproducible performance.
• Hudson River Trading, Sean Mann, Kevin Luo, Claire Donnat’s CV, University of Chicago profile, and linked SAMoSSA, ROTI-GCV, and No Free Prune papers, accessed 2026-08-28; personnel and academic-lineage evidence with current/former-role boundaries, not HRT model-ownership or performance proof.
• Optiver, Quant Research for Trading, accessed 2026-08-12; research lifecycle, AI Lab, LLM-based analysis, and production constraints; no model inventory or agent authority.
• DRW, AI Engineer and Data Developer, accessed 2026-08-12; AI/ML infrastructure, RAG, embeddings, fine-tuning, monitoring, and multi-asset systematic-team hiring; no filled-role or investment-deployment proof.
• Capstone Investment Advisors, AI Engineer, AI Infrastructure Engineer, and open positions board, accessed 2026-08-12; AI/LLM/RAG/MCP hiring and infrastructure signals; no filled-role, investment-authority, or return-attribution proof.
• IMC Trading, Software Engineer - AI Powered Engineering, accessed 2026-08-12; agentic developer-workflow engineering signal; no trading-system agent authority or AI-attributed P&L.
• Numerai, homepage, NumerCon 2026 recap, Faith release, MCP documentation, example-scripts repository, and AI Scientist role, accessed 2026-08-12; public AI-scientist workflow, Predictive LLM feature generation, open agent skills, MCP research/submission tools, and hiring intent; no autonomous firm-capital authority, independent evaluation, or AI-attributed return proof.
• QRT, Production Support Engineer - AI & LLM and Greenhouse job API, accessed 2026-08-13; production-support hiring language for front-office trading flows, LLM-based systems, monitoring, release management, algorithmic-trading platforms, and risk systems; no filled-role, model, permission, trading-authority, or performance proof.
• BlackRock / Aladdin, 2025 AI Engineer event schedule and LangChain Interrupt recording, accessed 2026-08-13; official conference abstract and recording metadata for “Agents in Investment Management: Aladdin Copilot”; no retrieved transcript, model inventory, user-permission map, investment-advice authority, or performance proof.
• High-Flyer / 幻方, homepage and fund page, accessed 2026-08-13; official Chinese-language AI compute, AI-used fund, basic AI research, technology naming, and mainland/Hong Kong entity evidence; no model weights, AI-company/fund operating-boundary map, permission map, or independent return attribution.
• Lingjun / 灵均, AI research-process article and Hong Kong/offshore article, accessed 2026-08-13; official Chinese-language research-agent, ML/deep-learning signal, AI risk-model tooling, multimodal-data, Hong Kong/offshore product, and supercomputing-center signals; no model inventory, deployment audit, trading permission, or AI-attributed return proof.
• Mingshi / 鸣石, homepage, accessed 2026-08-13; official Chinese-language private-fund manager and AI-factor/ML-weighting research-pipeline evidence; no GenAI/LLM disclosure, model inventory, live deployment proof, or independent performance attribution.
• DeepWin / 蝶威量化, official page, Chinese Securities Journal event report, and Chinese Securities Journal industry survey, accessed 2026-08-13; official Chinese-language private-fund identity, P1070101 registration statement, AI quantitative-investment/RL/Agent process language, and media-reported firm-event research-agent workflow; no directly retrieved AMAC record, model inventory, product/fund-vehicle boundary proof, permission map, independent deployment audit, or AI-attributed performance.
• Metabit Trading / Qianxiang / 乾象, official page, join page, and JuiceFS customer case, accessed 2026-08-27; official Chinese-language private-fund-manager registration statement, AI-driven quantitative-trading, RL/POMDP, HPC/GPU-cluster, DL/ML training/inference, data-engine, and risk-control language plus vendor-published cloud research-infrastructure workflow; no direct AMAC extract, model inventory, training-data license, production permission map, autonomous capital-allocation disclosure, independent performance audit, or AI-attributed return proof.
• WizardQuant / 宽德, careers, Machine Learning Platform Engineer, Machine Learning Performance Engineer, Strategy System Engineer, and AI-era indexing article, article dated 2025-06-22, accessed 2026-08-14; official Chinese and English hiring, infrastructure, and lab-positioning signals; no model inventory, live deployment proof, WILL/fund operating-boundary map, permission map, filled-role proof, regulated-product boundary proof, or AI-attributed returns.
• QTS Capital Management, homepage, Feature Selection in the Age of Generative AI, SEC Form D/A for QTS Partners, L.P., and QTS Capital Management LLC LEI record, accessed 2026-08-13; official methodology plus legal-entity separation between Delaware and British Columbia surfaces; no agent workflow, implementation audit, model weights, or AI-attributed performance proof.
• Dynamic Funds, Ahmad Golaraei profile, accessed 2026-08-13; official named AI & Quant Investment Analyst role and investment-research support signal; no model inventory, agent runtime, production authority, portfolio permission, or AI-attributed investment result.
• Connor, Clark & Lunn Investment Management, Quantitative Equity Data Science role, accessed 2026-08-15; official Canadian quant-equity data-science role naming ML/AI, NER, knowledge graphs, alpha-research support, and production deployments; no filled-role, model inventory, LLM/agent runtime, live trading authority, or AI-attributed performance proof.
• Longqi Scientific Investment / 龙旗, official page, accessed 2026-08-15; current firm-controlled mainland/Hong Kong entity, AI/ML, fund-family, and SFC Type 9 language with an unresolved CE-number discrepancy between the legal section and footer; no direct SFC-register extraction, model inventory, permission map, deployment audit, fund-level product document, or AI-attributed-return proof.
• Jasper Capital, Opalesque TV manager page, Opalesque manager article, Jasper Capital Hong Kong, Jasper Capital International, and Blackstone sub-adviser profile, accessed 2026-08-27; dated China-A-share strategy, Bo Huang personnel, entity, and factor-research/process evidence; no current AI/GenAI, model inventory, training-data, permission, or AI-attributed-return proof.
• Jane Street, Alok first-party profile and transcript, quantitative-research page, and video pointer, accessed 2026-08-27; first-party role and research-workflow evidence for an unnamed New York researcher; no surname, model inventory, GenAI system, production permission, or performance attribution.
• Jane Street, Machine Learning Performance Engineer, Hong Kong, accessed 2026-08-28; first-party Hong Kong ML-infrastructure hiring signal covering training/inference performance, CUDA/GPU-networking vocabulary, and distributed-training algorithms; no filled-role proof, model/provider identity, training data, production endpoint, autonomous order authority, or AI-attributed performance.
• Hudson River Trading, Code of Conduct video, Society of Women Engineers profile, PyLadiesCon speaker page, and public LinkedIn post, accessed 2026-08-28; historical Jessie Newman HRT algorithm-engineering, testing, recruiting, and community-leadership evidence; automatic captions add a timestamped role description, but no current employment, AI/GenAI, model, permission, or performance proof.
• Jane Street, first-party mock-interview page and Grace/Nolen recording, accessed 2026-08-27; dated recruiting and engineering-culture evidence for unnamed presenters Grace and Nolen; no surname, current team, AI/GenAI, model, permission, or performance proof.
• Numerai, Chai Time Data Science episode, iVoox episode page, and Numerai Signals announcement, accessed 2026-08-27; dated Richard Craib founder-media and platform-evolution evidence; no complete transcript, current architecture, permission, portfolio-authority, or independently audited-return proof.
• Numerai, Lex Fridman Podcast #159 publisher page, direct recording, and Numerai documentation, accessed 2026-08-27; dated Richard Craib founder interview and canonical source for two derivative clips; no current architecture, permission, portfolio-authority, or independently audited-return proof.
• Numerai, Other Life Episode 162, Apple listing, and YouTube recording, accessed 2026-08-27; dated founder interview and company-attributed performance framing; no current architecture, permission, portfolio-authority, or independently audited-return proof.
• Numerai, Office Hours S01E13 recording and Keno Leon public profile, accessed 2026-08-27; first-party community-participant workflow evidence with recovered timestamps; no staff, internal-operation, portfolio-authority, or independently audited-performance proof.
• Numerai, Office Hours S01E03 recording, description metadata, and Michael Oliver forum introduction, accessed 2026-08-27; dated participant-to-Data Scientist transition and historical tournament/hedge-fund interface role; no current staffing, current implementation, portfolio-authority, or independently audited-performance proof.
• Avangard Investments, performance, FAQ, governance, and our people, accessed 2026-08-15; official Australian fund, entity, named-system, and firm-reported data-scale evidence for A.L.F.R.E.D.; no independent ASIC extract, model design, live GenAI-agent, permission map, independent performance, or AI-attributed return proof.
• Antipodes, careers, job posted 2026-06-01, applications close 2026-06-27, accessed 2026-08-14; official Australian Investment Analyst - AI & Alternative Data hiring signal covering PM/analyst agent, skill, and MCP catalogues, systematic-research and portfolio-integration work, AgentOps, evals, observability, audit trails, and Anthropic/OpenAI SDKs; no filled-role proof, production catalogue proof, model inventory, permission map, investment authority, or AI-attributed returns.
• RQI Investors, long-short fund launch and AI paper page, accessed 2026-08-15; official Australian active-quant equities release and AI-use discussion; no model identity, production-stage map, permission map, audited signal contribution, or AI-attributed-return proof.
• Macquarie QIS, Quantitative Investment Strategies and machine-learning QIS index case, accessed 2026-08-13; official bank/QIS methodology disclosure and 2021 reinforcement-learning case; not a hedge-fund-manager peer and no current GenAI-agent, permission, independent-performance, or AI-attributed fund-return proof.
• IQuest/Ubiquant, UbiquantAI’s Hugging Face organization, Fleming-R1-7B, Fleming-R1-32B, Fleming-VL-8B, Fleming-VL-38B, Universal Reasoning Model paper, IQuestLab organization, IQuest-Coder model page, IQuest-Coder repository, and public researcher page, accessed 2026-08-18; public model, paper, and code signals plus qualified personal-page affiliation to IQuest Research/UbiQuant. The Fleming cards reviewed here state medical reasoning and multimodal medical use, not finance. No official Ubiquant fund disclosure, regulator record, finance-specific model, investment workflow, benchmark reliance, or performance attribution was found.
• RAM Active Investments, research page, company page, and EMNLP 2024 Industry Track paper, accessed 2026-08-15; official Swiss systematic-manager research and public paper evidence for LLM/newsflow and deep-learning alpha-input work; no live portfolio permissions, model weights, production deployment stage, data-rights controls, independent performance audit, or AI-attributed returns.
• T. Rowe Price Integrated Equity, How AI can open new avenues for systematic investment research, July 2026, accessed 2026-08-15; official systematic-research workflow note covering LLM-assisted scoring, controls, batch scoring, software-resilience analysis, and bias/overfitting caveats; no model vendors or weights, portfolio permissions, production adoption, audited signal contribution, investment recommendation authority, or AI-attributed returns.
• Mackenzie Investments, 2026 Market Outlook and How Arup Datta uses AI to help track volatile emerging markets, accessed 2026-08-16; official Canadian quant-equity materials covering ML, NLP, LLM, cloud-computing, human-overlay, earnings-call sentiment, alternative-data, and native-language financial-statement signals; no model inventory, permission map, autonomous-trading authority, independent performance audit, or AI-attributed returns.
• Pictet Asset Management, Quest AI evolving model, Innovation with intent, and Quest AI practitioner page, accessed 2026-08-16; official systematic-manager materials covering AI stock selection, forecast attribution, human oversight, retraining, data governance, boosted trees, and IBES data; no model weights, full feature list, permission map, independent performance audit, or AI-attributed returns.

BlackRock’s Aladdin Wealth Auto Commentary case account adds a distinct product-workflow signal. The page names Ted Stratigos as Global Head of Aladdin Wealth and describes a feature that uses hundreds of data points from Aladdin Wealth analytics, CIO outlook material, and client holdings and preferences. It says the prototype was developed in secure production environments with real data, controls, and advisor workflows, with risk, legal, information-security, and model-governance guardrails; the output is a first draft that an advisor reviews, edits, and approves. This is a company-reported wealth-platform workflow, not evidence about BlackRock Systematic, AI Labs, autonomous investment decisions, or hedge-fund trading. The page does not disclose the underlying model, training or retrieval data, evaluation denominator, client permissions, adoption, or investment performance.

BlackRock’s 2026 Chairman’s Letter supplies a separate parent-company strategy signal. Larry Fink’s letter places Aladdin at the center of BlackRock’s technology strategy, describes the integration of Preqin data into Aladdin’s public/private workflow, and frames AI as a potential efficiency layer across data-heavy investment and analysis processes. That is a shareholder-level statement about platform direction. It does not identify a model, a research team, a dataset-rights regime, user permissions, evaluation results, autonomous authority, or AI-attributed investment performance. The parent-company statement should therefore remain separate from the firm-controlled Systematic, Aladdin Copilot, Asimov, and AI Labs evidence layers.

The Tech Talks Daily episode with Man Group CTO Gary Collier, published February 26, 2025, has now been reconciled to its public Libsyn enclosure and timestamped local ASR. In the recording, Collier discusses historical nearest-neighbour, recurrent-neural-network, NLP, and reinforcement-learning work; a speaker-reported approximate quarter-share of machine-learning signal models at Man Numeric (06:31–07:02); ManGPT as an internal interface to multiple LLMs (07:58–08:16); an open-source-first Python platform and firmwide data/AI workflow (10:25–12:33); and Alpha Assistant/AlphaGPT, automated backtests, human review, and explainability (26:41–29:05). These are dated CTO statements, not an audited model inventory, user denominator, permission map, autonomous-trading claim, or performance record. See the podcast ASR recovery note.

The same title-blind Apple expansion recovered Ed Cole’s June 29, 2026 episode and its public Omny enclosure. Cole, identified by the publisher as Head of Multi-Strategy Equities within Solutions at Man Group, discusses the possibility of using open-source models on private infrastructure (08:46–09:16), smaller language models for discrete tasks (11:49–12:23), and model-cost/compute tradeoffs (12:24–13:18). The interview is a named practitioner’s market and enterprise-AI perspective; it does not establish that Man’s multi-strategy team uses a particular model, agent, vendor, or deployment permission.

The Big View episode with Gary Collier, published March 25, 2025, is also now audio-recovered rather than metadata-only. Collier describes a firm interface to multiple language models and monthly use by more than half the firm (12:10–13:23), Alpha Assistant and AlphaGPT (15:17–16:29), a possible document-to-idea-to-backtest workflow (16:41–18:08), and data lineage as part of explainability (27:31–28:37). This is a separate third-party interview and should be read as attributed practitioner evidence alongside Man’s first-party pages, not as proof that all named projects are one system or that any agent has capital authority. The recovery note records hashes, ASR method, and limits.

Three additional AQR routes are now indexed separately. A company LinkedIn post about the 2026 London Symposium names machine learning and recent innovations at AQR and identifies John Liew as a special-session participant; it is event and speaker metadata, not a transcript. A CFA Society Luxembourg event page lists Peter Hecht, Managing Director and Head of AQR’s North America Portfolio Solutions Group, for a September 29, 2026 session on machine learning, market timing, stock selection, complexity, and validation; the session is upcoming and no presentation was retrieved. Finally, the May 23, 2026 Excess Returns episode with Cliff Asness is now covered by a recovered public enclosure and private local ASR. The timestamped audio adds Asness’s distinction between a single AI market thesis and diversified quantitative positioning (00:21:40–00:23:08) and his discussion of economic intuition, overfitting, data mining, underfitting, and complex models (01:11:54–01:15:52). It remains a dated founder methodology account and does not establish a new AQR model, agent, permission map, deployment stage, or AI-attributed performance record. The capture note records the primary-audio hashes and boundaries.

GMO’s company LinkedIn post on the AI investment boom adds a distinct social route to its Sink or Swim research. The post frames rapid AI-infrastructure investment, historical investment-boom behavior, developed-market comparisons, and the interaction of investment and sentiment as questions for future equity returns. The underlying research and the social post are investment commentary; neither establishes a GMO internal GenAI platform, model, data permission, or production workflow.

An additional AQR-affiliated research route is the NBER Working Paper on Artificial Intelligence Asset Pricing Models, revised in June 2026. Bryan Kelly, Boris Kuznetsov, Semyon Malamud, and Teng Andrea Xu describe a transformer embedded in the stochastic discount factor, using cross-asset information sharing and nonlinear structure, alongside a linear-transformer surrogate for interpretation. The NBER-SAIF 2026 program places the paper in a session on AI and financial decisions and lists Xu with AQR Capital Management. The NBER paper’s disclosures are narrower than an AQR deployment claim: they identify Kelly at Yale, AQR, and NBER, describe Malamud as an AQR consultant, and state that AQR may or may not apply similar techniques and that the views are not necessarily AQR’s. The paper reports research-model pricing-error results; it does not disclose an AQR production model, portfolio permission, data-rights regime, live deployment, or AI-attributed performance. This route therefore strengthens the public research lineage around finance-specific transformer modeling while preserving the boundary between affiliated scholarship and firm operations.

Two older Brevan Howard partnership routes materially expand the firm’s public AI history. A 2019 Nomura announcement says a Brevan Howard-established venture, AIM2, used data science and machine learning to develop alpha strategies and was collaborating with Nomura on wholesale-market tools using historical and real-time high-frequency data, including client flows and market data, to provide quotes and recommendations. A 2021 Business Wire release says Brevan Howard and DRW entered a strategic partnership with Edge Focus, whose platform used proprietary ML and AI for credit underwriting, pricing, default forecasting, and institutional portfolio construction; the release names Aron Landy and Adam Garner in the partner statements. These are historical partner and technology-company disclosures, not evidence that either relationship remains active, that AIM2 or Edge Focus systems run inside current Brevan portfolios, or that any AI system produced attributed returns. They also sit alongside Brevan’s current supplier term restricting use of firm-provided information for model training or improvement, which is a data-governance signal rather than a model inventory.

Brevan Howard’s current About Us page provides a separate firm-surface record: it describes global macro and digital-assets activity, technology investment, quantitative and qualitative risk management, and a continuously updated stress-test bank. Its Careers page lists quantitative analysts among its public role families. The 2025 UK Gender Pay Report says the firm expanded early-career programmes into new geographies and quantitative, technology, and AI roles. Separately, its supplier terms restrict suppliers from using Brevan Howard-provided information or outputs to train, tune, test, or improve AI/ML systems, including through third parties and after anonymisation. The terms are a contractual data-governance signal, not evidence of an internal model stack. None of these sources identifies a current AI lab, model inventory, production endpoint, agent permissions, autonomous trading authority, or AI-attributed performance. See the capture note.

Voleon’s public seminar network adds a distinct academic-research route. The UC Berkeley CLIMB page labels a March 20, 2025 talk as part of the Voleon Seminar Series and identifies Manish Raghavan as the MIT Sloan/EECS speaker. The talk concerns information differences between humans and AI, homogeneity when people use the same AI, and applications in clinical decision-making and content production; the biography records a prior postdoctoral fellowship at Harvard’s Center for Research on Computation and Society. This is evidence of a Voleon-associated university seminar and an MIT/Harvard research-network route. It does not establish that Voleon uses Raghavan’s work, operates a human-AI system based on it, or has disclosed a related model, dataset, permission boundary, or performance result.

Winton’s first-party AFI podcast page adds a previously unindexed title-blind media route: the December 3, 2024 interview with CIO Simon Judes lists research in medium-frequency macro and equity strategies, systematic commodities, systematic credit, specialized rates, options, recruitment, and trading China. The page links the episode through Acast, Spotify, and Apple Podcasts, and publishes chapter metadata rather than a transcript. This expands the Winton media and regional-strategy trail; it does not establish a new AI system, model, data-rights regime, deployment permission, or performance attribution. The previously unresolved Acast audio is now recovered and locally transcribed. The recording adds dated operating detail: Judes describes research resources directed toward medium-frequency macro and equity strategies, systematic fundamental commodities, systematic credit, specialized rates, and options ([12:49–14:21], local audio); he describes systematic credit as a broad category in which relevant information can be hidden in bond covenants and other text ([15:11–15:41], local audio); and he describes a collaborative, non-siloed structure requiring data, IT infrastructure, statistics, and market knowledge ([15:55–16:40], local audio). The interview also discusses London talent and a Shanghai-centered regional footprint. These are speaker statements from a 2024 interview, not evidence of an NLP model, AI lab, LLM workflow, current headcount, production permission, or AI-attributed return. The Winton source note records the recovery hash and automatic-transcription boundary.

Winton’s first-party systematic-equities article adds an older but concrete NLP route. It describes learning topic mixtures from Section 1 of US 10-K filings and using those representations to classify companies with more nuance than a single sector label. The page says the model is trained on a broader company universe than the displayed S&P 500 example. This is historical methodology evidence, not evidence of a current Winton model, data license, execution linkage, or performance; the capture note preserves that boundary.

The first-party BNP Paribas Asset Management Talking Heads episode, published May 14, 2026, identifies Ram Rasaratnam as CIO for Quant Equity Strategies at AXA IM Core and Chris Iggo as AXA IM Core’s CIO. Rasaratnam says the team put a neural network into production approximately nine years earlier to forecast the probability of a stock’s short-term volatility rising over the following month; he gives a speaker-reported example involving Silicon Valley Bank and Signature Bank before the Q1 2023 regional-banking crisis. The publisher transcript also describes machine-reading earnings-call transcripts for a quality signal, an internally hosted open-source coding assistant with access to model code, and repository writes that were still being tested. Rasaratnam says current language models are not capable of model design and retains human control over that boundary. These are first-party speaker statements, not an independent model audit or return attribution. The episode is an asset-manager comparator rather than evidence about a tracked hedge fund. See the capture note.

A separate Australian The Greener Way episode, published March 31, 2026, gives a second public route to Ram Rasaratnam’s AXA IM / BNP Paribas quant-equity account. The recovered Captivate audio supports timestamped navigation: he places AXA IM’s first neural network at roughly nine years earlier and its NLP capability at roughly ten years earlier, describing text, sentiment, and tone extraction as inputs to investment signals (06:20–06:49). He also describes starting with a fundamental portfolio and overlaying ESG decisions so the effects can be separated and discussed (04:53–05:14). This corroborates the existing AXA route while adding a clear “machine learning toolbox” and transparency framing; it remains speaker-reported and does not disclose model weights, training data, providers, permissions, production telemetry, autonomous order authority, or independently audited performance. See the capture note.

The title-blind archive pass found a distinct Fidelity International governance episode with Sue Lyn Stubbs, Associate Director in Sustainable Investing, published July 8, 2026. The publisher says Fidelity engaged 31 ASX-listed companies across five domains: strategy and value creation, board oversight and skills, risks and controls, ethical AI governance, and workforce impact. Stubbs describes adding a “Stage Zero” category for companies not yet actively implementing AI, and the recovered audio adds public governance questions around executive/board narrative divergence, explicit no-go areas, the limits of a single human-in-the-loop check, and unauthorized employee “BYO AI” tools as possible IP or compliance risks (04:35–06:15; 07:45–12:20; 18:40–21:30). The accompanying Fidelity paper reports the sample distribution as Stage 0 13%, Stage 1 26%, Stage 2 9%, Stage 3 13%, Stage 4 16%, and Stage 5 23%. Those percentages are Fidelity’s assessment of the 31 surveyed companies against Microsoft’s framework; the companies are not identified and the sample is not presented as representative of the broader ASX. This is a public active-manager governance-research route, not evidence about Fidelity’s own portfolio models, providers, permissions, or performance.

The same archive adds an adjacent ISS STOXX Sustainability episode with Josh Gilbert, Head of Geospatial Strategy, published June 9, 2026. The publisher describes satellite imagery and sensor data being translated from reporting inputs into climate- and nature-risk information, with a chapter on “AI, numeric models, and financial integration” at 12:32. This expands the alternative-data modality map toward geospatial signals while leaving the model, data contract, customer, portfolio linkage, and deployment status unspecified. See the capture note.

A third The Greener Way route, published August 12, 2026, identifies Mans Carlsson as Head of ESG and Co-Portfolio Manager at Ausbil Investment Management. The publisher describes proprietary ESG research, company engagement, and on-the-ground investigation; its chapter list explicitly includes “Can AI replace ESG research?” at 04:44 and “Decarbonisation, AI and energy demand growth” at 12:28. This adds a named stewardship and ESG-process route, not a disclosure of an Ausbil AI model, provider, training data, permissions, production endpoint, or performance. See the capture note.

The July 27, 2026 Bloomberg Radio interview transcript with State Street Investment Management global CIO Lori Heinel adds another adjacent-manager control case. Heinel says State Street has used machine learning and natural-language processing in active strategies for more than a decade, while building governance, cybersecurity, privacy, and deployment infrastructure around a regulated institution (42:49–44:05). She describes early GenAI use in repeatable operational work such as RFPs, commentary, and client servicing, with a human owning the final output, and an experimental research copilot that lets portfolio managers survey dozens or hundreds of reports before pressure-testing the result and deciding what to do (43:27–44:55). These are interview statements reproduced by the publisher, not an independent audit of State Street’s systems or results. The route is not evidence about a tracked hedge fund, model inventory, agent permissions, or autonomous trading. See the capture note.

An April 8, 2026 ETF.com transcript with Pictet Head of Quantitative Investments David Wright adds a distinct media route to the existing Pictet Quest AI record. Wright distinguishes the disclosed strategy from an LLM and describes thousands of decision trees built with LightGBM-style technology, roughly 15 years of feature history, quarterly retraining, a 20-day residual-return forecast, daily scoring, weekly rebalancing, and risk/turnover constraints. He describes interpretability tooling for feature pathways and states that the automated strategy is not autonomous: portfolio managers define guardrails and can remove stale trades after material news (transcript sections “Bringing Machine Learning to Quant Investing,” “Digging Into the Methodology,” and “Building Human Constraints Around the Signals”). The page includes speaker-reported asset-scale and risk examples but no independent audit. This is an asset-manager comparator, not evidence about a tracked hedge fund, complete model inventory, data rights, or AI-attributed performance. See the capture note.

The UBP podcast with Campbell Managing Partner Joseph Kelly, published July 29, 2026, was present in the discovery registry but had not been promoted into this article’s firm ledger. UBP describes Campbell as a systematic multi-strategy manager; the page says Campbell began using AI tools in 2022 for research efficiency and code completion and later moved toward agentic coding tools, including Claude Code. The discussion is useful as a named current-executive workflow account and as a prompt for checking crowding and process-compression risks. It does not disclose Campbell’s model inventory, training data, data rights, evaluation fixtures, agent permissions, production deployment, or AI-attributed performance. See the title-blind discovery note.

The LSEG Post-Trade Solutions episode “Agentic AI in Quant Risk”, published April 29, 2026, adds a title-blind vendor route with a public transcript. Xabier Anduaga, Stuart Smith, and Joey O’Brien discuss agent-assisted model development, market-data analysis, backtesting, regulatory calculations, dynamic stress testing, and risk-system APIs. Their examples include synthetic missing-data or volatility-surface generation, test-backed implementation of prescribed regulatory rules, current-news-driven stress scenarios, and extending an open-source risk library; they also describe routine risk-data cleansing as a candidate for automation while retaining review of generated work (01:26–07:47; 07:47–14:40; 15:35–18:44). This is evidence about a vendor’s narrated workflow and research agenda, not about a hedge fund’s implementation, model inventory, data rights, production permissions, autonomous trading authority, or performance. See the capture note.

The title-blind pass also recovered a first-party BedRock Partners podcast archive, an adjacent private investment-manager route with a China/Hong Kong/global-investing footprint. Its founder profile identifies Cong Tan as CIO and founder and describes prior research and investment roles at Infineon, Morgan Stanley Asset Management (China), China Universal Asset Management, and HSAM, alongside Stanford GSB and Fudan training. The archive’s E25, E26, E27, E28, E29, and E30 now have an alternate public-audio recovery route through Xiaoyuzhou. Local Chinese ASR supplies navigation timestamps for E25–E30 where YouTube captions were unavailable. This improves discoverability but does not elevate the evidence to a verbatim transcript or model specification. See the capture note.

The newly recovered E19 transcript adds a named practitioner route. Bill Sun Qingyun describes Stanford mathematics training, early Google Brain work, and historical roles at Millennium, Citadel, and Point72; those affiliations remain guest-reported. The conversation focuses on methodology-specific AI assistants for repeatable, back-testable tasks, retrieval, private-data/tool integration, and Bayesian updating, while preserving human judgment for unprecedented long-horizon questions. A separate Beacon Pin AI profile, Chinese-language Bilibili repost, and Indigo Talk regional podcast listing expand the person and regional-media route into Gen Alpha, AIUSD, and self-evolving-agent themes. The repost and regional coverage are discovery/corroboration surfaces, not independent proof of product safety, regulatory status, live execution, or employer deployment. See the cross-surface capture note.

August 27 Johns Hopkins personnel and conference routes

The official Johns Hopkins AMS 2026 agenda lists Michael Baeder of Campbell and Company and Chaoyu Liu of Singular Square Management as keynote speakers on April 7, 2026. The organizer’s speaker page describes Baeder as working on global cash equities with economic theory, data analysis, and machine learning; it also says he chairs a biweekly research seminar, participates in an AI working group, and peer-reviews internal research. The page describes Liu as Singular Square’s founder and chief quantitative strategist and says he is building a platform using AI and alternative data around market microstructure and real-time information processing. These are organizer-published biographies, not independent audits of either firm’s technology or investment process.

The public personnel record needs title discipline. Johns Hopkins calls Baeder a director, while The Org lists Principal Researcher and describes global market-neutral equity model development; LinkedIn records Campbell and a Johns Hopkins thesis involving manifold embeddings in empirical asset pricing. Campbell’s first-party site separately describes proprietary data, high-speed infrastructure, contextual risk management, and a combination of human domain knowledge, statistical tools, and infrastructure. Together these sources establish a public research-personnel and operating-language route. They do not establish a current title without conflict, a specific AI system, a training corpus, production permissions, autonomous trading authority, or performance attribution. See the capture note.

The Jacobs Levy Center 2026 conference agenda now exposes a fuller professor–manager route for September 25, 2026. It lists “Machine Learning Meets Markowitz,” presented by Duke Professor Campbell Harvey and discussed by Wharton Associate Professor Winston Wei Dou; “Demand Propagation Through Traded Risk Factors,” presented by Wharton Assistant Professor Amy Wang Huber and discussed by State Street’s Jennifer Bender; and a fireside chat with AQR founder, Managing Principal, and CIO Cliff Asness. Other sessions cover achievable alpha and recession prediction markets. The page links research papers but does not provide a recording, model specifications, data rights, or evidence that the participants share a model, platform, or investment deployment. This route is tracked as conference and research-network evidence, not as firm implementation evidence.

The title-blind Bloomberg Odd Lots / Joe Peta episode, published September 5, 2024, identifies Peta as a former Point72 Head of Performance Analytics and discusses portfolio-manager evaluation, skill attributes, and noise in realized results. The capture note records the timestamped caption evidence and keeps this personnel/process route separate from Point72’s current AI hiring and infrastructure signals.

The Milken Institute Public Market Alpha panel, held May 5, 2026, adds a multi-firm conference route that was not previously in this article’s registry. The public transcript PDF identifies Eddie Fishman of D. E. Shaw, Philip Seager of CFM, and Joanna Welsh of Citadel. Seager discusses the relationship between expanding data, compute, and research directions (10:29–13:24), then describes statistical risk controls, cleaned correlation estimates, volatility forecasts, and an additional layer for directions that may sell off (13:47–17:05). Welsh describes a multi-strategy operating ecosystem that includes research, capital allocation, portfolio construction, risk management, and central technology (35:56–38:07). She also mentions Claude Code best practices as a personal reference point while discussing judgment around numerical outputs (22:48–26:11); the transcript does not identify a Citadel deployment. A later D. E. Shaw Group public follow-up identifies Fishman as an Executive Committee member, links the panel, and repeats a broad statement that the group has used machine learning in quantitative research and trading. These are event-date and firm-post evidence, not a model inventory, production audit, permission map, or performance attribution. See the capture note.

The Bodhi Research Group 2026 Annual Symposium provides a separate Canadian conference and personnel route. Its May 14, 2026 agenda lists a 10:40 session, “Reading the Market at Scale: AI in Quant Equity,” and identifies Trevor Mottl as “Portfolio Manager, AI Quant Equity, Magnetar” and a UC Berkeley lecturer. The organizer describes the session as examining LLM-based reading of company narratives across thousands of names, including design principles and practical limitations. No recording, slide deck, model description, or Magnetar-controlled biography was recovered, so this is retained as dated event metadata and a research-topic lead rather than evidence of a current firmwide system or investment result. See the capture note.

Minotaur’s first-party June 2026 monthly report adds a dated update to its already captured podcast and interview record. The report says Taurient exceeds 600,000 lines of code and describes 2,438 June commits, about 195 company directories scaffolded, and 96 completed initiations. It names four agent additions—Spectra for telecom, Vesta for regulated and concession infrastructure, Tempo for market structure and liquidity, and Vigil for portfolio-risk analysis—and reports 22 investment agents plus five operations agents. Minotaur describes cross-agent handoffs and a Vigil review that surfaced concentration and common-driver risk in the portfolio. These counts and workflow descriptions are firm-reported, not independently audited. The April report names Talos as a specialist AI-infrastructure and semiconductor analyst and says the system now reaches monitoring, modelling, thesis validation, transcript analysis, scenario work, and risk review, while humans retain trading decisions and outputs cite primary sources. The December 2024 quarterly report reports a model-agnostic layer integrating more than 20 models from 10 providers and approximately 10,000–20,000 daily API calls to providers including OpenAI and Anthropic. Finally, the public Thomas Rice engineering repository identifies him as a Minotaur co-founder and documents reusable agent briefings, parallel worktrees, and isolated development services; it is not the Taurient codebase. None of these sources discloses the complete model roster, prompts, training data, data licenses, agent logs, permissions, or independently attributed investment performance. See the temporal capture note.

August 27 adjacent investment-firm AI leadership routes

B Capital’s appointment announcement names Dr. Andrew Jackson as General Partner and Chief AI Officer and assigns him an AI strategy remit spanning investment, portfolio management, operations, AI tools, and AI-driven products. The firm describes his prior G42 and Inception roles, enterprise-deployment experience, responsible-AI work, and machine-learning PhD from Trinity College Dublin. A separate Axios account describes B Capital’s internal Bee Hive system as preparing investment-committee material, taking notes, structuring arguments around potential deals, and tracking passed decisions for later review. B Capital is a venture-capital firm, not a tracked hedge-fund manager; these sources establish named leadership and a firm-reported internal investment workflow, not model versions, data rights, evaluation results, final investment authority, or attributed performance. See the capture note.

Generali Investments’ announcement appoints Ole Jorgensen Chief AI Officer and gives him a coordination remit across the multi-affiliate investment platform, including AI strategy, productivity work, risk management, and governance. It identifies his prior twelve-year tenure as Director of Research at Global Evolution Asset Management and says he led AI integration into that affiliate’s investment process; it also records machine-learning and economics training plus two MIT strategic-AI programs. This is a current personnel and organizational signal with a systematic-macro affiliate lineage, not a model inventory, production map, permission record, or AI-attributed performance. The announcement’s platform-level asset figure is kept separate from Global Evolution’s strategy scope. The capture note preserves the entity boundary.

August 27 Strand Global Macro title-blind manager route

The Strand Global Macro interview with Daniel Gladiš, CIO of Vltava Fund, was found through an archive search that did not require AI or hedge-fund terms in the episode title. The publisher synopsis says the August 20, 2026 discussion covers attention versus information and AI’s possible effect on informational inefficiency and intellectual conformity. The embedded YouTube recording was subsequently recovered with automatic captions. Gladiš says Vltava uses AI as an accelerator for investigating companies, reading and comparing material, challenging assumptions, and finding missed possibilities (27:30–28:12); he places reliability judgments, assumptions, valuation, conviction, and responsibility with people (28:14–28:46). He also frames a possible tension between cheaper information processing and greater intellectual conformity (22:25–27:19). This supports a dated, speaker-reported research-assistance and judgment-boundary account, not a model inventory, data contract, production endpoint, portfolio permission, or investment-performance claim. See the capture note.

The same Strand archive exposes 17 dated manager and economist episodes across Vltava Fund, Quantica, Quantedge, Lombard Odier, Aberdeen Investments, Maat Investment Group, Crossbridge Capital, AXA Group, Iguana Investments, Frazis Capital Partners, Eurizon SLJ, Syz Group, JM Finn, Sionna Investment Management, Kernow Asset Management, Umbra Capital, and Oldfield Partners. The archive’s synopses explicitly mention AI in the Vltava, Aberdeen, AXA, and Frazis routes; the other episodes are valuable title-blind investment-process and personnel routes. The full enumeration, media IDs, and capture states are in the archive note. YouTube automatic-caption layers have now been recovered for the full 17-episode set. Diggle gives a speaker-reported account of AI use in Aberdeen’s own research process (08:36–08:48); Frazis discusses AI in biotech while emphasizing first-in-human and regulatory bottlenecks (40:03–41:32). Storno describes systematic model signals with a human decision point (09:28–10:00); Reid describes quantitative analysis after an idea is identified and the use of multiple models (13:43–15:00). These are bounded speaker-reported process or market-context observations, not proof of AI deployment or performance. See the AI caption-recovery note and systematic-archive recovery note.

The final five previously unresolved Strand recordings add useful negative controls and one new internal-workflow lead. Crossbridge CIO Manish Singh discusses AI investing, potential job displacement, and productivity effects (11:59–12:45; 43:16–43:39), while Sionna CIO Kim Shannon discusses AI-company concentration, debt, and valuation (02:06–02:11; 05:43–06:40); neither passage discloses an internal AI system. Kernow CIO Alyx Wood contrasts quant funds’ short-horizon semantic/relative-value activity with Kernow’s longer-horizon long/short process (44:12–44:34), again without describing an AI deployment. Oldfield Partners CIO Sam Ziff says AI and software-company effects are difficult to model and are handled largely through qualitative analysis (06:35–07:00), and describes AI as increasing information-processing efficiency while leaving investor judgment important (26:22–28:25). Umbra CIO Marcus Szemruk provides the only firm-specific implementation lead in this batch: he says Umbra has started trialling Anthropic Claude to consolidate internally aggregated data used in asset-allocation decision-making and client/prospect literature (40:20–41:05). That is a dated, speaker-reported trial—not evidence of the data sources, tenancy, retention controls, evaluation, production status, portfolio authority, or performance. See the caption-recovery note.

August 27 Atreides Management public media map

The title-blind expansion also resolved a previously absent Atreides Management media cluster around Gavin Baker, whom publisher pages identify as the firm’s managing partner and CIO. Capital Allocators episode 489 describes public, private, and crossover strategies focused on technology and consumer companies, and reports $7 billion under oversight; that figure is retained as a publisher claim, not independently reconciled AUM. The episode’s publisher description emphasizes fundamental understanding, hypothesis-driven research, debate, culture, execution, and risk management, while also disclosing the host’s LP and advisory relationship with Atreides. The full transcript is behind the publisher login.

The a16z Runtime conversation and Colossus episode 451 expose Baker’s public AI-investing discussion across infrastructure, frontier models, scaling laws, token economics, verification, enterprise adoption, and the application layer. The May 20, 2026 Colossus episode is particularly important as a discovery lead because its official show notes include a distinct 56:13 “How Gavin Uses AI in Atreides” segment. The episode-specific YouTube recording now supplies public automatic captions. The recovered window around 54:58–57:49 discusses narrow-domain proprietary data, model-training economics, frontier/open-source model dynamics, and distillation; 58:35–59:18 discusses cybersecurity and synthetic impersonation risk. The captioned material does not verify a specific Atreides tool, dataset, permission boundary, evaluation fixture, or production endpoint, so the show-notes label remains a recovery lead rather than evidence of an internal workflow. See the capture note.

The July 8 Generating Alpha episode and August 4 Colossus episode add later distribution and topic routes. The August episode’s official show notes and recovered YouTube automatic captions provide separate navigation for model routing (00:26:38), continual learning (00:23:55), and token spending as a share of compensation (00:30:51). A 12:06–12:33 Rogo/Felix advertisement is explicitly excluded from Atreides evidence. These are public executive-media and research-context signals; they do not establish Atreides model versions, training data, vendors, budget, evaluation fixtures, agent permissions, production endpoints, or AI-attributed performance. See the full Atreides media note.

August 27 Risk.net Quantcast: Imperial’s market-ML and industry-collaboration route

The previously unpromoted Risk.net Quantcast Master’s Series episode with Jack Jacquier is now backed by the public SoundCloud recording and local timestamped sidecars. Risk.net’s December 2025 page describes a shift toward market microstructure and machine learning. In the recording, Jacquier says Imperial’s programme added market microstructure and reinforcement learning, introduced generative modelling, and had previously added quantum machine learning in finance (04:14–05:05). He describes the design process as a response to industry conversations and an advisory board, while retaining mathematical foundations and practical group projects (20:31–22:38). These are programme-design statements, not evidence of any fund’s production system.

The episode also exposes a talent and research-network route. Jacquier describes industry-funded doctoral work involving QRT, HSBC, Deutsche Bank, BNP Paribas, Citigroup, and CFM-related quantitative-finance institutes (10:24–11:21). He describes a QRT collaboration around identifying the Sharpe ratio, funding a PhD student, and supervising the work with Johannes Muhle-Karbe (26:45–27:15). Imperial’s first-party profile identifies Jacquier as Professor of Mathematics and Director of the MSc Mathematics and Finance, while the research page lists market microstructure and learning algorithms among the institute’s research topics. The industrial-collaborations page provides a separate institutional record of collaborations, but does not establish that every listed relationship remains current or that any project entered a live investment strategy.

The source therefore adds a structured academic-to-industry pathway around ML, generative modelling, reinforcement learning, quantum ML, internships, practitioner teaching, and funded doctoral research. It does not establish a model inventory, training corpus, data license, named hedge-fund deployment, agent permissions, portfolio authority, or AI-attributed performance. See the capture note.

August 27 Shanghai regional-language quantitative-AI conference routes

The Shanghai Advanced Institute of Finance event notice schedules an “AI, Compliance, Evolution” quantitative-investing talent forum for September 4, 2026. Its program names Kan Rui of SAIF, Xu Hao of Zhongtai Securities’ technology R&D department, and Zeng Lingqi of Mingyue Fund. The notice assigns Zeng a session on AI applications in quantitative trading and describes his prior work on medium- and low-frequency factor research, followed by an organizer-published account of an AI-large-model-assisted factor-discovery approach. It describes Xu’s responsibility for Zhongtai’s XTP quantitative-trading platform, which is presented as a service for private managers and other professional investors. This is first-party event and biography evidence; no recording, slides, code, model version, data-rights statement, or independent test was available. See the capture note.

The Cailian Press report on a July 19, 2026 Lujiazui Financial Salon roundtable adds a second Chinese-language route. It names participants from Beijing Xinhong Tianhe Asset Management, Hangzhou Higgs Investment, the University of Science and Technology of China, DataGrand, and Huawei’s securities group. The report describes data validation, ontology and rule layers, agents translating unstructured information into factors and backtests, and staged controls from information extraction through co-pilot use and constrained execution. It also reports a requirement for out-of-sample stability, agreement between live and backtested behavior, audit trails, and a period of live validation before signal authorization. These are event-report statements, not evidence that any named organization has reached a particular stage or operates a shared system. The source family does not establish model inventory, training data, permissions, production status, portfolio authority, or attributed performance.

An independent Eastmoney account of a related July 2026 Lujiazui private-securities session adds a named Blackwing Asset route: founder and CIO Zou Yitian is quoted on AI in quantitative investing and the limits imposed by alternative-data density. The report also attributes to Hangzhou Higgs Investment general manager Tan Xiaojun a shift toward heavier models and fewer hand-engineered features, with distribution shift and limited samples as constraints. It reports a University of Science and Technology of China research report on AI across research, trading, risk, and operations, and repeats the event’s emphasis on reliable data, knowledge ontologies, human responsibility, and staged agent controls. This is Chinese financial-media evidence and should remain attributed; it does not establish a Blackwing or Higgs model inventory, data rights, production status, or performance. See the capture note.

The Australia EQD 2026 programme adds an Asia-Pacific event route dated August 11, 2026. Its advertised “AI, Risk Management and Portfolio Optimization” discussion lists people from Future Fund, Victorian Funds Management Corporation, the University of Sydney, Convex Strategies, NGS Super, and arcpoint OCIO; a separate systematic-macro session lists Brad Guynn of PIMCO. The page also exposes further Australian and New Zealand data, risk, portfolio-intelligence, and asset-allocation personnel for follow-up. The agenda establishes topics and advertised affiliations only, not attendance, session content, shared systems, model use, or investment outcomes. See the capture note.

August 27 Risk.net Quantcast: Swiss and Australian academic feeders

Two adjacent episodes from the same Risk.net Quantcast Master’s Series add regional talent and curriculum evidence that was absent from the article’s prior capture set. The episodes are useful for tracing public research and recruiting surfaces around quantitative finance; they do not identify a shared employer system or establish that an academic method is used by a named fund.

The Walter Farkas / ETH–University of Zurich episode was published December 12, 2025 and is available as a public SoundCloud recording. Farkas describes a 2022 curriculum refresh that retained core quantitative courses while adding machine learning, computational finance, and some data science, with more room for data-driven methods (07:10–08:07). He also describes project work using real data, model development, public presentations, research seminars, and industry-linked thesis work (13:40–15:25). The UZH faculty profile independently identifies Farkas as Professor of Quantitative Finance, Associate Faculty at ETH Zurich, and director of the joint MSc; it lists the DaDFiR3 data-driven financial-risk project, RiskON, and a finance roundtable concerning AI and Big Data. These sources expose a public academic and industry-contact route, not a fund’s model inventory, data licence, agent permission, or investment performance. The episode also provides a useful negative control: Farkas says the programme reviews applications without AI (23:20–24:10), so “AI” should not be inferred merely from a programme’s quantitative orientation.

The Kihun Nam / Monash University episode was published December 5, 2025 and is available as a public SoundCloud recording. Nam says the Melbourne programme shifted toward asset-management and buy-side topics in response to the local superannuation market while blending AI and machine learning into mathematically named courses (03:38–04:08). He describes neural networks being used to investigate PDE solutions and optimal-control problems (04:41–04:55), plus AIM days, MISG workshops, Q-Group conferences, invited lectures, and collaborations linking students with industry (05:52–07:06). He characterizes Australian employment as concentrated in superannuation, sovereign, pension, and insurance institutions, with smaller hedge funds also hiring PhDs, and names Singapore, Hong Kong, and China as additional destinations (09:50–11:28). The Monash research profile identifies Nam’s role and stochastic-control/PDE research, while the Centre for Quantitative Finance and Investment Strategies separately lists machine learning and data science among its public research areas. The recording does not establish a named manager’s deployment, proprietary data, model ownership, or AI-attributed return.

The accompanying timestamped capture note preserves the audio hashes, ASR sidecars, and uncertainty boundaries. Both episodes should be used as academic-feeder, personnel, and conference-discovery evidence. They should not be promoted into claims about any hedge fund’s current architecture or relative position.

August 28 Badass Capital: an AI-first launch claim across prediction markets and trading

The Badass Capital first-party site and a dated Global Economic Press episode add a new Miami-based manager route. The site identifies Mark Thomas as Founder & Managing Partner and describes BA Capital Fund I LP as a Florida limited partnership. It publicly positions the firm as an “AI-first hedge fund” using proprietary models and agents across sports betting, traditional finance, crypto, and prediction markets. The site separately describes a systematic workflow of modeling, pricing, trading, and risk management across those domains.

The publisher transcript repeats more specific operating claims: AI models and agents handling research, trade alerts, and analysis in the TradFi/crypto lane (02:18–03:05); maker/taker participation and hedging with event contracts (03:05–03:34); and simultaneous platform connectivity as part of the firm’s launch narrative (03:57–04:22). Thomas’s prior technology, gaming, sportsbook-operator, and sports-betting-analytics background is described at 04:22–05:10. This is useful as a new personnel, strategy-domain, and public-architecture lead, but the technical content is firm/publisher self-description. It does not establish model versions, providers, training or retrieval data, evaluation, error rates, agent permissions, execution venues, human approval gates, production endpoints, assets, or AI-attributed performance. See the timestamped capture note.

August 28 AI-native emerging-manager websites: Formenos, Nujum, AYVID, and Sentient Snake

Four current first-party sites add a distinct emerging-manager route. They are useful for tracking explicit architecture and operating-status language, but their public claims are not interchangeable with regulated-fund or independently audited evidence.

Route	Publicly stated design or status	Evidence boundary
Formenos	Describes an AI-native hedge fund and research lab that post-trains frontier models; agents conduct research and people make decisions; the page presents a small, internally tooled team. A separate self-authored Sebastian Tan site says he is building Formenos and records a Stanford gap-year connection.	The personal site is a qualified identity and education lead, not proof of formal title, ownership, or employment; the Formenos page still discloses no full personnel roster, model/provider inventory, data rights, evaluation, regulator record, or production evidence.
Nujum	Identifies Ijlal Hannan Hafeedz and Ariff Azraai as founders and states that the Kuala Lumpur venture is pre-revenue, pre-seed, and in ideation. Its Haystack design covers Southeast Asian equities, FX, and sovereign credit, multilingual evidence extraction, a knowledge graph, debate, version-pinned strategies, deterministic replay, and logged human overrides.	The page describes a design/prototype-stage system; it does not establish a registered fund, prime broker, administrator, live capital, data licenses, model weights, or performance.
AYVID	Describes a pre-launch manager running a paper system against Interactive Brokers. Its pages describe a Python/C++ modular monolith, Neo4j supply-chain graph, seven-agent debate, named financial-data connectors, and an independent C++ risk watchdog.	The About page says the manager is not yet a registered investment adviser and that production capital follows regulatory milestones. Architecture, paper-trading, and self-reported research figures are not independent deployment or performance evidence.
Sentient Snake	Describes agent swarms that research, simulate, and ship strategies on AI-native quantitative infrastructure; the site says it is hiring two quants and one engineer.	The two quants are unnamed, and the page does not establish a funded vehicle, live deployment, model inventory, or performance.
The full capture note preserves the status, personnel, architecture, and negative-evidence boundaries. These sites expose a useful new public surface—especially the distinction between research assistance, paper execution, and intended autonomous execution—but no relative conclusion is drawn.

August 28 additional public architecture routes: Zeropoint, TAQuant, Blackwave, Strike, and Alliela

The same web pass recovered several more detailed public architecture artifacts. They are kept separate by operating status and source type.

Route	Public disclosure	Status boundary
Zeropoint Capital	A self-authored architecture post names AlphaBeta as an agentic General Partner/Fund Manager and assigns research, quantitative strategy, risk, execution, data, compliance, investor-relations, and market-intelligence roles to named agent identities. It describes structured inputs/outputs, authority levels, risk vetoes, a BTC-denominated relative-value strategy, and prediction-market work.	Agent role names are not human personnel. The reviewed site does not independently establish the legal vehicle, assets, provider contracts, data rights, execution records, or performance.
TAQuant AISA	A January 2026 technical report attributed to Tyler Leonard / TA Quant Research Labs describes event-sourced logs, offline policy updates, regime classification, policy versioning, human validation, separate risk/execution layers, and multi-tier kill switches.	A self-published technical report; its future-work section and public pages do not establish funded capital, a regulated vehicle, independent reproduction, or current production deployment.
Conformal AI / Blackwave	The studio page presents Blackwave as a systematic multi-strategy fund with agentic research, signals, and execution bounded by people and real-time risk limits, alongside Stych data tooling and Relayer Interactive Brokers infrastructure.	First-party product/fund positioning. The page does not name the prior Austin fund, provide a full team or vehicle record, or independently verify live deployment or performance.
House of Quant / Strike	The site labels Strike as in internal testing and describes five equity-options strategies, specialist research/signal/execution/risk agents, a meta-agent allocator, regime-specific walk-forward gates, post-mortems, and explicit risk controls.	Internal-testing evidence only. The page says performance metrics will be published after the sample stabilises; no funded vehicle, named team, model inventory, or realized performance is established.
Alliela Research	Allan Gaigher’s personal research project exposes a multi-agent investment-committee design, typed mandates, risk/compliance roles, immutable archives, and paper experiments.	Explicitly a personal research project, not a verified hedge fund. Several agent specifications are labeled templates not yet built. It is a negative control for separating architecture demonstrations from operating-fund evidence.

See the expanded architecture capture note. No comparison or relative assessment is implied by the table.

August 27 Risk.net Quantcast: UBS finance-native neural-network route

The Risk.net episode with Stefano Iabichino is a previously unpromoted recording-level source for a named finance-specific AI architecture. Risk.net’s November 18, 2025 page introduces Iabichino as Director of the QIS team at UBS in London and links the paper AI as pricing law, published October 20, 2025. A public SoundCloud recording supplies the timestamped layer. The QuantMinds speaker page separately describes him as a Strategy Quant in UBS’s Quantitative Investment Solutions team and lists sessions on AI as pricing law and AI in markets. Iabichino states that the podcast expresses his views rather than those of UBS AG or its affiliates (00:15–00:44).

The architecture is materially more specific than a generic “AI in investing” label. Iabichino describes network depth as time, neurons as market states, and custom Markovian activation functions intended to encode financial constraints rather than relying on standard activations (03:00–05:14). He describes backward and forward passes for non-path-dependent and path-dependent assets, including Bermudan options, XVAs, and QIS portfolios (05:14–06:08). The case study is described as combining QIS and XVA concepts, using 800 layers over 60 years of exposure, 100,000 simulated neural paths, hedge-valuation-adjustment concepts, and reverse stress testing (15:01–17:30). Those figures remain speaker-reported until independently reproduced from the paper or code.

The proposed buy-side route is also explicit but remains hypothetical. Iabichino describes learning transition structure from market-implied probabilities, recomputing option values around a manager’s view, and identifying future economic states that could maximize portfolio losses so static hedges or position changes can be considered (19:31–22:16). When asked whether the framework had been put into practice internally or externally, he declines to answer (22:19–23:14). That is an important evidence boundary: the recording supports a named UBS-affiliated research direction and proposed applications, not UBS adoption or hedge-fund deployment. The related SSRN record provides the authored-paper layer; UBS’s firmwide AI page provides separate strategy and governance context, not a connection to this architecture. See the timestamped capture note.

August 27 Risk.net Quantcast: Courant and Point72 limit-order-book research route

The Risk.net Quantcast episode with Petter Kolm adds a market-microstructure route distinct from generic AI hiring language. Risk.net identifies Kolm as director of NYU Courant’s Mathematics in Finance programme and discusses his work with Nicholas Westray; the related Risk.net paper is authored by Kolm and Westray. Risk.net’s Risk Awards profile supplies the public personnel connection to Westray’s Point72 role. That establishes a research and personnel bridge, not evidence that Point72 adopted the paper.

In the recovered public SoundCloud recording, Kolm describes taking essentially every message from individual-stock order books, often many levels deep, to predict near-term mid-price changes and an alpha term structure across horizons (29:07–30:12). He describes full-book snapshots and order-flow or imbalance features, then adds the duration between updates, time-to-vector-style representations, and intraday-seasonality inputs (36:14–38:37). The design discussion treats LSTM as a practical baseline and emphasizes data, architecture, and training choices (31:08–33:42).

The scaling detail is operationally useful but remains speaker-reported. Kolm describes one model per stock, potentially across hundreds or thousands of stocks, a GPU farm for training, and a rough 10-minute-to-one-hour training range for a few weeks of look-back data (41:50–43:40). He also says the original experiment has not been rerun on updated data and reports only anecdotal evidence that predictability has declined (43:48–44:20). The public record therefore gives a concrete research and infrastructure pattern, but not a current fund model inventory, data licence, retraining policy, production permission map, live capital authority, or independently audited performance. See the timestamped capture note.

August 27 Risk.net Quantcast: Bayes/Cass practical quant-finance pipeline

The Risk.net Quantcast episode with Laura Ballotta adds a London academic-feeder and recruiting route. Risk.net identifies Ballotta as director of the MSc in Quantitative Finance at Bayes Business School, formerly Cass. Bayes’s current course page independently identifies her as course director, lists Python, MATLAB, and C++, and offers machine-learning and predictive-analytics electives with pathways covering portfolio management, risk, quant strategies, data science, and quant trading. Her first-party profile supplies the academic role, qualifications, research interests, and publication history.

In the recovered public SoundCloud recording, Ballotta says the programme shifted toward machine learning, quantitative trading, and algorithmic trading, while adding electives as industry topics change (01:48–02:26). She describes applied group projects involving portfolio valuation and risk, real-data or industry-informed coursework, and a programming module using Python and MATLAB (03:17–04:39; 06:16–06:43). She also describes GenAI as a coding aid that must be tested: generated output is not treated as self-validating, and copy-and-paste alone is insufficient (07:05–08:17). Bayes separately describes collaboration with financial institutions and data organizations through its quantitative-finance and data-science group.

This route supports a specific talent-pipeline observation: quantitative-finance training is adding ML and algorithmic-trading content while retaining programming and model-validation fundamentals. It does not identify a fund’s internal tools, model vendor, proprietary project data, production permissions, investment authority, or AI-attributed performance. See the timestamped capture note.

August 27 Risk.net Quantcast: Imperial/CFM execution-cost and quant-research lineage

The Risk.net Quantcast episode with Johannes Muhle-Karbe adds a finance-specific execution-control route. Risk.net’s August 1, 2025 page describes an intraday tool developed with Zoltán Eisler for evaluating broker performance and links the paper Optimising broker evaluation through intraday modelling of execution cost. In the recovered public SoundCloud recording, Muhle-Karbe frames the method as ex-post evaluation when execution is outsourced or constrained by another team (approximately 02:41–03:09). He distinguishes linear spread-related costs from size-dependent impact, then describes using intraday mid-prices and time weighting to reduce market noise in the estimation (approximately 04:13–10:20). Risk.net summarizes the method as prioritizing the beginning of the trade, while the recording reports a six-to-seven-fold signal-to-noise multiplier under one modeled case; that figure remains speaker-reported.

The data requirement is unusually concrete: individual intraday fills, not only start and end prices (10:44–12:56). The speaker discusses equity-index futures, other futures, and a related exchange-data study, but says the method has not been tested across all asset classes and is less suitable for very illiquid markets. He says linear-cost estimation is used in practice and suspects impact estimation is used, while declining to provide a concrete production example; broker routing is described as a separate problem (approximately 19:00–22:00). This is an execution measurement and validation route, not a new AI deployment claim.

The personnel bridge is independently visible on Muhle-Karbe’s first-party Imperial page, which identifies him as Head of the Mathematical Finance Section and Director of the CFM-Imperial Institute of Quantitative Finance. The same page lists prior faculty positions at Carnegie Mellon, Michigan, and ETH Zürich and publicly identifies academic destinations including SIG, Jump Trading, Cubist Systematic, and Squarepoint Capital; it also lists a PhD collaboration with Qube Research & Technologies. These are academic-supervision and recruiting signals, not evidence that those firms share a system, data source, permissions, or production implementation. See the timestamped source note.

August 28 regional title-blind manager routes: Varick, Cambridge Machines, and Ark Global Fund

Three additional first-party routes expand the regional manager set without implying that their public claims are independently audited. Varick describes a quantitative manager applying machine learning to systematic futures across equities, fixed income, currencies, and commodities. Its process language runs from hypothesis to validation and continuous learning, with a proprietary platform spanning data collection through execution. The leadership page names CEO Duncan Valentine, whose biography records a 2007–2025 Capstone Investment Advisors tenure; Founder and CIO Bob Arends, whose biography records systematic work at Shell Asset Management and Henderson; Head of Research Nathan de Vries, whose biography assigns model development, data infrastructure, and AI and records prior Shell AI and Quant Research work plus a Leiden astrophysics PhD; and CTO Marcus Malak, whose biography mentions LLMs, agentic AI, synthetic data for market simulation, and Delft/Imperial training. These are firm-controlled role and process statements; the page does not identify model families, features, training data, evaluation fixtures, permissions, or AI-attributed returns.

Cambridge Machines Asset Management describes a Singapore manager using AI, machine learning, and Bayesian inference in global financial markets, with a team drawn from market practitioners, Cambridge’s Cavendish Astrophysics Group, and software engineering. The site dates the Constellation quantitative multi-strategy programme to September 2022 after four years of research, initial live trading on the Orion platform to June 2020, and a 2021 scholarship supporting a PhD on machine learning in cosmology. It states that the Singapore entity holds a capital-markets-services licence; that remains a first-party status statement here. The page does not name a current AI owner or expose model, data, evaluation, or permission details.

Ark Global Fund describes an Australian wholesale unit trust whose Fund One global-macro algorithm uses machine-learning techniques and automated trading across liquid futures in major currency, equity-index, bond, and commodity markets. It says risk layers and alerts constrain autonomous operation, and names AI Funds Management as investment manager and Quay Fund Services as trustee. This is a product-page account of automation and entity structure, not independent evidence of system design, controls, personnel, or performance. The three capture routes are preserved in the source note.

August 28 European title-blind manager routes: Genio, Bateson, GA Asset Management, and Agami

Genio Capital describes a systematic manager with learning-based systems dating, in its account, to 2016 and a multi-agent architecture organized around universe selection, factor discovery, dynamic allocation, risk management, and intelligence/reporting. Its public personnel page names Ronnie Söderman over investment and execution, Yan Pu leading research and the technical platform, and Harri Kairavuo leading quantitative solutions, alongside management and advisory biographies tied to UBS, FIM, Goldman Sachs, Citibank, Pivot Capital, UCL, and Nordic trading firms. The site says investment vehicles applying Genio strategies are managed by FCA-authorised Vittoria & Partners LLP; the FCA Register, FCA FOI11477 Annex C, Annex G, and Vittoria terms page corroborate the Vittoria & Partners LLP name and FRN 709710 route. These are first-party architecture, role, and legal-structure claims plus regulator identity evidence; the pages do not disclose model versions, training data, evaluation fixtures, agent permissions, vehicle-level permissions detail, or independently reconciled results.

Bateson Asset Management presents a London quantitative manager that says it is FCA-regulated and applies AI and machine learning across crypto and private-credit as well as other asset classes. Its contact/regulatory page repeats the FCA-regulated statement, and FCA FOI9651 Annex A lists Bateson Asset Management Limited, FRN 747038, with “Managing investments” and “May control but not hold client money”; FOI11477 Annex C and Annex G also list the Bateson name/FRN. The firm identifies Dr. Richard Bateson as founder and describes him as a former Head of Man AHL’s Dimension fund with physics connections to Cambridge and CERN. This creates a useful named-person, lineage, and UK-registration route, but the public pages and FCA CSV snapshots do not disclose an AI team, model inventory, data rights, controls, deployment status, full current permissions, or independent performance.

GA Asset Management describes a Dutch manager offering global long/short equity and managed futures. Its process page says positions are model-sized and portfolio-constrained, and that machine-learning and statistical methods are tested on noisy data with relationships retained only when they survive out-of-sample checks. The site says it operates under the Netherlands’ small-manager registration regime and publishes no performance information on the page. No named AI personnel, model details, training corpus, or permissions are disclosed.

Agami Capital describes a London systematic multi-strategy manager spanning statistical arbitrage, equity market neutral, macro, rates, FX, and commodities. Its public research list includes papers on inference constraints for LLMs, adaptive learning beyond hand-crafted signals, cross-asset risk architecture, and factor congestion. The page exposes research themes and stated platform components, not proof that a named model or paper is deployed. It provides no current personnel, model-provider, data-rights, agent-permission, or performance detail. The four routes are preserved in the capture note.

August 28 title-blind quant-risk and ML lineage route: Nikolai Nowaczyk

The title-blind Podscan search recovered The Blushing Quants episode with Nikolai Nowaczyk, published May 18, 2026. The publisher describes a mathematician, published researcher, and quantitative-risk professional whose discussion covers counterparty credit risk, model development and validation, machine learning in quant finance, and production constraints. The LSE profile independently identifies Nowaczyk as a Visiting Senior Fellow in Mathematics and lists research interests spanning dynamic initial margin, deep learning and xVA, correlated-data backtesting, GenAI and systemic risk, and AI-model validation and regulation. It records a mathematics diploma from Bonn, a mathematics PhD from Regensburg, and an academic-visitor role at Imperial College London.

The public role signal is bounded. The LinkedIn page is titled “Nikolai Nowaczyk - NatWest Group,” and the podcast speaker says he is currently working for NatWest and looking after xVA models (58:44–59:14). The automatic transcript misrecognizes the firm and acronym in places, so this is retained as a LinkedIn-title and speaker-self-report route, not as a first-party NatWest personnel-page confirmation. The route does not establish a NatWest model inventory, internal reporting line, or production AI deployment.

The AI discussion provides a useful boundary on what the public record supports. Nowaczyk says classical quant methods remain the basis of most production processes, while selected ML applications include estimating credit spreads for illiquid counterparties and improving traditional American Monte Carlo techniques (58:21–59:14). He also discusses deep pricing and deep calibration as possible speedups, alongside bias–variance diagnostics, train/test splits, learning curves, out-of-sample testing, explainability, documentation, and model validation (53:41–58:21; 01:00:39–01:03:37). These are practitioner statements about model-development patterns, not evidence of a specific institution’s implementation.

The coding-agent material is similarly bounded. In the episode, he describes financial institutions exploring coding agents for quant development while retaining controls, procedures, review, and human responsibility (58:44–01:00:38). His public LinkedIn posts discuss agents as design sparring partners for xVA and counterparty-credit-risk code, agent-assisted pair programming, and testing agentic development against difficult legacy quant problems. They name no provider, confidential codebase, permission model, production endpoint, or performance result.

The related Nowaczyk–Piterbarg backtesting presentation, presented in London on October 24, 2024, adds a research-methodology route. It covers strategy returns and risks, VaR, SIMM, and counterparty-credit-risk quantities; explains the problem of overlapping and cross-correlated samples; and gives a 10-day sliding-return example with 241 samples and up to 90% correlation (PDF pp. 3–5). It compares filtering, Monte Carlo treatment of the null distribution, and Cholesky-based decorrelation (PDF pp. 11–14). The numerical power comparison on page 26 is a toy study under stated parameters, not a trading result or evidence of adoption.

Nowaczyk is also a co-author of the 2019 Frontiers paper on AI and systemic risk. The paper identifies his Quaternion Risk Management affiliation and describes combining graph models and machine learning to generate synthetic financial systems from realistic distributions of bank-trading data, then evaluate counterparty-credit-risk outcomes under regulatory scenarios. This is an academic/practitioner research artifact. It does not establish current deployment, proprietary data access, hedge-fund use, a model provider, autonomous authority, or investment performance. The full capture and evidence boundaries are in the source note.

The later 2025 SSRN paper on branching simulation adds a more recent public research signal. It lists Anas Bakkali, Andrew Greene, Nowaczyk, and Vladimir Piterbarg with NatWest Markets affiliations and describes a branching-simulation method for estimating mean squared pricing error in high-performance pricers used for valuation adjustments and counterparty credit risk. The abstract also describes CVA-error bounds, feature-selection and polynomial-degree examples, and automated goodness-of-fit assessment. Nowaczyk’s public publications page places the work in a September 2025 WBS Quantitative Finance Conference talk, while his public LinkedIn post connects the talk and paper. This strengthens the public chronology of the research agenda; it does not establish live deployment, proprietary code access, or investment performance.

The public professional cross-check also labels Nowaczyk “Quantitative Analytics, Director, NatWest Group” on a WBS speaker page and exposes a personal “Claude Code in Action” credential on his LinkedIn profile. That is evidence of a public individual credential and role description, not evidence that NatWest has standardized on Claude Code or that the credential maps to a specific internal project. The public GitHub repository surfaced in the search is a general C/C++ example collection; no public quant model or production-risk implementation was found there. The updated capture note records the chronology and negative result.

August 28 QuantZ/QMIT: quantamental ML and agentic-AI public route

The title-blind route around Milind Sharma’s QMIT/CQA recording adds a current public strategy surface outside the original hedge-fund seed list. QMIT’s April 14, 2026 post says the recording came from the April 7 CQA Las Vegas event and presents Sharma’s 2026 book, The Quantamental Revolution. The post describes QMIT’s “ML enhanced ensembles,” Enhanced Smart Betas, a “Hedge Fund in a Box” market-neutral index, and a later discussion of LLMs, agentic goal-seeking systems, alignment, and possible AGI effects. The embedded YouTube recording is now reachable through its public video ID and automatic English captions. Caption-level timestamps add discussion of the factor-zoo problem, ensemble-to-smart-beta construction, QIS constraints, the transition from earlier AI to LLM/reasoning models, and agentic-AI policy (00:08:11–00:09:23; 00:11:32–00:11:50; 00:14:20–00:15:21; 00:19:20–00:21:08). The video stream itself returned HTTP 403 from the current retrieval route, so the captions remain automatic and audio-unverified; they do not establish a QMIT production agent, model-provider relationship, data licence, capital authority, or investment result.

QMIT’s current home page identifies the organization as QuantZ Machine Intelligence Technologies and lists Milind Sharma as CEO, Pratik Sharda as COO, Esma Gregor as Head of Business Development, Sandy Warrick as CRO Emeritus, and Econophysica Ltd. as an affiliate team. The CFA Institute profile independently identifies Sharma as CIO of QuantZ Capital and CEO of QMIT, with Carnegie Mellon computational-finance and applied-mathematics degrees and a logic doctoral-program connection. These sources help resolve the person, firm, and academic lineage; they do not disclose a model registry, model providers, training corpus, data licences, evaluation design, production agent, investment permissions, or AI-attributed performance. The root QMIT.ai page displays a different personnel surface—Amit Sardar as CTO and Dr. Oleg Kolesnikov as Head of Research—while /home-1 lists Pratik Sharda as COO and Esma Gregor as Head of Business Development. Both pages carry a 2024 copyright notice without role dates, so the public record supports a roster discrepancy, not a definitive current organizational chart. QMIT’s Hedge Fund in a Box page displays dated June 28, 2024 charts labeled as five-year daily-rebalanced and 22+ year summaries with “19Y backtest with 5 YR LIVE.” The page provides image-based charts and disclaimers, not an independently auditable return series or model/data disclosure. The Wiley publisher page confirms the March 2026, 512-page book and its public scope around factors, smart betas, multi-factor models, and ML ensembles. A Google Books contents listing adds chapter-level public vocabulary around HFIB, LBO modelling, NLP sentiment, causality, agentic AI, agent-driven quant-fund design, portfolio monitoring, and agentic factor investing. That is a published research agenda and book outline, not proof that every listed method is deployed or independently validated.

The QMIT media archive is a separate discovery surface. It lists the 2022 podcast with Dimitri Bianco, a 2022 “Hedge Fund in a Box” presentation at a Re*Work AI conference, an IBKR factor-investing master class, a 2020 AI-and-data-science-in-trading panel, a 2019 “ML, Smart Betas & LBO predictions” presentation, and a 2019 RavenPack presentation on quantamental signals and sentiment. The 2022 YouTube mirror now has a private automatic-caption capture with 3,826 timestamped segments; it adds historical context on quantitative-finance education, factor investing, and the development of quant talent, but not a current QMIT model or deployment disclosure. The QuantStrats USA 2025 brochure separately lists Sharma as moderator of a panel on AI and LLM risk in fixed income and FX. Those records identify additional capture targets and public discussion topics; an archive listing or conference agenda does not establish attendance, adoption, implementation, or results. See the capture note. The EQDerivatives Global EQD 2026 programme independently dates a May 20–21, 2026 Las Vegas event and lists Sharma for an academic keynote titled “The Quantamental Revolution — Factor Investing in the Age of Artificial Intelligence.” The same programme displays adjacent public personnel routes for Akshay Padmanabha of Balyasny, Karen Luong of Man AHL, Arnab Sen of Paloma Partners, Vishnu Kurella of a Millennium platform company, and Mayank Gupta of Fidelity, alongside sessions on volatility prediction, systematic strategies, and QIS infrastructure. The page also advertises Asia EQD (October 27–28), Investor EQD Abu Dhabi (November 18–19), and Japan EQD (November 30) as future conference surfaces. This is event-programme evidence and a discovery map, not evidence of attendance, shared systems, firm adoption, or performance. See the capture note. The Asia EQD 2026 agenda provides more regional detail: its October 27–28 Hong Kong programme includes “Enhancing Portfolio Optimization — AI and Investment Decisions,” with April Fu of Ping An of China Asset Management Hong Kong and Daniel Xystus of AOP Capital listed. It also names personnel from Jain Global, Capula Investment Management, Convex Strategies, Life Asset Management, China Re Asset Management, Hang Seng Investment, and other regional institutions. The event is upcoming and the public page does not establish attendance, recording availability, firm adoption, or shared technology. See the capture note.

---

August 28 India title-blind quantitative-value practitioner route

The title-blind queue also recovered a CFA Society India webinar with Sharad S. Ramnarayanan, uploaded October 17, 2023. The Global Conference of Actuaries biography identifies him as Appointed Actuary at The New India Assurance Company and records prior investment roles at Aditya Birla Sun Life Asset Management, Pari Washington Company Advisors, and Tactica Capital Management, along with ICT Mumbai chemical-engineering and MDI Gurgaon finance training. The Institute of Actuaries of India book excerpt independently describes his quantitative-value investing framework.

In the webinar, Ramnarayanan separates return on equity, earnings growth, required return, and entry price; treats fair value as a moving estimate; and discusses Kelly-style sizing constrained by a stated risk cap (04:25–22:00; 54:00–64:17). The route adds Indian investment-practitioner lineage and portfolio-construction vocabulary. It does not contain AI, GenAI, LLM, agent, or model-provider evidence, and it does not establish a current hedge-fund role, production system, data rights, investment permissions, or performance. The capture note records the transcript and identity boundaries.

August 28 Re7 Capital: title-blind DeFi operating-model route

The title-blind queue also recovered Capital Horizons’ interview with Evgeny Gokhberg, published December 12, 2025. The recording identifies Gokhberg as founder of Re7 Capital. Companies House independently records him as an active director of RE7 CAPITAL LTD, appointed July 8, 2021. A Re7 first-party media post dated February 12, 2026 identifies him as Founder and Managing Partner and describes market-neutral DeFi income, smart-contract risk, and diversification across platforms and chains.

The interview describes a DeFi operating model centered on market-neutral stablecoin liquidity provision, overcollateralized lending, smart-contract and code risk, and continuous monitoring (01:27–16:05). Gokhberg says Re7 built a monitoring interface analogous to a Bloomberg terminal for DeFi yields, risk, and alerts. Later passages discuss tokenisation, an unnamed large traditional financial institution, and the operational burden of high-volume alerts (28:47–41:59). These are timestamped speaker statements; the recording does not identify the software stack, data vendors, alert quality, model types, or institutional counterparty.

The European Blockchain Convention profile adds a public Re7 Labs description focused on on-chain risk curation, vault management, and DeFi ecosystem design, as well as a DeFi-native yield portfolio and discretionary long-only fund. Its approximate asset figure is event-organizer and firm-context material, not an independently verified AUM record. Public LinkedIn hiring material also exposes liquid-token/DeFi investment hiring language.

The deeper first-party pass adds an explicit Re7 DeFi Ratings framework: Re7 Labs says it evaluates smart-contract and other DeFi risks, assigns holistic ratings, and feeds those ratings into DeFi asset-allocation models. Re7’s public research also describes proprietary data infrastructure, a “Re7 Risk Engine” underpinning an ETH Yield Strategy on Optimism, Re7 Labs as risk curator for a BUIDL DeFi integration, and a Dune co-authored tokenized-funds note. A Zodia Custody partnership announcement exposes on-chain fund-interest representation and off-exchange settlement. These are dated first-party architecture and partner disclosures, but the pages do not define the engine’s algorithms or establish AI/ML, autonomy, model providers, training data, or independently audited results.

This route adds a distinct digital-asset manager and a concrete data/operations surface that a generic hedge-fund/AI title search can miss. It is not AI evidence: no public source reviewed here establishes AI, GenAI, LLMs, autonomous agents, model providers, training data, production permissions, or independently audited performance. The unnamed institutional relationship remains unresolved. See the capture note.

August 28 Hedgeye/Millennium historical healthcare-PM route

The queue also recovered Hedgeye’s July 29, 2020 interview with Michael Taylor. The host introduces Taylor as a healthcare hedge-fund manager at Millennium who ran Critical Mass (00:10–00:31). The interview covers science-to-investing background, catalyst and horizon framing, PM mentoring, idea exchange, and position risk management (01:41–05:20; 10:49–12:18; 43:01–45:57). The later Simplify prospectus provides a separate public role record for Taylor as a Simplify portfolio manager and Critical Mass portfolio manager.

This is useful historical personnel and operating-model evidence, but it is not AI evidence. No reviewed source establishes a current Millennium role, AI/GenAI/LLM/agent use, model provider, data rights, production permissions, or performance attribution. See the capture note.

August 28 Cerebellum Capital: historical ML-fund and researcher route

The title-blind pass found a Cerebellum Capital YouTube lead whose public automatic English captions were recovered on recheck, then recovered additional public artifacts through the person and firm search. The 83-second clip introduces David Andre with Cerebellum Capital and adds historical conference context about collaboration, low-latency algorithms, statistics, and big-data use (00:00–01:20). David Andre’s first-party biography says he co-founded Cerebellum and describes a largely automated process for discovering trading strategies, testing their likelihood of generalizing to unseen data, and combining strategies. A public fund-brochure record describes statistical-machine-learning software, forecasts from historical market activity, and model-made buy/sell decisions in most cases, with a human-judgment middle ground.

The technical surface is unusually specific for a historical manager: the brochure describes roughly hour-to-month holding periods and equity, option, ETP, and futures exposure; the public Steven Pav CV describes Cerebellum backtest, execution, and research infrastructure, multi-factor models, genetic programming, stock clustering, regression, execution-impact analysis, and methods to correct overfit bias. The Hertz Foundation profile supplies Andre’s Stanford/UC Berkeley and Hertz lineage and dates his Cerebellum CEO/CTO tenure through 2020, while the NYU profile identifies Thomas Dubno as a board member with prior bank technology leadership.

This is historical ML-fund evidence, not a current deployment claim. It does not establish GenAI/LLM use, current employment, model weights, training-data rights, production permissions, or independently audited performance. The founder biography and Hertz profile differ in current-role framing, so that temporal discrepancy remains explicit. See the capture note.

August 28 Optiver: title-blind low-latency infrastructure route

The title-blind queue also recovered the Optiver-linked CppCon talk by David Gross, “When Nanoseconds Matter: Ultrafast Trading Systems in C++.” Direct YouTube recovery was blocked by login protection, but the official CppCon announcement and Optiver technology article identify the talk, Gross’s Options Tech Lead role, and its low-latency trading-system scope. A public transcript mirror provides searchable text, but not an official timecoded transcript.

The public material describes order-book data structures, cache locality, contiguous vectors, bounded shared-memory queues, producer/consumer separation, multi-process failure isolation, profiling against market-data distributions, and the trade-off between FPGA speed, software flexibility, and engineering cost. Gross frames low latency as a system-design constraint that includes model accuracy, hardware, concurrency, and time to market. The Q&A says the demonstrated queue is a building block rather than a durable write-ahead-log implementation and that structure-of-arrays techniques are used selectively. These are public engineering statements, not a complete production-stack disclosure.

This route is relevant to the AI map because Optiver’s separate research page describes an AI Lab and LLM-based analysis, and its 2026 agentic-SDLC article describes agent-led exchange-connectivity development with evaluation, isolated execution, and governance. Those later first-party surfaces should not be back-projected into the 2024 CppCon talk. No reviewed source here discloses an Optiver model, provider, training corpus, agent connected to orders, permissions, or independently audited performance. See the capture note.

August 28 Callanish Capital: title-blind historical quant-manager route

The title-blind queue recovered the OpalesqueTV interview with Eoin Murray, dated August 2, 2011. Opalesque identifies Murray as CEO and Chief Risk Officer of Callanish Capital Partners, a London-based quantitative hedge fund, with earlier quantitative-management roles at Northern Trust Global Investments, Deutsche Asset Management, and Old Mutual Asset Managers. The video cache supplied chapter markers but not a complete transcript, so the synthesis uses publisher text and cross-checks rather than treating the cached markers as verbatim evidence.

The Opalesque roundtable says Callanish was established in 2008 and launched the Callanish Global Macro Fund in May 2010 with seed capital from IMQubator, the seeding platform established by Dutch pension fund APG. A contemporaneous Opalesque report describes the challenge of recreating institutional infrastructure in a startup hedge-fund business and names Fabrice Connin as a portfolio manager. The HSTalks Quantmare catalogue adds historical quantitative-investing, factor-model, volatility, liquidity, and crowded-trade topics. Later Federated Hermes disclosure cross-checks Murray’s subsequent personnel chronology.

No reviewed source identifies AI, GenAI, LLMs, autonomous agents, model providers, training data, or AI-attributed results. “Models” in the Opalesque description is treated as conventional investment-model language. This is historical systematic-global-macro, personnel, seeding, and infrastructure context; it does not establish current Callanish operation, current GSA or Federated Hermes practice, production permissions, or performance. See the capture note.

August 29 Odd Lots: allocator view of multi-strategy operating models

The title-blind queue recovered Bloomberg’s Odd Lots episode with Albourne partner Ronan Cosgrave, published May 26, 2025. Apple Podcasts and Bloomberg cross-check the episode identity and scope. Direct YouTube recovery was blocked, but the Omny transcript API exposes 185 timestamped segments through 49:11; Tapesearch and Podscripts provide secondary search indexes.

Cosgrave frames allocator diligence as evaluation of people and trades together with the firm’s business, risk, and investment models (06:01–07:12). He distinguishes traditional multi-strategy structures from pass-through/pod-shop structures and discusses how compensation, fees, leverage, cost control, and alignment interact (08:06–09:10; 18:14–20:29; 28:41–30:12). This is allocator-level operating-model context; it is not a disclosure by Millennium, Citadel, or another named manager. The recovered API transcript uses generic speaker labels, so these are timestamped paraphrases rather than speaker-verified quotations.

The episode does not identify an AI model, LLM, agent, provider, training corpus, data rights, or AI-attributed performance. References to “models” are treated as conventional investment, risk, or business-model language. The route adds a useful diligence vocabulary for examining how research, risk aggregation, incentives, and execution might fit together, without inferring a firm’s internal technology. See the capture note.

August 28 Quantmatic Capital: first-party AI-native FX disclosure with a temporal discrepancy

The title-blind queue recovered a Meet the Fund Managers interview with João Guilherme Cruz, identified by the interviewer as an industrial engineer building Quantmatic Capital in Brazil. The interviewer’s LinkedIn post, Cruz’s public LinkedIn profile, and the Quantmatic company page cross-check the firm/person route. The The Org listing displays Cruz as “Cofounder & Managing Partner” but marks the organization record unverified; that title is not treated as independently confirmed.

The Portuguese-language InfoMoney report on XP Quant Summit 2026 adds a separate Brazilian systematic-manager route through Mauro Shinzato, identified as partner and CEO of Giant Steps Capital. The report attributes to him discussion of executive-call transcripts, sentiment analysis, inconsistency detection, geolocation data, data-to-trading latency, and signal diversification, alongside explicit caution about testing failure, noise, cost, and complexity. This is regional financial-media evidence rather than a firm-controlled model disclosure; the reported figures and implementation details are not independently audited. See the capture note.

A separate Brazil route is an open GitHub replication of “Mimicking Finance”. Its README says it reconstructs month-to-month CVM holdings for 174 Brazilian funds, supplies Python LSTM code and downloadable data, and reports headline predictability and performance statistics. The repository’s own audit file documents duplicate-position bugs, an unverified fund appearing in results, inconsistent trade counts, a small sample, threshold sensitivity, and methodological differences from the NBER paper. The code and data are therefore valuable replication and benchmark material, but the headline results are not promoted as validated until the data construction and evaluation are independently repaired and rerun. See the capture note.

Quantmatic’s first-party website states that proprietary AI algorithms, quantitative models, machine learning, and AI agents drive FX signal generation, position sizing, and risk management. The page says agents search the web, social media, and news for fundamentals, sentiment, sudden-change indicators, and expected volatility around economic events. It also says the system has traded real capital since 2025, a team monitors operations with escalation and a kill-switch, and engineering reviews and improves algorithms monthly. Its strategy description names more than ten technical indicators alongside fundamentals, price action, and machine-learning algorithms.

The interview adds implementation detail. The speaker describes MetaTrader expert advisors, three algorithms embedded across seven strategies and 21 signals (06:48–07:21), technical indicators, nearest-neighbor language, grid positioning, and global-news/sentiment scraping (07:51–08:17). He dates the coding and structuring phase to 2024 and says the fund went live in 2025 (04:02–04:36). At 16:02–17:15, he describes an AI copilot suggesting when to operate and what lot size to use, while the conversation retains a human operator and final judgment.

That last point does not cleanly match the current website’s statement that decisions are made “without human discretion.” It may reflect system evolution, different meanings of “decision,” or inconsistent public descriptions. The public record does not resolve the discrepancy, and it does not disclose model families, providers, training data, data rights, evaluation protocol, brokerage permissions, order authority, or independently audited performance. The capture note preserves the full boundary: Quantmatic Capital / João Guilherme Cruz.

August 28 Plettenberg Capital: research-workflow AI, a market-decision boundary, and an internal LLM interface

The Sophron Network episode identifies Alexander Levin as Founder and Chief Investment Officer for public markets and Daniel van Flymen as Co-Founder and Chief Technology Officer at Plettenberg Capital. Its publisher description says the New York firm runs roughly 500 live algorithms, uses one engine for research and live execution, treats point-in-time data and corporate actions as backtest-integrity controls, and keeps language models away from market decisions. These are publisher and speaker claims; the local audio was recovered and is being retained for timestamped spot checks before any verbatim quotations.

The public description places AI in discovery, testing, and validation workflow compression rather than direct market prediction or execution. It also says the discussed process uses price-focused inputs and deliberately excludes volume, fundamentals, and alternative data. The public record does not expose the data vendor, adjustment logic, simulator, model families, evaluation fixtures, execution controls, or the permissions of any research tool.

Local audio spot checks add precision to that boundary. Around 24:17–30:20, the speakers describe roughly 500 production algorithms, decision-tree and allocation layers, point-in-time pricing, ticker changes, and corporate actions as backtest-integrity concerns. Around 43:08–44:47, they say AI is used for hypothesis variations, reporting and anomaly triage, and visualization, while not measuring the market, touching an order, or sitting in the production line; they also describe deterministic outputs traceable to decision-tree paths. Around 50:04–50:50, they describe highly automated monitoring and execution with people retaining oversight and limited manual changes. These are speaker statements recovered from the public audio, not an independent system audit; the capture note records the ASR and spot-check boundaries.

There is a separate public van Flymen post describing internal tools and an autonomous engineering platform. The accompanying Levin post says an internal Plettenberg LLM allows natural-language interaction with portfolio and returns. Read together with the episode description, the record suggests a possible distinction between an LLM interface for internal portfolio/reporting interaction and models kept out of the market-decision path. Scope and timing are unresolved: no model name, provider, retrieval/control policy, audit log, or production permission is public. Levin’s launch post also names Clear Street, Middlegate, Riveles Wahab LLP, NAV Fund Services, and U.S. Bank as operating partners, without specifying their contractual or data roles. Full boundaries are in the capture note.

August 28 Alex Zhong: cross-platform quant research and crypto portfolio construction

The Sophron episode and its publisher LinkedIn announcement describe Alex Zhong’s route through WorldQuant’s BRAIN/ACE research environments, the DRW Crypto Predict competition, and research with Trexquant on machine-learning equity strategies and systematic signal development. The publisher describes a current focus on portfolios of more than 50 crypto strategies, signal validation, combination, and alpha decay. A public profile shows an AX Quantitative surface and a crypto/quant hiring signal, but does not independently confirm every historical affiliation.

This route is useful for the research-method layer: out-of-sample testing, correlation-aware combination, strategy decay, and portfolio construction are more specific than a generic “AI in finance” label. The reviewed sources do not identify an LLM, agent, model family, training data, current fund mandate, production permissions, or independently audited performance. Historical platform participation is not treated as evidence that WorldQuant, Trexquant, or DRW deployed a shared system. See the capture note.

August 28 Shell / Nikita Granger: reinforcement learning in gas and power markets

The Sophron episode identifies Nikita Granger as a Quantitative Developer at Shell and describes reinforcement-learning models for gas and power trading, price forecasting, and structured energy derivatives. Shell’s official research bibliography independently lists Granger as a co-author on 2023 papers for deep RL gas trading and domain-adapted/imitation-based power arbitrage. The public personnel route adds econometrics, applied mathematics, and computational-science training plus earlier energy-modeling work.

The papers expose specific model targets rather than a generic “AI” label: sequential decisions in natural-gas futures; dual-agent decision-making across Dutch day-ahead and intraday power markets; domain-informed reward engineering; imitation of trader behavior; and order tranching. The papers report simulation or backtest comparisons under their own assumptions. They do not establish live Shell deployment, a current model owner, data rights, trading permissions, or AI-attributed returns. Shell’s separate corporate AI page mentions Azure OpenAI GPT-4, Copilot, Shell e, and reinforcement learning for decision support, but it does not connect those platforms to Granger’s papers or a particular trading desk. Full source boundaries are in the capture note.

August 28 The Sophron Network: fund autonomy boundaries and a separate public LLM project

The newly identified Sophron Network publisher feed adds three relevant routes. The feed and Apple Podcasts listing identify the show as a finance and quantitative-trading interview series; direct Spotify/Anchor episode pages and Listen Notes records provide dates, descriptions, and navigation markers. These are publisher and platform records, not a substitute for a transcript.

In “How Close Are We to Fully Autonomous Trading?”, the publisher identifies Jeremy Kadouch as Portfolio Manager at MN Fund. MN Fund’s first-party team page separately describes his portfolio-manager remit, strategies, and capital deployment. The episode description says MN Fund operates a 24/7 systematic volatility engine, while deliberately keeping the trade decision away from AI. It also marks a move from manual trading to a systematic engine, an intelligence layer, and an algorithm that reads and reasons without AI in execution (08:39; 11:37; 16:41; 21:15). This is evidence of a stated separation between automation and AI decision authority, not an independent audit of the fund’s system.

Kadouch’s public LinkedIn post describes a separate project, Quorum Index: continuous crypto-pair monitoring; OHLCV, narrative, and macro inputs; OpenClaw orchestration and persistent memory; 22 scheduled skills using the Anthropic API; weekly Claude Code review; and a custom MCP server exposing market state, signal reliability, and output quality. Follow-up posts describe checking the last five outputs for repetition and event-driven repair with cooldown and scope limits (feedback post, repair post). These are self-reported details of a public market-intelligence project associated with Kadouch; they are not evidence that MN Fund grants this system trading authority. No repository, logs, benchmark, model configuration, cost record, or independent reproduction was reviewed. The recovered Sophron audio adds a bounded scope check: Kadouch describes AI-assisted stack construction and external enhancement, but says the executable strategy itself applies mathematical rules to market data without AI inside the execution path (16:56–23:30); he describes Quorum separately as an agentic market-intelligence project (52:17–56:34). See the ASR recovery note.

The Robert Carver episode is a historical Man AHL personnel route. The publisher describes Carver as an independent systematic trader and writer, formerly at AHL, most recently Head of Fixed Income, and describes his current personal system as covering about 250 futures markets with the open-source pysystemtrade framework. Recovered audio adds timestamped methodology detail: Carver discusses overfitting risk and robustness across regimes, explains a modular simulation/production/broker architecture (49:51–53:17), and warns against trusting AI- or “vibe”-coded trading software with money (65:19–68:05). It does not establish that AHL used the personal methods described, or disclose an AHL model, dataset, permission boundary, or AI-attributed result. See the ASR recovery note.

The Amay Patel episode identifies Patel as Partner and Head of Investments at Atomic Digital and describes a 2025-launched market-neutral digital-asset fund spanning OTC structured products, DeFi, quantitative, fixed-income, and credit markets. Local ASR closes the audio gap and captures one isolated GPT reference in a legal-knowledge context (13:17), but no Atomic Digital AI system, model, data pipeline, or agent authority. Full boundaries and routes are in the Sophron capture note and ASR recovery note.

August 28 Optica: founder-led AI-native quant disclosure found through a title-blind route

Optica’s first-party site describes the firm as “AI-Native Quantitative Investment Management,” while its public LinkedIn company page describes an AI-native quantitative hedge fund. The company page displays Seattle, an 11–50 employee size band, a 2024 founding year, and a partnership organization type. These are company-page fields, not independent verification of legal structure, assets, or live deployment.

The personnel route is unusually useful because it begins with founder activity rather than an AI-labelled vacancy. Co-founder Adarsh Nair’s public hiring post sought a quantitative data scientist and named MLPs, CNNs, RNNs, and time-series transformers as relevant experience, with risk modelling and portfolio-management algorithms as useful additions. That is a hiring-intent signal, not evidence that each architecture is used or that the role was filled.

Nair’s Beyond the Grind interview, uploaded September 27, 2025, is titled around a career transition and does not name Optica. The episode therefore demonstrates why the discovery program must search founder names, career narratives, chapters, and guest appearances. In the interview, Nair describes Optica as a pure-play US-equities investment construct (17:53–18:03), a prediction engine for selected stocks and market-entry/exit timing (30:54–31:25), and a system built around proprietary algorithms that determine when to buy and sell (29:24–30:22). He also describes a human role in designing algorithms and retaining portfolio-level judgment (18:52–19:09; 29:14–29:23).

The episode uses the phrase “our chief investment officer is AI” (29:43–29:49). That is conversational/promotional framing, not a verified personnel title or evidence of autonomous legal authority. The captions are auto-generated; the phrase “hundreds of milliseconds” appears in the same passage, but the exact technical wording should be audio-checked before being treated as a specification. The episode’s discussion of public-market and satellite-style data examples is similarly a research lead, not proof of an Optica dataset or data licence.

Nair’s separate “Phase Locking in Financial Markets” post supplies a publicly discussed hypothesis about synchronized cross-asset behaviour under stress, liquidity, and volatility. It does not establish a deployed Optica model, backtest, position, or performance result. The reviewed record does not name Optica’s model providers, training corpus, retraining process, evaluation design, risk limits, brokerage permissions, or order authority. Full provenance and failure boundaries are in the capture note.

August 28 Binomial Technologies: public finance-model stack, Sentry control surface, and OpenAI pilot claim

Binomial Technologies’ first-party firm page describes a long-short equity hedge fund launching in June 2026 and names Ilay H. Ibrahimzadeh as Portfolio Manager & Lead Engineer. The page says he oversees portfolio management, quantitative research, and technology infrastructure, and describes Lattice as a proprietary AI system for quantitative modelling, risk diagnostics, and financial workflow automation. The same page gives a Bocconi finance/mathematics background and an in-progress machine-learning/computer-science master’s at Reichman University. These are firm-controlled biographical claims; they are not independently verified employment or academic records.

The public research page is unusually explicit about the intended model boundary: fine-tuned open-weight LLMs, classifiers, forecasters, and task-specific networks trained on the firm’s own data pipeline. It describes these as small and auditable, and says the firm intends to expand the work into a dedicated AI research lab. The page names a “Binomial Marks” information-sheet release, a “Binomial Shannon” news-sentiment model, and a Sentry benchmarking item, but the linked PDFs were not recoverable in this pass. No benchmark result or model specification is therefore promoted.

The Sentry page describes a deterministic supervisor routing queries to separately prompted and tooled specialists for equity research, quant/risk, derivatives, capital flows, and market signals. It also describes persistent workspace state, citation trails containing source/timestamp/tool information, scheduled tasks, Excel and SDK surfaces, risk decomposition, and post-trade attribution. The approach page describes 700+ signals spanning value, quality, growth, sentiment, options activity, alternative data, and institutional flow; ML ranking by mispricing probability; stress testing; thesis breakers; and a human sign-off before capital moves. These are detailed first-party architecture and process descriptions, not an independent deployment or return audit.

The firm page also states that Binomial has entered an OpenAI pilot covering AI-model use inside hedge funds and advises on integrating OpenAI products into investment firms. That is a public partnership claim, not an independently corroborated OpenAI announcement; scope, model identity, production permission, and commercial terms are unknown. Separately, BinomialHash on PyPI is an independently visible MIT-licensed software artifact for structured-data compaction, retrieval, statistical/causal tools, and multiple model-provider adapters. Its existence does not prove use in the fund’s live stack or imply a provider partnership.

The public artifact layer goes further. The Binomial Marks model card describes a 400-million-parameter, 16,384-token encoder producing 23 structured earnings-call outputs. It says the model was trained on more than 80,000 transcripts covering more than 2,700 tickers and evaluated on a held-out 2,000-call set against frontier-model labels. The card calls it a Tier 2 research preview and explicitly separates language-model scoring from an alpha model. Binomial’s Confidence Premium note reports a peer-relative confidence factor built from the model, with a 63-day rank IC of +0.025 over a July 2011–July 2026 evaluation sample. Those metrics are firm-authored, dated, and backtest-based; they are not independently replicated or evidence of live returns.

The separate Binomial Shannon 2 note describes an approximately 150-million-parameter news model with ticker/macro routing and structured outputs for event type, direction, novelty, specificity, materiality, and claim type. It is presented as a distilled, locally deployable research preview and explicitly distinguishes agreement with teacher models from realized-return prediction. Together, Marks and Shannon expose two distinct modality lanes—earnings-call language and news—not just a generic “AI analyst” label. The public record still does not establish full corpus licensing, point-in-time joins, live model permissions, or independently audited signal contribution.

This route adds a concrete public model/platform vocabulary—fine-tuned open-weight models, specialist routing, persistent workspaces, tool/citation trails, human gates, and post-trade attribution—without establishing weights, training-data rights, evaluation fixtures, live AUM, order authority, or AI-attributed performance. See the capture note.

August 28 BedRock Partners: former buy-side practitioner route and methodology-specific AI discussion

The first-party BedRock Podcast E19 transcript, published October 21, 2025, identifies Bill Sun Qingyun as a mathematics PhD graduate of Stanford and describes his own prior work at Google Brain, Millennium, Citadel, and Point72. Those historical affiliations are guest-reported in the transcript and are not treated as independently confirmed personnel records. The separate BedRock founder page identifies Cong Tan as CIO and founder and supplies a Stanford/Fudan lineage and a first-party description of a standardized research framework.

The episode distinguishes repeatable, back-testable problems from long-horizon judgments about unprecedented events. The speakers describe AI as a research assistant for information retrieval, key-point extraction, methodology-specific question templates, and updating judgments as new evidence arrives (publisher transcript, approximately 00:00–19:00). They discuss turning a fund’s investment method into a junior-analyst agent template and supplying private-data access, expert-call records, and Bloomberg-like plugins. They also frame the current limitation as precise retrieval/tool integration and engineering packaging, rather than a missing foundation model.

The transcript separately introduces a Gen Alpha / AIUSD.AI crypto product concept in which AI acts as a wealth-management and execution assistant. That is a guest/product description and is not merged into BedRock’s investment-process evidence. The BedRock archive’s additional ML and cybernetics articles are retained as title-blind research-culture leads. No reviewed source establishes BedRock’s model inventory, data rights, live deployment, order authority, regulatory status of the crypto product, or AI-attributed performance. Full boundaries are in the capture note.

The newly recovered public audio changes the archive from metadata-only to timestamped navigation, with an important evidence boundary. In E25, Cong Tan and guests discuss AI as a broad-reading and research aid while retaining analyst-defined questions and judgment (approximately 41:09–47:14); the publisher separately positions the episode around global compounder investing and AI’s effect on fundamental research. In E26, a pseudonymous external engineer discusses model capability versus environment, instructions, tools, and verification loops (approximately 05:07–08:00). In E27, BedRock and Trader Jiu discuss the shift from model training toward inference and agents, and the gap between capability and adoption (approximately 01:25–04:43; 09:48–11:41). These are automatic Chinese-ASR paraphrases from public audio, not quotations or evidence of BedRock deployment.

E28 includes firm-process statements about retrieval, model updates, data tracking, basic research agents, monthly AI-use reflection, and researchers working more like product managers for AI collaboration (approximately 00:00–01:00; 07:36–08:14; 08:36–15:13). E29 and E30 are client/manager discussions of AI as an investment and infrastructure topic, including agents, compute, data transfer, and first-principles analysis (E29 approximately 11:24–13:30; E30 approximately 18:15–21:31). None of these episodes names a model, training corpus, data license, evaluation denominator, production endpoint, permission map, autonomous portfolio authority, or AI-attributed performance. The BedRock capture note records the audio URLs, ASR settings, and per-episode boundary.

The same public channel exposes earlier AI episodes that were previously under-tracked. E16 covers cloud, model-company, application, and hardware economics; E20 covers AI investment themes and valuation debate; E22 discusses agents and investment-research work; E23 discusses code execution and running several model updates for earnings work (approximately 06:52–07:28); and E24 introduces two BedRock analysts and their AI/agent-era research context. Local ASR provides navigation sidecars for these episodes, but they remain speaker commentary rather than evidence of a named model, data rights, production deployment, or investment authority.

August 28 AIMA Singapore: fund-manager briefing and title-blind AI leadership route

The AIMA Singapore Fund Manager Briefing page records a January 21, 2026 in-person session on AI in finance. AIMA framed the event around the shift from narrow machine learning to foundation and domain-specific systems, with deployment, governance, and measurable business outcomes as the stated questions for financial institutions. Its displayed speakers included Natalie Ang in Morgan Stanley Prime Brokerage Business Consulting; Hozefa Topiwalla, listed as Head of Asia Research Product and Asia Lead for Firmwide AI at Morgan Stanley; Matt Zhang, listed as Chief Risk Officer at Keystone Investors; and Gary Ang in AI Governance and Risk Management. This is a useful Singapore personnel and conference-discovery route, especially because the relevant firm roles are not all labelled “AI.”

The public page exposes event scope and displayed roles only. No replay, transcript, model inventory, vendor list, training corpus, data-rights record, production permission, investment authority, or AI-attributed performance was located. The route therefore does not establish that Morgan Stanley and Keystone Investors share an architecture or that any named speaker described a live system. See the capture note.

A public LinkedIn profile for Matt Zhang Changhao separately displays a Keystone Investors affiliation and lists him as a panelist at the AIMA briefing, with an NUS education record and a professional certificate in investing using machine learning and alternative data. A public profile for Hozefa Topiwalla associates him with Morgan Stanley in Asia, but does not reproduce the AIMA title. These profiles strengthen the event-person discovery route; they do not establish Zhang’s exact current remit, Topiwalla’s AI responsibilities beyond the AIMA event page, or either firm’s model implementation.

August 28 AIMA: AI meeting tools as a fund-manager governance route

The AIMA event page records an August 26, 2026 webinar with Charu Chandrasekhar of Debevoise & Plimpton and Suzan Rose of AIMA. AIMA says the discussion revisited the use of AI tools to generate transcripts and summaries of calls and meetings, with attention to adoption, compliance concerns, and fund-manager considerations. This adds a current title-blind route for meeting-data and voice-workflow discovery.

The page provides organizer, date, displayed roles, and stated scope. The linked replay was not publicly retrievable in this capture, so there is no transcript-level evidence, named fund implementation, product or model inventory, retention policy, data-rights record, production permission, investment authority, or performance claim. “More commonplace” is the organizer’s framing, not an independently measured adoption rate. See the capture note.

August 28 Middle East conference route: Abu Dhabi Finance Week 2026

The ADGM announcement and ADFW site place Abu Dhabi Finance Week 2026 on December 7–10, 2026. The first-wave announcement lists Dmitry Balyasny as Co-Founder, Managing Partner and Chief Investment Officer of Balyasny Asset Management, Steven Desmyter as President of Man Group, and an inaugural AIMA Global Hedge Fund Leaders Platform. The official ADFW site separately displays Balyasny as Managing Partner and Chief Investment Officer. This is a new Middle East conference and future-recording route for the tracked speaker and firm universe.

The evidence is limited to event identity, dates, displayed roles, and programme structure. It does not establish attendance, session remarks, AI or GenAI use, a model or data partnership, production deployment, investment authority, or performance. The event should be monitored for a public agenda, Arabic/English materials, recordings, and speaker-authored follow-ups. See the capture note.

The MSCI Institutional Investor Forum Doha page adds a September 14, 2026 MENA event route focused on AI, geopolitics, portfolio construction, public-market factors, and private-market data. It displays Ashley Lester, Seibert Kruger, Uday Karri, and Bentley Kaplan in MSCI research or institutional-investment roles. This is vendor/event context, not evidence of a hedge fund’s implementation. The page supplies no transcript, attendee confirmation, model or dataset disclosure, production authority, or performance evidence. See the capture note.

August 28 Arthur AI Fest: Balyasny speaker and model-evaluation route

The Arthur AI Fest page lists Charlie Flanagan as Head of AI at Balyasny Asset Management. Its displayed agenda includes LLM evaluation and benchmark design, efficient inference and controllable serving, AI-platform architecture, embeddings, autonomous agents, multimodal models, and governance. The page displays September 26 but omits the event year, so the date is unresolved. This is a speaker and agenda signal, not evidence of a Balyasny/Arthur partnership, model choice, dataset, deployment, investment authority, or performance. See the capture note.

August 28 UBP Hedge Fund Conference: AI panel and manager-discovery route

The UBP conference recap, dated July 17, 2026, says its London Hedge Fund Conference sold out with more than 170 attendees and included managers and clients from the UK, Asia, Europe, and the Middle East. UBP says panel discussions included representatives of Campbell, Brigade, and Shannon River and addressed the impact of artificial intelligence alongside macro and liquid-credit themes. UBP names Kier Boley and John Argi as Co-Heads of its Alternative Investment Solutions platform in accompanying quotations. The embedded video is blocked by cookie preferences, and the page does not identify the individual manager representatives. This is a new first-party conference and follow-up route, not evidence of a manager’s model, partnership, deployment, investment authority, or performance. See the capture note.

August 28 Institutional Investor FTF: CIO/CTO and Automating Alpha route

The Institutional Investor FTF Winter Workshop & Automating Alpha page displays February 24–25, 2026 in Fort Lauderdale and describes a seventh-year event for buy-side chief technology officers and chief investment officers. Its stated scope is applying artificial intelligence, new data sources, and analytical tools to investment management. The page simultaneously says registration is not open and does not expose a complete agenda or buy-side speaker roster. This is an operating-model discovery route, not evidence of attendance, a named fund’s system, model, data partnership, deployment, investment authority, or performance. See the capture note.

August 28 J.P. Morgan QIS: point-in-time data, hedge-fund overlays, and bounded LLM use

The first-party J.P. Morgan Making Sense episode on quantitative investment strategies, published November 11, 2025 and recorded October 15, 2025, introduces Arnaud Jobert as co-head of Strategic Indices for Markets and global head of Equity Structuring and Eloise Goulder as head of the Data Assets & Alpha Group. Jobert says the Strategic Indices/QIS platform crossed $100 billion of client notionals and spans equities, commodities, rates, and multi-asset strategies. He describes hedge-fund use cases including execution outsourcing and central-risk-book overlays, plus a Nexus structure in which clients retain their own models and intellectual property while outsourcing execution. These are J.P. Morgan’s provider descriptions, not named-client disclosures.

The transcript describes JPMaQS as a broad, long-history, point-in-time macro-fundamental dataset and discusses planned macro strategies across rates, equities, and FX. Its AI/LLM discussion is narrower: writing factor and performance commentaries and reducing time to market so teams can focus on subject-matter work and strategy development. The page does not disclose a model provider, fine-tuning, training corpus, client implementation, autonomous trading permission, or AI-attributed performance. Notionals are platform scale, not AUM or return evidence. See the capture note.

The related first-party Deepak Maharaj QIS episode, published November 12, 2024 and recorded October 7, 2024, adds an earlier model-development layer. Maharaj describes an LLM-enhanced Quest thematic-index workflow in which news articles are used to identify companies associated with a theme and an LLM refines the search vocabulary for those associations. He describes internal testing against an earlier NLP approach but gives no test design, sample, benchmark, numerical result, or evaluation threshold. The episode also names hedge-fund factor and thematic use cases and a Macrosynergies partnership for macro time-series data, while discussing social-media, sentiment, intraday, and credit-card data as exploration areas. This is provider-reported product-development evidence, not proof of a named client’s model, data rights, production permission, or performance. See the capture note.

The same first-party archive’s July 2025 intraday-QIS episode identifies Mathieu Boisot as head of Cross-Asset, Volatility and Intraday Product Development. Boisot describes live gamma maps, short-dated and zero-day-to-expiry options, levered ETFs, intraday data, and daily estimates of options gamma imbalance as inputs to understanding possible intraday momentum or mean reversion. He gives a speaker-reported observation that roughly 60% of relevant volume was less than one week to expiry, without publishing a methodology or denominator. This is a market-structure and data signal route, not evidence of a named client’s model, permission, autonomous authority, or performance. See the capture note.

The related first-party Rui Fernandes structuring episode, published February 21, 2025 and recorded January 29, 2025, describes a cross-asset structuring team of approximately 140 people across seven financial centers and three operating principles: innovation, customization at scale, and platform interoperability. Fernandes describes multi-strategy and multi-asset client needs, QIS delivery, and a Nexus structure in which clients retain their models and intellectual property while outsourcing execution. AI appears as a broad emerging trend, without a model, agent, data vendor, evaluation, or deployment disclosure. This is organizational and platform evidence, not evidence of a named hedge-fund implementation or performance. See the capture note.

August 28 Graham Capital: AI/ML hiring mandate anchored in systematic trading

Graham Capital’s current AI / ML Quantitative Research Manager posting is a firm-controlled hiring signal with a more specific scope than a generic AI vacancy. It places the role in Quantitative Strategies, reporting to and working with the CIO of Quantitative Strategies, and asks the manager to assess existing AI/ML capabilities, define a research roadmap for systematic trading systems, and create new AI/ML trading signals. The listed work includes production-grade models over large-scale alternative datasets, integration of new data sources, cross-functional implementation with data science, technology, operations, and trading, plus robustness, scalability, compliance, white papers, and conference presentations.

The Quantitative Strategies page supplies the investment context: Tactical Trend, Quant Macro, and K4D combine price-based and non-price-based models, including trend, fundamental macro, carry, and value/reversion, with portfolio construction, risk management, and execution capabilities. Graham’s careers page describes quantitative researchers as processing large datasets to detect signals and patterns, while the research archive exposes a dated public research trail.

That trail includes a 2017 Machine Learning note, a 2019 Model Interpretability in Machine Learning note, and the 2023 Chatting with ChatGPT note. In the last item, Graham researchers tested repeated sentiment prompts on financial-news text: the simple example produced a reported mean of -70.75 with standard deviation 5.91 over 20 responses, while the more complex example produced a reported mean of -28.50 with standard deviation 34.50. The authors concluded that out-of-the-box ChatGPT was unsuitable for systematic trading because non-trivial inputs produced unstable outputs and systematic strategies require repeatability and transparency.

Taken together, the public record shows a dated progression from ML and interpretability research, through an explicit LLM reproducibility test, to a current hiring mandate that includes alternative-data models and systematic-trading signals. It does not identify the person hired, a current provider or model family, training-data rights, evaluation fixtures, production endpoint, model permissions, autonomous order authority, or AI-attributed performance. The capture note keeps the hiring, strategy, and historical research layers separate.

August 28 AI-native market labs: Event Horizon Labs and Quantegies

Two newly indexed first-party sites describe market-research systems organized around AI agents, but neither is classified here as an established hedge-fund vehicle. Event Horizon Labs describes a San Francisco AI research lab for markets, a small researcher/engineer team, an agentic hypothesis-and-experiment loop, and an aspiration toward end-to-end automated research. Its public careers page names three founding role families—AI Research Engineer, Quantitative Researcher, and Infrastructure Engineer—and describes reward design, model improvement, knowledge accumulation, backtesting, reproducibility, checkpointing, versioned datasets, exchange connectivity, execution, and monitoring. The site aggregates prior Citadel, Jump Trading, Stanford, Caltech, and Berkeley affiliations without naming the individuals.

EHL’s research index adds public artifacts beyond generic positioning. A noisy-rewards note describes an LLM-driven evolutionary loop that edits mutable strategy code while holding the simulator, data, and PnL accounting fixed, then evaluates candidates on held-out backtests. An execution note, co-authored with the University of Chicago Project Lab, describes a researcher-agent/backtester/evaluator loop with GitHub versioning, S3 archival, self-hosted compute, scheduled island migration, and quality-diversity search across execution templates. A Hyperliquid note describes wallet-level reconstruction from public fills, realized versus open-position PnL separation, null-hypothesis checks, and out-of-sample analysis. These are public research and self-reported backtest artifacts under stated assumptions; they do not establish a conventional fund, external capital, live order authority, or audited performance.

Quantegies describes an AI research lab building systematic strategies and names a “Foundation Models for Finance” thesis. Its public architecture vocabulary includes temporal transformers for financial time series, counterfactual stress-regime simulation, neural execution policies, multimodal market and alternative-data ingestion, and adaptive training/inference capacity. The associated BAQLABS surface describes queue-aware simulation, latency controls, order-book analytics, MBO replay, a latency lab, a model lab, and results. Advertised roles include a Quant Researcher focused on feature engineering and alpha generation and an Execution Trader who manually executes systematic strategies and coordinates with research.

These sites expand the search vocabulary from “AI analyst” to objective-function design, evolutionary search, reward verification, population migration, research memory, microstructure simulation, and execution-policy evaluation. They are useful adjacent routes for discovering models, researchers, code, and future filings. The reviewed sources do not disclose model weights, providers, training-data rights, named current personnel, legal fund structure, production capital, permission maps, autonomous order authority, or independently audited returns. Full source boundaries are in the AI-native market-lab capture note.

August 28 QSentia and SEKUIA: adjacent investment-intelligence routes

QSentia presents an AI investment-intelligence platform for institutional users rather than a hedge fund. Its public architecture separates agent signals from portfolio context, human approval, decision history, mandate limits, and model governance. Its Shadow Mode description says an institution can connect a real portfolio, record proposed actions without capital authority, compare them with outcomes, review evidence, and expand authority only after institutional review. The page names a leadership and research/engineering roster: Lucas Zarzeczny (Founder & CEO), Clayton A. Cuteri (Chief Product Officer), AN Ali Navid (Chief Technology Officer), Spencer Ozgur (Quantitative Research), Ananth Seshadri and Anikesh Bhuvaneshwaram (Quantitative Development), and five software-engineering staff. These are first-party team-page claims; current employment and biographies were not independently corroborated in this pass.

The QSentia recruiting page labels the company pre-seed and advertises founding roles spanning quantitative development, data and research leadership, technology, and engineering. The advertised scope includes reinforcement-learning agents for allocation and position sizing, walk-forward and out-of-sample validation, transaction-cost and market-impact models, portfolio controls, agent monitoring, alternative-data research, data governance, broker/exchange integrations, and security. The role text also calls for controls against look-ahead bias, survivorship bias, leakage, overfitting, and unrealistic execution assumptions. This is a public specification of an intended governed agent workflow; it is not evidence of customer deployment, model weights, training data, data rights, live permissions, autonomous order authority, or audited performance. Full boundaries are in the QSentia capture note.

SEKUIA is a separate monitor-only route. Its careers page describes a small organization across Capital Solutions, Quantitative Research, Terminal, and Macro Reports, with roles for Quant Researcher—Macro Factors, ML Engineer—Forecasting, Risk & Scenario Analyst, product/platform engineers, and macro/editorial staff. It states an “AI as discipline” principle in which models, signals, and recommendations cite their reasoning, and lists Stockholm, Frankfurt, London, and EU-remote locations. The page footer calls the entity “SEKUIA AB” but displays Reg. 559-XXX-XXXX, which appears placeholder-like and was not independently verified. Accordingly, the record is retained as a macro-intelligence discovery lead, not as a confirmed hedge fund, asset manager, or current employer roster. No reviewed page discloses model families, providers, data rights, evaluation, customers, live capital, permissions, or performance.

These routes widen title-blind searches from “AI at a hedge fund” to governed investment terminals, shadow accounts, decision ledgers, macro-factor research, forecasting, and risk/scenario roles. They should remain separate from firm-level deployment evidence. See the SEKUIA/QSentia capture note.

August 28 AXQ Capital: Asia quantitative ML-infrastructure route

AXQ Capital is a named quantitative investment-firm route with a current firm-controlled Head of Machine Learning Infrastructure vacancy in Beijing or Shanghai. The posting reports to the CTO and describes GPU/CPU clusters, distributed training, research-compute and ML platforms, petabyte-scale feature engineering, tens of terabytes of training data, model versioning and deployment, batch and online inference, cost optimization, monitoring, access controls, and security. It also lists PyTorch Distributed, Ray, Slurm, NCCL, CUDA, RDMA, cloud GPU optimization, and Claude Code, Codex, and Cursor as relevant experience. AXQ’s board also lists algorithmic-trading, portfolio-management, financial-data, low-latency, production-engineering, and quantitative-development roles. This is a detailed hiring specification, not proof that the role is filled or that any named model or AI system is live. No reviewed source discloses AXQ’s model inventory, training corpus, data rights, evaluation benchmark, agent permissions, order authority, or AI-attributed performance. See the AXQ capture note.

Together with the already indexed Tensor provider route, this pass adds a regional infrastructure signal to the title-blind queue: search not only for “AI researcher” but also ML-infrastructure heads, distributed-training platforms, GPU-cost roles, real-time systems, model deployment, execution systems, and vendor-side agentic research. Keep firm recruiting, provider marketing, and live deployment as separate evidence classes. See the Asia AI-infrastructure capture note.

August 28 CFA and JM Finn title-blind investment-media routes

CFA UK’s May 7, 2026 webinar page adds a title-blind portfolio-construction route. It identifies Brian Pisaneschi, CFA, a Senior Investment Data Scientist at CFA Institute, as speaker and Elaine Xu, CFA, CAIA, as moderator. The advertised case combines factor analysis, economic-regime screening, sustainability scoring, retrieval-augmented agents, hybrid portfolio metrics, and agentic portfolio optimization. The page offers a gated recording that was not recovered in this pass, so it supports event and speaker metadata, not transcript-level claims or a named-manager deployment.

JM Finn’s March 18, 2026 Spring Investor Conference page supplies another first-party video index whose title does not advertise hedge funds or quantitative finance. The programme names J.P. Morgan portfolio manager Sam Witherow discussing energy, global supply chains, and AI, and Polar Capital partner Ben Rogoff discussing AI and technology disruptors. The two embedded assets—video 1 and video 2—returned cache misses. The page is retained as a speaker/date/topic and later-recovery route, not as evidence of either firm’s model, data, permissions, or performance. Full boundaries are in the CFA/JM Finn capture note.

August 28 Federated Hermes MDT and Lynx: two distinct research operating models

The first-party Federated Hermes MDT article, dated March 16, 2026, describes a daily quantitative portfolio-construction process through Daniel Mahr, CFA, Senior Vice President and Head of MDT Group. MDT says its workflow downloads vendor data overnight, recalculates company characteristics, runs a domestic-equity universe through a decision-tree model, feeds forecasts into a proprietary optimizer, and produces a trade list for human review. The optimizer combines alpha forecasts, hard risk constraints, statistical risk models for volatility and tracking error, and trading-cost models for market impact and liquidity. The team checks data accuracy and material news before review. This is conventional ML and portfolio-engineering evidence; the page does not claim GenAI or disclose model, feature, vendor, permission, or performance details.

The first-party Lynx Asset Management AI-Assisted Research page and its May 2026 PDF, authored by quantitative researcher Anton Holst, describe a separate GenAI research architecture. Lynx says it began serious development in early 2025 and built a centralized, structured knowledge base containing thousands of papers, reports, literature sources, and podcasts. Its described workflow pairs a human researcher with one or more agents across hypothesis formation, data discovery, code generation, and backtests; it also uses parallel agents for established-method review and recent-literature search. Lynx says its infrastructure can switch among models, that providers were pre-cleared by legal, compliance, and technology teams, and that domain-specific integration with data pipelines, market databases, backtesting, and reusable internal “skills” is the firm’s own development focus.

These disclosures should not be collapsed into a single AI category. MDT publicly describes a recurring ML forecast-to-optimizer process with data and trade-list validation. Lynx publicly describes a researcher-led GenAI implementation and knowledge workflow. Neither source identifies model versions, training corpora, data rights, benchmark fixtures, live user counts, agent permission maps, autonomous order authority, or audited AI-attributed performance. See the combined capture note.

August 28 Pensions & Investments: named asset-manager AI disclosures

The Pensions & Investments special report, dated June 16, 2025, adds a trade-publication survey route with several named investment leaders. P&I identifies Andrew Chin as AllianceBernstein’s chief artificial intelligence officer and reports his account of using LLMs to synthesize disparate company disclosures during the 2025 tariff announcement, reducing a process he said normally took weeks to days. The article also reports Chin’s description of NLP-generated equity signals in systematic strategies and his interest in reinforcement learning and eventual LLM-based time-series forecasting. These are speaker-attributed workflow and research statements, not an independent performance or deployment audit.

The same article reports Schroders’ descriptions of Genie, its proprietary AI assistant introduced in 2023, and ContextAI, a large-language-model system for sustainability-focused analysis of company reports and approximately 200 targeted questions. P&I also reports Schroders’ statement that ChatGPT Enterprise was being deployed across investment functions, that dozens of custom GPTs had been created and shared, and that they had been used thousands of times by investment decision-makers. Schroders further reported adoption above 80% across investment desks and an average saving of 50 minutes per day on research tasks. These figures are firm-reported through P&I; the PDF does not publish denominators, model versions, prompts, data-rights terms, audit logs, portfolio linkage, or performance.

P&I also identifies Raife Giovinazzo as FullerThaler’s managing partner and lead portfolio manager. His reported use cases are document summarization and surfacing common wisdom about companies or stocks with human verification. He frames a separate risk: generative models trained on human information may reproduce or amplify behavioral bias, which makes governance and verification relevant to a behavioral-investing process. Candace Shaw, identified as SLC Management’s chief operating officer and deputy CIO, describes a more administrative use boundary for private credit—meeting-minute summaries and document or dataset examination—while emphasizing relationship judgment and verification. These accounts add named personnel and dated operating boundaries; they do not establish production models, providers, permissions, or AI-attributed performance. See the capture note.

August 28 title-blind expansion: manager workflow, enterprise controls, and Boston quant history

The title-blind sweep recovered a MOI Global interview with Samir Patel, whom the publisher identifies as founder and portfolio manager of Askeladden Capital. Patel describes a three-stage research process, AI-assisted source finding and archive retrieval, thesis/watch-list comparison, and accounting investigation, while retaining source review and materiality judgment (09:36–14:57; 18:11–24:48; 45:56–52:01). He also names personal use of general-purpose model products, but the episode does not establish enterprise contracts, fine-tuning, a model registry, data rights, autonomous allocation, or AI-attributed performance. Askeladden is described here as a long-only public-markets manager, not treated as a hedge-fund deployment record. See the capture note.

A separate Odd Lots recording with Goldman Sachs CIO Marco Argenti provides an enterprise-control comparison point rather than a buy-side disclosure. Argenti describes a GSAI Assistant made available to approximately 47,000 people, data curation across hundreds of sources, a lakehouse/MCP path, agentic developer tooling, model gateways and token metering, provider-side forward-deployed engineers, and pull-request, security, technology-risk, and production gates (06:25–11:10; 19:52–26:51; 38:28–40:33). These are executive-reported bank-technology claims with automatic-caption navigation; they do not establish Goldman Asset Management or hedge-fund implementation, and the user figure is not independently audited. See the capture note.

The queue also recovered a historical Boston/MIT route: Sean Kruzel’s 2016 Boston Data Mining talk. The event description identifies him as Astrocyte Research’s founder and CEO, notes earlier quantitative global-macro and fixed-income-relative-value hedge-fund work, and links the talk to MIT mathematics and economics training. The recording’s public description and noisy automatic captions frame scientific practice, machine learning, trade sizing, fat tails, and strategy evaluation. It is retained as historical personnel and methodology evidence; it does not establish Astrocyte’s current status, a current model or data system, production permissions, or investment results. See the capture note.

The MIT CSAIL interview with Andrew W. Lo adds an academic route on AI and financial advice, analysis, and risk management. Although YouTube did not expose a local caption or audio path, a Podwise transcript mirror provides searchable timestamp navigation for financial-report analysis, hallucination and trust, risk assessment, sentiment, fraud detection, trading-algorithm testing, and regulation. The mirror is not treated as a first-party edited transcript or audio-verified quotation, and the route does not establish any hedge-fund deployment. See the capture note.

August 28 University of Chicago: academic research-agent feeder route

The University of Chicago Data Science Institute’s 2026 capstone page adds a concrete academic model route. Alejandro Canete Baez, Sebastian Rivera, Sami Naeem, and Campbell Taylor describe a fine-tuned LLaMA-based agent that combines transformer LLMs, external tools, agentic orchestration, retrieval-augmented generation, and domain-specific chain-of-thought instruction. The representative task is paper replication within a broader quantitative-research workflow involving literature interpretation, signal generation, and strategy validation. The page identifies Justin Kurland, PhD, as faculty advisor and describes him as a Goldman Sachs Engineering Division Vice President and Tech Fellow.

The page says the team presented the project in a linked University of Chicago recording at 1:47:20, but that Box asset was not retrievable in this pass. The university page’s statements about reliability and analyst-time reduction therefore remain project-page claims pending inspection of the paper, code, data, and recording. This route is evidence of an academic design and a finance-industry advisor affiliation, not evidence that Goldman Sachs or any hedge fund adopted the system, used its fine-tuning corpus, granted it trading permissions, or obtained audited investment results. See the capture note.

August 28 AllianceBernstein: current AI leadership and research controls

AllianceBernstein’s firm-controlled Andrew Chin biography and leadership page identify Chin as Chief Artificial Intelligence Officer and a member of the Operating Committee. The biography places him previously in Investment Solutions and Sciences and, from 2022 to 2023, as Head of Quantitative Research and Chief Data Scientist. A separate April 2024 AB research article names Yuyu Fan as Principal Data Scientist—Investment Solutions and Sciences and describes NLP applied to company documentation and earnings calls. The article discusses context-aware language analysis and changing management language, while stating that its views do not represent every AB portfolio-management team.

AB’s June 2024 prompt-engineering article describes role-specific context, iterative prompt refinement, step decomposition, and “step-back” clarification for LLM-assisted research. Its March 2026 hallucination-control article adds vetted-document boundaries through RAG, source citations, repeated or multi-model checks, explicit uncertainty handling, and expert feedback loops. The public material therefore exposes both a named AI remit and a control vocabulary, but not model providers or versions, training data, full feature or signal construction, evaluation fixtures, permissions, adoption counts, or AI-attributed returns. See the capture note.

August 28 China quantitative-research personnel route: PICC, Millennium CRTC, and SUFE

The public homepage of a China-based quantitative researcher, cross-linked to the researcher’s GitHub profile, adds a personnel and academic-lineage route not found by firm-name searches. The page describes a current quantitative postdoctoral role in an insurance-asset-management research department, a prior quantitative-research role at a Chinese fund, and a July–December 2023 junior quantitative-analyst role at Millennium Management’s CRTC. The employment history is self-reported and should not be treated as independently verified current headcount.

The same page describes earnings-call sentiment-factor work using FinBERT, FLANG-BERT, RoBERTa, and FinGPT, including comparisons among inference, fine-tuning, and pretraining-plus-fine-tuning approaches. It also describes later A-share stock research using text representations and graph neural networks such as GCN and GAT. The page lists publications on text classification, LLM representation post-processing, ChatGPT comparison/detection, and crowdfunding prediction. The official SUFE profile of Hailiang Huang confirms Huang’s professor, doctoral-advisor, and dean roles and describes big-data/AI research that includes quantitative-trading collaborations.

This route is useful for mapping model vocabulary, advisors, institutions, and career movement, but it does not establish employer adoption, reporting lines, model ownership, data licenses, point-in-time controls, independent replication, investment performance, or trading authority. Full source boundaries are in the capture note.

August 28 Edmond de Rothschild: Quartz quantitative platform and personnel lineage

Edmond de Rothschild Asset Management’s June 9, 2026 announcement describes the launch of the Quartz quantitative-equity range. The first three funds are Core EMU Equity, Core US Equity, and Defensive Global Equity; the announcement says the range was intended to comprise seven funds by the end of 2026, including additional Europe, emerging-market, global, and 3D Climate strategies.

The release identifies Bruno Taillardat as Head of Quantitative Management and says Yu Sun and Juan Sebastian Caicedo joined as quantitative analysts and portfolio managers in Paris, reporting to him. It also names Xavier Marconnet and Frédéric Girod as earlier additions in Geneva. The release describes value, momentum, quality, risk budgets, climate objectives, and market-environment adaptation. It says artificial-intelligence techniques are integrated to support diversification and responsiveness, while manager judgment remains central to defining and monitoring the approaches. The public announcement does not say whether the AI techniques are classical ML, neural, generative, or mixed.

The personnel biographies add prior quantitative-management lineage: Sun’s Amundi and Lyxor experience and Caicedo’s Natixis, Seeyond, and Ostrum experience, alongside École Polytechnique, EDHEC, Paris-Saclay, ESCP, and Universidad de los Andes training. This is a current platform, personnel, and stated-AI-scope disclosure. It does not identify model providers or versions, training data, features, evaluation fixtures, permissions, production endpoints, trading authority, or AI-attributed returns. See the capture note.

August 28 Maverick Capital: AI investment thesis versus internal AI disclosure

Maverick’s public-equities team page identifies Lee Ainslie as Managing Partner, Benjamin Silver and David Tykocinski as Co-Chief Investment Officers, and Jaimin Shah as Head of Quantitative Research. Its public-equities strategy page describes a fundamental long/short stock-selection process and a 20-year proprietary quantitative-research effort; Maverick’s home page dates that effort to 2006. These pages expose leadership and a quantitative operating surface, but do not identify an AI lab or a GenAI lead.

The June 18, 2026 Goldman Sachs Exchanges interview, recorded June 4, identifies Silver and Tykocinski as Co-CIOs of Maverick’s public-funds business. Tykocinski describes an investment thesis that AI-related value may migrate from training infrastructure toward infrastructure and application layers that deliver business productivity. He specifically discusses CPUs and databases as bottlenecks because AI agents are integrated into existing enterprise workflows and stacks. He says Maverick is focused on that migration. This is an investment view; the episode does not say Maverick uses those agents, databases, or models internally.

Maverick’s Ventures page separately says the platform brings quantitative and data resources and deep AI networks to portfolio companies. Its Silicon page describes private-company investing in compute-related chips, hardware, software, and processes. Those pages establish an AI-adjacent investment and network surface, not an internal model-training program or public-equities deployment. The capture note preserves the distinction and the negative evidence.

The public record therefore supports a named public-equities leadership team, a long-running quantitative-research function, and a current AI investment thesis, while leaving model families, data, providers, agents, permissions, internal GenAI adoption, and AI-attributed performance undisclosed. The Goldman transcript is publisher-provided and speaker-attributed; it is not an independent technology or performance audit.

August 28 Xantium: Tudor-affiliated quantitative hiring, personnel lineage, and an unresolved GenAI layer

Xantium’s first-party site describes the group as a Tudor-affiliated global multi-strategy investment firm. It says teams combine quantitative researchers, developers, fundamental analysts, and traders across public and private equities, futures, options, FX, bonds, and commodities; the site displays “200+ people across 4 offices” and dates the group’s founding to 2019 as part of Tudor. Its operating-model page describes systematic, quantitative, and discretionary/fundamental strategies, Bayesian statistics combined with human judgment, rapid experimentation, and short feedback loops. These are first-party organizational statements, not independently reconciled scale or performance data.

The Xantium team page identifies Oliver Watson as Chief Executive Officer and Sergey Levin as Chief Investment Officer, with Dmitriy Genkin and Christopher Voekler as co-founders and co-heads of volatility and nonlinear strategies. Their firm biographies expose a research and personnel lineage through D. E. Shaw, Harvard, Oxford, the University of Pennsylvania, Cornell, and MIT. Voekler’s biography specifically records prior research assistance at MIT’s Laboratory for Information and Decision Systems. These are displayed biographies and do not establish ownership of any undisclosed AI system.

Current quantitative-research hiring names financial and alternative-data analysis, research into new machine-learning techniques, quantitative signal/model implementation, modeling infrastructure, and production-trading support. The quantitative-developer posting names market-data delivery, integrated research and execution frameworks, cloud simulation, and Python/C++ across on-premises and cloud environments. A volatility-trader posting includes system monitoring and, for some roles, individual trade execution; an infrastructure posting covers compute, cloud, deployment, experimentation, and reliability.

This creates a new dated route for tracking Tudor-affiliated quantitative research, production-adjacent staffing, and academic lineage. The reviewed public pages do not name an AI lab, Chief AI Officer, GenAI lead, LLM provider, fine-tuning corpus, model weights, evaluation fixtures, data rights, research agent, or agent permission map. “Deep learning” in the team description and machine-learning work in a job description should not be converted into a claim of GenAI deployment. A related SEC Form D/A identifies the Xantium Partners Fund L.P. vehicle and Tudor Investment Corporation as general partner; it is a vehicle-level securities filing, not group AUM or technology evidence. See the capture note.

August 28 VolGAN: a researcher trail through Xantium, Balyasny, and Man Group

An Oxford researcher’s public CV records presentations of generative-finance work at Xantium Group in London on November 14, 2024, Balyasny Asset Management online on October 4, 2023, and Man Group in London on March 22, 2023. The same CV lists presentations at QuantMinds, J.P. Morgan, UBS, Société Générale, BNP Paribas, and academic venues. These entries establish a public seminar and discovery trail, not employment, collaboration, data sharing, adoption, or production use by the host firms.

The underlying VolGAN paper by Milena Vuletić and Rama Cont describes a generative model trained on time series of implied-volatility surfaces and underlying prices to generate joint scenarios. The authors apply it to SPX data and option-portfolio hedging. The public repository exposes training functions, arbitrage-penalty calculations, example code, and preprocessing guidance for OptionMetrics data; the preprocessed data are not shared. A follow-on data-driven hedging paper describes a January 2000–February 2023 SPX sample, a training cutoff in June 2018, later out-of-sample testing, and comparisons with delta and delta-vega hedging.

This is a materially different modality from LLM-based research assistants: the published artifact is generative financial ML for volatility-surface scenarios and hedging, with explicit no-arbitrage constraints. It gives the research program a concrete model-and-data route for derivatives, but it does not reveal a host fund’s model weights, data license, transaction-cost-adjusted live result, risk limits, execution system, or AI-attributed performance. Risk.net’s research profile identifies Rama Cont as Vuletić’s PhD supervisor and provides additional attributed discussion of the model; the paper and archive remain the controlling technical sources. See the capture note.

August 28 Xantium former-personnel route: scientific ML into quantitative trading and drug discovery

The public site of Shayan Iranipour describes a Quantitative Researcher role at Xantium Group from February 2022 to January 2025. He says he built algorithmic trading strategies and worked with time-series and statistical models on large, noisy, low-signal-to-noise datasets, using production Python code in mission-critical trading paths. The chronology is self-reported and the reviewed Xantium team page does not list him by name; it should therefore be treated as former-personnel evidence rather than a first-party roster entry.

The same public profile describes a 2025 move to Isomorphic Labs as a Senior Research Engineer working on machine learning for drug discovery. LinkedIn independently displays Isomorphic Labs and Cambridge. His academic record includes a Cambridge PhD thesis on precision QCD and effective field theories with machine learning, the JHEP deep-learning paper with Maria Ubiali, and membership in the NNPDF collaboration. These papers concern scientific inference rather than markets; they expose a research lineage involving neural-network parameterization, uncertainty-aware scientific modelling, and large collaborative data workflows, not a transferable trading recipe.

This is a useful title-blind talent and methods-transfer lead because it links a Tudor-affiliated quantitative organization to a scientific-AI laboratory through one public career chronology. It does not establish an Xantium AI lab, a Tudor–Isomorphic relationship, the researcher’s Xantium model or dataset, production permissions, performance, or any firm-wide conclusion. See the capture note.

August 28 Rock Bund Capital: crypto quantitative trading and explicit AI-direction hiring

Rock Bund Capital describes itself as a 2019-founded proprietary trading firm focused on digital assets across centralized and decentralized venues. Its first-party pages describe high-frequency liquidity provision, market making, statistical arbitrage, relative value, low-latency infrastructure, risk monitoring, stress testing, and scenario analysis. The firm’s home page also says its 2024 strategy set included machine-learning prediction and volatility arbitrage. Its displayed scale figures are company claims and are not reconciled here: the home page reports more than 100 employees, more than 1,000 symbols, and a $5 billion peak daily trading volume, while the current Greenhouse posting reports different volume and transaction figures.

The Shanghai-based Quantitative Researcher (AI Direction) posting names two research directions: deep-learning and representation-learning systems that create embeddings for complex market data, and AI-based alpha generation and strategy optimization intended to move research toward production monetization. It requests PyTorch or JAX, Python, optional C++, Transformer architectures or time-series foundation models, and the ability to implement work from major ML conferences. This is unusually specific hiring-intent evidence around market-data representations, foundation-model vocabulary, and research-to-production objectives; it is not evidence that the role is filled or that any named model is live.

Rock Bund’s News & Insights page separately reports a June 2025 strategic investment approaching $10 million in systematic multi-strategy firm GrandLine and a May 2025 integration with Hyperliquid. Those are ecosystem and connectivity signals. The reviewed sources do not show shared datasets, AI technology transfer, model providers, training corpora, permissions, or performance associated with either relationship. See the capture note.

August 28 Cedalion Capital: named ML personnel and an academic time-series lineage

Cedalion Capital describes itself as a Netherlands-based investment fund using machine-learning algorithms, natural-language processing, and machine-learning specialists in its investment process. Its public page does not identify a full team, model inventory, provider, dataset, regulator, or evaluation record. A current ACMAD profile and Wessel Bruinsma’s public CV identify Bruinsma as a quantitative researcher at Cedalion from July 2025, with the CV describing development of investment strategies with machine learning. Because the role evidence is primarily profile-attributed rather than a named Cedalion team page, it is retained as a medium-confidence personnel link.

The same CV records Bruinsma’s Cambridge PhD under Richard E. Turner, prior machine-learning research at Invenia Labs, a quantitative-research internship at G-Research, later work at Microsoft Research and the Alan Turing Institute, and a March 2026 Cambridge research-lead role. Cambridge’s ML Group page corroborates the doctoral supervisor and publications on scalable multi-output Gaussian processes, convolutional conditional neural processes, and Gaussian-process autoregressive regression. These works make the route relevant to probabilistic modelling, uncertainty, and multi-output time-series research, but they do not identify Cedalion’s implementation.

Bruinsma’s Microsoft-era Aurora work is a separate environmental foundation-model route. The public Aurora repository describes atmospheric, air-pollution, and ocean-wave forecasting, and explicitly excludes non-environmental prediction and direct operational decision-making without expert review. No transfer of Aurora, its code, or its training data into Cedalion is inferred. See the capture note.

August 28 Brooklyn Investment Group: explicit GPT-4o portfolio monitoring, then a Nuveen transition

Brooklyn Investment Group published a 2024 white paper on generative AI for portfolio monitoring in personalized direct-indexing accounts. The described system predicts whether an account may need rebalancing and why, so that portfolio managers can focus optimization work on accounts requiring attention. The paper reports firm-generated estimates of 82% portfolio-manager time savings and 63–85% net compute-cost reduction; these are controlled-circumstance, self-reported results rather than an independent audit or investment-return result.

The paper names OpenAI GPT-4o as the selected model, with GPT-3.5-turbo and GPT-4o-mini as baselines. It describes zero-shot and multi-shot evaluation, separate prompts for each possible reason to trade, prompt inputs excluding client-identifying information and security-level holdings other than cash, and human portfolio-manager review. The evaluation used historical and live account-day samples and reported balanced accuracy, F1, precision, recall, and self-consistency. The paper also records failure modes around small percentage comparisons and multiple independent conditions, with basis-point conversion and prompt decomposition as mitigations.\n+\n+The Form ADV separately describes BKLN as an investment-advisory affiliate using BAIR/Skopos Labs ML and NLP software under a royalty-free licensing arrangement, with direct-indexing, automated rebalancing, and thematic-index workflows. Brooklyn’s research archive now states that BKLN merged into Nuveen Asset Management effective May 1, 2026. The public record does not establish that the GPT-4o workflow continued unchanged after that reorganization. This is a high-detail asset-management control case, not a current standalone hedge fund and not evidence of autonomous trading. See the capture note.

August 28 Presto Labs: predictive ML, simulation, and execution controls without a disclosed GenAI layer

Presto Labs describes a quantitative-trading firm established in 2014 with data-driven research, proprietary automated execution, post-trade analysis, and separate HFT, portfolio-management, and market-making contexts. Its first-party operating description says data are sanitized and cross-validated before research; predictive models and strategies are made verifiable; in-house simulators are compared with real trading; and orders are monitored within agreed risk parameters.

The same public page displays personnel narratives for Zijin Kong, Jingyuan Zhou, Yuxuan Pan, and Hyungjun Kim. Zhou describes creating, evaluating, and improving strategies as a quant developer. Pan describes an algorithmic-trader and trading-team-lead role. Kim describes a KAIST doctoral and master’s background and a hypothesis–experiment–analysis process for trading strategies. These are first-party personnel narratives, not a complete roster or AI-lab structure. The reviewed sources disclose predictive-model and research-to-production vocabulary but do not name an LLM, GenAI system, foundation-model provider, agent, fine-tuning corpus, or AI lab. See the capture note.

August 28 Korean AI investment-platform routes: Quantit and Qraft

Quantit adds a Korean-language route that connects AI-assisted analysis, backtesting, strategy implementation, automated orders, portfolio construction, and risk management in one first-party product description. Its solutions page displays customer labels spanning the National Pension Service, Korean banks and securities firms, Samsung and Hanwha Asset Management, NH Hedge Asset Management, CK Goldilocks Asset Management, and other regional institutions. The page associates those labels with different products, including multiple financial-AI models, direct indexing, financial-risk early warning, portfolio trading, automated institutional portfolio trading, and an HFT order-strategy system. Quantit also displays KAIST and Seoul National University as financial-engineering model-research partners and NYU Stern as a financial-investment AI-agent research partner. These are Quantit’s own displayed customer and partner claims; the reviewed pages do not establish contract terms, shared datasets, model versions, deployment dates, client permissions, or fund-level outcomes.

Quantit’s careers page names Duck Hee Han as CEO and Junbok Lee as Technology Research Institute Lead. It also describes “Captain-Q,” an internal chatbot initially built with GPT to search a Confluence-based internal wiki. The public page does not specify its current model, retrieval benchmark, tool permissions, or use in investment decisions. This is a concrete internal GenAI workflow signal, but not evidence of an autonomous trading agent.

Qraft provides a separate Korea-founded AI asset-management and ETF route. Its first-party page describes deep neural networks, AI-determined security selection and weighting, human supervision of product design and controls, and intellectual property spanning trading and order-execution tools, asset-price prediction, and portfolio generation. The page’s risk disclosure says the funds rely heavily on a proprietary AI selection model and third-party data; its glossary identifies the Kirin API as integrating multiple vendors for point-in-time macroeconomic and company-fundamental data. A historical Qraft article names Marcus Kim as founder and CEO and describes point-in-time factor-search work. These are firm-reported ETF and technology disclosures, not proof of hedge-fund deployment, current model versions, data licenses, autonomous authority, or AI-attributed returns. See the capture note.

August 28 Japanese AI and quantitative-investment routes: Mizuho First Financial Technology and Sumitomo Mitsui DS

Mizuho First Financial Technology provides a Japanese-language quantitative-research and investment-advisory route with explicit technical vocabulary. Its equities page describes big-data and AI-enabled investment-method development using statistical analysis, machine learning, and multilayer neural networks, supported by GPU, Hadoop, and cloud environments. A related asset-allocation page describes machine-learning estimates of expected returns and risk for equities and foreign exchange, dynamic allocation, scored news as an explanatory variable, and a random-forest asset-allocation model. These are first-party research and service descriptions, not evidence of a hedge-fund deployment, named model version, data contract, or performance attribution.

Sumitomo Mitsui DS Asset Management adds a named Japanese quant organization. Its page identifies Takahashi Katsunori as group head from 2026, describes a nine-person investment group and 13-analyst investment-development group, and reports approximately ¥3.69 trillion in assets under management as of March 31, 2026. Its 2025–2026 sustainability report says an in-house AIR AI-support tool was released in October 2025 after roughly two years of development and made available to all investment personnel. The report describes RAG over internal databases/files plus current web/news and agent functions for long reports, fact checking, and trade proposals; it also displays named quantitative-analyst and group-head roles and estimates approximately 50,000 annual hours of investment-department efficiency improvement. This is a company-reported internal GenAI workflow and efficiency estimate, not independent measurement or evidence of autonomous execution, model-provider choice, permissions, or AI-attributed returns.

The capture note records the Japanese-language discovery route and keeps sector-context material separate from named-firm deployment evidence.

August 28 Brazilian Portuguese quant and AI-investment routes: Sarpen and AZ Quest Bayes

Sarpen Quant Investments provides a new Brazilian Portuguese first-party route. Its site describes a São Paulo asset manager that combines artificial and human intelligence, proprietary algorithms, and high data-processing capacity to extract information, develop analyses, and construct systematic investment strategies. The Sintropia Multimercado FICFI page identifies Sarpen Quant Investments LTDA as manager, Intrag DTVM as administrator, May 26, 2023 as the fund start date, and a policy spanning fixed income, equities, foreign exchange, and derivatives. The site’s description remains at the process level: it does not name model families, providers, training data, alternative-data vendors, evaluation splits, or live permissions.

Sarpen’s June 2026 Portuguese post identifies Sérgio Werlang as a founding partner and former Central Bank director and links his monthly Broadcast+ article. That establishes a named personnel and publication route. It does not establish that his macro commentary is an AI-generated signal or that the fund uses any particular model. The capture note preserves the Portuguese wording and the unresolved technical boundaries.

It is important to separate XP Investimentos from the managers described here. XP is a Brazilian financial-services and brokerage group, and in this route it is the publisher of the survey and the distributor-side source—not a hedge fund. XP’s April 6, 2026 page says its survey covered 71 Brazilian asset managers representing half of the Brazilian fund market, and reports 97% using AI in practice and 58% using it daily. Those are XP-reported survey results, not an independently audited census, and the page does not identify the participating firms. It also links a YouTube episode discussing AI use for reports, company results, news, central-bank speeches, macroeconomic data, and asset prices. The Portuguese auto-caption track was retrievable, with the survey-size discussion around 01:17; this episode therefore did not require local ASR recovery.

XP’s June 29, 2026 case study describes Bayes Capital Management, whose strategies are distributed under AZ Quest Bayes after a 2022 strategic agreement. The article attributes to Bayes a systematic platform with automated market-data capture, corporate-event handling, order execution, and portfolio rebalancing; long-only systematic, long-biased, long-short, and global-macro strategies; and a proprietary library of fundamental, trend, market-microstructure, macro, and risk indicators.

A separate Money TV / Market Pillars Uncovered playlist expands the India-facing title-blind route. Recovered Hindi automatic captions identify episode-level public presentations for MavenArk’s AI-driven wealth-technology platform (01:32–02:09), Savart’s AI-supported investment-advisory framing (01:20–01:59), and Wright Research’s quant/robo-adviser positioning (00:53–01:17). These are adjacent asset-management and advisory disclosures, not evidence about the Brazilian firms or about model performance. The captions are automatic and were not manually reconciled against audio. See the India regional source note.

The same XP article describes a multi-agent code workflow: independent agents, including models from different suppliers and proprietary systems, produce candidate solutions; another AI compares divergences and assigns a confidence score; and a manager makes the final decision. The public case study does not name suppliers, model versions, prompts, evaluation fixtures, repository controls, or authorization boundaries.

The article also attributes a generative-AI research workflow to Bayes dating from 2021. It says the manager processes more than 20,000 earnings calls, converts executive speech into embeddings, and examines tone, recurring topics, and narrative changes. The source presents this as a possible future strategy input, not as a published live book with timing, portfolio mapping, or return attribution. It does not expose the transcript vendor, language coverage, data rights, point-in-time controls, speaker normalization, model revision, or out-of-sample results.

Finally, XP describes a proprietary-transformer project intended to represent structured and unstructured information together as company “mathematical signatures,” alongside an allocation system incorporating portfolio risk, liquidity, and economic regime. These are source-attributed development objectives. No weights, training corpus, test results, deployment date, or completion evidence is public in the reviewed case study. This route adds useful evidence about the kinds of workflow and data representations being discussed in Brazilian systematic management, but it does not support a ranking of firms or a conclusion about model quality.

The capture note records the Portuguese source route, XP/Bayes attribution boundary, and the distinction between manager claims and independently verified evidence.

The French-language Les Investisseurs 4.0 episode with Franck Béon supplies a separate title-blind personnel and methodology route. The recovered public audio identifies Millennium Capital Partners as a prior quantitative-hedge-fund employer (07:15–07:39) and describes algorithmic/arbitrage work, reference-data and market-connection construction, and an ETF-relative alpha-analysis workflow using factor models, Fama–French-style models, and Random Forest (05:08–07:39; 17:32–19:02). Ploovers’ first-party team page independently identifies Béon and Marc Rousseau as co-founders and describes their institutional backgrounds; public profiles also name Millennium among Béon’s prior employers. The podcast and profiles do not establish a current hedge-fund operation, live model authority, training corpus, vendor stack, or independently verified performance. See the French ASR recovery note.

August 28 Spanish AI-in-finance route: AthenAI and Value School

Value School’s February 23, 2026 episode, “Selección de algoritmos de inversión con Inteligencia Artificial,” is a new Spanish-language title-blind route. The episode features Guillermo Meléndez, whom AthenAI’s official origins page identifies as founder and CEO. AthenAI is an education and technology institute, not a hedge fund; this route is included as adjacent methodology, talent, and market-infrastructure evidence.

The publisher page exposes an audio enclosure but no transcript. The audio was recovered from the Libsyn route and processed locally with Spanish MLX Whisper using word timestamps. The run produced 4,478 timestamped segments over approximately 6,037 seconds. These are local offsets against the recovered audio, not publisher-provided transcript links; the full transcript remains private and is not reproduced here.

The recording presents a broad menu of financial-AI techniques, including genetic and swarm algorithms, fuzzy logic, machine learning, deep learning, reinforcement learning, and hybrid quantum/classical methods. Around 04:45–05:09, Meléndez distinguishes generative AI from the wider set of financial-market services and automation. Around 17:37–21:21, he discusses execution, news-impact estimation, strategy and portfolio rebalancing, fund-risk thresholds, market-abuse detection, and automatic model/data-quality checks. Around 55:56–58:58, he describes agentic workflows for classifying news, assigning events to strategies, rebalancing strategy sets, and allocating capital through combinations of fuzzy logic, reinforcement learning, and genetic algorithms.

The episode description frames a challenge over more than 14,000 investment algorithms. Around 67:56–70:56, the recording discusses evaluating a supplied algorithm set using risk, trading frequency, capital allocation, and style. The public materials do not expose the underlying algorithm set, labels, data provenance, reproducible benchmark, or a current hedge-fund customer. The algorithm-selection language is therefore useful for identifying candidate financial-AI modalities and control points, not evidence of a named fund’s production system.

AthenAI’s public biography materials describe Meléndez’s prior BME Innova and AI-in-finance work and list postgraduate training in quantitative finance, financial markets, data science/big data, and deep learning. The team page adds public AI/data and quantitative-development affiliations, and the finance competition page names Marcos Aza of Santander Asset Management on an evaluation committee. These are self-published biographies and program pages, not independently verified academic records, peer-reviewed research, or evidence of a tracked hedge-fund employment relationship. See the capture note.

August 29 title-blind expansion: Talking Market Data

The Talking Market Data YouTube channel and Buzzsprout RSS feed add a useful title-blind discovery surface. A channel scan and alternate-video recovery produced automatic English captions for five previously untracked episodes: Mike Salk, Oracle speakers James Calise and Jason Murphy, AiMi’s Ollie Cadman, Proton Advisors’ Brennan Carley, and LSEG’s Debbie Lawrence. These are vendor and market-data industry routes, not proof of a hedge fund’s internal system.

The recovered material consistently points to the control layer around financial AI: MCP and data-vendor connectivity, data quality and metadata, model-context preparation, domain-specific notification handling, data lineage, licensing, and intellectual-property boundaries. The Oracle episode’s chapters explicitly index AI in market-data infrastructure, model/cloud relationships, deterministic versus probabilistic systems, and data sovereignty (23:35; 32:12; 34:35; 37:34; 42:24; 46:59). The LSEG discussion provides timestamp navigation for data rights and lineage (05:03; 08:28) and for the distinction between model training and governed contextual access (25:50–29:20). These timestamps are based on RSS chapters and automatic captions; they are not treated as edited transcripts.

This route adds a practical search rule: scan full feeds and channel uploads for data, cloud, infrastructure, licensing, notification, and governance terms even when the episode title does not say hedge fund, quant, or AI. The capture note records caption hashes, recovery failures, timestamps, and non-claims. No reviewed episode establishes a tracked fund’s customer relationship, model inventory, data rights, production permission, investment authority, or performance.

The second recovery wave adds Nimesh Bharadia, Tim Baker’s earlier Blue Sky Nexus episode, Leigh Walters, and AMD’s Alastair Richardson. Their automatic captions add navigation for Asia-Pacific and model-explainability discussion (29:44–42:09), large-language-model and identifier/symbology context (30:00; 60:14), agentic automation and data-readiness constraints (16:12; 20:34; 44:08), and hardware-local inference and latency constraints (11:45–12:28; 32:22–32:44; 51:30). These are useful context and discovery links, not claims about a tracked manager’s live AI stack.

The five formerly captionless clips add a more concrete operating lens, but their evidence is local ASR and has not been audio spot-checked. Tim Baker recounts a small filing-difference workflow that used the OpenAI API to compare new and prior EDGAR filings, describing roughly $20 of API spend and about 30 hours of work by a summer intern (00:05–01:10). He then connects coding assistants to faster construction of new datasets from unstructured documents (01:59–03:34). Nimesh Bharadia gives generalized examples of KYC process automation, alternative-data commodity-price analysis, fraud detection, and AI-based wealth-management workflows, including a speaker-reported estimate of reduced manual intervention that is not tied to a named organization (00:31–02:03). These observations suggest search avenues for workflow automation and dataset formation; they do not establish a tracked fund’s implementation, economics, or performance. The updated capture note records the private ASR archive and its evidence limits.

An alternate-upload pass also recovered Himanshu Gupta’s Solace episode and a second caption route for Tristan Dehaan. The Solace recording adds event-broker, Kafka, cloud, latency, analytics/model, and automation context (06:13–10:31; 39:18–41:46). The Dehaan recording adds market-data cost, standardization, and automation context, but the inspected captions did not yield a substantive AI-system disclosure. A second AMD upload was retained as an alias rather than counted as another episode. The capture note records the hashes and duplicate-upload boundary.

August 31 regional-language title-blind additions: Iberian and French media

A Spanish-language Renta 4 Podcast episode, published August 28, 2026, identifies Natalia Aguirre and César Sánchez-Grande as co-directors of Analysis and Strategy. Its chapter list explicitly separates the team’s AI-use discussion (16:27), AI limits (18:36), and efficiency/productivity (20:40). This is a first-party investment-analysis media route and a useful lead for the underlying recording. The page does not identify a model, provider, data license, agent permission, production endpoint, or AI-attributed performance.

The Portuguese Think Tank Carregosa episode, published August 26, 2026, names Miguel Ricon Ferraz as Financial Analyst and head of the bank’s investment-advisory service. It also explicitly says the short episodes are created from specialist articles with podcast voice and AI assistance through NotebookLM. That is direct evidence of a generative-audio publishing workflow. It does not establish NotebookLM use in portfolio research, model training, trading, or capital allocation.

The French BFM Bourse episode identifies Cyrille Collet as director of quantitative equity management at CPR AM and Olivier de Béranger as CIO and head of asset management at La Financière de l’Échiquier. Its public transcript is timestamped and provides a new French-language personnel and market-context route. The reviewed episode does not disclose CPR AM’s model inventory, GenAI tooling, training data, agent permissions, or performance attribution.

The Spanish Barcelona Finance School episode features Pep Martorell, described as a former associate director of the Barcelona Supercomputing Center and current investor/adviser in AI projects. The page frames finance-sector AI around non-determinism, explainability, production use, and organizational constraints. It is an adjacent methodology and talent route, not evidence of a current hedge-fund affiliation or deployment. The regional capture note records the source boundaries, the private BFM page capture, and the unresolved audio/transcript recovery queue for the other three routes.

The same non-English pass surfaced two additional adjacent routes. Italy’s RadioBorsa episode, published July 1, 2026, is authored by SoldiExpert SCF and describes a proprietary active strategy named “Momentum Expert IA & Humanoid Robotics.” The Spreaker page confirms the publisher organization. This is a product and wealth-management lead, not a hedge-fund disclosure; no model, data, evaluation, or performance details were admitted from the public description.

Germany’s Alphawave adds a first-party proprietary-trading and media surface. The firm describes a Düsseldorf-based systematic trading business with automated strategies, internal infrastructure, and model-validation materials; its technology page describes short-horizon price-inefficiency trading and dynamic risk management, while its investor-relations page lists dated corporate, webinar, and model-related materials. These are company claims and should remain separate from independently verified research. The reviewed pages do not disclose model weights, training data, AI/GenAI configuration, agent runtime, or external-fund permissions. The regional capture note records the classification and follow-up capture queue.

August 29 title-blind adjacent route: Foresee Markets

The Foresee Markets podcast was found through generic investing and AI/quantitative searches. Its public Riverside RSS feed contained 23 episodes on August 29, 2026, each with an audio enclosure and a publisher-provided transcript URL. Selected episodes describe sector-relative stock selection, automatic exits, short-horizon cross-asset signals, options overlays, and a planned interface for themed long/short portfolios. Targeted local ASR now adds timestamp navigation for Episode 4’s AI/ML selection and universe filtering (00:46–01:31; 04:08–05:50; 06:12–07:01) and Episode 17’s platform/allocation workflow (02:31–05:21), with planned liquidity-provider connectivity at 07:53–08:00.

This is adjacent-platform evidence, not evidence about GMO, Acadian, Arrowstreet, or another tracked hedge fund. The show’s backtest, return, accuracy, and automation figures are speaker-reported; the reviewed feed does not provide code, complete point-in-time universes, transaction-cost assumptions, broker records, model weights, or independent performance verification. The capture note records the selected transcript hashes, targeted local-ASR hashes, timestamp basis, and evidence boundaries.

August 30 regional title-blind follow-up: UK, Canada, and Australia

The new regional pass adds several useful routes that do not depend on a firm publishing an “AI” label. BlackShip Capital’s careers page describes London systematic-volatility research and engineering roles using supervised, unsupervised, and reinforcement learning, PyTorch, TensorFlow, kdb+/q, custom HPC, and real-time data pipelines. Its employer LinkedIn page uses the phrase “AI-native execution,” while Gregory Rubio Schmidt and Felix Hemmerling provide public founder and education routes. The public materials do not identify an LLM, model provider, fine-tuning corpus, model owner, or autonomous-execution permission map. The phrase remains a company claim, not a technical disclosure.

The same search found UK control cases. Maven’s systematic division and experienced-hire page describe shared research and execution infrastructure, computational scale, and multi-terabyte data; public profiles for Stefan Ivanov and Denis Glazachev expose relevant scientific-ML backgrounds without proving firm use. Westren Capital’s quantitative-research role names regression, tree models, and neural networks. EverestQuant’s legal record, careers page, and self-reported company announcement add a London/Budapest entity route, while Genesis Quant Capital’s legal record, firm pages, and Jinmin Cai’s profile add an early-stage algorithmic-trading and AI/ML lead. None of these routes establishes a current model inventory, filled role, production GenAI system, or performance result.

Canada contributes a particularly clear separation between research-agent disclosure and manager classification. Blackmark Dominion’s engineering page describes “Cody,” a local coding-agent system with manager/worker routing, persistent project memory, tool execution, testing, verification, provider failover, and stuck detection. Its verification page describes “NEMO” as a competing-agent/prover/skeptic layer, while the Azar research page describes held-out evaluation, placebo tests, and leakage checks. The firm explicitly says it does not manage client capital or solicit investment, so it belongs in the research/prop-adjacent control set rather than the external-manager set.

Pointus’s careers page combines trading-platform support with “AI-driven automation initiatives,” and its leadership page identifies Brendan Magill as COO responsible for operations and technology. A historical Tick Data case study describes an AWS backtesting architecture using global multi-asset tick data and distributed searches. The current AI language is hiring intent; the 2015 architecture should remain explicitly historical. DeepAim’s join-us page describes anticipated founding quant/engineering roles and “OMOS” trading models, but no filled role, AUM, named architecture, or live system is public. RBC Borealis and its AI research program add a bank-laboratory control: Brian Keng, Afsaneh Fazly, and an RBC release connect Canadian academic and financial-AI research, not hedge-fund alpha.

Australia adds operating-model and alternative-data routes. Systematic Trading Associates describes automated futures trading, machine learning, backtesting, and low-latency execution; Hugh McGuire’s STAC Summit route adds a September 10, 2026 conference checkpoint. Geometrica’s investor deck describes proprietary ML/deep-learning models and identifies James Bradley as Head of Data & Analyst; a 2025 podcast discusses WeChat and port data. These are manager-reported process and alternative-data descriptions, not disclosures of LLMs, agents, or autonomous portfolio decisions.

Akuna’s Sydney ML role and employer post name neural networks, tree models, ensembles, execution metrics, and portfolio optimization; Marc Verdiel’s profile is a personnel lead, not proof of a filled role. VivCourt’s careers pages describe automated strategies, large datasets, and ML techniques, while its academic-to-quant profile supplies a research-lineage route. The GroWise PDS provides vehicle and strategy facts, and Tibra’s careers page provides a systematic quant role family; neither source discloses a GenAI system.

These routes expand the map across three recurring layers: research agents and verification, conventional ML and alternative-data research, and infrastructure or hiring language that may precede a named AI program. The evidence does not support a ranking or a conclusion that any named firm is further along. The complete regional note, entity classifications, source tiers, and next-step recovery queue are in the regional capture note.

---

State of AI | August 2026

September 1 title-blind Man Group insight refresh

Three additional Man Group first-party routes were recovered through a title-blind insight-index search: Something Has Got to Give (Eventually), The Productivity Paradox: When Will AI Deliver?, and H2 2026 Credit Outlook: The AI Buildout — Boom or Bubble?. They add public discussion of AI-driven technology-sector dispersion, research productivity, the possibility of “digital researchers,” and AI-related credit-market risks. These are firm-authored investment commentary and research framing, not proof that Man operates autonomous researchers, uses a specific model or dataset, or attributes returns to AI. The current Chao Xia profile separately documents her Man Numeric research-leadership role and listed interests in machine learning, artificial intelligence, alternative data, event strategies, and macro timing. The detailed source note is here.

Source status: Three first-party Man Group insight pages and one current personnel profile were checked on 2026-09-01. Source files: sources/13-multimodal-sources/man-group-title-blind-insights-2026-09-01-raw.md; sources/13-multimodal-sources/media-discovery-coverage-2026-08-18.json

September 1 title-blind BlackRock Systematic refresh

Three additional first-party BlackRock Systematic routes were recovered through a title-blind strategy and insight search: the current US systematic-investing page, the new-frontiers-in-data page, and the 2026 Thematic Mid-Year Update. The first two describe combinations of alternative data, electronic text, analyst and management views, machine learning, LLM-assisted text and thematic research, and human investment expertise; the third adds a current public video route around AI-related thematic market analysis. Company-displayed scale figures and method descriptions are recorded as first-party claims. None of these pages discloses model weights, training/evaluation splits, data rights, agent permissions, deployment dates, or AI-attributed performance. See the source note.

September 1 title-blind personnel and research refresh

The latest title-blind pass added four bounded routes. Citadel’s Hua Zheng profile provides a current Head of Market Impact Research route and describes price formation, counterparty response, execution-cost modelling, and interdisciplinary research teams. GMO’s 2025 Systematic Equity year-end letter names machine-learning sentiment measures, network-aware momentum, corporate-behavior alerts, AI-based analytical tools, and uncertainty estimates in its public research agenda. D. E. Shaw Research’s technology page adds a distinct biomedical-computation route involving Anton, simulation-generated neural-network data, compound search, and physics-model refinement; it is kept separate from D. E. Shaw Group investment evidence. A Marshall Wace quant-investing insight day adds a non-AI-titled recruiting and research-implementation route. These sources establish public role, method, research, or recruiting descriptions only; they do not establish model ownership, data rights, production status, agent permissions, or performance. See the source note.

September 1 title-blind Point72 and Millennium media cross-check

The Point72 Academy announcement and official podcast catalog add an official, non-AI-titled talent and operating-model surface: STEM-to-investing training, analyst hiring and internship design, international Academy offices, analyst/PM team formation, and curriculum. The catalog identifies Jaimi Goodfriend as Academy director and expressly restricts copying or distribution without prior written consent. It is therefore used here as metadata and discovery evidence, not as a transcript source.

The Matthew Granade publisher page and Michael Recce episode add two historical Point72/data-science routes. The publisher description for Granade records a chief market-intelligence role, central-portfolio and Point72 Ventures responsibilities, and stated venture interests including enterprise automation and AI. Recce’s timestamped private ASR records a historical 20–30-person Point72 data-science-team plan and a roughly $5–6 million Neuberger Berman data-science-team estimate (00:13:27–00:13:39; 00:35:12–00:35:20). Both are historical/publisher or practitioner evidence, not current firm budgets, model inventories, data licenses, permissions, or performance.

Millennium’s Ross Garon mChats page provides a first-party non-AI title route. Garon is identified as Global Head of Quantitative Strategies and describes the researcher/developer split, high-throughput research and trading infrastructure, and firm-provided data, risk models, and execution algorithms. This improves personnel and operating-model coverage without supporting a conclusion about a specific AI or GenAI system. See the capture note.

The official Millennium Careers YouTube channel adds a stable video route that the page-level search did not classify separately. Its episode list includes mChats with Gideon Mann, Vlad Torgovnik, Ross Garon, and Scott Rofey, plus a short “Find Your Path: Anil” clip. YouTube exposed no captions for the checked videos, so local timestamped ASR was recovered privately. Four clips were too sparse or noisy to support substantive transcription; the Anil clip provides only a short statement that the speaker runs the systematic-data platform team and chooses technologies according to each problem. Because the title does not provide a surname, it is not resolved to Anil Chandroth without an independent first-party identity link. This closes a firm-controlled YouTube discovery route while preserving transcript-quality uncertainty; it does not create a podcast record or disclose models, training data, permissions, deployment, or performance. See the capture note.

September 1 title-blind Algert and Neuberger Berman cross-check

The North Square April 2026 Algert Global episode and its publisher transcript identify Ryan LaFond as Algert Global’s Co-Chief Investment Officer. LaFond describes machine learning for combining signals, LLMs for text features including whether management answers questions asked, hundreds of insights evaluated across thousands of stocks, and proprietary risk and cost models in optimization. This is first-party, dated process evidence; it does not disclose model vendors, training data, evaluation design, data permissions, agent authority, or AI-attributed performance. The capture note records the source hash and retention boundary.

The first-party Neuberger Berman quantamental transcript identifies Tim Creedon, Ray Carroll of Breton Hill, and host Anu Rajakumar. Its historical discussion describes a joint fundamental/quantitative process, roughly 6,000 tickers processed overnight, an integrated central-research data-science team, and use of credit-card, web-browsing, and job-posting data to test investment theses. The PDF footer is dated 2019, and the local fetch was rate-limited; no current organization, model inventory, or durable private transcript checkpoint is claimed. See the capture note.

September 1 Systematica official-surface and historical AI-talk refresh

Systematica’s current Our Firm page describes alternative data, proprietary technology and trading infrastructure, electronic trading, and an investment organization divided across Research, Technology, and Trading, with Product Management providing coordination. Its Culture page adds a public Data R&D route and identifies quantitative-analyst employee stories, while the careers page describes data-driven decision-making and a six-office recruiting footprint. These are first-party operating and recruiting surfaces; they do not identify an LLM, agent, training corpus, model provider, or AI-attributed investment result.

The historical QCon London presentation page identifies Antoine Pichot as a Systematica quantitative researcher and describes an AI-in-asset-management talk, automated stock-market algorithms, and a Telecom ParisTech post-doctoral research route in hidden Markov models. The page is useful for date-scoped personnel and academic lineage, but it is not a current CV, transcript, or disclosure of Systematica’s present AI architecture. The capture note records the boundaries.

September 1 Winton firm-media and academic-partnership refresh

Winton’s current official news index exposes a firm-controlled media route with 2026 updates, while its dated research pages provide unusually concrete methodological context. The quantitative-investing article says machine-learning methods can support slower strategies when the data task is large, using company-report text and long backtests as an example; it also discusses hypothesis-driven research, selection bias, multiple testing, and backtest-to-live degradation. The equity-classification article describes covariance clustering and NLP over annual reports as alternative ways to group companies and inform sector or risk views. These are dated firm-authored methods, not a current model catalog or performance record.

The differential-privacy project names John Funge from Winton and a UC Berkeley group led by Dawn Song. Winton describes an exploratory architecture in which privacy-preserving SQL queries ran against location data held by a third-party vendor and returned aggregate results, and explicitly says it had no plans at that time to use location-derived data in investment algorithms. This is a valuable academic-partnership and sensitive-data-governance route, but it does not establish current data access, trading deployment, GenAI, agent authority, or AI-attributed results. See the capture note.

September 1 XTX official firm-media refresh

XTX’s current official home page says the firm uses machine learning to produce price forecasts for more than 53,000 financial instruments and uses those forecasts in exchange and alternative-venue trading and liquidity provision. The page also displays self-reported figures for research-cluster GPUs, storage, daily traded volume, countries, and employees. Its careers page adds the research-to-production boundary: quantitative researchers design statistical models, while exchange-trading developers move prototypes into high-throughput tick-to-trade systems and trading analysts supervise automated activity. These are direct current firm statements, with scale figures treated as dated self-report rather than independent audit.

The news index is a useful title-blind discovery surface spanning XTY Labs, Formal Frontier, Project Numina, mathematics funding, technical infrastructure, and external coverage. The separate XTX Ventures page describes AI/ML founder investing, technical validation, infrastructure scaling, architecture guidance, code review, and an engineering network. XTX Ventures is kept separate from XTX Markets trading-system evidence: the reviewed pages do not establish that a portfolio company supplies the market-making business or that venture support became an internal dependency. The current official surfaces still do not provide a complete LLM/GenAI catalog, training corpus, agent permission map, or AI-attributed result. See the capture note.

September 1 Schonfeld AI implementation and conference refresh

Schonfeld’s current AI Strategy Analyst listing places a dedicated role in the DMFI COO Office at the intersection of investment workflows, data strategy, and applied AI. The advertised mandate covers end-to-end implementation for macro and fixed-income PMs, custom prompts and skill libraries, vectorised document stores, research embeddings, email ingestion, SchonAI/Claude access, and proprietary pod-level, Bloomberg, Citi Velocity, DTCC, and internal analytics inputs. It names research briefs, trade write-ups, behavioural-bias detection, position analytics, and idea-generation tools as example workflows. The listing also specifies PM/analyst training, Claude Code sessions, recurring AI Lab sessions, usage/token-spend/hours-saved/model-adoption metrics, audit trails, trading-context guardrails, and documentation of active tools and datasets. Its three-to-five deployed-workflow target is a hiring objective, not proof of completion. The official opportunities index separately shows AI Implementation, AI Technology, AI Data Engineer, and quantitative/data roles across New York, London, Hong Kong, Singapore, São Paulo, and Miami. These are unusually specific public hiring and operating signals, but they do not disclose exact model versions, training corpora, data contracts, filled personnel, or portfolio authority.

Two conference routes widen the personnel and academic graph. Databento’s Quant Night London report names William Dorsey of Schonfeld on a June 18, 2025 commodities panel discussing dataset scope, alternative data, regression, and whether ML/LLMs had entered workflows, without attributing remarks to him. The Fordham QuantVision programme identifies Natalya Dmitriyeva as Schonfeld’s Global Head of Data Content and places her in an investment-process fireside chat alongside panels on ML in quantitative finance and data for quant funds. These are event and speaker-network records, not Schonfeld system disclosures. See the capture note.

September 1 Robeco agentic-AI, Boston hiring, and ML-method refresh

Robeco’s current Agentic AI and alpha page identifies Mike Chen as Head of Next Gen Research and describes possible agentic support for idea screening, signal monitoring, hypothesis testing, and portfolio review. Its companion workflow article places data quality, validation history, portfolio constraints, mandate boundaries, review, and human accountability around the system. A current Workday listing describes a newly established Boston next-generation quant-equity team applying ML/NLP-driven models to stock selection, alpha generation, and portfolio construction, with collaboration across Boston, Rotterdam, and London. This is unusually specific public strategy and hiring language, but it remains firm-authored intent rather than a filled-personnel or production-system audit.

Robeco’s 2025 “black box to glass box” white paper says its quant strategies use ML signals for return and risk, over different horizons, including direct alpha signals and indirect idea generators; it also says some models use textual or audio inputs. The paper names Tobias Hoogteijling, Vera Roersma, and Matthias Hanauer, and describes tree models, neural networks, Shapley interpretation, and a proprietary return-attribution framework. The examples are illustrative or hypothetical where stated, and do not disclose a current model catalog, provider, training data, permission map, deployment date, or independently audited performance. The capture note records the source boundaries.

September 1 title-blind QRT firm, partnership, and research refresh

QRT’s current About page describes research teams spanning engineering, computer science, physics, mathematics, data science, and fundamental analysis; a global research and execution platform; and structured and unstructured data processing from tick-level to multi-year horizons. This is current firm-controlled operating language, but it does not name an LLM, model family, data vendor, training corpus, agent, or production permission map.

QRT’s public LinkedIn conference post describes an IHES-linked Hong Kong conference on algorithmic collusion and names Nan Chen of The Chinese University of Hong Kong as keynote speaker. The accessible post supplies a research-network and governance-topic route, not a recording, transcript, or QRT system disclosure. Separately, the January 29, 2026 MAQi press release names QRT as a supporting partner for an AI-for-markets master’s program. It describes program-level GenAI coursework for financial documents and structured data, ML for risk and portfolio construction, and partnership support for satellite imagery, maritime-traffic datasets, and HPC. None of that establishes QRT access to those datasets, contribution to a model, or transfer into a trading system.

An indexed public profile for David Schnurr associates him with QRT and displays the 2024 NeurIPS paper Drift-Resilient TabPFN. The paper studies tabular prediction under temporal distribution shifts and lists academic affiliations, not QRT. This is a profile-reported personnel association plus academic-method evidence; it is not evidence that QRT owns, uses, funds, or deploys the method. See the capture note.

September 1 Renaissance social and conference-surface refresh

The current route check adds a verified Renaissance Technologies LinkedIn company page, which confirms the public entity boundary, East Setauket location, linked rentec.com domain, and mathematical/statistical description. It does not add an AI or GenAI disclosure. A separate Quant Conference speaker roster lists Robert J. Frey as current CEO/CIO of FQS Capital and former MD of Renaissance Technologies; Stony Brook’s biography dates his Renaissance retirement to 2004 and describes historical fixed-income research and predictive/risk work. This is a historical alumni and conference-network route, not current Renaissance personnel or current-firm AI evidence. See the capture note.

September 1 Voleon official firm-media refresh

Voleon’s current official home page explicitly describes machine-learning innovation as central to its data-driven investment-management work and says it applies flexible statistical models to financial prediction. Its management page names Michael Kharitonov as CEO; Jon McAuliffe and Vasco Chatalbashev as Co-Chief Investment Officers; Prem Gopalan as CTO; Jeremy Rosenblatt as Managing Director of Strategic Projects and Head of Systems and Software; and Dave Tolliver as Chief of Technical Staff. The same page associates Chatalbashev with predictive models and equity-trading systems, Gopalan with portfolio optimization and market-impact estimation, and Tolliver with European and Asia-Pacific equity-strategy research. These are direct firm-authored role and strategy statements, adding a firm-media route to the existing Voleon record.

The pages also expose a Berkeley/Stanford/Harvard/Princeton/Carnegie Mellon/Purdue/BITS academic and technical lineage for named leaders. That lineage is useful for personnel and research-network mapping, but it does not establish that any university supplied Voleon’s data or production methods. The reviewed official pages still do not disclose an LLM, GenAI assistant, research agent, model provider, fine-tuning corpus, data-license boundary, production endpoint, or AI-attributed result. The capture note records the page-level evidence and disqualification boundaries.

September 1 Squarepoint official-media and podcast/video search pass

Squarepoint’s current official home page and About page describe a scientific, research-led systematic-investment organization. Its early-careers page names machine learning among relevant quantitative disciplines and says technology interns work on research and trading systems. The reviewed navigation exposes no dedicated public podcast directory or video archive.

Title-blind searches across Squarepoint, named personnel, quantitative research, machine learning, and systematic trading returned careers, profiles, conference routes, and secondary material, but no podcast or YouTube episode whose publisher and participant identity could be verified as Squarepoint media. This is a bounded negative result, not evidence that no unindexed, private, or later episode exists. Existing Vinit Adya, Peetak Mitra, and conference records remain separate personnel or academic routes; public statements about tools or APIs are not promoted to firm-wide deployment claims. The official surfaces do not establish a current LLM/GenAI catalog, training corpus, agent permission map, or AI-attributed performance. See the capture note.

September 1 Millennium title-blind podcast routes

An INSEAD Knowledge podcast, dated November 24, 2025, identifies Ben Charoenwong as an INSEAD finance professor and former pod-shop practitioner discussing the Millennium Partners case. Its description covers the separation of alpha generation from shared capital allocation, risk, compliance, and technology functions, plus concentration and liquidation risk. This is an academic case discussion of an operating model, not a current Millennium AI disclosure.

Two additional Pitch The PM routes were recovered. EP035, dated June 10, 2026, identifies Eric Moster as a former Millennium portfolio manager and current Portrait Analytics CEO; its AI-powered research-tool discussion belongs to the post-Millennium company unless the recording provides a direct contrary attribution. EP018, dated November 25, 2025, identifies Doug Garber as a former Millennium senior PM and supplies chapter metadata around research, risk models, factor overlays, market-neutral portfolios, and platform culture. These are historical former-employee perspectives, not current Millennium system evidence. None of the three routes establishes a current Millennium LLM, model provider, training corpus, production endpoint, agent permissions, or AI-attributed performance. The capture note records the boundaries and private-capture queue.

September 1 WorldQuant official firm-surface refresh

WorldQuant’s current How we work page describes a global quantitative asset manager developing and deploying systematic strategies through a proprietary research platform, millions of mathematical algorithms, and collaboration among researchers, technologists, and portfolio managers. The page also links BRAIN and Data Exchange as distinct public research and data pathways. Its current careers index exposes AI, data-acquisition, BRAIN-research, and technology role families across multiple regions.

The current AI Scientist posting explicitly seeks machine-learning research for the investment pipeline, predictive signals, LLM-based agentic technology for testing trading signals and algorithms, signal compression/combination, deep-learning architectures, and model frameworks for investment professionals. A June 2026 WorldQuant Ideas article describes agents digesting financial documents, generating hypotheses, running simulations, and refining algorithms in a largely autonomous loop, while retaining human direction and critical evaluation. These sources document current hiring intent and firm-authored AI positioning; they do not establish a filled role, named model/provider, training corpus, evaluation design, production endpoint, agent permissions, or AI-attributed performance. WorldQuant University, Ventures, Foundry, BRAIN, and Data Exchange remain separate evidence surfaces. See the capture note.

September 1 QRT podcast-search disqualification

The title-blind QRT podcast pass did not locate a verified QRT-hosted or QRT-attributed podcast. A Spotify lead reached through a Peter Buckland-Merrett LinkedIn profile route resolves to a QMUL “Just Sayin’ IT” episode with Check Point cybersecurity guests; it does not identify Buckland-Merrett or QRT. That lead is retained as a disqualification, not as QRT evidence. Separate QRT first-party, conference, vendor, academic, and AI/LLM hiring records remain valid, but none is converted into podcast evidence. The capture note records the bounded negative result and the next recheck route.

September 1 title-blind Man Group data-foundation webinar route

The A-Team Group “Data Foundation for Alpha” webinar page adds a title-blind vendor-media route dated June 23, 2026. Its public speaker biography identifies Matthew Bell as a Man Group Senior Data Scientist working on onboarding, analysis, and research across commodities, FX, and equity/fixed-income indices, and says he leads an effort to use AI to automate workflows and extract signals from textual datasets. The page’s broader agenda covers point-in-time accuracy, identifiers, validation, governance, lineage, and build-versus-buy decisions. Because the recording is behind an access form and no public transcript was verified, this is bounded personnel and publisher-description evidence. It does not establish speaker-attributed remarks, a named model, training corpus, data contract, production deployment, permissions, or performance. See the capture note.

September 1 G-Research ICML video-hosting cross-check

The title-blind conference pass found a distinct official G-Research Vimeo upload, “ICML 2026: What G-Research researchers found most exciting.” The public video page identifies G-Research as the uploader and says that three quantitative researchers discuss ICML 2026 in Seoul, conversations with machine-learning paper authors, and emerging optimisation ideas. This adds a stable video-hosting route to the firm’s existing ICML paper-review page. The speakers are not named and no captions were exposed in the reviewed page, so the route is retained as conference and research-culture evidence rather than a transcript-backed model disclosure. It does not establish a model inventory, data rights, agent permissions, production deployment, or performance attribution. See the capture note.

September 1 CFM title-blind J.P. Morgan transcript route

The J.P. Morgan Making Sense episode on alternative data, published May 23, 2025 and recorded April 29, identifies Mark Fleming-Williams as CFM’s Head of Data Sourcing. In the publisher transcript, he describes a three-person external-data function covering discovery, trial coordination, provider management, and commercial negotiation. He gives a dated qualification frame around long history, broad instrument coverage, and point-in-time correctness, and describes multi-month dataset trials before a production purchase. The episode also discusses LLMs and unstructured data as industry-level changes, including easier scraping and broader text analysis. This is concrete practitioner and workflow evidence, but it does not establish which datasets passed CFM’s trials, a named model or provider, a corpus, data rights, agent authority, live deployment, or performance. See the capture note.

September 1 HRT CoreWeave infrastructure-partnership route

CoreWeave’s August 20, 2026 newsroom announcement announces a multi-year agreement with Hudson River Trading for AI-driven trading research and model development. It names NVIDIA Vera Rubin NVL72 and HGX B200 systems, Spectrum-X networking, and connectivity between HRT’s on-premises environment and CoreWeave. The release quotes Kevin Lee as HRT’s Head of Research & Development and frames the need around larger models, growing training-data volumes, and researcher experimentation. This is a concrete vendor/customer infrastructure route, distinct from HRT’s media discussions of research workflows and token use. It remains vendor-reported evidence: it does not establish contract economics, utilization, model architectures, training corpus, data rights, agent permissions, autonomous trading, or performance. See the capture note.

September 1 CoreWeave customer-newsroom expansion: IMC and Flow Traders

The CoreWeave IMC announcement, dated August 6, 2026, says IMC will significantly scale its use of CoreWeave’s AI cloud for research infrastructure. It quotes Rob Burke, IMC’s Chief Technology Officer, and says the firm had run several CoreWeave clusters in production since 2025, with the expanded commitment intended to support larger models and datasets, parallel experiments, and faster iteration. This is vendor-reported customer evidence; it does not identify IMC models, data, evaluations, permissions, or which outputs reach trading systems.

The CoreWeave Flow Traders announcement, dated July 28, 2026, says Flow Traders selected CoreWeave as primary AI-cloud provider for its AI and deep-learning division and secured dedicated compute for foundation-model training supporting an AI-driven quantitative-trading strategy. It quotes Joshua Mathew as co-head of that division and describes preparation for a launch. That is unusually specific public organizational and infrastructure language, but it remains a vendor announcement: it does not establish the model architecture, objective, training corpus, data rights, evaluation, production date, agent authority, or performance. The capture note keeps the two firms separate.

September 1 HRT NVIDIA AI-factory and simulation route

NVIDIA’s GTC 2026 financial-services coverage and the related on-demand session describe an HRT AI factory using Blackwell and Spectrum-X, connecting data ingestion, model training, simulation, and deployment. The partner material names synthetic market data for backtesting, electronic-market digital twins, and a Lefdal Mine Data Centers training hub; it also publishes a 1.6x Blackwell-versus-Hopper research-iteration comparison. These details create useful infrastructure and modality routes, but they are partner-published and lack public workload definitions, controls, or reproducible benchmark artifacts. They do not establish HRT model architectures, training corpus, simulator validation, data rights, agent permissions, or AI-attributed returns. See the capture note.

September 1 Optiver IIT Bombay AI-lab route

Optiver’s October 28, 2025 corporate announcement announces the IIT Bombay–Optiver AI Innovation Lab. It says the facility will provide high-performance servers, GPUs, and specialised research software for faculty, doctoral scholars, and postdoctoral researchers, with joint research, workshops, and training led through IIT Bombay’s CMInDS and SJMSOM units. The announcement names Tish Ghosh as Optiver India’s Managing Director and Head and Prof. Shireesh Kedare as IIT Bombay’s Director. This establishes an academic-partnership and named-leadership route, but not a project list, principal-investigator roster, model or dataset use, publication policy, production handoff, or investment impact. See the capture note.

The follow-up IIT Bombay facility page adds a 2,000-square-foot facility, 15 workstations, GPUs and servers, Bloomberg and Refinitiv terminals, and real-time market-data sources. It names Rajendra Sonar as Professor-in-Charge and Pratik Jawanpuria as associated faculty, while a C-MInDS update says the lab had awarded initial grants across four projects in adaptive AI, online learning, and tabular foundation models by July 24, 2026. Optiver’s education page names derivatives, ML pricing, liquidity modelling, and algorithmic trading as educational research areas, and IIT Bombay’s BFSI symposium page records a March 28, 2026 event with keynotes, panels, papers, workshops, and an ideathon. These are university, company, and event disclosures; they do not identify the four projects, show model or data access, establish Optiver production use, or support performance attribution. See the updated capture note.

September 1 Flow Traders first-party AI-programme refresh

Flow Traders’ 2025 annual report says the firm was accelerating AI-supported analytics for trading efficiency and had launched a dedicated deep-learning programme to build quantitative capabilities for short- and mid-term alpha generation. In its February 12, 2026 results-call transcript, CEO Thomas Spitz describes an applied-research approach using large exchange and firm-flow datasets for trend and market analysis, with an intention to roll results into a live proof of concept over the following quarters. The call frames the work as multi-quarter and multi-year, not as a completed production result.

The June 23, 2026 Horizon 2030 release places AI and deep learning within research, technology, and connectivity, and separately names business-process automation and a 2027 investment horizon. Together, these first-party materials provide a dated strategy-to-implementation trail. They do not disclose model architecture, training corpus, labels, evaluation, data rights, agent permissions, live proof-of-concept results, or AI-attributed returns. See the capture note.

September 1 IMC official AI/ML workflow and hiring refresh

IMC’s current AI/ML careers page describes machine learning and AI across research and trading workflows, including predictive modelling on large noisy datasets, signals, algorithms, strategies, models, and “production agents.” It separately describes high-speed experimentation and agent integration across the technology stack. A current Machine Learning Research Lead listing specifies a centralized ML environment, market prediction, signal generation, portfolio optimization, structured and unstructured data, feature engineering, model architecture, and deployment, and names PyTorch, TensorFlow, and JAX as relevant tools. These are first-party careers and positioning surfaces; they do not establish a filled role, selected framework, model architecture, training corpus, data rights, agent authority, or AI-attributed returns. See the capture note.

IMC’s PhD Fellowship page adds an academic-talent route: it targets doctoral candidates in AI, ML, statistics, applied mathematics, computer science, and engineering, and describes funding, a researcher event with PhD advisors, mentoring, and conference support. The public page does not name recipients, advisors, principal investigators, projects, or publications, so it documents a research-network channel rather than an academic lineage or production handoff. The capture note records that boundary.

September 1 H2O GenAI leadership and workflow cross-check

H2O’s official October 3, 2025 announcement identifies Timothée Consigny as CTO and Head of Innovation in Generative AI. It says H2O began integrating GenAI in 2023 to improve internal processes and names external-research processing and analysis of cognitive and group biases in internal investment discussions as examples; the same announcement says H2O retains its discretionary approach and investment decisions remain with the investment team. A CFA UK community page says Consigny and CIO Vincent Chailley shared lessons learned in a presentation and Q&A whose recording was shared with members. The public page does not expose a transcript. Separate A-Team speaker metadata and a TradeTech FX programme cross-check Consigny’s GenAI leadership role and public event participation. This supports role and stated scope; it does not establish live adoption, model or provider, training corpus, data rights, agent authority, or performance. See the capture note.

September 1 Akuna official ML and prediction-markets routes

Akuna’s official Quantitative Researcher posting describes statistical and machine-learning algorithms for trading strategies, signal and feature research, predictive-model improvement, and scalable, reproducible ML workflows. An official Sydney recruiting post names neural networks, tree models, ensembles, execution metrics, and portfolio optimization for a regional ML-research role. These are first-party hiring signals; they do not identify filled personnel, models, providers, data rights, or live deployment.

Akuna’s separate Prediction Markets quantitative-research posting exposes a different modality: data-capture pipelines for sports contracts, historical-data pricing and calibration, internal research tooling, and a stated research-to-production bridge with development and trading. It lists machine learning, statistics, and optimization among the desired disciplines. This is a role specification, not evidence of a deployed model or realized performance.

An official Python-engineering posting adds a workplace AI boundary: Akuna describes itself as AI-friendly for daily work while prohibiting AI assistance in interviews and assessments. That is a job-listing policy statement, not a model-provider or enterprise-permission map. The capture note records the source classes and limits.

September 1 Quadrature title-blind research-technology route

The current Quadrature Platform Engineer posting and Quantitative Developer posting expose a system boundary that a search limited to “AI” titles would miss. Both describe an automated trading system that ingests data, forms views, executes in markets, and learns from results. The platform role names research infrastructure, low-latency stream processing, distributed data systems, trade execution, observability, and an internal developer platform; the quantitative-developer role says developers work across the system and names Python, C++, and Rust. These are current first-party job descriptions and therefore hiring-intent and operating-scope evidence, not proof that each component is live or that “learning” means online learning.

Quadrature’s internship posting separately describes an 11-week London/New York programme spanning Quant Development (Research and Technology) and Core Technology, with projects designed from scratch around real business problems. Its current openings page shows Platform Engineer, Quantitative Developer, internship, and open-ended hiring routes. A public self-authored profile for Salvo Scellato identifies him as Head of Research Technology at Quadrature, formerly an engineering director at Google DeepMind, and describes responsibility for AI-model research experimentation and paths to production. That personnel remit and lineage remain self-reported pending first-party confirmation.

The Quadrature regulatory page resolves the current legal-entity surface, while an April 2026 self-published hiring post adds a secondary account of team structure and autonomous systems. The combined public record supports a technology-and-research operating-model route. It does not disclose a model inventory, LLM provider, training corpus, data rights, evaluation gate, permissions, live utilization, or AI-attributed performance. See the capture note.

September 1 Seldon currentness and disqualification loop

Seldon Capital’s current public LinkedIn company page describes a San Francisco private investment firm taking a scientific approach to long-term macro and fundamental forecasting, and separately refers to machine learning and systematic analysis. The page visibly names Jay Mehta, Satish Ramaswamy, Matthew Fong, and Soyun Kim, but it also displays both a 2–10 company-size label and a link to 17 employees. Those page fields conflict, so neither is used as a headcount estimate and the visible names are not treated as a complete roster.

An indexed quantitative-analyst internship description provided a historical title-blind route into return, volatility, and regime forecasting across equities, rates, FX, and commodities; noisy financial and alternative data; feature construction; prediction and classification; and backtesting, cross-validation, sensitivity analysis, portfolio construction, risk, and position sizing. The indexed text named Python, NumPy, pandas, SciPy, scikit-learn, and PyTorch. The direct URL returned 404 on September 1, 2026, so this is retained as historical hiring vocabulary rather than a current vacancy or proof of a live system.

A third-party mirror repeats related language about ingestion pipelines, domain-specific features, automation, and backtesting, but its employer, location, and employment-type metadata are anomalous. It is a discovery lead only. A LinkedIn profile for Satish Ramaswamy exposes an indexed MIT education signal and a recruiting message, but the profile was rate-limited during review; a secondary Matthew Fong profile is likewise not first-party confirmation.

The public record therefore supports a date-scoped positioning statement about fundamental and macro forecasting combined with machine learning and systematic analysis. It does not establish a current filled role, named model, LLM or GenAI system, training corpus, data rights, provider, evaluation gate, production endpoint, agent permissions, or AI-attributed performance. See the capture note.

September 1 title-blind XAI and Quantica research routes

The Harrington Starr interview with Federico Fontana, Chief Technology Officer at XAI Asset Management, is a distinct 2026 media route found through the publisher’s systematic-trading archive rather than an AI-title search. The publisher describes XAI as a mid-frequency liquid-futures systematic manager and Fontana as involved in infrastructure, research and development, and product formation. Its account covers cloud-provider, programming-language, architecture, resource-allocation, incremental-design, and overengineering decisions in a firm built from the beginning. The publisher also describes a focused instrument set as part of XAI’s strategy design. This is named executive and publisher evidence; it does not disclose an AI model, training data, permissions, or performance attribution.

A separate TooBigToFail episode with Nicolas Mirjolet, CEO and co-head of research at Quantica Capital, was published June 25, 2026 and found through an Italian-language podcast route even though the interview is in English. The timestamped publisher text describes Mirjolet’s path from short-horizon statistical-arbitrage work to longer-horizon systematic trend following and discusses Quantica’s portable-alpha paper. Quantica’s 2026 Q1 research PDF provides a separate first-party research route: a hypothetical, benchmark-based study of diversifier construction using public proxies and volatility normalization. These sources add strategy and personnel context, not evidence of GenAI, model deployment, or realized performance.

The Women in Data Science official archive is an additional verification surface for a historical Leda Braga/Systematica episode. It explicitly records Braga’s dated caution that autonomous investing was constrained by randomness and sparse financial data. This is useful as a human/machine-boundary control, not as a current Systematica model or policy disclosure. See the capture note.

September 1 title-blind PDT ML and research-engineering route

PDT’s current careers page lists Applied ML Scientist, Quantitative Researcher, and Research Engineer in Research / Strategies, and Performance Engineer and Software Engineer in Technology. That role mix makes the public signal more specific than a generic “quant firm uses ML” statement: the firm is recruiting across statistical research, research infrastructure, model/system performance, data, controls, and execution. It is still hiring-intent evidence, not a complete employee roster or a claim that every listed role is filled.

The Research Engineer posting describes a team intended to accelerate machine-learning research and expand the space of ideas tested with quantitative researchers. The stated remit includes scalable infrastructure for alpha, signal, and portfolio-construction research; machine-learning, hardware-accelerator, and HPC advances; an internal research stack; and model optimization for inference in real-time trading systems. The posting explicitly asks for infrastructure for training and fine-tuning large ML models and names PyTorch/TensorFlow and hardware accelerators. This is a meaningful public disclosure of the engineering problems PDT wants to staff, but it does not identify model families, training data, model owners, deployment coverage, or performance.

The Applied ML Scientist posting describes a devise–implement–evaluate–iterate loop for statistical methods applied to trading-strategy improvement. It calls for large-dataset experience, applied-ML research, a strong publication record, and a research history detailed through ranked papers, a representative paper, model projects, and data-analysis projects. That application design is a public recruiting signal about how PDT evaluates research talent. It does not show which publications were hired, which methods reached production, or whether the role is filled.

The Performance Engineer posting links GPU and systems optimization to researcher throughput, model scaling, compute capacity, cost, telemetry, and production. The London Software Engineer posting adds external-data onboarding and validation, cloud and research environments, firm-wide frameworks, trading controls, market-data systems, and execution infrastructure. Together these postings expose a public architecture around the research loop and its operational dependencies; they do not disclose a specific cloud provider, spend level, foundation model, or agent permissioning.

An older University of Washington PDT recruiting event provides academic-lineage leads for two named event participants: the indexed description associates Seth with Stanford Statistics and MCMC variance-reduction research, and Kurt with a UC Berkeley CS PhD and an ML focus. The page also records a PDT start in 2011 and an NYU Center for Data Science adjunct role for Kurt. Because this is dated university recruiting material, it should be read as lineage evidence rather than current personnel or a map from academic research to live alpha.

See the PDT capture note. The evidence supports public hiring intent across ML research, research engineering, fine-tuning/inference, GPU and systems performance, data infrastructure, and trading controls. It does not establish a named model, training corpus, data rights, vendor, agent, production deployment, investment result, or comparative position.

September 1 title-blind D. E. Shaw GenAI operations route

The D. E. Shaw Group’s current Human Capital Analytics Developer posting describes Python-based automated analytics, dashboards, ETL pipelines, reporting infrastructure, and analysis for strategic and operational talent questions. It also asks for an interest in generative AI and using it to improve internal tools, workflows, and deliverables. This is a firm-controlled signal about enterprise analytics and workflow automation. The role is in Human Capital, not investment research; it does not disclose an investment-facing model, agent, provider, data source, or trading permission.

The The Org Generative AI page displays a five-person D. E. Shaw team, including titles such as Lead, Generative AI Skills; Generative AI Engineer; Vice President, Applied Generative AI; and Engineering. The page is explicitly marked “Unverified,” so these names and titles remain discovery leads rather than confirmed public personnel. They should not be merged into D. E. Shaw’s historical Pedro Domingos machine-learning-group route or into the separate D. E. Shaw Research organization.

The historical Domingos page identifies a 2019 Head of Machine Learning role and discusses alternative data, finance, and machine-learning research. That route remains distinct from the current Human Capital job and from D. E. Shaw Research’s interdisciplinary computational-science work. The combined evidence supports a layered public record—historical investment-oriented ML leadership, current internal analytics automation, and an unverified secondary GenAI-team lead—but not a current investment GenAI architecture, named model inventory, training corpus, deployment status, agent permissions, or AI-attributed performance. See the capture note.

September 1 Radix title-blind ML, research-platform, and HPC route

Radix Trading’s official firm page describes the company as a research firm monetized through trading, says it evolves strategies through machine-learning and statistical methodologies, and refers to an automated research platform. Its current experienced and university job boards expose a title-blind architecture: quantitative research sits alongside quantitative technology, DevOps and automation, HPC systems, FPGA, network engineering, and research technology and trading systems. These are current first-party surfaces and hiring signals, not a complete employee roster.

The Quantitative Researcher posting describes trading-opportunity research, alpha design, empirical testing, market microstructure, and a collaborative research platform. It also asks applicants for academic transcripts and links to publication or research lists, creating a useful academic-lineage discovery route. The DevOps and Systems Engineering posting explicitly connects live trading systems, research infrastructure, alpha generation, research-to-production translation, fully automated strategies, ML research and simulation infrastructure, distributed/HPC systems, and petabyte-scale data analysis. The C++ quantitative-technologist internship repeats the connection between production code, automated strategies, live trading, and ML research and simulation.

This supports a current first-party description of an integrated research, simulation, systems, and trading-technology operating model, and supplies ordinary titles for future title-blind searches. It does not disclose model families, training corpora, data vendors, foundation models, LLM or GenAI use, model owners, evaluation fixtures, production coverage, agent permissions, or AI-attributed performance. The capture note preserves the source boundaries.

September 1 NeurIPS Generative AI in Finance personnel route

The organizer-maintained NeurIPS 2025 Workshop on Generative AI in Finance page lists Dianqi Li with Citadel Securities in its panel and displays Pusheng Zhang as Head of Machine Learning and AI at Cubist Systematic Strategies. The same biography describes Zhang’s prior Citadel AI Research route and earlier research and engineering roles at Uber and Microsoft. This is a useful dated conference and personnel-discovery surface, separate from the J.P. Morgan post and Point72/Cubist first-party pages already tracked.

The page is event metadata rather than a first-party firm disclosure. It does not provide a recording, transcript, paper authored by either person, model inventory, training corpus, deployment status, agent permissions, investment authority, or performance evidence. The displayed affiliation is not by itself a current-employment verification or proof that a workshop topic was implemented at either firm. See the capture note.

September 1 title-blind media-route expansion

Seven exact routes were added after checking the local ledger for URL, video-ID, and named-person collisions. The KX Capital Markets Summit panel and its YouTube recording name North Rock Capital Management’s Thomas Yang alongside BTIG, Scotiabank, and KX speakers. KX describes historical and streaming data, kdb+ and q, real-time signal generation, research-to-production latency, causal discovery, reproducible pipelines, and limits of agentic AI in trading. This is vendor-hosted practitioner material; a recording-level capture remains a follow-up.

The Finteda / Deutsche Bank QIS talk provides a publisher transcript for a March 12, 2026 session with Wai-Chung Ip (Caio), Senior Quantitative Developer. It describes a Research SDK, serializable backtest pipelines and an Inventory, data catalogs, the stated NelData route to roughly 8,000 datasets, DataStore orchestration, agentic RAG, progressive MCP/tool disclosure, iterative evaluation, LangChain/LangGraph, and confidentiality routing between on-premises processing and GCP. The related panel page uses another “Caio” identity; the records are not merged. These are bank-QIS workflow claims, not hedge-fund deployment or performance evidence.

Additional title-blind routes include OneTick’s low-latency analytics webinar, the A-Team unstructured-data webinar, and the AWS Industry Forum Hong Kong customer-event page naming ActusRay Partners speakers Ken Poon and Arthur Yan. These expose vendor, market-data, cloud, and customer-event routes; capture status and source access differ, and none establishes a named customer model, data right, or production result.

The Top Traders Unplugged SI400 episode adds a publisher-transcript route with Yoav Git and Rob Croce, while the Professional Punters Spotify episode adds a Spotify-only route with Evan Semet. The latter does not name Semet’s current employer, so historical employers are not assigned to a current firm. These are process and discovery evidence, not AI-system or performance disclosures. See the consolidated capture note.

September 1 Aquatic and Seven Research title-blind ML routes

Aquatic’s official site describes applied scientific research, research engineering, a high-performance research-and-development platform, systematic strategies, and disciplined risk management. Its current early-career quantitative-researcher posting describes designing, implementing, and evaluating signals, models, strategies, and research-system components using data-intensive statistical work. The application requests a GitHub username and public SSH key, which adds a technical-workflow and code-review signal. These are first-party hiring and positioning statements, not evidence of a filled role, model inventory, or performance.

Seven Research’s official site describes a New York quantitative research and technology company launched in 2024, with scalable models, systematic strategies, data analysis, risk management, and automated execution. Its current job board lists quantitative research, deep-learning research, algorithmic development, and core-development openings. The Deep Learning Researcher role names advanced deep-learning models for quantitative finance, rigorous systematic research, large datasets, HPC, PhD-level training, and publications or conference presentations. The Core Developer role connects research and trading infrastructure to HPC, network optimization, hardware selection, latency-sensitive development, and distributed systems.

The two firms add title-blind vocabulary for future passes—research engineering, model evaluation, deep-learning research, core development, HPC, and research-system components. The sources do not establish model families, training corpora, data rights, vendors, agents, production coverage, autonomous capital authority, AI-attributed performance, or a causal link from prior affiliations to current systems. See the capture note.

September 1 conference and regional discovery routes

The IAQF/LSE DSI London event, held June 23, 2026, adds a named speaker route around AI/ML in financial markets: Álvaro Cartea of the Oxford-Man Institute, Peter Hafez of RavenPack / BigData.com, James Hamp of Citi, Petter Kolm of NYU, Luitgard Veraart of LSE, Johannes Ruf of LSE, and Hilary Till of Premia Research. The event description covers insight generation, risk, execution, market structure, and adoption challenges. It is speaker and conference evidence; the page does not provide a fund-specific model, transcript, deployment claim, or performance result.

The FOW Asia 2026 page and agenda add a Hong Kong route for September 24, 2026. Its “AI in Trading — From Experiment to Market Infrastructure” panel sits beside sessions on Chinese derivatives, market structure, and risk. FOW identifies hedge funds, asset managers, proprietary trading firms, exchanges, banks, brokers, and technology providers as intended participants, but the current agenda does not expose named panelists in text. It is therefore a regional conference route, not evidence about any participant’s system.

The Battle of the Quants London 2026 route was rechecked during this pass, but it was already covered in the August conference record, including its video archive and role-level Two Sigma/Berenberg signals. It is intentionally not duplicated here. See the existing capture note.

September 1 European title-blind quant and ML routes

The UK/Swiss pass found LMR Partners’ careers page, which is dated to June 30, 2026 and describes Technology and QuantLab across alpha research, derivatives modelling, portfolio analytics, technology, risk, and portfolio-manager collaboration. Its firm page names Ben Levine, Stefan Renold, and Danny Mitchell in senior roles. A The Org directory names quantitative roles but is explicitly unverified, and an ETH job artifact is hiring evidence. No named AI lab, model family, vendor, corpus, or production deployment was established.

An SGX customer story identifies Florin Court Capital as a quantitative hedge fund and describes automation across its FX workflow. Risk.net, The DESK, and an SGX post provide dated personnel and practitioner-media routes. The public record supports systematic and automated-execution context; the efficiency claim remains SGX-reported customer-story evidence and no ML, GenAI, or AI-attributed return claim is established.

Oxford Algorithms describes proprietary ML software, explainable AI, real-time data automation, and an AI-powered macro quantitative fund. Its approach page describes continuously fed data informing hold or reallocate decisions, while the team page names Dr Mohsen Zadeh-Koochak, Shirin Dehghan, Mobin Zadeh Kochak, Martin S. Hauge, and Professor Rama Cont. The firm’s appointed-representative relationship with JTC AIFM Solutions UK means the entity classification needs care; the reviewed pages do not publish model specifications, data, permissions, or independently verifiable results.

Ultramarin adds a German AI-oriented quantitative asset-manager route covering equity forecasting, stock selection, tactical allocation, managed accounts, ETFs, mutual funds, structured products, and an Ultrascope API. Its team page names Ronald Richter, Dr Mathias Eitz, Dr Benjamin Trendelkamp-Schroer, Cheng Chen, Dr Andreas Fuest, David Dümig, and Julien Florian Jensen across ML engineering, technology, asset allocation, equity selection, and quantitative strategies. The site records 2025 talks and team changes, while Goldman Sachs’ announcement supplies a product route. This is not treated as a confirmed hedge fund, and no model architecture, training data, or production permissions are shown.

mu Capital Management describes AI-driven stock selection, a long ML history, hundreds of signals, analysis of more than 8,000 stocks, compute resources, full automation, and a proprietary codebase. Its team page names Ulrich Carl and Marcial Messmer; Messmer’s 2017 SSRN paper studies deep feed-forward networks trained on 68 characteristics for cross-sectional return prediction. A 2024 factsheet describes an automated ML strategy in an actively managed certificate with mu as strategy advisor. The reviewed legal-purpose language excludes activities requiring a regulated asset-management license, so this is a strategy-advisor/product route rather than a confirmed hedge-fund manager.

Spinoza Capital’s careers page advertises an Investment Research & Machine Learning internship involving algorithmic or AI-driven strategies, backtests, reporting, and ML tools. Its management page names Dr Philip Schnedler and Benedikt Schöps, and its imprint identifies a German securities institution supervised by BaFin. This adds a regulated German systematic-asset-management route, but the vacancy is hiring intent and does not establish a filled role, model inventory, vendor, or live-trading authority.

SINTRO describes systematic traded funds, scientific research, data-driven decisions, algorithmic execution, and ML. Its careers, team, and company timeline expose quant, software, data, and research-platform roles and name Denis Keller and Jona Detjen. A June 5, 2026 research article discusses liquidity indicators, on-chain data, and agent-based stress models. Because the timeline describes a first AIF as “2026 Soon,” SINTRO is not treated as a verified hedge-fund platform; its self-published claims remain bounded accordingly.

Causality describes market-neutral long-short portfolios using ML and factor modelling. A 2023 interview names Mark Horvath, an SSRN paper supplies a firm-affiliated research route, and a public CV describes Gábor Balázs as a former quantitative analyst in the private fund. A 2021 university job posting describes terabyte-scale equity data, ML integration, simulation, hybrid cloud, automated trading, and a live-trading engine. Most detailed technical evidence is historical; current entity structure, personnel, and deployment remain unresolved. See the capture note.

September 1 personnel and academic-lineage expansion

Public personnel surfaces add research-lineage leads that are separate from firm deployment evidence. Fan Chen’s profile and the Seven company page show a Seven affiliation, while Princeton’s ECE profile identifies an ECE PhD path advised by William M. Jacobs. Zijia Cheng’s profile, CV, and research site associate him with Seven and document Princeton physics and Tsinghua training; Princeton’s thesis list confirms a 2025 doctorate. The Seven company page separately describes Yiqun Luo’s completed quantitative-research internship, market-replay/matching-engine work, and an exploratory reinforcement-learning overlay for intraday CME futures. These routes establish varying levels of affiliation and academic history, not Seven model ownership or production deployment.

Aquatic’s public personnel layer includes Ethan Jaffe, whose profile describes Aquatic quantitative research, ML in statistical arbitrage, and MIT training alongside an MIT thesis record and arXiv thesis; Shuntao Chen, whose UW mathematics PhD is documented through a dissertation record; Thomas Swayze, whose profile lists CMU training; and former research engineer David Lin, whose profile lists Stanford, ML/NLP coursework, and transformer/sparse-computation work. Currentness and firm-to-paper linkage vary; no Aquatic production claim follows.

For North Rock, Rakshith Kamath’s profile lists Columbia signals/information training and ML, NLP, speech-recognition, reinforcement-learning, and convex-optimization coursework, while a Manipal publication record confirms a computer-vision medical-imaging paper. For QFI, the official site names Ajay Biradar as CEO/founder and Baibhab Kumar Mustafi as Quant Backend Systems Head, while an Ajay Biradar build update connects Mustafi to trading-engine/backtesting work. Aditya Raj’s site and CV describe historical QFI internships involving forecasting, sentiment, pipelines, and AI research, but conflicting dates and titles keep currentness unresolved.

Sapientia personnel routes include a self-reported internship post by Marcus Tsz Hin Hung naming data-analysis/system-development work and supervisors, plus Jing Wang’s profile. Alpha Node’s Alpha Prime Trust page identifies Dr Andy Ting as CIO and describes AI in trading algorithms; his public profile lists NTU research and publications on computational situation awareness and time-critical decision-making. These are personnel and firm-positioning signals, not evidence that academic papers became live investment models. See the capture note.

A title-blind search surfaced a useful Campbell & Company personnel and research-engineering route. Hannes Vandecasteele’s public profile says he has been a Campbell Research Engineer since October 2025, following a Johns Hopkins postdoctoral fellowship and a KU Leuven Ph.D.; it names Ioannis Kevrekidis as a supervisor and describes work across machine learning, time-series analysis, stochastic modelling, and high-performance computing. The Org’s Campbell team page displays a nine-person Research and Engineering roster and describes predictive-model and trading-algorithm work, but labels the directory “Unverified.” These sources establish a personnel lead and a public team surface, not Campbell model ownership, data rights, portfolio authority, or performance. The source note keeps the two evidence classes separate.

Unlimited Funds adds a distinct former-Bridgewater and alternative-manager-replication route. The firm says its machine-learning technology uses index-return data to infer and replicate real-time positioning of alternative managers, and names Bob Elliott and Bruce McNevin as co-founders, with McNevin identified as Chief Data Scientist. This is first-party strategy and personnel positioning; the page does not disclose the training sample, replication error, permissions, or independently audited results. Aargo Trade’s team page is retained as a lower-confidence regional watchlist route: it names Ankush Shah, Joost de Ruijter, and Yi Heng Sun among quant, engineering, and operations roles and describes Aargo Capital VCC as a boutique quant hedge fund using automation. The promotional page does not independently verify the fund entity, team size, models, or outcomes.

September 1 Kronos Research internal-platform route

Kronos Research’s official careers page currently displays Machine Learning Researcher, quantitative research, quantitative trading, portfolio optimisation, and senior SRE roles across Taiwan, Singapore, Hong Kong, and remote locations. An indexed Machine Learning Researcher posting, now redirected to the current jobs page, describes time-series, order-book, and trade-data features; MLP, LSTM, RNN, Transformer, and reinforcement-learning architectures; backtesting and PnL attribution; model deployment; and automated retraining and monitoring. Because the detailed URL redirects, this is historical hiring language rather than a current vacancy.

The still-accessible Senior SRE posting exposes a particularly specific internal-platform route: Slurm HPC, Lustre/NAS storage, AWS/Alibaba Cloud/GCP, Terraform/CDK, Docker/Kubernetes, self-hosted GitLab, and CI/CD for research and production. Its GenAI section names LangChain, LangGraph, Bedrock, Elasticsearch RAG, MCP servers, chatbots, and AI agents. This is public staffing specification, not proof that every named component is live or that agents have investment authority. The portfolio-optimisation artifact is also redirected and therefore historical. See the capture note.

September 1 title-blind audio routes across Brazil, Australia, and the United States

The Stock Pickers / Bayes episode identifies Marcello Paixão as Bayes co-founder, CEO, and manager and discusses price, statements, text, images, and audio as possible signal inputs. This is a notable multimodal research claim, but the public transcript contains apparent automatic-caption/name errors and does not identify model families, training data, licenses, permissions, or verified performance.

The Plato episode and Livewire summary identify founder and managing director Dr Don Hamson and discuss quantitative long/short origins, systematic fundamental analysis, and an incremental view of AI tooling. The Vinva episode adds a distinct appearance by managing director and head of investments Morry Waked, covering his path into systematic investing and Vinva’s data- and technology-scaled research process. The publisher did not expose verbatim transcripts, but both public MP3s were captured privately and processed with timestamped WhisperX-MLX (1,008 and 1,258 segments respectively). These remain process and personnel routes; no model or deployment claim is established.

The Behind the Ticker / THOR Bridgeway episode is distinct from the existing Bridgeway Excess Returns record and identifies co-CIO Elena Khoziaeva. Its timestamped machine transcript covers small-cap universe design, negative-momentum exclusions, multi-metric valuation, monthly rebalancing, and research culture. It adds systematic-process context, not GenAI or model-deployment evidence.

The Avos Capital episode identifies founder Josh Blanchfield and discusses China-linked commodity flows, options, convexity, and AI as a possible research assistant. The Asset Management One USA episode covers Jiro Fujisawa’s 80-synthetic-market trend, carry, and skew strategy. The Tuesday Capital episode identifies CIO Dato Netto in a title-blind cross-border discussion, and the Springs Capital episode identifies founder/CIO André Caldas in a Portuguese discussion of Brazilian long/short, macro/fundamental research, and AI’s emerging role. None of these episodes establishes a named model, training corpus, data rights, agent permissions, or measured outcome. See the capture note.

September 1 global regional and vendor/customer expansion

The East Asian pass adds GBM Asset Management, whose Tokyo role page describes systematic macro research using ML for signal discovery and portfolio optimization, alongside John Fou’s biography and Takuya Sugimoto’s public profile. Saccade Capital and its Korean site describe AI/ML, automated research, petabyte-scale data engineering, and low-latency proprietary trading; public profiles for Minkyu Han and Munki Chung add financial-engineering research routes. Kronon Labs describes ML factors, high-frequency systems, stochastic-control execution, and reinforcement-learning market making, with Ali Ahmed Sheikh and Shiban Atif Khan as public affiliation routes. These firms describe proprietary or self-capital activity; the sources do not establish external hedge-fund vehicles, model ownership, data rights, or independently verified results.

UC Capital and Holdwin Capital add Taiwan routes through an NTU AI Quant Program notice and an NTU Holdwin job notice. The public materials mention model construction, backtesting, GPU resources, ML training, structured and unstructured data, and AI agents connecting research workflows. BlockTech’s quantitative-trader role adds crypto-system monitoring, execution logic, and market-microstructure work, with Bryan Wong Wei Heng’s profile as a personnel route. Nogle advertises Meridian as a financial-AI platform with citation-grounded RAG, private-cloud deployment, and large model/corpus figures; those figures are first-party claims and remain uncorroborated. Nissay’s Kota Takano profile adds historical ML-enhanced factor-investing and alternative-data context. None of these routes by itself establishes a live production system or external hedge-fund mandate.

The Africa and Latin America pass adds Differential Capital and Mazi Asset Management, with the Prescient Mazi disclosure supplying a named long/short fund-product route. Numoro’s JSE announcement describes an ML-powered ETF partnership with Prescient. Quantum Wave Capital and its fund regulation add Chilean ML pattern-discovery and private-fund claims, while a LarrainVial quantitative-equity job describes production signals, datasets, ML deployment, monitoring, and automation as hiring scope. Vector COMMODQ describes a Mexican AI fund product. Arqaam’s research page and company post, plus the DFSA register, add a MENA hybrid-ML model-portfolio route. The SEDCO Freestyle AI fund terms describe manager-verified AI recommendations and order direction, but this is a public-equity fund rather than a hedge fund and the English text is an unofficial translation. The consolidated route note keeps fund classification and disclosure boundaries explicit.

Vendor/customer material adds useful implementation vocabulary without identifying every underlying customer. FALGOM’s Amicorp case describes AI/ML strategies, commodity futures, execution, and ETH Zürich researchers in an actively managed certificate. NNAISENSE’s Lightning AI case and official site describe Bayesian Flow Networks, a Large Investing Model, multimodal data, EvoTorch, a risk-control room, Kubernetes, and persistent storage, with Jonathan Masci and Vojtech Micka named publicly. YC’s Prodigy Research page describes a quant-finance foundation model and AI agents but leaves the model, data, capital, and evaluation undisclosed. Auquan’s CloudThat case exposes a portfolio-intelligence and AWS architecture route. An S&P Global case describes an anonymous investment firm using a finance-tuned LLM on earnings calls and filings, while a Virtova case describes an unnamed multi-strategy fund using on-premises inference, specialized models, regime detection, behavioral factors, simulation, kill switches, and audit trails. These anonymous and vendor-controlled routes are leads, not independently verified firm disclosures.

Snowflake’s State Street Alpha case names Aman Thind, Jeff Shortis, and John Plansky and describes deep-learning anomaly detection, data-quality controls, and a conversational portfolio interface. An RSystems case adds Duality Group cloud/security infrastructure vocabulary without an AI claim. These are adjacent institutional and platform routes, not evidence that the named technologies are used by GMO, Acadian, Arrowstreet, or another covered manager. No comparative firm assessment is made in this expansion.

The title-blind conference follow-up adds event routes that a hedge-fund keyword search would miss: a J.P. Morgan QIS Cambridge recap, the With Intelligence Hedge Fund COO Summit Europe, the Fall JOIM AI in Finance conference, a Cornell Financial Engineering event, ICCF 2026 Oxford, the Asian Quantitative Finance Conference programme, and the London AI in Financial Services event. These pages expose dates, themes, and selected speaker or programme metadata, not attendance, model ownership, implementation, or performance. See the conference follow-up note.

The Odds on Open catalog reconciliation confirms that the catalog must be searched episode-by-episode rather than by podcast name alone. It adds the Neel Somani episode as a dated former-Citadel power-market route with an official YouTube mirror and private caption capture, while placing a conflicting Rich Falk-Wallace catalog entry on hold until its guest identity is reconciled. Several apparent catalog gaps map to existing episodes through alternate Spotify, YouTube, Apple, or publisher routes. This is a metadata-integrity and former-personnel workflow finding; it does not establish current Citadel systems or any model.

September 1 global regional and title-blind route expansion

The regional-language pass adds a Nikkei Financial route for Nissay Asset Management, whose visible metadata describes generative-AI use in portfolio-company analysis and recurring disclosure work. The full article is paywalled, so the model, data, controls, and performance details remain unavailable. A Huatai Securities/Taidu Voice episode describes factor factories, end-to-end models, post-training, and GenAI workflows, while explicitly stating it is not a Huatai research report. A Cailian Press roundtable names Lingjun, Pansong, and Mengxi practitioners and attributes discussion of filings, text, audio, earnings calls, video, LLM-assisted factor discovery, and human responsibility. These are regional media and practitioner routes; they do not establish a common production stack or independently audited outcome.

An Arabic-language route adds a different kind of regional signal: the Apple Podcasts episode “Trading stocks using artificial intelligence” names Predictfa CTO Maisara Hammouda and exposes chapter metadata around AI-based equity trading. It is retained as Arabic investment-technology discovery, not as evidence that Predictfa is a hedge fund or that the discussion reflects a live portfolio system. The capture note records the translation and transcript boundary.

The same pass found university-hosted hiring surfaces for Mingshi, Mengxi, and Turing Private Fund, plus a DTL Quant listing. Their public vocabulary includes deep-learning research, AI infrastructure, NLP/LLM work on research text and alternative data, signal construction, backtesting, simulation, and execution analysis. These are recruitment surfaces, not filled-personnel, live-model, or performance evidence. The Pi Associates/Vietquant history is more explicit about a claimed sequence of ML feature engineering, reinforcement learning, agentic systems, genetic algorithms, and multi-agent architecture from 2023–2025, but remains a firm-authored claim without independent system or data corroboration.

India adds three dated AlphaGrep reports (licence; retail launch). Together they trace proposed AI/ML quant products, reported licensing, and a quantitative multi-asset product, but they do not disclose model versions, vendors, training data, permissions, or isolate AI contribution. A separate academic-partnership pass finds two AlphaGrep research-lab routes. IIT Madras’s Wadhwani School lab page describes an AlphaGrep Quantitative Research Lab funded with reported CSR support of Rs. 5.65 crore and names N. Sudarsanam, Nirav Bhatt, Chandrashekar Lakshminarayanan, and Arun Ayyar under B. Ravindran, with research spanning financial markets and microstructure, quantitative investment management, and risk. The IIT MoU page lists intended areas including portfolio optimization, market microstructure, deep and reinforcement learning, alternative data, sentiment, explainability, and automated trading; IIT’s annual report additionally describes five-year funding, training, and dataset/resource development. Separately, the AlphaGrep x IIITH lab site describes a Precog-linked collaboration, displays an active “Portfolio Prediction” project using machine-learning and statistical models, and names no deployed model or data source. AlphaGrep’s public announcement names Ponnurangam Kumaraguru, Mohit Mutreja, and Hemang Mandalia in the initiative; a public personal profile adds a research-intern route focused on modeling human cognition in financial markets. These are academic, company-posted, and self-authored signals, not evidence of a live production model, data rights, agent permissions, or AI-attributed performance. See the AlphaGrep academic-labs capture note. In Australia and Canada, the Monash–Q Group colloquium names Alice Berriman, Armina Rosenberg, Michael Kollo, and Hasan Fallahgoul in AI/quantitative-investing discussion, while the RBC Borealis AI in Finance Summer School supplies an academic-industry talent route. Neither event establishes fund deployment.

Several current first-party role surfaces sharpen the operating vocabulary without exposing a model registry. Cubist/Point72’s London quantitative-research role names price-volume, order-book, alternative-data, anomaly-research, backtesting, and production-implementation work. Flow Traders’ Amsterdam role mentions ML, possible deep learning, large datasets, and live-trading deployment. BlackRock’s quantitative-investing careers page describes a 200-plus-person function, ML-based systematic equity research, internet-search and demographic data, peer review, and live strategy implementation. These pages are first-party hiring or function descriptions; they do not identify named hires, model weights, data rights, or AI-attributed results.

The upcoming TradeTech FX Europe draft agenda names Julian Gronau/QRT and Nir Vulkan/Oxford and lists AI sessions on PM/quant workflows, macro research intake, and decision support. A title-blind J.P. Morgan Making Sense episode discusses systematic commodities, signal generation, execution alpha, and discretionary/quant convergence. Rob Carver’s Top Traders Unplugged episode adds AI-assisted backtesting, overfitting, robustness, and due-diligence discussion. These media and event routes remain separate from any named firm’s production system.

Two research routes add methodology context. The Axyon/EuroHPC project describes an LLM layer intended to interpret numerical asset-ranking forecasts, alongside proprietary ML ranking; customer identity, weights, and production use are not disclosed. Deep Parametric Portfolio Policies studies neural-network portfolio policies and variable-importance analysis, but an author affiliation does not establish LIQID deployment. The Bank of Canada paper contributes a control reference on text, speech, image, and transaction data, reproducibility, residency, and operational risk rather than investment-fund evidence. The full route set and explicit exclusions are in the dated source note.

The podcast pass adds title-blind personnel and infrastructure routes. The Alternative Data Podcast episode with Mohsen Chitsaz identifies him as a Portfolio Manager at Eisler Capital and advertises discussion of news data, satellite data, LLM developments, and quantitative research. The recovered public audio includes a personal account of real-time news parsing, alternative-data examples such as satellite and corporate-jet data, LLM-assisted news-context analysis, and the importance of data curation and vendor-bias checks (18:49–19:08, 24:33–27:55, 36:01–37:36). These are automatic-ASR-mediated personal statements, not evidence of Eisler’s current model inventory, data licenses, vendors, permissions, production endpoint, or performance. See the capture note. A separate Avi Rosenbluth episode records a historical former-AQR role without establishing his current employer or current AQR practice. These are discovery and personnel routes, not current system disclosures.

The Unhedged “Trade like a bot” episode, with an FT transcript locator, discusses a University of Chicago experiment using an LLM and company financial statements. The FT page is subscription-gated, so any backtest claim needs paper-level verification. BMLL’s “Leveraging Data as a Differentiator” discusses level-3 market data, order-book analytics, data quality, Snowflake/Databricks, and possible AI/ML use cases; it is vendor positioning rather than evidence of a named fund’s purchase. SDS 485 adds a full transcript PDF covering entity resolution, knowledge graphs, bitemporality, and research pipelines, while client and live-alpha claims remain self-reported.

Finally, Can Machines Invest? identifies Bryan Kelly as AQR’s Head of Machine Learning in a dated educational discussion with Horst Simon of Lawrence Berkeley National Laboratory. Wharton Behind the Markets identifies Gareth Shepherd as Voya’s Co-head of Equity Machine Intelligence and discusses model competition, overlapping datasets, pattern recognition, and prediction horizons. Shepherd is already represented elsewhere; this episode is a new media route, not new personnel evidence. The full podcast batch is in the dated source note.

September 1 AI-lab and researcher-lineage follow-up

The personnel pass adds Anastasia Borovykh as a public CFM Executive Director working on ML sources of alpha, with prior Imperial College and Liquid AI work and a linked financial-forecasting paper. Thomas Eboli’s site identifies him as a CFM quantitative researcher working on ML and GenAI for trading signals and exposes a public GitHub account. The listed papers are primarily computer vision and image restoration; neither route establishes that those methods became CFM portfolio components. The lineage note keeps the academic and firm evidence separate.

Man’s Martin Luk biography identifies him as Head of Applied AI, Systematic, and says he leads LLM research while developing systematic strategies. A separate 2026 Man paper names Sumant Wahi and Alex Preston alongside MIT Research Fellow Paul Kedrosky in market commentary. These are direct personnel and research-context signals, not model weights, permissions, or return attribution.

Bridgewater’s AIA Labs article dated August 27, 2026 links a UIUC/Thinking Machines collaboration to the ReViSQL paper and open-source implementation. The topic is reinforcement learning and text-to-SQL data quality, not a disclosed trading model. A current Citadel PhD ML Researcher posting describes deep learning, NLP, unconventional data, backtesting, production-quality code, and investment-strategy model development; it is hiring intent and does not identify a hire or live system.

The Two Sigma regime-modeling paper by Alex Botte and Doris Bao applies a Gaussian Mixture Model to 17 factors and discusses risk/allocation use cases while explicitly saying the model is not predictive. AQR’s Financial Machine Learning survey, by Bryan T. Kelly and Dacheng Xiu, is explicitly not tied to an AQR investment strategy or official view. Jane Street’s Alok Puranik paper identifies an ML researcher and studies positional encodings for attention models without claiming trading deployment. A G-Research NextGen profile documents an Oxford DPhil, Mark van der Wilk supervision, Imperial AI/ML training, and prior unnamed multi-manager hedge-fund data-science work; the scholarship does not establish G-Research employment or model use. No comparative conclusion is drawn from these routes.

September 1 major-firm title-blind routes

The undercovered-firm pass adds a D. E. Shaw Venture Studio AI Product Engineer role naming OpenAI APIs, LangChain, LlamaIndex, vector databases, orchestration, workflow automation, and prototypes. It is hiring intent in a venture-studio context, not proof of a D. E. Shaw investment model. The AIMA-AITEC webinar replay identifies Marshall Wace CTO and Partner Conor Kiernan in a discussion of AI, cloud, cybersecurity, and vendor due diligence; a separate AIMA CyberTech agenda lists James McGinnigle in an AI/ML applications session. Neither route discloses a Marshall Wace model or production system. A historical Jim Simons archive adds secondary 2009 media metadata but no current Renaissance or AI conclusion.

Vendor infrastructure routes add detail without a model claim. A DDN announcement identifies Jump CTO Alex Davies and describes storage for HPC and AI-driven quantitative research. A ClickHouse customer story names Arnaud Adant and describes petabyte-scale trading logs, Iceberg, Redpanda, research analytics, and HPC workflows. Tom Marty’s profile says he is joining QRT as a quantitative research analyst in September 2026 and lists probabilistic modeling, reinforcement learning, generative modeling, compression, and web agents; the affiliation is self-reported. Akhilesh Narayan’s profile shows Squarepoint experience alongside a 2023 ChatGPT portfolio-selection paper, but the paper does not list Squarepoint and firm adoption is not inferred.

The ACI FMA AI-powered FX webinar page identifies XTX Global Head of Distribution Jeremy Smart and advertises a May 21, 2025 discussion of ML and AI in FX prediction and trading; no internal model or transcript was recovered. Goldman Sachs’ 2026 alternatives-conference page provides a February 13, 2026 Brevan Howard route identifying Brian Friedman, with AI appearing as a broader macro theme rather than a firm-system disclosure. See the major-firm route note. No comparative firm characterization is made.

September 1 Asia-Pacific and India route expansion

The Asia-Pacific pass adds a more concrete asset-management implementation record from Japan. A Digital X case study reports that Tokyo Marine Asset Management introduced its TMAM AI application across the company in April 2025 for investment analysis and company research. The report says approximately 70% of roughly 400 employees were using it by May 2025, describes a projected annual efficiency benefit above 10,000 hours, names Nowcast as development support, identifies Claude among multiple LLM services, and places the application in TMAM’s closed AWS environment. The public record is unusually specific about workflow, vendor, cloud, and adoption claims, but it remains a publisher account and does not expose model weights, data rights, permissions, or AI-attributed investment performance.

Nissay Asset Management now has a linked first-party research trail rather than a single generic AI reference. Its BERT analyst-report page names Takaki Yoshino, Yoshiaki Kimura, and Megumi Tsukamoto in the Investment Engineering Department and discusses BERT/Transformer representations, fine-tuning, and analyst-report sentiment scores. Adjacent machine-learning work on cyclical stocks, a deep-learning stochastic-volatility topic, and random-matrix market-structure analysis broaden the visible research vocabulary. These pages show public research activity and named contributors; they do not establish that one model drives a live portfolio or isolate any return contribution.

An important upcoming discovery point is IEEE CIFEr 2026, scheduled for September 10–11, 2026. Its programme lists Michinori Kanokogi, Head of Data, AI & Quant Research at Nippon Life Asset Management, speaking on analysts working with agents, production-development limits, failure modes, and human feedback loops. It also lists University of Tokyo Professor Kiyoshi Izumi on causal discovery over Japanese earnings reports, conditional diffusion for price paths, multi-agent analytical processes, and market simulation. These are keynote abstracts, not independent verification of the described systems; the conference should be rechecked for slides, recordings, and evaluation details after the event. The AIMA Japan 2025 agenda supplies an earlier related speaker route, including Nissay, Ryobi AlgoTech, Snowflake, and Japan’s Financial Services Agency in an AI and technology breakout. AIMA marks that forum as Chatham House Rule, so unrecorded remarks are not attributed.

India adds a customer-side partnership disclosure. In a February 10, 2026 release, Kotak Mahindra AMC says it deployed Pascal AI’s agentic research platform in institutional investment research. The release describes a Context Graph connecting internal frameworks with regulatory filings, earnings transcripts, and market updates inside a sovereign VPC. This supports a dated, named deployment claim and a data-integration clue; it does not disclose model versions, evaluation results, agent permissions, autonomous trading authority, or performance. A separate SMU SKBI discussion names Dymon Asia co-founder and Co-CIO Danny Yong alongside Professor Hong Zhang and discusses ML-discovered factors, explainability, liquidity, and turnover. That is practitioner/research dialogue, not a Dymon system disclosure.

The regional job and social surfaces are useful but weaker. An Enrise Family Office listing describes LLMs, RAG, quantitative agents, reinforcement learning, factor libraries, and explicit backtest controls, but the listing was no longer accepting applications and does not establish a hire or deployment. A Bareksa post names QUANTIT and STAR Asset Management and describes AI-assisted analysis, portfolio construction, early-warning signals, and backtesting; those statements remain social-post claims pending partner-side or technical corroboration. The full regional batch, including inaccessible pages and event-only leads, is in the dated source note. No cross-firm capability assessment is made.

The same academic-lab search surfaced the IIT Bombay–Citadel Securities Quantitative Research Lab, a university-industry research and education route distinct from Citadel’s internal systems. Its public project list names machine-learning options pricing, IPO/social-data analysis, multiagent bandits, deep-learning American puts, order-flow toxicity and liquidity, and high-frequency change-point detection. The page lists Sudeep Bapat as Professor-in-Charge, S.V.D. Nageswara Rao as first Professor-in-Charge, and associated faculty whose public interests include ML, reinforcement learning, neural networks, stochastic control, statistical signal processing, optimization, and Rohan Chinchwadkar’s stated focus on LLMs and agentic AI in finance. A PhD fellowship notice describes research grants, mentorship, and thesis pathways, and the lab page records an April 7, 2026 research symposium. These are university programme and faculty-interest disclosures; they do not show Citadel model adoption, access to proprietary data, production endpoints, trading permissions, or performance. See the IITB–Citadel capture note.

A second title-blind pass extends the map into less-covered regions and role families. Highfort Investment’s 2027 campus campaign lists quantitative research, AI-algorithm research, and systems engineering positions in Shanghai; Graviton’s Gurugram ML-research posting is more specific about order-book time series, in-house features, scalable predictive models, Python/C++ production integration, and infrastructure for ML research. Both are hiring evidence, not proof of filled roles or live capital deployment.

The PolyU–Quant Cloud Centre adds an academic-commercial research route in Hong Kong. PolyU names Professor Jingran Zhao and Going Quant Chairman David Wu; the associated project role describes LLM/NLP market-signal extraction, model integration, agent development, Level-2 Hong Kong/U.S. equity data, and H100/A100 resources. This establishes a research collaboration and technical project specification, not production performance or trading authority.

Canada adds a weak but trackable title-blind route through Infinitus Capital. Its home page describes a Toronto-based quantitative trading firm using proprietary multi-strategy and long/short algorithms with automated infrastructure; its careers page describes a Toronto research-centre and software roles supporting systematic trading, financing, and accounting, with computer-science, mathematics, physics, Java, C++, SQL, and Unix requirements. The public pages do not mention ML, deep learning, LLMs, GenAI, named researchers, model details, data rights, or audited performance. The firm’s separate performance page makes broad unsupported claims, which are excluded. This is a discovery lead and technology-hiring signal, not verified AI-strategy evidence. See the Infinitus capture note.

Australia and Europe add different operating signals. Avangard Capital’s FAQ describes an A.L.F.R.E.D. AI/ML process for Australian equities, while AE Capital’s careers page describes data science, ML, and automated systematic strategies; both remain firm self-reports. RAM AI’s public research/company pages and a Risk.net report connect historical ML and deep-learning work to a reported agentic-AI initiative, but do not disclose a complete model or independently audited outcome. The full route set and academic lineage leads are in the regional Europe/APAC note and personnel note.

The podcast and conference sweep also found source types that keyword search misses. Papers With Backtest exposes direct VTT captions for episodes on web data, 13F filings, factors, and backtesting. QuantMinds International 2026 lists future sessions on AI-driven alpha, agentic AI, RAG evaluation, synthetic curves, and deep hedging, while Oxford–Man’s 2026 archive provides direct academic replays. These routes widen the capture queue; event and podcast descriptions do not establish any named firm’s production system. See the vendor, conference, and podcast route note.

September 1 Australia and New Zealand title-blind routes

The State Super announcement, published October 7, 2020, describes an investment data-science program using machine-learning models for financial-market decision-making. It names Deputy CIO Charles Wu and an Academic Oversight Body comprising Michael Kollo, David Michayluk, and Alex Antic, with reporting to the investment committee. This is historical evidence of governance around investment ML, not a current model inventory, dataset disclosure, or performance record.

The Future Fund FY2024–25 review names its proprietary AI collaboration tool, LUMi, as synthesizing information to support decision-making. Earlier annual highlights describe AI-literacy work involving Disruption, Technology, and Investment teams and migration of historical data, models, and processes onto strategic platforms. These are first-party institutional disclosures; they do not name LUMi’s model/provider, permissions, evaluation, or whether it generates portfolio signals.

New Zealand adds a dated governance and strategy route. The Guardians appointed Anastasia Moskvina Head of Data Analytics on July 1, 2026, with a remit described as shaping AI governance, practical use cases, and scalable capability. The 2025 New Zealand Capital Markets Symposium describes a front-office data-analytics function created in March 2022 as a centre of excellence and lists Kathryn Kerner, Head of Data Analytics. An Otago conference profile identifies Shane Varn as responsible for a domestic-equity machine-learning strategy. These are dated organizational and speaker disclosures, not a complete current architecture or performance record.

Several title-blind role routes add implementation vocabulary. Alphinity’s biography identifies Richard Hitchens as Principal and Head of Quantitative Research, with UC Berkeley information-and-data-science training. Ardea’s biography identifies Dr Laura Ryan as Head of Research and Development and describes advanced statistical and machine-learning models supporting idea generation and portfolio technology. Catamount Trading describes systematic strategies and a Data Scientist role covering forecasting/classification, alternative data, backtesting, transaction-cost/latency analysis, and production deployment. The Catamount listing is no longer directly viewable. These pages do not establish current model inventories, data rights, GenAI use, or AI-attributed results.

The La Trobe industry scholarship describes an unnamed Melbourne quantitative-trading partner and a brief covering reinforcement learning, multi-agent cooperation, generative world models, tick-level prices, news sentiment, order-book depth, continual learning, uncertainty, interpretability, compliance, backtesting, and live shadowing. The partner and researchers are not named, so this is a follow-up/deanonymization route rather than evidence about a named manager.

The AI-Driven Super FundTech 2026 Summit, scheduled for October 29–30, 2026 in Melbourne, lists Dr Sheenal Srivastava (AI Lead, Vision Super), George Tranganidas (AI Product Manager, Cbus Super), Stephen Jackel (CTO, Future Super), and Daniel Selioutine (CIO, ESS Super). Its investment panel covers ML in asset allocation, risk modeling, scenario simulation, human–AI collaboration, and regulation. This is organizer-supplied speaker metadata; recordings, slides, and employer-controlled biographies remain to be recovered. See the regional capture note.

September 1 regional-language investment-AI routes

Regional-language searches add product and partnership evidence that generic hedge-fund keywords miss. MAFFinTech’s Arabic page describes Cortex as emerging-market AI investment infrastructure with multiple models, decision logging, and explainable recommendations, while QANTERION describes paper-first quantitative research, visible risk, human responsibility, and separately enabled live execution. Stock AI PSX describes Urdu and English Pakistan Stock Exchange reports built from continuous data capture, computed indicators, and an agentic reasoning layer. These are self-published product descriptions; customers, model architectures, data rights, validation, and outcomes are not established.

Quant.co.id describes an Indonesian AI/ML platform with IDX analytics, an AI signal engine, backtesting, automated risk controls, GPU compute, and direct-market-access infrastructure; its displayed quant.engine v0.1.0 is a product identifier, not evidence of live deployment. RHB’s official release and Malay Bernama account describe an AI-driven multi-asset fund with Qraft Technologies, naming Marcus Kim and Ng Chze How and reporting more than 80 datasets. A Maybank/Arabesque AI report describes a Shariah-compliant AI-powered discretionary mandate with monitoring and rebalancing; its RM100 million figure was a forward target. The releases do not disclose model families, contract terms, permissions, monitoring, or independently measured performance.

The Edge Malaysia describes PK Cypher’s Singapore-domiciled US Fund and its “Good Eye” model, attributing to Anthony Siau a daily ranking of 500 S&P stocks using 17 sub-factors, weekly batch orders, regime indicators, and covariance-based risk controls. These are founder-reported media claims without audited track record, model code, or independent live-system verification. In Vietnam, Rồng Việt Securities’ role names CNNs, RNNs, LSTMs, Transformers, GenAI, LLMs, RAG, agents, and workflow automation, while an FPT Securities role specifies embeddings, causal analysis, drift monitoring, A/B testing, post-deployment measurement, and MLOps/LangOps. Both are hiring intent, not proof of filled roles or production systems. See the regional-language capture note.

September 1 guest-name podcast expansion

Guest-name searches added media routes around known practitioners. The CFA Institute page identifies Acadian co-CEO and co-founder John Chisholm in a January 21, 2021 discussion of machine learning, big data, ESG, value investing, and data-science talent and provides a downloadable transcript PDF. This is historical executive context, not a current Acadian model inventory.

Odd Lots adds a November 13, 2025 Cliff Asness episode with AI/ML chapter metadata. Rational Reminder provides a full transcript for a 2020 Asness discussion of market efficiency, value, behavior, and allocation, while Meb Faber and a 2026 Antti Ilmanen episode add AQR personnel and research-philosophy routes. These do not establish current AQR AI deployment or performance.

Euan Sinclair’s August 2026 Low VIX episode, his 2023 options episode, and Imran Lakha’s Macro Dirt episode add audio routes for people associated elsewhere with Hull Tactical and Options Insight. The individual episode pages do not establish those firm affiliations or any firm-level AI use; the missing captions make local ASR the appropriate follow-up.

FirstMark’s event page and YouTube recording identify Gideon Mann as Bloomberg’s Head of ML at the time, while a separate Millennium profile is current-personnel evidence. The Bloomberg discussion must not be transferred to Millennium. See the guest-name capture note.

September 1 European quantitative AI-lab and platform routes

AKO Capital identifies Angus Lund as Head of Data Science and Patrick Hargreaves as CEO and Portfolio Manager. A 2022 data-scientist role described data sourcing, ML, portfolio analysis, automation, statistical evaluation, and data infrastructure; a 2025 J.P. Morgan podcast describes ML and LLM tools as additive to discretionary research. The public materials do not name production models, vendors, data rights, agents, or AI-attributed performance.

The ansa site names Kevin Jörg, Maxime Didascalou, Jan Gutjahr, Dr Maximilian Sauer, and Dr Sascha Mergner in portfolio and research leadership. It says machine learning is used on structured and unstructured data for alpha and risk forecasts, and biographies reference deep learning, high-frequency signals, GlassBox ML, dynamic return forecasting, factor timing, and ML-based stock strategies. These are first-party descriptions; model versions, deployment, permissions, and results are not supplied.

Sandro Felicioni’s Vontobel profile identifies him as a Quantitative Investment Manager and says he develops models behind the firm’s AI Powered Global Equity strategy. It describes statistical analysis and ML over large datasets, automation and optimization with portfolio managers, and public work on LLMs, explainability, and backtesting. The profile does not identify the provider, architecture, corpus, benchmark design, or degree of automation.

Berenberg Innovation & Data describes ML and alternative data drawn from written and spoken news, newspapers, television, and social media. It names an ALOS—Autonomous Learning Overlay Strategies—system for sentiment-based price prediction and claims approximately 600 million documents analyzed monthly and roughly €3.3 billion in strategies controlled by self-developed ML algorithms. It also describes cooperation with Google on a chatbot for discretionary portfolio managers. These are company-reported figures and capabilities; model architecture, licensing, current chatbot status, benchmarks, and independently verified AUM are not shown.

An Amundi Quantitative AI Portfolio Engineer role, published June 12, 2026, describes fixed-income ML model research, validation, and deployment using tree ensembles, LSTMs, Transformers, Hugging Face, sentence-transformers, PyTorch/TensorFlow, and data from news, central-bank communications, broker research, transcripts, and regulatory publications. It also specifies walk-forward testing, look-ahead and survivorship controls, transaction-cost modeling, drift monitoring, rollback, and governance. This is a job specification and research route, not proof of a filled role, a live model, data rights, or performance.

Quoniam describes a proprietary research platform integrating tested ML, new data sources, emerging factors, and academic collaboration, alongside a Snowflake data-management release. Its public interviews name human judgment, prompt quality, model supervision, and problem definition as important controls. The reviewed sources do not disclose a named LLM, weights, training corpus, agent permissions, or AI-attributed result. QUANT is retained separately as a technology vendor describing deep learning, deep RL, causal ML, LLM chatbots, RAG, and a claimed “Virgo” agentic architecture; that is vendor evidence, not evidence of a particular hedge-fund deployment. See the European capture note.

September 1 quant personnel and academic-lineage expansion

Ting-Wu “Rudy” Chin’s profile states that he has been a Citadel Securities quantitative researcher since July 2021, applying statistical and machine-learning techniques to market modeling. CMU’s lab alumni page independently lists him at Citadel; his academic work covers model compression, neural-architecture search, transfer learning, and efficient deep networks. This establishes a person and academic route, not transfer of computer-vision research into Citadel market models.

The University of Chicago placement page records Tingran Gao’s placement at Radix, while his UChicago profile documents harmonic analysis, differential geometry, topological data analysis, manifold learning, and high-dimensional statistics. His public papers include SelectNet and Gaussian-process landmarking. The Org directory labels its Radix profile unverified; no Radix model or production application is public.

Kieran Wood’s profile records Caxton quantitative-research employment from July 2023 to January 2026 in systematic futures and FX, followed by NVIDIA AI Solutions Architect work. His Oxford DPhil research includes DeRegiME, HANET, and a deep-financial-time-series benchmark, focused on regimes, probabilistic forecasting, causal changes, and transaction-cost-aware positioning. The papers do not establish use at Caxton.

Walleye’s firm-authored post identifies Ben Cook as a Senior Quantitative Researcher on its Quantic team. Harvard’s career profile documents his move from an astronomy PhD into quantitative finance, while his personal page describes GPU-based computational astrophysics and earlier Akuna experience. No Walleye model, dataset, or deployment process is disclosed.

Farrer’s team page lists Shawn Unger as Senior Quantitative Researcher. A firm appointment post describes ML, statistical modeling, fundamentally driven signals, prior Millennium systematic long/short work, and alternative-data products developed at Nasdaq, with University of Toronto training. No named current model or paper-to-production mapping is disclosed.

Keynum routes connect named personnel to public research. A Paris 1 profile describes Hoang-Viet Le as Lead Quantitative Researcher and Data Scientist across trading, risk, ESG, financial text, and sentiment; a 2025 Springer paper independently lists Le and Keynum. Public IPAG material identifies Hans-Jörg von Mettenheim as an Oxford-Man associate and Keynum CEO, with a journal archive adding applied-ML-for-quantitative-trading research. These sources support an academic/industry bridge, not a verified live model.

The NUS Graduate School profile identified Clement Yee in August 2026 as a Cubist quantitative researcher conducting global-equity alpha research and transforming data into predictive insights; his public profile corroborates Cubist and NUS affiliations. Verition’s official site confirms its multi-strategy and quantitative businesses, while a UCLA-hosted 2026 role description specifies LLM-powered filing extraction, earnings-call analysis, data aggregation, and research acceleration. Neither route establishes a named model, filled role, production permissions, or performance. See the lineage capture note.

September 1 title-blind Europe and quant-media routes

Svelland Capital’s team page identifies Pål Sundsøy as Head of Data Science and Laurent Hoffmann as Head of Quantitative Research. The biographies connect Sundsøy to systematic-equity research at Norges Bank Investment Management and Telenor’s Big Data & AI research team, and Hoffmann to quantitative-research and analytics leadership at Shell Trading & Shipping. This is evidence of named data-science and quantitative-research ownership; it does not identify models, datasets, vendors, deployment controls, or performance.

Rembrandt Capital’s public announcement identifies Jimmy Pang as Head of Quantitative Research. A separate Pang article, dated February 24, 2026, describes persistent Markdown journals committed to Git so AI coding assistants can recover context across sessions. The personal workflow is not attributed to Rembrandt; neither source discloses firm models, data sources, permissions, or live investment applications.

The April 9, 2026 DESK report describes Alison Hollingshead becoming CTO of Jupiter Asset Management and Mike Poole becoming Head of Investment Platform after leading automation and data integration across the trading desk and systematic franchise. It names a fixed-income dashboard using Power BI, APIs, Aladdin Data Cloud, TCA providers, Propellant, historical trade data, and RFQ data. This is concrete platform evidence from asset management, not evidence of an ML model, research agent, automated investment authority, or performance result.

Feynman Point Asset Management lists Matthew Harrington as Head of Quantitative Research & Applied AI. A separate biography describes prior work at GoldenTree, Risk Harbor, Virginia Tech’s Hume Center, and MITRE Labs, including an in-house backtesting engine, a crypto factor library, ML network simulations, and federated-ML research. The official title is the stronger evidence; the separate biography is corroborating context. Current production models, data rights, and investment outcomes remain undisclosed.

The Sophron Network adds a title-blind podcast route: Euan Sinclair is identified as Portfolio Manager and Senior Financial Engineer at Hull Tactical Asset Allocation, Imran Lakha as an Options Insight practitioner and former BlueCrest portfolio manager, and Martijn Bron as a former Cargill Global Head of Cocoa Trading. Chapter metadata includes a data-layer discussion for Sinclair and an “AI applied to fundamental research” segment for Bron. These are publisher descriptions and chapter metadata, not independently verified transcripts or proof of production AI systems. The capture note records the boundaries and regional-language follow-up routes.

Canadian institutional comparators add another title-blind pattern. PSP Investments’ FY2025 annual report says its digital-innovation team joined Global Alpha in July 2024 and was renamed Alpha Science, with a remit to validate or challenge investment theses using advanced technologies and share AI/data-science tools across public and private markets. Ontario Teachers’ May 4, 2026 appointment announcement names Feifei Wu as Senior Managing Director, Investment Technology & Applied Intelligence, responsible for AI strategy, governance, and adoption. PSP and Ontario Teachers’ are institutional investors, not hedge funds; neither disclosure identifies models, datasets, permissions, or AI-attributed investment outcomes. See the Canadian institutional capture note.

September 1 vendor and customer AI-platform routes

Vendor case studies add implementation details without proving model ownership or live alpha. ClickHouse’s QRT story, dated June 9, 2026, describes a centralized research-data platform, near-real-time risk/P&L, and ClickStack observability. Its D. E. Shaw story, dated May 15, 2026, names Mike Vasiliou and describes high-throughput compute-grid telemetry, historical backfills, OpenTelemetry, Grafana, and workload-level capacity planning. Neither discloses trading models, training data, permissions, or performance.

A Google Cloud case study names Ed Jeffery, Principal Software Engineer, Investment AI R&D at Schroders, and describes a multi-agent research prototype using Vertex AI Agent Builder, private analyst-note retrieval, BigQuery natural-language-to-SQL, Google Search grounding, LangGraph, Firestore, human checkpoints, analyst feedback, and ground-truth evaluation datasets. This is prototype evidence, not proof of production authority or investment returns. A separate BlackRock case describes Dan Wolf’s portfolio-data operations work: dataset profiling, anomaly alerts, and cloud services supporting risk, performance, portfolio, and alternative-data workflows. Its generative-AI language is architectural possibility, not deployment proof.

Greenland Capital Management’s Sigma case names Nan Xiao and Michael Englander and describes governed Databricks/Sigma access across 15 investment teams, PM reporting, risk/performance/alpha data, and Python notebooks. The AWS AQR case names Michael Raposa and describes an ECS platform serving research, portfolio implementation, trading, optimization, operations, and security. These are platform and workflow disclosures; no GenAI model or performance attribution is identified.

September 2 fresh title-blind role surfaces

The underexplored-firm route note adds two role-level records. A Jain Global Singapore AI Research Intern listing describes LLM and agent-system implementation, unstructured-document processing, open-source framework adaptation, workflow automation, and robustness/statistical evaluation. A Harvard FAS Old Mission listing provides a university-hosted route into 2027 quantitative-trader recruiting and repeats the employer’s proprietary-capital description. The first is employer job-listing evidence and the second is a secondary reproduction; neither establishes a filled role, a named AI owner, model deployment, or performance. The same search re-found existing canonical records for Event Horizon Labs, Binomial Technologies, HFR, and Hedge Fund Huddle, which remain single-counted.

Microsoft’s CSOP 2024 case and 2025 case name Melody He, Yi Wang, Emma Wang, and Cenny Cui Haofan. They describe Azure OpenAI for internal document, email, coding, and operations workflows and an Intelligence Hub using Azure AI Foundry, GitHub Copilot, and available models including OpenAI o1/o3 and DeepSeek R1. The disclosed investment workflow includes extracting trade-confirmation fields and analyzing charts and sell-side research for internal ideas, with human gatekeeping and no stated direct order authority. Model availability is not evidence of CSOP training or live alpha. A Microsoft PIC case similarly describes Copilot and Azure/Fabric/Power Platform/Purview for due diligence, investment-committee materials, historical retrieval, research, and drafting, with throughput—not predictive accuracy—as the stated benefit.

Two anonymous data-provider routes remain bounded leads. Parameta Solutions describes an unnamed global-macro hedge fund receiving real-time and historical OTC/alternative fixed-income data in Snowflake for systematic research. An Earnest Analytics use-case PDF describes an unnamed U.S. long-short hedge fund using consumer-spending data, short lags, cloud feeds, and predictions across public companies. The customers, exact contracts, models, validation, permissions, and outcomes remain undisclosed. See the vendor/platform capture note.

September 1 investment-technology comparator: MDOTM LAB and Sphere

MDOTM LAB is a vendor and investment-technology R&D operation rather than a hedge fund. Its public materials describe academic and industry collaboration around AI, behavioural finance, portfolio management, multi-asset allocation, factors, asset pricing, global macro, ESG analysis, market inefficiencies, and market microstructure. The careers page describes dissertation projects in multidisciplinary teams supported by financial professionals, AI specialists, and data scientists, and reports 10+ partner universities, 30+ external researchers, 50+ completed projects, and 90+ research papers. Its scientific-advisor page names Alessandro Sbuelz of the Catholic University of Milan/Bocconi and Viktor Elliot of the University of Gothenburg. These are company-reported lab and academic-network claims; the underlying project and paper register was not exposed.

MDOTM’s Sphere product description says its client-tailored AI engine supports market-regime analysis, risk monitoring, diversification, portfolio construction and rebalancing, and investment commentary while keeping professionals at the centre of the decision. A separate product/news page mentions AI-generated commentary and a ChatGPT integration. The team page names Tommaso Migliore, Federico Mazzorin, Giorgio Malchiodi, Mario Ciardulli, and Peter J. Zangari; Zangari’s company interview gives his prior MSCI and Goldman Sachs Asset Management roles and Rutgers economics/econometrics PhD. The reviewed surfaces do not disclose foundation models, fine-tuning corpora, evaluation fixtures, customer permissions, execution authority, or AI-attributed performance. The full bounded record is in the MDOTM capture note.

September 3, 2026 — German ACATIS podcast adds model-version and human-filter details

An August 16, 2026 German Bulle & Mensch episode with ACATIS founder Hendrik Leber adds a more granular, timestamped public account. The publisher attributes to Leber a funnel in which portfolio managers preselect roughly 300–400 companies and the firm’s AI selects roughly 50 (00:59:05 chapter). Recovered local ASR adds his description of quantitative fundamental data, company histories, thousands of 60- or 90-day scenarios, an evolutionary optimizer, a ten-person AI team, chained models, and a semi-automated portfolio-maintenance idea (28:52–29:15; 60:10–61:56; 64:22–65:56; 71:35–72:52). He describes a model at version 47 and says he does not trust conventional backtests, instead observing portfolio distributions and small test portfolios (66:14–68:46). ACATIS’s own AI-funds page independently says the firm has studied AI for portfolio management since 2014 and launched three AI-using fund strategies. A separate ACATIS 2026 Value Conference page describes NNAISENSE’s LIM as generating joint market scenarios and feeding a downstream optimizer for an ACATIS ETF; that is a separate first-party product disclosure. These claims describe a human-filtered, proprietary AI selection process, but do not disclose weights, full architecture, features, data rights, split design, audit controls, execution authority, or independently verified performance. The local ASR is a navigation aid and should be checked against the audio before verbatim quotation. See the German capture note.

September 3, 2026 — German NORD/LB/ELAN route exposes allocation and risk vocabulary

The official NORD/LB Investors Talk listing and direct timestamped episode transcript add a German-language portfolio-allocation route that was not in the hedge-fund queue. The January 5, 2026 episode features Marcel Leist and Ralf Schülein of ELAN Capital Partners discussing TOPAS and NORD/LB Smart Faktor. The publisher describes TOPAS as an AI-assisted model intended to maximize portfolio efficiency, dynamically optimize allocations, analyze markets objectively, and use a turbulence index as a possible early-warning input. The transcript describes a “digital fund manager,” financial-science foundations, and a speaker-reported 2017 evaluation by LMU Munich (03:34–05:43); it also describes the turbulence-index concept and claimed crisis-warning behaviour (07:03–09:40).

NORD/LB’s September 2025 Smart Faktor fund portrait describes a “self-learning AI-based algorithm,” multiple market-data inputs, dynamic weights across equities, bonds, commodities, gold, and cash, and a multi-factor ETF universe. It presents a gross 2009–2025 backtest and says strategic realignment began on October 22, 2024; the PDF also states that fees and commissions are excluded from the simulation. These are first-party product claims and marketing backtest material. They establish a public allocation-system description, not model architecture, training data, retraining, permissions, execution controls, or independently verified performance. See the German TOPAS capture note.

September 1 title-blind comparator routes: IFTA, ACATIS, and Rathbones

The IFTA 2026 speaker page and official event listing add an upcoming October 9–10, 2026 London route. Its programme includes an AI-driven equity-curve and position-sizing session, an AI-powered-trading session by Eoghan Leahy of Quant Market Intelligence, and Adam Sorab, identified as Partner at Aptior Capital. The event page also connects Leahy’s biography to prior Bloomberg work and early BloombergGPT involvement. These are agenda and biography disclosures; they do not establish a named fund’s model, dataset, permissions, or performance, and the event remains a capture target.

ACATIS’s first-party AI-funds page names ACATIS AI Global Equities, ACATIS AI US Equities, Kevin Endler, and Dr Eric Endress, while the fund page says AI is responsible for stock selection in ACATIS AI Global Equities. A dated German interview with Endress identifies him as Head of Quantitative Research and discusses language models for company-information analysis, portfolio monitoring, and combining value investing with AI. The public record establishes a firm-described AI-fund strategy and named role; it does not provide model versions, training corpus, validation design, permissions, execution authority, or AI-attributed returns.

Rathbones Asset Management adds a data-foundation route through COO Stephen Wood. The July 16, 2026 account describes Snowflake-based golden sources, quality checks, business rules, and data spanning performance, risk, sustainability, and unstructured investment-decision information, with possible use of Snowflake agentic services. Snowflake’s customer announcement lists Rathbones as an EMEA Data Driver of the Year. This supports a customer-reported data-platform signal, not an independent assessment of AI capability, model ownership, portfolio authority, or investment outcome. See the capture note.

The unfiltered Blushing Quants publisher feed adds four episodes that were absent from the existing ledger. Roger McIntosh’s #35 episode describes institutional factor construction, point-in-time data, weekly updates, ex-post testing, and speaker-reported use of machine-learning algorithms in factor allocation. Paul MacGregor’s #36 episode contributes electronic-market, commodity-benchmark, and market-maker connectivity context, including the migration of exchange activity to electronic platforms. Gilad Bar-Ilan’s #37 episode identifies him as CEO and co-founder of Crowd Wisdom Trading; the publisher description and private timestamped ASR describe LLMs and task-specific agents extracting assets and structured trade-plan fields from thousands of hours of financial video. Antonio Berenguer’s #38 episode covers options market making, higher-order Greeks, inventory and transaction costs, and the gap between academic models and executable signals. These are publisher and speaker-reported routes with private non-diarized ASR; they do not establish a hedge-fund deployment, model version, training corpus, data rights, trading authority, or performance. See the private recovery note.

Manuel Ritsch’s #19 episode adds a distinct AI-native asset-management comparator. Ritsch’s official profile describes Alpha Rho Technologies and an advisory relationship to the GBVIII Global Dynamic Allocation fund; his company research page says the research is implemented in that fund. In the private timestamped ASR, he describes AI-analyst teams, attempts to model discretionary-trader behaviour with machine/deep learning and LLM components, workflow/tooling as the principal IP, explainability as an investability constraint, and internal rather than client-facing AI access (02:02–04:33; 18:42–19:29; 22:20–26:36). These are self-reported and partly promotional claims from a non-hedge-fund comparator; they do not independently establish model scale, architecture, data rights, permissions, fund authority, or performance. See the capture note.

The same unfiltered inventory recovered three older entries that sharpen the evidence boundaries. Raffaele Ghigliazza’s #9 episode discusses using an LLM as a processing block for text, video, or audio before conventional structured trading components, plus model comparison and tradability checks. Paul Chalmers’s #30 episode describes speaker-reported dynamic algorithm adaptation and risk-focused backtesting, but also contains promotional claims that remain quarantined. Vincent Randazzo’s #32 episode supplies a rules-based breadth and regime-management comparator without identifying an AI system. The #9 ASR contains a repetition-quality warning, and all three are non-diarized; none establishes a current employer’s model, dataset, permissions, trading authority, or performance. See the additional recovery note.

The unfiltered feed also recovered Ryan Ling’s #1 episode, whose title contains neither “AI” nor “hedge fund.” The publisher-feed metadata identifies a market-maker interview; the private timestamped ASR records discussion of algorithm monitoring, changing behaviour in response to observed flow, execution-pattern analysis using data and machine-learning methods, and human interaction with fast algorithms (07:34–15:34; 56:23–56:32). The episode is a title-blind market-structure comparator, not evidence of a named fund’s production system: current employer, model architecture, training data, permissions, trading authority, and performance remain unresolved. See the Ryan Ling recovery note.

Haris Chalvatzis’s #13 episode adds a title-blind quantitative-research pipeline route. The private timestamped ASR describes signal generation and evaluation, portfolio construction, execution, risk-factor neutralization, continuous refitting, an alpha pool, and separation between live and research portfolios (04:45–08:59; 16:57–24:00; 39:48–41:48). It also ties model validation to realized live performance, holdout data, and untouched data (25:17–25:34; 29:52–30:14). The episode’s public introduction mentions experience including BlackRock, while current public profiles show conflicting role context; no employer-specific deployment claim is admitted. See the Haris recovery note.

Roman Isachenko’s #22 episode adds a title-blind derivatives-pricing and model-validation route. The private timestamped ASR describes open-source models and AI coding agents as possible aids for coding, testing, and production pricing work (11:39–12:09), stacking ML outputs and generating parameter-outcome sets (14:59–18:16), and warnings about reproduction, overfit, alpha decay, and monitoring (19:09–21:15; 22:13–29:10; 33:33–34:09). It also mentions CatBoost/boosting in a model-selection context (46:35–48:33). These are speaker-reported methodology comments, not evidence of a named employer’s system, model weights, dataset, permissions, execution authority, or performance; the current employer remains unresolved. The local private transcript is durable, while its second private remote checkpoint is pending a DNS retry. See the Roman Isachenko recovery note.

Generic-finance discovery is also adding institutional routes whose titles do not say “hedge fund” or “quant.” The Blunt Dollar’s Lucy Walker episode discusses fund-research tooling, personalization, and AI in fund selection; its public YouTube recording was recovered and privately transcribed with timestamped local ASR. The episode describes AM Insights as using algorithmic alerts, an AI meeting note-taker, and an OpenAI-driven fund summary; these are speaker-reported product details, not evidence of a named client’s internal system. The Bull of Wall Street’s Michael Hunstad episode covers quantitative investing, factors, AI, and tokenization; and TD Asset Management’s official podcast page lists an August 10, 2026 episode on AI, markets, and growth with Vitali Mossounov, Juliana Faircloth, and Tarik Aeta. None by itself establishes a model, dataset, production authority, or performance. See the follow-up note.

A second title-blind video pass recovered several adjacent but useful routes. In a QMIND lecture, Ernest Chan connects financial-ML practice with feature selection and interpretability, risk management, capital allocation, point-in-time data, and feature engineering (11:40–14:05; 27:45–32:10); this is practitioner methodology, not current QTS or PredictNow deployment evidence. An AuumAI interview with Chris Hartnoll describes an allocator workflow with folder-level access controls, legal review of GP-document protections, information sharing across distinct investment processes, and human-in-the-loop risk review (03:55–05:18; 11:41–14:06). A November 2025 Odd Lots episode adds dated AQR executive context on technology, investing, and alternative data, but does not reveal AQR’s current AI stack. Jane Street’s firm-branded mock interview and the Citadel recruiting-analysis video are retained as firm-media surfaces without being treated as AI-deployment evidence. See the recovery note.

The recovered 2019 Acadian interview with Katherine Glass-Hardenbergh adds historical process detail that the current Acadian role pages do not expose. She describes a global quantitative-equity team building economically motivated stock-return models, structuring inputs from sources such as satellite imagery, earnings-call text, web data, and real-time exhaust, and testing whether relatively simple signals are additive to an existing book (02:55–06:47; 12:57–13:48). She describes a roughly ten-person systematic team, limited research bandwidth, vendor trials, and a business-case approach to buying data rather than a dedicated alternative-data acquisition budget (09:05–10:20; 15:10–16:22; 25:06–28:35). This is dated speaker-reported evidence from 2019: it does not establish Acadian’s current model families, GenAI systems, data rights, agent permissions, or performance. See the capture note.

The 2025 Railpen interview with Richard Dudley adds a title-blind institutional systematic-investing comparator. Dudley describes an in-house global long-only equity team of about ten people, roughly six in PM/research roles, managing about £7 billion; he discusses end-to-end ownership of research, data, and code, an early textual-data direction, and simple-signal tests against the existing book (09:05–10:20; 14:42–16:22; 20:14–22:50). He also describes decentralized data procurement, business-case review, curated datasets, and the long history needed before scraped data can support a useful backtest (23:40–28:35; 36:23–37:44). This is a dated speaker account from a pension investor, not evidence of a current Railpen GenAI system, model, data license, research-agent permission, or performance result. See the capture note.

The recovered Quantbot interview with Paul White adds a more explicit data-and-compute operating signal. White describes a roughly 80-person firm whose research workforce includes quantitative research, data science, feature engineering, machine learning, and market-microstructure work; he says the firm reviews about 150 datasets each year in trial mode and evaluates whether their expected value justifies cost (15:51–18:49; 22:06–25:25). He also describes refocusing roles on ML/AI activity, AI-generated or AI-derived data products, regional data discovery through exchanges and vendors, and a possible path toward very large feature libraries (26:53–29:48; 30:33–38:26). These are speaker-reported 2025 statements, not a model registry or performance disclosure: no foundation model, training corpus, fine-tuning process, data contract, research-agent permission, production authority, or AI-attributed return is established. See the capture note.

September 1 Alternative Data Podcast catalog recovery

Comparing the public Acast catalog with the media ledger found 106 episode URLs that had not been indexed at the episode level. The first recovery batch adds five dated, title-blind routes. The Conor Taggart episode identifies Taggart as Eagle Alpha’s Head of Data Monetization Strategy after four years heading equities data sourcing at Millennium. He describes how a dataset can support a differentiated corporate-access question without entering a formal model (42:18–43:07), and describes Eagle Alpha’s engineering, data-science, delivery, legal, and compliance infrastructure and five live revenue-generating products at the time of recording (56:51–60:08). The recording also provides a disclosure-quality warning: vendor “AI” claims may not explain the concrete value added, while large-fund data teams seek advisory filtering and noise reduction (46:59–50:08). These are speaker-reported observations, not evidence of a current Millennium, Balyasny, or Eagle Alpha model or deployment.

The Ben Cohen episode identifies Cohen as former Global Head of Data Strategy at WorldQuant. He describes repeatable procedures for dataset trials, research, and information dissemination, with data strategy acting as a bridge between external providers and internal researchers (18:44–25:05; 28:02–30:15). His AI discussion separates workflow enablement and technology-stack improvements from meaningful alpha impact, which he had not observed clearly at the time; he emphasizes the importance of longer evaluation histories (43:22–47:44). This is a dated personal assessment and does not establish WorldQuant’s current system.

The Wenqi Zhou episode identifies Zhou as a quantitative researcher on gardening leave after roles at Balyasny and Citi. She describes a hybrid use of third-party and in-house sentiment data, with AI/NLP processing high-volume news, and discusses bank proprietary flow data, market-impact calibration, volatility and volume forecasting, risk management, and options/equity interaction (10:19–11:15; 14:45–20:54). These historical, speaker-reported details do not establish a named Balyasny model, vendor contract, current employer, or live deployment.

The ChinaScope episode identifies Tom Liu as ChinaScope’s founder and CEO. Liu describes NLP and machine-learning work on Chinese news beginning around 2012, including metadata extraction from fragmented news sources and later out-of-sample corpora for information extraction (10:48–18:57; 30:31–31:32). He also discusses China-specific data access, regulatory constraints, government data liberalization, and the domestic quant market (51:48–55:45; 1:14:43–1:17:48). This is vendor and market context, not evidence that a named Chinese hedge fund used a particular dataset or model.

The second Abraham Thomas episode adds an industry-level LLM perspective from the Quandl co-founder. Thomas places LLM value in large-scale document summarization and research assistance (28:43–29:26), while distinguishing routine data wrangling and feature engineering from investable-signal discovery (33:29–34:02). He discusses training-data demand, security-master data as an enabling key, and the growing difficulty of web scraping (54:35–59:12). These are industry observations, not evidence of a named fund’s model, training corpus, permissions, or returns.

All five recoveries use automatic, non-diarized ASR. Episode identity, dates, and named roles are supported by the publisher catalog; substantive statements remain medium-confidence navigation evidence. Full transcripts remain private. See the catalog recovery note and the corresponding ledger records.

September 1 additional Alternative Data Podcast catalog recovery

Six further episodes extend the public record beyond named hedge-fund titles. The Flywheel Alternative Data episode identifies James Griffiths as General Manager and describes daily SKU-level e-commerce observations, data stitching across categories and countries, and initial long-short fundamental demand (09:29–14:32). Griffiths also describes external partners for defined R&D problems, later in-house R&D, machine-learning and regression workflows, and caution around LLM category classification because investment users require traceability (15:14–19:11; 30:30–35:12). This is dated provider testimony, not evidence of a named customer’s current model or performance.

The Ideate Capital episode identifies Tom Liu as founder, with prior hedge-fund and private-equity experience. He describes web scraping, credit-card and import/export data, and alternative-data signals for quantitative and long-short fundamental clients (07:33–13:27). His current AI thesis connects data transformation and agentic workflow to EBITDA improvement, while emphasizing that fragmented systems and conflicting sources of truth often precede useful AI deployment (30:08–31:50; 38:37–40:05). The example of automated form completion is illustrative and not independently audited.

The Norges Bank Investment Management episode identifies Mark Thompson in Primary Research. He describes a small data-science unit supporting active-equity PMs, surveys in 11 major markets, raw-data analysis increasingly performed in house, Snowflake as the data environment, and LLM search over years of corporate meeting notes to support research and management questions (10:13–12:26; 18:59–20:20; 30:31–32:59). This is a dated speaker account; it does not disclose model names, permissions, trading authority, or AI-attributed results.

The Battlefin episode identifies Tim Harrington and discusses Battlefin’s acquisition of Exabel. It describes a data-exploration workflow combining providers such as Placer, ConsumerEdge, and Datos, a Snowflake-linked data-science track, and attempts to shorten data trials through preloaded signals and KPI exploration (34:54–39:26). It also records security and licensing concerns around putting customer data into AI systems (36:58–37:31). The episode is company/speaker-reported product context, not an independent deployment or performance disclosure.

The Tim Baker episode adds a market-data infrastructure route. Baker describes acquiring IEX Cloud assets and a platform for normalizing streaming and slow-moving data, point-in-time availability, complex-event processing, and API delivery across prices, fundamentals, and news (28:32–32:39). The episode describes a rapid technical lift-and-shift while the post-acquisition product plan was still forming (38:27–39:30); it does not establish an AI model or trading system.

The Brian Peltonen episode identifies Peltonen as a former Fidelity Director of Data Analytics and Parcosm co-founder. He describes demand-driven web scraping, credit-card data, compliance review, noisy signals, and a data-science request backlog at Fidelity (23:08–28:54; 34:30–40:19). He describes Parcosm’s proposed entity-resolution and taxonomy/knowledge-graph work across e-commerce, social media, and corporate filings, plus technical and advisory collaborators and meetings with funds and data providers (45:57–53:34). This does not establish current Fidelity systems, Parcosm model details, data rights, permissions, trading authority, or performance. See the third recovery note.

All six are automatic, non-diarized ASR recoveries. Episode identity and dates come from the publisher catalog; substantive statements remain medium-confidence navigation evidence and full transcript bodies remain private.

September 1 second Alternative Data Podcast catalog recovery

The same catalog comparison identified a further six useful episodes. The Live@Neudata panel names System2, AI Liftoff, Cybersyn, and Schulte Roth & Zabel participants. It discusses international and Arabic-language data, LLM-assisted structuring of unstructured material into features, and task-specific use of multiple LLMs (02:08–02:40; 07:33–08:55; 18:23–22:59). A panelist’s $15–20 million annual data-effort figure is explicitly a back-of-the-envelope opinion, not a named fund budget (11:51–13:27). The panel does not disclose a customer, model, contract, permission, production endpoint, or investment result.

The Mark Ainsworth episode identifies him as former Head of Data Science at Schroders. He describes a Data Insights Unit growing from one person to more than 25 and serving dozens of investment teams, with a question-led workflow for alternative-data analysis (23:45–25:41; 30:36–33:32). He also distinguishes data used to fill fundamental research blind spots from short-horizon quantitative signals (35:57–37:43). This is historical speaker-reported context, not evidence of current Schroders staffing, a current model, an agent, permissions, or performance.

The Alex Izydorczyk episode identifies Izydorczyk as former Partner and Head of Data Science at Coatue and later founder of Cybersyn. He describes a shared Coatue data platform and a Cybersyn service that acquires, cleans, maps, transforms, and combines enterprise data for delivery through Snowflake (07:14–08:58; 15:49–18:42). He discusses properly licensed proprietary data as potentially useful for LLM training, but does not identify a customer corpus, model, or fund deployment (20:50–23:04).

The RavenPack episode identifies Peter Hafez as Chief Data Scientist and describes RavenPack’s conversion of unstructured text into structured financial analytics. The discussion covers rules and ML, entities, events, sentiment, roughly 6,800 event types, and point-in-time/look-ahead controls (08:40–10:17; 16:21–22:17). These are dated vendor-reported product and methodology descriptions, not a customer-specific model, contract, permission, or result.

The Synthesis episode identifies Olga Kane and describes U.S.-equity statistical arbitrage across thousands of names, machine-learning research, and out-of-sample checks for overfitting (12:48–16:00). It emphasizes data history, coverage, price, documentation, and portfolio fit when combining alternative and conventional signals (09:28–10:11; 25:02–27:23). No foundation model, training corpus, agent permission, or AI-attributed performance is disclosed.

The System2 episode identifies Matei Zatrianou as a former King Street analyst and System2 founder. He describes starting a data team at King Street and using credit-card data to forecast company revenue, with legal and compliance questions arising as the work entered investment discussions (14:25–16:14). He describes System2 as an external data-science team for fundamental investors and hedge funds, including work on data engineering, modeling, vendor negotiation, and a collaboration with Eagle Alpha (24:50–27:57; 37:39–38:18). This does not establish current clients, model versions, data rights, agent permissions, trading authority, or performance.

These six recoveries use automatic, non-diarized ASR. Episode identity, dates, and named roles are supported by the publisher catalog; substantive statements remain medium-confidence navigation evidence. Full transcript bodies remain private in the durable vault. See the second catalog recovery note and the corresponding ledger records.

September 1 firm-specific Alternative Data Podcast cohort

The Balyasny episode identifies Carson Boneck and describes a cross-strategy data group, an enterprise dataset catalog, and Antenna, a provider-submission workflow that returned risk-model and residual-return backtests (24:16–27:24; 36:41–38:20). It also describes PM data use as confidential IP, with only some data broadly catalogued (39:59–40:40). This is dated speaker-reported evidence and does not establish current Balyasny systems, permissions, or performance.

The Man Group episode identifies Hinesh Kalian as Director of Data Science and describes AHL’s six-month data-science pilot becoming a group-wide function spanning sourcing, ingestion, platform construction, evaluation, storage, transformation, cataloguing, and data products (15:35–19:20). It emphasizes point-in-time history, validation, entity tagging, and tools that let PMs combine data without becoming data scientists (20:08–26:29; 36:00–37:23). No current model or performance disclosure is made.

The AQR episode identifies Avi Rosenbluth as a former portfolio manager and describes systematic research using historical data and statistical models. It highlights entity mapping and corporate-action metadata as integration risks for systematic users (01:02–02:19; 30:20–32:33). Rosenbluth describes AI/ML research as exploratory in the dated account, not as a disclosed AQR production system (35:30–36:28).

The PanAgora episode identifies George Mussalli and describes alternative data, machine learning, and NLP in quantitative fundamental research. It includes a Boston Company quantitative lineage, manual entity-to-ticker mapping from 10-K data, and a cloud ML workflow over terabyte-scale pharmaceutical data in the episode’s account (01:09–02:33; 15:30–18:29; 32:01–34:28). These are dated speaker descriptions, not current model or return evidence.

A newer regulatory surface supplies a more current description of PanAgora’s investment process without turning it into a GenAI claim. The August 19, 2026 SEC prospectus supplement says PanAgora is scheduled to become a 25% sub-adviser to the Morgan Stanley Pathway Emerging Markets Equity Fund on September 21, 2026, with George Mussalli named as portfolio manager and Global Chief Investment Officer. The filing describes the quantitative Dynamic Equity process as producing a global-stock ranking through the Contextual Alpha Model, followed by team review, portfolio construction, and controls for regional, country, stock-specific, liquidity, and thematic-risk exposures. This is useful current strategy and governance evidence: it does not identify model class, features, training data, weights, LLMs, agents, permissions, or performance, and the appointment was prospective as of September 2. See the capture note.

PanAgora’s first-party Senior Analyst, Equity Data Science role adds a separate platform signal. The role sits across Alpha Research, Portfolio Construction, and software, and names data integration and cleansing, data-quality confidence and transparency, vendor evaluation, data-science tools, CI/CD, and structured and unstructured datasets. Python and SQL are required, with R, big-data tooling, containers, Kubernetes, and Airflow listed as relevant skills. This exposes the research/data operating interface without establishing an LLM, GenAI system, filled headcount, or investment authority. See the role capture note.

The Maven episode identifies Aaron Cooper as Head of Mid-Frequency Alpha. He describes a systematic-alpha group in which data, infrastructure, algorithms, machine learning, and data science are core components, with shared execution and data ingestion infrastructure (01:02–01:49; 38:07–40:41). The discussion also highlights data reliability, history, and update frequency as constraints. It does not disclose current model families, permissions, or performance.

The Point72 episode identifies Adam Braff as former Chief Data Acquisition Officer. He describes a demand-driven process that starts from an investor question and triangulates datasets, plus longer-horizon talent-related data use cases (23:08–28:54; 23:52–24:34). He also describes teaching and advising across asset-manager, pension, family-office, and hedge-fund settings (20:36–22:44). This is dated former-employer context, not evidence of current Point72 systems or results. See the firm-cohort recovery note.

All six are automatic, non-diarized ASR recoveries. Episode identity and dates come from the publisher feed and linked surfaces; full transcript bodies remain private in the durable vault.

September 1 data-vendor and alternative-data-market cohort

The Exorde episode dates the interview to July 2024 and identifies Mathias Dail as co-founder and CEO. Dail describes a crowdsourced social-media capture network spanning X, Reddit, decentralized platforms, and more than 120 languages, with roughly 40% of collected posts reported as non-English at the time (09:37–10:04). He describes specialized developer-operated workers, protocol checks, and incentives for timely capture (16:32–17:25). This is a vendor-reported description, not an independent completeness or legality audit and not evidence that a named hedge fund uses Exorde.

The ValueStream Ventures episode dates the interview to May 2024 and identifies Greg Neufeld. He describes a venture thesis around proprietary data flywheels, in which software-generated data improves a product or customer relationship and can become defensible (02:39–05:46). He distinguishes companies willing to license data to funds from companies using data primarily to improve their own products, and describes bespoke introductions between portfolio companies and hedge-fund partners when commercially appropriate (11:07–15:10). The episode does not identify confidential fund clients, a live model, or performance.

The Maiden Century episode dates the interview to May 2024 and identifies Qaisar Hasan, founder of Maiden Century, with prior Point72 experience. Hasan describes a dated Point72 workflow in which data was cleaned and converted into KPI forecasts for research teams, followed by a hybrid data-science/fundamental-analyst team (15:27–19:40). He describes Maiden Century’s cleaning, normalization, featurization, and statistical-model process for hedge-fund and mutual-fund customers (27:12–28:42; 31:25–32:02). This is dated speaker-reported context and does not establish current Point72 systems, current customers, model versions, or returns.

The Glacier Network episode dates the interview to March 2024 and identifies Don D’Amico as CEO and founder of Glacier Network, following Neudata. D’Amico describes alternative-data governance as a supply-chain problem involving insider-trading risk, privacy, intellectual property, web-scraping terms, counterparty diligence, written policies, and retrievable records (03:43–10:23). He says the same framework must adapt as technologies such as AI change collection and use (10:23–11:13). This is legal and operational context, not evidence about a particular fund’s controls or AI deployment.

The Fluid Markets episode dates the interview to February 2024 and identifies Kenneth Book as CEO and founder. Book describes a Goldman Sachs corporate-derivatives background and a co-founder with quant experience, then frames the product as a way to obtain exposure to a business segment without owning the entire parent company (02:00–03:17; 08:27–11:29). This is market-structure and founder context, not a hedge-fund AI disclosure; it does not establish adoption, liquidity, model use, or performance.

The Placer.ai episode dates the interview to September 2023 and identifies Ed Lavery in a new role at Placer.ai after Similarweb. Lavery describes de-identified, aggregated mobile location data, AI and machine learning for visitation estimates, and features including demographics, loyalty, dwell time, and movement between locations (10:19–11:18). He describes mapping visits to stores, brands, and tickers, then combining location data with transaction, census, survey, and other geocoded data (12:28–16:52). This is a concrete vendor-reported modality and entity-mapping route; it does not establish a specific fund’s use or performance.

These six recoveries were title-blind at the fund/AI level and are retained as private timestamped ASR. See the data-vendor cohort recovery note and the corresponding coverage-ledger records.

September 1 fund-personnel and international quant cohort

The Darren Voges episode dates the interview to May 2024 and provides a retrospective account of the Yodlee bank-account and credit-card dataset entering investment research. Voges describes income, spending, retailer wallet-share, and customer-journey data, then a long contracting and engineering process with Majestic Research before the data was sold through research to hedge funds (02:10–06:18; 09:30–18:50). This is historical, speaker-reported lineage, not evidence about current fund clients or current systems.

The BRAIN episode identifies Francesco Cricchio and Matteo Campellone as physics-trained co-founders. They describe using AI and NLP to structure news, earnings calls, and company filings for quantitative-fund customers, with consultancy to help clients apply the data (04:07–05:22). Campellone describes complex-systems and statistical-modeling lineage, while treating AI as one modeling approach among several (06:11–10:59). This is vendor-reported methodology and personnel context, not a named client’s deployment or performance.

The Barclays episode identifies Ryan Preclaw and Adam Kelleher and describes a Data and Investment Sciences structure with separate investment-sciences and data-science functions. The account covers central data infrastructure, alternative-data integration, algorithm development, entity-tagging methods work, and a PhD-heavy hiring mix that included physics (01:42–02:25; 04:31–06:54; 10:43–15:14). Adam also describes prior network-theory, information-diffusion, and causal-inference work. It is a dated sell-side account and does not establish current staffing, model inventory, client usage, or investment returns.

The Joe Hung episode identifies Hung as a former QVT Financial and Point72 analyst who later joined Exabel. He describes a bridge between fundamental and data analysts and a move from simple revenue signals toward margins, operating costs, job postings, ESG, and other multi-dataset questions (01:32–02:30; 07:43–09:35). He also says that more commonly available credit-card data requires more careful slicing and combination (09:53–12:22). This is dated former-employer context, not a current Point72 system disclosure.

The Rayliant episode identifies Vivek Viswanathan as Global Head of Research and Portfolio Management at Rayliant Global Advisors. He describes China and emerging-markets quant strategies using machine learning and alternative data, supported by a Hangzhou research team that sources and interprets local material (01:22–02:31; 06:55–08:21). He discusses China-specific daily net-small versus net-large buying categories from a local financial-data platform (08:21–10:00). This is a dated speaker account and does not establish current models, data rights, authority, or performance.

The Bright Data episode identifies Or Lenchner as CEO and describes public-web data infrastructure, large-scale collection, and the then-current Meta litigation (01:24–02:06; 06:28–16:03). It is useful for the alternative-data supply-chain and access layer, but it does not establish a hedge-fund customer or a fund’s AI use. See the fund-personnel cohort recovery note.

September 1 quant-platform and signal-vendor cohort

The CenterBook episode identifies Ross Fabricant and describes an alpha-capture strategy using information about external hedge-fund positions and research to build quantitative portfolios. Fabricant describes a programming-to-quant path, quantitative hedging of fundamental exposures at Third Point, and a later data-science push after an initially factor-based systematic strategy (01:03–01:35; 02:24–03:42; 06:50–07:47). The account also describes Alpha Theory as an internal tool for positions and portfolio constraints before its later commercialization (08:11–10:34). This is dated speaker and publisher evidence, not a current CenterBook model, data-rights, authority, or performance disclosure.

The YData episode identifies Fabiana Clemente, founder and Chief Data Officer, and describes synthetic data for privacy-preserving sharing, class balancing, and simulation of unobserved behaviors (01:29–06:38). It connects synthetic data to financial simulation and rising AI-driven demand for data (03:20–04:32). This expands the research program’s modality and governance map but does not show that a named fund uses YData or synthetic data in production.

The Nowcast episode identifies Masashi Tsujinaka as CEO and describes a Japan-based alternative-data provider using big-data methods for faster economic indicators. The account names a customer mix including the Bank of Japan, research institutes, sell-side economists, and buy-side hedge funds, and describes expanding data sources and partners (02:35–05:08; 07:47–08:46). This is dated vendor and market context; it does not identify a specific fund’s model, data contract, or performance.

The SigTech episode identifies Vera Shulgina, with prior operations and data-strategy roles at Citadel and Two Sigma. She describes SigTech as a cloud-hosted, Python-based front-office quant-research platform covering data normalization, backtesting, and strategy deployment (01:02–01:36; 09:31–11:09), with lineage in an internal systematic platform developed at Brevan Howard (02:23–03:52). This is a dated vendor and personnel account, not evidence of current customer deployments or returns.

The ExtractAlpha episode identifies Vinesh Jha as CEO and founder and describes curated alpha signals and alternative datasets for discretionary and quantitative investors. The episode revisits an earnings-estimate workflow using multiple sources, analyst weighting, and separate gross-add and churn models for KPI forecasting (01:04–03:17). It also points to the HFM European Quant Summit and NNIP as a route to further data-acquisition and combination material. These are dated vendor and conference references, not a current customer model or performance disclosure.

The EPFR episode identifies Cameron Brandt as Director of Research and describes fund-flows and allocations data across mutual funds and ETFs, tracing the provider’s development from post-Soviet Eastern Europe to a global dataset (01:02–01:58). This adds an institutional-flow source and personnel route to the coverage graph; it does not disclose an AI system or hedge-fund deployment. See the quant-platform cohort recovery note.

September 1 AI, satellite, and data-platform cohort

The AI Liftoff episode identifies Tjeerd van Cappelle and describes Spoiler as a deep-learning system trained on quarterly and annual reports plus related press releases. The account says the output combined textual and numerical company information, including revenue and earnings-surprise scores, with backfill from 2002 and earlier data reserved for training (09:22–10:41; 15:12–16:19). It also describes AWS delivery and security concerns (21:08–22:04). This is dated vendor evidence, not evidence of a current fund deployment or model performance.

The SpaceKnow episode identifies Anu Murgai, with prior macro portfolio-management experience including Alphadyne. The episode describes imagery sourced from satellite providers and an in-house computer-vision workflow that converts observations such as material piles, vehicle activity, and loading docks into daily time series. One example describes a three-to-four-day lag and a Russian cement index intended to supplement discontinued official statistics (16:11–19:38). This is dated vendor and personnel evidence, not a current customer, rights, or performance disclosure.

The SynerAI episode identifies Travis Nixon and describes a multilingual-model, English-ingestion NLP workflow that turns large quantities of information into predictive company signals. The distinctive public claim is “directional NLP”: questions about entities are meant to evolve through machine-led updating, with stock-price outcomes used to attach direction rather than relying on generic sentiment (07:56–09:38; 10:34–15:51; 16:03–19:49). The episode says the initial scope was 1,100 NYSE/NASDAQ companies, with roughly 7,000 companies observed for proxy relationships. These are dated speaker claims; current weights, data rights, customers, and performance are not established.

The Michael Watson episode identifies Watson as founder of Chained Metrics and former Managing Director in Equities Engineering at Citadel. He describes distributed computation over datasets reaching tens or hundreds of billions of rows, and an operational requirement that alternative-data systems be reliable, consistent, and ready for every market open (05:33–07:52). He emphasizes connecting engineers to the KPI, ticker, and economic use case, and reducing silos among PMs, data scientists, and data engineers (12:53–15:10; 20:36–21:06). This is dated former-employer context, not a current Citadel system disclosure.

The M Science episode identifies Spenser Marshall and traces M Science to Majestic Research. The interview gives a concrete multi-source pattern: transaction data, web-technology observations, and domain-registration data can each inform the same KPI, providing redundancy and a way to detect source failures (20:30–24:10). It also describes licensed, exclusive, and case-by-case routes for acquiring data (18:50–20:03). This is dated vendor evidence; it does not identify current fund customers, exclusivity, or performance.

The Thinknum episode identifies Gregory Ugwi and describes a web-crawling platform that structures economic activity into queryable tables, charts, alerts, and an API. Examples include job postings, store locations, prices, employee reviews, customer reviews, and other operating indicators; the episode says most datasets reached back to at least 2015 (05:49–11:54). This adds a direct public-web operating signal route to the coverage map, but does not establish current coverage, permissions, models, customers, or forecast accuracy.

The six transcripts and timestamped SRT sidecars are retained privately under the AI/satellite/platform recovery note and archived at the private checkpoint. Full transcript bodies are not published.

September 1 platform-history and data-vendor cohort

The Jason DeRise episode describes UBS Evidence Lab’s early model as a small technical team working with sell-side analysts: domain experts supplied the questions, technical staff tested data approaches, and successful approaches were scaled. The episode also describes expansion from one-project analyst support into a broader data product and direct buy-side/private-equity access around 2019 (09:55–13:44). This is dated organizational history, not current UBS architecture or evidence of investment performance.

The Omri Shtayer episode describes Lagoon’s founding thesis: many investors could see the value of alternative data but lacked the data-science capacity to use it. Shtayer describes multilingual data work, a shift from a marketplace concept toward insight generation, and a 360-degree company view combining alternative and financial data (04:29–07:00; 07:13–12:47). This is a dated founder account; current company status, customers, models, and returns are not established.

The Etna Research episode describes a data-scouting and quantitative-signal platform used predominantly by hedge funds. Marco Jean Aboav identifies data discovery, acquisition, ingestion, testing, compliance, and productionization as sources of time and cost, with a possible one-to-two-year delay even for quantitative shops. Etna’s described response is to assess combined datasets and deliver customized indices rather than simply resell raw data (09:41–16:24). These are dated vendor claims, not independent proof of alpha or current client use.

The Exchange Data International episode adds a market-data and licensing dependency. Jonathan Bloch describes reference data and corporate actions supplied to banks, hedge funds, asset managers, and data-service providers, emphasizing redistribution rights, field mapping, and the operational cost of changing vendors (07:17–10:51). It does not identify a specific fund’s model or deployment.

The G2 episode describes B2B software reviews, buyer-intent activity, ratings, and review counts being used for company discovery, prioritization, and diligence. The episode says an investor-data business grew alongside G2’s core commercial product and included feeds beyond the public interface (12:08–17:16; 18:47–22:49). This adds a software-adoption and enterprise-demand signal route, but does not establish current data access, customer identities, or forecast performance.

The TenderAlpha episode describes government contract-award data across North America, Europe, Australia, South Korea, and selected other markets. The interview reports near-100-million-contract historical coverage since 2010 and daily delivery, then points to digitized bidding data as a possible route to tender-outcome and competitive analytics (07:38–10:27). This is a dated vendor account; it does not establish a fund customer, model, permission boundary, or performance.

The six transcripts and timestamped SRT sidecars are retained privately under the platform-history recovery note and archived at the private checkpoint. Full transcript bodies are not published.

September 1 international and sentiment-vendor cohort

The Starz Data episode describes a Paris-based data-discovery community with a recommendation engine for matching external data sources to corporate, consulting, and investment use cases (01:45–05:23). This is a dated discovery-workflow account, not evidence of current fund customers, model quality, or performance.

The Moojing episode adds a China-specific e-commerce and public-web route. Philip Beck describes crawling public e-commerce pages from 2013 and a historical connection to a Chinese engineer with a Stanford PhD who had built near-real-time data software for U.S. hedge funds (13:08–18:22). The account does not establish current models, data permissions, customers, or returns.

The RepRisk episode describes ESG-risk identification from information outside companies, including machine-learning techniques in the provider’s data-generation history. The interview connects those issues to portfolio entities (08:08–13:29). This is dated vendor evidence, not a current fund deployment or model disclosure.

The SESAMm episode describes a Paris NLP data lake built from millions of sources, with the interview reporting a multilingual mix that included English, simplified Chinese, and Japanese. The system evolved from experiments on text and market prediction toward client-usable NLP tooling (02:07–07:41). These are dated vendor statements; current weights, rights, customers, and performance remain unverified.

The RIWI episode describes low-latency random-domain-intercept surveys and ad tests designed to reach non-habitual respondents across a large international footprint. The interview connects the data to predictive analytics and trend tracking while discussing machine-learning-driven aggregation (03:56–08:29; 13:12–13:27). It does not establish a hedge-fund customer, current model, consent boundary, or investment performance.

The Unacast episode describes aggregated app and publisher location signals, an Oslo data-science team, and models that infer mobility patterns such as store visits. Privacy, aggregation, and delivery into business-intelligence and decision systems are explicit parts of the described workflow (04:18–08:31). This is dated vendor evidence, not proof of current fund use or performance.

The six transcripts and timestamped SRT sidecars are retained privately under the international and sentiment recovery note and archived at the private checkpoint. Full transcript bodies are not published.

September 1 macro, labor, crypto, and private-company vendor cohort

The QuantCube episode describes macroeconomic nowcasting from four data families: text, satellite, geolocation, and structured data. The interview names social media, blogs, news, job-opening text, Earth-observation and atmospheric imagery, vessel and aircraft tracking, and observed prices; it also describes direct Arabic and Chinese-language analysis (06:45–07:05; 07:39–12:20). The speaker reports real-time tracking of more than 80,000 vessels with 15-second updates and says the nowcasts can precede some official releases. These are dated speaker claims, not independent evidence of current models, customer use, or performance.

The SafeGraph episode describes a place-data product centered on high-confidence facts such as operating hours, geometry, category, and location attributes. Auren Hoffman reports more than 30 million places and emphasizes the quality-versus-breadth trade-off, including the edge cases required to model schedules accurately (05:52–09:57). He describes hedge funds and private-equity funds as a customer category, while saying that many investment users treated data as project research rather than an operational product (13:02–14:27). The episode does not disclose a named fund deployment, model inventory, rights, or performance.

The Tansu Demirbilek episode adds a macro-quant personnel route through prior Campbell & Company, Man Group, Soros Investments, North Asset Management, Anderson Global Macro, and Windhaven Investment Management contexts (00:21–01:38; 04:57–06:34). Demirbilek distinguishes discretionary experimentation from systematic use that requires data quality, a long history, and a backtest. He describes PMIs, nonfarm payrolls, yield-curve information, field observations, and parking-lot counts as examples of information that can become a productionized decision tool (03:32–04:42; 08:31–12:49). This is historical personnel/process evidence, not a current employer-system disclosure.

The Santiment episode describes crypto-asset analysis that combines public discussion—from Reddit, Bitcoin Talk, TradingView, and blogs—with on-chain activity and fundamental-style measures such as developer and network use (03:16–06:52; 10:58–11:37; 22:50–24:42). The episode identifies Maxim Balashevich as founder and CEO, but does not establish current model architecture, a named hedge-fund customer, data rights, or audited trading results.

The LinkUp episode describes daily web collection of job openings from company sites and aggregation into company-, sector-, geographic-, and macro-level labor-demand measures (02:07–02:54). Toby Dayton says a forecasting model built during the 2008–09 recovery estimated nonfarm-payroll growth from the job-opening dataset. The capital-markets product required historical storage and was developed with Smart Market Data before a 2014 rollout (06:39–10:46; 13:14–16:27). The episode does not establish current fund use or performance.

The Soleadify episode describes public-web crawling, entity resolution, and private-company profiles across 200 countries and 70 million profiles. Florin Tufan says the system uses large-scale NLP and other AI to extract descriptions, categories, activity tags, locations, news, social profiles, and ESG categories; some tasks use regression and statistics, while most are described as using AI (04:52–11:17). The product is delivered through an API and is described as serving financial services, private-equity, and hedge-fund users. This dated vendor account does not establish a named fund deployment, current model inventory, data rights, or investment performance.

September 4 — AI-generated covariates: a research-quality control for text signals

The NBER Working Paper Inference with AI-Generated Covariates, by Junting Duan and Markus Pelger, adds an important methodological route for any investment process that turns news, earnings calls, or social text into model features. The authors describe how hallucination, look-ahead bias, and model- or prompt-specific error can distort downstream regressions when generated sentiment or classifications are treated as ordinary observed data. Their AI-Powered Inference framework combines human-labeled calibration, bias correction, adaptive weighting across model-prompt pairs, and targeted label allocation. The NBER page reports a news-sentiment and stock-return application; this is a working-paper result, not evidence of a manager’s deployment or a tradable outcome.

The NBER 2026 Methods Lecture and Summer Institute agenda expose a professor network around inference with AI-generated data, including Melissa Dell, Ashesh Rambachan, Junting Duan, Markus Pelger, and Whitney Newey. A public LinkedIn post by Pelger provides a practitioner-facing summary. A public Zihan Lin profile, identified as an algorithm developer at Hudson River Trading, displays engagement with that post; engagement is recorded as a discovery edge only and does not establish HRT adoption, endorsement, collaboration, or a firm research position.

For the tracked-firm research, this creates a concrete quality gate: preserve the original text and publication timestamp, freeze model and prompt versions, calibrate against human labels, test multiple model-prompt pairs, and report estimator sensitivity before assigning economic meaning to a generated feature. The full evidence boundaries are in the academic source note and coverage ledger.

September 4 — T. Rowe Price podcast: agentic research plus the staffing behind it

A title-blind podcast search recovered T. Rowe Price’s first-party The Angle episode on agentic AI, recorded in July 2026. The page publishes an HTML transcript and names Dominic Rizzo, CFA, as a portfolio manager and Frank Shi as an investment analyst. Jennifer Martin, a global equity portfolio specialist, hosts the episode. The transcript is stronger than agenda metadata because the workflow statements are attributed to named firm personnel, but it remains a firm-authored account rather than an independent audit.

Rizzo describes agents that break goals into steps, use tools, check results, and continue working. He says his agents can build charts and models, conduct research against his investment framework, and make trade recommendations. Shi says he uses agents to gather information, review filings and expert transcripts, identify changes in the technology landscape, and compare accounting or management-incentive changes across earnings seasons. These statements establish claimed workflow scope; they do not establish model versions, training data, evaluation design, production access, or order authority.

The most operationally specific disclosure is staffing. Rizzo says reaching a “junior analyst” workflow required two full-time support roles: “John,” described as the data-cleaning and integration “plumber,” and “Albert,” described as an AI specialist who helps build skills. The transcript gives no surnames, formal titles, reporting lines, or public personnel records for either person. This creates a concrete personnel-recovery route without turning first names into asserted identities.

The episode also exposes a useful infrastructure vocabulary: CPU scheduling, sub-agent launch, Python execution, database queries, memory/context, and networking. Those comments are an investor’s public view of agent economics, not proof of T. Rowe Price’s internal architecture. The page does not disclose prompts, model inventory, retrieval sources, permissions, observability, security review, adoption across investment teams, or AI-attributed performance. The full capture and evidence boundaries are in the dated source note and coverage ledger.

September 4 — professor-led embeddings and finance-programme research routes

The academic search now adds a distinct idea layer alongside firm personnel and media evidence. The Chicago Booth Center for Applied Artificial Intelligence finance catalogue lists work on agentic asset pricing, recurrent neural networks for nonlinear time series, financial machine learning, generative AI for transcript risk extraction, bloated disclosures, and asset embeddings. This is a map of public research topics and authors. It does not establish that an investment firm operates any listed model.

One especially useful route is the Becker Friedman Institute paper Asset Embeddings by Xavier Gabaix, Ralph S. J. Koijen, Robert J. Richmond, and Motohiro Yogo. The paper represents firms and investors from institutional portfolio holdings, then compares recommender systems, Word2Vec-style models, and BERT-style transformers. Its benchmarks cover relative valuation, stock-return comovement, and managed-portfolio choice using US holdings from 2005.Q1–2022.Q4, including hedge funds. It also discusses crowded trades, stress testing, generative portfolios, and LLM-generated economic narratives for groups of similar firms. Those are paper-level methods and author-reported benchmark results, not a disclosed hedge-fund system or after-cost performance result.

The research design suggests a different build path from a text-only sentiment pipeline: holdings encode investor behavior, while filings and earnings calls can be used to interpret an embedding neighborhood. A research programme could therefore test point-in-time holdings, filing text, calls, and market data in separate layers, with reporting lags and data rights recorded explicitly. The paper names fixed income, currencies, commodities, and derivatives as possible extensions; each remains an unverified research opportunity until its historical data and out-of-sample design are established.

The Market Genome Project is a public implementation surface linked from Koijen’s research archive. MktGen describes a multimodal model trained on institutional portfolio data, AI-discovered investor and stock groupings, SEC filings, and connections among earnings calls, holdings, and rebalancing, with interpretable narratives. The site also says it is intended for non-commercial research, education, and information, and warns that AI-generated analysis may be incomplete or inaccurate. The reviewed pages do not identify the product team, disclose model architecture or data licensing, name a hedge-fund customer, or establish that the paper authors operate the product. It is tracked as a public product claim and discovery route, not as firm deployment evidence.

The Stanford AFTLab ABFR archive adds a recurring professor-and-video route. Stanford describes the AI & Big Data in Finance Research Forum as a monthly collaboration with the Cornell FinTech Initiative. Its 2026 archive lists Ralph Koijen on optimized agentic AI for asset pricing, Álvaro Cartea on AI bubbles with LLMs, Hui Chen on LLM uncertainty quantification, and a symposium on recovering economic agents’ forecasts with AI. Andrew Chen’s discussion archive links a direct YouTube video for the Koijen session and further discussions on LLM measurement error, machine-learning limits, factor replication, and AI-intermediated-market fragility. The pages establish event, speaker, and video-discovery edges; they do not establish fund attendance, model use, or production performance.

These professor and finance-programme routes are useful for generating testable research questions and finding technically trained people. They should be kept separate from current employment, firm ownership, and deployment claims. The academic source note and coverage ledger record the evidence tiers, unresolved questions, and next captures.

September 4 — FSU professor route: communication style and hedge-fund data quality

The direct Florida State University profile for Ilias Filippou identifies him as an Assistant Professor of Finance and Dean’s Emerging Scholar. His public presentation list connects two research directions that are easy to miss in a firm-name search: Unusual Financial Communication: ChatGPT, Earnings Calls, and Financial Markets, presented at Wolfe’s Global NLP and Machine Learning in Investment Conference and a hedge-fund-strategies conference, and Improving Hedge Fund Returns Predictions: Dealing with Missing Data via Deep Learning, listed at 2026 finance meetings.

The working-paper page and SSRN paper describe a prompting strategy that identifies 25 dimensions of unusual earnings- call communication, including executive and analyst behavior, unusual content, and technical issues. The abstract reports relationships with returns, volume, volatility, option-implied uncertainty, and analyst forecast revisions. These are research findings, not evidence of a fund’s production process.

The hedge-fund-return paper describes BRITS deep learning for missing fund returns and 23 predictors. Its abstract reports improved prediction and a selected-fund exercise with 13.4% annual alpha net of costs. That figure is kept as an author-reported result, not treated as validated investment performance. The same description refers to past and future time-series values, so any live reconstruction must separately test a strictly causal, lagged version; future-value imputation would otherwise create a look-ahead problem.

This route broadens the research questions beyond sentiment: communication style, data missingness, and model-imputation policy may each affect signal quality. The public record still does not establish hedge-fund adoption, model permissions, code, or production results. The academic source note and coverage ledger track the source boundaries and next captures.

September 4 — MIT PI lineage for AI peer grouping

The MIT CSAIL student-poster archive identifies Manish Singh as an MIT EECS PhD student advised by Andrew W. Lo and associated with MIT’s Laboratory for Financial Engineering and CSAIL. It describes research interests spanning healthcare finance, investment management, and sustainable investing, alongside an AI-based industry peer-grouping project.

The Andrew Lo publication record and paper DOI identify the 2022 Artificial Intelligence-Based Industry Peer Grouping System. The public abstract describes AI-based company clusters, continuous similarity scores, and hedged-portfolio construction. It reports higher out-of-sample return correlation but lower stability and interpretability than a standard industry classification in the authors’ tests. This is an academic artifact; it does not establish a hedge-fund implementation or current live strategy.

The useful discovery is the lineage edge: a named finance-AI project, its PI and doctoral researcher, and a possible healthcare-finance application. Poster, code, feature construction, turnover, costs, and point-in-time portfolio formation remain open research questions. Academic affiliation must remain separate from current employment or commercial deployment.

The six complete transcript bodies and timestamped SRT sidecars are retained privately under the macro/labor/private-company recovery note and archived at the private checkpoint. Full transcript bodies are not published.

September 1 regulation, quantamental, China, shipping, NLP, and GameStop cohort

The SEC episode adds a regulatory route. Adam Storch of the SEC Division of Examinations describes the April 2022 Investment Advisor MNPI Compliance Issues risk alert, which prominently highlighted alternative data and Section 204A policies for preventing misuse of material nonpublic information (19:06–21:40). Kelly Koscuiszka of Schulte Roth & Zabel adds that a written policy needs ongoing diligence on data sourcing and proactive application (22:35–24:40). This does not establish any named fund’s compliance program or model.

The S&P episode describes an alternative-data marketplace with schemas, sample data, linked research, trials, and production pipes. Daniel Sandberg describes identifier linking across Capital IQ, CUSIP, CEDOL, and ISIN, plus FDA-approval data linked back to financial identifiers (09:39–17:27). He also describes a Scripps Asia partnership for Japan and Hong Kong earnings-call transcripts and a quantitative research process that vets datasets, prototypes products, and publishes white-paper evidence (16:18–22:23). These are dated publisher and company statements, not evidence of a particular fund’s use or returns.

The DataYes episode describes a Shanghai-based provider combining financial data with AI-finance solutions, NLP, news and research-report sentiment, and sampled supermarket cash-register data (01:47–14:40). Its “quantamental” workflow uses a knowledge graph for analyst-described revenue structure as input to a machine- learning model that combines fundamental and quantitative processes (16:18–17:46). The episode does not establish current model architecture, named customers, data rights, or performance.

The Suez Canal episode contains MarineTraffic and Shipfix interviews. MarineTraffic describes AIS receiving stations, satellite partners, vessel metadata, weather data, historical movement storage, predictive arrival services, port-congestion analysis, APIs, and aggregated exports (02:07–06:30). The event discussion connects movement data to investors, banks, and insurers and records route diversion, delays, fuel, freight, and cargo effects around the 2021 blockage (07:04–14:11). These dated provider accounts do not establish current fund use or model performance.

The Alexandria episode adds historical financial NLP detail. Dan Joldzic describes more than 500,000 finance-analyst-labeled sentences for sentiment, events, and topics, with the goal of replicating expert judgment over larger text collections rather than using returns as both input and output (15:02–17:44). He names early Dow Jones and Thomson Reuters/Lexalytics relationships and describes backtesting data for return, risk, and Sharpe before showing it to hedge funds (07:53–08:07; 12:00–14:27). A claimed FinBERT comparison remains a dated company account pending independent review; it is not treated here as a ranking or current performance result.

The GameStop episode describes SESAMm’s monitoring of Reddit and WallStreetBets at message and stock level. Sylvain Forté says message volume and sentiment appeared before the GameStop price move, and describes rapid dashboards and quantitative files for thousands of stocks (01:33–07:22). The discussion extends the search route to Twitter, 4chan, and local communities in Russia and Japan, while noting the short history and event-specific limitations of systematic use at the time (08:44–11:45). This is a dated speaker account, not an independently reproduced event study or evidence of a named fund deployment.

The six complete transcript bodies and timestamped SRT sidecars are retained privately under the regulation/quantamental recovery note and archived at the private checkpoint. Full transcript bodies are not published.

September 1 bank, ESG, marketplace, data plumbing, and privacy cohort

The Jefferies episode adds a dated bank-side data-strategy workflow. Rayne Gaisford, identified in the episode as Jefferies’ Global Head of Data Strategy, says the first internal customers were fundamental research analysts. His team started with the analyst’s unanswered question and looked for observable proxies—such as mobile foot traffic, web activity, or supplier/customer behavior—to add evidence to an existing company thesis (10:00–15:17). He describes a team of approximately nine people and a “Jeff Data” / “The Compiler” research stream intended to show the source, transformation, and conclusion so clients could adapt the method to other companies (15:49–19:32). The interview places this alongside Jefferies’ prior M Science history but does not disclose current model architecture, model ownership, or named hedge-fund deployments (09:41–10:00; 19:45–21:05).

The ESG Analytics episode describes a SaaS product founded in July 2020 by Qayyum Rajan. The dated account combines news feeds, scraping of news/media sources and company webpages, and social inputs with an in-house NLP taxonomy and sentiment analysis to extract ESG signals outside traditional corporate disclosure (05:17–08:54). Rajan also describes a lower-cost, API-oriented product for smaller teams and startups (05:58–06:47; 12:17–13:07). He notes that ESG-provider scores can diverge because frameworks and coverage differ, with a size bias toward companies that disclose more (23:58–25:23). This is a dated product and model-process account, not evidence of current ownership, fund use, or returns.

The Datarade episode adds a marketplace and data-discovery route. Thani Shamsi describes prior work at Zeotap monetizing telecom-operator data in a privacy-compliant manner and a Berlin-founded marketplace intended to combine personal trust with searchable listings, reviews, and standardized data descriptions (02:48–05:12; 05:53–10:38). The interview identifies hedge funds as one demand center for non-traditional data, but says the marketplace was designed to serve multiple industries (09:20–10:57). The proposed metadata includes history length, attributes, update frequency, missing fields, and delivery format (18:17–20:30).

The JP Gravitt episode records a historical Seven Park Data productization pattern. Gravitt describes Vista-backed proprietary datasets, including web scraping, credit-card, and e-receipt data, packaged with context, data-quality caveats, visualization, dashboards, and “beacon” reports (03:49–06:12). He distinguishes raw-feed buyers from investors who need interpretation and describes combining multiple datasets into a company-level view (06:19–07:37; 14:19–15:36). The episode also records Jumpshot dependency and privacy-related commercialization risk (11:28–12:39; 33:21–34:05). A speaker estimate that data represented about 80% of investor input is retained only as a dated opinion, not a market statistic (08:06–09:47).

The Crux episode adds a data-infrastructure signal. Philip Brittan describes Crux as an integration layer rather than a data creator or analytics terminal. The problem is repeated onboarding across APIs, FTP, loaders, S3 drops, schemas, validation, and monitoring (05:52–08:09; 16:45–17:05). He gives a dated company/founder estimate that upfront engineering and quality-control work consumed roughly 80% of firms’ data effort, leaving the remainder for research, models, or risk systems; this is not treated as a controlled benchmark (17:18–20:05). He describes one API, Python access, cloud-warehouse integrations, and connections to AI platforms, and names Two Sigma as an early customer without specifying its datasets or models (25:11–29:16).

The LeapYear episode adds a privacy-preserving analytics route. Garrett Long describes differential privacy software for statistical analysis over sensitive datasets, including query-level privacy accounting and the possibility of differentially private regressions and machine-learning models (03:23–05:46; 07:04–08:36; 11:26–12:47). The proposed “trustless analyst” workflow lets data owners retain control of raw records while external users receive protected results or API access (15:18–18:56; 25:57–26:58). References to related techniques at LinkedIn and Apple are historical context, not evidence of LeapYear deployment or of any hedge-fund use.

Together these episodes expand the research checklist: trace the route from an investment question to an observable proxy; record how raw data is turned into reports or models; search for schema, validation, and cloud-integration roles; and treat privacy, source continuity, and rights as first-class constraints. None of these accounts establishes a named fund’s live model, alpha, returns, or comparative position.

The six complete transcript bodies and timestamped SRT sidecars are retained privately under the bank, ESG, marketplace, data-plumbing, and privacy recovery note and archived at the private checkpoint. Full transcript bodies are not published.

The feed reconciler then recovered two remaining historical Acast gaps whose RSS links incorrectly pointed to the sponsor site. The TrueSource episode identifies Edward Egusev and describes prior work at Deutsche Bank supporting a trading platform, Mail.ru search, Google building a machine-learning and data-analytics platform, and Glean enterprise search (01:44–02:56). He describes TrueSource as a self-serve commercial layer for data providers: storefront UX, subscription or per-record pricing, and API delivery, rather than a source of data or an initial distribution marketplace (03:05–04:53; 08:09–10:21). The episode gives a dated company estimate that alternative-data providers were approximately 20–25% of customers at that time; it is not an independent market-share measure (14:15–14:31).

The DataBoutique.com episode identifies Andrea Quattrito and describes a Milan-based luxury-retail data operation that collected online price, availability, distribution, and product information, built historical series, and cleaned large volumes of web data (00:27–02:02; 06:35–12:47). The episode describes corporate, private-equity, fundamental-research, and quant use cases, with quant users seeking clean raw data for sector-specific models (26:41–28:55). It also cautions that website structures and interpretable signals change over time (29:15–32:20). No named fund deployment, current model architecture, data-rights claim, or performance result is established.

The two additional complete transcript bodies and timestamped SRT sidecars are retained privately under the bank, ESG, marketplace, data-plumbing, and privacy recovery note and archived at the private checkpoint. Full transcript bodies are not published.

An earlier metadata-only China route was also repaired. Chinatown 2.0 — Ep. 2 identifies Robbie Yan as a co-founder of a China-based quantitative fund that was primarily trading Chinese futures at the time (01:40–01:49). The fund is not named in the recording and is not assigned an identity. Yan describes a workflow covering data collection, bias-aware cleaning, missing-data handling, storage, hypothesis testing, signal selection, and fault-tolerant production engineering (02:52–04:19). He separates statistical errors such as overfitting from software bugs and describes graceful shutdown as a safety requirement (05:17–06:22). The episode also compares the data, machine learning, backtesting, and infrastructure loop with Google advertising systems (09:27–10:22). These are dated practitioner statements, not evidence of a named model, current fund status, live permissions, or performance.

The timestamped transcript and SRT are retained privately under the Chinatown 2.0 capture note and archived at the private checkpoint. Full transcript bodies are not published.

September 2, 2026 — title-blind conference and personnel expansion

The Quaint Quant Conference 2026 speaker page adds a newly tracked personnel surface. Its organizer-published biographies describe Daniel R Barrera as a quantitative researcher at LMR Partners working on statistical models and proprietary reversal signals, with prior MSCI-Barra factor-model experience; Jacob Bowers as a BlackRock vice president supporting systematic funds with inflation hedging, backtesting, and machine-learning exploration for portfolio optimization; and Jeffrey Ryan as a former Citadel GQS builder associated with data, risk, alpha, and HPC systems before his current QUANTkiosk work. The page also describes John DeTore’s ARGA strategic-R&D role, including tools that bring additional AI resources to investment professionals and his historical MIT Sloan teaching. These are dated organizer biographies, not evidence of current firm-wide deployment, model ownership, data rights, permissions, or performance.

The conference agenda places Barrera in sessions on scientific methods and ML/AI use cases, lists Bowers for a conformal-methods session, and identifies a Point72-sponsored optimal-index-options project presented through Lehigh University. Sponsorship, programme placement, and co-presence do not establish collaboration or live investment use. No authorized recording was located in the reviewed pages, so this route contributes personnel and conference metadata rather than a transcript.

The MacroMinds 2026 agenda adds a title-blind New York conference route. Its June 4 programme included “Finding Value — On Both the Long and Short Side — in AI,” with Jim Chanos of Chanos & Company and Val Zlatev of Analog Century Management, plus a risk-management discussion featuring Joanna Welsh, identified as Citadel’s chief risk officer. The official speaker biographies separately describe Welsh’s prior Tudor risk role, Zlatev’s historical quantitative-tool work at Quentec, and Ulrike Hoffmann-Burchardi’s former Tudor portfolio-management role. These are organizer-published programme and biography claims, not evidence of current firm-wide AI deployment, model ownership, permissions, or performance. No authorized recording or transcript was located.

An academic-to-manager route adds Zhilin Zhang’s affiliation with Lumos Alpha in a 2026 arXiv review. The review covers sentiment, financial reports and earnings calls, cross-stock relationships, price-series tokenization, and multi-agent systems, while emphasizing timestamp leakage, horizon design, illiquidity, evaluation, and limits of predictability. Lumos Alpha’s About page identifies Zhang as Founder & CEO and describes AI-assisted portfolio allocation and a risk-mitigation process; his public profile links the Lumos Alpha and UC San Diego affiliations. This establishes a public researcher-to-manager and positioning route, not proof that Lumos Alpha deploys every method in the review or that its self-reported performance is independently verified. See the capture note.

A title-blind search also surfaced the 7 Systematic workshop, an education and data-business surface associated with former Citadel quantitative researcher Jeffrey Ryan and University of Illinois Chicago lecturer Justin Shea. Its June 2026 agenda names Gemini and NotebookLM as examples of AI agents for turning discretionary ideas into a research assembly line, alongside SEC-filings and other unstructured-data ingestion, feature engineering, backtesting, and risk review. The public Apple Podcasts listing for a Jeffrey Ryan episode provides a separate former-Citadel podcast lead. Neither page names the crypto hedge fund referenced in the workshop biography or establishes current Citadel deployment, customer adoption, model ownership, agent permissions, or trading authority. See the capture note.

The Gandalf Symposium 2026 adds an Italian/English regional route that a firm-name search would miss. Its June 27–28 programme covers open-source quantitative finance, AI-driven asset allocation, trading agents, evolutionary methods, XGBoost, regime classification, and walk-forward analysis. The organizer names Tom Starke of AAAQuants, Giovanni Trombetta of Gandalf Project and Rocket Capital Investment, and several practitioners with historical or consulting relationships to quantitative and algorithmic funds. The biographies and programme establish public speaker and method descriptions; they do not establish current fund affiliations, production deployment, data rights, permissions, or performance. No authorized recording or transcript was located in the reviewed pages. See the capture note.

The Turnleaf Analytics recap of the Imperial College Hedge Fund Conference adds a separate research-method route. It reports presentations on decision-aware end-to-end portfolio learning, machine-learning forecasts designed to avoid lookahead bias, and narrative-attention features built from long news histories. These are reported conference takeaways, not evidence that Turnleaf or a named fund implemented the methods, and the recap is not a substitute for the underlying papers or proceedings. See the capture note.

September 2, 2026 — Citadel GQS and T. Rowe Price personnel routes

The Citadel GQS personnel route adds a first-party recruiting interview with Navneet Arora, dated April 20, 2023, alongside his current Citadel leadership profile and the GQS PhD Fellowship. Citadel identifies Arora as head of Global Quantitative Strategies, says he joined as a senior quantitative researcher in 2013 and was named GQS head in 2019, and lists prior model-based-credit and quantitative-research roles at American Century, BlackRock, and Barclays Global Investors. The fellowship page adds a research-talent route spanning mathematics, statistics, physics, electrical engineering, and computer science. These pages establish personnel, academic-recruiting, and research-language signals; the embedded video was not transcript-recovered, so they do not establish a GQS generative-AI system, dataset, model, agent permission, or performance result.

The T. Rowe Price AI leadership note records a separate first-party disclosure dated August 13, 2026. T. Rowe Price names Vinit Agrawal to lead Investment AI Solutions, with a remit covering AI strategy, agentic products, education and adoption, strategic partnerships, and experimentation and research. It names Sal Dhanani for Global Distribution AI Strategy and Transformation; Mathieu Lorentz for T. Rowe Price Labs within AI Engineering led by Chris LePre; and Prabhakar Bhogaraju for Enterprise AI, Model, and Technology Risk. The firm’s April investment-process article also describes agent-based workflow experimentation and tool integration across research, content analysis, financial modeling, and coding, while its July alternative-data article describes web-scraped pricing/inventory data combined with field observations and primary research. This is unusually explicit public organizational evidence, but it still does not disclose model inventory, training data, partner contracts, permissions, live adoption by every team, or AI-attributed returns. T. Rowe Price is included here as an asset-management comparison point, not as a hedge-fund classification or a ranking.

The MSCI Private Assets R&D job description adds a vendor-side research and hiring signal. The Executive Director, AI Research role specifies embedding-based similarity over unstructured text and filings, LLM-assisted nowcasting and cash-flow forecasting, agentic workflow prototypes, and recurring automation such as commentary generation and model-parameter re-estimation. It names Burgiss, IPD, and RCA data assets as part of the private-markets research context and asks for production ML, RAG, unstructured-data, and agentic-framework experience. This exposes the task vocabulary and data lineage that can be searched across institutional investment vendors; it does not establish that the role is filled, that the capabilities are live, or that a hedge fund uses a particular model.

The Arcesium investment-operations note provides a complementary implementation vocabulary. Abhishek Agarwal’s June 2026 article describes structured data and ontologies, a clean access layer such as MCP, encoded domain skills, and a purpose-built orchestrator. It describes using an LLM to develop a workflow, validating it against real data, and promoting the accepted logic into deterministic Python or SQL with human review before consequential action. This is vendor architecture guidance, not evidence of a named customer’s deployment, model inventory, data rights, or investment outcomes, but it gives a practical way to interpret “Head of Automation,” data-platform, and operations-engineering titles in future firm searches.

The title-blind queue also recovered L&G Asset Management’s Episode 408, “Unpacking the US Investment Outlook,” from its public Audioboom feed after the YouTube locator required an authenticated request. L&G identifies Jason Shoup as US CIO, Jason Becker as Head of US Credit Strategy, Anthony Woodside as Head of Multi-Sector Fixed Income & Investment Strategy, and Dan Dreher as Solution Strategist. In the recording, Shoup frames AI as a multi-year competitive force that can move from infrastructure spending into industry-level disruption and credit dispersion (01:22–03:29; 07:19–08:01); Becker discusses second-order AI exposure in investment-grade credit (09:09–10:39); and Woodside links AI infrastructure spending to capital-market supply and listed-infrastructure portfolio considerations (11:34–12:01; 18:46–20:54). This is asset-manager market commentary, not evidence of L&G’s internal AI systems or a hedge-fund deployment. The timestamped transcript and SRT are durably retained privately at the private checkpoint; full transcript bodies are not published. See the capture note.

September 2 title-blind media and personnel expansion: Voloridge, Hudson Bay, and Quest

The MIT CBMM seminar page adds a dated Voloridge route that does not appear as a hedge-fund or AI podcast. The page records a 2017 seminar by David Vogel and exposes searchable discussion of ensemble machine-learning methods, portfolio construction, risk management, and a historical MIT-alumni staffing statement. The 2026 Vitality Vision episode supplies a separate current-era personnel and career-progression route naming David Vogel, Barry Miller, and Daniel Hammack. Its public MP3 was recovered and locally transcribed into 1,096 timestamped ASR segments, with private archival checkpoint de3a1382. The local ASR records a speaker-reported statement that LLMs were actively used for coding velocity, clearly naming Codex and Copilot; a third product is rendered as “Cloud Code” and is not normalized without human audio confirmation (approximately 00:25:06–00:25:21). It also describes separate alpha, risk, volume/trading-cost, and cross-asset-correlation models (approximately 00:31:21–00:32:11). The ASR is navigation evidence, not automatically exact quotation. These sources establish dated public statements and personnel clues; they do not establish Voloridge’s current model inventory, GenAI architecture, permissions, or performance. See the capture note.

Hudson Bay’s first-party site now contributes explicit but bounded AI-process language. Its Our Approach page describes the Deal Code System, Gerber Statistic, and RMon as proprietary systems, while the Research & Insights archive exposes a research and academic-collaboration surface. In a timestamped Sander Gerber interview, Gerber discusses AI as useful for research and drafting while emphasizing checking for errors and the continuing importance of human judgment (approximately 39:48–43:34). The combined record describes process, risk analytics, and a human-review boundary, but does not identify an LLM, training corpus, agent permission map, or AI-attributed investment outcome.

The Blockworks episode with Mike Harris adds a Quest Partners route found through generic quant-trading media. Its August 2024 metadata and timestamp markers cover CTA trend following, statistical arbitrage, multi-strategy trading, AI trading models, and risks in training models with AI. The YouTube locator and Apple catalog entry provide additional discovery paths. The public record does not establish which methods Quest uses, whether any named model is in production, or any performance result.

September 2 title-blind personnel and research routes: Verition, Systematica, Squarepoint, and Rokos

Verition’s public research trail now includes Peter Pommergård Lind’s profile and the Aalborg University thesis record. The profile reports a quantitative-research role and discusses neural networks, tree regression, differential machine learning, and option valuation; the university record confirms the 2024 thesis and names Frederik Steen Lundtofte and Orimar Sauri Arregui as supervisors. This is useful personnel and academic-lineage evidence, not proof that the thesis methods are Verition production systems.

Systematica gains a public Shanghai equity-alpha and research-infrastructure route through Yanzhong Huang’s BagelQuant profile and its linked GitHub organization. The profile describes factor research, predictive modeling, portfolio construction, robustness, point-in-time discipline, earnings-call NLP, and volatility forecasting. The activity is self-authored and partly open-source; it is not treated as evidence that Systematica owns or deploys those personal projects. A separate Raymond Ji profile identifies a Squarepoint quantitative-research role and describes dated personal projects involving newspaper NLP, recurrent networks, sentiment, volatility, and Twitter text. The profile does not establish that those projects were Squarepoint work or remain current.

The current SBAI London Annual General Assembly page adds a dated event and personnel route for Systematica. It places Grégoire Dooms, identified as a Portfolio Manager, in a session on what investors expect from managers regarding AI policy and AI-use concerns. The biography says he leads the firm’s Data Science effort and links his background to Brown University computer-science and AI research and a UCLouvain PhD. This sharpens the public personnel and governance map, but it does not identify a model, training corpus, evaluation record, agent permission, or live investment deployment.

Finally, the AEA JOE listing for Rokos records a Senior Economic Data Scientist vacancy posted in November 2025 and inactive in January 2026. It describes macro forecasting, alternative datasets, data-science leadership, and infrastructure supporting economists and investment decisions. It is evidence of a dated hiring requirement, not evidence that the role was filled or that a particular model or dataset is live. See the personnel-route capture note.

September 2, 2026 — German allocator workflow route

The German-language BAI newsletter adds an allocator-side view of AI in alternative investments. Its article maps ML and LLM use across three stages: planning, including liquidity forecasting and scenario analysis; implementation, including screening databases and extracting PPMs, LPAs, and track records into structured investment material; and monitoring, including qualitative updates and sentiment. It emphasizes a consistent data layer and digitized institutional workflows as prerequisites for useful AI.

The article cites private-equity research on text-based fund selection and tone in GP reports. Those cited findings are retained as research leads, not as evidence about a named hedge fund or liquid-market alpha. The BAI route does not disclose a manager’s model, training data, data rights, permissions, deployment, or performance. See the German-language capture note.

September 2, 2026 — Italian Euklid technical and entity route

An Italian-language search surfaced a dated Money.it interview with Antonio Simeone and linked first-party material from Euklid. The 2019 interview attributes a historical Euklid description to Simeone: more than 260 variables, multiple algorithms per asset, and genetic-algorithm and swarm-intelligence methods. Euklid’s current site describes customizable time-series forecasting for financial assets and names Modelomni as an optimization partner. A public Euklid/Modelomni technical document sets out a claimed train/test split, preprocessing, soft-computing models, meta-heuristic optimization, signal aggregation, and daily or weekly signal delivery to Modelomni for adaptation and execution. These are historical or firm-authored descriptions; no model code, feature history, leakage test, permission map, live-capital evidence, or independent performance audit is published.

The public Companies House record adds a useful continuity check: EUKLID LTD is active and classified under fund management, while Simeone’s recorded directorship ran from January 2021 to January 2023 and Pierfrancesco Savona is the current recorded controller. The FII 2022 programme identifies Simeone as co-founder and algo trader. A separate ASIC notice names an Antonio Simeone in an Australian licence-cancellation matter; the Australian entity and the UK Euklid record are not merged without entity-level corroboration. See the Italian capture note.

The same Italian-language pass recovered a separate historical product route: the HI QuantWave disclosure describes a 2018 Hedge Invest quantitative, multi-strategy, multi-asset fund managed by Numen Capital and associated with Marco Jean Aboav. The publication attributes to the product a proprietary AI system using Big Data, including Twitter, cloud computing, and simulated trader styles. Its stated return and volatility objectives are product targets, not realized performance. The source does not establish that the fund remains live, identify the models or data rights, or verify execution, current personnel responsibility, or results; see the capture note.

September 2, 2026 — Omphalos Fund: Polish R&D and Dutch research surfaces

The foreign-language pass found a substantial Omphalos Fund trail that was absent from the ledger. Omphalos’s fund page describes a Luxembourg special limited partnership using software from AI Investments and managed by AITR Sarl. Its technology page describes testing, training, optimization, simulation, paper trading, alpha evaluation, lifecycle monitoring, and human oversight of risk and exposure.

A Polish National Centre for Research and Development project record names AI Investments Sp. z o.o., inbestMe, the University of Oslo, and Ulm University in a 2019–2021 project. It identifies Transformer and Differentiable Neural Computer technologies in an AII portfolio and risk platform and reports fund results from January 2022 through May 2024. Those performance figures are beneficiary-reported on a government-hosted project page, not independently audited here.

The original Polish 99 Twarzy AI episode with Paweł Skrzypek adds a podcast-level route. Its LLM-generated transcript summary describes five-to-ten-day forecasts and many smaller trades, but those details remain speaker-reported until the audio is independently checked. The AI Investments news archive also records a 2019 presentation on hybrid statistical/ML time-series forecasting, serverless computing for model workloads, and a 2021 announcement that Omphalos began using AI Investments technology. These are historical development and media signals, not a complete current model or control record.

An Eurex derivatives-forum report attributes a further account to CEO Borno Janekovic: reinforcement-learning “artificial portfolio managers” trained in simulators, approximately 5,000 monthly trades, and no human portfolio managers for investment decisions, with human oversight of agents and execution context. This should be read alongside the fund page’s statement that experienced traders monitor risk and exposure; the public material does not provide an auditable control diagram.

The firm’s Dutch Behind the Cloud index and 2026 chapters expose a public research vocabulary around market sensors, point-in-time data, revision and timestamp leakage, survivorship bias, microstructure, liquidity, volatility regimes, dispersion, tradability, and sensor-health monitoring. The leadership page names Paweł Skrzypek as co-founder/CTO/COO, Borno Janekovic as CEO, and Tomasz Przeździęk as CDO/Quant IO, alongside additional founders and advisers. This is the firm’s public description of its architecture and people; it does not establish model quality, current permissions, audited returns, or collaborator endorsement. See the Omphalos capture note.

September 2, 2026 — regional and vendor expansion

The regional-language pass adds a Brazil–Mexico–Chile route that is not visible in an English-only firm search. Kadima’s Portuguese site describes systematic, algorithmic, and factor-based products across Brazilian equities, international markets, long-short, fixed income, and other strategies. The FGV EMAp RiO 2025 recap and EBFin 2026 programme add academic coverage of generative financial time series, deep hedging, reinforcement learning, factor testing, credit-risk ML, and futures-price prediction. The UNAM seminar page attributes a generative-neural-network portfolio method to Fintual. These are regional manager, academic, or attributed company descriptions; they do not establish a common production architecture or independently verified outcomes.

Additional regional routes include Stance’s Kyle Balkissoon profile, which describes machine learning and portfolio construction for systematic sustainable equity; the FIX Japan speaker page, which adds Japanese algorithmic-trading and AI/DX personnel; and the KPMG Union Investment episode, which describes AI-supported price validation with human four-eyes control. Command Capital’s Chinese interview describes a future intelligent-agent direction for combining research and trading information, which remains an intention rather than evidence of current deployment. See the regional source note.

The vendor pass adds architecture and data surfaces that are useful for interpreting future firm disclosures. Microsoft Research’s R&D-Agent-Quant article and Trade in Minutes paper describe research loops and multi-agent trading research as published technical work. Arcesium’s MCP article describes structured ontologies, packaged skills, deterministic execution, and auditability. Databento’s Blue Ocean ATS release exposes a market-data route with depth, trades, and reference data. Google Cloud’s Rogo case study and the Yields case study add named research-retrieval and model-governance accounts. These are vendor or academic reference points, not proof of deployment by any tracked firm. See the vendor route note.

September 2, 2026 — employer, lab, and regional route expansion

Additional title-blind searches add public technical and organizational clues across the wider tracked universe. CFM’s public deep-reinforcement-learning repository accompanies an arXiv-linked portfolio-optimization study, while a CFM technology project-management role explicitly mentions AI agents for project-management workflows and collaboration with Trading, Predictions, Research, Infrastructure, and Compliance. The repository is historical technical evidence; the vacancy is an intended operational workflow. Neither establishes a current investment model or deployment.

Arrowstreet’s AI Platform Engineer and AI Security Engineer vacancies expose a detailed intended control-plane vocabulary: managed LLM inference, Bedrock, Azure OpenAI, MCP, model routing, RAG, vector databases, agent frameworks, non-human credentials, delegation, audit logs, sandboxing, prompt-injection controls, and cost dashboards. The recruiting post corroborates hiring intent. These pages do not establish that the roles were filled or that the systems are live.

The MIT CSAIL Applied AI internship page for Balyasny explicitly describes DSAI support for investment teams, generative-AI solutions for textual data, LLM and embedding-model fine-tuning, LangChain, RAG, vector databases, and a training block before desk placement. Point72’s Cubist ML Engineer role and AI Solutions Architect role separately mention production-support agents, synthetic data, MCP research workflows, forward-deployed use-case discovery, and adoption tracking. These are employer-authored requirements and operating-model signals, not model inventories or performance evidence.

QRT’s Cambridge partnership announcement describes QRT Labs activity across Cambridge, Imperial, and Oxford, involving more than 70 early-career researchers across quantitative and computing disciplines. A QRT research-community post adds public links to ICLR, ICML, CppCon, LLMs, reinforcement learning, interpretability, graph neural networks, reasoning, and optimization. This establishes an academic-network route, not trading-system use. Man Group’s applicant privacy notice is a separate governance route that discloses operational AI for research, process streamlining, meeting records, compliance, and internal training while describing safeguards and retention. It does not name an investment model or vendor.

Regional and vendor searches add further discovery paths. The AlphaGrep–IIIT Hyderabad programme names Franklin Templeton and AlphaGrep speakers in a curriculum covering neural networks, transformers, deep reinforcement learning, and portfolio optimization. The KCMI programme contributes Korean-market coverage for transformer investing, agentic AI, AI quant, direct indexing, and governance. The IDEA profile for Guo Jian describes Quant5.0, e2eQuant, Alpha-GPT, automated quant research, scenario generation, and context graphs. These are event or institutional research signals; the pages do not prove firm deployment.

Vendor surfaces expose implementation vocabulary without resolving customer internals. Databento’s practitioner profiles cover former Squarepoint/AQR practitioner Alex Reyfman and related profiles covering corporate-bond automation, market microstructure, data cleaning, headline processing, and AI/LLM limits. Its Temple Capital customer page names a systematic crypto/futures data use case. The Snowflake quantitative-research webinar and NVIDIA multi-agent signal-discovery article add portfolio optimization, backtesting, Snowpark, and a reference architecture with signal, code, and evaluation agents. These sources are useful for vocabulary and architecture comparison; they do not establish customer-specific deployment, permissions, or results. See the expanded source note.

September 2, 2026 — AI-native fund reporting, an agentic-investment competition, and an AWS reference repository

A new title-blind route is a March 2026 Bloomberg News report reproduced by Advisor Perspectives about Epicenter Capital, a reported three-person fund launched by former Coatue trader Rahul Kishore. The report says an AI system named Eve was connected to email and trade information, created and assigned bot tasks, wrote code, processed more than 13,000 company disclosures, listened to podcasts, monitored social and news inputs, and generated a morning podcast. It also reports a document-synthesis task completed overnight and a Claude Code suggestion to move infrastructure from AWS to Cloudflare. These details are attributed to people familiar with the matter and an investor letter seen by Bloomberg; they are not a first-party technical specification. The report does not disclose Eve’s architecture, training data, data rights, evaluation design, audit trail, agent permissions, trade-approval boundary, fund size, or performance. Reported information-consumption and cost-reduction figures are not independently verified. See the source note.

The Society of Quantitative Analysts’ Agentic AI in Investment event page provides a separate event-level route. It lists Jonathan Berkow of AllianceBernstein, Reha Tutuncu of Point72, Rui Ding of Graham Capital Management, Revant Nayar of FMI Technologies, and Christos Koutsoyannis of Atlas Ridge Capital on a panel, alongside a keynote on agentic AI and knowledge-graph reasoning. The page says 120 participants across 62 teams worked on questions involving asynchronous data, ML/LLMs in fundamental research, transaction costs, and portfolio construction. The public winner announcement names projects involving a Point72/Cubist streaming application, an AllianceBernstein Edgar-plus-LLM challenge, Principal’s corporate-narrative task, and a Northfield market-impact project. This establishes a competition and speaker network, not sponsor adoption, production deployment, model quality, or investment results.

The AWS sample-tech-for-trading repository adds a public implementation surface adjacent to the vendor’s hedge-fund factor-modeling article. Its examples cover factor mining, factor trading, agentic backtesting, and an AI fund-manager pattern; the repository lists contributors and an MIT-0 license. It is a vendor reference repository, not evidence that a named fund uses the code or that its sample factors or agents have predictive value or production authority.

September 2, 2026 — regional prospectus, Australian practitioner, and Canadian research routes

The SGX-hosted prospectus for a Lion Global Investors fund advised by Nomura Asset Management provides a date-scoped regulatory disclosure for an actively managed Japanese-equity ETF. It says portfolio selection is based primarily on proprietary or licensed AI and machine-learning models using fundamental, technical, qualitative, quantitative, and other datasets. It describes monthly model updates using market data, company financial information, filings, announcements, and news; valuation and technical factors; a one-to-three-month middle-term horizon; risk constraints; a typical 50–100-stock model portfolio; and manager discretion between scheduled rebalances. This is a stated product policy, not a model registry: the prospectus does not name architectures, weights, training history, vendors, validation protocol, GenAI components, or AI-attributed performance.

An Australian title-blind route, i3’s “Building AI Agents in Investment Research” program, describes a Sydney seminar with Michael Kollo, Chief AI Transformation Officer at Qualitas and a former senior quantitative practitioner at BlackRock, Fidelity, and AXA Investment Managers. The page describes an equity-analyst demonstration that reviews financial statements, derives forensic metrics, flags anomalies, generates competing hypotheses, checks them against competitor data, product-market evidence, news, and filings, and applies a challenge layer. This is a public event description, not proof that Qualitas or any former employer uses the described system in production, and it does not disclose models, datasets, permissions, or results.

The RBC Borealis AI in Finance Summer School adds a Canadian research-talent route: the Vancouver program is described for PhD students at Canadian universities and covers foundation models, generative AI, reinforcement learning, causal inference, time-series foundations, model risk, responsible AI, and research-to-production. Its displayed faculty and industry network includes UBC, the University of Toronto, Waterloo, Alberta, UC San Diego, Polytechnique Montréal, Vector, and Amii. This is evidence of an academic-industry ecosystem and recruiting surface, not evidence of a shared investment system, private collaboration, or production deployment. See the regional route note.

September 2, 2026 — Australian systematic-investing transcript

An Australian fund-management transcript adds a specific, date-scoped signal. In the Future Generation Global FY2025 Q&A, Dr David Allen, Head of Long/Short Strategies at Plato Investment Management, describes automated red flags across a broad company universe and says AI represented approximately 10–15% of the process (pages 3–4). He describes using LLMs to compare earnings-call sentiment with prior quarters and peers, and to score whether management answered or evaded analyst questions (pages 4–5). The transcript is stronger than generic AI commentary because it names the workflow and its intended use, but it does not disclose the model version, labels, point-in-time construction, evaluation split, costs, permissions, or independently reproducible results. The full transcript is retained privately; see the capture note.

September 2, 2026 — CFM audio recovery and title-blind personnel routes

The Alternative Data Podcast episode with Yves Lemperiere identifies him as CFM’s Head of Research, Alpha Strategies. In the recovered 2022 audio, he describes CFM’s preference for performing NLP internally, selective use of outside providers when they add value, and explicit concern about crowding when a vendor sells the same transformed signal to many managers (approximately 18:36–21:20). He also describes satellite data for energy and grain forecasting beginning around 2010 (approximately 14:15–16:20), and a research process that connects alternative data to macro-variable forecasts, cross-country coverage, and longer historical evaluation (approximately 25:10–28:05). These are speaker-reported statements from a dated interview. They do not establish CFM’s current models, vendors, data rights, production systems, or returns. The original audio and timestamped sidecars are retained privately; the full transcript is not published. See the capture note.

Two Balyasny routes add distinct public roles without collapsing them into one AI claim. The Bloomberg Odd Lots episode page identifies Giuseppe Paleologo as Head of Quantitative Research and dates the recorded conversation to June 12, 2025. Local ASR of the recovered audio captures his description of centralized factor-model, hedging, portfolio-advisory, and risk services (approximately 05:40–07:15), productivity/document workflows as current baseline AI use and more agentic workflows as an anticipated direction (20:25–23:10), and the need for point-in-time controls when evaluating AI-generated factor ideas (28:40–29:15). The audio is retained privately; the transcript body is not published. The A-Team Insight report identifies Samantha Mait as Data Science Operations Lead and reports her discussion of point-of-use data quality, exposure-specific checks, anomaly and drift monitoring, and licensing safeguards. The public record supports personnel and operating-process signals; it does not reveal Balyasny’s model inventory, agent permissions, or production outcomes. See the capture note.

Additional title-blind routes include QRT’s Adrien Hardy profile, which describes a Quantitative Research Director and ML-focused technical problems, and Thomas Le Menestrel’s self-authored research site, which reports QRT quantitative research involving LLM-based trading signals and an LLM pipeline. The latter is not QRT-endorsed and is kept separate from employer evidence. The IAQF Boston listing names Acadian Senior Vice President and portfolio manager Bin Shi in a biography mentioning quantitative analysis, ML, and AI. A CFA Society India announcement names GMO Senior Quantitative Researcher Anshul Jain for a systematic-equity and valuation discussion. An Arrowstreet Harvard FAS listing describes a partner-led research introduction and quant-development and quant-research panel. These routes establish public roles or event participation, not current AI deployment.

The NVIDIA GTC on-demand session with Cohen & Steers adds a technical presentation route: Yigal Jhirad is identified as Portfolio Manager and Head of Quantitative and Derivatives Strategies, and the session description names sheaf neural networks, probabilistic forecasts, and quantum portfolio optimization. NVIDIA exposes an AI-powered transcript, but the presentation description does not establish live-book use or audited results. The FIX EMEA speaker archive broadens title-blind discovery across AI/GenAI, LLMs, agents, RAG, pre-trade analytics, execution analytics, data governance, and alpha capture. Its Tom Doris route links Apple Podcasts, Spotify, and YouTube coverage of machine learning and alternative-data preparation; no public captioned transcript was located. Finally, the NeurIPS 2025 sponsor roster creates a research-ecosystem route spanning Cubist/Point72, QRT, Two Sigma, WorldQuant, Virtu, Voloridge, Hudson River Trading, Jane Street, Jump Trading, and Optiver. The associated public projects cover auditable multimodal retrieval, financial-agent orchestration, financial RL environments, and LLM-assisted bond retrieval. Sponsor presence and project pages do not prove firm adoption.

September 2, 2026 — title-blind institutional media pass

Generic leadership and infrastructure searches produced several additional firm-linked routes. HBR’s January 2026 Ray Dalio interview exposes a full transcript in which Bridgewater’s founder describes computerizing decision criteria, comparing human and computer decisions, and treating AI as a decision partner rather than a substitute for causal understanding. Lehigh Business’s October 2025 Carter Lyons episode identifies Two Sigma’s co-CEO and links a complete transcript, with a public framing that keeps human experience and judgment alongside technology and data science. Neither route discloses a current model inventory, data rights, production permissions, or attributable investment outcomes.

The S&P Global transcript with Man Group CEO Robyn Grew adds a more specific organizational disclosure: Grew describes machine learning, AI, and large language models across data, data science, signals, execution, and operations, alongside a move toward more agentic “build” workflows and expert supervision. The Ritholtz transcript with Balyasny founder/CIO Dmitry Balyasny separately reports more than 500 technology staff and more than 100 data/AI staff supporting internally built tools for investing, trading, research, risk, and some operations. Both are speaker-reported public accounts; they do not provide model registries, vendor contracts, permission maps, or AI-attributed returns. See the capture note.

The Anyscale Ray Day recap identifies Todd Gaugler as a Cubist Systematic Strategies / Point72 quant developer and describes on-premise multi-tenant Ray clusters paired with Anyscale cloud infrastructure for time-series and windowed model fitting. It names data-loading, scheduling, and memory failure modes and links Point72’s raydar and csp projects. This is vendor event material, not an independent deployment or performance audit.

September 2, 2026 — title-blind roles, conferences, and vendor routes

The expanded title-blind pass adds public evidence that sits beside, rather than inside, the firms’ named AI programs. Voloridge’s people page exposes named research-technology leadership and describes deep-learning and AI methods used by the firm. Two Sigma’s investment-management page describes a data-to-execution workflow and publishes firm-reported scale figures. Neither page supplies a current model registry, agent permission map, or performance audit.

Several employer pages expose intended workflow scope without naming a product: DRW’s quantitative-AI strategist role, AI-engineer role, and commodities researcher role mention research platforms, validation and monitoring, RAG/fine-tuning, and LLM use in research and trading. Verition’s sector data-science listing mentions alternative data, predictive signals, and generative-AI research tooling. WorldQuant’s software-engineering listing mentions AI coding agents and LLMs in research workflows. Job descriptions establish hiring intent and desired scope, not filled roles or live capital authority.

Brevan Howard’s Behavox announcement adds a distinct non-investment use: AI for text, voice, and communications compliance. The ExodusPoint hiring report and a public SEC-filing-delta project provide personnel and research-route clues, but remain third-party or self-authored evidence and are not treated as proof of firm sponsorship.

The conference search found additional title-blind routes: Columbia’s MAFN archive names Tower’s Global Head of Core AI & ML in an AI-agents-in-HFT session; Cornell’s Quantmate seminar describes multi-agent investment analysis; and NVIDIA’s GTC trading archive links sessions involving HRT, E Fund, Cohen & Steers, Barclays, and Jump. These pages establish named speakers and session topics, not implementation or results. The full route inventory and evidence boundaries are in the capture note.

September 2, 2026 — regional, lab, and governance routes

The regional-language pass adds Chinese, Japanese, Korean, German, Australian, and Indian routes that do not depend on an English-language “AI hedge fund” title. Eastmoney’s Chinese roundtable names several quantitative-investment founders and research leads discussing foundation models and multimodal systems in research. Tokyo Finance Forum No. 40 identifies a Nippon Life Asset Management research leader and former AQR Japan investment-office head in a generative-AI and asset-management session. M.M. Warburg’s German podcast archive provides an older discussion of ML, Big Data, portfolio management, and causality. These are regional research and practitioner routes, not evidence of current fund deployment.

Several first-party pages make stronger but still bounded architecture claims. Hildene’s team page and CLO episode identify a quantitative-AI role and the CLOver analytics platform. BlackRock AI Labs lists a lab remit spanning statistics, ML, optimization, stochastic control, and decision theory, with named academic advisers. QRT Labs describes research partnerships with Imperial, Cambridge, and Oxford across foundation models, agents, decision-making, and HPC. Vanguard’s University of Toronto partnership names small-language-model, causal-AI, responsible-AI, and autonomous-agent research, including applications to earnings calls and public corporate communications. These pages establish public research remits and personnel networks; they do not connect specific papers to live portfolios, production permissions, or returns.

The India and UK sweep found Sixteen Alpha AI, Quanentry, Stokhos, and Evore Labs. Their sites describe AI/ML-driven quantitative research or governed autonomous research, but the claims are company-authored and lack public model, evaluation, or capital-authority evidence. PIMCO’s stewardship report adds a governance route describing customized AI tools, training, data-quality and hallucination controls, security, privacy, MNPI, recordkeeping, and human accountability. The full route inventory and uncertainty labels are in the regional/lab capture note.

The native-language follow-up adds several Indian and Chinese routes. AlphaComet describes a Category III AIF process using proprietary AI, alternative data, portfolio optimization, and an ingest/predict/optimize/execute loop. Prometéi Labs’ FireSight page claims a fully AI-managed pilot fund with no human in the loop and names sequence modelling, world modelling, reinforcement learning, and order-book signals; these remain unverified first-party claims because the page provides no legal fund identifier, model card, risk limits, or audited record. A DTL Quant recruitment posting exposes Chinese-market hiring vocabulary spanning ML/NLP research, LLM-aware data and low-latency infrastructure, GPU optimization, market-impact modelling, routing, and execution analysis. A CITIC Prudential Fund recruitment route is retained as a search-index lead describing alternative-data modelling, prediction-model iteration, live integration, attribution, and LLM-inference optimization; the page still requires a reliable direct capture. A China Securities report adds an unresolved Chinese quant-AI-lab organizational lead. These sources add discovery and technical vocabulary, not a basis for ranking firms or inferring production deployment.

September 2, 2026 — Citadel and Schonfeld media expansion

New firm-controlled and publisher-hosted routes add current leadership and workflow context. Citadel’s Stanford Leadership Forum listing and its YouTube recording describe Ken Griffin discussing agentic AI and complex research workflows; the public record does not name a model, dataset, evaluation, or permission boundary. Citadel Securities’ “A Day with Ryan” describes an equity-options quantitative-development project in which LLMs help an intern explore, test, and validate research questions while core interpretation remains human-led. This is a firm-produced account of one internship, not a firm-wide model disclosure.

Milken’s capital-markets panel identifies Citadel Securities President Jim Esposito and links a transcript PDF; the panel is about AI, data infrastructure, and capital-markets modernization rather than a specific Citadel Securities model. Milken’s Ken Griffin conversation links a separate transcript PDF and records broad AI-enabled business-process discussion. Both PDFs are retained privately and are not reproduced in this article. Alpha Exchange’s Colin Lancaster episode adds a current Schonfeld macro-leadership account with timestamped transcript access; its public discussion covers platform, systems, data, risk, and AI context without disclosing a named Schonfeld model or permission map. See the Citadel/Schonfeld capture note.

September 2, 2026 — French, Indian, and Chinese first-party routes

A French-language Pictet Asset Management fund note provides a more product-specific disclosure than a general AI commentary page. It describes Pictet TR-Quest AI as an international-equity long/short strategy, says its model uses approximately 400 market-activity, fundamental, and sentiment signals, and describes boosted decision trees trained on decades of data to forecast one-month stock performance. It also describes relative-risk penalties and factor-neutrality constraints in portfolio construction. These are Pictet’s first-party and marketing statements; the page does not publish weights, point-in-time training splits, raw data rights, independent validation, or a complete permission map. It is a comparator route, not a ranking.

La Française’s quantitative-management page states that its quantitative hub uses machine learning and generative AI in market analysis and to support group investment teams. The page displays €3.7 billion in quantitative assets under management, more than ten dedicated employees, and four managers as of June 30, 2026. It does not identify models, vendors, datasets, evaluation, or AI-attributed results. IndiQuant adds a newly verified Indian fund-formation route: its site describes crowdsourced predictions on anonymized Indian-equity data, meta-model aggregation, and live-outcome scoring, while explicitly displaying that Round 001 scoring is not yet live. ZenX Quant adds a Chinese manager identity page with a displayed Shenzhen private-securities-fund entity and registration number but no captured AI, model, or strategy detail; it is retained as a negative control rather than an AI-use claim. The regional-language follow-up note records the original-language evidence and boundaries.

The same note adds Spain and Portugal routes. Global Gradient describes a Spanish global-multi-asset fund using low-frequency macro, fundamental, and market data in machine-learning models and displays CNMV, adviser, and dated performance metadata; its references to methods used at large quant firms are marketing claims, not relationship evidence. Afi describes Spanish quantitative-model development and validation using ML and AI as a consulting service. A Portuguese Banco Best page for Acatis AI Global Equities preserves a historical partnership route involving Acatis, Nnaisense, and Quantestein, with deep learning, approximately 4,000 stocks, 50 selections, and 80 factors. FINTAI adds a Spanish practitioner/vendor surface covering factor investing, backtesting, portfolio construction, production LLM architectures, and financial-market surveillance. These pages add regional discovery and historical partnership clues; they do not establish current model maintenance, customer permissions, or independent performance.

September 3, 2026 — exact regional role and personnel URL reconciliation

The regional source note contained several public URLs that described distinct organizations but had not yet been admitted to the central ledger. The new records include Fulcrum’s quantitative-analyst role, which explicitly names LLM/NLP work, model monitoring, process automation, and Claude Code and Codex as desired tools; BNK Asset Management’s Korean Data-Scientist notice, which names ML/deep-learning algorithm R&D for quantitative portfolios; and a China Securities profile of Hefu Investment, which names its founders, describes research across news, retail sentiment, and analyst expectations, and says AI/ML factor research is screened with cross-validation and an observation period. These sources add specific role, workflow, and research-process vocabulary in the UK, Korea, and China; they do not establish filled roles, model ownership, data rights, production permissions, or performance.

The same reconciliation adds a Dubai-based Fibration Research profile describing ML dataset generation, model training and serving, out-of-sample evaluation, and Rust/C++ low-latency systems; IWM Quant’s Mumbai firm page listing an AI/ML Engineer role and an in-house data, simulation, risk, and execution stack; and Axcyon’s Australia/New Zealand trading-firm page describing machine learning and pattern recognition across futures, options, commodities, and equities. Index Solutions adds a South African founder/CIO route describing in-house ML, optimization, backtesting, and automated portfolio management. These are first-party positioning and personnel surfaces; none publishes a complete model registry, training corpus, permission map, or independently audited result.

The personnel reconciliation also retains Felix Leibfried’s current iSAM profile, a self-authored account of LLM-based text research at iSAM and point-in-time ML training and evaluation at former employer Eisler, plus Major Capital’s official page, which names its CTO and describes algorithmic AI/ML fund selection and monthly rebalancing. The iSAM and Eisler statements remain personal claims, while Major Capital’s statements remain firm-authored descriptions. The regional role note and personnel note preserve the remaining routes and source boundaries.

Additional exact-URL records preserve separate hiring and product surfaces: Voloridge’s 2027 quantitative-research fellowship describes PhD-level work on large datasets and ML with data scientists; Winton’s MENA equities role describes a London-to-Abu Dhabi research, backtest, live-trading, and risk workflow; Citadel’s Asia ML-research role names deep learning, NLP, unstructured data, backtesting, and Python/C++; and Inquant’s Mumbai careers page names ML and time-series experience for quantitative research. ACATIS’s Spanish-language AI-funds page dates direct AI-based portfolio selection to 2014 and lists three AI strategies. These are current or dated public role/product descriptions, not proof of filled hiring, model ownership, production permissions, or results.

September 2, 2026 — vendor architecture and data-intelligence routes

A vendor/platform pass found several sources that explain how investment research is being packaged without naming a particular tracked manager. AWS’s GenAI factor-modeling workflow and linked sample repository expose a reference architecture for collecting market data, SEC filings, financial reports, and web-search results, then running factor mining through S3, ClickHouse, Lambda, Step Functions, and AWS Batch. This is provider-authored sample code, not evidence of GMO, Acadian, Arrowstreet, or another manager using it.

AWS’s Boosted.ai case study supplies a more specific model-economics disclosure. It says Boosted Insights processes millions of documents from 150,000 sources—including filings, earnings calls, trade publications, international and local news, and nontraditional sources—and reports a move from a costly general model to a smaller domain-specific fine-tuned model. AWS describes Invisible as the annotation partner, reports a 90% cost reduction, and says private-VPC deployment enabled five-to-ten-minute update targets. These are vendor/customer claims: the page does not identify the 180 asset-manager customers, the tuned model, the training corpus, data licenses, or an independent investment-performance study.

Anthropic’s current Aura Intelligence case study adds a workforce-intelligence route relevant to title-blind personnel research. The page says Aura uses Claude in Amazon Bedrock to classify 200 million titles and industry pairings, analyze hiring and exit patterns, detect workforce anomalies, and support sentiment and multilingual workflows. It also says Aura is launching APIs with major high-frequency-trading hedge funds. The named customers, contracts, permissions, reproducible test set, and any portfolio use are not disclosed; the reported accuracy and customer-scale figures remain vendor-reported.

LangChain’s Captide customer architecture describes a disclosure library covering more than 14,000 public companies, domain-specific research and equity-modeling agents, stateful concurrent execution through LangGraph, and trace/evaluation workflows through LangSmith. Its separate agentic-AI ROI post describes token-level cost, latency, error, KPI, and budget monitoring. These pages add concrete search terms—state persistence, parallel invocation, regression evaluation, model routing, cost governance, and source-linked outputs—but do not establish a named fund’s adoption, model inventory, trading authority, or returns. The vendor-platform capture note records the evidence tiers and disqualification boundaries.

September 2, 2026 — Momentum show expansion

A show-level RSS and Apple expansion found four previously untracked episodes whose titles do not identify a particular hedge fund. The Matterfact episode identifies CEO Ashutosh Agarwal, describes prior Millennium quant experience, and discusses prompt scaffolding, sector-specific research playbooks, and AI-generated research artifacts. The Big Data Federation episode identifies founder and CEO Pouya Taaghol and discusses satellite imagery, geolocation, transaction and receipt data, sales forecasting, and alternative-data decay. The Equity Data Science episode identifies Sandeep Varma and Benjamin Lieblich and covers disaggregated research workflows, factor models, portfolio construction, attribution, and risk management. The CarbonArc episode identifies CEO and founder Kirk McKeown and discusses alternative-data licensing, consumption-based API access, MCP tools, and agent-oriented data structures.

These are speaker- and vendor-account signals, not proof of any named hedge fund’s current deployment. The recovered audio and timestamped ASR sidecars for the first three episodes are retained privately in checkpoint d7af4b68, and the CarbonArc episode is retained in checkpoint bfa1a887; full transcript bodies are not published. The capture note records episode dates, capture hashes, segment counts, and the boundaries on claims. The routes are added to the coverage ledger as separate records so show-level discovery is not confused with firm-level evidence.

September 2, 2026 — generic finance media and quant-research lineage

A title-blind search of generic finance media added ClearAlpha founder Brian Hurst’s Ritholtz interview. The publisher transcript describes his earlier Goldman Sachs Quantitative Research and AQR context, automated public-data and multi-signal workflows, model-compute constraints, organizational standardization, and the boundary between public research and unpublished proprietary detail. These are attributed practitioner statements, not a disclosed ClearAlpha model inventory or performance result. Corey Hoffstein’s Investipal episode and the CFA Institute event with Marcos López de Prado add quant-research and causal-inference routes; their embedded YouTube recordings were not recoverable in this pass, so only publisher metadata and listed topics are used. The capture note records the access failures and evidence boundaries.

September 2, 2026 — Bridgewater AIA personnel and CFM ML Lab refresh

The Bridgewater/CFM route note adds a current Bridgewater personnel surface and a separate CFM publication route. Bridgewater’s Aspen profile identifies Nina Lozinski as Co-Head of AI & ML Investment Strategy and reports a firm-stated increase from six AIA Labs members in 2023 to more than 50 scientists, engineers, and investors. A current Greg Jensen profile connects his Managing CIO remit to AIA Labs; a separate Bridgewater article identifies Suri Bandler as an Architect on the Technology Team. CFM’s October 2025 publication describes an ML Lab embedded in research teams and linked to academic work.

The headcount and titles are date-scoped, firm-reported evidence. The CFM description is text in a CFM-hosted publication and is not an independent audit. None of these routes establishes a complete roster, a model inventory, data rights, production permissions, portfolio authority, or performance.

September 2 fresh title-blind firm and regulatory routes

The fresh underexplored-firm pass adds four public surfaces that were not canonicalized in the central ledger. AlphaSimplex’s July 2025 Form ADV Part 2A describes the adviser’s process as model-driven and highly automated, using proprietary quantitative models approved by its Investment Committee and fundamental, technical, and macroeconomic data. The filing also lists managed futures, global alternatives, replication, model-portfolio, and private-fund services. This is useful process and service-scope evidence, but “automated” is not converted into “AI” or “GenAI,” and the filing does not disclose model weights, training data, vendors, or performance attribution.

Rokos’s company recruiting post states that its 2026 graduate programme recruited across Technology, Quant, and Analyst streams. PanAgora’s official careers archive exposes an Equity Data Science role working with Alpha Research, Portfolio Construction, and software teams; the page also identifies Boston location and ownership context. Capula’s public jobs surface displayed Quantitative Strategist (PhD), Technology Graduate Analyst, Networking & Security Engineer, and Portfolio Manager postings, including London and New York locations. These pages establish dated recruiting and organizational signals only. They do not establish filled roles, an AI lab, model ownership, data rights, deployment, authority, or investment outcomes.

September 2, 2026 — regional and title-blind process evidence

The latest regional pass adds several public signals that should remain separate from named AI-lab evidence. Dymon Asia’s careers page displays an AI Intern (Fundamental Equities) description embedded in investment teams. It names evaluation of GPT-4, Claude, Gemini, and open-source tools; Python, SQL, prompt engineering, LangChain, and Hugging Face; prototype work, governance, risk/benefit analysis, and deployment techniques. The page also says no positions were open at capture. This supports a dated hiring and workflow vocabulary signal, not a filled role, a production model, or an investment-authority claim. Dymon’s public LinkedIn jobs surface separately displayed AI Engineer, Data Engineer, and AI Management Fellowship titles in Hong Kong and Singapore.

Arrowpoint’s ML-engineer description specifies predictive ML/deep learning, alternative-data ETL, AWS deployment, monitoring, and an end-to-end lifecycle for systematic strategies. Brevan Howard’s front-office data-engineer role describes data infrastructure supporting research, trading, and portfolio decision-making. Hadron Capital publicly describes a gradient-boosted ML alpha model, 60-plus cross-sectional features, purged walk-forward validation, multiple optimizers, and executable trade generation. The first two are employer hiring intent; the third is self-description. None supplies independent verification of a live model, data rights, or performance.

RQI Investors’ May 26, 2026 transcript is unusually specific about process. Andrew Francis, Dr Joanna Nash, and Dr David Walsh discuss extensive ML use in research, nonlinear patterns, alternative and unstructured data, human-led idea generation, portfolio and risk construction, and a possible future role for agentic systems. This is a dated first-party account, not a model registry or reproducible performance record. Robeco’s quantitative-investing page adds named quantitative research leaders and a broad first-party description of alternative data, ML, NLP, and AI supporting return, risk, sustainability, and portfolio implementation; it is included as an asset-manager comparator, not as evidence about a hedge fund.

PanAgora’s research archive, chatbot/quantitative-investing article, and Crowell Prize announcement add Boston-linked research routes involving chatbot adoption, conference-call text, and deep learning on executive presentations. They establish published research and judging activity, not current PanAgora production deployment. The full cross-regional route list and evidence boundaries are in the source note.

September 2, 2026 — frontier lab and role vocabulary

The next title-blind pass adds G-Research’s official careers page, which separates Quantitative Research & Machine Learning, Engineering, Technology Innovation Group, and Open-Source Software teams. Grace Investment Machine explicitly describes itself as a technology company and research lab, with public AI Research Engineer and AI Systems Engineer roles covering agentic methods, signal discovery, model evaluation, historical data, backtesting, execution, risk controls, automated training, deployment, and inference. Both pages are organization and hiring signals. They do not establish a filled roster, model inventory, data rights,

An Arabic-language PRNewswire distribution of the company announcement dated July 9, 2026 reports that Grace Investment Machine raised a $20 million Series A and names Jiahao Xu as founder and CEO. It describes the company’s CogAlpha research as receiving an ACL 2026 oral recommendation, a seven-layer agent architecture, purpose-built financial-market foundation models, and multi-agent systems that generate, test, validate, and iteratively adapt signals. It also says the company is testing AI-supported investment products in real financial environments. These are company-announcement claims in an additional regional-language distribution channel; the release does not independently verify the financing, establish a live external-capital vehicle, disclose model weights or data rights, or show performance.

WorldQuant’s official Book Portfolio Manager listing adds a Singapore/Sydney route that mentions systematic strategies, a proprietary research platform, broad datasets, internal research conferences, cross-asset execution, and AI/ML opportunities. Amundi’s Fixed Income Investment Lab listing is an asset-manager comparator whose role description names rates and credit signals, alternative and unstructured data, tree ensembles, LSTM/Transformers, embeddings and fine-tuning, bias-aware backtesting, drift monitoring, rollback, and model-acceptance criteria. It is detailed hiring intent, not proof of a filled role or live model.

One Eleven Capital describes a Paris quantitative hedge-fund manager using ML in equities and listed derivatives, while SINTRO exposes systematic-trading roles in Berlin, Frankfurt, and Zurich. These first-party pages extend the European fund surface but do not independently validate model weights, training data, or performance. The complete route list and boundaries are in the frontier capture note.

September 2, 2026 — AXQ agent-harness and production vocabulary refresh

The new AXQ role-surface capture note adds three separate employer-controlled postings to the existing AXQ machine- learning-infrastructure record. An Agent Harness Research Engineer role describes prompt, tool-calling, memory, workflow orchestration, and evaluation modules for quantitative research; reusable workflows for data analysis, feature generation, signal validation, backtests, and reports; and tracking of agent-driven backtests, live performance, failure cases, and market adaptation. A Lead Financial Data Engineer role describes multi-asset data lineage, temporal semantics, alternative data, automated quality checks, and research/backtest/live-trading data support. A Production Engineer role describes model deployment support, execution and risk monitoring, incident response, runbooks, observability, and operational automation.

Together, these pages expose a public hiring vocabulary spanning research agents, data correctness, and model operations. They do not establish that the roles are filled, that the agent harness is live, that any named model exists, or that an agent can alter a portfolio or place an order. The pages also do not disclose training data, model weights, vendors, evaluation results, or AI-attributed performance.

September 2, 2026 — agentic-trading evaluation controls

The agentic-trading evaluation capture note adds a university benchmark and four academic implementation/evaluation routes. HKU Business School’s Agentic Trader report describes a common-environment live-market evaluation with ten language-model agents and explicitly cautions that static reasoning or coding benchmarks do not necessarily predict market behavior. The Agentic Trading survey codes study protocols and highlights missing time-consistent splits, cost models, and reproducibility in much of the reviewed literature. The newer quantitative-trading workflow survey organizes research across factor mining, signal discovery, portfolio construction, execution, and risk. TradeLens connects trading records, runtime traces, and deployment configuration for cost and decision-value diagnosis. AlphaCrafter and its public repository provide a Miner/Screener/Trader harness design with policy constraints and verification.

These are control and vocabulary sources, not evidence that GMO, Acadian, Arrowstreet, or another named manager uses any of them. Their results and protocol claims require reproduction; no model ranking, firm ranking, or deployment inference is made.

September 3, 2026 — multilingual agentic-quant and product disclosures

The latest Chinese-language pass adds a practitioner account of AI agents moving from isolated assistance toward coordinated roles in factor mining, report analysis, data cleaning, strategy modelling, execution support, risk support, and research productivity. The China Financial Information Network / China Securities report also records warnings about overfitting, data-cleaning bias, and errors being amplified across collaborating agents. It attributes an AI-augmentation programme to 蒙玺投资 and describes GokuTech / 念空科技’s account of AI-assisted factor mining and routine software work. These are regional media and quoted practitioner statements; they do not establish a common production stack, model inventory, data rights, or return attribution.

QuantaAlpha’s public research and product pages add a separate research-lab route. Its public description covers hypothesis generation, factor construction, executable-code generation, backtesting, trajectory-level mutation and crossover, constraint gates, and factor-pool maintenance. The associated Chinese-language Orient Securities report links the work to researchers across Chinese universities, Stanford, and the QuantaAlpha team. The RePEc record for a Chinese-futures LLM-factor paper adds a separate publication route. These records document public research and architecture claims, not hedge-fund deployment. Reported backtest results remain unreplicated here.

The German-language sweep adds a dated Quoniam product disclosure. Quoniam’s 20 August 2026 release says its Global Data Sentiment public fund exceeded €100 million and describes daily AI-assisted news analysis for changes in market expectations across equities, bonds, and currencies. The release names Markus Ebner as Head of Multi-Asset. It does not identify the model family, training corpus, data contracts, human approval gates, or AI-attributed results.

Quantmade’s fund page provides a distinct operating-entity disclosure: it says Axia Asset Management operates the Quantmade AI Quant Fund while Quantmade supplies the strategy and AI-based quantitative signals. Quantmade identifies Dr Michael Geke as CEO on its company page. This separates the regulated asset manager from the signal provider in the public record, but the pages do not disclose signal models, training data, validation, order authority, or

independent performance. See the multilingual capture note.

September 3, 2026 — Korean hiring surfaces expose the AI-to-trading bridge

Two Korean-language recruitment records add role vocabulary that is easy to miss in an English-only search. A Mirae Asset Management listing names an AI financial-engineering asset-management division and combines fund operations, quantitative and AI strategy research, and AI-based workflow efficiency and database construction. A Hana Bank AI Quant listing places the role in Treasury Markets and asks for LLM- and agentic-AI research workflow automation, large time-series and text-data processing, and—among preferred qualifications—experience connecting an AI investment model to a real trading system.

These are recruiting specifications, not filled-personnel records. They do not establish the model family, dataset, approval boundary, production status, or investment result. Mirae Asset and Hana Bank are retained as financial- institution comparators, not as hedge-fund evidence. See the Korean hiring capture note.

September 3, 2026 — localized first-party product pages

Acadian España’s Spanish-language site exposes a more detailed public operating description than a generic corporate overview: data and signals, quantitative alpha models, constrained portfolio construction, and real-time monitoring of factor exposures, liquidity, concentration, and tail risk. It says the process integrates price, fundamental, alternative, macroeconomic, and sentiment data across global markets. The displayed €108 billion AUM, strategy figures, and performance figures are Acadian claims; the page does not name model families, data contracts, feature definitions, or approval gates.

A French Pictet product page describes an AI-enhanced active-ETF range. Pictet says its proprietary AI engine proposes stock selections that investment experts validate, using an integrated machine-learning model trained on hundreds of characteristics and boosted decision trees. The page also describes a factor-neutral mandate and displays an annual after-fee alpha objective of 1–1.5%. That objective and the methodology are product-page statements; no weights, feature list, validation split, or independent forecast audit is provided.

These localized pages add operating and product vocabulary, not a ranking. They do not establish that all firm strategies use the same system or disclose live permissions, model ownership, training-data rights, or AI-attributed results. See the localized product-page capture note.

September 2, 2026 — Italian and Dutch routes add distinct disclosure layers

The Italian-language Pictet page adds a separate distribution and product route. It describes four AI Enhanced equity ETFs listed on Borsa Italiana and says the proprietary model analyzes more than 400 variables per company, including fundamental, market, and sentiment inputs. It describes removal of common style, sector, and geographic factors, small systematic overweights and underweights, beta of 1, a tracking-error limit around 2%, and an annual net- alpha objective of 1–1.5%. Those are first-party product statements and targets. The page does not expose the model weights, exact feature definitions, training or validation periods, data contracts, or an independent live forecast audit. See Pictet’s Italian product disclosure.

The Dutch-language route produced both sector-level evidence and a separate manager lead. The AFM’s April 2026 sector note describes AI use in analysis, price prediction, and trading-strategy improvement, while reporting that policy and ethics controls do not uniformly match usage. It specifically flags data quality, bias, explainability, and concentration among technology providers. This is regulator context, not evidence about a particular tracked manager.

Felidae Investment Partners is a separate Dutch FX fund lead. Its site describes selecting algorithmic strategies from the Profectus AI platform, ten-to-twenty-year historical testing, limited-exposure live testing, and continuous portfolio monitoring and replacement. A Spanish central-bank entity list also lists Felidae Fonds and Felidae B.V. The fund’s algorithmic and AI workflow claims remain self-described; the public record does not independently verify its strategy population, performance, data rights, or production controls. A secondary fund-directory record also distinguishes registration from supervision, so the two should not be collapsed.

The foreign-language capture note records these routes alongside the earlier Japanese, Chinese, Korean, French, German, Spanish, Portuguese, Arabic, and Hindi findings. The overall pattern is descriptive rather than comparative: regional pages reveal vocabulary and operating claims that English-only searches miss, while the missing fields are still usually model identity, training data, permissions, validation, and independently attributable results.

September 2, 2026 — Nordic-language searches add two fund-specific routes

The Swedish search added ORCA Hedge, a market-neutral long/short equity fund described by Markov Capital as using alternative data, a broad factor universe, machine learning, and around 500 positions. Finserve’s fund page describes the same fund as systematic and quantitative, with machine learning used in portfolio construction and continuous optimization. The Swedish Financial Supervisory Authority register lists ORCA Hedge as an active special fund managed by Finserve Nordic AB and records fund-rule approval under LAIF in January 2025. Markov’s team page names Mikael Andersson as CEO and co-founder, Philip Bogdanffy as Quantitative Engineer and co-founder, and Anton Reimbert and Filip Heikkilä as Quantitative Engineers. The public material does not assign model ownership to individuals or disclose model families, features, training windows, or data vendors.

The Finnish route added a temporal operating trail for Mandatum Managed Futures. Mandatum describes a systematic multi-asset managed-futures fund using mainly momentum-oriented long/short strategies. Its Finnish manager interview names Ville Rantanen as portfolio manager and describes daily use of historical prices, country-level data, and macroeconomic variables for trend and recurring-pattern detection. Mandatum’s 2021 disclosure describes AI and probability-based machine learning forecasts across equity indices, bonds, and currencies, with the portfolio manager executing trades and monitoring risk within predefined parameters. The pages do not disclose model architecture, feature engineering, forecast horizon, retraining, or independent AI-attributed performance.

These Nordic results add two different evidence structures: a Swedish technology-provider/fund-manager split with regulator confirmation, and a Finnish first-party history connecting model-generated forecasts to a named manager and explicit human execution controls. See the Nordic-language capture note.

September 3, 2026 — Norwegian and Swedish searches add research-funding and data-provider routes

The Norwegian-language search found a material status distinction around NorQuant. NorQuant’s public research page says it received NOK 14 million from the Norwegian Research Council for a four-year machine-learning and sustainable-investing project with Nord University and the Norwegian Computing Center. The stated research scope includes ML portfolio construction, climate and ESG risk prediction, greenwashing detection, and transparent ESG indices; the page lists six publication routes. NorQuant’s current company page also says its Multi-Asset fund was transferred to Centaur Fondförvaltning AB on April 11, 2025 and that NorQuant returned its investment-firm authorization on May 27, 2025. The public evidence therefore supports a research and technology lineage, not a claim that NorQuant is currently a licensed fund manager. See the NorQuant capture note.

A separate Swedish route adds Alfakraft’s Allocator Global Macro, where the fund page distinguishes Alfakraft Fonder’s management, execution, compliance, and risk responsibilities from Luccai’s macro-data analysis, models, and predictive signals. The page names John Ricciardi, Ravi Kishore Booka, Pramila Prasingu, and Sarath Kotamarthi and displays 27 countries, 28 economic factors, 50M+ datapoints, 800K series, and four-times-per-hour updates. Luccai’s Global Macro Matrix describes daily forecasts for 28 variables across 28 economies, two- and six-month horizons, and API/cloud delivery for allocation, hedging, sector rotation, and stress testing. These are provider and fund-page claims; the pages do not disclose model families, features, training windows, data contracts, validation design, or customer-specific deployment.

The Norwegian Association for Quantitative Finance also provides a new conference-discovery route, with archives covering alternative data, machine learning in finance, and portfolio management and a 2020 webinar featuring former AQR ML head Marcos Lopez de Prado alongside NorQuant. These event and speaker records expand the search graph but do not establish shared employment or fund deployment. The Aalto record for Sina Seyfi’s 2026 dissertation adds methodology context on high-dimensional characteristic space, basis-portfolios, and nearest-neighbor predictability, not manager evidence.

September 2, 2026 — Turkish hiring language exposes the research-to-execution boundary

A Turkish-language Tera Yatırım recruitment posting describes a Quantitative Strategies & Algorithmic Trading team being formed in Istanbul. The advertised scope runs from statistical and time-series modelling, alpha research, and backtesting through order-book analysis, execution quality, event-driven and low-latency systems, real-time risk and PnL monitoring, and FIX, OUCH, and API integration. It names equities, VİOP, options, FX, alternative data, macroeconomic data, BISTECH, VİOP, and SPK controls, and lists reinforcement learning, deep learning, and AI-driven trading as preferred experience. This is a useful Turkish vocabulary and intended operating-scope route, but it is a hiring specification: it does not identify the incumbent, model family, data vendors, production status, or live results.

Tera Yatırım’s 2025 first-party report separately describes TRA Bilişim as developing end-to-end software and algorithmic-trading infrastructure for financial markets. The same report mentions an investment in KavAI and characterizes it as AI-based “active physical intelligence.” The report does not connect that investment to the advertised quant team or establish that either description corresponds to a model used in investment decisions. See the Turkish capture note.

September 2, 2026 — Edge Focus adds a current private-credit ML route to the older partnership record

Edge Focus was previously represented here mainly through the historical 2021 Brevan Howard/DRW partnership disclosure. Its current first-party site now provides a separate, current route: it describes an ML underwriting suite paired with institutional capital, portfolio and underwriting software, and consumer-credit access through joint ventures, funds, and securitizations. The site displays company-reported figures of more than $3.5 billion in capital deployed, more than 150 billion trained data points, and more than ten lending partners. Those figures are not independently audited, and the page does not define the data-point measure or expose model details.

The same first-party page names Elliott Lorenz (CEO), Frank Jones (President and Chief Architect), Kevin Hennessy (CIO), Jeff Andrews (CRO), and Sean Mills (CTO), with biographies assigning ML credit-engine, investment-research, platform-architecture, partnership, and technology responsibilities. A current India Quantitative Researcher posting adds hiring intent around predictive AI/ML models, strategy development, investment-committee work, production Python, and loan-originator/capital partnerships. Neither page establishes a filled role, a model inventory, data rights, validation design, or AI-attributed returns.

A timestamped August 12, 2026 Bloomberg discussion with Lorenz and Atlas Merchant Capital’s Bob Diamond adds an executive media route: around 03:57–04:51, Lorenz describes broad consumer, platform, credit-bureau, and alternative data, plus PhD researchers focused on consumer modelling and current trends. Edge Focus’s current press index and partner releases separately connect the platform to Fortress, SoFi, Prosper, Happy Money, and Nelnet. These sources document a technology-enabled private-credit operating model and partner graph; they do not show that one model runs across all partners or that the system generates market alpha. See the Edge Focus capture note.

September 2, 2026 — current low-coverage recheck: Schonfeld and Marshall Wace

Schonfeld’s current FE AI Lab announcement describes a firm program for fundamental-equity portfolio managers and analysts. It names earnings preparation, idea generation, document analysis, inbox triage, and spreadsheet workflows as automation targets, and says the program is backed by SchonAI, a proprietary platform with model partnerships including Anthropic and OpenAI. The page also says new tools pass through a structured pilot and downside-evaluation process before rollout. This is a current first-party operating statement. It does not disclose model versions, training data, evaluation denominators, user counts, agent permissions, or portfolio authority. The current careers page adds AI Training & Education, AI Data Engineer, AI Strategy Analyst, and AI Technology roles across New York, London, Hong Kong, and Singapore; those are hiring signals, not filled-personnel evidence.

Marshall Wace’s current Quant Research Association Programme posting describes machine learning, large-scale data analysis, novel-dataset research, predictive-signal design, validation, realistic backtesting, and collaboration with portfolio managers to bring selected ideas into production. Its separate AI Placement posting says AI-focused teams build services, models, and data assets that bring generative AI into investment and operations workflows. These descriptions expand the public AI vocabulary around Marshall Wace, but they remain recruiting intent: they do not identify filled roles, model providers, live systems, permissions, or AI-attributed performance. See the Marshall Wace capture note.

September 3, 2026 — Brazilian survey and title-blind allocation podcast

A Portuguese-language InfoMoney report summarizes XP’s first survey of Brazilian asset managers. It reports 71 respondents representing more than R$5 trillion in assets, with generative AI used chiefly for coding, data analysis, and report preparation, and non-generative AI used for monitoring, alerts, predictive models, and portfolio selection. It also reports a mix of ChatGPT, Gemini, Claude, and internally developed systems. The XP research page confirms the survey date and sample scope. These are sector-level, publisher-reported results: the public material does not identify respondents, publish raw responses, separate hedge funds from other managers, or establish investment-decision authority.

The title-blind sweep also recovered Santander Asset Management Brazil’s Minutos a Fundo archive and episode 23, “Alocação de ativos utilizando modelos quantitativos e inteligência artificial”, published November 5, 2024. YouTube identifies Clayton Calixto as host and Luiz Felix as Santander Asset Management’s Global Head of Asset Allocation. The episode establishes a named institutional discussion of quantitative allocation and AI; its public metadata does not disclose model family, inputs, validation, permissions, live use, or AI-attributed performance. See the Portuguese capture note.

The Korean-language Marketin / Edaily report dated November 26, 2024 adds a separate allocator route for Korea Investment Corporation (KIC). It identifies a dedicated quantitative-equity office, describes a mathematical/statistical algorithmic excess-return model that developed into an enhanced strategy, and says KIC actively uses big data and machine learning. The report places this function within KIC’s wider direct, delegated, fundamental, quantitative, strategic, and tactical allocation structure. This is dated secondary reporting about a sovereign allocator—not evidence of a hedge-fund deployment, current model inventory, personnel roster, data rights, or AI-attributed performance. See the Korean capture note.

The Russian-language Alfa Capital Quant product page adds a current mutual-fund comparator. Alfa Capital describes an exchange-traded fund whose Russian-equity selection and position-closing signals use machine-learning and AI technology with professional-manager participation; the stated signal horizon is several weeks or days. The page names Nikita Elenberger as manager. This is a first-party product description, not hedge-fund evidence. Its model family, features, retraining, permissions, validation, and AI-attributed performance remain undisclosed. See the Russian capture note.

An Arabic/MENA media pass also found TIC Talks, a publisher-described trading podcast whose company page claims 72 episodes; Apple Podcasts displayed 70 at retrieval. It is distributed through Spotify, Apple Podcasts, Anghami, and YouTube. The page identifies Ahmed Tahsin as founder of Tahsin Investments Co., displays a personal UAE Securities and Commodities Authority financial-influencer licence, and uses systematic and AI-assisted trading language. This establishes a regional discovery surface, not a hedge-fund vehicle or proprietary-model disclosure. Episode-level claims, client mandates, model identity, data rights, trading authority, and audited

results remain unresolved. See the Arabic capture note.

The Russian-language Alfa Lodes episode, dated June 6, 2024, adds a person-first route for Yakov Shlyapochnik, whom the episode describes as founder of Algo Capital and a quantitative-methods and algorithmic-strategy practitioner. A public biography adds MIPT and INSEAD education and attributes early Russian ML/AI strategy-use claims to him. These are historical publisher- and self-authored statements; they do not establish a current hedge-fund vehicle, current model, production authority, or independently verified result. See the Russian podcast capture note.

The Spanish-language FundsPeople Spain roundtable page, dated April 30, 2024, adds a gated event/personnel route. Its public metadata names Ana Concejero and Daniel Ung of State Street Global Advisors, Alberto García of ACCI, Gisela Medina of BBVA, and Marcos Aza of Santander Asset Management in a quantitative-investment discussion about AI. The article is registration-gated, so this is event and speaker metadata only—not evidence of any participant’s model, deployment, authority, or performance. See the Spanish capture note.

An Indonesian-language search adds a product-level route for Sinarmas Asset Management. A July 2024 Sinar Mas release names Genta Wira Anjalu as Chief Investment Officer and says the firm collaborated with an artificial-intelligence company from Canada in connection with the Simas Danamas Saham mutual fund. It calls the approach “Simas Quantamental,” combining human and machine input, and discusses a possible intellectual-property filing for the investment techniques. A public CIO profile later describes processing more than 1,000 variables, exposing drivers behind stock rankings, and retaining human oversight over portfolio construction and risk management; a portfolio-manager profile provides a separate personnel route. These are company and self-authored claims: the Canadian partner is not named, and the public record does not disclose model architecture, training data, validation design, permissions, or independent performance. The release’s return, backtest, and prediction-count figures remain attributed rather than verified. This is mutual-fund and asset-manager evidence, not a hedge-fund classification. See the Indonesian capture note.

The chronology now includes a June 2023 Indonesian report that calls the product-level approach “Boosted AI,” and a 2025 seminar record naming Anjalu for an AI presentation covering real-time analysis, sentiment tracking, and algorithmic trading. These additional surfaces support a continuing public AI narrative, but do not prove that one model or partner persisted across the dates.

The Spanish-language pass also found Noctua, a technology-and-finance bulletin described by Apple and its RSS feed as directed by Andromeda Capital EAF. Its archive and feed expose AI-titled episodes about OpenAI, Nvidia, DeepSeek, CrowdStrike, IBM, data centers, and related technology themes, and identify Flavio Muñoz, Juan de Dios Gómez Gómez-Villalva, and Silvia Lanzarote Vargas in presenter or collaborator metadata. A separate Píldoras del Conocimiento interview dates an 88-minute AI discussion with Muñoz and Gómez Gómez-Villalva of Andromeda Value Capital and provides chapter timestamps. These routes connect named personnel and a fund-owned media archive to AI-related commentary; they do not show that Andromeda uses AI or GenAI in its portfolio process, nor do they disclose a model, data rights, permissions, or AI-attributed result. See the Andromeda capture note.

The recovered Noctua E05 audio adds a process-adjacent clue but not a system disclosure. Around 02:05–02:09, the speakers discuss continuing the series under a possible Andromeda-branded name. Around 14:16–14:54, an inserted sponsor message uses “data analysis” and “intelligent investment” language and describes a U.S.-technology focus; the automatic transcript renders the legal entity inconsistently, so this is not used to resolve the firm or infer an AI portfolio engine. The timestamped Spanish ASR is retained privately and is used for navigation, not as a verbatim public transcript.

A French-language Tendances de Fonds episode dated March 17, 2025 names ODDO BHF portfolio manager Brice Prunas and says he co-manages ODDO BHF Artificial Intelligence with Maxence Radjabi. ODDO BHF’s first-party fund page describes the product as a global thematic equity fund using AI on top of fundamental analysis to identify AI-related sub-themes and companies. The public wording is important but bounded: it documents a thematic product and named managers, not an internal GenAI laboratory, a disclosed model stack, or an autonomous research or trading agent. The public pages do not disclose weights, features, training data, validation, permissions, or AI-attributed performance. See the French capture note.

The recovered French recording adds public industry context rather than a technical ODDO disclosure. At approximately 01:11–02:09, Prunas discusses the rapid improvement of reasoning models and productivity implications. Around 04:26–05:08, he discusses distillation, mixture-of-experts, and reinforcement learning in the context of DeepSeek and compute constraints; around 07:54–08:11, he discusses post-training and inference compute. These are speaker-reported views about the AI theme and model ecosystem, not evidence of ODDO’s private model, training data, deployment, or portfolio authority. The timestamped French ASR is retained privately.

A Russian title-blind search added КВАНТОВЫЙ РЫНОК 2026, a May 18, 2026 deep-tech investment episode from “Квантовый шум | СПАРК.” Its publisher description names Alexander Borisoglebsky and mentions AI agents on a platform alongside quantum-market commercialization, IPOs, and private capital. No hedge-fund affiliation, vehicle, agent architecture, or financial result is established. This is retained as an adjacent discovery route because the AI-agent reference appears under a quantum-investment title rather than a hedge-fund or AI title. See the Russian capture note.

The recovered Russian recording adds a concrete agentic-workflow lead. Around 22:30–24:20, Borisoglebsky describes a WQ/DoubleQ ecosystem in which an automated flow of verified-source signals is filtered, validated, classified, registered, and connected by chained AI-agent layers. Around 24:31–25:08, he describes a venture filter for product value, investment potential, analytical leverage, and software products that do not require heavy initial hardware. Around 28:20–29:49, he describes daily human review of triggered signals, a planned quality-monitoring loop, external data plus a separate data layer, and a machine-learning engineer on the team. This is a speaker-described research-and-venture pipeline, not evidence of a hedge fund, public-market authority, model weights, training permissions, or performance.

September 3, 2026 — French LFIS internal GenAI and quant-technology route

The French-language search surfaced a current LFIS recruitment specification for an internal AI project that reads structured-product term sheets in PDF form with language models. The advertised work covers LLM document pipelines, prompt design across product types and issuers, automatic comparison with reference data, extraction-quality and model monitoring, architecture, agents, and cloud infrastructure. This is project intent expressed in a vacancy, not evidence that the role was filled or that the system reached production.

LFIS’s first-party biography of François-Xavier Sapa identifies him as Chief Quant & Technology Officer from 2024 and describes his work on quantitative tools, medium-frequency algorithmic trading, strategy research and monitoring, and technical leadership of the Wyse platform. LFIS’s company page separately describes Wyse Technology’s proprietary support for pricing, risk management, backtesting, order management and execution, operations control, and trade lifecycle management. The reviewed sources do not connect the term-sheet project to the execution layer, disclose a model/provider, or establish portfolio-decision authority.

LFIS also documents a partnership with the Quantitative Management Initiative covering AI and signal generation, portfolio construction and risk management, and implementation challenges. The same page records the 2018 Quant Vision Summit, organized with QMI. These are research and organizational surfaces; they do not provide model weights, training data, data rights, or independently reproducible investment results. See the capture note.

September 3, 2026 — Chilean title-blind AI-investment software route

A Spanish-language search for Latin-American systematic managers and AI investment software adds Quantsoft as an AI-investment-software and fund-vehicle lead. Quantsoft’s page describes Turing as internally developed software that receives data, analyzes it, decides, and executes orders for a globally invested fund; it calls the system fully automated while also describing continuous human monitoring and risk management. Its separate EyeQuant product page describes daily Uptrend signals across equities, ETFs, cryptocurrencies, and currency pairs, generated by quantitative algorithms processing tens of thousands of data points per day. The EyeQuant site also uses out-of-sample-backtesting language. A public Chilean corporate record records the 2022 incorporation of Quantum Software Systems SpA, but the reviewed public sources do not establish an authorized fund, regulated-manager status, named vehicle, model family, training data, data rights, execution venue, or independently audited performance. Turing’s automation language and EyeQuant’s out-of-sample wording remain company claims. This is not verified hedge-fund evidence. See the capture note.

The Indonesian-language search also enriches the existing Pinnacle Investment personnel route. Pinnacle’s official January 2026 English factsheet states that Pinnacle Strategic Equity Fund uses quantitative investment strategies, data-driven insights, and an in-house AI model. It describes a “Human x Machine” process in which investment insights are validated through backtesting and risk-management frameworks. The Indonesian product and strategy pages describe quantitative and fundamental research, computer models, portfolio construction, and optimization. The factsheet also displays PT Pinnacle Persada Investama’s OJK authorization. These are first-party product claims; they do not disclose model family, features, training data, data rights, validation splits, live permissions, or AI-attributed performance. The public job/profile route names Ricky The Ising, but does not establish that he owns or maintains the in-house model. The mutual-fund, offshore-vehicle, personnel, and AI-process evidence remain separate. See the capture note.

The Malay-language search also recovered a distinct Kenanga Investors fund route. Kenanga’s August 2024 release describes its Islamic wholesale fund as feeding into the Chicago Global Responsible Strategies target fund and says the target fund uses big data and AI to identify financial data, market trends, social sentiment, and alternative data for signal generation. The Kenanga-hosted information memorandum identifies Chicago Global Portfolios VCC as a Singapore vehicle, Chicago Global Capital Pte. Ltd. as target-fund manager, and July 25, 2024 as the target-fund inception date. This is a product and vehicle disclosure that complements the existing Chicago Global personnel/podcast route; it does not establish that Kenanga operates the research engine. Model family, feature definitions, training data, data rights, validation, execution permissions, and independently audited AI performance remain undisclosed. See the capture note.

September 3, 2026 — South Africa Apex AI-futures route

The South Africa route also adds Apex Capital Management through a 51-minute August 1, 2026 episode of The Day Trading Show, also available on Spotify. The publisher identifies Dylan Maltman and describes Apex’s intraday futures operation, Sierra-based automated order-flow strategies, Chicago colocation, SMA/LPGP capital structures, sub-account risk allocation, and no manual or discretionary trades. The episode’s chapter metadata also surfaces a reported risk-sizing error during an order-management-system import. Apex’s first-party site describes AI-assisted research and strategy development, low-latency futures infrastructure, and separate proprietary and client-SMA divisions. Its About page names Dylan Maltman as Co-Founder & CEO, Jonathan Quenet as Co-Founder & CTO, and states that Apex Capital (BVI) Limited is a CFTC-registered CTA and NFA member; that registration statement remains an Apex self-claim pending direct NFA BASIC confirmation. No model family, training data, data rights, reproducible validation, customer-level performance, or exact live-system boundary is public. Target risk metrics are not treated as realized results. See the capture note.

September 3, 2026 — Chilean Fintual research and GenAI route

The Spanish-language pass surfaced a materially different comparison point: regulated Chilean asset manager Fintual, rather than a verified hedge fund. Fintual’s investment-process page describes a modified CTGAN that generates synthetic return scenarios conditioned on points along the US Treasury yield curve. Those scenarios feed a CVaR-constrained portfolio optimizer; Fintual says the resulting model portfolios are then adjusted by portfolio managers and monitored by investment and risk teams. The same page identifies a 14-year-plus backtest over ten asset classes, but those are Fintual’s stated research results, not an independently audited live-performance record.

The underlying 2024 Quantitative Finance paper names José-Manuel Peña, Fernando Suárez, Omar Larré, and Domingo Ramírez with Fintual affiliations and Arturo Cifuentes of CLAPES UC. Its public arXiv record describes contextual synthetic scenarios, ten asset classes, January 2008–June 2022 evaluation, and out-of-sample comparisons; the publisher record points to public code and states that historical data came from Bloomberg. This links an identifiable investment team to a concrete generative-model research artifact, while leaving current production use, model revisions, feature-level data, and live permissions unresolved. See the capture note.

The same sweep found a separate GenAI surface. Fintual’s terms for Fintual IA describe a customer-facing tool using a network of LLM-based conversational agents for Fintual services, funds, basic market analysis, personal finance, and regulation. The terms describe authorization for personalized account information, model-error warnings, restrictions on access to other users’ data and reverse engineering, and external technology providers. This is customer-service GenAI evidence; the source does not say that the agents select trades or generate alpha. Fintual’s Fernando Suárez biography identifies him as Senior Portfolio Manager and describes machine-learning and quantitative-finance work, teaching at PUC Chile, and academic publications, but does not assign him sole ownership of any production system. The reviewed sources do not name model providers, prompt/retrieval architecture, training data, data rights, or independently measured AI performance.

September 3, 2026 — Brazilian Kinea current AI/ML personnel surface

The Portuguese-language pass also found a current first-party personnel and process surface at Kinea. Its Kinea Gama fund page displays Henrique Pires with the role label “Inteligência Artificial e Machine Learning” and Rodrigo Zobaran as Sócio e Head de Pesquisa Quantitativa. The page separately describes process automation through dashboards built from multiple economic, sector, and company-data sources to support market reading. The same role labels recur on Kinea’s Prev Mont Blanc page.

This is stronger evidence for named current responsibility and an investment information workflow than a generic AI-themed fund page, but it is still a role and process disclosure. The pages do not establish reporting lines, headcount, model families, features, training data, data rights, vendor relationships, evaluation design, agent permissions, or AI-attributed performance. They also do not establish autonomous trading or a hedge-fund strategy. See the capture note.

September 3, 2026 — Brazilian Kadima workflow and ML disclosures

The Portuguese-language pass also recovered a useful first-party history from Kadima, a Brazilian systematic asset manager. Kadima’s ChatGPT management letter says the firm used ChatGPT for coding, plotting, translation, text review, commercial copy, and social-media drafts. It discusses possible investment-research, risk, communication, and algorithmic-trading applications, but separates proposals from tested uses and emphasizes human review, data quality, latency, cost, and reliability. The source does not establish an LLM signal or live trading integration.

Kadima’s 2023 letter describes machine-learning techniques used to improve how some models interpret data, hybrid models, asset-class adaptations, and asset-allocation research. It also describes controls comparing realized trades with backtests, checking execution assumptions and drawdowns, and interrupting a model when discrepancies appear. The 2024 letter adds feature-engineering, variable-importance, hyperparameter testing, and the firm’s own statement that it avoids black-box models. These are useful first-party workflow and governance disclosures, but they do not identify exact algorithms, model weights, features, training data, data rights, provider contracts, live permissions, or AI-attributed performance. See the capture note.

September 3, 2026 — Brazilian Cognus / Cartesius AI-managed-fund route

The Portuguese-language pass also found Cognus Capital’s current first-party manager and fund pages. Cognus says it develops proprietary machine-learning models and methodologies for listed, liquid global assets. Its displayed timeline places the project’s start with researchers and academics in 2009, a first operational model in 2012, further model development in 2018, test-fund and manager structuring in 2022, team expansion in 2023, and new Cognus strategies in 2025. The manager page says research is mainly scientist-led, collaborative, and owned by the manager, with human control of outliers and extreme events.

The current fund page identifies PRIMUS MODERATUS FIC FIF MM RESP LTDA, CNPJ 48.038.268/0001-49, as an open fund managed by Cognus and administered/custodied by BNY Mellon. It describes systematic management using AI, quantitative and computational methods, global liquid assets, short-horizon operations, and external human control of outliers. An earlier Forbes report describes the Cartesius vehicle as using predictive AI while stating that investments were not decided automatically and that trading robots were not used.

This evidence establishes a distinct Brazilian AI-managed-fund and governance route, but not a complete current model inventory. The public pages do not name AI researchers, model families, features, training data, data rights, vendors, evaluation design, execution permissions, or AI-attributed performance. The historical Cartesius naming and current Cognus branding remain separate pending corporate-record reconciliation. See the capture note.

September 3, 2026 — Brazilian AZ Quest Bayes first-party ML validation trail

The Portuguese first-party AZ Quest systematic-funds page adds a dated model-development trail to the existing XP/Bayes case study. It describes a proprietary database built in 2010, more than 100 indicators, a traditional factor-family method first developed in 2012 and now at version F2022, an alternative machine-learning method, and machine-learning methods in portfolio optimization. These statements describe the firm’s published process; they do not identify the algorithms or model weights.

The May 2024 manager letter says the first ML version was implemented in Bayes FIA paper trading. The February 2025 letter names the alternative model IPCA, describes monitoring through the Bayes Reports system, and says its analytics were compared with backtests and the core Fama–French 2022 model. The July 2025 letter and December 2025 letter continue to describe IPCA as being validated in paper trading, with possible gradual introduction to a small sleeve.

This is useful because it exposes a research-to-production gate: named model, paper-trading observation, analytics comparison, and only then a possible limited live rollout. The public record still does not establish that IPCA entered live production, nor does it disclose its features, model family, training data, validation split, data rights, execution permissions, or AI-attributed performance. See the capture note.

September 3, 2026 — Chilean Holdo AI-managed mutual-fund route

The Spanish-language pass found Holdo’s Chile Smart Fund product page, which describes a Chilean-equity mutual fund periodically managed by the firm’s AI portfolio manager, “Harry.” The page identifies Toesca as the Administradora General de Fondos and says the product was created using more than 300 national and international economic variables and millions of simulated portfolios. Holdo’s process article adds historical-volatility inputs, efficient-frontier filtering, correlation analysis, and a long-term portfolio objective.

Local Diario Financiero coverage names Vicente Icaza as CEO, Alejandro Brücher as Head of AI, and Matías Humud as CTO, and attributes algorithm creation to Brücher. A public Dalal Chahuán profile provides an additional CTO personnel route. The sources establish a named AI product, public leadership, and a regulated mutual-fund wrapper; they do not disclose the model family, whether the system is generative, features, training data, validation design, vendor stack, permissions, or independently verified AI-attributed performance. This is a Chilean asset-management comparator, not a hedge-fund disclosure. See the capture note.

September 3, 2026 — Chinese BaoYing machine-learning quant-researcher route

The Chinese-language CUHK-Shenzhen career posting reproduces a BaoYing Fund quantitative-investment role focused on machine learning. The posting asks for stock-selection factor mining over large, multidimensional data using neural networks, tree models, and reinforcement learning; prediction models adapted to the Chinese market; and explicit attention to overfitting and data timeliness. It describes the full path from data analysis and feature engineering through strategy backtesting and support for investment staff, and names NLP, Graph ML, deep learning, alternative data, XGBoost/LightGBM, CNN/RNN, and Transformers as relevant methods.

The listing closed on August 15, 2026. It therefore provides a concrete research-hiring vocabulary and a route for finding later personnel, but does not establish that the role was filled or that any named model reached production. It also does not disclose training data, data rights, evaluation design, trading permissions, or AI-attributed performance. See the capture note.

September 3, 2026 — Chinese Southern Fund ML, large-model, and RAG hiring route

The Chinese University of International Business and Economics employment notice reproduces Southern Fund’s 2026 financial-technology recruitment. It lists a machine-learning quantitative researcher for quant-fund strategy models and evaluation systems. A separate quantitative-trading role says the employee would use modelling for trading research, apply large models to trading decisions, build trading systems, optimize processes, execute orders, and control risk.

The notice also lists AI-algorithm and AI-application roles involving data collection, cleaning, feature engineering, model validation, public-fund research applications, prompt design, RAG workflows, vector retrieval, APIs, and data integration. This separates predictive research, potential large-model trading research, and enterprise retrieval/application work. The notice was updated August 27, 2025 and does not establish filled roles, production deployment, model versions, data rights, permissions, or AI-attributed performance. Southern Fund is an asset-manager comparator, not a hedge-fund disclosure. See the capture note.

The Asset Management Association of China award material adds a later project record for Southern Fund’s multi-agent trading assistant, called “小南同学” or “小喃同学” in different public PDF copies. It describes a generative-model and multi-agent layer with a unified conversational entry point for cross-system execution, plus a terminal/role/network permission model and AI-based grading of real-time instructions. The award credits a 15-person team, including Yang Xiaosong, Li Haipeng, Wang Ke, Yang Dongbo, Huai Quan, and others, but does not assign individual technical ownership.

A May 2026 report from a Fudan finance event identifies Wang Ke as Southern Fund’s transaction-management general manager and reports his statement that the team built a multi-agent trading assistant in 2024 and revised it by the time the award arrived. His description focuses on organizing agent teams, tool calls, task division, process monitoring, and result verification. These sources strengthen the personnel and governance trail, but they still do not disclose model identity, training data, current production permissions, autonomous order authority, or AI-attributed returns. See the capture note.

September 3, 2026 — Chinese CUFEL finance-AI laboratory and agent-arena route

The official CUFEL laboratory page identifies a joint Central University of Finance and Economics–Beihang research lab. It names Zhang Xueyong as director and Wang Yunhong as co-director, and lists financial time-series foundation models, multimodal risk governance, high-frequency-trading governance, and multi-agent systems as research priorities.

CUFEL publicly describes a finance model based on Qwen3 8B with incremental domain training and alignment, including analyst-report and financial-reasoning data, self-distillation, LoRA, GRPO, and model merging. It also presents CUFEL-Q, a multi-agent ETF/FOF arena with common evaluation metrics such as net value, Sharpe, drawdown, and turnover. These are official lab descriptions and public evaluation surfaces; the reviewed capture does not establish reproducible benchmark data, a completed technical report, external-fund deployment, capital permissions, or audited performance. This is an academic research ecosystem route, not a hedge-fund claim. See the capture note.

The CUFEL page also exposes a small academic-to-industry bridge: its institution-published 2025 intern summary lists destinations in data-science graduate study, large-model inference and high-performance deployment, deep-learning derivatives research, and Beijing private-fund CTA, mid/high-frequency, equity multi-factor, and derivatives strategy research. The students are anonymized, so this is a cohort-level placement signal, not a personnel graph. It does not show that CUFEL models or methods transferred to those employers, that the placements remain current, or that any employer granted investment authority. The capture note records the boundary.

September 3, 2026 — Saudi Keheilan–SahmAlgo Arabic partnership route

An Arabic-language search found a dated partnership route that was absent from the earlier Saudi product sweep. Keheilan’s company page describes a June 2026 strategic collaboration with SahmAlgo to provide AI-supported trading tools and institutional financial technology for Saudi public markets. The announcement frames the intended combination as Keheilan’s public-equity and portfolio-management activity with SahmAlgo’s AI-for-financial-data and local-market technology, covering quantitative analysis, investment decision support, and execution.

The announcement is repeated in Keheilan’s public LinkedIn post and Arabic financial-media coverage, while Ahmed Abdelhamid’s public profile adds a named executive route. Keheilan’s site separately presents KAIF as an AI-managed Shariah-compliant global-equity strategy. These materials establish public positioning and a named partner, not a disclosed production stack.

The sources do not establish model architecture, training data, data rights, evaluation, execution permissions, legal fund structure, or independently audited AI-attributed performance. Keheilan is retained as a regional asset manager and technology-investment route, not as verified hedge-fund evidence. See the Arabic capture note.

September 3, 2026 — Indian Fourier Capital Management Hindi/English route

The Hindi/English regional pass found Fourier Capital Management’s public LinkedIn company page. It describes a Chennai-based quantitative investment firm using statistical and machine-learning approaches across commodities, foreign exchange, credit, and European fixed income. The page says the current fund is private to the founder and selected partners and displays a 2025 founding date and a 2–10 person size range.

This is useful as a title-blind Indian manager-discovery lead, but the only reviewed evidence is self-authored company positioning. It does not establish a named or regulated vehicle, named research personnel, model architecture, data rights, live deployment, execution permissions, or independently verified performance. Fourier remains separate from the larger Indian quant managers and is not treated as verified hedge-fund evidence. See the capture note.

September 3, 2026 — Italian regulatory and institutional AI map

The Italian-language pass also found a 2026 OECD–Banca d’Italia report covering AI and GenAI adoption, experimental versus production use, third-party dependence, governance, and applications in forecasting, market analysis, and settlement. It cites a CONSOB–Assogestioni survey of eight large asset managers representing 60% of Italian assets under management. The report provides survey scope, not a current adoption rate or performance result.

The report adds concrete supervisory ML examples: CONSOB’s unsupervised clustering methods for potential insider-trading patterns near price-sensitive events, and Banca d’Italia tools using NLP/ML to cluster more than 10,000 annual complaints, named-entity recognition, and automated reasoning over knowledge graphs. These examples expand the map of where financial AI is operationally useful, but they are surveillance and administrative systems rather than hedge-fund alpha models.

The report does not identify the eight asset managers, their model inventories, portfolio permissions, or AI-attributed returns. It is retained as an Italian institutional comparator, not firm-specific hedge-fund evidence. See the capture note.

September 3, 2026 — Vietnamese XCap/XNOQuant platform and competition route

The Vietnamese-language gap pass found a more substantive first-party route at XNOQuant. The platform describes AI-assisted signal construction, backtesting, deep-learning architectures for complex market data, continuous recalibration, algorithmic risk management, and a process in which high-performing submitted alphas may be reviewed for capital allocation by XCap Fund. It presents price, volume, fundamental, and alternative datasets as inputs.

XCap’s separate company page names Hoang Minh Thang as founder and distinguishes XCap Fund’s proprietary high-frequency activity, XNOQuant’s research and automated-execution platform, and XCap Ventures. The Vietnam Quant Challenge 2026 adds a concrete talent and data surface: Vietnamese market data, factor research, portfolio optimization, order-book simulation, tick-level data, historical replay, and a staged path from signal construction through backtesting and possible capital allocation.

These pages establish public positioning, a named founder, and an unusually explicit research-platform and talent-acquisition route. They do not establish legal fund status, actual capital allocation, model versions, training data or rights, agent permissions, execution controls, participant identities, or independently verified performance. XCap/XNOQuant is retained as a Vietnamese quantitative-technology and fund-discovery route, not as verified hedge-fund evidence. See the capture note.

September 3, 2026 — Vietnamese XCapital/XNOQuant AI-agent hiring surface

The XCap follow-up found public job descriptions that expose a more detailed advertised GenAI architecture. The quantitative-researcher posting connects statistical and ML research across equities, derivatives, and crypto with price, order-flow, sentiment, macro, and alternative data; backtesting; signal metrics; and model integration into production.

The AI-researcher posting adds price prediction, risk modelling, portfolio construction, production experimentation, and collaboration with a Head of AI and PhD-level quant researchers. The AI Engineer posting specifies multi-agent orchestration, memory, tool routing, portfolio and backtest tools, multi-tier LLM routing, financial-document RAG with embeddings and vector search, quant-signal integration, golden-set tests, hallucination detection, and token/latency/cost/audit observability. A separate quant-trading AI-engineer posting adds reinforcement learning, LLM systems, distributed execution, Python/C++, data pipelines, and MLOps.

The postings mention Claude, OpenAI, Gemini, LangChain, and LangGraph as relevant production experience. The XCapital company page names several employees and identifies Bùi Minh Đức as Head of Research in an event recap; Minh Thang Hoang’s profile shows a Tech Lead label, while the XCap page describes Hoang Minh Thang as founder. These public labels are retained as separate evidence and are not treated as a resolved reporting structure.

This is detailed hiring and architecture language, but it remains advertised scope. It does not establish filled roles, deployed components, model versions, data rights, provider contracts, live permissions, or measured investment outcomes. See the capture note.

September 3, 2026 — Brazilian Quantique fund and research route

The Portuguese-language pass found a separately grounded Brazilian manager in Quantique’s company profile, which names Fernando Vieira Santos Filho, Renato Ometto, and Camila Fairbanks as founding partners. The Quark fund sheet identifies a multimarket fund managed by Quantique M3 Investments, while the fund regulation names BTG Pactual as administrator and records CVM authorization Ato Declaratório 20.059. The sheet describes a systematic strategy using quantitative analysis and technology across asset classes and geographies.

A dated InvestNews interview describes the stated data combination as macroeconomic reports, company financial statements, specialist analysis, mathematical and statistical models, and machine-learning processes. Quantique’s site also records an ITA-linked “Machine, Math & Mind” data-science challenge. These sources provide fund, founder, and research-network evidence; they do not disclose model versions, training data, live permissions, or independently verified returns.

September 3, 2026 — Brazilian Newfoundland Eagle historical ML route

The same Portuguese pass found a fund identifier that requires explicit entity and vintage reconciliation. The historical Newfoundland Eagle LB FIA product page describes a Brazilian quantitative fund whose strategies were said to use machine-learning and optimization algorithms, and displays CNPJ 42.747.231/0001-03. A BTG Pactual fund file with a January 29, 2026 reference date independently displays the same fund name and CNPJ and classifies it as an equity fund for non-qualified investors.

MaisRetorno reports an operating fund snapshot and attributes a layered algorithmic architecture and a data lake with more than two decades of alternative data to the fund. Those architecture and data-lake statements are secondary and unverified. The current Newfoundland Capital Management website now presents Pan LatAm and Pantera products and a Newfoundland Iron entity surface, without displaying Eagle on its accessible homepage. The old Newfoundland Malibu product page, BTG file, secondary fund record, and current website are therefore kept as separate time/entity-specific evidence.

This establishes a dated ML-positioning trail and a fund identifier, not a current model inventory, manager identity, data-rights record, production deployment, or performance attribution. See the Quantique capture note and Newfoundland capture note.

September 3, 2026 — Thai SCB Machine Learning fund and personnel route

The Thai-language pass found a regulated asset-management comparator with a more concrete personnel surface than a generic “AI fund” label. The official SCB Machine Learning China All Share prospectus lists a Machine Learning investment group responsible for quantitative equity management. The prospectus names Dr. Poonsak Lohsunthorn as Executive Director of that group, with a USC Ph.D. in Electrical Engineering and USC master’s degrees in Electrical Engineering, Mathematical Finance, and Mathematics. It also lists Pairich Nityanupap in quantitative and Big Data analysis, Satitpong Chantrachirawong as Director of the Machine Learning Equity department with a University of Washington computational-finance degree, Krit Jan-nak as a Machine Learning-group Principal with LSE and University of Reading degrees, and Nattawut Dechabindr as an Associate Director with a Chulalongkorn financial-engineering degree.

The same filing records prior-firm lineage for the group. Krit Jan-nak is listed with previous QIS Capital portfolio-management, J.P. Morgan Securities Hong Kong quantitative-research, and WorldQuant Research Thailand VP of Research roles. Satitpong’s entry lists prior Senior Investment Analyst work at the Employees Retirement System of Texas. These are prospectus entries tied to the fund-manager disclosure snapshot, not a current independent personnel verification or proof that each person works on the same model.

An official SCB Thai article, written by Poonsak in his Machine Learning investment-group role, gives unusually specific public research examples: satellite night-light intensity, ship counts at ports, and NLP applied to documents to extract qualitative company information. It says these inputs need to be converted into structured data and explicitly warns about overfitting relationships that work in historical samples but fail in live conditions. A Thai media archive also identifies Satitpong in a January 2023 Machine Learning equity interview.

The current SCB Machine Learning Thai Equity fund page states that SCBMLT(E) uses quantitative analysis and Machine Learning for Thai equity selection through a system developed by the manager. The page displays data dated September 1, 2026 and a December 2017 fund establishment date. It adds current product-policy evidence, but not the system’s feature set, model family, training window, turnover, or live decision permissions.

This is mutual-fund and asset-management evidence, not hedge-fund evidence. The sources do not identify model versions, training corpora, data licenses, production endpoints, live permissions, or AI-attributed returns. See the capture note.

September 3, 2026 — Vietnamese XNOQuant event, speaker, and talent route

The Vietnamese-language event pass found a second public layer that the job search did not expose. The XNOQuant event page for a December 6, 2025 workshop names Bùi Đức as Quant Research Lead and ThS. Ngô Hiển Dương as an AI specialist in quantitative investing. Its agenda covers quant-fund workflow, strategy development, risk management, market microstructure, transaction costs, slippage, capacity, and regime change. The speaker page for Ngô Hiển Dương describes him as a Product Owner at Vietnam Invest Tech and reports previous data-science and quant-research affiliations including WorldQuant Associate, IRD Vietnam, Gnosis Tech, ICLS Tech Vietnam, and Yeager Technologies. It also lists AI/ML, generative AI, reinforcement learning, and financial modelling as expertise labels. These are organizer-published biography claims, not independent evidence of a named model, dataset, paper, or live investment permission.

The XCapital Vietnamese recap also refers to an Xbot Research Lab team supporting quant-community sessions. That creates a new lab/personnel search route, but the public page does not define the lab’s legal status, headcount, reporting line, research output, or repository. Separately, an XCapital recap identifies Bùi Minh Đức as Head of Research at XNO Quant. Because the public pages use both “Bùi Đức” and “Bùi Minh Đức” without independent identity evidence, the article keeps those labels source-specific rather than assuming they are the same person.

The CTE FTU competition announcement and later XCapital update add a talent and data-exposure route: individual alpha simulation, financial datasets, team strategy development, training, mentor review, and final strategy presentation on XNOQuant. A public participant profile reports a third-place finish for Hiếu Nguyễn Ngọc. This is evidence of a public competition and recruiting surface, not evidence that the participant joined XCapital, received capital, or disclosed a reproducible strategy.

The competition pages disagree on timing: the initial announcement gives a July 13 registration close and August 15 final, while later company and event notices show a July 26 Round 1 close, a top-100 progression, and an August 21 final-night notice. Those are retained as conflicting snapshots. The capture note records the route and its evidence boundaries.

September 3, 2026 — Turkish Tera TMV algorithmic-fund comparator

The Turkish-language reconciliation found a regulated quant vehicle that should not be confused with an AI claim. KAP identifies Tera Portföy Algoritmik Stratejiler Serbest Fon, ticker TMV, as a perpetual free fund established by Tera Portföy Yönetimi A.Ş. It names Engin Döğenci as fund manager, appointed August 27, 2024, and records the fund’s June 28, 2022 issue date.

The KAP strategy wording describes algorithm-supported short-term capital-gain systems across BIST equities, rights, warrants, futures, options, and other derivatives, with hedging through correlated domestic and international instruments. The prospectus records the original statistical-arbitrage structure and 2023 name change.

The reviewed filings do not claim machine learning, generative AI, alternative data, model architecture, or AI-attributed performance. “Algorithm-supported” is therefore retained as algorithmic evidence only. Tera TMV is a useful Turkish quant comparator, not an AI-fund classification. See the capture note.

September 3, 2026 — Korean Exponential Asset Management AI and multi-manager route

The Korean-language pass found a firm-controlled technology page that is more specific than an AI-themed product label. Exponential’s technology page describes a deep-learning investment model and a multi-algorithm platform that combines algorithms within a fund, evaluates risk and profitability, measures correlation, dynamically allocates size, and uses global-market and corporate-fundamental data. It also says portfolio managers’ knowledge and experience are used jointly with the algorithms. These are Exponential’s own descriptions; the page does not expose model versions, training data, or live permissions.

The manager’s 2026 regulatory personnel report lists Kim Taesun as CEO with overall responsibility; Namgung Jaehoon, Kim Junsu, Jung Jaewoo, and Jang Gichan with investment responsibilities; and Gong Dahyeon and Seong Junmo in operations and marketing. The Korea Financial Investment Association quant-research posting advertises regression/correlation-based model development, PM-level performance and contribution decomposition, capital-allocation analysis, risk monitoring, stress tests, scenario analysis, and factor screening. It is hiring intent, not proof of a filled role or production deployment.

A 2026 ESTsoft filing separately identifies Exponential as a group asset-management subsidiary and describes quantitative models, AI-based research, a multi-manager system, and risk-monitoring tools. It also describes ESTsoft’s own Alan LLM, RAG search, and agentic-AI activities, but does not say those systems are used by Exponential. The parent-company stack is therefore kept as adjacency, not investment-system evidence.

TheBell’s dated reporting adds that Exponential applied an AI-described manager-characteristics system to the SQUARE 1 multi-manager fund, using profitable/loss-day and realized-gain/ loss measures for manager review and allocation support; it also reports a securities-firm B2B contract. A separate IPO-fund report describes a roughly fifteen-input workflow using accumulated IPO and market-response data to estimate attractiveness, volatility, and lock-up horizons, with managers retaining participation and sizing decisions. The public record does not disclose the tool’s model type, the B2B client, validation protocol, data rights, current status, or independently verified returns.

This route adds a named Korean personnel and hiring surface plus explicit deep-learning, algorithm-allocation, manager-analytics, and IPO-model claims. It does not establish current model ownership, autonomous trading, production scale, or AI-attributed performance. See the capture note.

September 3, 2026 — Japanese MUFG AI Japan fund and MTEC model route

The Japanese-language pass connected previously separate MUFG/MTEC records into a single route beyond the existing Mizuho and Sumitomo Mitsui DS capture. MUFG’s AI Japan Equity Open product page identifies a Japanese-equity absolute-return product, while the 2026 prospectus links its investment advice to MUFG Trust Bank and identifies the MUFG Trust Investment Technology Institute (MTEC) and MUFG Trust Bank as the developers of its AI models. The product combines individual-stock selection with stock-index-futures allocation. The bank page uses a hedge-fund category, but the prospectus describes a regulated investment-trust wrapper; it should not be treated as a private hedge-fund organizational disclosure.

MUFG’s 2017 first-party account names historical personnel Noriyuki Okamoto (岡本訓幸), Masato Ishibe (石部 真人), and Yasuhiro Kōnomaru (鴻丸靖弘). It describes news-text mining, deep learning for short-term market moves using roughly 300 data types, random forests for medium- and long-horizon returns, and fractal analysis for turning points. It also says humans check transaction information and execute trades, with AI handling the data-analysis portion. The article’s 2008.4–2016.3 comparison with TOPIX is a historical company-reported simulation, not current live performance or independent validation.

The 2026 monthly report names stable-high-dividend, news-pick, daily-prediction, monthly-prediction, and turning-point-prediction models. It describes separate stock and futures allocation roles and changes in effective domestic-equity exposure. These are dated product-report labels, not reproducible architecture or evidence that one model is better than another.

MTEC’s current asset-management page describes research into news and important events, economic networks, and investor behavior, with Japanese/global equity, FX/rates, screening, risk, portfolio-optimization, smart-beta, and multi-asset model work. Its data- analytics page lists deep-learning index prediction, facial-expression analysis, earnings-call audio analysis, reinforcement-learning parameter estimation, large-scale news/social-text analysis, TSE order-book processing, cloud HPC, and RPA. The page is a capability surface, not proof that every modality is live in the AI Japan product.

The MTEC internship page reveals a broad research pipeline: futures prediction, economic-indicator surprises, buybacks, short-data equities, HFT volume, news text, VaR, integrated reports, LLM evaluation, generative-AI document extraction, optimal stopping, RAG, box embeddings, JAX Greeks, and GNN robustness. These themes indicate where the institute publicly recruits and experiments; they do not establish production deployment, a named model owner, data rights, or AI-attributed returns.

The route therefore adds a long-running, named institutional model lineage and a current multimodal research surface. It does not establish current model ownership, autonomous execution, model superiority, or a complete production stack. See the capture note.

September 3, 2026 — Russian and Thai title-blind fund routes

The Russian-language pass added a mixture of first-party product evidence and attributed media claims. These routes are included as public-signal records, not as a ranking or a conclusion about comparative capability.

The Russian-language Gate Learn description of Gate Quant Fund, last updated March 27, 2026, adds a crypto-platform comparator. Gate describes AI, deep learning, big-data analytics, market-neutral hedging, and real-time monitoring of NAV, position sizes, and maximum drawdown, and lists several named strategies. Its example yields are dated November 21, 2025 and are promotional, Gate-published figures rather than independent performance evidence. The page establishes a public product vocabulary, not a separately verified hedge-fund manager, model inventory, named personnel, data rights, execution permissions, or audited returns. See the Gate capture note.

The Alfa Capital Quant product page describes a Russian-equity exchange-traded mutual fund whose machine-learning and AI signals contribute to security selection and the timing of position closures, with professional managers participating. It names Nikita Elenberger as manager and describes horizons of days to several weeks. The page does not identify model family, features, retraining, validation design, trading permissions, or independent AI attribution.

The ATON Portfolio of Strategies page describes a qualified-investor interval mutual fund combining seven strategies, including market-neutral, event-driven, risk-parity, and an AI-based strategy. It names Isuf Atskanov as responsible for a long/short market-neutral strategy and an AI-based strategy. ATON says the AI strategy analyzes prices and forms a market-neutral portfolio, while an optimization algorithm allocates capital across strategies using return and drawdown measures. It also says managers supervise and can adjust the model. This is a concrete governance description from a product page, but it remains a company claim without a public model card, data-rights record, or independent performance attribution.

A Russian Finam report also surfaced several recently formed managers whose principals described AI workflow use. MQT founder Dharmesh Maniyar was quoted on AI expanding the research capacity of small teams; MQT’s public company profile describes global discretionary and systematic macro strategies and a Tudor Investment Corporation lineage. Palinuro Capital was described as using large language models in research and options-hedging work; its public profile describes a discretionary global macro hedge fund. Alpha Curve’s principals were reported as using Claude to analyze international consumer-price data; its official site confirms the Geneva credit office and fixed-income fund but does not confirm that workflow. Calibrate Management’s Eric Lonergan was reported as using AI to screen a broad global asset set into a smaller set of ideas; Companies House confirms the firm’s fund-management business. These AI statements are retained as attributed media claims, not as verified model or production disclosures.

The Greek-language Newmoney rendering of a Bloomberg report adds Osmosis NL as a separate small-manager route and names Victor Verberk as CEO and Head of Investments. The report describes an approximately 15-person fixed-income operation using AI to compress research timelines, and names Azure and Bloomberg as its two core platforms. It also attributes a reduction in a three-year hiring plan to AI, but provides no measurement method. This is publisher-reported operating evidence, not first-party confirmation: no model, agent architecture, training data, data rights, permissions, execution logs, or independently measured outcome is disclosed. See the capture note.

The Thai pass exposed an important product-structure distinction. Asset Plus’s Point Artificial Intelligence Hedge Fund page describes a non-retail open-ended product that invests 96.82% in the Wellspring GBL-Q unit trust, with displayed size of THB 459,254,882.99 as of June 30, 2026. The title describes exposure to global hedge-fund strategies focused on AI; the page does not establish that Asset Plus itself builds or operates the underlying models.

The Thai First Plus FP QUANT announcement adds a separate quantamental route rather than an AI claim. It describes China A-share selection using quality, valuation, volatility, momentum, turnover, analyst consensus, and market-sentiment inputs, with monthly rebalancing and outsourced management by First Plus Asset Management Pte. Ltd. in Singapore. The DAOL Quantum Computing Fund page provides a useful false-positive control: “quantum” in that product name refers to a feeder fund investing mainly in a VanEck quantum-computing ETF, not proof of quantum or AI decision-making by the Thai manager.

The capture note records the full route and evidence boundaries: Russian and Thai title-blind pass.

September 3, 2026 — German, Chinese, and Japanese model routes

The next regional pass added model-specific language that is easy to miss when searching only for English AI or hedge-fund terms.

The German Ampega product page for GET Capital states that machine-learning algorithms and an AI-driven adviser model are used for global-stock selection and for controlling the fund’s equity-versus-cash allocation. It also says that the fund did not yet have sufficient performance history for useful historical-performance information. This is product-level evidence from a fund page, not private hedge-fund evidence, and it does not identify the model, features, owner, training data, validation, or permissions.

A separate Deka-PB Multi Asset Quant page describes machine learning as one complementary strategy within a diversified multi-asset product and refers to modern risk-control systems. It does not establish autonomous investment decisions or a particular model.

The Chinese-language CITIC Prudential quantitative-investment job posting specifies research on machine learning, deep learning, alternative data, large-scale prediction, live-strategy integration, performance attribution, overfitting, stability, regime adaptation, and large-model inference optimization. This is an unusually detailed statement of intended capability, but a job posting does not prove a filled role or production deployment.

The Hainan Shengfeng Private Fund posting names price-prediction and risk-control models, paper reproduction, multi-GPU and multi-machine training, high/medium/low-frequency research, and LSTM, Transformer, and MLP model families. It also calls for standardized experiment logs and live visualization. These are hiring requirements, not evidence of deployed models, live capital, or validated returns.

The Japanese Society for Artificial Intelligence paper on LLM feedback in automatic stock-strategy generation describes a hypothesis-to-code-to-backtest-to-feedback loop on TOPIX 500 stocks excluding financials. It evaluates returns, Sharpe ratio, drawdown, IC, and factor exposures, and reports both useful feedback interpretation and cases where revisions did not improve results. The affiliations are University of Tokyo, Matsuo Institute, and Osaka Metropolitan University. This is an academic methodology route rather than evidence about a named fund.

The capture note records the evidence tiers and the JoinQuant deduplication decision. The new routes broaden the public model vocabulary to portfolio-level allocation, distributed training, live-strategy integration, attribution, and iterative LLM research, while leaving proprietary weights, data rights, production telemetry, and AI-attributed performance undisclosed.

September 3, 2026 — Portuguese Magnetar AI-vehicle report

The Portuguese-language Bloomberg Línea rendering of a Bloomberg News report adds a dated, highly specific but source-qualified Magnetar route. The article says the firm planned a new vehicle, expected by the end of 2026, in which hundreds of AI bots would scan investments, analyze companies, recommend opportunities, and project trends. It says humans would retain final trade authority. The article attributes these details to people familiar with the matter who requested anonymity, and Magnetar declined to comment; this is therefore reported intelligence, not a firm-controlled launch announcement.

The report describes a predominantly long-biased, buy-and-hold strategy plus a smaller component seeking signals on millisecond horizons. It attributes to the system a large-scale signal-processing function, multiple NVIDIA servers running continuous workloads, and an inference layer coordinating different AI agents for task-specific work. It names Trevor Mottl as Magnetar’s Head of AI Quant and links his reported background to Fusion Fund, Walleye Capital, Lazard Asset Management, Balyasny Asset Management, and Man Group. The Bodhi 2026 agenda independently provides a dated public listing for Mottl as “Portfolio Manager, AI Quant Equity, Magnetar,” but does not establish the planned vehicle or its architecture.

This route materially expands the public vocabulary around AI investment infrastructure—agent orchestration, continuous GPU workloads, signal processing, and human trade approval—while leaving launch status, model identities, training data, permissions, and performance unresolved. It is kept separate from the older Odd Lots Magnetar financing record. The full evidence boundary is in the Portuguese Magnetar capture note.

September 3, 2026 — Portuguese NeoFeed Brazilian asset-manager AI workflows

The Portuguese-language NeoFeed investigation is one of the most information-dense foreign-language routes found in this pass. Published July 22, 2026, it says it interviewed eight independent managers representing more than R$100 billion in investments. The details below are attributed interviews and manager statements, not independent audits.

At Galapagos Capital, macro manager Jorge Dib reportedly built an overnight briefing that routes work across Claude, Grok, and Gemini, assigning each model a different function. At RBR Asset, technology director Thomaz Pougy described a corporate Claude account, deeper training for one employee in each area, and agents integrated into proprietary systems. He said a fund-of-funds workflow scrapes hundreds of public documents for pre-analysis and that operation registration fell from more than a day to approximately half a day. The article does not provide the prompts, data rights, evaluation protocol, or independent timing study.

The article also reports different coverage and organizational patterns. Paramis Capital CEO Danilo Ribeiro said AI lets an analyst cover roughly two or three times as many stocks without adding staff. CVPar analyst Davi Costa described using AI to build Python and R tools for operational work. At Reach Capital, Henrique Lara and Iara described a daily update agent and an internal idea repository combining reports, earnings-call transcripts, and internal meetings, with manager judgment retained.

The most explicit staffing and agent detail came from Drýs Capital (formerly Equitas). CEO and co-founder Luis Felipe Amaral said the equity team moved from nine analysts and a manager to two managers, each with a self-built agent “orchestra.” The reported system scans about 20,000 stocks, collects calls, interviews, and company information, and returns material for manager review. Amaral also attributed a daily Claude Code update cycle and a monthly cost of US$2,000–3,000 to the agent system. NeoFeed separately reports performance and staffing figures; this report does not treat them as independently verified or AI-attributed.

For quantitative research, Daemon Investments CEO Sérgio Schirato described AI for strategy simulation, parameter training, cross-market testing, and variation analysis, saying work that previously took months could be done in days. He also described Anthropic enterprise tooling connected to the shared code base. Legacy Capital quantitative manager André Dias described agents crossing time series with internal analyst comments, evaluating historical analyst hit rates, and classifying earnings-call tone as optimistic or pessimistic. These disclosures identify workflows, not model performance or production permissions.

This route adds public evidence for multi-model routing, agentic document ingestion, broad-universe monitoring, shared research memory, code-generation assistance, quant experimentation, and earnings-call text classification. It does not establish a proprietary foundation model, autonomous trading authority, legally cleared training corpus, reproducible backtest, or independently measured AI-attributed alpha. See the capture note.

September 3, 2026 — Arabic and Gulf AI asset-manager routes

The Arabic and Gulf regional pass added several different kinds of evidence that should not be merged into one “AI hedge fund” category.

Derayah’s 2025 Arabic annual report describes enriching its digital platforms with AI solutions and identifies a closed-ended Sharia-compliant fund focused on AI and future technologies that invests through HOF Capital’s Strategic Opportunities Fund. This is an AI-themed product and an operational-AI disclosure; the report does not identify an internal portfolio model or AI-attributed performance.

The Arabic Algotoria site provides an unusually explicit boundary: it says “AI-Native” refers to AI-supported operations under human oversight, while investor trading decisions remain deterministic and rules-based. It describes programmed risk limits and automatic position reduction. These are first-party claims, with no public model inventory or independent performance audit.

Deep Finance Capital identifies itself as a brand of Rasameel Investment House Ltd and describes AI engines, permissioned portals, human review, and immutable audit logs. Its related Deep Finance Analytics site claims three agents monitor markets, filings, and regulatory changes, alongside LLM narrative-to-math tools, factor/VaR/time-series/valuation engines, confidence scores, and API or on-premise delivery. These are firm-authored capability claims, not proof of client deployment or investment outcomes.

The Invesense site exposes an AI Focus Strategy alongside Global Equities, GCC Quant, Multi Strategy, and Global Sukuk. The DFSA register identifies Invesense Asset Management Limited as a DIFC company with reference number F002331 and a November 17, 2014 licence date. The public pages do not resolve whether the AI strategy is thematic exposure or internally generated signals, and do not disclose architecture, holdings, or AI-attributed results.

Finally, the Saudi Capital Market Authority permit list lists Lamha Tanbu for “the use of A.I. in advisory,” and the Saudi Exchange announcement records completion of the experiment’s commencement requirements. This is regulatory evidence of an AI-advisory test, not evidence of a fund, model, execution authority, or returns.

The route therefore adds a sharper taxonomy—AI-themed products, AI-native operations, quant/risk infrastructure, and regulated advisory experiments—while leaving personnel, current deployment, data rights, model versions, and independently measured outcomes unresolved. See the capture note.

September 3, 2026 — Central and Eastern Europe and Israel title-blind routes

The next regional pass added regulator records, academic pipelines, product documents, and a new office-specific hiring route.

The University of Warsaw AI in Investments course describes the complete systematic pipeline from data and features through signals, forecasts, portfolio, risk, execution, performance, and AI agents. Its syllabus includes classical ML, sequence models, transformers, Chronos, reinforcement learning, LLM/RAG, and multi-agent architectures with veto logic, alongside walk-forward and double-out-of-sample validation. The course’s illustrative results use public or synthetic data and are not evidence of fund deployment. The University of Warsaw Quantitative Finance Research Group adds a research and talent route focused on automated transaction systems.

The Czech National Bank AI-crypto warning is a useful control against false positives: AI branding around trading profits does not establish a regulated strategy. Czech Fond Quant and RSJ careers add algorithmic-investment and quant-trading surfaces without public model or independent-performance detail. Romania adds the UPB Financial Computing program and a Romanian-language ESMA warning covering AI recommendations and unlicensed tools.

In Greece, the FTSAI laboratory connects quantitative finance, financial engineering, trading algorithms, AI, and automated decisions, with Achilleas Zapranis identified as director. The QFRA 2026 symposium adds an event route covering AI in asset pricing, deep-RL allocation, high-frequency data, and quantitative risk. These are academic and conference routes rather than fund-deployment evidence.

The DRW Tel Aviv ML Engineer role adds a new office/personnel surface into DRW’s Algorithmic Trading Research Group. Final and 44 Quad add Israeli systematic/HFT hiring routes involving ML research, signals, execution, and risk. Vacancies do not establish filled people, model ownership, or live permissions.

The Serbian TeleTrader quantitative-analyst page describes Belgrade-based work on hedge-fund strategies, algorithmic trading, machine-learning methods, and financial-data models. The Slovenian University of Maribor Institute of Finance and Artificial Intelligence adds academic work spanning optimization, VaR, Monte Carlo, neural networks, transformers, LLMs, and supercomputing.

This pass expands the talent, regulatory, and product-document map more than the verified list of proprietary fund models. It does not establish proprietary weights, legally cleared training data, current production deployment, live capital authority, or AI-attributed performance. See the capture note.

September 3, 2026 — South Asian finance-agent and language routes

The South Asian pass added an open evaluation artifact, a dense finance-AI personnel trail, and multilingual financial-NLP research.

The FrontierFinance preprint, submitted August 12, 2026, introduces 220 expert-written queries and 11,543 source- attributed rubrics across six investor workflows. It evaluates long-form finance-agent work with public data and studies how tool harnesses affect quality, cost, and efficiency. It reports that screening and sector/industry/ macro work remain difficult across evaluated systems. This is an offline benchmark, not evidence of live capital, investment authority, or returns.

The Samaya team page describes Expert AI Agents for global financial-market information and investment research. It publicly lists Maithra Raghu, PhD (Google Brain), Ashwin Paranjape, PhD (Stanford), Richard Diehl Martinez, PhD (Cambridge), Yuhao Zhang, PhD (AWS AI), Ozan Koyluoglu, PhD/MBA, Patrick Folan (Barclays), Jeffrey Tha (JPMorgan), Rainbow Chik (CC&L/GoldenTree/Orbis), Rajul Bothra (Goldman Sachs), and Arash Alidoust (Verition Fund Management), among others. These are self-reported roles and lineage; the page does not establish customer deployment, capital authority, or portfolio permissions.

Arise Labs describes an autonomous investment-research platform and names Hengxin Fun as founder and CEO, Xueguang Ma as co-founder and CTO, and Yuansheng Ni and Shengyao Zhuang as founding members. Their public biographies connect market/HFT, Amazon and ByteDance systems work with University of Waterloo and University of Queensland information-retrieval, multimodal, and LLM-ranking research. The site says it launched August 28, 2026; deployment, regulatory status, and performance remain unresolved.

The published Frontiers in Artificial Intelligence paper describes SSABE-TSCM for low-resource Bangla financial-news sentiment using semi-supervised learning, temporal contrastive modeling, and Temporal-SHAP explanations over a five-year, eight-sector corpus. Its reported metrics are offline historical-news results, not evidence of a live trading system or an LLM agent. An IIT Bombay research profile adds an Indic-language reasoning and reward-shaping route under Prof. Ganesh Ramakrishnan. The SPJIMR paper adds Indian AI-investing governance and systemic-risk analysis.

This route adds personnel lineage, an open evaluation artifact, multilingual financial NLP, and governance research. It does not establish which models a hedge fund uses, whether an agent can trade, what data rights exist, or whether offline metrics transfer to live decisions. See the capture note.

September 3, 2026 — Spanish-language Allaria quant-fund disclosure

An additional Spanish-language route comes from the regulated-product pages of Allaria Fondos. Its Allaria Equity Selection — Acciones Latam product is labeled a “Fondo Quant”; the manager says it combines fundamental analysis with advanced AI and machine learning for portfolio construction, focused on Brazilian CEDEARs. The manager’s launch announcement is dated February 23, 2025. The product page names Allaria Fondos Administrados SGFCI S.A. as administrator and states that it is registered with Argentina’s CNV as no. 29.

This is a concrete public disclosure of an AI/ML-labeled investment product, but it does not reveal model families, features, training data, validation, retraining, personnel, execution autonomy, or AI-attributed performance. It is also a mutual-fund product rather than evidence of a hedge fund. The page’s portfolio and performance fields are point-in-time and should be rechecked before any quantitative use. See the capture note.

September 3, 2026 — Portuguese, French, and Japanese podcast routes

The podcast search added an ANBIMA Portuguese episode with Pedro Simonetti of Giant Steps Capital. The publisher summary says Giant Steps applies an algorithm to US-company earnings-call transcripts to extract points used in investment decisions. It also emphasizes problem definition, reliable and protected data, and platform choice before AI adoption. This is a publisher summary; the full transcript and model details remain unresolved.

A French Bodic episode frames fund AI adoption as a distinction between individual prompting and systemic data infrastructure. Its operating vocabulary includes versioned assumptions, common KPI definitions, golden-source data, reproducibility, and traceability. It is an industry discussion, not evidence of a named firm’s deployment. Quants Research’s Japanese robo-adviser page adds a vendor route describing algorithmic portfolio support, automated rebalancing, and simulation tools for asset managers and banks, without naming customers or models. See the capture note.

The German podcast search recovered a previously untracked BIT Capital episode with a publisher-hosted timestamped transcript. At 09:22–10:01, Jan Beckers, CIO and founder, says BIT Capital built approximately 15 agents that monitor separate portfolio-risk factors each morning and prepare updated exposure information. At 10:26–10:51, he extends the agentic-use-case description beyond research into areas such as marketing and says the firm’s AI bill had multiplied over a period of months. The episode describes monitoring and briefing, not trade placement or autonomous portfolio authority; it does not identify models, data sources, evaluation gates, or permissions.

The same search recovered a historical Antiloop/Lynx episode listing from October 7, 2020. It identifies Martin Sandquist as Lynx founder and CIO of Antiloop Hedge and lays out a 3M framework—pattern recognition, machine learning, and macro—with chapter markers on replacing patterns and detecting when they stop working. This is dated practitioner evidence about model decay, not evidence of current Antiloop or Lynx implementation. See the German podcast capture note.

September 3, 2026 — Finnish, Dutch, and Turkish fund routes

HCP Quant’s Finnish product page describes a quantitative value fund using an algorithm to screen tens of thousands of small and mid-cap companies. It exposes fund documents and audited portfolio- return history, but does not identify an AI model, features, training data, or model-level attribution.

GA Asset Management’s Dutch site describes alternative funds in global long/short equity and managed futures. Its process disclosure says positions are sized by model, research and risk are documented and repeatable, and ML/statistical methods test whether relationships survive outside the sample. It publishes no model inventory or performance.

Azimut Portföy’s Turkish ITY page names a venture fund focused on AI, ML, big data, cybersecurity, blockchain, fintech, software, and robotics companies, and identifies Koray Ucuzal as portfolio manager. This is thematic venture exposure rather than evidence of an AI-driven public-markets trading process. See the capture note.

September 3, 2026 — Polish search finds a licensed AI/ML asset-management route

The Polish-language pass found Aixon Investment Management, which publicly describes professional-client asset management, the Aixon AI Multi-Factor closed-end fund managed for Baltic Capital TFI, and a separate route with Trigon Dom Maklerski. Aixon says it uses machine learning and deep learning for equity selection and portfolio optimization, forecasts monthly returns for more than 2,000 U.S.-listed companies, and processes more than 100 predictive signals through monthly factor analysis. Its site states that the company operates under Polish Financial Supervision Authority permission DIF-DIFZL.4010.18.2021 dated December 6, 2024. These are first-party legal, product, and methodology statements; they do not disclose model families, features, training or validation windows, data rights, production permissions, or independently audited AI-attributed performance. Aixon is an asset-management route, not evidence of a private hedge fund. See the Polish capture note.

September 3, 2026 — Japanese and French model-surface routes

The local-language pass also found model disclosures that do not use “AI” in their titles. ALAMCO’s Japanese product page describes a domestic-equity mutual fund built around a proprietary stock- evaluation model. The page does not identify the model family, features, training data, validation, or personnel, and does not itself make an AI claim.

A Tokio Marine Asset Management publication uses Voleon as an example of machine-learning models seeking finer-grained market nuances beyond simple value and momentum factors. This is secondary industry commentary, not a Voleon technical disclosure or independent audit; it does not establish current model details, permissions, or returns.

A French Vivienne Investissement research posting proposes a deep-learning study for quantitative asset management, including generation of synthetic multivariate financial series. It is a dated research and recruiting artifact rather than evidence of a current production system. See the capture note.

September 3, 2026 — Southeast Asian institutional and product routes

KWAP’s institutional announcement describes a 2023–2025 AI-adoption programme with proofs of concept, sandboxing, a Gen-AI Lab, and AI-enabled analytics for investment and retirement services. It names CEO Datuk Hajah Nik Amlizan Mohamed and UMK’s Prof Ts Dr Arham Abdullah. The announcement does not establish that AI makes portfolio decisions or executes trades. Malaysian Securities Commission technology-risk guidance adds explicit governance topics—explainability, auditability, validation, privacy, resilience, and portfolio/trading-algorithm use cases—but is not firm deployment evidence.

In Indonesia, an HSBC product note states that OJK-supervised Batavia Prosperindo Aset Manajemen used big-data processing through AI/ML to support management of a 2021 global ESG Sharia equity product, with BlackRock as technical adviser. This is product-level AI/ML assistance, not proof of an internally built model, live execution, or AI-attributed returns.

Vietnamese title-blind searches found Investify, a research-agent and backtesting product that explicitly says it is not an asset manager, plus SENAI and Revo AI’s launch article. The latter claims a public AI-operated fund launch, but no corresponding regulator record was resolved in this pass; it remains a lead, not a verified fund or production-system fact. A Thai SEC algorithmic-trading study adds a useful distinction between execution algorithms and profit-seeking algorithms. See the capture note.

September 3, 2026 — French, Italian, and Portuguese institutional routes

The foreign-language pass also exposed several distinct disclosure types. The French AMF report describes surveyed financial-entity uses or near-term plans across NLP, image recognition, transcription, PDF extraction, market analysis, compliance, and supervisory eDiscovery. It is useful for mapping workflow categories, but is largely self-reported and does not identify proprietary models or performance.

Edmond de Rothschild France’s 2025 annual report records recruitment of an active quantitative-management team using structured decisions, mathematical models, algorithms, and machine learning, with a product line planned for 2026. Ossiam’s Europe ESG Machine Learning factsheet describes a systematic equity strategy that reassesses ESG/financial relationships as new information arrives. Neither source discloses model families, training data, validation, or current deployment details.

In Italy, ANIMA Europa AI and its prospectus state that allocation may be determined by proprietary models that can include AI modelling and name Luca Libralato, Simone De Leonardis, and Dario Giacomelli in the quantitative/systematic context. A Generali Quant Data Scientist posting specifies predictive signals, portfolio construction, validation, monitoring, cloud deployment, and Python-to-C++ production responsibilities. These are product-policy and recruiting signals, not proof of live autonomous trading.

Portuguese-language sources add a historical engineering and governance trail. An H2 Asset role brief describes a Brazilian proprietary trading system using data analysis and ML, while a KPTL fund document describes mathematical models, large-scale data processing, and ML for thesis testing and short-term spreads. NeoFeed’s report on Clave’s Cortex describes daily PM review, manual risk reduction, constraints, and a stop function around an automated multi-market monitoring workflow. These materials are dated and do not establish current code, permissions, or returns. See the capture note.

September 3, 2026 — French Itavera media separates thematic AI exposure from internal AI

French video coverage adds Itavera Asset Management, where Rolando Grandi presents the launch of an actively managed AI-themed fund and describes his investment focus on the AI economy. A related News Asset Pro segment frames the manager’s public thesis around AI-market size, company categories, valuations, and AI-washing. The launch coverage identifies Thibault Saint-Raymond, Grandi, and Charles Beriot as part of the team and identifies Itavera as a Valoria Capital subsidiary (launch report).

Grandi’s public LinkedIn update reports more than €200 million of subscriptions across three funds and more than €500 million of assets under management at the end of 2025. Those figures are self-reported and not independently verified here. The route is useful because it prevents a common false positive: an AI-themed fund and an AI investment thesis are not evidence that the manager trains or deploys AI for security selection. The reviewed material does not disclose an internal model, training data, model provider, research lab, data rights, portfolio permissions, or AI-attributed performance. See the capture note.

September 3, 2026 — French, Danish, and German searches expose different disclosure layers

The regional title-blind pass added three distinct routes rather than one uniform “AI fund” category. Valeyre Research is a Cannes-based research and signal provider that says it supplies hedge funds with automated, systematic mid-frequency signals across 3,000 stocks and 70 futures in Europe, the United States, and Asia. It describes one-week reversal and one-month trend signals, hedge-fund advisory work, and research in neural networks, NLP, and computer vision. The page does not identify clients, models, datasets, personnel, permissions, execution authority, or audited results, so the public evidence supports a signal/advisory route—not an attribution to any client fund.

Methodica Management is a Danish-regulated AIFM (registration 25065) whose public page describes fully automated systematic trading in liquid indices and FX. Its FX fund description uses diversified statistical-arbitrage strategies across major G8 currency pairs and a market-neutral objective. The page does not claim machine learning or GenAI; it is therefore retained as an automation and quantitative-trading route, not classified as an AI program.

Hamburg-based GRADTEC / Gradient Technologies provides a new agent-workflow disclosure in this pass. Its AI Research Agent page says agents read filings, transcripts, news, and fundamentals across 15,000+ stocks; form testable hypotheses; gather evidence on companies and peers; validate against fundamentals and quantitative signals; and output conviction-scored signals with rationales. The advertised delivery surfaces include dashboards, watchlists, earnings forecasts, REST APIs, alerts/webhooks, and data export. The company says its corpus contains more than one million documents over 20+ years, but that is an unverified first-party claim about the training/data environment, not proof of data rights, model weights, or end-to-end training. The site names Pablo Hebestreit and Jan Wöltjen as managing directors; its jobs page reported no open positions when checked. The pages do not establish autonomous trading, live capital authority, independent evaluation, or audited performance.

The pass also rediscovered Genio Capital and One Eleven Capital’s French page, but both already have ledger entries and article coverage. The duplicate audit therefore kept them as existing records rather than inflating the foreign-firm count. See the regional capture note.

September 3, 2026 — Dutch searches separate ML strategy, operating AI, and service-provider routes

SysCat Capital describes a Dutch quantitative firm trading systematic global-equity and derivative strategies in the minutes-to-hours interval. Its public research vocabulary includes predictive engineering, nonlinear dynamics, complex adaptive systems, turbulence modeling, signal processing, and ML. The team page names Mahbod Elmi (CIO), Max Waaijers (CTO), Arjaan Ringeling (CRO), Arie Schoordijk (COO), Piotr Jarmołowicz (software engineer), and Daros Nakai (data analyst), with prior firms and university backgrounds. The site does not claim GenAI or LLM use; it also does not disclose model families, training data, permissions, or validation design. Its performance and award statements are linked to Hedge Fund Journal material and are not independently audited in this record.

Qliq describes an Amsterdam-based, internally funded proprietary firm using fully automated, mid-frequency algorithms on derivatives across U.S. asset classes, with expansion toward other regions. It explicitly says AI is part of day-to-day work across research, systems, and operations, while not using AI as a strategy label. That is evidence of stated operating use, not a disclosure of LLMs, agents, model weights, or AI-generated alpha. The firm says signals are tested and retired when they stop working, but gives no formal maintenance or evaluation specification and names no personnel.

Affor Analytics provides a different route: its reviewed backup-domain page says it started as a systematic trading firm in 2019, had a live strategy with cloud/data infrastructure by 2021, and pivoted in 2023 to AI-driven quantitative solutions for financial firms. It names Koen Ripping, Jasper Kousen, Jelle Willekes, Isaac Braam, and Jonathan Ybema. Because this page is a backup domain with a 2022 footer, the timeline and current status require primary-domain and corporate-record reconciliation. No model, customer, data-rights, permissions, or independent-performance detail is disclosed.

The search rediscovered Machine Capital and Felidae, both already present in the article and ledger, so they were not counted twice. Clarion Capital Fund, Stable Fund, and QVA Fund remain unpromoted leads because the pages did not survive the same verification pass. See the Dutch capture note.

The Dutch De Aandeelhouder podcast episode adds Antaurus AI Tech Fund and portfolio manager Marc Langeveld. Antaurus’s first-party fund page describes a global long/short technology strategy spanning AI, cloud infrastructure, fintech, medtech, robotics, gaming, and semiconductors. Its November 2025 launch article frames the process around fundamental analysis, data-driven insights, and a macroeconomic framework, while a 2026 participant presentation reports the fund live from January 2026, with €40 million in assets and 4.5% net return through April 2026. Those figures are first-party statements, not independently audited here. Langeveld’s public biography includes Robeco, BNP, Van Lanschot Kempen, Barclays, Bankhaus Metzler, Petercam, Antaurus, and Econopolis. The surfaces support a Dutch thematic/product and personnel route, but do not establish an internal AI model, LLM, agent, training corpus, data rights, model permissions, or AI-attributed performance. See the Dutch Antaurus capture note.

September 3, 2026 — Italian Eurizon machine-learning overlay

An Italian-language Eurizon Fund II prospectus adds a regulated-product disclosure that is more precise than a generic “AI fund” label. The product Eurizon Fund II — Q-Multiasset ML Enhanced was formerly Epsilon Fund — Q-Multiasset ML Enhanced; Eurizon’s notice to investors confirms the March 1, 2025 Epsilon SGR merger into Eurizon Capital SGR and the continuation of the strategies.

The prospectus describes a strategic portfolio using macro, market, and fundamental analysis and a tactical long/short overlay using proprietary machine-learning models and AI-generated market signals. It states that the strategic portfolio should generate at least 85% of expected total volatility, while the tactical portfolio should generate no more than 15%. It also states that the manager makes all portfolio decisions and does not rely on AI-based automated trading strategies.

This gives the public record a concrete human-control and risk-budget boundary: ML signals may inform a limited tactical overlay, while final portfolio authority remains with the manager. The prospectus does not disclose model families, features, training data, retraining, research personnel, validation splits, turnover, or AI-attributed performance. See the capture note.

September 3, 2026 — Korean and Japanese title-blind routes

The regional pass added two operating-manager routes and one high-claim watchlist route. Ridge Asset Management’s Korean strategy page describes market, fundamental, and alternative-data collection; quantitative model design and historical robustness checks; automated execution; and real-time monitoring of VaR, volatility, slippage, and risk limits. Its organization page separates quantitative research and automated trading from solution development, FEP/DMA systems, monitoring, compliance, and risk. The firm does not make an LLM or GenAI claim on these first-party pages.

A public Ridge AI/ML hiring listing, marked closed when checked, describes ML/DL analysis of time-series, order-book, tick, and news-text data; LLM/agent-assisted collection and feature engineering; backtesting; model-drift detection; retraining; and monitoring. It lists LangChain, LangGraph, CrewAI, local-model fine-tuning, PyTorch or TensorFlow, time-series databases, orchestration, MLOps, and Grafana. This is specific hiring intent, not evidence that the role was filled or that the listed stack is deployed. The posting does not disclose a model, data license, evaluation split, agent permissions, or live performance. See the capture note.

LINE Investment Technologies provides a separate Japan/Korea systematic-manager route. Its public site describes a global macro program trading more than 100 liquid futures, with strategy categories spanning macro, technical, cross-asset, AI, and intraday methods. It also describes a real-time risk system monitoring more than 80 risk metrics, and publicly lists Soyeon Song (CEO), Hyoungjoo Lee (Portfolio Manager), Seung Min Lee (Head of Risk), Jeffrey Kang (Head of Research), and Jeongae Han (Head of Engineering), along with prior employers and university credentials. The site says the firm has more than 20 employees across Japan and Korea and that more than 70% hold doctoral or master’s degrees. These are displayed company biographies and strategy claims; they do not establish model ownership, training data, deployment permissions, or AI-attributed returns. The displayed Japanese commodity-trading-adviser permit is separately captured in the source note.

TRAI is retained as a high-interest but unresolved public claim. Its technology page names Graphon Mean-Field Games, Kalman/EnKF state-space models, Bayesian updating, Gaussian-mixture risk classification, a meta-learner, metaheuristic allocation, anomaly detection, and stochastic execution controls. Its FAQ describes an intended end-to-end AI fund lifecycle, including execution without human intervention, while its site displays incomplete or unindependently verified performance figures. Those claims are not treated as established deployment or results. The page does not provide model cards, source code, training data, independent evaluation, legal vehicle verification for the described activity, order logs, or a reproducible performance record. See the capture note.

September 3, 2026 — Hong Kong Value Partners internal-AI workflow route

The bilingual Value Partners final-results announcement filed through HKEX adds a listed-asset-manager comparator that was not found in the earlier English-only sweep. Value Partners says it is adopting AI progressively across the organization and has integrated AI into the research workflow through deep-research tools intended to increase stock-research output efficiency. It also describes earnings-call scheduling and attendance bots, AI marketing, realistic AI digital-human videos for investor education, and an internal fund-document knowledge base searchable in natural language. The traditional-Chinese filing confirms the same categories and preserves the original-language route.

This is evidence of stated internal workflow adoption at a Hong Kong asset manager, not evidence of an autonomous investment system. The filing does not name model providers, versions, retrieval or training data, data rights, agent permissions, evaluation results, production endpoints, or AI-attributed returns. It also does not identify a named AI owner. The route is therefore kept separate from hedge-fund deployment claims and from evidence about any specific fund’s trading authority. See the capture note.

September 3, 2026 — Chinese JF SmartInvest agent and research-platform route

The Chinese-language JF SmartInvest annual-results announcement adds an adjacent listed fintech and investment-advisory platform rather than a hedge fund. Its filing names JF Robo-Advisor, FinSphere Agent, and FinSphere Report, and describes AI use across investment research, investor education, compliance, and customer service. It says FinSphere Agent Large Model Assistant V3.0 includes tool invocation, user-memory construction, chain-of-thought decision reasoning, and multimodal text-and-image responses. The filing reports approximately 664,000 customers served, 22.58 million cumulative services, and more than 159.2 billion large-model tokens during the reporting period; these are company-reported usage figures, not independent measurements.

The filing also describes stock-diagnosis agent 4.0, an AI assistant connected to theme, value, and quantitative-investing modules, AI compliance and content inspection roles, and a shift from a single general Q&A model toward a general-foundation plus scenario-intelligence architecture. It names a technology subsidiary, Jiufang Zhiqing, strategic partnerships with Suntime, Tencent Cloud, and Nonconvex, approximately RMB356 million of R&D spending, 624 R&D personnel, and five AI-related papers covering dialogue systems, large language models, financial-information retrieval, and stock analysis.

This is public evidence of a substantial AI product and research platform, not evidence that its agents autonomously manage a hedge-fund portfolio. The filing does not disclose model weights, provider contracts, training or retrieval corpora, data rights, evaluation fixtures, tool permissions, human-approval rules, production incident records, or independently attributed investment outcomes. See the capture note.

September 3, 2026 — Chinese JF SmartInvest research and model-implementation detail

The Chinese-language JF SmartInvest AI-team page adds a more specific public research surface. It names the AI center’s FinSphere Agent, FinSphere model family, FinSphere Image, FinSphere Video, FinSphere Retrieval, and FinSphere Report, and describes collaborations with Huawei, iFlytek, Johns Hopkins, Columbia, Fudan, and Hong Kong University of Science and Technology. The page also names Wang Bing as the AI-team contact for the interview. These are first-party descriptions of an AI program, not an independent audit of current production ownership or model access.

The open FinSphere paper links JF SmartInvest to a five-author research group spanning JF SmartInvest, Columbia, Shanghai University of Finance and Economics, and Johns Hopkins. It describes AnalyScore, a 100-point human rubric; Stocksis, a 5,000-pair expert-refined training set; and a real-time stock-analysis agent that invokes specialized quantitative tools before synthesis. The paper says ten senior analysts iteratively reviewed the training examples and reports full-parameter tuning of Qwen2-72B on 16 NVIDIA A100 GPUs with a 32K context window, learning rate 1e-5, batch size 16, and two epochs. It evaluates 100 queries using human expert scoring. The ACM ICAIF’25 programme independently confirms the FinSearch title, the JF SmartInvest affiliations, and the Johns Hopkins affiliation for Yiqing Shen.

This is unusually concrete evidence about model-building workflow, evaluation design, and academic personnel in a Chinese financial-technology platform. It still does not establish live hedge-fund portfolio authority, training-data rights, released benchmark data, production error rates, or independently verified investment performance. The comparison claims in the paper and the 1,500-question FinSearchBench-24 description remain author/company claims. The expanded evidence is preserved in the capture note.

September 3, 2026 — Chinese quant-agent operating model and control language

Two additional Chinese-language routes add operating and governance detail without proving a specific production system. A China Securities News survey quotes NianKong Technology general manager Wang Li on AI-assisted IT and factor mining, domestic-model research, and NVIDIA, Alibaba, and Huawei GPU capacity. It quotes a Mingshi representative describing AI across data processing, factor mining, strategy research, prediction, execution, risk assistance, text/sentiment, microstructure, order timing, and transaction-cost control. The article also reports a 90–180-day versus seven-day research-cycle claim from an unnamed quant founder through a third-party researcher; the firm and measurement protocol are not identified, so the comparison is not assigned to any manager.

A July 2026 Lujiazui financial-salon transcript names participants from Xihiggs Investment, the University of Science and Technology of China, Daguan Data, and Huawei Securities. The transcript describes an agent layer for converting unstructured information into backtestable signals, generating factors and backtest code, and checking out-of-sample stability. It lays out four stages—efficiency tools, researcher copilots, constrained execution, and highly autonomous execution—and associates the constrained stage with allowlists, risk budgets, human circuit breakers, and complete logs. It also states that agent-generated signals need at least six months of live validation, including backtest/live consistency and traceability, before receiving production trust. These are panel-described operating and control ideas, not evidence that every named participant has implemented every stage.

The new material is valuable because it gives the research queue concrete control terms—live shadowing, out-of-sample stability, factor-code generation, risk budgets, circuit breakers, and auditability—while preserving the boundary between an industry proposal and a verified firm deployment. The sources do not disclose model weights, training data, data rights, permissions, production logs, portfolio authority, or AI-attributed returns. See the capture note.

The May 2026 Chinese Fund News forum report adds a named-personnel and organization timeline. It identifies Lingjun founder/CIO Ma Zhiyu, Mingshi founder Yuan Yu, Mengxi chairman Li Xiang, Umeili chairman He Jinlong, and LuXiang executive director Su Xu. Yuan Yu is quoted describing a Mingshi AI laboratory formed in 2021 for machine learning, deep learning, and reinforcement-learning research, followed by dedicated supercomputing hardware in 2022. Ma Zhiyu is quoted describing large-model use for code generation, logic implementation, strategy testing, and extracting information from research reports, news, and financial reports, including multimodal inputs. These are dated forum and interview statements; no model weights, training-data rights, evaluation splits, agent permissions, or AI-attributed returns are disclosed. The new record is retained in the Chinese quant capture note.

September 3, 2026 — Chinese agent-factory and financial-agent benchmark routes

An additional Chinese Securities Journal report, also available through a Sina Finance mirror, names Wei Mingsan of DeepWin / 蝶威量化 and describes a firm-hosted quantitative-research agent workflow. The report says DeepWin began building a purpose-specific framework from the bottom up in 2025, with 46 operator libraries and more than 6,400 underlying features. The described loop runs from broker-report factor extraction and factor optimization through code generation, backtesting, performance evaluation, and factor-library admission, with analysis, research, coding, evaluation, and fund-manager review roles. It also reports a 7-by-24 operating claim. These are speaker- and media-reported details; the public material does not provide source code, agent logs, model inventory, data rights, or an independent deployment audit.

A separate Cailian Press report says a Shanghai Jiao Tong University team led by the Scalable Computing Institute, together with Caiyue Xingchen, StepStar, and Ruiyuan Fund, released a financial agent evaluation framework and practice platform at WAIC 2026, alongside an 84-page position paper titled Rethinking Financial AI Benchmarks. The framework is described around capital–asset matching rather than research-report similarity. Its six dimensions are return, risk, horizon, liquidity, return structure, and constraints; its three capabilities concern understanding funding requirements, analyzing assets through economic substance, and issuing auditable match/mismatch/insufficient-information decisions. The report says refusal and “not a match” can be valid outputs, and describes task design, asset coverage, scoring, human/machine review, and continuous monitoring.

This is an evaluation-infrastructure route, not evidence that a named hedge fund has adopted the framework. Neither source discloses a reproducible public dataset, code repository, versioned scoring manifest, model list, production permissions, autonomous portfolio authority, or AI-attributed returns. See the capture note.

The related Chinese-language Cailian report and the public FinResearchBench II preprint add the benchmark’s research and personnel trail. The WAICA 2026 accepted-paper list places Beidi Luan and Rui Sun at Stepfun, Sinuo Wang at the University of Adelaide, Yan Gu and Chao Li at FinStep, Zhenliang Xiong at Shanghai Jiao Tong University, and Jing Li and Zuo Bai at Stepfun and FinStep. The paper reports 104 real-world financial queries, 1,040 generated reports, and 14,450 query-specific candidate rubrics. After a human-versus-LLM judge check and two filters, it reports 3,687 consistency-passed rubrics and 2,600 final consensus-derived rubrics, with 98.67% label-level agreement on jointly unanimous sampled items. Those are reproducible research claims only to the extent that the paper’s data and code become available; they are not evidence of live trading or portfolio authority.

The Cailian report additionally says the rubric pipeline was embedded into Caiyue Xingchen’s AI小财神 product to guide report planning, source tracing, reasoning, coverage, visualization, and reinforcement-learning iteration. That product-use statement is company/media evidence. It should be kept distinct from the arXiv paper’s benchmark methodology and from any claim about investment outcomes.

September 3, 2026 — German Quoniam podcast gives a model-and-governance route

The German-language Generation Data episode, published August 25, identifies Carsten Rother as Quoniam’s Co-Head of Research Forecasts and publishes a timestamped transcript. Rother describes ML as an additional component beside a linear factor model, aimed at nonlinear relationships and “tipping points” such as changing leverage/value behavior (03:50–04:50). He describes training and validation samples, rolling historical checks, and economic-expectation gates across long histories (06:07–10:16). He names boosting or a neural network—not an LLM—for this component (08:48–09:07).

Rother also says the European model’s ML component carries a 5–10% weight and delivered roughly 3–4% of “Mehrwert” over eight years (11:41–12:20). That is a speaker/company-reported contribution claim, not an independently audited return attribution. The episode describes decomposition into Value, Quality, Momentum, and ML contributions down to the stock level, with the ML component able to reduce a position’s weight when it identifies a concerning pattern (14:04–15:43). It explicitly retains portfolio-manager judgment for risk, fiduciary care, and geopolitical crises, and says the system should not independently manage portfolios or make investment decisions (16:21–18:49).

The related Episode 6 with Quoniam co-founder Thomas Kieselstein adds a German-language discussion of setup cost, training, control obligations, provider dependence, and the choice between hosted and self-controlled models. Together these episodes provide public model and governance vocabulary, but do not disclose weights, training data, vendor contracts, current production configuration, permissions, or independently verified performance. See the capture note.

September 3, 2026 — Japanese FOLIO allocation and Mizuho reporting-AI routes

The Japanese search also found a current SMBC DS–FOLIO AI Multi-Asset Fund document that separates AI prediction from financial-engineering allocation. The May 20, 2026 document says FOLIO uses selected leading market data—prices, rates, foreign exchange, and commodity prices—and excludes economic indicators, news, and social-media data from this model. It describes one model for roughly one-month returns across seven asset classes and another for detecting when an extraordinary rebalance may be appropriate. AI output is used as a relative return ordering, while a financial-engineering optimizer applies risk, allocation-cap, turnover, and diversification constraints. The fund’s public material describes a hybrid process and notes that human intervention may be required. It does not disclose model families, weights, training-data rights, or independently measured AI-attributed returns.

A separate Japanese Mizuho Trust Bank × Minkabu case study, dated May 12, 2026, describes the formally launched “Robot Report AI” for investment-trust reports and market commentary. The case study names Mizuho participants Ryo Matsuzaki and Katsuyuki Kobayashi and Minkabu participants Nobuya Saito and Akio Tajima. It describes a workflow grounded in Minkabu’s verified market and news data, dedicated programs for numerical calculation, human-provided definitions of broad market phases, and source links attached to generated text for fact-checking. It says the proof of concept was tested with reporting staff before the February 2026 launch announcement. Tajima’s profile identifies Amazon Bedrock agent experimentation and SaaS operation as part of his work. This is evidence about a reporting workflow and data partnership, not about FOLIO’s allocation models or a hedge-fund portfolio.

See the Japanese capture note for the separate evidence boundaries.

September 3, 2026 — Portuguese, Spanish, and Russian title-blind routes

The Portuguese XP Investimentos report adds a Bayes Capital / AZ Quest route that was not visible in the English search. It describes a proprietary indicator library covering fundamental, trend, market-microstructure, macro, and risk signals, and reports that Bayes uses independent AI agents from different providers, including internal solutions, to produce separate code implementations. A further AI compares the outputs, maps disagreements, evaluates consistency, and assigns confidence before a manager reviews the consolidated result. The report also says Bayes has processed more than 20,000 earnings calls into embeddings since 2021 to track tone, recurring themes, and narrative changes. A proprietary transformer and an allocation system combining risk, liquidity, and economic regimes are described as development projects. These are publisher-reported claims; the source does not disclose model weights, provider contracts, embedding model, call-corpus rights, evaluation splits, agent permissions, or independent performance attribution.

The Spanish Global Gradient fund page describes machine-learning strategies using macroeconomic and market data, company fundamentals, and technical indicators, with separate algorithmic components for medium-term trend, short-term trend, and active risk management. It states a working universe of approximately 3,000 companies and 2,000 ETFs, and describes risk-allocation and liquidity controls. José Suárez-Lledó is identified as adviser, with Andbank Wealth Management named as manager. This is first-party product evidence, not a public model card or independent performance attribution.

The Russian HSE AI in Mathematical Finance Lab page adds an academic route focused on multi-agent systems, reinforcement learning for market making and asset hedging, and generative models. It names Yaroslav Lyulko as laboratory head, Mikhail Zhitlukhin as scientific supervisor, and Luiza Koychueva as manager, and lists research on evolutionary and agent-based finance and algorithmic-market dynamics. This is an academic research and talent signal, not evidence of a hedge-fund partnership, production system, or live trading authority. See the capture note.

September 3, 2026 — Chinese IFQuant and QP Alpha hiring routes

The Chinese job-board pass added an IFQuant / 亦赋量化 route through a Zhejiang University career posting. The posting seeks a deep-learning quantitative-research specialist to combine factors into equity and derivatives models, explore sequence modelling, multi-objective methods, graph neural networks, and reinforcement learning, and improve prediction generalization. It also mentions model quantization, GPU-operator development, training/inference speed, and live-strategy metrics, alongside PyTorch/TensorFlow and research-publication experience. The page describes IFQuant as an AMAC-registered private manager, but the role is hiring intent; it does not establish a filled position, production model, data rights, or AI-attributed performance.

The same search found a separate Peking University Financial Engineering Laboratory recruitment page that describes QP Alpha / 量派投资. It names founders Sun Lin and Yu Hang and attributes prior experience at Barclays Capital, Two Sigma, Knight Capital, and Tower Research Capital. The page describes index-enhancement, market-neutral, CTA, and quantitative long-only strategy lines, plus Hong Kong Type 4 and Type 9 licensing claims. Its recruiting language covers quantitative strategy research, machine learning, data mining, live signals, low-latency systems, and trading-system support. The page does not disclose a named model, current headcount, live permissions, or independently verified performance.

See the Chinese hiring-route note.

September 3, 2026 — Hindi-facing QuantReign exposes a paper-trading agent stack

The Hindi/Indian title-blind pass found QuantReign’s first-party platform page. It describes five parallel agents for market, portfolio, options, risk, and explanation tasks, running on AWS Bedrock with Anthropic Claude. The page also states that explanations are grounded in a 50,000-chunk retrieval base using FAISS and Amazon Titan Embeddings. Other advertised components include options pricing, Monte Carlo simulation, GARCH and other VaR methods, regime detection, strategy backtesting, and a separate 1,000-agent market-participant simulation.

The same page says QuantReign is launching in paper-trading mode and is an educational and analytical platform rather than a SEBI-registered adviser. It therefore provides a concrete public vocabulary for agent orchestration, retrieval, simulation, and compliance guardrails, but does not establish a hedge-fund deployment, live capital, training-data rights, evaluation results, or investment performance. See the capture note.

September 3, 2026 — FactorEngine adds an academic program-synthesis route

The title-blind academic search found the arXiv preprint FactorEngine: A Program-level Knowledge-Infused Factor Mining Framework for Quantitative Investment. The paper treats a factor as executable code and separates logic revision from parameter optimization, LLM-guided directional search from Bayesian hyperparameter search, and LLM proposals from local numerical computation. Its knowledge-infused bootstrapping pipeline extracts ideas from financial reports, verifies them, generates factor code, and records successful and failed search trajectories for later refinement.

The paper reports backtests on real-world OHLCV data and reports IC, ICIR, Rank IC/ICIR, annualized-return, and Sharpe comparisons against baselines. Those are academic, paper-reported results. The public record does not establish use by a tracked hedge fund, production deployment, proprietary-data rights, model provider, trading permissions, or independently audited performance. The route is useful because it adds concrete search terms—report-to-code extraction, verifier agents, program-level factor mining, Bayesian tuning, and experience memory—for future personnel, GitHub, conference, and hiring searches. See the FactorEngine capture note.

September 3, 2026 — French media route adds an agentic-governance control framework

The French Option Finance report led to the underlying Risk AI Center SSRN paper, which proposes Policy, Engineering, Composition, and Systemic governance layers for agentic AI in finance. It specifically treats agent reward functions as a policy and model-risk issue, proposes statistical drift alerts, and adds market-level monitoring for correlated or crowded agent behavior.

The paper reports that 88% of respondents in an informal LinkedIn survey lacked an operational agentic-AI governance framework, and that its review identified formal governance policies in 24 of 75 large U.S. money managers disclosing AI use in Form ADV. These are study-specific observations, not an audited industry census or a ranking of firms. The route adds control and monitoring vocabulary, but does not establish any tracked manager’s internal policy, model, vendor, training data, permissions, or return attribution. See the capture note.

September 3, 2026 — Polish UniCredit release exposes a manager-controlled Axyon route

The Polish-language UniCredit Poland release describes a planned active-management module using algorithms developed by UniCredit-affiliated Axyon AI. The release says the model would generate buy and sell signals while the portfolio manager decides how much weight to give the recommendations. It names Tomasz Wróblewski as Head of Treasury for UniCredit NV/SA and describes a portfolio service built around more than 70 ETFs and MiFID risk profiling.

The wording describes a planned bank/wealth-management feature, not a confirmed hedge-fund deployment or necessarily a live module. It does not disclose the Axyon model family, training data, signal frequency, contract terms, evaluation, or outcomes. The capture note keeps this product-partnership route separate from Axyon’s academic and infrastructure records.

September 3, 2026 — Dutch AFM survey supplies a sector-level AI baseline

The Dutch regulator AFM’s March 2026 report reports an exploratory questionnaire of 323 Dutch asset-management-sector institutions, distributed in the first half of 2025. It says 170 respondents (53%) used AI or planned to use it within twelve months. Among those 170, reported uses included information retrieval for analysis and decision support (48%), unstructured alternative-data analysis (48%), and research-report writing (45%). Proprietary traders reported uses including trading-algorithm parameter optimization, strategy improvement, and price or market prediction.

The report also provides model and infrastructure vocabulary: NLP was reported by 80% of AI-using respondents; supervised, unsupervised, and reinforcement learning by 80%, 44%, and 29% of ML users respectively; and general-purpose models by 94% of AI-using respondents. Commercial-cloud hosting was reported by 68%, private or dedicated hosting by 9%, and hybrid hosting by 20%.

The control gap is equally relevant to the agent search: 89% of all respondents reported no formal policy covering AI-agent use, 78% did not plan to implement agents or multi-agent systems within twelve months, and 2% reported a dedicated agent policy. These are self-reported, survey-specific results across a broad sector—not a hedge-fund ranking or a firm-by-firm deployment record. See the Dutch AFM capture note.

September 3, 2026 — French Drakai route links systematic credit, earnings-call analysis, and learn-to-rank research

The French-language Drakai Capital podcast episode identifies Guillaume Boulanger as co-founder and CRO and describes a machine-learning-based systematic-credit strategy, a quant team with Polytechnique lineage, and an AI use case involving hundreds of financial- communications conference calls. The episode is historical, dated November 2020, and its publisher description does not disclose model architecture, training data, permissions, or performance.

The firm’s 2020 launch coverage names Samer Comair as founder and former Société Générale proprietary credit trader and Boulanger as director of risk. A 2025 Journal of Financial Data Science paper by Mathis Linger, Thibaut Metz, Khalil Sbai, and Boulanger studies learn-to-rank methods for long-short portfolio construction. The HAL/RePEc record links the authors to Drakai and Linger to the University of Orléans LEO laboratory. These are research and personnel-lineage signals, not proof that the published ranking method is Drakai’s live credit-fund implementation. See the French capture note.

September 3, 2026 — French Bodic archive expands the workflow and governance vocabulary

The Bodic French-language archive adds a cluster of asset-management-specific episodes that were not visible from the single previously indexed Bodic route. The investment-committee episode proposes standardized memos, traceable retrieval sources, formalized red-team contradiction, and a Quality-of-Decision Score to address information overload. The Claude Cowork and Dispatch episode describes persistent project context across local files and communications tools for report synthesis, financial-data extraction, and regulatory monitoring, with privacy and staged deployment constraints.

The OpenClaw browser-agent episode connects browser interaction to sourcing, due diligence, monitoring, and CRM enrichment, while the Portfolio Portal episode emphasizes validated and historized KPI data as the base for benchmarking and AI. A separate cost episode frames infrastructure, integration, governance, and data quality as costs in addition to tokens. These are vendor/editorial descriptions, not evidence of any named hedge fund’s deployment, model permissions, or investment results. See the Bodic capture note.

September 3, 2026 — Spanish title-blind episode on AI fund selection

The Spanish La ruta del dinero episode, also available through an elEconomista YouTube mirror, discusses experiments with general-purpose AI systems for selecting investment funds. Around 05:07, the conversation describes filtering a large fund universe by risk, return, losses, fees, benchmarks, and peer categories. Around 06:01–08:08, it describes different shortlists from repeated prompts, including an institutional share class that would not be accessible to a retail investor.

The speakers then separate fund selection from portfolio construction. Around 09:10, they emphasize specifying beta and sub-beta exposures such as size, value/growth, currency hedging, and sectors. Around 11:51, they raise the possibility that similar AI prompts could converge on similar funds and create concentration or capacity risk, particularly in less-liquid markets.

This is a Spanish-language consumer and industry discussion, not evidence of a named hedge fund’s production system. The automatic captions support timestamped navigation but do not establish model accuracy. The episode does not disclose model weights, training data, data rights, evaluation design, portfolio permissions, or AI-attributed performance. The capture note records the private transcript-retention and evidence boundaries.

September 3, 2026 — Japanese Pictet event links token economics to market concentration

The Japanese Pictet Theatre LIVE episode was recovered through a YouTube recording with a timestamped Japanese transcript. The event discusses AI-market pessimism, technology-equity valuation, debt-funded capital expenditure, data-center depreciation assumptions, and LLM token prices. Around 23:37, the speaker links falling token prices to possible changes in model-company revenue assumptions. Around 26:45, the speaker connects AI-related memory stocks with increased co-movement between the SOX and KOSPI indices.

This is institutional market-outlook commentary, not evidence of Pictet’s internal portfolio model. The transcript supports navigation and paraphrase, but does not disclose model weights, training data, validation, portfolio permissions, or AI-attributed performance. The capture note records the source and evidence boundary.

September 3, 2026 — Chinese title-blind quant-podcast routes

Two additional Chinese-language episodes were absent from the prior ledger. Qishi Formula, with additional Apple Podcasts chapter metadata, identifies the guest only as Max, a Qishi member. Its chapter metadata covers the researcher-to-PM transition, low-correlation portfolio construction, merger-arbitrage backtest blind spots, ML’s interpretability boundary, alternative-data commoditization, and POD versus centralized research organization. It does not identify a fund employer or disclose a live model.

厚雪长波, also available as a YouTube mirror, identifies Kong Xianzheng as General Manager of Noah Fund’s Multi-Asset and International Business department. The recovered creator-caption transcript discusses multi-factor, ML, and neural-network approaches across equities, CTA, bonds, FX, and asset allocation at 10:48. It contrasts high-frequency and low-frequency validation horizons at 26:23, discusses interpretable versus non-interpretable factors at 31:28, and describes human-designed alongside machine-discovered factors at 48:12. These are speaker statements and methodological discussion, not independently verified performance or evidence of an AI production system at Noah Fund.

The Qishi record remains metadata-level; its Apple page supplies chapter metadata but no recovered full transcript. The 厚雪长波 record now has a recovered creator-caption transcript, retained privately. The capture note records the evidence boundary and follow-up queue.

September 3, 2026 — Chinese interview adds an attributed neural-network route

A Chinese published interview with Kong Xianzheng adds a separate model-surface claim. The interview attributes to his quantitative work a preference for end-to-end neural-network construction and describes machine learning as a way to capture nonlinear relationships. It also discusses the choice among multi-factor models, traditional ML, deep learning, factor construction, and end-to-end approaches. The same interview says that the team’s core research direction is deep learning, particularly improving model architectures, and that portfolio managers participate in ML and factor research rather than only consuming model outputs.

Noan’s first-party investment-research page independently confirms a quantitative and systematic multi-asset organization covering index, index enhancement, active quantitative, FOF, asset allocation, duration, convertible-bond, and derivatives strategies. It does not confirm the interview’s architecture claim. The interview does not disclose code, model versions, feature definitions, training data, validation splits, permissions, or independently audited performance. This route is therefore retained as attributed practitioner evidence, not as proof of a live model or a comparative capability claim.

September 3, 2026 — Spanish audio route on general-purpose AI and investing

The Spanish Humanos episode was recovered through a YouTube mirror with a timestamped Spanish transcript. The speakers separate “investing with AI” from investing in AI-related assets and discuss three general-purpose-model failure modes: stale information, hallucination, and sycophantic agreement. Around 11:05, they discuss organizational controls for keeping sensitive context under the user’s control. Around 14:12, they describe the risks of treating a model as an oracle or sole source of truth. Around 25:42, they describe using models for summarization, screening, benchmarking, stress testing, adversarial review, and multi-model auditing.

This is generic Spanish finance and technology media, not evidence of a named manager’s production system. The transcript supports navigation and paraphrase, but does not establish model accuracy, training data, data rights, portfolio authority, or AI-attributed performance. The capture note records the private transcript-retention and evidence boundaries.

September 3, 2026 — Singapore routes expose an AI-native protocol and a named AI team

The search added two Singapore-facing routes that should not be conflated with regulated hedge-fund disclosures. Calais Markets’ about page describes a Singapore-headquartered quantitative investment fund with stated United States and China offices, multi-asset and options strategies, and AI-powered modeling. It names Lily Yan as CEO and co-founder, Max Guo as Chief Risk Officer with an NUS economics doctorate and prior quantitative-strategy experience at an unnamed hedge fund, and Richard Ding as CMO with an AI doctorate and prior Alibaba and PwC experience. Ding’s public description explicitly mentions deep learning, large language models, agent systems, and finance. Its careers page adds quant-researcher and quant-developer openings in Singapore and Xi’an and a young-talent program that mentions quantitative research, live trading, and strategy development. Its services page adds digital-asset liquidity aggregation, algorithmic execution, MPC custody, and audit-trail language. These remain first-party claims: the pages do not disclose model names, training data, permissions, fund vehicles, regulatory status, or independently verified performance.

A dated Hubbis report also reports a Calais operational partnership involving UBS Asset Management’s tokenized uMINT money-market fund, ByCustody, Bybit, and DigiFT. The reported structure used uMINT as off-exchange settlement collateral during live crypto trading, while the asset remained in custody and potentially continued to earn money-market yield. UBS documents uMINT and Bybit lists it among supported real-world-asset collateral. The public record supports the partnership route and infrastructure components, but not transaction size, returns, contemporaneous haircuts, or general adoption. This is an execution, custody, and capital-efficiency signal—not evidence about Calais’s AI models.

Secondary coverage leaves the stress terms unresolved: collateral haircuts, valuation, redemption timing, liquidation rights, and legal control across the custody, distributor, and exchange layers. A reported uMINT asset-size and holder-count snapshot is retained as secondary, time-stamped context—not as a measure of Calais’s balance sheet or exposure.

OpenForage’s documentation describes a different category: a closed-beta synthetic-dollar protocol in which AI agents search a library, submit trading signals, pass stated in-sample and out-of-sample checks, and contribute to crypto and prediction-market systematic exposures. Its strategy page discusses agent search, reinforcement learning, genetic algorithms, and large ensembles of signals. The public GitHub library adds a Python registration, local-evaluation, submission, and callback surface. Its terms expressly state that the company is not regulated or licensed under Singapore’s Securities and Futures Act or equivalent regimes. This makes OpenForage useful evidence of a public agent-economy design, but not evidence of a regulated hedge fund, live capital, or validated investment performance.

The OpenForage founder’s public technical essay adds an attributed but unnamed former-hedge-fund research route. It says a general language model trained to predict stock returns from news initially appeared predictive because of look-ahead bias in pretraining; the described correction paired news articles with realized future returns and fine-tuned the model toward return-prediction error. A separate alpha-mining essay frames the loop as candidate generation, evaluation, rejection, retention, and repeat, with out-of-sample correlation and portfolio-management constraints. These are founder-reported methods and design claims. The former employer, data, splits, model, and performance are not independently identified.

See the Singapore capture note for the entity, regulatory, and evidence boundaries.

September 3, 2026 — AgonAlpha exposes an auditable alpha-research loop

An August 2026 AgonAlpha paper and its public repository add a distinct agentic-research route. The system searches over evidence-bearing research artifacts—not only formulas—and separates a proposer from a fresh- context reviewer that can re-run an evaluation and veto unsupported evidence. Its scheduler allocates pending-aware search budget across candidate lineages using Monte Carlo tree search. The paper reports five WorldQuant BRAIN users, 60 submissions, and 17 platform-assigned SPECTACULAR grades; the supplement reports a highest observed Fitness of 9.50 and full-universe Sharpe of 3.48. Those are author-reported outputs from an external evaluation platform, not an independent audit or evidence of a hedge fund’s live system.

The repository makes the implementation boundary unusually explicit: the orchestration layer is public, while data, credentials, evaluator integrations, generated candidates, and evaluation results remain local. A separate public S&P 500 reproduction repository was found in the follow-up search, but its partial CogAlpha reproduction and claims are kept separate from AgonAlpha. Neither route establishes a named manager’s model inventory, proprietary data rights, live capital, execution permissions, or net performance. See the capture note.

September 3, 2026 — Brazil and quant-technology routes add implementation detail

The July 17 Sophron episode with Luciano Boudjoukian França identifies him as a founding partner, CIO, and portfolio manager at Avantgarde Asset Management in São Paulo. The publisher description places the discussion around the gap between a Brazilian factor backtest and a live book, including liquidity, implementation costs, point-in-time data, survivorship, corporate actions, capacity, and model override; it marks an AI-research-process segment at approximately 53:56. Avantgarde’s first-party site describes a data-driven, systematic factor process that cleans company data, constructs factors, ranks securities, and manages risk and implementation costs. The episode’s AI segment remains a recovery lead pending audio review; the sources do not establish a named model, production deployment, or AI-attributed result. See the capture note.

The same Avantgarde team page names André Gouldbaum Lichtenstein as the partner responsible for the AInvest Capital Inteligência Artificial Global FIA, a thematic fund the page describes as focused on semiconductors, memory, energy, and data centers. It separately lists Brendo Henrique in Quantitative Research and describes his UFAL economics and informatics training and research in mathematical, econometric, and statistical methods. Avantgarde’s academic-engagement page also exposes partnerships with Brazilian finance and university groups. The AInvest site supplies an additional product and media surface led by Lichtenstein. These pages establish a named AI-themed fund and personnel structure; they do not disclose model architecture, training data, internal research agents, or AI-attributed performance.

The May 29 Sophron episode with Yves Hilpisch adds a quant-workforce and tooling route rather than a hedge-fund disclosure. The publisher discusses agentic coding, the backtest-to-live last mile, and the distinction between outsourcing code and outsourcing understanding. His first-party pages corroborate leadership of The Python Quants and The AI Machine, mathematical-finance training, AI-in- finance publications, and open-source quant-finance tooling. This does not establish adoption by any tracked manager. A separate Alphaparty episode with Rob de Rozario is retained as an adjacent digital-asset route; its AI reference is publisher metadata, not evidence of a hedge-fund model or authority.

The Portuguese-language follow-up also surfaced iVi Technologies’ first-party team and technology pages. iVi publicly describes an internally developed, fully systematic Brazilian-equity process using AI, data science, mathematical models, backtesting, portfolio and asset-selection optimization, and automated decisions. Its page names Lendel Vaz Lucas in CEO/Risk/Compliance, Max Schoppen as COO, Antonio Bertuccio in portfolio administration, Guilherme Brasil as Tech Advisor, Leonardo Miranda as Quantitative Developer, Enzo Passarini as Data Analyst, and Kai Schoppen and Ionan Fernandes as partners. The site states that it collects more than 80,000 data points daily. Those are company-reported descriptions; the public pages do not disclose model families, feature definitions, training data, agent permissions, fund-level deployment, or independently audited AI-attributed performance. See the regional follow-up note.

A dated Exame profile adds a historical product route: iVi described a proposed quantitative crypto ETF, an AI/machine-learning decision process without analyst-defined weights, monthly rebalancing, and an intended index relationship with a Miami exchange. This is a 2023 media account of company statements. It does not establish that the proposed product launched, that the partnership continued, or that the reported historical outcomes were independently audited.

A separate Abrapp Portuguese event report from May 29, 2025 names Giant Steps’ Glauber Guarinello, DataV’s Odemir Depieri Jr, and Harumi’s Miriam Koga. The report attributes to the panel a workflow spanning clean point-in-time backtests, signal research, transaction costs, risk constraints, portfolio optimization, data-quality barriers, and AI for operations research and code generation. This is attributed event-report evidence, not a technical disclosure from any one firm, and sponsorship is not treated as adoption evidence.

The same Portuguese search also found a CEIA/UFG project page describing research and evaluation of machine-learning methods for economic indicators and asset selection, alongside an effort to improve an unnamed company’s data-capture and processing framework. Because the company, researchers, and agreement are not identified, this remains an academic discovery lead rather than evidence of a hedge-fund relationship.

A Spanish title-blind search also recovered Value Investing FM episode 168, an April 2021 interview identifying José Iván García as CEO of Zona Value and portfolio manager of Kau Markets EAFI. The publisher description covers quantitative filters, data sources, decision-tree selection of an AI, and model bias. The audio was not recovered, so this remains publisher-metadata evidence; it does not establish a current production system, named model, or audited result. See the Spanish capture note.

The Spanish regional pass also added ConoSur Asset Management, an Argentine manager whose first-party pages describe quantitative models, algorithms, AI, adaptive algorithms, backtesting, and automated back-office processes across six named funds. Public company and personnel surfaces name Pablo Berenbaum, Juan Martín Yanzon, and Mauro Ezequiel Fiorenzano in early, trading, and management roles; one public profile describes order-execution bots and a duration optimizer. These are company and profile claims. They do not identify model architecture, training data, current permissions, or independently audited AI-attributed performance.

The same search found Inmatek, a Mexican algorithmic MAM and MT5-robot route. Its site describes 46 internally developed robots and an investor-account structure, but does not establish a regulated hedge-fund vehicle, AI/ML model, named operator, or independent results. It is retained as an algorithmic-infrastructure discovery record, not classified as an AI hedge fund. Numerical return and win-rate statements on the site are not promoted because they are unverified marketing claims.

An AzValor podcast listing with Director of Innovation Carlos Camps, together with a timestamped public transcript mirror, adds a concrete governance route. The transcript describes filtering AI output, using public documents for first-pass summaries, keeping client and proprietary accounting data outside cloud models, testing AI and traditional reconciliation systems in parallel, and retaining human investment authority. The transcript is a third-party capture pending audio review; it does not disclose providers, prompts, retrieval design, production completion, or investment performance.

See the Latin-American and Spanish route note.

September 3, 2026 — Chinese-language roundtable and governance routes

The China Fund News roundtable report, published after a May 28 Shenzhen forum, names Lingjun’s Ma Zhiyu, Mingshi’s Yuan Yu, Mengxi’s Li Xiang, and Umeili’s He Jinlong. The report attributes to the participants discussion of machine learning, deep learning, an Mingshi AI laboratory established in 2021, dedicated supercomputing hardware described as built in 2022, LLM-assisted code and strategy testing, and use of reports, news, filings, and other unstructured data. This is Chinese financial-media reporting of a public roundtable; it does not establish that every statement describes a current live system, or disclose models, data rights, permissions, or performance.

A separate International Finance News interview identifies Lingjun chairwoman Cai Meijie and reports her interest in applying quantitative-technology experience to AI applications outside investing. It also describes a post-2024 governance change in which she took final overall management responsibility after a previous split between business and investment-technology leadership. This adds leadership and governance context, not evidence of a trading-model change or current AI deployment.

See the Chinese-language route note.

September 3, 2026 — Japanese practitioner and talent-pipeline routes

The Japanese Technical Analysts Association archive lists a seminar on AI in quantitative management with Kei Nakagawa and Mitsuyoshi Imamura of Nikko Global Wrap. The page describes pattern recognition over historical prices and investor-behavior patterns and connects the topic to a deep-learning chaotic-time-series paper. The archive year is not clear in the reviewed text, and the page does not disclose a current model, data rights, portfolio authority, or performance.

The Matsuo Institute AI Quant Trading course adds a 2026 Japanese talent route covering machine-learning return prediction, large-language-model applications, and agents for coding and strategy exploration. Its rules explicitly state that the competition is educational and is not used for investment decisions or asset management. This is a recruiting and skills signal, not evidence of a live fund or production model.

See the Japanese route note.

September 3, 2026 — Korean former-personnel and adjacent fintech route

The Korean AIM profile and Edaily interview identify Jihae Jenna Lee as AIM’s founder and display a prior Citigroup and Acadian investment-career path. AIM’s public product pages describe an AI wealth-management algorithm, named “Esther,” and a customer-account model in which assets remain in the customer’s brokerage account. The displayed biography also claims NLP/machine-learning adoption and portfolio-review automation experience.

This is a former-personnel and adjacent-fintech route, not evidence of Acadian’s current system. The public sources do not establish Lee’s exact employment dates, that AIM’s architecture came from Acadian, or Esther’s model, data, permissions, evaluation, or performance. Company-reported outcome claims are excluded. See the Korean AIM route note.

September 3, 2026 — Japanese practitioner-built assistant and quant research routes

The Japanese SMBC DS Asset Management interview identifies Takehide Hirose as head of the Investment Development Group and describes AIR, an internally developed assistant for investment research. The interview says AIR reads news, earnings materials, and EDINET/TDnet disclosures; operates inside a closed network; routes work across Azure, AWS, and Google Cloud models; and includes specialized agents for long-report drafting, fact checking, and news/price analysis. The stated purpose is to compress information gathering and hypothesis preparation so human investment staff can spend more time on ideas and risk scenarios. The interview explicitly retains human forward-looking judgment and does not disclose model versions, retrieval design, training data, entitlements, evaluation fixtures, or portfolio/order authority.

Nissay Asset Management’s Japanese 2025 Quant Topics archive adds a named research-team route. Its BERT article names Takaaki Yoshino, Yoshiaki Kimura, and Megumi Tsukamoto and describes contextual BERT representations and sentiment scoring for analyst reports. A cyclical-equity series describes machine-learning estimates based on economic-cycle similarity and a semiconductor/SOX switching example. A deep-learning volatility series and its MCMC follow-up describe a Statistical Recurrent Stochastic Volatility model and Bayesian parameter estimation. These are detailed first-party method publications and personnel disclosures, but they do not establish live deployment, model ownership for a particular fund, or independently attributed performance.

See the Japanese SMBC/Nissay capture note.

September 3, 2026 — German DekaBank and ACATIS routes

The German Sparkasse interview, dated August 28, 2024, identifies Prof. Dr. Dominik Wolff and Dr. Michael Wegener as leaders of DekaBank’s quantitative teams. They describe NLP applied to ad-hoc disclosures, earnings calls, and news, and say Deka intensified AI research from 2017 for equity and later bond forecasts. The interview describes a daily short-horizon forecast of market segments, but states that Deka does not use fully automated trading signals: a portfolio manager reviews and translates the AI proposal before it reaches the market. Model family, training corpus, validation, permissions, and performance are not disclosed.

The German FINANZDIALOG episode listing dates an episode to November 24, 2021 and presents Kevin Endler as ACATIS’s quantitative-management lead discussing AI in investment and whether it can replace a fund manager. The audio was not recovered, so the episode remains metadata-level evidence. ACATIS’s current team page displays Endler and Dr. Eric Endress in quantitative-AI roles, while its AI-fund media page links additional interviews and press coverage. These pages establish a named personnel and media route, not a public model inventory or autonomous trading system.

See the German DekaBank/ACATIS capture note.

The newer German Bulle & Mensch interview with ACATIS founder Hendrik Leber, published August 16, 2026, adds a current founder-level disclosure. Its publisher describes an ACATIS-developed AI that participates in selecting roughly 50 stocks from hundreds, and gives dedicated chapters for the AI model and the firm’s skepticism toward backtests. The page also attributes to Leber the claim that ACATIS introduced an early AI-assisted fund in 2016. The audio was recovered and locally transcribed for timestamped review; the ASR remains a navigation aid rather than a reviewed verbatim transcript. These are speaker/publisher claims; they do not disclose architecture, data rights, validation, permissions, order authority, or independently verified AI-attributed performance. See the capture note.

September 3, 2026 — Swedish SEB AI-fund podcast route

The Swedish Förvaltarpodden episode page and its English companion listing date a 13-minute episode to July 8, 2026 and identify Pavel Lupandin as one of the SEB AI fund’s portfolio managers. The discussion concerns the fund’s origin, the definition of an AI company, surprises after three years, and the opportunities and challenges of the theme. SEB’s first-party fund page names Lupandin, Johan Söderström, and Alexander Winberg and describes a human-selected universe spanning AI developers, infrastructure suppliers, and companies applying AI outside technology. The audio was not recovered, so this is a personnel/product and podcast-discovery route, not evidence of a named model, proprietary signal, agent, or autonomous trading system.

See the Swedish SEB capture note.

September 3, 2026 — Arabic Earthian AI-native fund watchlist route

Earthian’s Arabic AI-native hedge-fund page, updated September 1, 2026, describes an AI-native fund concept and names Project Alpha-Index as its foundational research initiative. The page lists model labels including Technology Tenet-0, Geopolitics Axiom-0, Lucid Climate-0, NatCat Lighthouse-0, Cybersecurity Ichnos-0, and Policy Evergreen-0, and claims that they would combine climate, geopolitical, technology, macroeconomic, and asset-level risk reasoning in dynamic portfolio construction. The related Alpha-Index page describes continuous optimization of risk exposure, liquidity, and asset selection.

Earthian’s governance page displays Shayan Shokri as founder and CEO, Cristina Arango and Yamina Damil on a board surface, and several research/education associations. Its LinkedIn page also describes an internal AI research fund. These are first-party positioning and biography claims. The reviewed pages do not provide a regulator record for the described hedge-fund vehicle, model cards, weights, training data, source code, order logs, portfolio statements, or independently verified performance. The route is therefore a watchlist and personnel-discovery lead, not a verified manager classification.

See the Arabic Earthian capture note.

September 3, 2026 — Turkish Magnus portfolio-platform and fund-partnership route

Turkish title-blind searches surfaced Magnus, a public AI portfolio-optimization platform, rather than another disclosed hedge-fund model. Magnus describes quant tools for portfolio construction, rebalancing, reporting, and custom simulations. A Turkish iDeal Data product listing adds algorithmic asset allocation, automated monitoring, backtesting, algorithm comparison, and automatic order transmission. The pages do not disclose model families, features, training data, validation, client entitlements, or order controls.

The Anadolu Agency report, dated February 15, 2024, reports a Trive Portföy partnership for the KIB Robotik Teknolojileri Değişken Fonu and attributes to Trive and Magnus executives the use of Magnus’s AI applications and planned machine-learning work. Dr. Esra Ulaşan is identified in industry coverage as Magnus co-founder and CEO, while İskender Ada is identified by Galatasaray University as its deputy general manager for business development.

The current KAP summary confirms KIB as a perpetual Trive Portföy fund with current periodic portfolio disclosures and lists Mehmet Yağız as an assistant portfolio manager/contact. KAP’s general-information page describes its AI-and-robotics investment mandate. These filings establish the fund vehicle and public disclosure surface; they do not identify Magnus’s specific holdings, model, decision rights, or AI-attributed results.

This is vendor and thematic-fund evidence. It does not establish that Magnus made every KIB decision, that the advertised tools were fully deployed, that the fund is a hedge fund, or that any result was caused by AI. See the Turkish Magnus capture note.

September 3, 2026 — Turkish AI venture-fund registry routes

The Turkish KAP registry also exposes Aktif Portföy LİMA Yapay Zeka Girişim Sermayesi Yatırım Fonu, code XXQ, with current reporting entries under Aktif Portföy’s venture-fund umbrella. The SPK issuance list lists additional AI-named venture vehicles, including İstanbul Portföy Akson Yapay Zeka Teknolojileri and İstanbul Portföy XCapital. These are registry and venture-investment routes, not evidence of public-market AI trading.

Public biography material identifies Ahmet Berter Argun with Lima Ventures and AI/fintech startup investing, but does not establish management of XXQ or an AI investment model. See the Turkish AI venture-fund registry note.

September 3, 2026 — Slovak ČSOB / KBC names the QAISAR allocation system

ČSOB Slovakia’s AI-fund page publicly names QAISAR as the AI system proposing portfolio composition for the mixed Optimum Fund Enhanced Intelligence Global Allocation Responsible Investing. The page says QAISAR uses incoming data to predict returns and select asset classes, regions, and sectors; it says decisions are monitored daily and that the system is trained on new data. ČSOB also describes more than 1,200 economic and market indicators and almost 200 mathematical models. Those counts are company-reported product claims, not an independent model-card or benchmark.

ČSOB’s November 28, 2025 explainer adds the public feature vocabulary: macroeconomic and company data, market sentiment, regional and sector trends, historical cross-asset relationships, risk factors, and ESG criteria. It describes a 55% equity / 45% bond neutral allocation that can deviate on QAISAR’s recommendation, with data scientists and investment experts monitoring and developing the system.

KBC’s English newsroom account provides historical context. KBC says a multidisciplinary team began exploring AI-assisted investing in October 2017, the software went live in-house on June 29, 2018, and the firm monitored and optimized the system for more than two years before launching the fund. KBC describes more than 1,000 parameters, more than 100 AI models including machine learning and deep learning, and sentiment inputs from news articles and publication volume. It says KBC experts selected eligible asset classes, regions, sectors, and themes, while the fund manager could ignore or partially follow the model and intervene in exceptional circumstances.

The FSMA fund register and 2025 prospectus separate the legal and contractual surface from the product marketing. The prospectus identifies the sub-fund’s October 20, 2020 incorporation date, delegation of intellectual management to KBC Fund Management Limited in Ireland, permitted asset classes, target allocation, and a long-term expected tracking error of 2%. The prospectus does not name QAISAR or disclose its architecture.

This is a regulated, bank-sponsored mixed-fund route rather than a hedge-fund classification. The reviewed material describes AI and machine learning, not an LLM, generative-AI workflow, agent framework, retrieval system, fine-tuning, or public code. It does not disclose model versions, training or validation procedures, data licenses, current staffing, decision latency, order controls, or AI-attributed performance. See the QAISAR capture note.

September 3, 2026 — QuantBeats adds a Central-European title-blind quant-media route

The QuantBeats series page and its YouTube channel expose nine videos uploaded between November 2024 and June 2026. The series identifies Radovan Vojtko as Quantpedia CEO and Head of Research and describes his prior role as a Tatra Asset Management portfolio manager responsible for more than EUR 300 million across quantitative funds focused on asset allocation, multi-asset CTA/trend-following, market timing, and volatility trading. It identifies Dan Hubscher as founder of Changing Market Strategies and describes his electronic, algorithmic, quantitative-trading, and fintech background. These are public biography claims; the page does not provide audited AUM or a complete employment record.

Several episodes are useful precisely because their titles do not name a hedge fund or AI. The Wayne Ferbert / Alpha DNA episode describes website visits, search trends, and app usage as alternative data, then describes machine learning turning those inputs into trading signals and long, hedged, and equity-market-neutral portfolios (02:00–02:23). The Quantmatix episode describes a big-data and machine-learning “market GPS” for market turning points. The Sid Ghatak episode emphasizes curated, trusted data and the risks of AI-generated or contaminated training content, with a historical data-integration account (03:28–04:21). The Alex Gillula episode links Tocaya Capital Management to high-level signals, market-neutral approaches, and AI in strategy development. The Jiří Mrkva episode describes automated systematic crypto strategies.

The remaining episodes cover rules-based quantitative value investing and bubble detection, including links to LPPLS-related research. They do not add a verified GenAI disclosure, but they widen the guest and vocabulary graph for future searches. These are speaker, publisher, and auto-caption descriptions; they do not establish a live model, proprietary data entitlement, current employment beyond the public episode framing, order authority, or verified performance. The route does not disclose LLM providers, fine-tuning, retrieval, training corpora, agent permissions, or order logs. See the QuantBeats capture note.

September 3, 2026 — Hungarian EverestQuant podcast and academic-lineage route

The Hungarian Economx Bázispont episode adds a person-first and regulated-wrapper route that extends the existing EverestQuant legal-and-careers record. The publisher names Márton Price as investment director and Gábor Fáth as quantitative-analysis and risk-management lead, and frames the episode around quant-fund data, the human role in research and decisions, the physics-to-finance path, and AI in quantitative hedge funds. The same recording has a YouTube mirror, but the public caption pass found no original or automatic caption track, so no audio-level claim is promoted here.

EverestQuant’s first-party site describes a London headquarters and Budapest Quant Research Office and lists systematic macro and commodities, volatility, equities, equity derivatives, and European electricity strategies. The APELSO product page separately identifies a public open-ended fund-of-funds wrapper, its depositary and ISINs, and a policy allowing up to 100% in MPP&E Capital Flagship Algorithmic US Limited, described there as a systematic multi-strategy quantitative vehicle. The displayed approximately USD 4.0m is a date-scoped wrapper figure, not total EverestQuant or underlying-fund AUM. The Hungarian National Bank publication record confirms the January 2026 prospectus/rules-document route; it does not independently validate broader media AUM or “first” claims.

The Eötvös Loránd University profile for Gábor Fáth identifies him as director of ELTE RiskLab, a senior research fellow working across quantitative finance, machine learning, and quantum computing, and a former Morgan Stanley Hungary managing director leading Budapest Fixed Income Strats. It lists recent work on neural-network option pricing and quantum Monte Carlo, including the Deep Weighted Monte Carlo paper. This establishes public academic and personnel lineage; it does not establish that those methods or prior-employer technology are used in EverestQuant’s current strategies. The route does not disclose a GenAI system, model owner, training corpus, data rights, permissions, validation design, or AI-attributed performance. See the Hungarian capture note.

September 3, 2026 — Polish Third Dot podcast links a fund manager to AI drug-discovery research

The Polish-language Procent Składany episode adds a title-blind route connecting Robert Florczykowski, managing partner of closed fund Third Dot, with biotechnology and life-sciences investing. The publisher describes the fund as investing in technology companies in Western Europe and the United States, with particular interest in biotechnology. Its chapter markers place AI’s effect on drug discovery at approximately 28:00–34:00 and the bottlenecks in AI drug discovery at approximately 34:00–37:00. The page supplies a public Spreaker audio enclosure and Apple/Spotify show routes, but no transcript was recovered in this pass; the chapter markers are not treated as transcript-level quotations.

The publisher biography describes Florczykowski’s prior management of approximately PLN 3 billion at PKO TFI, his mathematics and quantitative- methods education at the University of Warsaw and SGH, postgraduate molecular biology at Jagiellonian University, and a supervisory-board role at Captor Therapeutics. These are publisher biography claims requiring first-party or regulatory cross-checks before quantitative personnel use. The route does not establish Third Dot’s clinical-trial dataset, model family, use of voice or imaging data, drug-approval model, trading model, decision authority, or AI-attributed performance. See the Polish capture note.

September 3, 2026 — Hungarian OTP quantitative-strategist and time-series research route

The Hungarian OTP Multi-Asset manager page names Péter Nemesi as a quantitative strategic analyst in the manager surface for the OTP Multi-Asset range. A Portfolio Conference profile states that he joined OTP Asset Management in 2024 after approximately two and a half years as an MSCI quantitative researcher in securitized-product research, and that he conducts ELTE research on financial-time-series analysis and forecasting with AI.

The ELTE AI Research Group seminar page and its presentation PDF identify a June 17, 2025 presentation by Nemesi and László Varga titled “Variational autoencoders to mimic and forecast time series.” The material describes attempts to mimic ARMA, GARCH, SARIMA, and fractional Brownian-motion processes, with planned extensions to stochastic volatility and real-world forecasting. The Portfolio AI in Business 2025 program also lists Nemesi in a session on ChatGPT, DeepSeek, and AI-model selection.

This is a useful academic and personnel lineage route, not evidence that OTP uses a VAE, LLM, or the seminar code in a live fund. The public sources do not disclose model ownership, data rights, validation splits, production permissions, order authority, or AI-attributed performance. See the Hungarian OTP capture note.

September 3, 2026 — Romanian personnel route for central-bank text signals

The Romanian public profile for Marius Iancu identifies him as a Quantitative Portfolio Manager at Erste Asset Management Romania. His public activity describes a “Hawkishness Index” for Romanian National Bank press releases, using sentiment or language-model techniques to measure hawkish and dovish communication and examine potential shifts in monetary-policy stance. The post cites a BIS working paper and says an interactive notebook is available on request.

This adds a person-first route linking a named quantitative portfolio role to central-bank text analysis, a useful extension of the earnings-call and news- sentiment search graph. The notebook, code, labels, evaluation window, point-in-time market join, model version, and data rights were not publicly resolved. The profile does not establish that Erste approved or deployed the index, that it informs a particular fund, or that it has predictive power. See the Romanian capture note.

September 3, 2026 — Slovenian Quantix systematic-options route

The Slovenian title-blind pass found Quantix Capital, whose first-party page describes a registered alternative investment fund manager specializing in fully systematic, algorithmic options trading. It describes intraday strategies, minimal overnight exposure, and team expertise in financial mathematics, machine learning, and algorithmic trading. These are firm statements about its stated strategy and capabilities; the page does not name models, data sources, model versions, validation windows, or AI/GenAI systems.

The QX00 Fund LEI record identifies an active Slovenian fund created June 23, 2025, with Securities Market Agency registration-authority metadata and Quantix Capital as managing fund parent. The manager LEI record identifies the corresponding active Slovenian legal entity, created July 16, 2024. These records corroborate an entity/fund relationship, not current assets, live trading, investor eligibility, model ownership, or performance.

This route adds a Slovenia-based systematic-options manager to the discovery map. It does not establish production use of LLMs, retrieval, fine-tuning, autonomous agents, or any specific ML model. See the Slovenian Quantix capture note.

September 3, 2026 — Serbian Risk Free Capital Management Pandora route

The Serbian title-blind pass found a public Risk Free Capital Management LinkedIn page describing a Belgrade-based private investment-management firm as a multi-strategy quantitative hedge fund using algorithms, machine learning, and statistical analysis. A public employee route identifies Radovan Martic as a quantitative trader and developer.

Martic’s public activity includes a company description of Pandora, an in-house black-box engine claimed to generate, evaluate, and prioritize trading strategies at scale. The post describes Random Forest pattern detection, a technical/statistical/behavioural feature pipeline, event triggers for unusual market behaviour and volatility, and downstream choices involving direction, risk limits, position size, and trade duration. It also says high-level intent, triggers, search space, and instruments remain user-defined.

This is a useful public architecture claim, but it is not an audited model card or independent deployment record. The sources do not establish Pandora’s current status, data, retraining, validation, live/paper separation, execution venue, human approval process, capital authority, or performance. They also do not establish that every part of the described workflow applies to a named fund. See the Serbian Pandora capture note.

September 3, 2026 — Saudi Green Rock and Pakistan hiring route

The Arabic/English regional pass found Green Rock’s first-party site, which describes a Riyadh-based private investment company using a proprietary AI engine for electronic stock trading, minute-level volatility prediction, and system-to-system execution. It also describes board and committee oversight for compliance, risk, and AI governance. These are first-party claims; no regulator record or fund vehicle was resolved in this pass.

Green Rock’s LinkedIn company page adds a Tadawul-equity and architecture vocabulary: deep learning, terabytes of unstructured alternative data, satellite imagery, semantic-web sentiment, adaptive neural networks, and a research-to-production pipeline. Pakistan job postings for an AI Software Engineer and Quantitative Developer mention NLP, PyTorch/TensorFlow, Spark, distributed or low-latency systems, Python/R/C++, statistical modelling, and quantitative education. A separate company hiring post mentions reinforcement learning, Bayesian statistics, optimization, signal processing, FastAPI, Kubernetes, and microservices.

The combined surfaces expose intended capabilities and a Pakistan build-out route, but not a filled roster, production status, named model, training corpus, data rights, validation design, execution venue, capital authority, or performance. The separate Saudi business-consulting entity using the GreenRock name remains an entity-resolution risk. See the Green Rock capture note.

September 3, 2026 — Turkish KALFA | BEYO systematic-fund route

The Turkish title-blind pass found KALFA | BEYO’s GNH Fund page, which describes the TEFAS-listed Global MD Portföy Algoritmik Model Hisse Senedi Fonu. The page says the fund keeps at least 80% in Borsa Istanbul equities, covers approximately 200 stocks, and uses proprietary financial-econometrics-based portfolio-optimization models. Inputs described include historical prices, domestic and international macroeconomic indicators, company financial data, and statistical portfolio-construction methods.

The page describes both scheduled and quantitatively triggered portfolio updates, with final allocation combining model outputs and a fund manager’s economic assessment. The KAP disclosure provides an official public reporting route for the GNH fund. Neither source identifies an AI/GenAI component, model family, code, feature definitions, training or validation windows, data rights, execution system, or AI-attributed performance.

The KALFA | BEYO company page adds a public personnel/company route involving S. Yanki Kalfa and Jeffrey Beyo and describes a hybrid process in which quantitative models generate signals while analysts review execution. Those are company/profile claims, not independent evidence of current employment or individual model ownership. See the Turkish GNH capture note.

The KAP fund summary identifies Global MD Portföy Yönetimi A.Ş. as founder and lists Barış Subasar as general manager and portfolio manager. The Global MD fund page links the prospectus, investor information form, risk-measurement principles, audit report, financial statements, and performance reports. The prospectus is a primary contractual document, but the reviewed documents still do not identify the proprietary algorithm’s model family, training data, validation design, or AI/GenAI implementation.

September 3, 2026 — title-blind multilingual expansion adds research-lineage and partnership layers

The latest multilingual and title-blind pass adds several distinct public surfaces. They should be read as evidence about stated processes, personnel, products, and research conversations. They do not support a ranking of firms, or a conclusion that any named method is deployed, permitted to trade, or responsible for performance.

ReSolve: a transcript-bearing research process

The ReSolve Riffs episode with Andrew Butler identifies Butler as ReSolve’s CIO and describes his responsibility for the firm’s machine-learning and portfolio-optimization ecosystem. The publisher transcript is unusually useful for process research: Butler discusses surrogate modelling from computationally expensive oil-reservoir simulations (03:59–07:49), noisy financial time series (10:51–18:59), and the need to treat universe selection, portfolio construction, and other hyperparameters as validation questions rather than incidental choices (28:32–32:53). The transcript explicitly describes training, validation, and final holdout sets for such design decisions (31:37 onward).

ReSolve’s Osprey strategy page separately describes a proprietary machine-learning process that searches thousands of signal combinations and filters candidate sub-strategies for overfitting, trading friction, and costs. These are first-party descriptions of a research and product process. They do not disclose model weights, current features, data vendors, permissions, or independently audited AI-attributed returns.

Taaffeite and T24: academic lineage is visible, model mapping is not

The T24 Capital about page identifies the T24 Complex as a machine-learning price-prediction and portfolio-risk- optimization system created by founder Desmond Lun. It also records Lun’s MIT EECS graduate degrees, computational-biology work at the Broad Institute of MIT and Harvard, Harvard Medical School research fellowship, Rutgers computer-science professorship, and University of Melbourne undergraduate training. A Livewire profile of Taaffeite provides the historical bridge from Lun’s academic work on complex biological networks to the Taaffeite long-short strategy and identifies Howard Siow as Taaffeite’s CEO.

This is a meaningful academic-to-investment lineage route: computational biology, network methods, machine learning, and optimization appear in the public biography surrounding a systematic investment business. The sources do not establish that every paper or patent is used by the current investment system, that Taaffeite and T24 have identical current operations, or that the historical system remains live in the same form.

Canada and Europe: CFM, Quantica, and Metori in a dated quant panel

The CAASA “Future of Quant” archive dates a May 23, 2023 webinar and lists Yves Lempérière of Capital Fund Management, Nicolas Mirjolet of Quantica Capital, Nicolas Gaussel of Metori Capital Management, and Christophe L’Ahelec of University Pension Plan Ontario. The programme frames the discussion around the progression from filters and trend-following systems to AI and marks a replay as available. The reviewed page does not expose a transcript, so it is a speaker and topic route rather than evidence of what each participant said or implemented.

Japan: partnership and product surfaces

Two Japanese SBI Holdings releases and SBI/AlpacaTech release add partnership routes. The January 2026 release announces joint research and development involving Vertex and AlpacaTech with investment from a Dai-ichi Life innovation fund. The May 2026 release announces a planned MixSeek AI data lake connection to Anthropic financial-agent skills. The accessible publisher shells verify the release identities, dates, and announced parties; the full Japanese bodies were not recovered in this text pass. Accordingly, they are partnership and architecture-vocabulary leads, not evidence that an Anthropic model is used in a fund or that an investment agent has live trading authority.

An older but separate Rakuten Big Data Japan Equity Fund page dates its product material to June 2019 and describes anonymized aggregated Rakuten ecosystem statistics, data-science and AI analysis, human qualitative judgment, and signals used to adjust the effective equity ratio. The page says not all Rakuten business data is used and presents parts of the process as an illustration. A Japanese Investment Management Association reprint names Shintaro Nagao of Rakuten Investment Management and discusses AI, data-science, and the limits of direct market prediction. That paper is a practitioner perspective, not a current model card or product audit.

Korea: AI appears in ordinary organizational structure

The Korea Financial Times profile dated February 17, 2025 describes KB Asset Management’s AI Quant & Direct Indexing headquarters, including domestic and overseas AI-quant teams, an AI financial-engineering team, and direct-indexing AI algorithm, strategy, and technology teams. It also describes Mirae Asset’s AI financial-engineering unit, Shinhan’s quant center, NH-Amundi’s AI-quant team naming, and Samsung’s AI-quant team. Named personnel include Kim Hong-gon, Lee Hyun-kyung, Kim Ki-deok, and Kwon Young-hoon.

The route matters because it captures AI in unit names and role structures, not only in product marketing. It remains a dated publisher account: it does not establish current staffing, model versions, training data, permissions, or AI-attributed returns. First-party role and hiring pages remain the preferred cross-check for individual claims.

China: LLM exploration alongside explicit quant controls

The STCN interview with Taiping Asset’s Wang Zhenzhou attributes to him a statement that Taiping Asset invested resources in AI and large language models to improve research efficiency and used large pre-trained models in technical implementation. The same interview discusses overfitting, changing market structure, factor controls, strategy diversification, and risk management. It therefore adds a useful public description of LLM exploration alongside ordinary quant governance language, but it does not name a model provider, training corpus, agent, code, or live decision permission.

The Blackwing first-party site describes quantitative equity, macro-hedge, and CTA products; mentions mathematical statistics and machine learning in the equity strategy; and uses “full-process AI quantitative investment” language in its company description. It names Chen Zehao and Zou Yitian as founders. These are explicit company claims and should remain separate from the existing Chinese event and media records. The page does not disclose a model family, data rights, validation design, AI-lab roster, or independently verified performance.

Vendor and allocator conference routes

The Neudata New York Traditional and Market Data Summit 2026 page dates the event to September 17, 2026 and displays Danielle Castelli as Head of Research at The Baupost Group, Dhagash Mehta as BlackRock’s Head of Applied Artificial Intelligence Research for Investment Management, Ilya Voytov as a Lazard quantitative research analyst, John Bauer in Jump Trading data strategy, Jordan Rubin as Trexquant’s Director of External Alpha, and Larry Komenda as Campbell & Company’s CTO. The organizer’s biographies describe research/data processing, AI and GenAI, responsible AI, machine learning, alternative data, and market-data operations. This is event-organizer biography evidence, not a complete roster or deployment audit.

The A-Team/Eagle Alpha event report dated April 8, 2026 identifies a New York panel involving Atlas Ridge Capital, UBS Asset Management, Jump Trading, Arctium Capital Management, Sensor Tower, and MacroX. It reports discussion themes including AI-driven data processing, data synthesis, centralized AI efforts, quality monitoring, faster data-source acceptance or rejection, and proprietary underlying data. The discussion was held under the Chatham House Rule, so this is a publisher account of themes and displayed roles rather than speaker-by-speaker evidence.

The Applied Investment Conference 2026 programme lists Avi Turetsky as Partner and Head of Ares Management’s Quantitative Research Group and Bill Kieser as Principal and Co-Head of its Research & Data Science team. The biographies describe quantitative tools developed with investment teams, institutional clients, and academics, and list Turetsky’s Bar-Ilan/INSEAD/Case Western Reserve lineage and Kieser’s Pennsylvania State/Villanova/University of Georgia training. The programme also lists a paper on forecasting private-equity NAVs using machine learning. It is a dated programme and academic-lineage route; it does not identify Ares models, data, code, permissions, or production attribution.

Brazil: Carbon Asset exposes ordinary quant titles and academic support

The C6/Carbon Asset first-party page identifies Carbon Asset as C6 Bank’s investment manager and names a resource-management team including Tiago Quixadá, Vinicius Martins, Fernando Fortunato, Larissa Frias, and Guilherme Ribeiro. Ribeiro is described as responsible for quantitative strategies focused on market trends, yield curves, and risk control. The page also describes systematic funds built from systematic models across domestic and international assets.

A C6 systematic-funds article, updated May 21, 2026, says the models are developed and monitored daily; describes a systematic futures fund spanning commodities, currencies, equity indices, and rates; and says the fund uses statistical models, daily rebalancing, and long/short positions. It reports approximately nine months of product development, assistance from researchers at Columbia and Stanford, and testing over roughly 20 years of asset history. These are first-party claims. They do not identify the researchers, model family, code, data rights, validation protocol, or AI-attributed performance.

Latin America: title-blind searches expose the implementation layers

The FAPP Chile quant-investment posting, published August 12, 2026, explicitly calls for asset-allocation and portfolio-optimization models, ML/deep learning/AI, training and validation, simulation, backtesting, and moving models from research into tools used by the investment team. This is recruitment evidence of intended scope, not evidence that the role was filled or that a model is live. A separate AFC Chile quant-markets posting is a useful control: it describes ordinary quantitative risk, attribution, optimization, stress-testing, and monitoring without an explicit AI claim.

The SFA Investimentos data-and-operations posting exposes a different layer at a Brazilian fundamental-equity manager: PostgreSQL, Python, automation, n8n, vector databases, AI, agents, and RAG appear as desired capabilities, alongside a stated June 2024 strategic partnership with Porto Asset. The posting does not establish production use or investment-decision authority. The Alpha Wave Capital site supplies concrete non-GenAI strategy vocabulary—implied-versus-realized volatility, probabilistic and regime filters, IV Rank, Z-scores, term structure, kurtosis, defined-risk options, VaR/CVaR, and stress testing—without identifying the model family or training data.

The iVi Technologies page names displayed roles spanning risk/compliance, operations, portfolio administration, technology, quantitative development, and data analysis, and describes proprietary platforms, AI, data science, optimization, simulation, and more than 80,000 daily data points. The Equus quantitative-fund index claims to cover ten Brazilian quantitative funds using AI and ML, but does not disclose the constituents on the reviewed page; it is a discovery index, not a ranking or independent verification. The Safra AI fund page is a separate category again: an AI-sector thematic fund, not a disclosure of internal AI-enabled security selection.

The ConoSur Asset Management site describes Argentine funds using quantitative models, algorithms, AI, adaptive algorithms, continuous backtesting, and automation from back office through the trading operation. The IMB Capital Quants page describes proprietary algorithms, ML, broad-market trading, backtesting, risk management, and human supervision, while explicitly stating that it does not directly raise capital or manage third-party funds. These records are useful precisely because they keep regulated managers, adjacent advisers/vendors, and marketing claims distinct.

The expanded Latin-American source note records the remaining Chilean, Brazilian, Argentine, Mexican, Colombian, and regional routes, including unverified candidates. None supports a firm ranking or a conclusion that a particular model is superior.

MENA and European-language searches: more operational AI, fewer “AI fund” claims

The BMCE Capital Cap’AI Reverse programme and its Data Scientist posting expose an activation pipeline in Morocco: business use cases, startup and FinTech engagement, MVP development, a hackathon, and use cases spanning compliance, risk, investment-decision support, liquidity prediction, and ordinary reconciliation/settlement automation. These are programme and recruitment signals, not proof that a finalist or investment tool reached production.

Hermes Capital describes licensed Iranian algorithmic wealth management and names portfolio-management and trading/asset-management roles. Dorfak Intelligent Systems describes an Iranian algorithmic-trading platform and portfolio-management products. Both are first-party descriptions without model architecture, training data, live permissions, or independently verified returns. Algorio’s quantitative-research role similarly describes systematic research, large financial datasets, backtesting, risk analysis, and production algorithms, but does not make an AI or GenAI claim.

The Danish and Nordic pass exposed a useful ordinary-title layer: Alipes’ careers pages describe ML research, high-frequency data, time-series architectures, and a quantitative team; Nordea’s Trading Quant role mentions order generation, execution, trading-data models, and AI use cases; and MFT Energy’s annual report describes algorithmic power and gas trading, data from more than 50 providers, and ML models. These sources speak to intended or company-reported infrastructure, not comparative performance.

The AI Alpha Lab materials are one of the clearer European model disclosures: a Bayesian-neural-network strategy and monthly retraining are described, while the portfolio report labels historical results simulated. The Elo practitioner article is a useful counterweight, describing ML applications in sentiment, quantitative-model construction, auditing, and allocation while discussing overfitting and regime instability. The expanded MENA/European source note records the remaining Persian, Arabic, French, Hebrew, Nordic, Central European, and regulatory routes.

The title-blind control search also found a CFA Institute conversation with Johns Hopkins and Yale finance researchers, a CFA UK panel including Quantmate’s co-chief AI officer, and a Lord Abbett portfolio-manager transcript. These broaden the media corpus across portfolio construction, governance, research productivity, and fundamental investing; none by itself establishes a hedge fund’s live AI system or performance. The mainstream-investment source note preserves those boundaries.

The new title-blind job pass adds public capability signals across the operating stack: Bridgewater’s commodities data role mentions vendor-data quality, ontologies, alternative data, AI/ML, and development assistants; Citadel’s data-strategy role mentions knowledge graphs and AI/ML research; Winton’s quantitative-platform roles mention lineage, LLM-assisted data investigation, and research-to-live validation; Schonfeld’s role mentions APIs for quants and AI agents, feature pipelines, and RAG; Man Group’s data and streaming roles mention AI/LLM tooling and agentic engineering; and Robeco’s quant-platform role mentions text, documents, audio, and GenAI/LLM-powered solutions. The full job/infrastructure source note also records Arrowstreet, Balyasny, Dimensional, G-Research, Jane Street, Point72, WorldQuant, AQR, Winton, and Schonfeld roles. These are recruitment and platform evidence, not proof of filled roles, live deployment, investment authority, or performance.

The same title-blind pass recovered a new Acadian transcript and new Man Numeric appearances (John Lidington, Greg Bond). It also found an Arrowstreet governance/lineage route through John Campbell, a historical Renaissance route through Howard Morgan, and the Colossus “Alpha In Podcasts?” episode, whose public excerpt makes niche podcasts and industry commentary an explicit discovery surface while keeping its claims secondary and login-gated. None of these routes supports a firm ranking or a claim about current model quality.

The academic/personnel expansion adds a separate talent-network layer. Hao Zhang’s public research site describes a Balyasny Senior Research Scientist remit spanning financial AI, LLM post-training, agent harnesses, AutoML, and alpha/feature mining; this is self-described and does not establish deployment. Pierre Laforgue’s profile links current CFM quantitative research to online learning, bandits, federated learning, kernels, robustness, and sample-bias work, while Jun Yan, Vitaly Kuznetsov, and Lang Liu expose HRT and Citadel Securities research affiliations without naming production systems. Other university and personal pages add Point72 internship/placement routes, AQR macro and former-NLP personnel, Citadel Securities RL/transformer and LLM-lineage routes, Squarepoint data-driven-strategy personnel, and a Two Sigma ML-in-finance placement. Historical records for Acadian, Arrowstreet, and CFM are kept separate from current headcount. The academic-lineage source note records advisors, papers, dates, and identity caveats; no source in this pass maps a paper directly to a named live investment model, dataset licence, or portfolio permission.

The conference route expansion adds four follow-up surfaces without converting event marketing into firm evidence. Battle of the Quants London is scheduled for October 21, 2026 and explicitly advertises AI, LLMs, agentic AI, alternative data, and quantitative hedge-fund programming, but its reviewed page does not yet publish speakers. Arena/GlobalData’s AI in Financial Services event is scheduled for September 8–9 in London and publicly lists BlackRock personnel among its featured speakers; the 2026 agenda is access-controlled. RE•WORK’s Fall AI in Finance brochure advertises October 21–22 programming and on-demand access in Hoboken, while Agentic Finance Singapore lists Monarq Asset Management and Pantera Capital personnel for October 8. These are conference and person-discovery routes; they do not establish attendance, model ownership, data rights, live deployment, or performance. See the conference expansion note.

Family offices and long-horizon capital are a missing evidence layer

Family offices and similar capital pools often publish less than hedge funds, but the public language can still expose a different operating model. Pylon Investment Group identifies itself as a Hong Kong family office investing proprietary capital in global equities and derivatives. Its site says the firm invests in AI and uses machine intelligence in its research process, with a focus on Korea and the United States. That is a first-party description of mandate and process; it does not name the model, data, team, permissions, or outcome.

Two additional proprietary-capital routes widen the map. Confluence Capital describes a Calgary-based private investment office that manages proprietary capital and lists quantitative/systematic strategies, algorithmic trading, data engineering, and applied machine learning among its research interests. Apex Capital Holdings describes a family office with more than 20 professionals and a dedicated quantitative-algorithmic division developing proprietary models and tools. Apex also publicly connects its family-office operation to a separately named seeded fund and software that originated in-house for pre-market workflow, idea generation, and algorithmic position sizing. These pages disclose operating language and entity relationships, not model families, training data, permissions, evaluation design, or performance.

Iron Hall Capital adds a Portugal-based proprietary- capital route. Its public site names Bernardo de Ascensão as founder and describes AI/ML-assisted testing of academic ideas, strategy backtests, regime and sentiment inputs, and in-house research, decisions, and trading. The site also says it does not manage external capital. These are first-party descriptions: no model family, training corpus, independent evaluation, execution log, or audited performance is publicly established.

The founder’s separate public professional site adds a self-authored technical surface: Bernardo de Ascensão lists CNN, RNN, LSTM, Transformer, and reinforcement-learning methods for financial time series, along with backtesting, statistical validation, risk modeling, position sizing, and data pipelines. It also describes a financial-research product called QFI Terminal. The page does not establish which techniques are used in Iron Hall’s live process, nor does it disclose training data, evaluation splits, or independently audited results.

The allocator-service layer adds further routes without changing the entity classification. Henderson Rowe describes Rayliant machine-learning algorithms in institutional mandates and a white-label OCIO service for family offices; its reported asset figure and model description are first-party claims. Caelion describes AI delivery, data governance, and investment-risk work in a live single-family-office engagement and anonymized sovereign-fund work. Neither page identifies a named family office’s model, permissions, or investment results.

SFDT Capital describes a family-office investment mandate covering foundation models, generative AI, AI infrastructure and MLOps, applied machine learning, and intelligent automation. The page establishes an explicit investment focus, not an internal trading system or independently validated performance record. Tang Ventures adds a venture-capital family-office route focused on frontier AI, compute, and robotics; it is relevant to capital formation and technical-company access, not evidence of public-market quantitative research.

The adjacent institutional layer has more explicit systems language. Mubadala’s MAIA case study describes a proprietary AI-enabled platform for investment intelligence and operational efficiency and gives a future vision for investment-committee use. The Public Investment Fund’s strategy lists advanced AI with strong data foundations among its 2026–2030 objectives, while its HUMAIN announcement describes a PIF-owned company operating and investing across the AI value chain. These are sovereign-investor and ecosystem-building disclosures, not hedge-fund model evidence.

Soros Fund Management identifies itself as a family office and describes AI risk management, responsible data use, and a corporate AI-risk assessment framework. This is a stewardship and governance signal rather than a disclosure of internal alpha research. At the allocator level, MITIMCo provides an MIT-linked endowment-investment route with external-manager due diligence and a generalist research model. Harvard Management Company remains an important endowment watchlist candidate, but the reviewed public pages did not provide a current AI-specific investment-system disclosure.

Two ecosystem sources provide context without turning these cases into a census. Citi’s family-office report describes privacy as a central constraint and reports that family offices commonly prioritize operational leanness before alpha generation. Morgan Lewis’s guidance emphasizes documented use cases, approved tools, data controls, and human review. The family-office source note records the survey and guidance boundaries, including the fact that neither source identifies a complete set of participating offices. See the family-office and long-horizon-capital source note for the full route inventory and recovery queue.

These routes should be tracked separately from hedge funds: family offices, endowments, sovereign investors, foundation-linked asset managers, venture platforms, and service providers have different disclosure incentives and decision rights. The public evidence now supports a broader map, not a ranking or comparative quality conclusion.

A second pass found a further layer that should not be collapsed into family offices. Gabah Research describes an AI-assisted prediction- market research lab in which the model attacks hypotheses while the researcher retains deployment authority; its public Russian, German, and French pages also make it a useful multilingual discovery route. RS Investment and Lehman Capital describe proprietary-capital research and quantitative/ML stacks, but their public claims require independent entity and operating verification. Prodigy Research names an AI quant-research lab and two founders, while its performance and model comparisons remain unverified and are not used here.

The vendor layer is equally important. Empiric’s investment-committee case study describes an unnamed private office using an AI review layer over investment motions. Alpha Intelligence Labs, altHQ, and QINV expose the product vocabulary being sold to family offices and allocators: document ingestion, institutional memory, manager diligence, portfolio aggregation, supervised recommendations, audit trails, and reporting. These are vendor claims or anonymized cases; they do not identify a named investor’s internal system. The GPIF AI-impact report adds a historical pension-investor baseline spanning operational efficiency, quantitative risk analysis, intrinsic-value assessment, macro forecasting, and tracking-error reduction. The expanded proprietary-capital and allocator source note records the verification and disqualification boundaries.

The media search also recovered an allocator-specific event and podcast layer. The FOX 2025 Family Office AI Roundtable in Cambridge named Cody Crowell of Frisbie Group for a family-office AI-investing spotlight, Cogo Labs leaders for live AI demonstrations, and MIT-linked speakers including Ramesh Raskar, Alexander Amini, and Daniela Rus. The agenda establishes public association with the topic, not any named office’s internal deployment. The FOXcast Tide Cycle episode and Mack Podcast OCIO panel add title-blind routes about virtual family offices, OCIO, and buy-versus-build decisions. Their public metadata does not establish AI tools or investment results; audio recovery and speaker-level extraction remain queued. The family-office media source note records the dates, distribution routes, and capture boundaries.

The university layer is also now tracked as a distinct research source. Wharton’s Winston Wei Dou directs its AI in Finance Lab and publishes on asset pricing, AI market behavior, liquidity, competition, and financial stability. MIT’s Hui Chen lists financial machine learning, structural models, LLM interpretability, uncertainty quantification, and dynamic private-asset allocation. Yale’s Theis Jensen lists machine-learning asset pricing and an implementable-efficient-frontier paper. UCI’s Jinfei Sheng lists textual analysis across news, earnings, reviews, crypto whitepapers, and fund prospectuses, and a public speaking route that includes Citadel, BlackRock, and Citibank. These are academic and speaking signals, not evidence of employment or live deployment at those firms. The academic finance source note records the programme, lineage, paper, and disqualification fields.

The expanded university pass adds two dense idea and talent surfaces. Chicago Booth’s Dacheng Xiu is described as developing statistical and machine-learning methods for financial data and as having built Booth’s AI Essentials course. Booth’s applied-AI finance pages and Xiu’s public research archive cover nonlinear asset pricing, neural nets, regression trees, CNNs for price-chart representations, autoencoders, sentiment, text and image data, and an LLM-related expected-returns working paper. Columbia’s Harry Mamaysky combines a finance-program director role with research on machine learning and the information in news, earnings calls, and central-bank communications; Columbia’s MSFE materials list Text Data in Finance and Machine Learning. These sources expose research directions and training pipelines, not named-firm adoption, proprietary data rights, live portfolio permissions, or investment outcomes. The academic source note records the paper, lineage, and evidence boundaries.

Berkeley Haas adds an unusually explicit programme route: its public faculty page lists a three-course Machine Learning for Finance sequence progressing from supervised and unsupervised learning to reinforcement learning, then generative AI, causality, and LLMs. The same page connects Ali Kakhbod’s MIT economics/Michigan EECS training to research on information, beliefs, and AI-related investment perception. Anastassia Fedyk’s profile adds research on financial news, employment records, firm technology investment, and AI/workforce composition, alongside prior Goldman Sachs Asset Management research and portfolio management roles. These are university curriculum and research signals; they do not establish participation or adoption by any named fund, proprietary data access, live model authority, or performance.

The European academic layer adds implementation constraints and recruiting paths. London Business School’s Victor DeMiguel describes machine learning and AI in portfolio optimisation and asset pricing with parameter uncertainty and trading costs; Svetlana Bryzgalova’s profile adds empirical asset pricing, financial econometrics, macrofinance, and machine learning. Imperial’s MSc Mathematics and Finance explicitly separates machine-learning-in-finance electives, including reinforcement learning and deep learning, and describes supervised projects that may involve external sponsors. Yufei Zhang’s Imperial profile connects mathematical finance, machine learning, stochastic control, and games. These pages reveal curriculum, research, and talent-pipeline structure; they do not identify a sponsor’s live model, proprietary data, deployment permission, or result.

The international academic route now includes Princeton’s Bendheim Center for Finance, whose programme materials connect finance training with machine learning and AI, and NUS’s MSc Finance curriculum, which lists Machine Learning in Investments across idea generation, allocation, execution, and evaluation. Ke-Wei Huang’s NUS profile adds a named academic and programme-leadership route spanning data mining in finance, causal inference, machine learning, and digital finance; NUS’s AI in Capital Markets programme also identifies generative-AI instruction. These pages establish education, research, and talent-pipeline scope, not participation by a named fund, proprietary data access, production permissions, or investment outcomes.

September 3, 2026 — academic finance routes expose model ideas and talent channels

The academic layer is useful when it is read as an idea and validation map rather than as a league table. Stefan Zohren’s Oxford profile connects Oxford–Man, commercial projects with Man Group, and research directions including diffusion models for limit-order-book simulation and forecasting, graph auto-encoders, LLM-embedding compression, LOB benchmarking, and LLM-integrated portfolio optimisation. Maike Osborne’s profile adds Gaussian processes, active learning, Bayesian optimisation, probabilistic numerics, and changepoint-tolerant inference, while Samuel Cohen’s profile connects Oxford–Man and the Alan Turing Institute’s Machine Learning in Finance theme to decision-making, control, and uncertainty aversion. These are explicit research routes; they do not establish Man Group adoption, a named production model, proprietary data rights, or investment results.

Markus Pelger’s Stanford research archive adds economically constrained deep-learning asset pricing, statistical arbitrage with learned representations and portfolio constraints, fund-manager-skill prediction, and large-panel multiple-testing controls. The Stanford Advanced Financial Technologies Laboratory archive provides historical context on nonlinear asset-pricing models and interactions among momentum, liquidity, and volatility. These pages show what to test and what to control for in a research design; they do not identify a current fund system or attribute returns to a named employer.

Vasant Dhar’s NYU Stern profile provides a direct academic/practitioner lineage through systematic investing, prediction, AI governance, and the historical founding of SCT Capital Management. Columbia’s MS Financial Economics curriculum shows a finance-program pipeline combining computing and big data in finance, empirical asset pricing, systematic investment strategies, text data, machine learning, artificial intelligence, generative AI, and a quantitative thesis seminar. These are talent and curriculum signals, not evidence of current SCT architecture, student-project content, proprietary datasets, or employer deployment. The expanded academic source note records the source boundaries and the next extraction fields.

Columbia’s AI for Business & Finance certificate adds an executive-education and speaker-discovery route distinct from the MSFE. Its official programme page explicitly names investment analysts, portfolio managers, quantitative analysts, hedge-fund analysts, risk and compliance professionals, and finance technologists as relevant participants. It describes scenario modelling, predictive analytics, alternative-data analysis, portfolio and risk work, fraud detection, and financial APIs, and lists guest-speaker roles at BlackRock, Morgan Stanley, T. Rowe Price, Capital Group, and Lexington Partners. The page is evidence of programme scope and public speaker metadata; it does not establish session attendance, employer sponsorship, a speaker’s current internal system, or deployment of any course method.

Two additional routes connect academic finance to practitioner implementation. Ben Charoenwong’s public INSEAD profile identifies him as an Associate Professor of Finance, while his research archive lists financial regulation, financial technology, networks, alternative data, and machine-learning work. The Blushing Quants interview identifies his Chicago Global role and discusses theory-driven quant research, feature engineering, explainability, model simplicity, alternative data, and AI’s effect on finance education. This links an academic profile, a fund-management practitioner route, and concrete research questions; it does not link a paper to a Chicago Global production model or disclose proprietary data, permissions, or performance.

The CFA Institute Research Foundation’s 2025 AI in Asset Management volume adds a professional-finance education route. Its public chapter map spans unsupervised learning, network theory, deep learning, reinforcement learning, natural-language processing, commodity-futures ML, and ethical AI. The Paul Bilokon episode adds implementation checks around trustworthy backtests, data normalization, corporate actions, market microstructure, execution lags, transaction costs, market impact, simple baselines, and explainability. This is a useful research design and talent-development route, not evidence of a named fund’s deployment.

September 4, 2026 — finance-specific labs and professors expose additional model ideas

The University of Toronto’s Rotman FinHub is a university finance lab, teaching platform, and industry-facing discovery surface. Its public research list includes synthetic data for regulatory stress testing, neural-network valuation using volatility-surface features, and “Agents Are Not Algorithms,” a paper with Royal Bank of Canada co-authors that studies decision-time reasoning in a trading simulator for tender selection and execution. The same page lists finance courses covering supervised, unsupervised, and reinforcement learning, Python, financial APIs, and data-driven case studies. These are publicly described research and teaching artifacts; they do not establish a hedge fund’s live implementation.

The Rotman route also exposes practitioner and personnel links. The John Hull Financial Innovation Fund forum frames new funding around derivatives, hedging, risk management, and AI/ML in finance. Jacky Chen’s profile identifies him as Managing Director of Completion Portfolio Strategies at OPTrust and an Adjunct Professor, with stated ML research interests in portfolio hedging, derivatives pricing, and risk management. Zissis Poulos’s York profile adds financial NLP, generative models, volatility, and derivatives hedging, plus his prior Rotman postdoctoral role and Tartan AI co-founder history. These pages establish affiliations and research topics, not a shared employer system or production authority.

Dimitris Bertsimas’s MIT profile adds a methods and talent route through optimisation, machine learning, and finance. It identifies him as MIT’s Associate Dean of Online Education and Artificial Intelligence, Vice Provost for Open Learning, Boeing Professor of Operations Research, and faculty director of the Master of Business Analytics. His listed finance applications include dynamic portfolio theory, asset allocation, risk management, optimal execution, and derivatives pricing. This is evidence of a research and training pipeline connecting optimisation and statistical learning to financial decisions; it does not identify a named fund’s deployment.

Tim Bollerslev’s Duke profile adds a title-blind empirical-finance route. The page lists a 2026 Management Science paper on forecasting time-varying correlations with realized-correlation features and LASSO, and a 2025 paper on intraday market-return predictability using modern machine learning. His stated research surface includes high-frequency data, volatility, and financial econometrics. The ideas to carry into the research queue are feature selection for correlation risk and intraday signal construction, while the public page remains insufficient to infer any investment-manager implementation or performance.

Carnegie Mellon’s MSCF academics page describes a computational-finance curriculum jointly built across business, statistics and data science, mathematical sciences, and computer science. It combines quantitative finance with hands-on projects, industry-sponsored competitions, internships, and collaborations with firms. This is a talent and applied-project route to inspect for public speakers, student projects, recruiting patterns, and research topics; the page does not identify a particular employer’s model or data access.

The Baruch MFE programme lists machine learning for financial engineering, market microstructure, time-series analysis, algorithmic trading, risk, commodities and futures, and derivatives hedging and valuation. Its 2026 ML seminar syllabus names asset pricing via ML, factor models, regime detection, credit-risk classification, and portfolio optimisation. This is curriculum evidence, not evidence that a fund hired a participant or adopted a classroom method.

The HKUST MSc in Financial Mathematics catalogue is an Asia-Pacific quant-finance route with reinforcement learning, AI in fintech, Python for medium- and high-frequency multi-factor models, trading simulators, abnormal-trading detection, and financial-firm capstone projects. It states learning outcomes around derivatives pricing, portfolio management, trading strategies, risk monitoring, and investment opportunities in fund management. The HKUST AI in Finance course record adds NLP, predictive analytics, fraud detection, portfolio optimisation, case studies, and hands-on financial-model development. These pages reveal programme scope and potential employer-project channels; they do not establish project sponsors or live deployment.

The Swiss Finance Institute at EPFL faculty roster connects named finance faculty to quantitative risk, machine learning in finance, portfolio selection, nonlinear filtering, and liquidity and asset prices. Damir Filipovic and Semyon Malamud provide professor-and-paper routes for extracting model assumptions and doctoral lineages. Their public expertise labels are discovery seeds, not evidence of a fund relationship, data licence, model ownership, or investment result.

Travis L. Johnson’s McCombs profile identifies him as an Associate Professor of Finance at the University of Texas at Austin, with research spanning AI and machine learning in finance, options, asset management, market microstructure, and financial econometrics. He teaches “AI Powered Investments,” which treats modern AI tools as partners in the investment management process while emphasizing capability limits and failure modes. His public 2026 working-paper list also links a long-horizon financial-statement forecasting project, a public ProForma-20Q benchmark, and code. These materials expose a research-and-evaluation route for financial-document forecasting; they do not establish fund adoption or investment performance.

Pietro Bini’s Boston University profile adds a Boston-linked academic route. It identifies him as an Assistant Professor of Finance at Questrom, with empirical asset pricing and portfolio-choice work, and lists risks of generative AI and machine learning in financial decision-making as a current research interest. The page also links his Fintech@Cornell affiliation, DEFT Lab organising role, and AI & Big Data in Finance Research Forum affiliation. The profile establishes research and network links; it does not establish an employer’s AI system, dataset, or live portfolio use.

Anthony Sanford’s HEC Montréal profile connects options-based return forecasting and portfolio construction to the Machine Learning in Finance (Fin-ML) CREATE programme. His teaching list includes financial econometrics, portfolio management, computational finance, and developing, coding, and evaluating financial trading systems. This creates an academic route for examining option-implied uncertainty, system evaluation, and Canadian ML-in-finance talent, while leaving employer adoption and production authority unresolved.

Anton Lines’s Copenhagen Business School profile frames machine learning in finance around asset pricing, asset management, and the trading and investment decisions of institutional investors. This is an institutional-investor research route that points beyond prediction toward how organisations make allocation decisions. The public profile does not identify a particular manager, proprietary dataset, model, or performance result.

Peking University HSBC’s 2025 Machine Learning in Asset Pricing syllabus names Lingxiao Zhao as instructor and explicitly combines arbitrage pricing, factor models, portfolio analysis, return prediction, investor learning, and quantitative investment with R/Python exercises. It also identifies a reproducible data stack: WRDS, CRSP, Compustat, the French data library, Goyal–Welch predictors, Fama–French and momentum factors, and FRED macro predictors. The syllabus is a course and data-access map, not evidence that students or any fund used those data in production.

The Canadian Fin-ML CREATE programme team connects six universities—Université de Montréal, HEC Montréal, Concordia, Waterloo, Queen’s, and Calgary—with machine learning, quantitative finance, and business analytics. Its graduate programme describes training future researchers and professionals, collaborative applied research, and access to research-internship offers from an industrial network. The curriculum page adds intensive ML-in-finance and insurance modules and problem-solving workshops where partners present R&D problems. This is direct evidence of a structured academic–industry talent channel, but not evidence that a named hedge fund supplied a problem or hired a participant.

The ProForma-20Q benchmark repository turns the financial-statement forecasting route into a reproducibility target. It states that the benchmark is a protocol rebuilt from a user’s own WRDS access, using Compustat Fundamentals Quarterly, Compustat industry mappings, and the CRSP CCM link table. Its submission specification requires forecasts for 78 statement items at horizons one through twenty quarters and supports an optional predictive standard deviation for probabilistic evaluation. This creates a concrete model-and-evaluation surface for long-horizon forecasting; the repository does not prove any fund’s use of the benchmark, data licence, or investment authority.

September 4, 2026 — further academic and programme routes

The academic map now extends to dedicated finance-technology labs and applied programmes outside the previously tracked US/UK/Canada routes. Stanford’s Advanced Financial Technologies Laboratory describes research spanning big data, machine learning, computation, optimisation, stochastics, and financial institutions, while exposing researcher, seminar, conference, and industrial-affiliate pages. This is a source-discovery surface, not evidence of any firm’s model ownership or deployment.

Stanford’s 2026 applied-AI macro-finance session also surfaces a specific research design: retrieval-augmented generation over filings, patents, earnings calls, peer disclosures, and news to construct an AI-derived measure of marginal q. A separate Stanford GSB working paper describes graph-based deep learning over financial-intermediary holdings, with economic priors and inductive representations of assets and investors. Both are academic research claims and useful evaluation ideas; neither establishes a live investment system or fund adoption.

MIT’s Andrew W. Lo profile adds a finance/AI/healthcare-finance research route through the Laboratory for Financial Engineering, CSAIL, machine learning and LLM applications, and the AUTOCT clinical-trial prediction paper. Singapore Management University’s Peng Liu profile adds explicit course routes in machine learning and reinforcement learning for quantitative finance, quantitative trading, financial data science, and risk management. These pages show research and talent channels, not firm implementation.

India adds three different signals. IIM Calcutta’s Financial Research and Trading Lab describes real-time market-data infrastructure and historical NSE tick data. IIT Madras’s CIFIL projects include machine-learning comparable-company valuation, and M. Thenmozhi’s profile links finance, hybrid AI, SVM prediction, and a fintech innovation lab. NISM’s data-science programme covers ML and deep learning in finance, econometrics, time series, algorithmic trading, and risk. Adelaide University’s Hao Zhou profile adds algorithmic trading, market microstructure, execution, and finance/data- analytics teaching. These are academic and professional-training routes; they do not establish a named fund’s model, data rights, permissions, or returns.

September 4, 2026 — firm-adjacent media and academic bridges

The title-blind pass recovered a useful set of routes between personnel, recruiting, technology vendors, and academic research. An OxWoCS Marshall Wace event names a Quant Developer and Quant Researcher and frames doctoral computer-science training as part of hedge- fund work. A Columbia IEOR Squarepoint session uses language about systematic strategies, large-scale data analysis, high- performance trading platforms, and compute farms, while an IAQF event adds a named quantitative-researcher career route. These are event and recruiting surfaces; they do not disclose production models, data rights, or returns.

For Brevan Howard, the Neudata data summit page identifies a US Deputy Head of Data Strategy and describes enterprise data, alternative data, and AI-solution responsibilities. An industry-summit recap also places the firm’s CTO in an AI/technology panel. These pages establish role and event context, not a model inventory, vendor list, or live investment workflow.

The TWIML Alpha Lee episode is a useful Winton-adjacent discovery route because it identifies a Winton Advanced Fellow and discusses uncertainty estimation, optimisation, and scientific ML. The Schonfeld Quantbot spotlight and International Mathematical Union report add technology-scaling and mathematics/data/AI event routes involving Schonfeld personnel. None of these sources establishes that the academic or vendor material is used in a named firm’s investment system.

The NVIDIA GTC24 financial-knowledge-graph session names a BlackRock speaker and describes document retrieval and RAG for financial knowledge. It is a concrete architecture-discovery route, but the session abstract does not establish production deployment, data permissions, or investment impact. The full evidence table and recovery queue are in the firm-adjacent source note.

September 4, 2026 — benchmarks, open data, and allocator-side routes

The academic pass also recovered resources that can support actual evaluation rather than only personnel discovery. NEFIN at the University of São Paulo states that it provides nine free, machine-readable datasets and exposes the last observation for each series. Its page says the Brazilian Fama–French/momentum risk factors and short-interest data were updated in July 2026, while some other series remain dated 2020–2023. That makes NEFIN useful for Brazilian factor and liquidity experiments, but it also creates a direct staleness-control requirement: a backtest cannot silently treat every downloaded series as current.

VisFinEval provides a public Chinese multimodal-finance benchmark with 15,848 annotated question–answer pairs, eight financial-image modalities, and 15 scenarios across front-, middle-, and back-office work. The repository reports organizer-run zero-shot evaluations of 21 multimodal models, including a 76.3% overall score for Qwen-VL-max and a gap of more than 14 percentage points against financial experts on advanced domain tasks. Those are benchmark-owner claims under the repository’s protocol, not a model leaderboard or evidence of investment value. The scenario counts make it possible to test chart reading, financial-statement extraction, strategy backtesting, allocation, and risk control separately rather than collapsing them into a single “finance AI” score.

FAMMA and the CLEF 2026 FinMMEval Lab add multilingual financial-document and multimodal evaluation routes across Asian, European, and other language contexts. They should be treated as benchmark inputs whose releases, splits, licences, and metrics require pinning; benchmark scores do not by themselves establish factuality, tradability, or portfolio value.

The programme layer now includes the Simon Fraser Quantitative Finance Lab, the University of Calgary trading and market-data course, the ETH Zurich ML in Finance and Insurance course, FinEML 2026, and the Concordia AI in Finance 2026 conference and PhD replication workshop. Their public pages expose distinct method-and-workflow routes: professional market data and simulation; regularisation, trees, boosting, neural networks, graph models, transformers, pricing and hedging; factor discovery and stochastic discount factors; and WRDS-based replication of published ML-finance work. These are talent and research routes, not evidence of named-manager deployment.

For the article’s idea queue, the useful crosswalk is: professor or course page → method and target variable; benchmark or open data → task and split definition; replication workshop → disqualification test; and only then a manager source → possible organisational relevance. The crosswalk should record point-in-time data, language, licensing, costs, capacity, human review, and whether a claim is a course description, benchmark result, paper result, or live disclosure.

The personnel map is also expanding across Australia and Switzerland. Bond’s Rand Low profile connects quantitative finance, portfolio optimisation, systematic trading, ML, and model-risk governance, with prior bank and asset-management modelling work. Basel’s Dietmar Maringer profile lists computational finance, algorithmic trading, high-frequency markets, regime- switching reinforcement learning, and market-impact-aware portfolio optimisation. These are faculty biographies and publication routes, not evidence of any named manager’s live system or results.

Allocator-side sources reveal a different operating layer. The Family Office Exchange AI in Action Workshop explicitly covers reporting, research, document review, workflow agents, confidentiality, governance, and deciding where AI should not fit. The Sohn 2026 conference page adds public video and family-office-investing discovery routes. These pages identify agendas and use cases, not participating families’ internal systems or investment outcomes.

The complete evidence table and recovery queue are in the academic finance, benchmark, and allocator source note.

Evidence boundary and recovery queue

The expanded source note records the full route set and negative findings. The next recovery steps are to obtain authorized Japanese release bodies, locate the CAASA replay and any CFM/Quantica/Metori materials, verify current T24/Taaffeite status, cross- check Korean organizational claims against first-party pages, and locate Taiping or Blackwing technical and hiring documents. None of those open tasks should be filled by inference from a job title, conference appearance, or marketing phrase.

September 4, 2026 — college professors and finance programmes as idea routes

The academic layer is now part of the firm-research map because it can expose model ideas, datasets, evaluation designs, and talent channels without pretending that a course or professor represents a hedge fund. Stanford’s Advanced Financial Technologies Laboratory explicitly joins finance with big data, machine learning, computation, optimisation, stochastics, and algorithms, and publishes separate research, people, events, and industrial-affiliate surfaces. This is a route for finding papers and speakers; it is not evidence of GMO, Acadian, Arrowstreet, or another manager using a particular system.

Several academic routes map directly to testable research questions. Stanford’s 2026 applied-AI macro-finance session describes retrieval over filings, patents, earnings calls, peer disclosures, and news to construct an AI-derived marginal-q measure. The Stanford GSB financial regulation paper uses graph-based deep learning over intermediary holdings and economic priors. Those are research designs that can be reproduced and stress-tested; neither page establishes live trading, proprietary data access, or hedge-fund adoption.

Professor and programme pages also expose different implementation layers. MIT’s Andrew Lo profile joins machine learning and LLM applications with quantitative finance and healthcare finance. NHH’s FIN544 course combines real datasets, causal identification, paper replication, LASSO/Elastic Net, NLP, local LLMs, and an individual empirical-finance project. ISB’s Centre for Analytical Finance describes applied ML tools for credit scoring, customer retention, and transaction anomaly detection with associated financial institutions. These sources distinguish research training and applied analytics from evidence about a named quant manager.

The finance-programme map extends into execution, portfolio construction, and market microstructure. FGV EMAp’s algorithmic-trading syllabus covers stochastic control, Almgren–Chriss execution, market making, inventory risk, feature engineering, supervised learning, reinforcement learning, portfolio, and risk. ETH Zurich’s ML in Finance and Insurance course lists regularisation, trees, boosting, neural networks, autoencoders, graph neural networks, transformers, pricing, hedging, and credit analytics. CBS BIGFI adds market, register, regulatory, and institution-provided data to a research programme covering ML in finance, replication, market stability, and social networks. Syllabi and lab pages reveal the idea supply and data vocabulary; they do not reveal any firm’s production stack or performance.

For multilingual and regional coverage, SUFE’s FinEval repository provides a Chinese financial-domain evaluation route, while CFLUE adds financial-language tasks and Fin-R1 adds an open finance-reasoning model route. IIM Calcutta’s Financial Research and Trading Lab exposes real-time and historical market-data infrastructure. These are useful for language coverage, data access, and evaluation design—not evidence that any tracked firm uses them.

The research note Academic finance AI and programme routes contains the expanded professor, programme, benchmark, conference, and recovery queue. For every paper or course, the next pass records the forecast target, horizon, modality, data vintage, point-in-time controls, split, transaction costs, and portfolio mapping before treating an idea as backtestable. Academic evidence and firm evidence remain separate, and no route is ranked.

September 4, 2026 — additional university talent and research surfaces

The follow-up academic search found several more public routes that are useful for idea discovery and personnel mapping. Columbia’s MS Financial Economics programme lists machine learning, AI, systematic investment strategies, text data in finance, empirical asset pricing, and thesis or internship routes. Its Text Data in Finance course names Harry Mamaysky and covers Python-based NLP for news, earnings calls, and central-bank communications. This is curriculum and talent-pipeline evidence, not evidence of a particular fund’s model or employer project.

The Oxford-Man Institute provides another public research surface spanning machine learning, data-driven models, seminars, conferences, industry participants, and a data library. Its financial-statement analysis paper is a concrete replication route involving abnormal-return forecasts around earnings announcements and comparisons among model families. Any reproduction still needs point-in-time features, dated splits, costs, and robustness checks.

Berkeley Haas’s MFE curriculum explicitly combines Python, machine and deep learning, financial data science, standard and alternative data, and independent studies with financial institutions, hedge funds, and fintech firms. Wharton’s Michael Roberts profile adds a faculty and initiative route focused on machine learning in finance and financial analytics. These pages expose academic training and possible industry contact surfaces; they do not disclose counterparties, proprietary data, production permissions, or investment results.

The full additions, including NHH, ISB, CBS BIGFI, regional programmes, benchmarks, and conferences, remain in the academic finance source note and the coverage ledger. The academic queue now records papers, syllabi, datasets, doctoral advisers, seminars, and industry-affiliate language separately from firm disclosures. No university route is used to infer or rank a tracked manager.

September 4, 2026 — regional professors and finance programmes

The regional academic sweep added several routes for identifying model ideas and future personnel. UNSW’s Data and Algorithms in Trading course states that it is taught with Optiver and has students submit algorithms to an Optiver-built simulated marketplace. That is explicit partnership and simulation evidence, not real-capital access or evidence about any manager’s production system.

Seoul National University’s Statistical Learning & Computational Finance Lab names Jaewook Lee and describes financial forecasting, derivative pricing, portfolio optimisation, risk, generative and simulation-based financial intelligence, and trustworthy-ML concerns such as robustness and verification. HUFS’s Kong Hyungwoo profile names research in financial graphs, synthetic financial data, fairness, AML, and investment behaviour, alongside projects involving Korean financial institutions. These pages establish academic and project routes; they do not establish a tracked fund’s use, model permissions, data rights, or performance.

Queen Mary’s MSc Finance and Machine Learning curriculum, USP’s finance-ML courses, and UCT’s thesis index add text and LLM analysis, asset pricing, optimisation, financial time series, volatility forecasting, market-state clustering, portfolio construction, and fraud networks. These syllabi and thesis pages are discovery surfaces for methods, students, supervisors, and papers. They are not evidence that any named firm has adopted the work.

The academic source note records the full regional URL set and the extraction queue: recover papers, code, recordings, theses, adviser lineages, project partners, data vintages, splits, and cost assumptions before promoting an idea into a backtest or linking it to a manager. The route map remains descriptive rather than ranked.

September 4, 2026 — further academic and conference routes

John Cartlidge’s University of Bristol profile connects a Financial Engineering Lab to AI, agent-based simulation, market microstructure, automated trading, limit-order-book modelling, BondBERT, and trusted ML for asset and credit managers. The page also identifies UKRI-funded research and named supervision routes. This is academic and grant evidence, but it does not establish a tracked firm’s system or investment performance.

The Princeton Conference on Asset Demand Systems adds a research route around portfolio holdings, asset-demand identification, machine learning on holdings data, out-of-sample evaluation, and fixed-income, currency, and derivatives applications. NUS’s 2026/27 quantitative-finance curriculum adds data science in quantitative finance, AI and FinTech, financial time series, portfolio selection, data engineering, NLP, networks, and internships.

ICSF 2026, ICAIF ’26, and the Monash Q Group Finance Colloquium add conference routes covering foundation models, synthetic data, financial decision support, robust control, asset pricing, trading, portfolio strategies, automated stock recommendations, and financial reinforcement learning. These agendas are useful for recovering papers, speakers, and recordings; they are not evidence of any speaker’s employer system.

The source note and ledger now include these routes with geography, language, source type, discovery route, and evidence boundaries. No academic route is used to infer or rank a tracked investment manager.

September 4, 2026 — finance-AI conference bridges and professor-led idea routes

The Georgia Tech AI and the Future of Finance conference is a useful public personnel bridge. Its official 2026 programme lists Sudheer Chava and Agam Shah for an agentic-AI presentation; a practical asset-management panel with Gerardo Rodriguez (BlackRock Systematic), Vladimir Zdorovtsov (Acadian), Mahmoud Hajo (PIMCO), and Linus Marco (Wellington); and a hedge-fund panel with Charlie Flanagan (Balyasny), Vaibhava Goel (Millennium), Mike Moreau (Schonfeld), and Ramit Sawhney (Tower Research). The page establishes the event, date, advertised topics, and those role associations. It does not establish any employer’s model, dataset, vendor relationship, permissions, or investment use.

The University of Cincinnati Johnson Investment Institute AI Symposium adds a professor-and-practitioner discovery route. Its May 2026 agenda names Dacheng Xiu for AI/ML limits in investment, Clifton Green for AI-model extrapolation of stock returns, Winston Dou for AI in financial markets and systemic risk, Alejandro Lopez-Lira for AI and trading strategies, and Andrew Robinson for systematic investing at BlackRock. The same agenda includes speakers from JPMorgan, S&P Global, Clarivate, and Cincinnati Insurance. This is dated programme evidence, not proof of an employer’s production system or a speaker’s unpublished work.

The SoFiE 2026 programme provides a concrete set of research questions for the finance-AI queue: expected returns and large language models; lookahead bias in LLM forecasts; transformer CoVaR using textual information; and structured knowledge joined with data. The listed authors include Yifei Chen, Bryan Kelly, Dacheng Xiu, Zhenyu Gao, Wenxi Jiang, Yutong Yan, Junyu Chen, Lingwei Kong, Weining Wang, Tom Boot, Yi Cao, Zexun Chen, Lin William Cong, and Heqing Shi. The immediate diligence implication is to separate forecast leakage, systemic-risk text, and structured-knowledge tests. A conference listing does not establish a tradable signal or manager use.

The ABFER–JFDS Conference on AI for Finance adds an Asia-Pacific route co-organised by CKGSB, HKUST, Tsinghua PBCSF, SUSTech, USTC, ABFER, and JFDS, with the 2026 event dated August 18–19 in Hefei. Its public page describes AI applications in finance, investment, fintech, and capital-markets research, but currently exposes limited programme detail. The source is therefore logged as a discovery route pending paper, author, and recording recovery.

Lancaster’s Financial Technologies and AI MSc is a programme-level route with unusually specific finance applications: text and sentiment analysis, quantitative trading, high-frequency and algorithmic strategies, neural-network risk models, NLP for fraud detection, market-manipulation identification in high-frequency data, and backtesting controls for overfitting and selection bias. It names Bloomberg and Datastream as teaching databases and includes a research dissertation. This describes training and curriculum, not a named firm’s deployment or performance.

The University of Chicago Financial Mathematics ML concentration, Sofonias Alemu Korsaye’s Johns Hopkins profile, and the Emory Finance Lab add separate academic routes around ML for finance, market frictions and asset-pricing methods, and real-market-data simulation of dynamic trading and portfolio risk. They are useful for paper, professor, and talent discovery; they do not establish a tracked manager’s model ownership, data rights, live authority, or returns.

The underlying academic source note records each route’s evidence class and the recovery queue. Academic ideas will be promoted to a backtest only after the target, horizon, data vintage, point-in-time controls, split, transaction costs, and portfolio mapping are documented.

September 4, 2026 — professor networks and finance-programme idea supply

The academic route is expanding beyond isolated course pages. NYU Stern’s SoFiE lecture archive exposes a financial- econometrics speaker network, while its quantitative-finance specialization lists forecasting, volatility, credit risk, portfolio management, financial data science and AI/ML, and systematic trading. NYU’s Spring 2026 course schedule also exposes machine learning, NLP, generative models, strategy evaluation, and a named instructor. These pages are academic and talent evidence; they do not show a tracked manager’s deployment.

Berkeley Haas MFE electives add deep learning, reinforcement learning, generative AI, and standard and alternative text, image, video, and audio data. Berkeley’s Applied Finance Project and employer-research route describe faculty-supervised projects and confidentiality agreements; the surfaced examples include news-language methods and an industry contact. The MFE advisory board provides additional personnel and firm-affiliation leads, including contacts associated with Fabric, Kepos, CPP Investments, Two Sigma, Morgan Stanley, and BlackRock. An advisory affiliation is not evidence of research transfer or live use.

The Oxford ML research group, Stephen Roberts profile, and Oxford-Man research overview surface Bayesian and graph ML, reinforcement learning, NLP, limit-order books, execution, market making, uncertainty, causal inference, and multi-agent markets. The Deep Learning for Quant Finance Strategies project adds deep momentum, learning-to-rank, and transformer-ranking research. These published research routes can generate replication candidates, but reported backtests still require point-in-time data, cost, turnover, capacity, and split checks before any investment interpretation.

The open ABIDES market simulator and JPMorganChase ABIDES implementation provide code routes for exchange agents, latency, limit-order books, and reinforcement-learning environments. They are useful for execution and market- impact experiments; they are not historical data or evidence of live performance.

Princeton’s Data Science and Financial Technologies programme and ORFE ML research area expose efficient trading, high-frequency data, retrieval, AI, deep learning, and optimization routes. Jianqing Fan’s research page adds high-dimensional statistics, graph learning, generative AI, reinforcement learning, financial econometrics, asset pricing, risk, and portfolios. Yale’s International Center for Finance publication index adds papers on LLM-based mutual-fund-skill measurement and neural-network portfolio selection, while its public confidence series and historical securities endpoint are candidate data routes with licensing and vintage constraints.

The resulting idea queue is specific enough to test: news-language reaction; graph-based asset pricing; LLM lookahead-bias controls; transformer text-risk; fill-probability and market-impact models; RL execution in simulated books; deep momentum and cross-sectional ranking; generative asset pricing; and rare-outcome tests for data-driven screening. The academic source note keeps these research routes separate from firm disclosures and records the evidence boundary for each.

September 4, 2026 — direct academic-to-market bridges and title-blind routes

Columbia Engineering’s Ali Hirsa profile is a direct personnel route. It lists Hirsa as director of Columbia’s MS Financial Engineering programme, Chief Scientific Officer at ASK2.ai, and Managing Partner at Sauma Capital, described on the page as a New York hedge fund. The biography also records prior roles as Global Head of Quantitative Strategy at DV Trading and Partner/Head of Analytical Trading Strategy at Caspian Capital. His stated research areas include algorithmic trading, machine learning, deep learning, data mining, optimisation, computational finance, and AI applications in asset management. The profile establishes publicly reported affiliations; it does not establish which methods the firms use, their data sources, or any performance.

David Lariviere’s University of Illinois profile adds an inspectable educational and code-discovery route. It identifies him as founder and director of the university’s FinTech Lab, with affiliations across the MS Financial Engineering, ISE, and Finance programmes. His public teaching archive links high-frequency-trading coursework to agentic-AI projects and earlier projects to AWS/Wireshark market-data capture, PCAP-to-backtester and OneTick loading, virtual-reality order-book visualisation, IEX matching-engine speed, NASDAQ feed analysis, and FPGA options pricing. It exposes LinkedIn, GitHub, GitLab, and Google Scholar routes. These are public course and code surfaces; student projects do not establish a fund’s production technology or live-market access.

Iowa State’s Alan L. Zhang research site is another title-blind route. Zhang identifies his research as AI in finance, investment, corporate finance, and FinTech. His 2026 publication “Generative AI and Asset Management” describes a measure of investment-company GenAI reliance and a study of fund outcomes. The site also lists work on machine-readable corporate disclosure, manager risk assessment from textual disclosure, greenwashing funds, and private information in fund-manager narratives. It links papers, SSRN, LinkedIn, X, conferences, and media coverage. The descriptions are researchers’ stated methods and findings, not independent evidence that a named manager uses the measure or that the result survives point-in-time replication. Sample construction, paper appendices, and code/data availability remain recovery tasks.

SFSU’s Yi Zhou profile adds an international and climate/alternative-data route. It lists finance research across asset pricing, derivatives, credit, volatility, weather risk, and AI/LLMs, and names “Using Generative AI to Predict the Weather Impact on Future Stock Returns” (2025) and “Weather Risk and Financial Markets” (2026). The biography records UCLA PhD training in finance, statistics, and economics, a Berkeley astrophysics master’s, and a public academic website with faculty interview videos. This is a concrete weather-to-equity research route, but the profile does not establish production use, data vintages, geographic coverage, or out-of-sample trading results.

UCLA Anderson’s MFE pages add a programme-to-project route. The programme identifies Lars Lochstoer’s “Advanced Financial Data Analytics and Applications of AI,” an Applied Finance Project in which faculty-supervised student teams solve problems for companies or organisations, an industry advisory board, and collaboration with the Fink Center for Finance. The faculty page exposes additional faculty biographies and practitioner-linked teaching content. The current 2026–27 course description is unusually concrete: it says students use LLMs for data analysis, model development, and task automation; implement Kalman, state-space, and hidden-Markov models as forecasting benchmarks; study neural methods and implied-volatility surfaces; and examine how LLMs are built, tuned, and deployed for signal generation, text analysis, and automation. These statements describe curriculum, not a manager’s production system, data rights, or investment results. These pages identify a structured employer and research surface; they do not disclose confidential project outputs, data permissions, or manager deployment.

NYU Shanghai’s Volatility Institute 2026 conference call is an upcoming conference route dated December 4, 2026 in Shanghai. Its theme is AI and financial markets, with calls for work on GenAI/LLMs in asset pricing, sentiment, portfolio optimisation and forecasting; AI-driven market microstructure and execution; hallucination, herd behaviour, cyber and operational risk; and capital- markets governance. It lists Robert Engle, Bryan Kelly, Lin William Cong, and a scientific committee spanning NYU, Fudan, Shanghai Jiao Tong, HKUST, Xiamen, and UT Dallas. This is a programme and speaker-recovery route; the call does not establish acceptance, recordings, firm affiliation, or investable results.

The academic queue now includes specific implementation ideas: measuring GenAI reliance from public disclosures; manager narrative risk assessment; machine-readable corporate disclosure; packet-level market-data and order-book research; weather and climate variables joined to fundamentals and returns; post-trade allocation; and AI-induced liquidity, execution, operational-risk, and governance studies. Each remains separate from firm disclosures and requires point-in-time data, leakage controls, out-of-sample splits, costs, capacity, and data-use review before backtesting.

September 4, 2026 — published university project briefs and additional finance-AI routes

The USC Capital One Center for Responsible AI and Decision Making in Finance (CREDIF) adds a project-level academic route rather than only a course or lab label. Its public research agenda covers noisy and drifting financial data, structured synthetic data, explainable multi-agent LLM systems, graph knowledge representations, and robust GenAI evaluation. The 2026–27 call for proposals specifically describes work on data quality, privacy, multi-agent explanation, graphs, and LLM evaluation. This is evidence of a joint USC–Capital One research agenda, not evidence of a hedge fund’s model or production use.

The published 2024–25 projects include conformal uncertainty for imperfect financial datasets, knowledge-graph RAG for multi-hop financial questions, differential-private synthetic text and in-context leakage audits, graph anomaly detection with LLM reasoning, and a multimodal trust-and-emotion dataset. The 2025–26 projects add latent-confidence probing for hallucination detection, topological analysis of GraphRAG, multimodal reasoning over stochastic differential equations and American options, human-interruptible multi-agent dialogue, gaze-based causal imitation learning for forecasting and trading, cost-per-accuracy Mixture-of-Experts serving, and symbolic-reasoning benchmarks. Named PIs and fellows are listed on the project pages. These briefs reveal concrete data structures, evaluation problems, and research outputs; they do not establish live trading, proprietary-data rights, or performance.

The 2026 USC fellows page provides an additional personnel and paper-recovery route, listing Duygu Nur Yaldiz, Ke Xu, Yuxin Yang, Mohammad Shahab, Tejas Srinivasan, Yuan Xia, Yutai Zhou, and Shaoyu Wang. The page alone does not establish each fellow’s adviser, employer, or project ownership, so those links remain a verification queue.

Cornell’s AI in Finance programme adds a practitioner-connected quant-finance route. The page describes predictive and generative AI, finance tools and data, and applied exercises. It names Victoria Averbukh, Director of Cornell Financial Engineering Manhattan, whose biography records fixed-income research at Salomon Brothers and structured-MBS research at Deutsche Bank, and describes CFEM’s Future of Finance conference and industry colloquium. Vera Chau is listed as a visiting assistant professor with a Chicago Booth finance PhD. This is programme and personnel evidence, not proof of a fund’s use of course content.

Cambridge Judge’s finance-technology and AI curriculum connects machine learning to trading, asset management, accounting, auditing, operations, and quantitative asset allocation. It names Andrei Kirilenko, founding director of the Cambridge Centre for Finance, Technology and Regulation, and external lecturer Alejandro Reynoso, whose public biography combines MIT economics training with algorithmic trading and machine-learning-for-finance teaching. This is an academic and talent route; it does not show a manager’s implementation.

UC San Diego Rady’s Fabrizio Ghezzi profile adds a title-blind research route: Ghezzi is listed as a finance postdoctoral scholar under Allan Timmermann, studying real-time information extraction from complex datasets with ML/AI and teaching AI in financial applications. Emory’s Tucker Balch profile provides a career bridge across Lucena Research, JPMorgan AI Research, and academic finance. It publicly describes work spanning machine learning, cryptography, multi-agent simulation, high-frequency electronic markets, and synthetic data, and records Georgia Tech and UC Davis computer-science training. These biographies establish career and research facts, not current ownership or deployment of any method.

Duke’s David Ye profile surfaces a finance programme whose title does not require an AI-lab label. Ye leads the Quantitative Finance Concentration in Duke’s Master of Interdisciplinary Data Science and teaches algorithmic trading, risk and derivatives, AI in Finance, and the impact of LLMs on finance business models and work. His profile also records prior senior risk roles at Nomura and State Street. The FSU Truist Beach Conference 2026 brochure adds Andreas Neuhierl’s ML/finance research route and Kuntara Pukthuanthong’s work using ML, image analysis, LLMs, and long-run text in international asset pricing. Conference and faculty pages establish stated research fields and affiliations, not manager adoption or investment outcomes.

The academic queue now includes specific implementation ideas: uncertainty under noisy labels and drift; privacy-preserving synthetic data; graph retrieval and topology-based abstention; multimodal reasoning over stochastic dynamics and options; human interruption and saliency supervision; efficient MoE serving; and long-run text/image/LLM representations in international markets. Each remains separate from firm disclosures and requires point-in-time data, leakage controls, out-of-sample splits, costs, capacity, and data-use review before backtesting.

The Illinois route also produced a recoverable public repository artifact. The IE421 HFT project README describes a C++ pipeline from Databento PCAPs through Wireshark, exchange-specific parsers, OneTick PRL/TRD schemas, and Strategy Studio. It covers CME MDP 3.0, Nasdaq ITCH, NYSE Pillar, Cboe ADAP, IEX, L2/L3 order books, packet-level issues, and loader limitations. The preserved extract records the retrieval hash and technical boundary. The upstream GitLab notice says VPN or campus access will be required from October 5, 2026 and the service is scheduled for retirement in May 2027, so the linked student projects are now a concrete preservation queue. This is educational project evidence, not evidence of production trading, proprietary-feed access, or performance.

September 4, 2026 — NBER research-conference bridges

The academic route now reaches the NBER conference network. The NBER–SAIF Conference on AI and Financial Markets was scheduled for June 15–16, 2026 and lists James M. Poterba and Yongxiang Wang as organizers. Its page links research on AI asset-pricing models and AI in financial regulation. Because the page was searchable but returned an access denial to direct retrieval in this pass, this entry is limited to official page metadata and a recovery route; it does not establish a recording, firm participation, or an investable result.

The NBER conference on Big Data, Artificial Intelligence, and Financial Economics is scheduled for November 13, 2026 in Cambridge and is organized by Itay Goldstein, Tarun Ramadorai, Chester Spatt, and Mao Ye. Its call covers large unstructured datasets, AI methodology and applications, market decision-making, market conduct, regulation, and accountability, with an optional submission route to the Review of Financial Studies. This supplies a professor and paper-discovery route for unstructured financial data; it does not show which papers will be accepted, who will attend, or whether any method is deployed by a manager.

The NBER Summer Institute Financial Market Structure agenda provides a paper-level route. The July 18, 2026 programme includes “The Limits of AI Trading” by Winston Wei Dou, Itay Goldstein, Zigang Li, and Liyan Yang, along with papers on multi-asset market making and the price of exchange data. The agenda links slides and names researchers and discussants from Stanford, Cornell, NYU, HEC Paris, Oxford, MIT, Berkeley, and the Federal Reserve. These are routes for recovering papers, slides, and academic lineages; they do not establish live trading, a fund relationship, or performance after costs and capacity.

The idea queue from these academic sources is now separated into four testable questions: strategic AI effects on information aggregation, liquidity, competition, and stability; incremental information from unstructured corporate, regulatory, and market text; changes in execution, adverse selection, market making, and the economics of exchange data; and measurement of AI-related regulatory and operational risk. Each needs its own point-in-time data, split, cost, capacity, and licensing record. Academic presence is a discovery signal only, not evidence of a tracked firm’s deployment or a tradable result. The full source records and recovery boundaries are in the academic finance source note.

September 4, 2026 — Boston and Stanford programme/repository routes

The MIT Sloan Master of Finance curriculum adds an operational programme route. Its Analytics of Finance options cover financial econometrics, dynamic optimisation, derivative pricing, machine-learning methods, structural information extraction, and applications in portfolio management, risk, derivatives, algorithmic trading, and fintech. MIT’s Finance Lab description says student teams work on practitioner-defined problems and that partners can include investment managers, hedge funds, private equity, venture capital, risk, and consulting organisations. This establishes curriculum scope and an industry-project mechanism, not the identity of confidential partners, student work product, data permissions, or production deployment.

Stanford’s Advanced Financial Technologies Laboratory profile for Markus Pelger and his teaching archive expose research and training routes around financial risk, high-dimensional statistics, machine learning in empirical asset pricing, high-frequency statistics, and financial-data analytics. The profile records UC Berkeley economics training and organising roles in AFTLab and the AI & Big Data in Finance Research Forum. These are academic lineage and research-vocabulary signals; they do not establish a current fund relationship or a deployed model.

Luyang Chen’s Stanford research page links public code, data, and papers on deep learning in asset pricing, mortgage-risk prediction, and optimal execution with unknown volume limits. The linked asset-pricing repository contains comparison notebooks and README tables for GAN, feed-forward, linear, and regularised-linear models. The mortgage-risk repository contains an older neural-network implementation and sensitivity-analysis scripts. The preserved artifact note records retrieval hashes and the reproducibility boundary. The repositories are inspectable research routes; their historical metrics are source claims until the sample, costs, capacity, data rights, and environment are independently rebuilt.

Boston University MET’s Hanbo Yu profile adds a Boston personnel and programme route. It describes prior energy-market analysis and asset-management work, investment-department data science, and teaching in Python/SQL, enterprise risk analytics, deep learning, and applied business-analytics capstones. It also describes research on signal interpretation, beliefs, information frictions, macro-finance, and reliable AI agents. This is self-described institutional biography evidence; it does not establish a current hedge-fund role, a validated trading result, or adoption by a named manager.

The idea queue now includes practitioner-defined finance projects, structured and unstructured information extraction, economic constraints inside learning objectives, high-dimensional asset pricing, public deep-learning code, execution under incomplete market information, mortgage and credit risk, and signal interpretation with agent reliability. The full evidence table and recovery steps are in the academic finance source note.

September 4, 2026 — Stanford applied-AI research routes

The official Stanford SITE 2026 session on Applied Artificial Intelligence in Macro-Finance adds two concrete, title-blind research routes. A paper presented by Bradford Levy and Ralph S.J. Koijen frames evaluation of agentic asset-pricing systems around look-ahead bias and market reflexivity: models trained on historical data can be contaminated by future information, while adoption can change the patterns being measured. The programme says the authors introduce a real-time, out-of-sample benchmark for agents that extract structured signals from earnings calls and optimize over them, with an SDK and open competition. Its reported explained- variation comparison remains a programme claim; the page does not provide the SDK, full data, permissions, or an independent reproduction. It is therefore a benchmark-recovery route, not evidence of a deployable fund system or live returns.

The same session lists “A Financial Brain Scan of the LLM”, presented by Hui Chen with researchers from the University of Melbourne and ESSEC. The abstract describes probing LLM representations to map economic forecasts to concepts such as sentiment, technical analysis, and timing, then steering outputs toward different risk preferences while holding other factors constant. This is a separate interpretability and scenario-control route from ordinary sentiment classification. The programme does not establish forecast improvement, portfolio control, manager adoption, or investment performance.

Kay Giesecke’s Stanford teaching archive adds a project mechanism alongside the paper routes. MS&E 446 has teams build a financial-technology project, research paper, or prototype and present work through the AI in Fintech Forum. MS&E 246 covers ML-based credit, market, mortgage, asset-backed-securities, derivatives, and systemic-risk analytics using real-data case studies. MS&E 444 describes market-data-rich projects co-developed with industry partners. These pages establish courses and project surfaces; they do not identify confidential sponsors, reveal student outputs, or establish production use.

The resulting recovery queue is explicit: obtain and audit the asset-pricing benchmark and SDK; reproduce the LLM concept-probing and steering design with frozen prompts and held-out finance tasks; and recover public Stanford course projects, posters, syllabi, and sponsor metadata. Each task needs model and prompt versions, data vintage, split, leakage controls, costs, capacity, and licensing before it can inform a hedge-fund research question. The full records are in the academic finance source note.

September 4, 2026 — Harvard and Boston finance-programme routes

Boston University’s QST MF 815 Advanced Machine Learning Applications for Finance provides a finance-programme route with an explicit methods list: financial-data features, deep learning, supervised learning, clustering, classification, reinforcement learning and optimal control, text mining, asset allocation, backtesting, and strategy risk. This establishes curriculum scope and a syllabus- discovery route. It does not establish the quality of a method, a student project’s result, or its use by a manager.

Irena Vodenska’s Boston University profile adds a professor and programme-leadership route. It describes research in financial networks, systemic-risk propagation, intraday pricing, news streaming, NLP, neural networks, and deep learning, and links work on shared-portfolio networks, sovereign-bank interconnectedness, global assets, and responsible ML in fintech. These are stated research interests and publication routes; they do not establish a current fund relationship, proprietary data, or live deployment.

The Harvard Law School / Program on International Financial Systems programme on the impact of AI on the global financial system is scheduled for February 2–4, 2027. Its public curriculum spans machine learning, deep learning, generative and agentic AI, adoption across financial institutions, AI in trading and market intermediation, supervisory use, explainability, and accountability. This is an upcoming executive-education and speaker-discovery route; the programme does not establish participant attendance, an employer’s system, or performance.

The Harvard Business School Finance Unit research index adds a cross-domain investment-data route. A July 2026 entry describes algorithms that identify potentially deprioritized clinical-stage drug-development programmes from multiple microdata sources to catalogue shelved assets for alternative development. The page reports 5,523 candidate programmes, but that figure and the underlying labels require independent checks of sample construction, dates, and false-positive rates. This is a concrete healthcare-investment data route, not evidence of a hedge-fund model, investment decision, data licence, or independent validation.

The recovery queue now includes the BU syllabus and instructor history, Vodenska’s linked papers and datasets, Harvard Law programme speakers and recordings after February 2027, and the clinical-stage asset paper, appendix, code, and data definitions. Each route should retain source vintage and legal status before being connected to a fund or backtest. The detailed records are in the academic finance source note.

September 4, 2026 — Yale, Columbia, Durham, and Connecticut routes

The SoFiE Financial Machine Learning Summer School at Yale adds a professor-and-programme route spanning high-dimensional finance, double descent, benign overfit, deep neural networks, financial NLP and LLMs, factor pricing, stochastic discount factors, portfolios, alternative data, CNNs, AI pricing theory, and transformer asset-pricing models. The page identifies Bryan Kelly as a Yale finance professor and head of machine learning at AQR, and Dacheng Xiu as a Chicago Booth professor working on ML solutions to big-data problems in empirical finance. It also names guest researchers from UCLA, Princeton, Chicago, Purdue, Wharton, and Washington University. This is dated academic-programme and personnel evidence, not evidence of live portfolio use.

Columbia’s Wall Street Voices adds a recurring industry-academic archive. Its 2026 listing names Eric Jaffe of BCA Research, Yin Luo of Wolfe Research, and FeiFei Wu of Macquarie Asset Management for an “AI in Investment Management & Research” session. The archive also lists Atlas Ridge Capital and PGIM Quantitative Equity personnel in a separate hedge-fund interview session, plus earlier data-and-analytics sessions. The page establishes speakers and topics; it does not establish recordings, internal systems, datasets, model ownership, or performance.

Durham’s FINN41615 Financial Modelling with Artificial Intelligence module combines ARMA/ARIMA, structural breaks, VARs, ARCH/GARCH volatility, supervised ML forecasting, deep learning for asset returns, and cross-sectional asset pricing. Its learning outcomes explicitly include critical evaluation of AI limits and an empirical project. The module is a syllabus and talent route, not evidence of an industry sponsor, proprietary data, or production results.

The 2025 University of Connecticut Finance Conference programme lists Bryan Kelly’s AQR machine-learning role, a keynote on model complexity, “Generative AI and Asset Management,” and “AI-Powered (Finance) Scholarship,” alongside work on trading heuristics and asset pricing. It is a dated programme record; it does not establish model deployment, evaluation quality, or fund use.

The next recovery targets are the Yale lecture materials and guest papers, Columbia recordings and speaker biographies, Durham’s full syllabus and project archive, and the papers behind the UConn AI sessions. Each should retain model and prompt versions, data vintage, point-in-time splits, licensing, costs, capacity, and independently verified employment links before informing a fund-level claim.

September 4, 2026 — Oxford-Man, Oxford AI in Finance, and Singapore routes

The Oxford-Man Institute events archive is a durable conference-discovery surface. It lists the June 2026 Machine Learning and Finance Conference, the October 2025 Big Data and Finance Workshop, the June 2025 Machine Learning in Quantitative Finance Conference, and a March 2027 Finance and AI event. Older entries include NLP for economic and financial modelling, AI and financial markets, algorithmic collusion, and machine-learning workshops. The archive supplies dates and event links; it does not by itself supply recordings, full papers, or evidence of fund adoption.

Oxford-Man’s research overview describes a multimodal and agentic research agenda: combining tabular financial data with text and audio, extracting signals from social media, news, central-bank statements, analyst reports, and filings, and studying interactions among autonomous agents, prices, liquidity, collusion, and bubbles. This is an institutional research description, not evidence of a manager’s data licence, production system, or investment result.

The Oxford Computer Science Artificial Intelligence in Finance activity adds a lab/activity surface covering asset pricing, risk analytics, systemic-risk modelling, autonomous financial AI, and governance of digitally connected financial systems. Its linked Algorithmic Markets route is a follow-up path for papers, people, and seminars; the activity page itself does not establish commercial deployment.

Mihai Cucuringu’s Oxford Statistics and Machine Learning in Finance page names concrete project classes: news-sentiment propagation in financial networks; deep asset pricing with news and technical factors in Chinese equities; limit-order-book and price simulation with GANs; order-flow and cross-impact models; graph-based asset pricing; option-volume signals; volatility and co-volatility neural networks; and SEC-filing classification of non-bank financial institutions. These are group project and publication leads, not evidence of live fund use or validated returns.

The 2026 Five-Star Asia Pacific Workshop in Finance at Singapore Management University adds an Asia-Pacific route. Its AI/ML session lists “Limits To (Machine) Learning,” “AI ‘Errors’,” and “Asset Pricing Using KAN,” with participants from NTU, SMU, and related institutions. The wider programme includes investor-information, hedge-fund, bond, and macro-finance research. The page exposes highlight images, but no complete recording or underlying code was located in this pass.

The next recovery targets are the Oxford-Man event pages, papers, slides, recordings, and speaker rosters; Oxford’s Algorithmic Markets and SMLFin people and publication links; and the SMU paper PDFs, authors’ code, and any posted video. Each route should retain source date, affiliation type, language, data rights, and the distinction between research-topic visibility and production use.

September 4, 2026 — Oxford-Man linked recordings and speaker rosters

The Oxford-Man Machine Learning and Finance Conference 2026 page exposes direct MP4 replay links rather than only a programme. It names David Hirshleifer, Siew Hong Teoh, Stefan Nagel, Ansgar Walther, Shuang Chen, Carol Alexander, Jesús Gorrín, and Patrick Chang and links recorded talks for several of them. This is a recoverable media corpus for local ASR and paper-to-talk alignment; it does not establish that a listed academic’s method is used by a fund. A command-line range smoke test returned HTTP 202 with a small HTML response rather than video bytes for sampled MP4 URLs, so these are verified publisher-linked assets awaiting local capture.

The Oxford-Man 2025 Machine Learning in Quantitative Finance page similarly links replay URLs for Andrew Lo, Ruixun Zhang, Shumiao Ouyang, Mihail Velikov, Lin Peng, Xuedong He, Xiao Xiao, Valentina Raponi, and Yuantao Shi, with slides exposed for some talks. It also provides a bridge to MIT, Peking University, Oxford, Penn State, CUNY, CUHK, Cambridge, IESE, and the Oxford-Man Institute. The page is a speaker/media and research-lineage source, not evidence of investment deployment. Sampled direct URLs returned the same HTTP 202/HTML response and remain browser or alternate-delivery recovery targets.

The Oxford-Man 2022 Artificial Intelligence and Financial Markets event contains concrete idea and personnel metadata. Its listed talks include algorithmic learning equations and tacit collusion in dynamic games; data-sharing and platform competition; boosted-tree selection of mutual funds from holdings; NLP over job postings and AI-driven job redesign; and machine-learning analysis of high-frequency stock returns and durations. Named speakers include José Penalva, Simon Mayer, Alberto Rossi, Matthias Qian, and Jianqing Fan. These are academic abstracts and biographies; reported findings remain source claims pending paper, code, data, and point-in-time replication.

The October 2026 Oxford-Man Financial Econometrics Workshop page adds a current speaker-roster route: Carsten Chong (HKUST), Federico Bandi (Johns Hopkins), Kim Christensen (Aarhus), Oliver Linton (Cambridge), Qiyuan Li (Hong Kong), Viktor Todorov (Northwestern), Cecilia Mancini (Verona), Seok Young Hong (NTU), Peter Reinhard Hansen (UNC), Roberto Renò (ESSEC), and Shifan Yu (Oxford-Man). It is scheduled for October 22–23, 2026; the page exposes a paper link for one speaker but no complete recording in this capture.

The next recovery targets are the permitted 2025 and 2026 MP4s, their captions or audio, and the linked papers, slides, code, and data definitions. Any transcript should retain talk-clock timestamps, model family, data modality, sample period, licensing, and the distinction between research visibility and production use.

September 4, 2026 — Professor-led finance ideas and programme routes

The Oxford-Man news archive adds a useful discovery layer for professor-led finance research. It links machine-learning conferences, an AI-and-financial-markets workshop, an ICMA reinforcement-learning keynote, the Risk.net collusion podcast, and research articles on model complexity and automated market making. The most relevant material is often indexed under a professor, event, or market-microstructure topic rather than under “hedge fund AI.” The archive is a discovery route, not evidence of a manager’s adoption.

Several linked ideas are unusually concrete. OMI’s complexity discussion maps the debate around Bryan Kelly, Kangying Zhou, and Semyon Malamud’s return- prediction research against criticisms from Álvaro Cartea, Qi Jin, Yuantao Shi, and Stefan Nagel. The useful research design is a capacity/data-quality experiment: compare simple, nonlinear, and high-capacity models under point-in-time splits, noisy-data perturbations, and economic controls. The page mentions a 12-month exercise with 12,000 parameters, but that is a description of the academic debate, not a claim that complexity produces a live fund result.

The recent OMI publication on decentralised finance and automated market making describes stochastic-control strategies for large-position execution and statistical arbitrage using competing-venue prices, stochastic liquidity, and Uniswap v3 data. The page says the authors run consecutive in-sample estimation and out-of-sample liquidation/arbitrage experiments. This suggests a distinct research lane around liquidity-aware execution and AMM microstructure; it does not establish a live crypto mandate, a named fund’s use, or independently audited returns.

The ICMA reinforcement-learning keynote identifies Professor Álvaro Cartea and links a YouTube recording. The Risk.net interview adds a market-integrity route around ML-powered trading algorithms, granular ETF and FX data, possible signalling, reward/punishment dynamics, and regulatory accountability. Its public index gives listening points at 00:00, 07:57, 13:50, 20:21, 30:53, and 37:40. These are source-attributed research and risk discussions; they are not findings about a named hedge fund.

The Graphcore/Oxford-Man case study and arXiv paper expose a public research-to- compute route: Zihao Zhang and Stefan Zohren’s DeepLOB-Seq2Seq and DeepLOB-Attention models forecast multiple limit-order-book horizons, using FI-2010 and LSE order-book data, with Graphcore IPU acceleration. Graphcore reports the speed result and says Man AHL considered commercial applicability. The public repository is a reproducibility lead. These materials do not disclose Man AHL production deployment, live permissions, model weights, data rights, transaction costs, or investment performance.

For the wider academic map, the MIT Sloan AI in Finance syllabus shows how a finance programme translates the space into lending, investing, market-making, risk, domain-knowledge-informed learning, generative AI, and algorithm fragility. MIT’s Hui Chen, Stanford’s Advanced Financial Technologies Laboratory, Wharton’s AI in Finance Lab, and Columbia’s financial-economics curriculum provide complementary professor, lab, and talent-pipeline routes. They reveal research questions and training surfaces, not proprietary fund systems or a league table. The detailed records and disqualification fields are in the academic finance source note.

September 4, 2026 — additional university labs and finance-programme routes

The second academic pass found several routes that add distinct institutions and methods. The UCL AI for Finance Lab describes liquidity-risk estimation, risk profiling, risk management, AI validation, benchmarking, and financial-data analytics. It names collaborations involving the Saudi Central Bank, Santander UK, Consob, the UZH Blockchain Center, and the DLT Science Foundation. Its public publication list includes agent-based credit modelling and noisy-environment reinforcement learning. This establishes a lab and industry-collaboration surface, not a fund model or production result.

The University of Toronto RiskLab adds a separate Canadian training route. Its public site describes financial risk, banking, and trading research, a Machine Learning & AI track, and RiskLab-Prep training that combines live instruction, self-study, finance and mathematical foundations, and an oral assessment. This is a talent and programme signal; it does not identify a manager sponsor, proprietary data, or live trading use.

PolyU’s 2026 MSc in Quantitative Finance and FinTech lists algorithmic trading, investment science, derivatives, deep learning, machine learning in finance, statistical ML, and quantum computing. The catalogue names Yu Xiang, Jiang Zhaoli, and Selena Qian Yihe in programme-leadership roles and describes a dual-degree path with NYU Financial Engineering. The HKU course catalogue adds supervised, unsupervised, and reinforcement learning for optimal trading, asset management, portfolio optimisation, high-frequency trading, and crypto, plus topic modelling, policy search, alternative data, and performance evaluation. These are curriculum and recruiting surfaces, not evidence of a named firm’s use.

The University of Sydney School of Finance explicitly includes machine learning and artificial intelligence alongside asset pricing, market design, fintech, and financial regulation, and describes engagement with financial institutions and regulators. Freiburg Quantitative Finance adds model-independent valuation, robust calibration under uncertainty, credit and liquidity risk, and machine-learning methods; its current public news also points to deep-learning option calibration, robust credit mixtures, and deep-learning concentration-risk research. These pages identify research and talent routes, not proprietary systems, deployment, or performance.

The full URL set, faculty/programme extraction queue, and disqualification fields are in the academic finance benchmark and allocator source note.

September 4, 2026 — named university researchers, sponsors, and public artefacts

The deeper UCL pass found a personnel-and-project layer behind the AIRiskLab description. UCL names Ramin Okhrati as lab head, Aldo Lipani as an Associate Professor in Machine Learning, Raad Khraishi as a UCL PhD-trained quantitative researcher and NatWest Lead Data Scientist, Pin Ni as a deep-learning/NLP/knowledge- graph PhD researcher, Zihao Liu as an ESG/ML researcher, Viktor Kazakov as an ML expert at EBRD, and Daniil Bargman as a UBS Global Wealth Management CIO-office director and UCL PhD student. These are public institutional biographies; current employment and research ownership should be rechecked before drawing a firm link.

Two projects are especially useful for idea discovery. A NatWest-sponsored UCL project explores offline reinforcement learning for contractual pricing, learning a policy from historical data before any online fine-tuning. A Consob-sponsored project uses unsupervised methods to flag investor discontinuities and synchronised groups around Italian takeover bids, with a linked public paper. These are concrete sponsor/method/data-context signals, not evidence of a hedge fund’s deployment or investment returns.

The RiskLab Toronto ESG-scoring paper adds a dated public artefact with named authors Alik Sokolov, Jonathan Mostovoy, Jack Ding, and Luis Seco. It describes BERT and deep-learning NLP for converting social-media text into ESG scores, and discusses aggregated scores and semi- autonomous scoring systems. The paper’s biographies identify ML, research- partnership, doctoral, programme-director, and investment-company roles. It does not establish that a named hedge fund used the scores, nor does it provide an independently audited investment result.

The full personnel, sponsor, paper, hardware, and disqualification fields are in the academic finance benchmark and allocator source note.

September 4, 2026 — direct professor, programme, and finance-practitioner routes

The Chicago Booth Center for Applied Artificial Intelligence paper page adds a direct 2026 research artefact to the earlier conference route. Booth attributes “Assessing the Benefits of Optimized Agentic AI Systems for Asset Pricing” to Bradford Levy and Ralph S. J. Koijen. The page describes a real-time, out-of-sample earnings-announcement benchmark intended to control for look-ahead bias and market reflexivity. It says the evaluated agents extract structured signals from earnings-call transcripts, optimize over those signals, and are paired with an open SDK. Booth reports a publisher-stated increase in explained variation relative to standard benchmarks. The page does not expose the SDK, data licence, prompts, agent workflow, or independent replication, so this is a public research claim and recovery lead—not evidence of a named fund’s system, authority, or performance.

The Cambridge Judge MFin programme article adds an alumni-to-employer route. Cambridge profiles Weijie Li, a 2021 MFin graduate, as a BRAIN Researcher at WorldQuant working on quantitative research and training research consultants, and says the route into the role included WorldQuant’s Global Alphathon. The same page says Li used Cambridge machine- learning and Python teaching and worked with BlackRock London on back-testing trading algorithms as part of a graduate capstone project. This is Cambridge’s public programme/alumni account; it does not establish current employment beyond the publication context, the capstone’s data or code, WorldQuant portfolio use, or an investment result. It is a concrete talent-pipeline lead for recovering alumni, capstone sponsors, and practitioner transitions.

The LSE Data Science Institute and International Association of Quantitative Finance event page records a June 23, 2026 public event titled “AI/ML in finance: advancing the future of financial markets.” The displayed roster includes Oxford-Man director Álvaro Cartea, RavenPack/BigData.com Chief Data Scientist Peter Hafez, Citi FX data-strategy leader James Hamp, LSE professor Luitgard Veraart, LSE deputy DSI director Johannes Ruf, and Premia Research principal Hilary Till. LSE describes short talks and a panel spanning research, data, risk, execution, market structure, and financial stability, and points to its YouTube archive for past events. The page does not provide a transcript, slides, model inventory, data rights, employer implementation details, or investment performance.

The full records and recovery boundaries are in the academic finance benchmark and allocator source note.

September 4, 2026 — additional faculty and programme routes with direct quant relevance

Ilias Filippou’s Florida State profile lists him as an Assistant Professor of Finance and Dean’s Emerging Scholar working on asset pricing, AI, machine learning, international finance, and macro-finance. Its presentation list includes “Unusual Financial Communication: ChatGPT, Earnings Calls, and Financial Markets,” shown at the 2024 Wolfe NLP and Machine Learning in Investment Conference and a hedge-fund-strategies conference, and “Improving Hedge Fund Returns Predictions: Dealing with Missing Data via Deep Learning,” shown at the 2026 Midwest Finance Association and other finance meetings. This is a faculty and presentation index, not the underlying papers or evidence that a fund uses the methods. The recovery targets are communication extraction, missingness mechanisms, deep-learning imputation, and separating predictive fit from investable return.

NYU’s Andrew Arnold profile identifies Arnold as an Adjunct Professor in Finance and Risk Engineering. His public biography describes a current Principal Applied Machine Learning Engineer role at Shopify and a prior Chief Scientist role at Oracle Alpha, where he says he led ML and NLP research and production for an emerging systematic fundamental hedge fund. It records Carnegie Mellon PhD training in Machine Learning and a Columbia bachelor’s degree in computer science and AI, plus teaching in NLP and ML for quantitative trading. His stated research interests target low-signal and low-sample settings, distribution shift, robust time-series models, and unstructured-data features. The biography does not name the fund, models, data, permissions, or outcomes; this is a former-personnel and academic-training route.

SKEMA’s Alexandre Landi profile lists Landi as a Senior Lecturer and 2025–2026 MSc Financial Markets & Investments programme director, with prior Lead Data Scientist work at IBM France and quantitative-research roles at Balanced Research and Brevan Howard. The profile dates the Brevan Howard role to 2022 and describes mid-frequency trading models and data-driven strategies. It also lists working papers on walk-forward- optimized FX pairs trading and an LSTM versus PatchTST comparison for asset-price prediction, alongside Georgia Tech coursework in ML for trading and knowledge- based AI. This does not establish which methods were used at either manager or any live performance.

KIT’s Chair of Financial Economics and Risk Management identifies Prof. Dr. Maxim Ulrich and separates an AI Finance group from the C-RAM group. The page describes mixture-density neural networks for equity-return distributions, reinforcement learning for robust portfolio decisions, millisecond- resolution option data, a multi-year option-signal dataset, and NLP analysis of central-bank communication. It also links these activities to the KABFI doctoral training group, KIT’s Computational and Data Science graduate school, and Financial Data Science and Advanced ML/Data Science teaching. The public AI-Finance page and team page name current postdoctoral and PhD researchers, including Joytirmayee Behera, Ardalan Azarnejad, Vishalini Balakrishnan, Mahsa Hajlotfalia, Yao Huang, Zhuojia Li, Jakob Maisch, Asen Mariov Karakanovski, Giang Vo, Alexander Walter, Xiaolong Wu, Hanqin Ye, Ziwei Zhang, and Lukas Zimmer. The KABFI PhD programme page further says the programme has HPC, tick-by-tick option analytics, and full order-book data for Xetra, Eurex, and EEX via Deutsche Börse Group, and announces 4–6 month research internships with Deutsche Börse and Allianz Global Investors culminating in a jointly developed research paper. These are academic lab, named-personnel, and programme-partnership clues, not evidence of a named fund’s deployment, data licence, or performance.

Johns Hopkins Carey’s Sudip Gupta profile lists Gupta as Professor of Practice, formerly faculty and director of Fordham’s MSQF programme, where he introduced big-data and ML in finance. His displayed research includes alternative credit scoring with big data and ML, alternative- data ESG ratings and portfolios, and nowcasting with alternative data and ML; current teaching includes Machine Learning for Finance. The profile records a Wisconsin economics PhD and consulting/advisory roles. It is a faculty CV-style source, not evidence that a named asset manager uses the listed models or that a paper produced investable performance.

These routes add five idea families to the recovery queue: missing-data-aware hedge-fund return prediction; robust NLP and time-series modelling under low signal and distribution shift; mid-frequency validation with walk-forward controls; millisecond option-implied density and tail-risk modelling; and alternative-data credit, ESG, and nowcasting. KIT adds jointly supervised academic–industry work on options and exchange order books with explicit market-data provenance. The underlying papers, code, data definitions, recordings, and doctoral lineages remain separate recovery tasks.

September 4, 2026 — additional international professor and finance-programme routes

The RiskLab IISc faculty and lab page adds an India-based quant-finance research route spanning derivatives pricing and hedging, XVA, high-dimensional portfolios, and point-process models of high-frequency order flow. It also lists an annual Recent Trends in Quantitative Finance symposium, current postdoctoral and doctoral researchers, and placements into Morgan Stanley, HSBC, and Wells Fargo. The linked public papers cover neural Hawkes-process estimation and order-flow forecasting, event-time quoting, order-book filtration, and robust path-dependent-option hedging (quoting, order-book filtration, and robust hedging). The page names Saurabh Bansal with neural-network option-pricing solvers, Sumanjay Dutta with low-sample portfolio analysis, and Aditya Nittur Anantha with event-time order-flow modelling. This is evidence about an academic lab, its public research surface, and listed placements; it does not show that any tracked fund uses these methods, has access to the data, or grants them live portfolio authority.

The University of Queensland Digital Finance Research Hub capability statement adds an Australia-Pacific route focused on AI and ML in capital-market pricing, trading, risk management, and information flow, alongside cyber-risk stress tests, algorithmic decision-making, AI governance, digital assets, and training. It names Min Zhu and Sergeja Slapnicar as co-leads and describes intended partnerships with government, regulators, financial institutions, and technology firms. This is a research-hub and partnership surface, not proof of a hedge-fund model, a partner dataset, production use, or investment performance.

The University of Reading ICM520 module record shows a 2026/27 Henley Business School module titled “Machine Learning, Artificial Intelligence, and Big Data in Finance.” The record lists 20 credits, Level 7, Semester 2, and Dr Mininder Sethi as convenor; its aims cover predictive financial modelling and AI/big-data application at business-problem, complex-project, and whole-organisation levels. This is a curriculum and talent route only. It does not identify student projects, employer partners, proprietary data, production systems, or fund deployment.

These additions expand the idea queue toward event-time liquidity and quoting, neural option pricing and robust hedging, low-sample portfolio construction, cyber-risk and algorithmic-governance controls, and enterprise-scale finance projects. The next evidence step is to recover papers, code, data definitions, conference recordings, faculty lineages, and authorized industry-partner material; academic or programme evidence remains separate from any manager’s deployment.

September 4, 2026 — allocator, family-office, and professor-programme routes

The discovery surface now includes capital owners and finance programmes that publish AI clues without using hedge-fund vocabulary. These are comparator and talent routes, not additions to any tracked manager’s capability record.

The Berkocorp family-office site identifies the Canadian office as the investment office of the Berkowitz family, with Vancouver roots and Toronto operations, investing family capital through fund commitments and direct investments. Its linked Mantle Mondays episode identifies Joshua Berkowitz as Managing Principal and says the February 9, 2026 conversation covers a one-person investment-office workflow and daily use of Mantle, Attio, and ChatGPT. The source exposes a concrete small-office tool surface; it does not establish permissions, model evaluation, decision authority, or performance.

The University of California Investment Office FY2024–25 annual report describes AI as informing risk monitoring, performance analysis, and anomaly detection in pension flows. It also says the office co-founded an AI Futures Lab with UC Berkeley professors Ken Goldberg and Fernando Pérez through the Berkeley Institute for Data Science. This is an unusually direct allocator-to-professor partnership disclosure, but it remains a first-party report: model architecture, data rights, evaluation logs, and investment-level attribution are not public there.

The University of Kentucky Investment Committee minutes name CIO Todd D. Shupp and Investment Director Nancy Rohde and record a June 12, 2025 AI discussion. Rohde described ways the Investment Office had used AI in workflows; Shupp connected hardware, hyperscalers, learning models, applications, and infrastructure to portfolio exposure. The minutes also record a Tola Capital AI presentation. This is governance and allocator evidence, not a named internal model or live trading claim.

Brown’s FY2025 Endowment Report contains a dedicated “AI Implementation” section. Brown says the Investment Office is examining, with investment partners and the Brown community, how generative AI could change the investment model. It separately describes digital infrastructure within real assets as a way to participate in generative AI growth. Neither passage identifies a model, data source, implementation, permission structure, or AI-attributed result.

Two title-blind Capital Allocators episodes add institutional-technology routes. The January 20, 2026 Washington University Investment Management Company episode identifies CTO David Xiaoxi Li and discusses a modular Snowflake-centered stack, governance, security, “walled gardens,” data quality, and practical AI pilots. The January 6, 2026 Yale Investments episode identifies CTO Bill Krueger, including a prior 13-year Bridgewater history, and discusses vendor partnerships, infrastructure modernization, and pragmatic AI in an allocator context. Both publisher pages mark transcripts as premium, so the metadata establishes episode identity and stated scope, not the full remarks or any model inventory.

The Commonfund episode “Responsible Investing in the Age of AI” published August 10, 2026 features George Suttles, Amanda Novello, and Georges Dyer of the Intentional Endowments Network. The publisher transcript separates AI questions across investment-office use, manager use of AI and data, and portfolio companies’ AI development, while also surfacing governance, reputational, environmental, workforce, and systems-level concerns. It is a network-level allocator framework, not evidence about a named fund’s system.

The Milken Institute’s “A Practitioner’s Guide to AI in the Investment Office” was an invite-only May 6, 2025 panel with Jason Klein of Memorial Sloan Kettering, Ned Brines of Arnel & Affiliates, Roben Dunkin of PGIM, Jonathan Larkin of Columbia Investment Management, and Umesh Subramanian of Citadel. The public page scopes discussion of internal use cases, decision-making, talent, risk, cybersecurity, and data integrity. It supplies a speaker and topic route, not a transcript or evidence of implementation.

The University of São Paulo EAD0830 course page lists Leandro dos Santos Maciel for a 2026 course on AI and ML applied to finance, including regression/forecasting, optimization, clustering, and quantitative financial-data work. USP’s MAP5922 graduate catalogue entry lists Christian Dieter Jakel and explicitly covers supervised learning in asset pricing, cross-sectional pricing, investor-belief formation, computational projects, and a research agenda. These are Brazilian professor and curriculum routes; they do not establish a fund’s data access, deployment, or performance.

Together these routes add concrete idea and recovery surfaces: small-team AI operations; allocator-to-university lab partnerships; AI value-chain exposure mapping; Snowflake-centered investment-office data governance; and ML research training in forecasting, asset pricing, optimization, and clustering. The full records and disqualification boundaries are in the family-office source note and the academic route note.

September 4 — professor and finance-programme idea routes

The academic programmes add useful context for interpreting public manager signals. They expose the ideas and skills being taught around quantitative investment without proving that any named manager uses them.

• Duke / Campbell R. Harvey: the 2026 Finance 656 syllabus places AI and machine learning in asset management alongside risk, factor identification, drawdown control, due diligence, quantitative stock selection, and a final research project. The syllabus also publicly states Harvey’s Research Affiliates and Man Group roles. This is curriculum and affiliation evidence, not evidence of a particular firm’s model or performance.
• Columbia / Ali Hirsa: the Summer 2026 IEOR E4737 directory covers fund, manager, and security selection, asset allocation, risk, fraud, climate finance, real-estate data, ML/deep learning, explainability, adversarial ML, resilience, and industry utilization. It is a direct route to the professor, syllabus, projects, and guest speakers; it does not show which methods an employer uses.
• Stony Brook AMS 520: the course page joins Bayesian models, Gaussian processes, nonlinear factors, autoencoders, HMMs, particle filters, RNNs, and deep reinforcement learning to traditional quantitative finance. The instructor is not named on the public page, which is itself a recovery gap.
• Illinois Gies: the 2026–27 finance catalogue exposes an idea sequence: causal analysis and reproducible workflows; ML, options, portfolios, and fraud; replication and overfitting tests; market microstructure; financial data management; Python projects; and AI applied to SEC filings, sentiment, discount rates, and risk.
• Stevens / Zonghao Yang: the FA690 page combines deep learning, asset pricing, risk, and portfolio optimization with embeddings, Transformers, RAG, fine-tuning, and LLM-agent workflows. Its visible schedule is dated 2025, so current offering status is unresolved.
• NYU: the Courant profile for Christos Koutsoyannis links data-driven modelling—regularization, PCA/SVD, cross-validation, Python, and public web scraping—to a profiled quantitative-investment practitioner. Separately, the 2026–27 Tandon bulletin lists ML applications to asset management, trading strategies, weight optimization, and risk management. Neither page establishes live deployment.
• International routes: the Monash 2026 handbook links financial ML and statistical learning to industry placement or research projects; the CUHK Finance brochure lists ML in finance, Python, financial econometrics, and AI applied to investments. These create Australia-Pacific and Greater China talent-recovery paths, not firm-level deployment evidence.

The resulting idea map is broader than news sentiment: causal and replication design, nonlinear factors and latent states, sequential decision models, microstructure and execution, document extraction, and agentic research workflows. Each requires separate point-in-time data, leakage controls, and an out-of-sample test. The source ledger records the underlying pages and their limits; no firm is ranked on the basis of academic curriculum.

September 4 — research labs and professor-led idea routes

The academic layer adds methods that public manager pages often omit. These are research and talent signals, not evidence that a named firm has deployed them.

• MIT’s 15.S06 AI and ML Research in Finance, maintained by Hui Chen, publishes projects on embedding-based political risk from earnings calls, simulated limit-order-book reinforcement learning, millisecond price-impact prediction, volatility-aware execution, and model interpretability. The page reports project-level results, but those results need paper/code recovery before being treated as reproducible signals.
• Princeton BCF’s Master’s Research Project page lists a practical finance-research environment with commercial databases and project topics including Transformer NLP, regularized momentum, Bayesian and LSTM option pricing, FX volatility, and regression-tree futures strategies.
• NJIT’s Ajim Uddin research page links graph-based asset pricing, dynamic networks, tensor completion, earnings forecasting, financial sentiment, bond prediction, and portfolio optimization to public code and named student research routes.
• Iowa’s Ashish Tiwari profile focuses on ML-based hedge-fund benchmarks, nonlinear time-varying exposures, Bayesian factor sparsity, and controls for false positives in alpha evaluation. This is a measurement and attribution route, not a claim about any manager’s internal system.
• USC’s CREDIF center documents a joint USC–Capital One AI-in-finance research center with doctoral fellowships, workshops, a symposium, a 2026–27 proposal call, and Petros A. Ioannou as director. The page does not disclose funded project methods or production deployments.
• Texas A&M’s Kangying Zhou profile provides a named professor route across asset pricing, investments, ML, and NLP, with publicly stated Yale and Chicago training. Monash’s staff directory adds faculty routes tagged with ML, HFT, market impact, hedge funds, media in finance, and earnings management.

The most useful new idea clusters are embedding-based risk measures, graph and tensor finance, microstructure and execution, Bayesian false-positive control, nonlinear factor models, and research infrastructure linking commercial data, open code, faculty supervision, and industry guests. They broaden the search program beyond sentiment while preserving the distinction between academic evidence, personnel lineage, and disclosed manager deployment.

September 4 — finance-AI researchers, labs, and model-evaluation routes

The professor and programme search surfaced additional public evidence about the research agenda around quantitative finance. These sources describe academic work, lab infrastructure, or public commentary; they do not establish that a named manager uses the methods.

• Boston College’s Miao Liu profile describes human–AI corporate valuation, LLMs and Transformers for filings and earnings calls, and autonomous AI research agents for hypothesis discovery in a human-designed lab. It also provides a Wuhan–SUNY Buffalo–Columbia–Chicago education route.
• Auburn’s Stace Sirmans research report describes controlled tests of 48 AI models on investment questions with biased and neutral framings, exposing framing, anchoring, narrative, and loss-aversion effects. The linked paper and exact model versions remain the verification path.
• Washington University’s Songrun He dissertation combines high-frequency jump risk, reasoning-LLM news narratives, chronologically consistent language models, and crypto-perpetual-futures benchmarks. It is particularly relevant to look-ahead-bias controls.
• Tulane’s Landon Ross profile describes small language-model analogues for estimating future stock returns and other financial information, alongside robust optimization and NLP research.
• Columbia’s Paul Glasserman publication page links LLM chronology, GPT-sentiment look-ahead bias, overnight news/returns, stress testing, options, and simulation in one research surface.
• Maryland’s Smith research page describes graph-based conversational modelling of earnings calls, using topic novelty, cross-references, sentiment, and GNNs for financial-risk forecasting. It also lists Sean Cao’s work on AI analysts, alternative data, disclosure, and human–machine complementarity.

The new model families and research questions are: agentic hypothesis discovery; chronological language models; behavioral stress tests of financial AI; graph representations of earnings calls; compact domain models; and ML benchmarks that separate factor exposure from manager skill. Each remains an idea or academic evidence route until the underlying paper, data provenance, point-in-time split, and independent result are recovered.

September 4 — additional professor and academic-media routes

The academic sweep also recovered several public sources that are useful for interpreting personnel, model, and media signals. They are not evidence of any named manager’s internal deployment.

• Boston College’s Miao Liu profile describes LLMs and Transformers for filings, earnings calls, and corporate text; human–AI corporate valuation; autonomous research agents for hypothesis discovery; and a Wuhan–SUNY Buffalo–Columbia–Chicago education route.
• Auburn’s Stace Sirmans report describes controlled tests of 48 AI models on investment questions with biased and neutral framings, reporting framing, anchoring, narrative, and loss-aversion effects. The exact prompts and model versions remain to be recovered.
• Washington University’s Songrun He dissertation combines high-frequency jump risk, reasoning-LLM news narratives, chronologically consistent language models, and crypto-perpetual-futures benchmarks, making it a useful route for leakage-control research.
• Tulane’s Landon Ross profile describes small language-model analogues for estimating future stock returns and other financial information, alongside robust optimization and NLP.
• Columbia’s Paul Glasserman research page connects LLM chronology, GPT-sentiment look-ahead bias, overnight news/returns, stress testing, options, and simulation.
• Maryland’s Smith research page describes graph-based conversational modelling of earnings calls using topic novelty, cross-references, sentiment, and GNNs for risk forecasting, alongside Sean Cao’s work on AI analysts, alternative data, and human–machine interaction.
• The Tippie “Can AI Beat the Market?” episode is an example of title-blind discovery: a university research story exposed an otherwise easy-to-miss podcast route to Ashish Tiwari’s hedge-fund benchmark work, with YouTube and Spotify links.

These routes add model-evaluation, chronology, human–AI collaboration, graph conversation, compact-model, and benchmark-attribution surfaces. They also improve the discovery process: university news pages, dissertations, faculty pages, and course-linked media can reveal relevant episodes without hedge-fund or AI terms in the title. The ledger records the recovery gaps and keeps academic evidence, personnel lineage, and manager deployment separate.

September 4 — additional college and finance-programme routes

The university layer is useful for mapping the idea supply around quantitative investment. It should be read as a research, curriculum, or talent signal. It does not establish that any manager uses a method, owns a dataset, or earned a return from it.

• Chicago undergraduate AI-in-finance pipeline: the University of Chicago AI in Finance Program describes separate “Traders Learning Tech” and “Techies Learning Trading” cohorts. The two groups receive targeted training, then build an AI-driven portfolio together. The page also says participants create public GitHub portfolios and connect with industry through visits and a trek. This is a direct route to student projects, alumni, and recruiting signals; the page does not name the projects or show their results.
• North Texas finance-ML route: Stephen Owen’s UNT profile spans financial machine learning, U.S. and international asset analysis, portfolio optimization, cryptocurrencies, high-performance computing, and textual analysis. It also records a graduate Python and Data Analytics for Finance course, Penn State finance doctoral training, and machine-learning-in- asset-pricing study at Chicago’s Stevanovich Center. The public page creates a professor, course, and lineage route; it does not identify an investment-firm relationship or live model.
• Stevens FinTech and market-structure route: a Stevens seminar page for Jingrui (Victoria) Li connects AI, LLMs, asset pricing, derivatives, investments, and market microstructure. Its research summary describes cryptocurrency pricing across 80 exchanges from 2019 to 2023 and variation in spreads by venue type and domicile. The reported numbers need the underlying paper, fee assumptions, execution constraints, and point-in-time venue checks before being used in a backtest.
• Chicago asset-pricing and model-risk route: Stefan Nagel’s research page links a machine-learning-in-asset-pricing textbook with work on real-time anomaly discovery, high-dimensional factor timing, weak signals, arbitrage limits, and expectations. This is useful for testing capacity, model risk, and discovery discipline alongside return prediction. The page does not establish a manager’s implementation or performance, and individual working-paper metadata should be rechecked because the page was intermittently unavailable during automated retrieval.
• Cornell behavioral-AI route: Lawrence Jin’s research page lists work on model-free and model-based learning in investor behavior and on LLM biases and corrections, alongside asset pricing and behavioral economics. The associated research direction treats financial AI as something to measure and correct under controlled experiments, not simply as a sentiment generator. It is a model-governance and human–AI decision route, not evidence of a fund’s use of the work.
• Conference author graph: the 34th Conference on Financial Economics and Accounting schedule lists a special session on the economics of LLMs and applications in finance and accounting. The programme names work on ChatGPT and corporate policies, LLM biases and corrections, and LLM-generated M&A exposures, with authors and a session chair. It gives the discovery process a paper, author, and conference graph beyond earnings-call sentiment; it does not provide code, recordings, or evidence of portfolio deployment.

The resulting academic idea surface now includes student-built portfolio projects, international asset analysis, crypto venue microstructure, factor capacity, behavioral corrections, corporate-policy extraction, and M&A exposure generation. These are routes for further paper, code, student, and employment recovery—not a basis for ranking firms.

September 4 — international finance-AI programmes and labs

The international university sweep adds research and talent routes across Europe, Asia, Australia, and the United Kingdom. These sources describe public curricula, laboratories, and academic work. They do not establish that any tracked manager uses the methods.

• INSEAD: the Master in Finance AI programme describes a 14-month France/Singapore curriculum covering generative and agentic AI for research, advisory, and client interaction, plus automation in trading, risk, and operations. It names Python, TensorFlow, PyTorch, scikit-learn, and cloud analytics, and lists projects in forecasting, fraud, credit risk, portfolio optimisation, algorithmic trading, and workflow automation. The page advertises an August 2027 intake but does not name student projects, instructors, employers, or live systems.
• HKUST: Dong Lou’s profile identifies him as Co-Director of the HKUST-DXM AI for Finance Joint Laboratory and Director of the Institute for Financial Research. It lists asset pricing, investment management, and behavioral finance, with Google Scholar and ORCID routes. The profile establishes a lab and researcher node; it does not disclose lab datasets, models, or industry deployment.
• RPI: Aparna Gupta’s profile connects Stanford and IIT Kanpur training with quantitative finance, risk, optimization, simulation, and ML for interconnected financial systems. It also lists NSF/DOE-funded financial-innovation projects and a 2026 paper on semantic-graph learning for trend prediction from long financial documents. The page does not show a hedge-fund implementation or tradable result.
• KAIST: Seunghun Shin’s profile lists asset pricing, asset management, corporate bonds, market microstructure, AI in finance, and big data. It also records Aalto and HKUST appointments after KAIST doctoral training, creating a cross-border personnel route into fixed income and liquidity research. It does not establish a fund relationship.
• HUFS: the Financial Analytics Lab is a Korean-language research surface organized around asset-price data, factor identification, explainable valuation, and financial-system analysis. It names Sihyun An and student researchers, and reports LLM-based synthetic sentiment work, a Bank of Korea–Naver AX conference route, and 2026 real-estate ML projects. The page does not establish live investment use or causal price impact.
• Glasgow: ACCFIN5230 teaches neural networks, evolutionary programming, meta-heuristics, and deep learning in finance through case studies and research papers. It is a reading- list and lecturer-recovery route; the public page does not name the current lecturer or show project outcomes.
• Southampton: the Financial Technology and AI MSc starts in September 2026 and combines AI in finance, computational finance, risk, responsible AI, Python, Bloomberg/LSEG data, model governance, and a dissertation. It names Manuel Nunes as course lead and says industry-linked projects may be feasible. It does not identify partners or imply portfolio use.
• Melbourne: the Hanqing Tian PhD seminar abstract adds LLM interpretability, predictable-content removal from financial news, cross-firm spillovers, and chronological text modelling. The page reports an academic Sharpe-ratio estimate, which requires the underlying paper, costs, multiple-testing treatment, and replication before reuse. It is not evidence of a manager’s signal or performance.

These routes extend the idea map into agentic finance education, Asian finance-AI labs, semantic graphs for long documents, corporate-bond and market-microstructure research, model-risk curricula, and chronology-controlled news signals. They are discovery and verification paths, not a ranking of firms.

September 4 — Canadian professor and industry-linked finance routes

The Canadian university sweep adds professor, programme, conference, and industry-funding routes that extend beyond generic earnings-call sentiment. These are academic or educational signals; they do not establish manager deployment.

• Guelph: Fred Liu’s profile spans AI, ML, financial econometrics, asset pricing, and risk management. It lists a 2026 deep-learning factor-timing paper, a billion-observation intraday-predictability project, courses in machine learning and AI in financial markets, government and industry partnerships for PhD projects, and conference routes including Wolfe’s Canadian Quantitative and Macro conference and Generative AI in Finance in Montreal. It is a route to papers, code, students, and media; it does not establish a fund’s use of the work.
• Toronto / iA Financial Group: this Rotman announcement records a C$600,000 five-year commitment to a quantitative-finance professorship held by Redouane Elkamhi, including asset pricing, risk, portfolio construction, and AI/ML integration. It is an explicit insurer–university funding route, not evidence of a hedge fund’s internal model or a specific investment result.
• Manitoba: Yu Xia’s profile provides McGill–Duke–Wuhan University training and research across asset pricing, financial institutions, and ML, with teaching in investments, fixed income, derivatives, and data analytics. No industry deployment is disclosed.
• Calgary: Adam Upenieks’s profile connects Bayes Business School and Waterloo training to ML in asset pricing. His public project uses Dow Jones Newswire intensity to study news-conditioned beta–return relationships. The design needs its paper, costs, timing, and replication before being used in a backtest.
• Ottawa: Adelphe Ekponon’s profile gives a HEC Montréal–Boston College–Toronto–Cambridge–Liverpool lineage and lists ML, FinTech, asset pricing, macro-finance, and crypto. It also records a 2024–2026 SSHRC grant on asset prices, business cycles, and ML and a MITACS “Qinvest Platform Engine” project. The platform’s implementation and data are not public on the profile.
• McMaster: Qian Yang’s profile covers empirical asset pricing, ML, NLP, alternative data, cyber risk, analyst disagreement, retail behavior, and LLM predictive capability, with Michigan State doctoral training. It is a research and personnel route, not evidence of a manager relationship.
• Toronto: Liyan Yang’s profile lists AI/ML, asset management, asset pricing, big data, China, and disclosure among public media topics; Cornell doctoral training and availability for industry projects and graduate supervision add collaboration routes. It does not identify a specific investment system.

These additions expose Canadian routes for news-conditioned asset pricing, semantic and deep-learning research, fixed income, crypto, cyber risk, retail behavior, public-institution funding, and university-to-industry project recovery. They add coverage without ranking firms.

September 4 — conference recordings and vendor-media recovery

The media pass converted several previously loose discovery leads into explicit capture records. The artifacts are useful for topic and personnel mapping; a public replay or vendor page is not evidence of proprietary deployment.

• Wolfe Research: the 2024 Canadian Quantitative and Macro programme names “AI Agents in Event Analysis” by Yin Luo, “Artificial Intelligence and Composite Systematic Strategies” by Sheng Wang, Javed Jussa’s Factor Awareness Dashboard, and Hallie Martin’s risk work. It adds named personnel and topic routes to the existing Wolfe search, but provides no recording or implementation disclosure.
• EDHEC replay cluster: the Future of Finance speaker archive exposes YouTube links for David Rapach on Shapley attribution of ML portfolio performance, Rama Cont on generative models for scenario simulation and hedging, Bryan Kelly on AI pricing and complexity, and Dacheng Xiu on weak signals in financial ML. The video URLs are now separately tracked, while transcript recovery remains open because automated video retrieval returned cache misses.
• Direct videos: Rapach, Cont, Kelly, and Xiu.
• Montreal: the Generative AI in Finance presentation is a separate November 16, 2024 YouTube artifact uploaded by Jesús Villota Miranda. Its public metadata has no description or transcript, so it is a recovery target rather than a content claim.
• Dauphine-PSL: the French-language “Finance et IA : repousser les frontières” replay describes a panel on AI and generative AI in finance, academic presentations, and a keynote on complex AI models. The page says the talks were in English; browser capture is still needed for speaker and video metadata.
• Boosted.ai: the 2024 GenAI-in-finance webinar names CEO Joshua Pantony and describes capital-markets use cases, financial chat, investment agents, API access, and institutional/professional-investor solutions. These are vendor claims and do not establish customer adoption or performance.
• Wolfe webcast: the Seven Sins of Investing with GenAI page confirms a first-party webcast surface, but the retrieved HTML contains no substantive abstract or speaker metadata. It is tracked as a recovery target, not as evidence of Wolfe’s internal AI strategy.

This recovery adds source types that were underrepresented in the professor search: event-page-to-video edges, vendor webinars, French regional wrappers, and named buy-side research topics. The article continues to separate public topic signals from verified implementation evidence.

September 4 — professor-led ideas and finance-programme routes

The academic layer adds useful research questions and personnel edges, but it is not a proxy for any firm’s internal system. Cornell’s AI in Finance keynote dates a November 21, 2025 practitioner event with Victoria Averbukh and Andrew Chin. Cornell describes Averbukh as Professor of Practice and director of Cornell Financial Engineering Manhattan, with prior Salomon Brothers and Deutsche Bank fixed-income research experience. It describes Chin as AllianceBernstein’s Chief AI Officer and Operating Committee member, previously Head of Investment Solutions and Sciences and, from 2022–2023, head of quantitative research and chief data scientist. The public programme focuses on selecting tasks for AI assistance, preserving human judgment, and integrating tools into workflows. The full keynote is gated after registration; no transcript, model inventory, data source, permissions, evaluation, or investment result is claimed.

Harvard Business School’s AI Institute recap describes a May 10, 2023 D^3 Assembly Talk moderated by Suraj Srinivasan, with Alexandra Mousavizadeh, Glenn Hopper, and Sanjay Srivastava. The recap surfaces natural-language and conversational interfaces, privacy and data-use controls, combining GenAI with conventional ML, CFO workflows, proprietary data, and continuous learning. It also summarizes a Jeff McMillan interview as discussing Morgan Stanley’s ChatGPT deployment and an OpenAI/Microsoft partnership. That partnership point is retained as a verification lead because the page is a secondary recap rather than a primary Morgan Stanley release. The recording area did not expose a transcript in retrieved HTML.

HEC Paris’s Christophe Pérignon profile adds an academic validation route: Professor of Finance, Deloitte Chair in AI for Business Innovation, Associate Dean for Research, and scientific co-director of Hi!PARIS. The linked work covers predictive-performance decomposition in credit scoring, fairness, computational reproducibility, stochastic LLM sentiment measurement in filings, AI-exposure stress testing, and AI-model audit. Those topics help define how public claims should be tested—measurement uncertainty, sensitivity to modelling paths, reproducibility, and stress—not whether a named hedge fund uses a given model. The profile says he advises banks, fintechs, and AI startups but names no fund or confidential system.

These routes expand the personnel graph and the research queue. Faculty roles, practitioner teaching, advisory statements, event participation, current employment, and firm deployment remain separate evidence types. No firm is ranked, and no academic or programme signal is treated as proof of production investment use. See the academic-finance source note and the media coverage ledger for capture status and disqualification boundaries.

September 4 — conference programmes that expose the next research queue

The expanded conference search found several routes whose titles omit hedge-fund and quant keywords but expose research ideas, evaluation designs, and personnel connections.

The official ICAIF 2026 programme schedules the ACM International Conference on AI in Finance for November 14–17, 2026 in Milan. Its workshops cover reinforcement learning for LLM agents, responsible-AI operations, AI-driven market microstructure, financial AI security and privacy, private-markets unstructured data, and workflow-level LLM evaluation. The competition page describes an autonomous trading-agent testnet with historical training data, held-out validation, and newly released live market data. The organizing committee also links academic, bank, vendor, and asset-management personnel. These are public programme and evaluation signals; they do not establish real-capital trading, production deployment, or performance.

The NBER–SAIF 2026 agenda lists “Artificial Intelligence Asset Pricing Models” by Bryan T. Kelly, Boris Kuznetsov, Semyon Malamud, and Teng Andrea Xu of AQR Capital Management. Other sessions address missing data, AI advising, AI at work in finance, browsing behavior, innovation, and regulation. The AQR affiliation is an author-network signal, not evidence that AQR operates the paper’s model or that the result is tradable after costs.

The UCLA Human × AI Finance experiment asked authors to use AI extensively while writing finance papers and required a machine-readable paper plus an AI-workflow description. Its UCLA Fink Center agenda includes research on unobservable lies, corporate silence, and signal homogenization and fragility in AI-intermediated markets. The site says AI agents selected four papers and that results and reviewer comments would be published; the retrieved page still showed those results as forthcoming. This is a rare public research-agent evaluation route, not evidence that AI reviewers select investable research reliably.

The Investment Association 2026 agenda adds a UK route titled “How Agentic AI Can Reinvent the Active Manager.” James King and d-fine managers Richard McIntosh and Kira Huneke are scheduled to show an AI-supported investment-research workflow and an AI-readiness framework. The agenda verifies the planned session and speakers, but not the demonstration’s tools, data, controls, or any firm’s production use.

The IIT Madras Pan-IIT finance conference material and conference report add an India regional route covering AI in finance, NLP for finance research, market surveillance, and keynotes by Avanidhar Subrahmanyam, Tarun Ramadorai, and Viral Acharya. The report states 650+ submissions and 376 selected papers; those are conference-reported figures, not independently audited counts. No paper, recording, dataset, or fund implementation is inferred.

Together, these routes add agent-evaluation protocols, academic-to-buy-side coauthorship, workflow demonstrations, and regional finance-AI coverage. They are retained as research and capture leads, with separate primary-source verification required for any model, dataset, partnership, personnel claim, or deployment conclusion.

September 4 — following programme titles into papers and regional seminars

The conference pass also recovered underlying research artifacts. The current SSRN version of Artificial Intelligence Asset Pricing Models shows a July 2026 revision and describes a transformer embedded in the stochastic discount factor, cross-asset information sharing, nonlinearity, and a linear-transformer interpretation layer. Its author block connects Bryan Kelly to Yale, AQR, and NBER and lists AQR for Teng Andrea Xu. The paper reports lower pricing errors against prior machine-learning models, but this remains a research result: it does not establish an AQR-operated model, live portfolio authority, or an after-cost trading result.

The Detecting Lies When Truth is Unobservable paper and Wenhao Li’s discussion-slide page make the UCLA Human × AI route more concrete. The paper describes downside-variance and total-variance compression statistics for misreporting, with examples involving funds and corporate earnings; Li’s page records public discussant slides for the UCLA Fink Conference. Labels, sample construction, statistical power, false-positive behavior, and replication materials remain open. The proposal is not presented as a validated deception detector or evidence of fund use.

The Sungkyunkwan University Korean/English seminar listing dates a hybrid June 20, 2026 seminar on Seongkyu “Gilbert” Park’s “Correlated Intelligence: Signal Homogenization, Price Discovery, and the Fragility of AI-Intermediated Markets.” Park’s research page also lists a related smart-limit-orders paper previously titled “Employing Artificial Intelligence to Create Smart Limit Orders.” This adds a Korean regional-language route connecting AI-signal commonality, market microstructure, and execution research. It does not establish a live fund system or empirical market effect.

The evidence chain is now more granular: conference programme, underlying paper, author page, discussant artifact, and regional seminar. Each edge remains separate in the academic-finance source note and coverage ledger.

September 4 — direct recovery of Stanford and QMUL academic routes

The Stanford Report feature now resolves directly. It describes Antonio Coppola and Matteo Maggiori using LLMs on more than 780,000 earnings-call transcripts and analyst reports covering more than 21,000 companies over more than a decade, with processing on Stanford high-performance computing clusters. The article points to a public Global Capital Allocation dashboard and says the LLM could identify geoeconomic pressure even when documents did not use “tariffs,” “sanctions,” or “export controls.” It reports that more than 30% of calls mentioned negative tariff effects and more than 60% of U.S. calls reported negative effects. This is a university account of a research workflow and findings; it does not establish a manager’s data rights, model versions, prompt design, dashboard access, or investment use.

The QMUL 2026 Finance and Machine Learning programme exposes a useful talent pipeline: corporate finance; asset pricing, trading, and portfolio construction; introductory machine learning; quantitative methods in R; big-data applications for finance; and large language models and textual analysis in finance. It also includes a supervised dissertation or research project, with possible external-organization activity. This is curriculum evidence, not evidence of student results, partner identity, production models, or manager adoption.

The NBER Working Paper 35431 route is now cross-linked to the Chicago Booth agentic-asset-pricing record. It identifies the July 2026 Koijen–Levy paper and its real-time out-of-sample earnings-announcement benchmark. The paper’s quantitative claims remain attributed to the authors’ research design and publisher summary, not generalized to any firm or treated as investment performance.

September 4, 2026 — professor-led controls and finance-programme artifacts

College professors and finance programmes are useful idea-discovery surfaces because they publish papers, syllabi, code, seminars, and named research lineages. They are not evidence that a tracked manager uses the idea. Three artifacts now deserve a separate recovery path.

The Chronologically Consistent Large Language Models replication package provides code, derived data, synthetic stand-ins, and checkpoint documentation for the ChronoBERT and ChronoGPT paper. The authors train on timestamped corpora and test a news-based asset-pricing application to address look-ahead bias. The full reproduction still requires licensed Dow Jones Newswires and CRSP; the public package is therefore a reproducibility route, not proof of a tradable result or any fund’s model.

The American Finance Association’s Common Task Framework paper by Oliver Hellum, Theis Ingerslev Jensen, Bryan Kelly, and Lasse Heje Pedersen proposes shared data, fixed metrics, and a leaderboard for comparing factor and machine-learning asset-pricing models. The author affiliations include Yale, Copenhagen Business School, NBER, and AQR. That creates an academic-to-practitioner research edge; it does not establish AQR deployment, sponsorship of a live contest, or performance of any submitted model.

The AFA paper on AI as conditional probability by Hui Chen, Antoine Didisheim, and Luciano Somoza studies token-level conditional probabilities in financial-news classification rather than relying on a model’s self-declared confidence. The proposed replication idea is to measure probability margins, reversal rates as more of an article is revealed, and calibration across model vintages. The paper is an academic result; it does not show that any of GMO, Acadian, Arrowstreet, or another manager uses the feature.

The expanded academic-finance source note now records the professor, programme, PI, code/data, and industry-connection edges separately. Future idea candidates will be admitted only with a dated forecast target, modality, data vintage, point-in-time split, costs, and a clear boundary between academic evidence and firm disclosure.

September 4, 2026 — programmes exposing agentic and multimodal finance ideas

The programme layer adds useful idea and talent routes without being mistaken for firm disclosure. Mannheim’s FIN 6080 course explicitly covers RAG, function-calling, agent-based systems, evaluation, financial document and earnings-call NLP, asset pricing, trading, risk, and generative AI in capital markets. Its final assessment can be an empirical paper with replication code and AI-workflow documentation or a finance MVP. This is a programme design and research-recruiting signal, not evidence that a named manager uses those systems.

Sydney’s FINC6028 unit lists portfolio optimisation, default prediction, factor strategies, option pricing, fraud detection, and risk management as AI application targets, with assignments and group projects. The Yonsei MLCF publication list adds Korean financial text-table agent evaluation, chart reasoning, dynamic-factor reinforcement learning, and social-signal research. Its chart-paper route is available through the DOI; the publisher abstract separates directional forecasting from cross-sectional ranking and reports market- and chart-representation dependence. Those are academic results and recovery leads, not evidence of hedge-fund adoption or live performance.

Chicago’s 2026–27 Financial Mathematics concentration lists machine learning for finance, high-frequency data, Generative and Agentic AI for Finance, reinforcement/deep learning, and modern optimisation. The listing is subject to change and does not expose faculty assignments, project code, or employer links. The expanded academic-finance source note records these programme, professor, modality, and artifact edges separately from any firm claim.

September 4, 2026 — new professor, programme, and agent-evaluation routes

The latest pass adds five source classes that help connect finance ideas to people, training programmes, and reproducible evaluation. Columbia’s AI Investment Mgmt course covers RAG, agentic frameworks, risk-factor extraction, news monitoring, stress testing, multi-agent research and trading, guardrails, and human oversight under instructor Cyril Shmatov. This is curriculum evidence, not evidence that a tracked manager uses those methods.

The Stanford Digital Economy Lab seminar with Chicago finance professor Suproteem Sarkar studies how agents shift work toward delegation, planning, and supervision, using Cursor data. Its linked recording is now a media-recovery target. The NYU AI talk adds Bhaskarjit Sarma, Head of AI Research at Domyn and former BlackRock Director and Principal Data Scientist, discussing evaluation-centric, risk-aware finance agents. Neither event establishes a named fund model, deployment, or investment result.

Two benchmarks extend the research side beyond market-prediction datasets. FINCH tests long-horizon, multimodal finance/accounting workflows across spreadsheets, PDFs, charts, and other artifacts. The Vals AI Finance Agent Benchmark describes 537 SEC-based questions across nine finance-research task categories, with expert input from banks, hedge funds, and private equity. FINCH’s evaluation and the Vals record are benchmark evidence; neither is evidence of a tracked firm’s internal system or performance.

The expanded academic-finance source note keeps each professor, programme, media, benchmark, and employer edge separate. The recovery queue now treats course guest rosters, linked recordings, repository history, and benchmark licenses as first-class evidence fields.

September 4, 2026 — finance curricula expose tools, tasks, and research controls

Several university surfaces now provide useful idea and talent signals. The University of Chicago FINM 33200 course names Claude Code, Cursor, OpenAI and OpenRouter APIs, LangGraph, vector databases, and MCP. Its stated applications include SEC-filing extraction, deep-research reports, connections to financial databases and analytical pipelines, and reinforcement learning for trading. Jeremy Bejarano and Mark Hendricks are listed as instructors, with weekly exercises and a final project. This is curriculum evidence, not evidence of any tracked manager using that stack.

Georgia Tech’s Fall 2026 PhD syllabus lists Sudheer Chava as instructor for “AI, Technology and Finance.” Students are asked to analyze an end-to-end research workflow—data creation and scraping, ETL, analysis, visualization, reporting, literature review, and hypothesis generation— and then specify where agents help or harm, what guardrails they require, and how they should be orchestrated. The syllabus also requires banking-data exercises and a final research proposal. It is a research-training route, not a firm disclosure.

The University of Illinois Gies account identifies Clinical Assistant Professor Tony Zhang and describes FIN 580 students using agentic and multi-agent systems to build trading platforms and analyze markets. The account supplies a project and teaching route for further capture; it does not provide a validated strategy, student performance result, named production system, or hedge-fund relationship.

The current Durham FINN41615 handbook entry updates the archived module already tracked. It retains the bridge from classical financial econometrics and volatility modelling to supervised ML, deep learning for asset-return forecasting, and cross-sectional asset pricing, with an empirical project as the full summative assessment. The current entry does not identify an instructor, dataset, sponsor, or live investment use.

The Johns Hopkins AI and Agentic AI in Finance programme is a Great Learning collaboration and is labeled here as provider-assisted professional education. Its public curriculum names earnings-call sentiment, SEC text, alternative data, RAG, embeddings, portfolio-risk monitoring agents, multi-agent RAG, model-risk governance, build-versus-buy, cost analysis, and Claude-based workflows using Claude Chat, Claude Code, and MCP. It advertises faculty-led masterclasses, industry mentorship, and hands-on projects, but does not identify all faculty, mentors, datasets, or production systems. The curriculum is a discovery lead rather than evidence of university or manager adoption.

The academic-finance source note and coverage ledger preserve these as separate professor, programme, tooling, workflow, and evidence edges. No cross-firm ranking is inferred from the presence or specificity of a course description.

September 4, 2026 — Asia-Pacific programmes and multimodal finance ideas

The NUS Business School finance-leader programme adds a Singapore route with named academic and Google Cloud-linked instructors. Its public description covers AI use-case prioritisation, ROI and data readiness, governance, prompt and workflow design, and prototypes for reconciliation, exception handling, and forecasting loops. It is executive education and does not establish a manager’s production architecture or investment results.

The University of Cincinnati Applied AI in Finance certificate lists a 12-credit sequence spanning financial econometrics, quantitative equity investing with factor construction and backtesting, fintech and cryptocurrency, algorithmic trading, and ML/AI applications. Its advanced course names news and earnings-call text, sentiment-based trading signals, Python, and real financial datasets. The programme page provides a training route; it does not validate student results, data rights, or any firm deployment.

A Purdue-hosted self-authored CV for Xinde Zhang claims an Applied AI in Finance MSF track, new AI-finance courses, a Walmart faculty partnership proposal, and a Purdue Office of Investments practicum involving Senior Investment Officer Sam Fehrman. Those are useful personnel and programme leads, but the document is self-authored; the programme and partnership details require independent confirmation.

The preprint Uni-FinLLM proposes a shared multimodal Transformer with modular heads for text, numerical time series, fundamentals, and visual data, with separate stock, credit-risk, and systemic-risk targets. It is a concrete architecture for cross-modal research, but the public record does not yet provide a full reproducible package or independent replication.

The recent paper Multimodal Visual Reasoning in Asset Pricing by Mengtao Chen, Yue Fang, and Hua Wang uses prompted multimodal models to read candlestick charts for the top 100 cryptocurrencies. Its abstract describes pre-specified seven- and 30-day strategies with monthly rolling-origin evaluation, and the paper discloses ChatGPT 4o use for trade-signal generation. This is a notable visual-modality research route, not evidence of hedge-fund use; prompt sensitivity, chart construction, turnover, costs, and independent replication still need examination before any transfer to international equities.

The academic-finance source note records these as separate programme, personnel, preprint, and peer-reviewed-paper edges. No firm ranking or adoption inference is made from them.

September 4, 2026 — regional programmes, agent-market experiments, and finance video

The HKUST 2026–27 GFIN 5600 entry describes an AI-in-finance course covering ML, NLP, generative models, investment strategies, and business analytics. It names implementation projects with ChatGPT coding assistance, current-event and industry cases, and evaluation of AI limits. This is curriculum evidence, not a manager disclosure.

The HKUST profile for Seth H. Huang connects finance faculty, Boston University and Cornell training, and a self-described practitioner role at Aris Capital. It states interests and claimed experience spanning algorithmic trading, HFT, market making, arbitrage, NLP sentiment, bond pricing, and rare-event prediction, as well as prior AI Applications Research leadership at Huawei. These are profile assertions; the public page does not provide a model, dataset, audit, or attributable performance record.

A Renmin University finance seminar records Pengfei Sui of CUHK-Shenzhen presenting “Dissecting AI Trading: Behavioral Finance and Market Bubbles” in English and Chinese. The abstract describes an experimental market populated by LLM agents, a 20-mechanism reasoning-text score, and prompt interventions in simulated bubbles. This is an academic market- simulation route, not evidence of a live strategy or fund deployment.

The CUHK-Shenzhen AI course page describes project-based quantitative-finance and AI classes covering portfolio optimisation, derivatives, volatility, statistical arbitrage, AI-assisted Python, backtesting, financial NLP, LLM statement analysis, reinforcement-learning trading simulations, multi-agent systems, and Data Spaces for access control and lineage. It provides a useful regional curriculum and tooling route without establishing student results, industry data access, or production use.

A Purdue-hosted self-authored CV for UBC professor Jan Bena lists a 2025–2027 grant as PI for adaptive learning and collaborative agents in finance education, plus 2026 Applications of AI in Finance teaching at UBC. It is faculty and programme evidence, but not evidence of a fund relationship or deployed system.

The NTU “What If” episode with Will Cong embeds Episode 1 and names Cong as President’s Chair Professor of Finance, Computing, and Data Science and GIFTS director. The page discusses HoloBit socio-economic simulation, Global InferenceNet, and related episodes on AI boards and human investors. These remain university-page statements and discovery leads; the page does not establish code, deployment, investment results, or hedge-fund use.

The academic-finance source note keeps regional language, academic affiliation, self-reported practitioner claims, video metadata, and firm evidence as separate edges. No cross-firm ranking is inferred.

September 4, 2026 — additional professors and finance-programme pipelines

The academic search also recovered five programme routes that are relevant to researcher lineage and talent-pipeline mapping:

• The CUHK-Shenzhen MSc in Finance describes an AI Finance concentration covering quantitative methods, open-source programming, machine learning, and generative AI in practical finance applications. It names Dan Li (York University Ph.D. in Finance) as programme director and identifies Cong Wang and Yongxiang Wang among the faculty. This establishes a curriculum and faculty route, not a fund system or investment result.
• Durham’s 2026–27 FINN3071 module combines time-series finance, statistical arbitrage, asset-pricing models, prediction/classification/risk-regime tasks, interpretability, slippage, and model comparison with classical approaches. It explicitly allows agentic AI coding assistants with documentation and attribution. Its assessment is an empirical student assignment; the handbook does not identify a fund partner or live-trading result.
• Duke’s Quantitative Finance Concentration joins data science and mathematics, with ML, AI, and decision optimisation across markets, banking, and insurance. Its 2026 offerings name David Ye, Hengzhong Liu, Hanchao Yang, and Yimin Yang across algorithmic trading, AI in finance, risk and decision optimisation, crypto trading, and financial-crime detection. Industry speakers and projects are part of the stated design; employer-specific systems are not disclosed.
• PolyU’s BSc in Financial Technology and Artificial Intelligence spans software engineering, security, AI, ML, cryptocurrency, e-finance, and big data. It requires at least 312 hours of work-integrated education and names Wu Xiao-Ming, Cheng Ran, and Fong Chi Kit Ken as programme leaders. This is evidence of a Hong Kong talent and placement pipeline, not evidence of any employer’s production models.
• Fudan International School of Finance’s FMBA curriculum places “Machine Learning and AI in Finance” beside hedge funds, derivatives and risk, fintech, behavioural investing, and big-data analytics. Its compulsory LIVE project, industry mentors, summer internship, and thesis create routes for recovering project artefacts and employer links. The page does not name the AI course instructors, partner firms, tools, or resulting systems.

These are academic and talent-pipeline signals. They should be used to recover faculty papers, student projects, guest speakers, internships, and research collaborations—not to infer that a named manager has adopted a course or that any academic project produced live alpha. The academic-finance source note keeps those evidence classes separate.

September 4, 2026 — Chinese finance-AI laboratories and model-research personnel

The Chinese-language search also recovered three lab and personnel routes that expose model and research vocabulary beyond generic “AI in finance” programme labels:

• Shanghai University of Finance and Economics’ AI Finance Laboratory describes a pipeline spanning financial corpora, the Fin-R1 finance-reasoning model, FinAgent finance agents, and FinEval evaluation. It also names reported collaborations with Shanghai AI Laboratory, KuPas, HSBC, Huawei, ICBC, Zhipu, and Orient Securities. These are university-reported lab and partnership statements; they do not establish data rights, independent benchmark results, partner scope, or live investment authority.
• The SWUFE profile for Qing Li identifies him as director of the Sichuan Provincial Key Laboratory of AI & Digital Finance. His listed work covers financial risk early warning, graph neural networks, multimodal news-driven stock prediction, option pricing, and web-media effects. The page lists PI projects and papers including lead-lag prediction, graph learning, option pricing, and multimodal event-driven LSTM methods. Publication counts and impact language remain self-reported and are not treated as a ranking or performance result.
• The SAIF profile for Jian Guo lists research interests in deep and reinforcement learning for financial time series, execution, quantitative investment and risk, financial knowledge graphs, large sequential models, and hybrid cognition systems. It describes his IDEA, HKUST Guangzhou, and Tsinghua affiliations and names Digital Economy Brain and Quant4.0 as projects. These are profile claims and research leads, not proof of a deployed large model, portfolio permissions, or attributable returns.

This adds lab, corpus, model, evaluation, agent, partner, and personnel fields to the regional map. The underlying repositories, technical reports, papers, and Chinese-language recordings remain the next verification targets. No firm ranking or adoption inference is made.

September 4, 2026 — Taiwan and African finance-ML faculty and programme routes

The regional academic sweep also found four additional routes:

• National Chengchi University’s Bingzheng Luo page lists financial machine learning, quantitative investment, financial econometrics, and earnings quality. It also lists 2026 research on Taiwanese risk premia and Transformer distillation for financial time-series forecasting, plus PI projects linking accounting knowledge and analyst scoring to ML. The page’s institutional and publication metadata should be checked against the underlying papers.
• National Kaohsiung University of Science and Technology’s Xuan-Qi Su profile identifies a professor and director of a cross-disciplinary intelligent-finance research centre. It lists AI quantitative investment, portfolio management, and asset pricing, with a National Taiwan University finance Ph.D. The profile does not disclose models, datasets, code, or live results.
• National Taiwan University’s Jau-er Chen profile identifies a Cathay Financial Holdings Elite Professor with an NYU economics Ph.D., prior MIT visiting-scholar experience, and research in causal ML and applied econometrics. His 2026 course listing shows a multilingual causal-ML course with data handling and programming. This is a research and training route, not evidence of fund adoption.
• The University of the Witwatersrand’s MCom Financial Technology programme is coordinated by Prof. Moinak Maiti and includes big-data analytics and ML, quantitative research, crypto, portfolio management, investment, derivatives, and financial risk. Its structure includes a substantial research report and a January 2027 intake. It does not disclose student datasets, employer projects, or production investment systems.

These routes expand the language and geography coverage of the academic map while keeping faculty biography, course design, paper claims, and firm deployment as separate evidence classes.

September 4, 2026 — India and European practice-linked finance programmes

Five additional professor and programme routes add specific research ideas and practitioner lineage:

• Raunaq Pungaliya’s FLAME University profile connects asset management and AI-in-markets research to a University of Iowa finance Ph.D., prior SKK AI in Business Innovation Lab co-direction, and the published “Machine Invasion” paper on automation and stock returns. It also lists an MIT Sloan fellowship and RBI CAFRAL visiting-scholar role. This is a paper and biography route, not evidence of fund adoption.
• Arun Gupta’s Ahmedabad University profile lists Yale financial-economics, Carnegie Mellon, and Berkeley EECS training and research interests in machine learning and agentic AI for finance, structured finance, reinsurance, valuation, and M&A. “Agentic AI for Finance” is listed as work in progress; the profile does not expose a system or result.
• Francesco Rotondi’s Bocconi profile lists machine learning for pairs trading using clustering, hidden-Markov-model statistical arbitrage in international crude-oil futures, and teaching in Python, computational methods, and ML for finance. His listed lineage runs from Padova mathematics through Bologna/LMU quantitative finance to a Bocconi Ph.D. These are research and teaching signals, not validated live performance.
• Narend Subramanian’s Krea profile describes an IIT Madras Ph.D. and prior Barclays rates-linear, CRISIL quant-team, Morgan Stanley, Mizuho, and World Bank experience. It names Murex and Numerix platforms and teaching in high-frequency finance and news analytics. These are self-described practitioner and teaching claims, not current-firm disclosures.
• Shivani Bali’s Jindal profile adds an operational-research lineage from the University of Delhi and stated teaching and consulting experience in AI, ML, optimisation, and decision-making. The page does not identify a finance model, fund collaboration, or paper result.

The new routes add model-family, paper, platform, faculty-lineage, and course signals to the research map while keeping academic claims separate from hedge-fund deployment evidence.

September 4, 2026 — Canada, Australia, Switzerland, Italy, the UK, and US academic routes

The expanded professor-and-programme search recovered six additional public routes that are useful for generating hypotheses and tracing talent. They should not be read as evidence that any referenced hedge fund uses the methods or people.

• The University of Toronto Rotman RSM2328H course page names Jun Yuan as instructor for “Machine Learning and Financial Innovation” in Fall 2026. The syllabus explicitly combines supervised, unsupervised, and reinforcement learning, NLP, GenAI, Python, prediction, clustering, adaptation to changing environments, and applications including algorithmic trading, hedging, private equity, fraud, and asset management. It is a curriculum signal and practical talent route, not evidence of a fund partnership, student performance, or production model.
• Monash University’s 2026 BFF5555 unit names Dr Hoa Briscoe-Tran as chief examiner and moves from regression and classification to tree models, unsupervised and deep learning, NLP in finance, model evaluation, regularisation, cross-validation, and overfitting. Project-based assessment and Python implementation make this a route for recovering student research and applied-finance artefacts; the handbook does not identify an employer dataset or live strategy.
• HEC Lausanne’s AI & Digital Economy Lab team page identifies Roxana Mihet as head of the lab, with Luca Gemmi and Kumar Rishabh as postdoctoral researchers and Luise Eisfeld, Ziwei Zhao, and Georgii Zvonka among affiliated researchers. It also publicly names SimTrade, Effixis, and SWZD as private-sector project partners. The page exposes personnel and partnership edges, but not partner data rights, model details, investment use, or results.
• Cardiff University’s Cardiff Fintech Research Group page describes a 13-academic group spanning finance, quantitative methods, financial mathematics, algorithmic and high-frequency trading, ML and AI, robo-advisers, and computer-assisted textual analysis. It names Qingwei Wang as director, Dudley Gilder and Yizhi Wang as co-directors, and Maggie Chen and Arman Eshraghi among the academic staff. This is a personnel-and-research-surface lead; it does not establish a fund relationship or deployable trading system.
• Bocconi’s profile for Andrea Beltratti identifies a finance professor with a Yale economics Ph.D. and dissertation supervision by Robert Shiller, and lists teaching in equity portfolio management, factor and long-run investing, asset management strategy, financial markets, and an “Artificial Intelligence for Finance Lab.” This connects classical asset-pricing lineage to a named AI-finance teaching surface, but the page does not disclose the lab’s projects, code, partners, or investment outcomes.
• The Richmond Fed episode “What CFOs Say About AI Adoption” features Sonya Waddell and Duke finance professor John Graham discussing CFO Survey responses from 750 firms. The transcript supplies testable organizational hypotheses: respondents described production efficiency and decision-making as major motivations; more than half had invested in AI in 2025 and more than 80% expected 2026 investment; and the researchers distinguish expected productivity from revenue-per-employee evidence. It also records a useful implementation boundary: human researchers had to map open-ended job descriptions to official occupations after an AI attempt performed poorly, while AI helped build the resulting dashboard. This is a cross-industry measurement source, not a hedge-fund disclosure.

These routes add explicit curriculum, lab-personnel, project-partner, model-evaluation, and organizational-measurement edges. The defensible next step is to recover the named faculty papers, lab project pages, course assignments, conference recordings, and student-to-employer links before making any inference about a fund’s internal research.

September 4, 2026 — UK market-microstructure labs and an industry-linked doctoral route

The UK search also recovered three routes that add explicit microstructure, privacy, and industry-collaboration detail:

• Ben Moews’s University of Edinburgh profile identifies him as Lecturer in Predictive Analytics, director of the Domain-Driven Machine Learning Lab, and director of the FinTech PhD Programme. It records prior investment-management work as a machine-learning research engineer and current projects in generative modelling for high-frequency market microstructure and privacy-preserving central-bank data disclosure. The profile does not identify a fund employer, production model, or live strategy.
• Raju Chinthalapati’s Goldsmiths profile connects computational finance and AI/ML with market microstructure, liquidity risk, algorithmic trading, risk monitoring, and automation. It also records prior Deutsche Bank FX-strategy research and lists work on nonlinear stochastic-volatility filtering, high-frequency FX directional-change nowcasting, and international trade-network/stock-market connectedness. These are academic and biography signals; they do not establish current fund use or results.
• UCL’s open doctoral thesis by Konark Jain, “Microstructural Financial Modelling: Point Processes and Reinforcement Learning,” describes limit-order-book dynamics with point processes, impulse-control formulations, and RL for electronic trading. The abstract specifically names collaboration with JPMorgan Chase and highlights Hawkes-process modelling for memory, adaptation, and tractability. The repository establishes the thesis and stated collaboration, but not JPMorgan deployment, data rights, trading authority, or out-of-sample performance.

Together these routes expose a research stack around generative market simulation, Hawkes and point-process order-book models, stochastic control, RL, and privacy-preserving financial data. They are candidate idea and lineage sources, not comparative evidence about any tracked manager.

September 4, 2026 — Dutch quantitative-finance chairs and finance-AI lab routes

Three further European academic surfaces add sustainability, financial-crime, and mathematical trading angles:

• VU Amsterdam’s appointment of Norman Seeger establishes a Quantitative Finance chair focused on empirical asset pricing, volatility, derivatives, and risk around scheduled announcements. The announcement also identifies Seeger as director of an MSc Finance programme with roughly 500 students and six specialisations, and describes an AI-aware curriculum agenda. It is institutional and programme evidence, not a disclosed hedge-fund system or performance record.
• Utrecht University’s AI & Finance Lab describes collaboration among data-science, AI, and finance researchers with practitioners on financial-crime detection such as money laundering, sustainable investment, risk management, and extracting knowledge from large financial datasets. The page is partly Dutch and does not identify all researchers, partner data rights, models, or investment deployment.
• Antonis Papapantoleon’s TU Delft profile identifies him as Professor of Mathematical Finance and “Machine Learning in Trading” lead at Delft FinTech Lab. It records a Ph.D. at Freiburg supervised by Ernst Eberlein, prior roles at NTU Athens, TU Berlin, and Mannheim, and industry experience at Commerzbank and the Quantitative Products Laboratory, a Deutsche Bank/HU Berlin/TU Berlin joint venture. His listed research includes model-free finance, systemic risk, and ML applications; the profile does not establish a current fund relationship or production trading authority.

These routes add public evidence about how academic finance programmes connect model design, market risk, sustainability, financial crime, and talent formation. The links do not justify cross-firm ranking or an inference that a named method is used by any tracked hedge fund.

September 4, 2026 — current university programmes and finance-AI idea routes

The university layer is also useful for recovering concrete research hypotheses, course mechanics, and personnel lineages. These sources describe academic work or training design; they do not establish deployment by GMO, Acadian, Arrowstreet, or another manager.

• MIT Learn’s AI and Finance course, taught by Andrew W. Lo, Jillian A. Ross, and Paul F. Mende, explicitly pairs reinforcement learning and large language models with narratives, market learning, investment analysis, trading, and language tasks. It also teaches bias, interpretability, accountability, and responsible deployment. The useful research split is between sequential decision problems and language/reasoning workflows, not one undifferentiated “AI” bucket.
• LSE’s 2026/27 FM484 Big Data and Finance course, convened by Tarun Ramadorai, separates tree-based credit and mortgage analytics—default, selection, and refinancing—from asset-management work using time series, large unstructured datasets, text, and quantitative hedge-fund strategies. It is available across several MSc finance pathways and uses continuous assessment. That makes credit/default research, portfolio design, and text-based research separate recovery lanes; it does not show student or manager deployment.
• Chicago Booth’s Bradford Levy profile identifies an Assistant Professor of Accounting and Applied AI with research at the intersection of AI, information processing, and financial markets, prior Wharton research, and stated AI and safety patents. Booth’s 2027 AI and Financial Information course detail names knowledge bases, language and multimodal models, preference alignment, RAG pipelines, and rerankers, with hands-on labs, cloud resources, team projects, and an industry-expert presentation. This is unusually specific programme and talent-pipeline evidence, not a manager model inventory.
• Wharton’s publication list for Winston Wei Dou lists the August 2026 “The Limits of AI Trading,” the November 2025 “Financial Market Fragility in the Era of AI Planning,” and related work on AI-powered trading, algorithmic collusion, and price efficiency. Together with Dou’s faculty profile, these sources expose hypotheses about AI-driven interaction, coordination, liquidity, price efficiency, fragility, and regulation, plus a named AI-in-finance lab and MIT/Yale/Peking University training. They do not expose code, data rights, portfolio permissions, or performance.
• NYU Tandon’s 2026 Machine Learning in Financial Engineering syllabus starts with regression and Bayesian methods, moves to neural networks and deep learning, and introduces reinforcement learning. It explicitly anticipates more image and text data in finance and requires Python/Jupyter assignments and substantial projects; its AI-assistant policy emphasizes reproducibility and student understanding. This is a modality and research-training signal, not evidence of a firm workflow.
• Booth’s May 2026 account of “A New Era for Finance” says Ralph S. J. Koijen convened roughly 35 people from AI labs, AI-investing startups, asset managers, and academia. The university’s summary describes specialized agents for earnings calls, financial statements, and macro news, with evaluation, benchmarking, privacy, traceability, and cost as open design issues. It is a published summary and discovery route—not a transcript, independent survey, or proof of any participant’s internal system.

The resulting idea map is more precise: sequential decision and control, credit/default and refinancing, unstructured-text portfolio research, multimodal financial-information systems, AI interaction and market fragility, and reproducible agent evaluation. The source note records the evidence class and the remaining recovery work; no firm is ranked on the basis of a professor, programme, or conference.

September 4, 2026 — additional university idea surfaces and validation baselines

The academic search uncovered several exact pages that make the “top ideas” queue more operational without turning university work into hedge-fund evidence:

• Stanford’s Advanced Financial Technologies Laboratory ML map separates deep-learning explainability, incomplete-information filtering, path-dependent risk, financial time series, equity-factor models, RL for order execution and portfolio selection, risk-preference identification, robust stochastic optimisation, and classifier-versus-human evaluation. This is a practical taxonomy for splitting future research tasks into prediction, execution, allocation, risk, and interpretability.
• Stanford’s 2026 applied-AI macro-finance programme describes a real-time/out-of-sample earnings-call benchmark designed around look-ahead bias and market reflexivity, adaptive multimodal inflation forecasting, structural RL for macro models, and graph learning over intermediary holdings during stress. Its abstract reports performance figures, an SDK, and an open competition; those remain conference claims pending paper, code, data, and independent replication.
• Cambridge’s 2026 Global AI in Financial Services report surveys 628 organisations across 151 jurisdictions, including fintechs, traditional financial institutions, AI vendors, and regulators. It provides a cross-sector baseline for workflow location, external foundation-model use, data quality, talent, explainability, privacy, and value measurement. It is not a hedge-fund sample and cannot establish any tracked firm’s AI maturity or vendor stack.
• Oxford’s finance research newsletter adds ML for intraday volume forecasting and VWAP implementation, a generative-AI/financial-economics research map, and evidence that feature engineering and inductive bias affect real-time investment strategies. The Cambridge 27-experiment review adds the required skepticism around backtest overfitting, cherry-picking, trading costs, feasibility, and the experiment-to-practice gap.
• Columbia’s Ciamac Moallemi profile adds a named digital-finance-lab director with MIT, Cambridge, and Stanford training, prior fixed-income-arbitrage experience, and research spanning market microstructure, quantitative trading, and blockchain. That is lineage and research-network evidence, not evidence of a current fund client or production system.

The resulting research queue is now more specific: earnings-call agents need point-in-time, reflexivity-aware evaluation; execution ideas need volume and cost measurement; holdings-network models need stress-period validation; and every academic result needs an explicit treatment of selection, trading costs, and reproducibility. These are candidate research directions, not comparative judgments about managers. See the academic frontier capture note.

September 4, 2026 — Asia-Pacific and European finance-programme routes

The regional academic pass adds programme detail and a market-structure risk surface:

• HKUST’s MSc in Financial Mathematics faculty page names courses and faculty spanning statistical ML, empirical Bayes, scalable AI algorithms, AI in FinTech, RL with financial applications, mathematical market microstructure, trading-behaviour modelling, financial-data mining, portfolio optimisation, and quantitative-finance software. It is a talent and curriculum map, not evidence of any tracked manager’s production system.
• HKUST Associate Professor Yan Ji’s January 2025 AI algorithmic-trading article describes a working paper with Winston Wei Dou and Itay Goldstein in which basic RL agents in a trading laboratory learn coordinated behaviour under some conditions. The article discusses price-trigger rules, imperfect monitoring, noise, and algorithmic homogenisation, while explicitly noting no evidence at that time of harmful AI collusion in live financial markets. This is a market-design and monitoring hypothesis, not a hedge-fund disclosure.
• HEC Paris’s Olivier Bossard profile links MSc Finance leadership and teaching relationships in Japan and China with earlier derivatives-trading leadership and listed interests in cryptography, blockchain, big data, and AI applications in finance. It is biography and programme evidence, not a current fund affiliation or model inventory.

These routes expand the research map across model design, RL market interaction, microstructure, derivatives, data mining, and talent formation. The regional source note records the HKUST FinStaR page as a recovery lead because its current URL returns a page-moved response; no inference is drawn from that failed capture.

September 4, 2026 — MIT-linked and European quant-conference routes

The conference search also recovered several useful bridges between finance programmes, vendors, and firm personnel:

• MIT CSAIL’s Quadrature Tech Talk page dates an October 2, 2025 event and names Daniel Goldbach, a Quadrature quantitative developer who had worked as Head of ML Infrastructure, and Hanna Yakubovich, whose listed work included low-latency infrastructure, portfolio construction, and alpha-forecasting research. The page does not provide a recording, model details, data sources, or investment results; it is a firm-linked personnel and MIT event route.
• The Quaint Quant 2025 page records an April 18, 2025 event at the University of Texas at Dallas with speakers from Morgan Stanley, BlackRock, Agora Data, Bertram Capital, Databento, H2O.ai, NYU, UT Dallas, and Lehigh. It is a programme-and-personnel discovery surface spanning finance, risk, data, and AI; it does not establish the employers’ internal systems.
• Politecnico di Milano’s ALGODEFI25 programme links algorithmic trading, market making, DeFi, and AI, with keynotes from Euronext’s Head of Quant Research Paul Besson, HEC’s Thierry Foucault, Stuttgart’s Martin Herdegen, and Imperial’s Eyal Neumann. It provides a dated academic-industrial agenda, not a recording or production-deployment record.
• The same QFinLab has a current ALGODEFI26 successor page for October 8–9, 2026. It names Eduardo Abi Jaber of École Polytechnique and Ryan Donnelly of King’s College as keynotes and announces a dedicated industry-oriented afternoon. Registration is open, but recordings, accepted papers, and implementation details are not yet public on the page.

The conference-route note keeps these event rosters separate from firm disclosures and academic results. The next capture step is to recover authorised recordings, speaker slides, papers, and current personnel pages before treating any topic as more than a public research or talent signal.

September 4, 2026 — professor-led research hypotheses and finance-programme routes

The university search adds a separate idea-supply layer. The LSE faculty-specialisms page identifies Giulia Livieri’s work across mean-field games, causal deep learning, graph learning, financial econometrics, and path-dependent contracts; Kostas Kalogeropoulos’ work on Bayesian and sequential learning for partially observed systems; Clifford Lam’s work on tensor and spatio-temporal factors, covariance, and shock propagation; and Chengchun Shi’s work on offline reinforcement learning, off-policy evaluation, and LLM trustworthiness. These are distinct hypotheses around partial observability, causal structure, contagion, and policy evaluation—not one generic “AI trading” claim. LSE also lists kernel methods, online learning, bandits, and multi-agent systems as adjacent routes. The page does not establish a fund sponsor, production system, or investment result.

The Swiss Finance Institute at EPFL faculty roster and Semyon Malamud’s publication page connect quantitative finance and machine learning with portfolio selection, nonlinear filtering, liquidity, implementable efficient frontiers, principal portfolios, and the role of complexity in return prediction. This is a useful path for testing capacity- and liquidity-adjusted representation learning; it is not evidence of a manager’s model or returns.

Princeton’s ORFE machine-learning research page maps high-dimensional financial econometrics, portfolio theory, high-frequency finance, graph/network modelling, real-time optimization, and control. Jianqing Fan’s teaching page records a 2026 “Deep Learning and Generative AI” course alongside financial econometrics and statistical learning. The programme is a talent and curriculum signal; it does not disclose student destinations, employer projects, or production deployment.

The professor-led idea-route note turns those sources into bounded research tests: offline-RL evaluation under partial observability, graph/tensor models for cross-market shock propagation, implementability-adjusted complexity, causal restrictions for sequential decisions, and provenance-aware evaluation of finance research agents. These are hypotheses for validation, not rankings or claims about any covered firm.

September 4, 2026 — Quoniam’s explicit AI workflow and a Bridgewater-linked macro route

Two title-blind manager-media routes sharpen the separation between public process disclosures and broader investment commentary:

• Quoniam’s May 2026 interview with Dr Maximilian Stroh, Head of Research, describes ML as a refinement to a linear forecasting framework built around value, quality, and sentiment. It names nonlinear signal interactions, NLP over news and reports, and agentic tools for code migration, internal-document search, research summarisation, and code/text review. The same account names holdout data, multiple-testing controls, sensitivity analysis, implementation-cost and liquidity assumptions, live monitoring, and human review of code changes. Quoniam’s July 2026 follow-up identifies Dr Desislava Rakova’s further machine-learning training at Stanford and stresses problem definition, parameterisation, validation, and critical review. These are unusually concrete first-party descriptions of research-tool and control categories, but they do not disclose model files, feature schemas, data permissions, vendor contracts, portfolio weights, or independently audited attribution. See the capture note.
• PGIM’s Outthinking Investor episode with Ray Dalio identifies George Patterson as PGIM’s Chief Investment Officer of Quantitative Solutions and says the conversation was recorded at the 2025 Greenwich Economic Forum. The official transcript places AI, quantum computing, and advanced automation within a wider discussion of technology, geopolitical competition, and long-cycle macro forces. This links a Bridgewater founder to a PGIM quantitative-solutions media route; it is macro/AI framing, not evidence of a current Bridgewater model, PGIM production architecture, data rights, portfolio authority, or performance. See the capture note.

The new manager evidence suggests a useful research taxonomy: forecasting and nonlinear interaction, unstructured-information extraction, research-agent operations, validation and portfolio-level evaluation, and macro scenario framing. The academic routes above supply candidate methods and personnel pipelines; these first-party firm routes show how some of those categories are described in practice. The two evidence classes remain separate, and no cross-firm ranking follows from them.

September 4, 2026 — Dresden and Lancaster academic idea routes

The TU Dresden 3rd Conference on AI in Finance is scheduled for September 17–18, 2026 and includes a PhD/postdoc workshop on “AI in Financial Decision Making” plus a workshop on running and fine-tuning local LLMs. The page names Lars Hornuf as organiser, Matthias Mattusch as the local-LLM workshop instructor, and Alejandro Lopez-Lira as keynote. This is a concrete route for recovering local-inference economics, fine-tuning choices, and academic research questions; it does not establish a hedge-fund implementation.

The Lancaster FoFI 2026 programme adds paper-level ideas around time-series foundation models, ML factor-strategy performance, nonlinear momentum, global news networks, treasury-bond return prediction, and bottom-up capacity constraints. The titles are useful discovery seeds for papers and code, but the programme itself does not provide a validated model, data licence, fund relationship, or performance attribution. See the Dresden/Lancaster capture note.

September 4, 2026 — Bologna, ETH, and finance-model evaluation routes

The Bachelier World Congress at the University of Bologna adds a dated academic–practitioner route spanning machine learning in finance, market design, market microstructure, volatility, liquidity, algorithmic trading, and mean-field or agent-based models. Its speaker roster includes researchers from Sorbonne, EDF, CUHK, EPFL, Imperial, Toronto, LSE, Waterloo, and Stanford; the committee page also names NatWest Markets’ Vladimir Piterbarg. This is a recovery surface for papers, recordings, and academic lineages, not evidence of a firm’s model use.

The ETH Zurich ML for Finance and Complex Systems course connects financial applications with graph methods, matrix factorisation, RL, neural ODEs, PINNs, attention, transformers, and Black–Litterman through a coding project. The Kronos AAAI paper adds a Tsinghua-authored finance time-series foundation-model artifact, while the FinVerse benchmark separates point forecasting, cross-sectional ranking, and portfolio-level evaluation across a broad financial series universe. These are research and evaluation routes; the papers do not establish live trading, data rights, fund adoption, or performance outside their published protocols.

The Bologna/ETH/Kronos/FinVerse capture note records the SWUFE MLiFE2026 page as a separate recovery lead because its programme page was not reliably retrievable during this pass.

September 4, 2026 — CMU’s practitioner speaker-series bridge

The CMU MSCF Speaker Series states that it draws buy-side and sell-side practitioners discussing quantitative finance applications, career skills, and organisational experience. Its dated entries provide a personnel and topic trail: September 3, 2026 lists Yumi Oh as Global Head of Trading Strategy at Citadel Global Quantitative Strategies, with a description of translating alpha and monetisation research into scalable products; November 6, 2025 lists former Two Sigma alternative-data leader Tony Berkman; and October 30, 2025 lists Two Sigma Fast Engineering leader Ryan Jerchau alongside Citadel Securities content-data-science leader Ryan Preclaw. The November 21, 2024 entry lists Reha Tutuncu as Head of Portfolio Research at Point72 and records prior Carnegie Mellon faculty, AQR, and Goldman Sachs Asset Management roles.

The page presents multiple dated biographies for Yumi Oh with different titles and descriptions, so those records remain time-scoped rather than being merged into one undated current claim. The page exposes routes around market microstructure, alternative data, execution engineering, content data science, and portfolio research; it does not disclose model inventories, datasets, permissions, recordings, or investment results. See the academic finance source note.

September 4, 2026 — sponsor-facing quant programmes and finance-ML curricula

Carnegie Mellon’s MSCF Machine Learning Capstone provides an unusually explicit academic-to-industry route. The programme says companies can propose 14-week projects with faculty, working with teams of four to six students; sponsors provide a project manager and data under an Educational Partner Agreement. Its public examples include earnings-announcement reactions, 10-K-based dividend-cut prediction, stock clustering, anomaly detection, transaction surveillance, and image/document data. The page also discusses sponsor privacy and secure data transfer. This exposes project categories and information boundaries, not sponsor adoption, live trading, or performance.

Toronto Rotman’s Fall 2026 Machine Learning and Financial Innovation course combines supervised and unsupervised learning, reinforcement learning, NLP, GenAI, Python, and hands-on finance applications including credit, algorithmic trading, fraud detection, hedging, and asset management. The University of Toronto Master of Mathematical Finance course list places machine learning beside numerical methods, risk management, optimisation, portfolio management, and an industry internship. These are talent and research- scoping signals; they do not identify student employers, private datasets, or live production systems. See the academic finance source note.

September 4, 2026 — University of Chicago applied-ML and finance-research routes

The University of Chicago Financial Mathematics programme exposes a concrete curriculum-to-experiment route. FINM 33160: Machine Learning for Finance describes using fundamental stock factors, decision trees, random forests, gradient boosting, Python, and backtesting to construct an algorithmic trading strategy, with statistical analysis of backtest results. The academic catalogue adds a roughly 4,000-company, ten-year quarterly-report factor exercise through the next earnings release, plus hyperparameter optimisation, feature selection, and a PyTorch neural-network introduction. This is a course specification and research idea route, not evidence of a live strategy, proprietary data, or performance.

FINM 33165: Reinforcement Learning and Deep Learning adds Bellman equations, RL, neural networks, and language-model training context. Taken together, the two courses expose a useful decomposition for research: cross-sectional prediction and backtest statistics, followed by sequential decision methods and deep learning. The programme pages do not identify student employers, private datasets, model providers, agent permissions, or production systems.

The Stevanovich Center conference archive adds recurring routes around high-frequency trading, large financial datasets, causal inference, generative modelling, and AI in econometrics and finance. Its 2026 calendar lists “Market Microstructure, Quantitative Trading, High Frequency, and Large Data” on May 28 and “Big Data and Artificial Intelligence in Econometrics, Finance, and Statistics” on October 8–10, with organisers including Rina Foygel Barber, Chao Gao, Roger Lee, Cong Ma, Per Mykland, Niels Nygaard, Dacheng Xiu, and Lan Zhang. It also lists an October symposium on statistical foundations of generative modelling. These are paper, speaker, and recording recovery surfaces; they do not establish a hedge-fund implementation or shared dataset.

The Niels Nygaard faculty profile records his Chicago mathematics role, earlier Princeton teaching, MIT mathematics PhD, and founding/director roles in Financial Mathematics. Per Mykland’s profile records his Chicago statistics-and-finance chair, scientific-director role, prior Oxford and Princeton appointments, and high-frequency/time-dependent-process research. These are academic lineages and method-recovery routes, not evidence of employment by or adoption at any tracked firm. See the academic finance source note.

September 4, 2026 — Warwick investment-AI practice and UCL generative-finance routes

The Warwick Gillmore Centre/CFA Institute event held on March 31, 2026 combines a research session on agentic AI and responsible investment research with a practitioner panel. The page names Brian Pisaneschi (CFA Institute), Julan Al-Yassin (CFA Institute), James Hadfield (former CVC, Ninety One, and Close Brothers), Simon Legrand-Green (WTW Head of Multi Asset and Systematic Strategies Research), and Carlos Salas (Portfolio Manager, Investment Research, ML/AI). The advertised topics include research automation, NLP-driven insights, data-quality checks, model risk, compliance, data lineage, interpretability, and how firms evaluate tools and workflows. This is event metadata and practitioner commentary; it does not disclose a named fund’s architecture, data permissions, agent authority, or performance.

UCL provides a more technical model-idea route. Its March 25, 2026 seminar on generative modelling of financial time series records Professor Hao Ni’s work on path signatures and path-development networks for irregular, noisy, and heterogeneous financial series. The event description names Sig-Wasserstein GAN, PCF-GAN, High-Rank PCF-GAN, and Sig-DEG, a compacted diffusion generator, and describes an intersection of stochastic analysis, machine learning, and multimodal data. UCL records Ni’s Oxford DPhil, prior Brown postdoctoral work, and Oxford-Man Institute postdoctoral affiliation. The seminar is a research and lineage source, not evidence of a hedge fund using these generators or of a tradable signal.

The UCL MSc Finance with Data Science programme adds a current talent-pipeline route: time-series analysis and forecasting, big data and machine learning, Python-based work with Bloomberg and Refinitiv data, and a concrete finance research project. It also identifies a CQF Institute partnership and describes intended roles including quantitative analysis at hedge funds. These statements describe curriculum, data-access context, and intended career pathways; they do not identify students, employers, private datasets, or production systems. See the academic finance source note.

September 4, 2026 — Imperial’s practitioner and recruiting routes

Imperial’s Practitioners’ Lecture Series is a first-party archive aimed at MSc and PhD students and explicitly focused on short lectures with direct industry relevance. Its dated 2025–26 entries include Alexis Yannakou (Citadel), Irene Perdomo (DRW), and Leonardo Marroni (DRW). Earlier entries include Alexandre Pinto (Signal AI), Cecilia Auburn (Capital Fund Management), Marco Dion (Qube Research and Technologies), and Alexis Bellot (DeepMind). This creates a concrete personnel-and-recovery queue around Citadel, DRW, CFM, Qube, and adjacent technical employers. The archive does not provide lecture transcripts, model inventories, datasets, permissions, production status, or investment results; the names and dates are therefore retained as programme metadata, not as evidence of firm deployment.

The separate Careers in Quantitative Finance series adds a title-blind recruiting route. Its current calendar lists Quantbot Technologies and Taula Capital, as well as Barclays, BNP Paribas, Deutsche Bank, and NatWest; its 2025–26 archive includes G-Research, MerQube, and Qube-RT. The calendar shows that these firms participate in a university-facing talent and practitioner channel. It does not show that a named speaker was employed by the programme, that a student joined a firm, or that any AI system was discussed.

The Imperial Quantitative Finance Centre events page records a July 1, 2026 hedge-fund conference hosted by Imperial and Goldman Sachs and a November 10–11, 2026 CFM–Imperial quantitative-finance workshop hosted by Deutsche Bank and Imperial’s Centre for Excellence in Quantitative Finance. The CFM–Imperial Institute description describes academic/practitioner collaboration, events, doctoral funding, and postdoctoral opportunities. These pages expand the source map for professors, students, and industry speakers; they do not support a cross-firm capability comparison. See the academic finance source note.

September 4, 2026 — Luiss model stack and Edinburgh foundation-model partnership

Luiss’s 2026/27 Data-Driven Models for Investment course publishes an unusually explicit curriculum map. It names satellite, geospatial, web, and news/media data; LLMs and generative AI; qualitative and quantitative AI; agentic and transformer architectures; and Python labs. The course contents name FinBERT, AlphaAgents, TradingAgents, TimeGPT, Chronos, and TabPFN, alongside zero-shot time-series forecasting, multi-agent portfolio and trading systems, ensemble models, genetic algorithms, and quantum-annealing/Ising exercises. It also describes a session on techniques associated with an unnamed UK hedge fund and its Head of Capital Markets. This does not identify the fund or speaker, nor does the course’s “live trading strategy” objective establish real-capital authority, production continuity, or performance.

The University of Edinburgh Business School’s sequential-foundation-model PhD studentship adds a named industry-partnership route outside securities trading. Co-designed with Lloyds Banking Group, the four-year project proposes pre-training a foundation model on ordered customer-event histories and reusing its learned representation for risk assessment, pricing guidance, and customer-behaviour prediction. It names Dr Victor Medina-Olivares and Dr Zexun Chen as supervisors, describes insurance and pensions as a transfer domain, and says the student may spend time at Lloyds’ Technology Centre in Hyderabad. This exposes sequence modelling, transfer learning, industrial data, and supervision structure; it does not identify the student, private data fields, model size, weights, evaluation results, or hedge-fund use.

The Edinburgh Financial Machine Learning II course catalogue should remain date-scoped: its 2025/26 description covers HFT and limit order books, feed-forward networks, autoencoders, CNNs, RNNs, RL, and transformers, but the catalogue currently says the course is not being delivered. This prevents a course description from being misread as current teaching or industry deployment. See the academic finance source note.

September 4, 2026 — Bocconi model implementation and Southampton finance-automation routes

Bocconi’s Data Science and Machine Learning for Finance course record provides a dated 2025–26 syllabus under Francesco Corielli. It moves from regression, principal components, canonical correlation, and clustering into neural networks, autoencoders, factor models, recurrent networks, attention, and LLMs, with applications to pricing, hedging, and trading. The course uses Python and Google Colab notebooks, practical exercises, and collaborative assignments; the learning outcomes explicitly include building and testing these models while avoiding common ML/DL pitfalls. Bocconi’s MAFINRISK page adds a numerical-methods and transfer-learning context under Francesco Rotondi, including Monte Carlo, lattices, PDEs, unstructured data, and basic neural networks. These are curriculum and instructor signals, not a manager’s model registry, data licence, production endpoint, or performance record.

Southampton’s Financial Technology and Applied AI module is a separate operations-and-governance route. Its 2026–27 syllabus covers prompt engineering for finance, AI-assisted data work, lightweight prototyping, APIs and data contracts, RPA for onboarding/KYC/reconciliation, document and knowledge automation, regulatory monitoring, risk/compliance triage, agent oversight, AI use logs, prompt registers, testing protocols, tool-risk classification, vendor selection, KPI design, and implementation planning. It explicitly distinguishes model building, tool use, agentic workflows, and process automation, and requires students to evaluate controls, auditability, bias, and operational resilience. This is evidence about what a finance programme considers delegable and governable work; it does not identify a hedge fund, proprietary data, production agent, or live investment authority. See the academic finance source note.

September 4, 2026 — Harvard practitioner-teacher and Maryland ML-finance routes

Harvard Division of Continuing Education’s Applied Quantitative Finance and Machine Learning course adds a practitioner-teacher route with explicit biography metadata. The Fall 2026 course covers data management and analytics, quantitative investment strategies, portfolio management, risk management, and ML/AI in quantitative finance. Harvard lists MarcAntonio Awada as instructor and identifies his current role as Chief Innovation and Digital Strategy Officer for Portfolio Investment at Brown Capital Management. The biography also describes prior quantitative and systematic-trading leadership at Morgan Stanley, Chase Manhattan, BNP Paribas, Dresdner Bank, Sunofia Quantum, and Balyasny, plus research/data-science roles at Harvard’s D^3 Institute and LISH. These are institution-published career and teaching claims; they do not reveal Brown Capital’s current model inventory, data rights, research process, or portfolio permissions. The page says recorded sessions are available to enrolled students, so the syllabus and recording-access boundary is a follow-up route rather than public deployment evidence.

The University of Maryland’s BUFN650 Machine Learning in Finance course listing adds a dated public course-and-instructor route. The Fall 2026 section is listed under Serhiy Kozak and describes hands-on financial modelling with lasso, deep learning, TensorFlow, Python, Google Colab, and big-data experimentation. The listing gives a prerequisite and course dates, which make it useful for tracking the academic talent pipeline and locating a syllabus or student-project archive. It does not provide Kozak’s research profile, project datasets, employer links, or evidence that any class exercise became a live trading system.

These routes are useful for idea extraction, not firm ranking: Harvard links systematic-trading career history with portfolio/risk framing and teachable ML workflow; Maryland provides a dated course surface for comparing regularised models, deep learning, and notebook-based experimentation. Neither route establishes a fund’s present AI or GenAI system. See the academic finance source note.

September 4, 2026 — Illinois automated-trading and Karlsruhe empirical-finance routes

The University of Illinois Urbana-Champaign’s 2026–27 Financial Engineering: Automated Trading Practices catalogue exposes a concrete programme architecture. The concentration pairs High Frequency Trading Technology with Algorithmic Trading Systems Design and Testing or Algorithmic Market Microstructure, then adds stochastic and learning foundations such as deep learning, algorithms for data analytics, and queueing systems. Illinois also describes a corporate-sponsored practicum for real-world financial-modelling problems using analytic tools and software. This is programme design and a stated industry-project mechanism, not evidence of a named firm’s data access, project outcome, production permission, or trading performance.

Karlsruhe Institute of Technology’s Advanced Quantitative Finance course page adds an empirical workflow. The 2026 course combines factor models, asset pricing, options, and credit risk with bootstrap, panel and cross-sectional regressions, instrumental variables, GMM, neural networks, and random forests. Students use Python and three hackathons to implement empirical tests on large datasets, including replication and related-question exercises with peer feedback. The page names Julian Thimme as the 2026 instructor and says course materials are held on the internal ILIAS system. It does not disclose the datasets, student projects, employer participants, or any live investment system.

The research-design ideas exposed here are: separate market-microstructure and HFT infrastructure from model selection; require replication before novelty; and use staged, peer-reviewed notebooks before production handoff. These are derived from programme structure, not claims about any fund’s internal process. See the academic finance source note.

September 4, 2026 — Singapore and Japan academic-to-practice routes

The National University of Singapore’s Centre for Quantitative Finance adds an Asia-Pacific institutional route. Its public mission connects mathematics, data science, computer science, economics, finance, and engineering, and names AI/ML theory and applications, digital finance, computational finance, seminars, workshops, an annual conference, and industry collaboration. The CQF team page adds concrete research vocabulary: Julian Sester is listed with model uncertainty, stochastic control, reinforcement learning, and credit risk; Zhang Yijiong with spatial-temporal econometrics, robust RL, nowcasting, and portfolio optimization; and Zhang Jingguo with ML in finance and transformers. It also lists NLP, deep RL, systematic market making, Gaussian processes, and quantum computing among public research topics. These are research-interest and institutional mission signals, not evidence of a named manager’s system, data access, or results.

Keio University’s Applied Research Institute of Finance (K-ARIF) is a dated FY2026 research project led by Principal Investigator Takaki Hayashi. The June 10, 2026 page frames K-ARIF around practical asset-management research, stock selection, portfolio construction, rebalancing, and empirical analysis. Its themes include volatility and tail-risk forecasting, high-frequency data, market impact, liquidity, optimal execution, derivatives, crypto assets, ESG, alternative data, AI, and generative AI. The planned data surface includes prices, volume, order books, news, corporate and macro data, and social media, with statistics, econometrics, data science, ML, and GenAI as methods. This is a research agenda and academic-to-practice structure; it does not disclose a completed model, partner fund, dataset license, production endpoint, or performance.

The Keio researcher record for Takaki Hayashi adds a deeper lineage and paper route. It records graduate doctoral-course work in Statistics at the University of Chicago, quantitative finance, financial engineering, time-series analysis, and high-frequency-data research. Listed outputs include a 2026 paper on two-fund separation under hyperbolically distributed returns, work on strategic liquidity provision in high-frequency trading, collaborative filtering for limit-order book market-quality evaluation, and earlier projects on covariance estimation, lead-lag analysis, and high-frequency stock-price formation. The record also lists Hayashi as an advisor to the Bank of Japan Financial Markets Department from December 2025 and as a Financial Services Agency adviser from 2022. These are academic and public-policy interfaces, not evidence of K-ARIF outputs, fund sponsorship, data rights, or live investment authority.

Yoshiyuki Suimon’s Keio profile provides a separate personnel lead: the university records him as a Graduate School of System Design and Management associate professor and former Senior Economist and Head of the Data Science Department at Nomura Securities. His stated research interest is mathematical and data-driven analysis of human and corporate behaviour in economic and financial systems. This creates a university-to-bank data-science route to investigate through papers, seminars, students, and former colleagues; it does not connect Suimon to K-ARIF, a hedge fund, a specific model, or a production system.

The transferable ideas are research hypotheses: combine robust control with portfolio construction under model uncertainty; treat market impact and execution as first-class labels rather than post-trade diagnostics; and join market, corporate, macro, news, and social data only with point-in-time controls. These are deductions from published research themes, not claims about live deployment. See the academic finance source note.

September 4, 2026 — Bangor/Bond execution benchmark and faculty lineage

Bangor University’s 2025 peer-reviewed trade-execution paper by Isaac Tonkin, Adrian Gepp, Geoff Harris, and Bruce Vanstone evaluates the decision layer rather than only return prediction. It describes a framework with benchmarks, diagnostics, and analytics for deep-RL execution strategies across two trading horizons. Its abstract reports higher median performance and lower standard deviation for the tested deep-RL approaches versus the benchmark strategies in that study, with a separate result for the most flexible action space. This is bounded to the paper’s experimental setup; it is not evidence of a live fund system, scalable execution, or future performance.

Vanstone’s Bangor faculty profile adds a 2006 Bond University PhD in Computational Finance, supervision in algorithmic trading, data science in finance, and ML in finance, plus publications on regional transfer, ELM/RNN/LSTM comparisons, news analytics, sentiment, statistical arbitrage, and execution. The faculty page and benchmark create a follow-up route for code, data, horizon definitions, transaction-cost assumptions, and replication—not a claim about any manager’s adoption.

The research idea is to score an investment workflow on separate axes: forecast quality, execution quality, robustness across horizons, action-space constraints, and diagnostics. That decomposition avoids treating a predictive model as a complete trading system. It is a methodological inference from the paper, not a firm comparison. See the academic finance source note.

September 4, 2026 — FinRL open framework and reproducible research surface

The Columbia-hosted FinRL paper by Xiao-Yang Liu, Hongyang Yang, Jiechao Gao, and Christina Dan Wang (2021) provides a public research artifact more specific than a generic model survey. It describes a three-layer deep-reinforcement-learning framework: an application layer for trading tasks, an agent layer with configurable DRL algorithms and reward functions, and an environment layer wrapping historical data and live-trading APIs. The paper models OHLCV, technical and fundamental indicators, smart-beta and NLP sentiment features; buy/sell/hold, long/short, and portfolio-weight actions; market frictions, liquidity, risk aversion, and multiple time granularities; and equity, crypto, FX, futures, and options environments.

The discovery implication is architectural: the public framework makes state, action, reward, environment, and backtest interfaces inspectable. That supports a replication or audit plan in which data provenance, transaction costs, reward design, and point-in-time splits are tested independently. The paper is a 2021 conference artifact; its framework description does not establish a current fund user, proprietary data, production authority, or live performance. See the academic finance source note.

September 4, 2026 — Chicago practitioner-teacher route

The University of Chicago Booth profile for Brian Boonstra records a practitioner-teacher route. Booth describes him as an adjunct associate professor of finance and quantitative researcher who served as senior or chief quant at Jump Trading, Cognitive Capital, UBS O’Connor, Delaware Street Capital, JPMorgan, and Helios, and who managed Thureos Capital, an equity/credit-arbitrage hedge fund. His stated interests are contingent-claims pricing, numerical analysis, and statistical learning; his education includes a Chicago mathematics degree and a Michigan Ph.D. in complex analysis. Booth’s 2026–27 schedule lists Advanced Models of Security Pricing and Credit Risk. This is biographical and teaching evidence, not a disclosure of any former or current employer’s AI stack, data, or results.

The route suggests a specific research hypothesis to track: statistical learning may sit alongside derivatives, credit-risk, and contingent-claims modelling rather than only language or return-prediction workflows. That is an inference from the public profile and course schedule, not a claim about any fund’s implementation. See the academic finance source note.

September 4, 2026 — university archives linking quant firms, labs, and researchers

The NUS Risk Management Institute recruitment-talk archive records a 30 September 2019 Squarepoint event with Jean-Damien Artarit, whom the page identifies as the firm’s CEO for Asia. The archived firm description says Squarepoint used scientific, data-driven research and automated strategies across global markets and asset classes, from millisecond horizons to longer-term strategies, with arbitrage and statistical approaches supported by infrastructure for large trading volumes. This is dated university-hosted recruitment material and firm description; it does not disclose current systems, model families, data permissions, or performance.

The NUS Centre for Quantitative Finance industry-seminar archive records a 5 September 2020 webinar with Dr Wang Yu of Squarepoint scheduled for 3:00–3:25 p.m., alongside Standard Chartered speakers. The page confirms participation and timing but exposes no recording or transcript. It is therefore a personnel-and-event route, not method evidence.

Justin Sirignano’s current Oxford Mathematics profile lists him in Machine Learning and Data Science and Mathematical and Computational Finance groups, with work on high-frequency limit-order books, loans, and options. The profile identifies him as course director of Oxford’s MSc in Mathematical and Computational Finance and records a Squarepoint Foundation grant of $800,000 supporting two postdoctoral researchers and two DPhil students for four years. The grant and research topics establish an academic funding and talent-pipeline connection, not Squarepoint access to the research, model ownership, or production use.

The GLIndA2022 organizer page provides an older but explicit personnel bridge for Naftali Cohen. The 2022 conference biography describes him as Senior Data Scientist at Schonfeld and an NYU adjunct professor, with a prior role as Vice President and Research Lead of AI Research at JPMorgan. It also connects his work to mathematical modelling of climate systems and data mining. Because this is a historical conference biography rather than a current employer page, it should be used for lineage and discovery only; it does not establish current Schonfeld employment or a Schonfeld model.

These routes expand the academic search beyond course syllabi: university recruiting pages expose dated firm language, seminar archives expose personnel timing, grant pages expose research funding and talent channels, and conference biographies expose career transitions. None should be treated as evidence of a live fund system or ranked as a quality signal. See the academic finance source note.

September 4, 2026 — current finance-programme and professor routes

The University of Illinois Urbana-Champaign’s IE 517 Machine Learning in Finance Lab is distinct from the university’s broader automated-trading concentration. The Fall 2026 listing describes a two-credit lab designed for first-semester MS Financial Engineering students, with rigorous Python exercises using pandas, NumPy, and scikit-learn and unique real-world financial datasets. It names M. Murphy as instructor and dates the section from October 19 to December 9, 2026. This exposes a current educational workflow and a stated dataset format; it does not identify the datasets, employer partners, student projects, or any production trading system.

The University of Chicago’s Machine Learning for Finance professional-education course is an eight-week online course scheduled to start August 31, 2026. Its public description covers data collection and organisation, feature engineering, train/test splits, cross-validation, backtesting, monitoring, tree ensembles, clustering, risk and portfolio selection, Monte Carlo out-of-sample testing, deep learning, and Bayesian inference. The page identifies Lara Kattan as a clinical assistant professor of operations management and describes her data-science, ML-engineering, and financial-modelling background. This is a continuing-education and instructor signal, not evidence that any fund uses the listed methods.

NYU’s 2026–27 Finance and Risk Engineering catalogue supplies a broader programme map. FRE-GY 7703 connects multivariate statistics, ML generalisations of linear factor models, classical and Bayesian econometrics, backtesting, model and estimation risk, distributional stress testing, and portfolio or business construction. FRE-GY 7773 covers supervised, unsupervised, and reinforcement learning in financial engineering. The same catalogue also lists statistical arbitrage, algorithmic trading and high-frequency finance, news analytics, forensic financial technology, thesis, and faculty-supervised project routes. These course descriptions are useful idea and talent-pipeline signals; they do not establish which students, employers, data sources, or live strategies are involved.

Yale SOM’s Bryan T. Kelly faculty profile is a high-value academic-to-practitioner record. Yale identifies Kelly as Tanner Professor of Finance, NBER Research Fellow, Associate Director of the International Center for Finance, and head of machine learning at AQR Capital Management. The profile places his research across asset pricing, financial econometrics, financial ML, volatility, tail risk, correlation, systemic risk, and financial networks, and lists recent work on credit-implied volatility, the implementable efficient frontier, option-return distributions, and replication. His education is listed as NYU Stern, UC San Diego, and the University of Chicago. This supports a dated personnel and research-lineage claim; it does not reveal AQR’s internal models, data permissions, deployment stage, or performance.

Boston University Questrom’s official Pietro Bini profile is the authoritative replacement for the previously linked personal page. BU identifies Bini as an Assistant Professor of Finance whose empirical asset-pricing work examines institutional investors and fixed-income markets. It separately describes research on the behaviour and risks of AI models in financial decision-making, his role in the AI and Big Data in Finance Research Forum, and DEFT Lab membership. BU lists a Cornell finance PhD in 2025, Bocconi economics-and-finance degrees, and a 2026 publication on LLM biases and corrections. This is a professor, lab, and research-network route; it does not connect those academic projects to a fund’s live system or investment authority.

Taken together, these routes add five idea families to the research queue: financial-document and portfolio-model evaluation with explicit leakage controls; factor-model generalisation and distributional stress testing; supervised, unsupervised, and reinforcement-learning comparisons; institutional-investor and fixed-income behaviour; and audit of LLM bias in financial decisions. Those are hypotheses to test against papers, code, data definitions, and point-in-time backtests—not claims about any firm’s capabilities or relative position.

September 4, 2026 — applied European and undergraduate finance-ML routes

The University of Glasgow’s ACCFIN5229 Advances in Machine Learning in Finance course is listed for the 2026–27 session in the Adam Smith Business School’s MSc Financial Technology programme. The 10-credit, two-week course explicitly combines algorithmic and pairs trading with bootstrapping, multiple-hypothesis testing, false-discovery and family-wise error rates, variable selection, and factor analysis. That combination is a useful research-design signal: the curriculum treats statistical-selection discipline as part of the trading problem. It remains a course description and does not establish any fund’s research process, data, or deployment.

KIT’s HECTOR School Alternative Data and Machine Learning for Business Applications certificate is scheduled for October 26 to November 6, 2026. The eight-ECTS programme moves from text mining and NLP for financial applications to neural-network design, training, validation, and pattern recognition, and explicitly targets quantitative analysts, investment professionals, asset managers, risk managers, and data engineers. It is a professional-training signal around alternative-data plumbing and model validation, not evidence of a named firm’s vendor stack or model usage.

Illinois’s FIN 453 Introduction to Machine Learning in Finance exposes an undergraduate pipeline that is separate from IE 517. The Fall 2026 listing covers neural networks, regression trees, gradient boosting, clustering, PCA, introductory deep Q-networks, option pricing, and credit-card fraud detection, with Python, PyTorch, scikit-learn, and XGBoost. The section is restricted to Finance or Finance + Data Science majors and lists V. Duarte as instructor. This is evidence about entry-level finance-ML training and application surfaces, not a claim about subsequent employer adoption.

The UMass Boston Machine Learning in Finance course page was returned by the search index with a Fall 2026 schedule, Python/NumPy/Pandas/scikit-learn, web scraping, textual analysis, and instructor metadata, but the official page was not retrievable in this pass. It is retained as a recovery lead only and is not used to support a verified programme claim here.

These routes add three testable idea families: multiple-testing control before promoting factors; alternative-data extraction with an explicit validation stage; and a staged progression from undergraduate models to execution and production-risk questions. They are academic hypotheses and talent signals, not a ranking of schools or firms.

September 4, 2026 — industry-funded university research and project surfaces

IIT Bombay professor Piyush Pandey’s public projects and grants page names three 2025–27 projects funded through the IITB Citadel Securities Quantitative Finance Lab: change-point detection in high-frequency financial time series; modelling order-flow toxicity and liquidity in high-frequency markets; and pricing options and computing implied volatilities with machine learning. The same page lists adjacent SBI Foundation projects on VaR, interest-rate forecasting with statistical/ML tools, and GenAI for climate-risk stress testing. This is unusually specific project-title evidence and a route to papers, students, and outputs; it does not establish Citadel Securities’ internal model, data rights, production use, or performance.

The research ideas exposed by the Citadel-linked titles are concrete: detect structural breaks before a signal is handed to a trading model; treat order-flow toxicity and liquidity as prediction or execution labels; and model the implied-volatility surface as a structured object rather than a single option-price target. Those are hypotheses derived from the project descriptions, not claims about implementation.

University of Guelph professor Fred Liu’s public faculty and research page adds a Canadian academic-to-market route. Liu lists AI, ML, financial econometrics, asset pricing, and risk management as research interests; affiliation with CARE-AI and Guelph’s data-science programme; funded PhD supervision using government and industry partnerships; and presentations of “Stock Distribution and Risk Premia Predictability with Quantile Machine Learning” at Quantbot Technologies in 2026. The page also lists Bank of Canada, Co-operators, and ADIA Lab connections and courses in ML for economics/finance and AI in financial markets. These are public academic, funding, presentation, and talent-pipeline signals; they do not establish Quantbot or any other organisation’s deployment or investment authority.

The University of Queensland’s ECON7333 Big Data and Machine Learning for Economics and Finance course profile is scheduled for Semester 2, 2026. It names Associate Professor Dong-Hyuk Kim as coordinator and lecturer, teaches supervised and unsupervised learning, support-vector methods, splines, nearest neighbours, and neural networks in R, and assigns a 30% research project with code and a report. The course page also makes the distinction between open assessments where AI tools may be used and a closed-book, identity-verified exam. This creates a useful evaluation and talent-pipeline route, not evidence of a fund’s use of R, its assessment projects, or its AI policy.

Across these sources, the next research queue is to recover project briefs, papers, student-safe code, data provenance, and conference slides; then test break detection, liquidity labels, volatility-surface modelling, and quantile return distributions with point-in-time splits and transaction-cost assumptions. The source material supports those experiments as ideas, not conclusions about any firm.

September 4, 2026 — bank ML-lab and portfolio-rebalancing research routes

Morgan Stanley’s first-party Machine Learning Research page publicly describes an applied research team working across fixed income, investment management, and electronic trading. It names time-series analysis, recommender systems, network theory, LLMs, finance NLP, fairness, and privacy as research areas; lists papers associated with ICLR, ICML, NeurIPS, and AISTATS 2026; identifies Nicholas Venuti as an Executive Director in Technology; and advertises a quantitative-technology ML research internship. This is unusually direct evidence of a bank research-lab structure and hiring surface that can influence the broader quant talent market. It does not identify customer or desk deployment, data permissions, model ownership, or performance, and it is not evidence about any hedge fund.

The recovered Wilfrid Laurier academic seminar page on AI and quantitative investing records an April 30, 2026 announcement for a May 5 hybrid talk by Arun Muralidhar. The page identifies Muralidhar as co-founder of Mcube Investment Technologies, client CIO of AlphaEngine Global Investment Solutions, adjunct professor of finance at Georgetown, and an MIT Sloan economics PhD. Its stated focus is AI-assisted portfolio rebalancing and the use of momentum, valuation, macro, and sentiment inputs. These are speaker and event metadata plus a stated research framing; they do not establish a specific client system, dataset, live authority, or independently verified return result.

The combined idea route is to treat rebalancing as a separately modelled decision layer: estimate predictable flows, distinguish signal construction from implementation, and evaluate turnover, costs, risk, and drift under point-in-time information. That is a research hypothesis derived from the public pages, not a claim about adoption by Morgan Stanley, Mcube, AlphaEngine, or a tracked fund.

September 4, 2026 — professor-led ideas and finance-programme signals

The academic search now includes public pages that expose research directions and training workflows even when neither a fund nor “hedge fund” appears in the title. Mihai Cucuringu’s UCLA/Oxford page lists 2026 work on lead-lag networks, cross-market US/China forecasting, asset pricing with heterogeneous data, limit-order-book foundation models and mixture-of-experts, rejected-order-flow prediction, generative order-book simulation, multi-agent reinforcement learning for execution, and LLM embedding priors. The page also links named papers and identifies Cucuringu as a UCLA Professor of Mathematics, Oxford-Man associate member, and co-organiser of two finance-ML seminar series. This is a professor and paper-discovery route; it does not establish any tracked firm’s data access, model ownership, deployment, or performance.

Columbia’s Spring 2026 Mathematics of Finance practitioners’ seminar records a Goldman Sachs presentation on the mathematical approach to GenAI and a Trexquant session by Denis Lapitski, Head of Strategy Research. The Trexquant abstract says the session covers portfolio construction, mixture-of-experts, alpha-forecast networks, CNNs, contrastive learning, transfer learning, and reinforcement learning in quantitative trading. The same programme lists a Stevens professor’s NSF-funded work on option-surface smoothing and quantum risk measurement. These are speaker abstracts and academic event metadata. They are not independent evidence of model weights, training data, production permissions, or audited trading results.

The current Columbia Business School profile for Harry Mamaysky adds a finance-programme and professor route tied to information-rich research. Columbia identifies him as Faculty Director of the Program for Financial Studies and MS in Financial Economics, with expertise spanning AI, asset management, market microstructure, news, and data analytics. The profile links the Big Data in Finance and Text Data in Finance courses and papers on news, earnings-call transcripts, fund alpha/beta dynamics, and technical analysis. This helps locate papers, syllabi, and potential academic lineages; it does not show a fund’s live system or investment authority.

Georgia Tech’s 2026 Machine Learning for Trading syllabus describes an applied workflow in which students analyse market data, implement and evaluate trading models, and document practical limitations; projects account for 71% of the listed grade. The programme is a talent-pipeline and artefact-discovery signal. The syllabus does not disclose student projects, employer adoption, data licences beyond course infrastructure, or live strategy results.

The ideas exposed by this pass are testable but not ranked: rejected-order and order-flow labels; graph and mixture-of-experts representations; multi-agent execution; embedding priors; and explicit separation of forecasting, portfolio, execution, and risk evaluation. The next evidence step is to recover the linked papers, slides, code, data definitions, and point-in-time evaluation details before connecting any method to a firm.

September 4, 2026 — finance-ML syllabi and professor-programme routes

Rutgers–Camden’s Generative AI and Textual Analysis in Finance syllabus is unusually explicit about the public research workflow it teaches. Professor Tengfei Zhang’s syllabus covers collection from filings, earnings calls, news, corporate websites, social posts, and APIs or web scraping; bag-of-words and topic models; sentence and document embeddings; FinBERT; GPT embeddings and fine-tuning; RAG; multimodal learning; and AI-based hypothesis generation. It also assigns an AI-frontier presentation, creation of a student AI agent, and a project using sources such as RavenPack, Compustat, CRSP, Glassdoor, Reddit, and earnings-call data. The syllabus says students choose and test a finance question with a written report and presentation. This is course-design evidence, not evidence that a tracked manager has access to the same data, uses the same models, or permits autonomous investment decisions.

The syllabus gives a concrete replication route: freeze source timestamps, compare lexicons, FinBERT, embeddings, fine-tuned models, and agents under identical splits, then audit source revisions, transaction costs, and multiple testing. That is an experiment design inferred from the course structure, not a result reported by Rutgers or the instructor.

The Wharton CV for Itay Goldstein adds a professor and programme-leadership route. The April 2026 document identifies Goldstein as Wharton Professor of Finance, Finance Department Chair, co-director of the Stevens Center for Innovation in Finance, and director of the Wharton Initiative on Financial Policy and Regulation; it records a PhD in Economics from Tel Aviv University under Elhanan Helpman. It also lists a 2026 Journal of Financial Economics paper on market feedback, recent work on information sharing and big data in finance, and a 2026 Wolfe Research AI in Finance invitation. These are academic lineage and research-surface facts; they do not connect a paper or talk to a hedge fund’s internal model, data rights, or performance.

Chicago Booth’s AI in Finance executive-education page exposes a current Hong Kong programme scheduled for November 23–26, 2026. Booth lists Dacheng Xiu and Juergen Rahmel, Director of AI Applied Research at HSBC and an HKUST adjunct professor, among the teaching faculty. The outline includes investment decision-making, reasoning models, enterprise deployment, agentic systems, scientific discovery, financial modelling, risk assessment, automation, bias, hallucination, model risk, and governance. This is a public executive curriculum and personnel route; it is not evidence of HSBC or any fund’s production architecture.

The academic search now spans research inputs and model construction, programme and faculty lineage, and executive deployment and governance language. Those routes support retrieval of papers, slides, project artefacts, and biographies while keeping educational content separate from commercial system claims.

September 4, 2026 — Asia-Pacific quantitative-strategy and faculty routes

The University of Queensland’s FINM7104 Quantitative Investment Strategies course profile describes a postgraduate course built around multiple instruments and asset classes, with workshops delivered by practitioners in quantitative funds management. Its public assessment asks students to evaluate a quantitative strategy, consider the size effect, test robustness and implementability, and explain how misleading conclusions can arise. The page also records a policy allowing appropriately referenced AI or machine-translation assistance. This is a strategy-research and evaluation workflow, not evidence of a fund’s live process, data licence, or performance.

HKUST’s 2026–27 MSc in Financial Mathematics catalogue lists MAFS 5440 Artificial Intelligence in Fintech as a hands-on course in machine-learning applications and says financial firms are invited to offer capstone projects combining mathematical models, statistical analysis, machine learning, and emerging financial technologies. The catalogue identifies Kani Chen and Lixin Wu as programme leadership. It does not name participating firms, project datasets, student outputs, or production systems.

PolyU’s potential-supervisor list lists Jie Jay Cao for sustainable finance, quantitative trading, asset management, crypto, and AI in finance; Chishen Wei for LLMs and ML in financial markets; Yong Zhang for media and social-media effects, LLMs, insider trading, and governance; and Wenqi Fan for recommender systems, LLMs, trustworthy AI, graph neural networks, and data mining. These are faculty-stated supervision interests and discovery leads. They do not establish a shared dataset, employer relationship, or commercial model.

UNSW’s FINS5566 Trading in Financial Securities handbook entry adds an Australian quantitative-trading route. The 2026 handbook describes quantitative analysis, algorithmic trading, market microstructure, risk management, and backtesting through practical exercises and real-world applications. The page is indicative and exposes no named instructor, student project, employer partner, or live strategy result.

These Asia-Pacific pages add a cross-check: the public academic layer can reveal the research loop, named faculty interests, capstone interfaces, and AI-use policies without revealing commercial implementation. The next evidence step is to follow faculty papers, capstone announcements, student-safe artefacts, and event recordings while preserving those boundaries.

September 4, 2026 — professor-led ideas and finance-programme routes in Greater China and Japan

The international academic search found several routes that expose model ideas and talent channels without turning them into claims about any tracked fund. IDEA’s official profile for Jian Guo identifies him as the institute’s executive dean and chief scientist for AI finance, lists adjunct or practice appointments at several universities, and describes research in AI quantitative investment, deep and reinforcement learning, and financial knowledge-graph reasoning. The page names four distinct projects: Quant5.0, a financial time-series foundation model spanning exchanges, instruments, and frequencies; e2eQuant, an end-to-end model from market or order-book inputs to orders; Alpha-GPT, an interactive agent for factor mining and quantitative modelling; and Quant4.0, which targets automated feature engineering, deep-learning tuning, deployment, and inference acceleration. It also describes a “financial wind tunnel” that combines generative models with market data, news, and corporate announcements to simulate stress scenarios. These are institute-profile and project-description claims. They do not establish live fund deployment, data rights, model weights, latency, or investment results. The same page lists a 2025 lead-lag paper and an RL-trained natural-language-to-SQL paper, creating concrete paper and code-recovery routes.

Peking University’s April 2026 seminar notice records a talk by BigQuant founder and CEO Ju Liang on “Agentic Quant.” The notice says the talk covered an AI-native “one-person research team,” multi-agent collaboration, and an automated path from data acquisition and factor mining through strategy development, backtest optimisation, and risk control. The same notice identifies Central University of Finance and Economics associate professor Jingyi Wang as deputy director of Beijing’s key financial-AI laboratory and lists courses on financial foundation models and deep learning/large language models. This is a dated university event record and a personnel-discovery route; the notice does not provide a recording, implementation details, or independent validation of the platform’s claimed scale.

Hunan University’s faculty profile for Guohao Tang provides a different kind of bridge. It identifies Tang as an associate professor, doctoral supervisor, and deputy head of the university’s Financial Technology and Engineering department; records a Central University of Finance and Economics doctorate and a Washington University in St. Louis visiting appointment; and describes research in financial technology, financial machine learning, and empirical asset pricing. The profile lists a machine-learning asset-pricing virtual-simulation system, a machine-learning finance textbook, and papers in finance and economics journals. It is evidence of an academic training and teaching route, not evidence that a named manager uses the system or that the academic results transfer to live trading.

CUHK-Shenzhen’s FIN-6130 course page lists a three-credit, graded School of Management and Economics course for the 2025–26 second semester. Its description covers AI and machine learning in fintech, linear and deep learning, quantitative investment, trading, and student projects applying AI in finance. The page also exposes the university’s adjacent AI and data-science schools. This is a programme-level talent-pipeline signal; it does not name students, project sponsors, datasets, or commercial systems.

Japan adds an important modelling and lineage route. Kyoto University’s current profile for Masahiko Egami lists a PhD in Operations Research and Financial Engineering from Princeton, an NYU Courant master’s degree, a Wharton MBA, prior banking experience, and current professorship in financial engineering. The profile’s research focus is stochastic optimisation applied to financial engineering, control, and credit risk, with recent peer-reviewed work on diffusion last-passage times. Chuo University’s profile for Jun-ya Gotoh records a 2026 move into its Department of Data Science for Business Innovation, Tokyo Institute of Technology engineering degrees, and research interests spanning machine learning, sparse optimisation, robust optimisation, portfolio selection, and credit-risk scoring. Its publication list includes work on out-of-sample data-driven optimisation and the interaction between financial risk measures and machine learning. These profiles point to robustness, model misspecification, and risk-sensitive optimisation as research surfaces to follow; they do not identify hedge-fund employment or live strategy use.

The practical discovery implication is to search academic sources for model objects rather than only “AI in finance”: financial time-series foundation models, order-book-to-order systems, financial wind-tunnel simulation, agentic factor research, financial knowledge graphs, robust portfolio optimisation, and credit-risk scoring. The next evidence step is to recover papers, code, talks, student-safe project artefacts, and dated affiliations. No professor, programme, or firm is ranked here.

September 4, 2026 — European academic routes for uncertainty, execution, and market structure

The next academic pass adds several non-duplicate routes. The University of Basel Computational Economics and Finance group describes work combining optimisation, simulation, machine learning, and data analysis for asset management, risk management, high-frequency trading, financial networks, and market stability. This is a group-level research map; it does not disclose a public dataset, codebase, or live-fund system.

Christa Cuchiero’s University of Vienna profile links quantitative risk management with data-driven risk inference, stochastic volatility, robust portfolio optimisation, mean-field games, signature methods, deep neural networks, and reservoir computing. The research route suggests testing path or signature representations alongside robust risk and portfolio models, but the profile does not identify a manager, proprietary data, or production deployment. The Vienna Graduate School of Finance doctoral listing separately names Aron Bodisz, supervisor Nikolaus Hautsch, and a topic on decentralised finance and financial machine learning. It is a supervised research topic, not a published result or fund affiliation.

The University of Amsterdam project page for Oskari Veijalainen frames a doctoral project around return predictability, predictive uncertainty, cross-sectional risk factors, momentum, and investor under- and over-reaction. That creates a testable route for comparing point forecasts with predictive distributions in portfolio decisions. The page does not disclose its dataset, code, results, or employer links.

The ETH RiskLab news page records Moritz Weiss’s July 27, 2026 doctoral examination on machine learning for algorithmic trading and risk management and Urban Ulrych’s announced move to an assistant-professor role in quantitative finance at the University of Ljubljana. ETH’s Spring 2026 course catalogue also lists deep hedging, deep calibration, neural architectures, ML time-series analysis, reinforcement learning, GANs, economic games, and LLMs in finance with coding excursions. These pages expose talent and method vocabulary; they do not disclose student artefacts, data licences, or live trading authority.

The Alliance Manchester Business School doctoral-supervisor directory adds a recruitment and lineage surface spanning quantitative finance, machine learning, deep learning, NLP, embeddings, transformers, LLMs, volatility forecasting, and ESG or text analysis. Names such as Ser-Huang Poon, Eghbal Rahimikia, and Aleksandr Ermakov are listed with relevant interests. The directory is not a project result and does not establish a firm relationship.

Andrei Kirilenko’s Cambridge Judge profile connects a finance professor and Cambridge Centre for Finance, Technology and Regulation leadership with earlier work on dynamic machine learning for discovering electronic-market ecosystems and multiscale models of high-frequency trading. The research idea is to infer interacting trader types and market-impact dynamics jointly. The profile provides publication and lineage evidence, not code, current employer use, or trading performance. Christian Borch’s University of Copenhagen profile adds an adjacent market-structure route through automated trading, ML, quantum computing, interviews with traders, regulators, and developers, agent-based simulations, and NLP collaboration. It is useful for studying strategic adaptation and institutional behaviour, not evidence of predictive alpha or investment deployment.

Two programme and lab pages add further implementation context. Southampton’s 2026–27 Equity Investing with Artificial Intelligence module names Mohamed Bakoush as module lead and covers AI for security selection, portfolio optimisation, factor discovery, risk modelling, and performance evaluation, with critical treatment of assumptions and limitations. The University of Geneva’s 2026 finance-AI chair announcement proposes a role spanning asset pricing, quantitative finance, machine learning, quantitative asset management, and doctoral supervision; its stated application deadline has passed, and the page does not establish the eventual appointment.

Taken together, these routes broaden the recovery queue from return prediction to uncertainty quantification, robust optimisation, deep pricing and hedging, trader-interaction modelling, and human/NLP evidence about automated markets. The next step is to retrieve papers, dissertations, code, course projects, recordings, and dated industry links, keeping academic curricula and research agendas separate from claims about any investment manager.

September 4, 2026 — expanded academic routes: uncertainty, market structure, and talent channels

The next discovery pass adds a set of university and research-group routes that are useful because they expose specific modelling objects rather than generic AI labels. The University of Basel Computational Economics and Finance group describes optimisation, simulation, machine learning, asset management, risk management, high-frequency trading, financial networks, and market stability. Christa Cuchiero’s University of Vienna profile adds data-driven risk inference, stochastic volatility, robust portfolios, mean-field games, signature methods, deep neural networks, and reservoir computing. The Vienna finance doctoral listing names Aron Bodisz, supervisor Nikolaus Hautsch, and a project on decentralised finance and financial machine learning. These are academic routes, not evidence of a fund’s system or performance.

The University of Amsterdam project page for Oskari Veijalainen focuses on predictive uncertainty alongside return predictability, cross-sectional risk factors, momentum, and investor reaction. That supports a concrete comparison between point forecasts and predictive distributions. The ETH RiskLab news archive records Moritz Weiss’s July 2026 doctoral examination on machine learning for algorithmic trading and risk management. ETH’s Spring 2026 catalogue lists deep hedging, deep calibration, neural architectures, ML time series, reinforcement learning, GANs, economic games, and LLMs in finance with coding excursions. The Alliance Manchester Business School supervisor directory exposes named doctoral supervisors and interests across quantitative finance, NLP, embeddings, transformers, LLMs, volatility, and ESG/text analysis.

Two market-structure routes widen the idea set. Andrei Kirilenko’s Cambridge Judge profile links finance research with dynamic ML for discovering electronic-market ecosystems and multiscale high-frequency-trading models. Christian Borch’s University of Copenhagen profile lists automated trading, ML, quantum computing, interviews with traders and developers, agent-based simulations, and NLP collaboration. These pages suggest research on interacting trader types and strategic adaptation; they do not disclose predictive alpha, code, or live deployment.

September 4, 2026 — regional programme and practitioner-linked routes

Singapore and India produced several additional primary routes. The NUS quantitative-finance seminar page records Julian Sester’s robust-portfolio work using Wasserstein-ball and parametric-normal ambiguity sets. An NUS event page describes Lin William Cong’s GOALS work around transformer-based reinforcement learning or panel trees for portfolio management, latent-factor generation, and regime-dependent return predictability. The NTU AI-for-Finance programme lists penalised regression, deep neural networks, LLMs, deep-learning asset pricing, stochastic discount factors, reinforcement learning for trading, NLP, and structural models.

IIM Bangalore’s capital-markets programme exposes Bloomberg/NSMART analysis, simulated trading, and an ML bankruptcy-prediction project. IIM Calcutta’s explainable-AI programme lists XAI, fraud, credit default, portfolio management, and LLM analysis of reports, speeches, and filings. A Curtin Singapore research report describes a self-reported prototype combining market, macro, technical, and LLM-extracted narrative signals, with agents for retrieval, feature engineering, model execution, and reporting. These pages reveal curricula or early project claims; they do not establish production systems.

September 4, 2026 — China, Japan, Korea, and Australia model routes

The Tsinghua profile for Yan Liu exposes financial-text, market-microstructure, time-series, graph-neural-network, reinforcement-learning, and ML asset-pricing research. Central University of Finance and Economics’ Xiangyu Zhang profile adds high-dimensional factors, financial econometrics, forecasting, and network-augmented conditional factor pricing. A University of Tokyo CIRJE paper describes diffusion models for synthetic stock-return distributions, learned reverse-process variance, and regime-conditional tail-risk scenarios. Kanagawa University’s Hideharu Funahashi profile lists deep learning for derivatives pricing and AI-enhanced SABR work.

Yonsei’s Jun Kyung Auh profile combines finance-and-AI faculty work with prior Cubist Systematic Strategies and Deutsche Bank roles, and lists portfolio optimisation, cross-sectional returns, derivatives, structured finance, and corporate bonds. HUFS’s Sihyun An profile lists LLM-based synthetic sentiment indices and financial-news research. UNIST’s Yongjae Lee profile describes a Financial Engineering Lab spanning market modelling, optimisation, ML, and AI. RMIT’s Steven Li profile lists Australian stock-movement and neurosymbolic stock-selection projects, while the University of Sydney’s Chao Wang profile covers realised volatility, Bayesian adaptive MCMC, and ML/deep learning for financial time series. These are personnel and research directions, not claims of employer adoption.

September 4, 2026 — public artefacts to ingest and test

The Yale LLMs and Financial Markets research-assistant page specifies Spring 2026 projects that clean text datasets and analyse them with LLMs on the Yale SOM cluster, with RAG/NLP skills requested. The FinText Hugging Face organisation publicly links a podcast, GitHub loading support, finance-specific time-series foundation-model checkpoints, chronological training collections, and model repositories. The official TSFM_Finance implementation provides notebooks and loaders for Chronos, TimesFM, and FinText checkpoints. The underlying TSFM paper compares zero-shot, fine-tuned, and finance-specific pre-training under its stated protocol. The LOBERT paper adapts BERT-style tokenisation to multi-dimensional limit-order-book messages. These are unusually actionable public research artefacts, but reported counts and results still require independent replication, point-in-time controls, costs, and capacity tests.

The expanded queue now covers robust ambiguity sets, predictive uncertainty, deep hedging and calibration, agent-based market simulation, financial narratives, diffusion-generated return distributions, neurosymbolic selection, limit-order-book tokenisation, and finance-native time-series foundation models. Follow the papers, code, recordings, course projects, and dated affiliations before connecting any method to an investment manager. No firm, professor, programme, or method is ranked here.

September 4, 2026 — Canadian faculty, university archives, and personnel lineages

Canada adds several public research routes that are useful for separating model ideas from firm disclosure. HEC Montréal’s profile of David Ardia connects financial econometrics with machine learning and asset-pricing research; the page is a faculty and publication route, not a disclosure of a manager’s data or production stack. McGill’s event record for Ruslan Goyenko adds an attention-guided deep-learning asset-pricing route. The event page supports the topic and speaker, but not a claim about out-of-sample tradability or institutional deployment.

Waterloo’s repository record for a finance thesis is a dated student-research artifact and a route to methods, data descriptions, and citations. NC State’s profile of Shawn Mankad and CMU Tepper’s profile of Bryan Routledge add faculty routes around financial data, accounting information, machine learning, and empirical research. The public pages establish research interests and lineage; they do not establish a fund relationship or live investment use.

HEC Montréal’s Tolga Cenesizoglu profile and Waterloo’s Peter Forsyth profile extend the search into asset pricing, quantitative finance, numerical methods, stochastic modelling, and computational finance. These are candidate paper and doctoral-supervision routes for testing portfolio, pricing, and risk ideas with point-in-time controls.

University archives expose a different kind of evidence. The Oxford-Man Big Data and Finance Workshop lists talks on cross-sectional learning and stochastic discount factors, LLMs and investor disagreement, LLMs in finance, and multi-agent learning or algorithmic collusion. Two public recordings are available for direct transcript recovery: Chukwuma Dim and Felix Drinkall. The roster and recordings support topic and speaker discovery; they do not show an investment manager’s internal implementation.

The University of Minnesota MCFAM seminar archive adds deep policy-gradient reinforcement learning for S&P 500 option implied volatility and deep recurrent or reinforcement-learning execution on multi-year limit-order-book data. Stanford AFTLab’s Shilong Yang seminar page describes hidden-factor extraction with MLP, RNN, and LSTM models and hierarchical tracking portfolios, with a J.P. Morgan affiliation stated on the page. These are specific research objects and an industry-linked seminar, not evidence of a hedge fund’s current model or results.

UChicago Financial Mathematics career outcomes lists 2025 placements at firms including Marshall Wace, Brevan Howard, Qube, and Squarepoint across AI/ML engineering, data engineering, quantitative research, and trading-systems roles. Duke’s MATH 585 course page describes datasets, Python, risk models, and live-trading projects with industry guests. These pages are talent-pipeline signals and should not be read as evidence that every listed employer uses every method taught or discussed.

One additional lineage route is Harvard Biophysics’ 2003–04 graduate alumni page, which lists Alexander Fotin as Head of Quantamental Research at Marshall Wace. It is useful for person-level verification and publication searches, but the alumni page alone does not establish the scope of that role, its AI methods, or any current firmwide policy.

These routes make the academic layer more operational: faculty profiles generate model hypotheses, repositories and recordings provide inspectable artifacts, and career pages expose talent channels. The evidence still has to be recovered and tested separately from claims about any investment manager. No firm, professor, programme, or method is ranked here.

September 4, 2026 — explicit course projects and AI-generated finance theories

Tsinghua’s AI + Quantitative Finance course page is unusually specific about the intended student work. The graduate course links portfolio theory, factor models, time series, risk models, stocks, futures and derivatives with ML/DL, LLMs, and agents. Its schedule names alternative-data factors, factor-zoo work, genetic-programming factor mining, OpenFE, and introductory LLM/agent methods; grading includes coding homework and a course project with a substantial code component. This exposes a talent and idea-supply route, not a commercial strategy, data licence, or evidence of live trading.

Wilfrid Laurier’s Sina Seyfi profile exposes a public research-and-code surface around AI, machine learning, and asset pricing. The page describes AutoTheory, an LLM-plus-evolutionary-search system that generates economic theories, mathematically audits them, translates pseudocode to code with review, calibrates and scores candidates, and uses simulated peer review. It names applications to price-multiplier and equity-term-structure puzzles and links papers, presentations, data, code, and public talks. The page also reports performance for several of Seyfi’s research projects; those are author-reported claims that require independent replication, point-in-time data controls, multiple-testing correction, and cost or capacity analysis before being treated as an investable result. The source establishes the public artifact and research agenda, not hedge-fund deployment.

Xi’an Jiaotong-Liverpool University’s FIN426 module catalogue adds a 2026–27 Level 4 course titled Large Language Models and Digital Finance Analytics. The catalogue describes weekly lectures and practical seminars using LLMs and textual analysis on real-world finance problems. It is a curriculum and recruiting signal; it does not identify student outputs, data vendors, model weights, or production systems.

Together, these routes make the academic search more concrete: course projects can reveal implementation conventions, while public research pages can expose the chain from hypothesis generation to mathematical audit, code review, calibration, and evaluation. None of the sources supports a cross-firm ranking.

September 4, 2026 — US programme and lab routes with explicit implementation detail

MIT Learn’s AI and Finance course provides a public bridge between finance education and the lab ecosystem. The self-paced course frames reinforcement learning and large language models across financial narratives, market learning, investment analysis, trading, and language tasks, while explicitly covering bias, interpretability, accountability, and responsible deployment. The page identifies Andrew W. Lo as an instructor and describes his roles as MIT Sloan professor, director of the MIT Laboratory for Financial Engineering, and principal investigator at CSAIL. This is a public course and faculty route; it does not establish a particular manager’s use of RL or LLMs.

The USC Capital One Center for AI in Finance’s 2026 fellows page lists Duygu Nur Yaldiz, Ke Xu, Yuxin Yang, Mohammad Shahab, Tejas Srinivasan, Yuan Xia, Yutai Zhou, and Shaoyu Wang. The page says the center supports AI and ML research applied to finance by Viterbi School faculty, but it does not attach a project title, dataset, model, sponsor deliverable, or result to each fellow. The roster is therefore a lab-personnel discovery route, not evidence of any fellow’s specific research output or commercial deployment.

Boston University’s finance catalogue lists QST FE 555, AI for Financial Analysis. The course description names LLMs, AI-assisted research, ethical considerations, and practical workflows for financial decision-making. The same catalogue exposes adjacent quantitative and investment analytics courses, which makes it a useful programme-surface route for finding syllabi, instructors, and student-safe projects. The catalogue does not identify a production system, data licence, or fund partnership.

KIT’s HECTOR School Fundamentals of Financial Machine Learning course is more implementation-specific. Its June 2026 course page names ARMA, GARCH, Kalman filters, state-space models, factor models, volatility forecasting, risk premia, backtesting, cross-validation, and hyperparameter tuning. It identifies Maxim Ulrich, Professor of Risk Management and Financial Economics at KIT, and describes academic and practice links including the ECB and Eurex. The page is an executive-education description; it does not establish participant access to those institutions’ data or systems.

The University of Cincinnati’s Applied AI in Finance graduate certificate combines financial econometrics, quantitative equity investing, factor backtests, algorithmic trading, NLP and LLMs, return prediction, risk management, portfolio optimisation, and paper-trading implementation. Its curriculum is a useful test design because it places data preparation, factor construction, model evaluation, and trading constraints in one sequence. It is a programme description, not evidence of student performance, employer adoption, or live capital.

These routes add two useful dimensions to the academic map: a named AI/finance lab with a current fellows roster, and curricula that expose concrete transitions from classic econometrics and factors to LLMs, agents, and responsible deployment. They remain talent and research signals, not firm rankings.

September 4, 2026 — academic-to-market lineage routes

Harvard CARES’ profile of Guillaume Basse records a Harvard Statistics PhD advised by Edo Airoldi and a subsequent Stanford faculty route before a Citadel Securities role. The public profile supports a causal inference and experiment-design lineage, including interference as a research concern; it does not disclose what the firm role covers or any trading application.

MIT Sloan’s profile of Juan Antolín-Díaz links an LBS PhD, an AQR fellowship, ECB work, and the founding of a quantitative research group at Fulcrum Asset Management. His public research route is Bayesian econometrics and macro or asset-pricing analysis. The biography and publications do not establish a current firm remit, model inventory, or performance.

Bodhisattva Sen’s Columbia CV records a faculty and doctoral-supervision route whose public placement evidence includes a Columbia PhD advisee moving to Two Sigma quantitative research. The research themes include nonparametric empirical Bayes, optimal transport, kernel methods, and distribution-free inference. This is a lineage and methods route; the CV does not disclose the researcher’s current scope or how those methods are used at the firm.

Yale’s event page for Madalina Persu describes a Harvard-to-MIT-to-Two-Sigma path and identifies her as a quantitative researcher and vice president. The page frames the research route around AI for predictive mid-frequency equity signals. It is an event biography, not a paper, model specification, data disclosure, or performance record.

Wharton doctoral career outcomes provide a programme-level route into firms including Cubist, Millennium, Trexquant, Tudor, Citadel Securities, Two Sigma, AQR, Jump, and Chicago Trading. The page names employers and placement years but not graduate names, advisers, roles, papers, or hiring causality, so it should be used for discovery rather than inference.

The Oxford-Man machine-learning-in-quantitative-finance seminar page lists Hans Buehler of XTX Markets and Slavi Marinov of Man Group alongside academic speakers, with themes spanning ML applications, execution, regulation, and market data. A separate Oxford-Man funding announcement describes Man Group support for the institute and identifies Stefan Zohren as a Principal Quant and deputy director. These pages establish institutional links and research themes, not a map of internal models or evidence of investment results.

Stony Brook’s quantitative-finance programme page records Robert Frey’s route from an applied-mathematics PhD through Renaissance Technologies and back to academia, where he helped build the quantitative-finance programme. The page describes student projects using tick data across equities, commodities, indices, and derivatives. It does not establish current Renaissance involvement, student placements, or strategy performance.

UChicago Financial Mathematics’ Jeff Greco profile adds a practice-to-classroom route through Citadel risk management and UChicago instruction in hedging, volatility and return distributions, and managed-risk strategies. A UChicago event with John Overdeck adds a public D.E. Shaw-to-Two-Sigma co-founder lineage route with an AI and quantitative-investing discussion. Neither page discloses production architecture, proprietary data, or live portfolio authority.

These routes connect methods, training institutions, and public industry affiliations without collapsing a professor’s research into a firm’s internal strategy. Each connection remains a lead for paper, talk, and role recovery, not a comparative judgement.

September 4, 2026 — regional academic artifacts and finance-ML curricula

IIT Bombay’s August 2026 Deep Learning in Finance event describes a presentation by Ovidiu Calin on CNNs, recurrent models, LSTM/GRU architectures, and Transformers for stock-price prediction, including hands-on financial-data training, hyperparameter selection, loss visualisation, and forecasting. The university event record establishes the topic and date; it does not provide a recording, code, dataset, or evidence of predictive performance.

Peking University’s Financial Engineering Lab teaching page names Li Xinping as the lab lead and instructor for its 2025–26 “Machine Learning and Asset Pricing” course, alongside Wang Xi and Wang Ranran. Li’s page-level biography describes prior quantitative-investment roles, Stanford economics training, and research in quantitative trading, asset pricing, ML, and financial text analysis; the page’s biography claims remain self-reported. The syllabus spans Chinese market and fund data, factor and Barra models, HFT, autoencoders, RNNs, CNNs, GANs, GNNs, Transformers, LLMs, financial-image analysis, and a millisecond-data API. A separate empirical-finance course names Zhao Liuyan and Li Xinping. The lab also advertises industry-linked talks, including Optiver technology leadership, BigQuant’s agentic-quant presentation, and Chinese brokerage quant researchers. This is unusually detailed talent and curriculum evidence, but it does not independently validate the code, data rights, or resulting strategies.

National Taiwan University of Science and Technology’s ML and Financial Applications syllabus names Chou Chih Lung as lecturer for the three-credit course. Its stated objectives cover stock-price and yield-curve forecasting, fraud and loan-default classification, PCA and clustering for allocation, pairs trading and risk stratification, and reinforcement learning for trading, hedging, and portfolio construction. The outline names ARIMA, deep-learning time-series methods, derivatives pricing, Alpha Vantage, CMoney, and Taiwanese financial databases, and says practical examples use Python. This is a dated syllabus, not evidence that the tools remain available or that student models traded capital.

POSTECH’s Bong-Gyu Jang profile adds a Korean faculty route covering asset pricing, credit risk, derivatives, ML/NLP in finance, and an interpretable deep-learning stock-returns model. The profile exposes a paper and researcher lead; it does not expose code, current data, employer use, or live deployment.

The Sogang University 2025 ML Finance Lab repository is an inspectable educational artifact. Its notebooks and materials cover time-series modelling, backtesting, causal factor investing, alternative financial data, deep learning, and agent-LLM workflows. The repository is useful for reconstructing how these methods are taught together; it does not establish audited results, current data-pipeline access, or live-trading permissions.

Keio’s asset-price-formation report describes using statistical and ML methods on R&D-related text and corporate activity for domestic listed companies to study firm value and asset-price formation. The report is a dated university artifact; its model families, data access, and out-of-sample results require document-level recovery.

CMKL University’s 2026 seminar on RL for option pricing names Chaniporn Nerunchorn, an MS AiCE student, as the April 3, 2026 speaker. The hybrid seminar frames reinforcement learning as a decision method for option-pricing under uncertainty. The page does not provide slides, a recording, the state/action space, reward function, market data, or experiment results, so the route is useful for speaker and topic discovery only.

The Vietnamese-German University thesis on ML in finance and economics compares ARIMA, XGBoost, LSTM, attention-LSTM, and hybrid models with VN30/ACB data, volume, macro variables, and technical indicators. It is a public regional artifact and a route to data and citation recovery, not independent evidence of tradable alpha or production use.

Fudan DISC-FinLLM’s English README describes a four-module Chinese financial LLM built from Baichuan-13B-Chat with LoRA experts for financial advice, document analysis, financial calculation, and retrieval- enhanced current-affairs analysis. The README reports about 246,000 instruction examples across consultation, financial tasks, computing, and retrieval-enhanced subsets; its retrieval subset is generated from a knowledge base containing 18,000 research reports and 69,000 financial-news items. It exposes Black–Scholes and EDF calculations through tool calls, and says the project released samples, weights, and a finance benchmark. These are lab-reported design and release claims. Data provenance, licensing, benchmark reproducibility, current model behaviour, and investment use remain unverified.

This regional pass adds model-training artifacts, course-level tool choices, and language-specific research surfaces that generic English-language searches tend to miss. They are evidence for discovery and replication queues, not comparative judgements about firms or schools.

September 4, 2026 — additional US and Nordic finance-ML programme routes

NYU’s FRE-GY 9073 syllabus places optimal investment and consumption, optimal execution, reinforcement learning, stochastic control, diffusion models for financial time series, and LLM fine-tuning in one finance course. It is a detailed teaching artifact with an instructor and project surface; it does not establish completed results, deployable models, or any fund’s use of the material.

Columbia’s MS Financial Engineering curriculum adds explicit programme routes for causal inference, AI applications in finance, networks and contagion, deep learning, big data in investment research, model-based trading, and algorithmic trading. Illinois’s MSFE academics page lists Machine Learning in Finance and a corporate-sponsored practicum for real-world financial modelling. Neither programme page supplies detailed assignments, sponsor identities, student models, deployment evidence, or performance.

NTNU’s BBAN4010 course page describes a 2026/27 course requiring a student project applying reinforcement learning to a self-selected finance or economics problem. This is a future-course signal: it does not yet expose the projects, datasets, readings, or outcomes.

MIT OpenCourseWare’s AI and machine learning in financial services module provides a dated finance-specific teaching surface covering NLP, chatbots and voice interfaces, compliance, risk management, underwriting, and investment strategies. The lecture and reading page does not establish a production system or investment performance.

Open research infrastructure adds inspectable implementation vocabulary. The TradeMaster-NTU repository describes multimodal market data, data-driven market simulators, more than 13 reinforcement- learning algorithms, and tutorials spanning DJ30, BTC, SSE50, and limit-order-book tasks. FinRL provides market environments, deep-RL agents, benchmarks, portfolio and stock-trading tutorials, and paper-trading examples. FinGPT documents sentiment, relation extraction, NER, headline analysis, RAG, LoRA fine-tuning, and financial instruction datasets. FinRobot separates SEC, market, and news ingestion from research, valuation, debate, synthesis, and report generation. These repositories are research or educational projects; their READMEs do not establish proprietary data rights, production controls, live authority, or investment efficacy.

Microsoft’s TimeCraft repository adds a financial-market simulation and controllable financial time-series-generation route, including diffusion and foundation-model methods and meta-agent guidance, with links to ICLR 2025 and AAAI 2026 work. The repository and linked papers require separate validation of market-microstructure preservation and downstream trading value.

The current repository surfaces also make the personnel and architecture trail more precise. TradeMaster identifies Sun Shuo of HKUST (Guangzhou) as first author and advertises research positions under his supervision; its public materials describe automatic feature generation, diffusion-based financial-data imputation, market simulation, and six evaluation axes with 17 measures. FinRL now explicitly separates the original educational/research framework from FinRL-X, which it describes as the next-generation AI-native, modular, production-oriented project. FinGPT’s repository credits Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang for its original paper, while FinRobot’s citation block names Hongyang Yang, Boyu Zhang, Neng Wang, Cheng Guo, Xiaoli Zhang, Likun Lin, Junlin Wang, Tianyu Zhou, Mao Guan, and Runjia Zhang among its authors. These are public repository and paper-lineage facts; they do not establish current institutional roles, proprietary data, production permissions, or investment results.

This pass expands the idea-origin map across execution, causal inference, simulation, voice interfaces, financial language models, and agentic research workflows. It does not compare schools, firms, or methods.

September 4, 2026 — further US professor and programme routes

NYU Stern’s Robo Advisors and Systematic Investing syllabus identifies Vasant Dhar and describes the use of price, fundamental, and unstructured news data in systematic computer-based models, with Python/Excel templates and machine-learning-informed projects. The syllabus establishes course content and evaluation intent; it does not establish current fund operations, proprietary data, or performance.

An NYU Tandon Peter Carr seminar describes Bruno G. Kamdem’s work with ESG time-series factors, a six-factor Fama–French model, shallow neural networks, explainability, and comparison with London Stock Exchange ESG ratings. The event abstract does not provide code, full data, out-of-sample results, or live portfolio use.

UT Austin’s Clemens Sialm profile lists a May 2026 paper, The Growth and Performance of Artificial Intelligence in Asset Management, with Shuang Chen and David Xu, alongside research on mutual funds and institutional trading. The public profile identifies a paper route but does not expose its model family, sample, data, or findings.

Michael Sury’s McCombs profile describes teaching in analytic finance and ML, student investment-strategy projects, peer examination of analytical steps, and prior financial-services and family-office work. This is a teaching and biography route; it does not establish a public strategy, codebase, investment mandate, or returns.

Cornell’s ORIE 5256 roster places Marcos López de Prado in a financial-engineering course described as teaching scientifically sound ML tools used in financial industry. His Cornell profile also records an ADIA quantitative-R&D and Cornell route. The roster does not expose algorithms, datasets, student output, or current employer authority.

Cornell M.Eng. project coverage describes sponsor-linked projects across equities, fixed income, FX, and digital assets using AI trading, NLP macro forecasting, market microstructure, and alternative-asset prediction. It names sponsor categories but not individual project datasets, code, or sponsor deployment.

Georgia Tech’s Xindi He profile describes combining holdings and textual data with theory, experiments, ML, and LLMs to study investor deviations from rational benchmarks and effects on asset prices. The profile does not expose a completed implementation, dataset, or investment use.

NYU’s Volatility and Risk Institute conference page lists a 2026 AI and risk-management conference with Robert Engle, Richard Berner, Michael Hsu, William Goetzmann, Bryan Kelly, and Marcos López de Prado among the speakers. Its public agenda covers AI and financial stability, AI research, inefficient news pricing, and causal factor-investing protocols, and says full videos are available. The event page does not establish speaker agreement, implementation detail, or investment performance.

These routes add faculty methods, course design, sponsor-linked student projects, and conference recordings to the idea-origin map. They remain separate from evidence of any firm’s internal deployment or comparative standing.

September 4, 2026 — Europe and UK finance-ML research routes

Oxford’s Mathematical and Computational Finance research page lists work spanning deep learning for continuous-time finance, risk estimation, option pricing/calibration/hedging, data-driven portfolio choice, optimal execution, limit-order-book models, and ML-based liquidity equilibria. The page identifies Sam Cohen, Rama Cont, Blanka Horvath, Jan Obloj, Álvaro Cartea, Leandro Sanchez-Betancourt, Justin Sirignano, and others. It is a research-group map, not evidence that one fund uses all listed methods.

Oxford’s LOBIN project adds an implementation artifact: Xinpeng Hong, Changgang Zheng, Stefan Zohren, and Noa Zilberman map limit-order-book features and trained ML models into P4-programmable switches using BMv2 and Intel Tofino/Tofino2. The page reports microsecond-scale latency and a server-benchmark comparison, but this is systems research rather than a disclosed trading strategy or P&L result.

The Oxford-Man ML-for-Finance reading archive records a 2018–19 reading group led by Matthias Qian and Jan Calliess, covering ML portfolio optimisation, predict-then-optimise, text-based return prediction, online market making, deep-learning asset pricing, tail dynamics, and trading-oriented dynamic-mode decomposition. These are dated seminar routes, not current Oxford-owned systems.

EPFL’s Finance and Technology Programme combines asset pricing, risk, microstructure, and game theory with data science, ML, AI, computational science, and cryptography, and records support from Swissquote and Liquity. Its publication index links work on topological-data-analysis clustering for sparse S&P portfolios, term-structure estimation, StockTwits sentiment, causal neural networks, and ML portfolio valuation and risk management. Programme and publication pages do not establish partner use or production deployment.

The EPFL-linked Discount Bond Database, by Damir Filipović, Markus Pelger, and Ye Ye, provides US Treasury zero-coupon yields, discount-bond returns, excess returns, four term-structure factors, and market- complexity measures from 1961 onward, with public companion examples and bond-return code. The derived data and code are public; raw CRSP extraction requires WRDS access.

LSE’s Probability in Finance and Insurance group lists Giulia Livieri’s high-frequency econometrics and statistical/ML time-series work, Gelly Mitrodima’s Bayesian quantile and risk modelling, and Yiwei Wang’s ML/time-series finance, calibration, and derivative-pricing interests. The LSE repository record for a 2026 deep-surrogate option-pricing paper links implementation code and describes high-dimensional structural option models, daily re-estimation, tail risk, liquidity, and parameter instability. The record does not establish live use or investment performance beyond the paper’s own evidence.

Warwick’s repository record for a 2021 active-portfolio review surveys ML applications in signal generation, portfolio construction, and execution, with attention to reinforcement learning and mixed active-ETF results in the examined sample. HEC’s public research route for a 2026 working paper on LLM randomness studies repeated sentiment classifications of corporate filings and variation from sampling, silent model updates, rounding, and routing. It recommends treating outputs as draws and reporting reproducibility details. Neither route establishes a production financial language model or a fund’s implementation.

These European routes add hardware-aware market-data processing, public datasets, financial ML methods, and measurement discipline to the academic map. They do not compare institutions, firms, or methods.

September 4, 2026 — Asia-Pacific finance-ML professors and programmes

NUS’s MSc Digital Financial Technology core-course document lists supervised ML, time-series forecasting, NLP, unsupervised learning, ML-derived systematic trading and backtesting, and agentic/generative-AI risks in finance for the August 2026 intake. It is a programme document; instructors, projects, and outcomes are not identified.

SMU’s Peng Liu profile lists Bayesian optimisation for Sharpe-ratio objectives, reinforcement learning for portfolio optimisation and derivative hedging, financial text mining, human-in-the- loop risk analytics, and explainable neural networks. SMU’s 2025 finance research camp programme also lists an LLM study of mutual-fund skills and a 30-year daily hedge-fund-trades study. The public pages expose methods and titles, not datasets, recordings, or empirical findings.

CityUHK’s P-Trees announcement names Professors Guanhao Feng and Jingyu He and describes a machine-learning model for asset pricing intended to improve portfolio-construction prediction and interpretability. This is a university-reported summary without code, data, benchmarks, or deployment evidence.

PolyU’s HedgeNet seminar abstract describes a model mapping features directly to option-hedging strategies and minimising hedging rather than pricing error, tested on end-of-day and tick prices for S&P 500 and Euro Stoxx 50 options. The abstract reports a comparison with Black–Scholes while also identifying a leverage-aware linear-regression baseline; the document does not link code or a recording.

NTUB’s Traditional-Chinese faculty profile for Chih-Hsi Hsieh lists AutoGluon for USD exchange-rate trading, factor analysis for high-dividend ETFs, PCA and autoencoders for NTD interest-rate-swap curves, automated ML for Taiwan futures, random forests for index options, and deep learning for option pricing. NCCU’s Luo Bing-Zheng profile adds characteristic-free Taiwanese risk-premia capture, theory-guided exchange-rate forecasting, and transformer knowledge-distillation for financial time series. These profiles list publication or conference routes without exposing datasets, code, or out-of-sample results.

OpenXAI’s Korean automatic-stock-trading repository contains a dated XGBoost workflow using Reuters OHLCV data for Korean stocks and KOSPI, historical train/validation/test periods, backtesting, and real-time simulation. It is an old public code artifact with no evidence of live deployment or independently verified returns.

Hitotsubashi’s Toshiaki Watanabe profile connects Finance and Data Science teaching with daily and intraday modelling of volatility, option pricing, VaR, and expected shortfall, while describing machine learning as a future extension to a Bayesian approach. This is explicitly an academic research direction rather than a completed ML deployment.

TongjiFinLab’s CFGPT repository describes Chinese-finance continual pretraining and supervised fine-tuning over prospectuses, announcements, research reports, news, and social media, with sentiment, event detection, summarisation, risk alerts, investment suggestions, and stock-movement tasks. The repository reports that CFGPT3 training code is available while weights are not open-sourced; corpus sizes and benchmark claims remain self-reported.

XJTLU’s Financial Machine Learning module for 2026/27 covers finance-data feature construction, signal generation, trading algorithms, scenario evaluation, Python implementation, production-line deployment, and risk-management interpretation. It is a module specification without a named lecturer, student project, dataset, or deployed system.

The University of Melbourne’s 2026 FDU programme lists Zhongtian Chen’s “Memory and Beliefs in Financial Markets: A Machine Learning Approach,” with Bryan Lim as discussant. The programme verifies the session and participants but provides no method, data, recording, or transcript.

The University of Canterbury’s Mithushana Ravindran profile describes a PhD project on ML-based early-warning models for financial risk across banking, corporate, and sovereign institutions, supervised by Kuntal Das and Mona Yaghoubi. The profile does not specify algorithms, datasets, or predictive results.

LNMIIT’s Argha Das profile records Morgan Stanley and Deutsche Bank experience through 2026 and describes applying AI/ML to risk frameworks, process automation, and quantitative strategies with trading desks. This is a practitioner biography and interest statement, not a paper, dataset, or reproducible result.

The Asia-Pacific pass adds local curricula, multilingual faculty surfaces, open code, and model ideas across portfolio construction, options, risk, text, and trading. It does not compare regions, institutions, firms, or methods.

September 4, 2026 — undercovered regional finance-ML routes

Universidad Austral Rosario’s quantitative-finance diploma starts in August 2026 and lists GARCH, Monte Carlo, VaR, Black–Scholes–Merton, Random Forest, XGBoost, SHAP, LSTM, Python, and real financial-market data, with Rodrigo Del Rosso, Braian Drago, and Ezequiel Nuske named as faculty. It is a Spanish curriculum and faculty route, not evidence of cohort activity, proprietary data, live trading, or hedge-fund deployment.

Universidad de los Andes’ Andrés Mora Valencia profile lists 2024–25 work on ML forecasting of skew, S&P 500 ETF returns, VIX, carbon, and oil returns, including Bayesian deep learning. The profile supplies publication metadata without code, dataset detail, performance validation, or fund affiliation.

Universidad Peruana Cayetano Heredia’s finance data-science seminar report identifies Nelson Castro, a faculty researcher and financial engineer at Banco de Crédito del Perú, in an April 2023 seminar on data science for financial-risk management. The university report does not expose a transcript, model specification, dataset, or deployment evidence.

HEC Montréal’s Canada Research Chair in Finance and Technology names Vincent Grégoire and lists ML/big data, textual data, AI effects on markets and information diffusion, cyber-resilience, high-frequency trading, and algorithmic trading. HEC’s applied-finance ML course adds structured prices/returns, unstructured text, nonlinear filters, ML, and high-frequency data. These are chair and curriculum surfaces without a named production model, dataset, or investment-firm deployment.

The University of Calgary’s Alexandru Badescu profile connects financial econometrics, quantitative finance, actuarial science, and ML with a Winter 2026 financial-economics course and publications on GARCH option pricing, hedging, and tree-based algorithms. It does not establish hedge-fund affiliation or live-system use.

American University of Sharjah’s Anis Samet profile records a faculty grant titled “Assessment of Machine Learning Models for Liquidity and Stock Market Predictions,” funded from June 2022 to May 2025. This is a university personnel and grant record; it does not expose specific models, datasets, results, or continuation.

FUOYE’s Tochukwu Timothy Okoli profile lists digital finance, fintech, financial econometrics, risk, AI, and digital transformation, alongside GARCH/VaR work and an AI/Africa conference route. The public biography has unresolved current-employment detail and does not show ML code or production finance deployment.

National Ain Shams University’s Financial Technology programme lists data analytics, machine learning, blockchain, cryptocurrency, algorithmic trading, Python, R, and applied financial-scenario projects. It is a programme-level description without named professors, a dated syllabus, datasets, or deployed artifacts.

These routes add Spanish-, French-, and region-specific academic surfaces across risk, market prediction, text, liquidity, and algorithmic trading. They are coverage and discovery evidence, not comparative assessments.

September 4, 2026 — additional European datasets, theses, and programmes

Aalto’s company earnings-call dataset offers quarterly JSON data from January 2002 through September 2024, with company metadata, speaker/session labels, transcript text, and Azure Search, plus MATLAB, Python, and R tutorials. It is a concrete route for earnings-call language, speaker, and event research, but access and Refinitiv licensing constraints remain explicit; it is not automatically a freely redistributable dataset.

Aalto’s Quantitative Finance and Machine Learning Applications course combines financial-data analysis, empirical asset pricing, portfolio programming, and basic ML applications. Sina Seyfi’s 2026 doctoral thesis adds interpretable nonparametric ML, characteristic-space similarity, centroid assignment, basis portfolios, and nearest-neighbour return prediction. Its reported FF5-alpha result is dissertation evidence, not independent replication or deployment.

Stockholm School of Economics’ semantic-regression thesis describes supervised-LDA company-characteristic classification, interpretable sentence-cluster semantic regression, and latent risk-factor analysis. The SSE Center for Data Analytics lists ML, neural networks, decision trees, Bayesian classifiers, simulation, forecasting, causality, data quality, and text/number analysis. These sources expose methods and a research centre, not a specific asset-manager deployment.

UZH/ETH’s quantitative-finance thesis archive lists 2026 projects on crypto limit-order-book prediction, LLM macro signals for dynamic allocation, LLM earnings-call sentiment and short-term volatility, heterogeneous-agent asset pricing, multi-agent hedge-fund systems, ML execution, and RL market making, with named supervisors including Erich Walter Farkas, Thorsten Hens, Markus Leippold, Yucheng Yang, Patrick Cheridito, and Josef Teichmann. The archive provides titles, presenters, supervisors, and dates, not code, datasets, or results.

UZH’s 2023 Quantitative Finance Winter School schedule lists ML risk prediction, portfolio optimisation, dynamic-factor risk estimation, news variables, volatility changepoints, shrinkage covariance, NLP, ClimateBERT, and greenwashing detection. Its Finance for Risk Management syllabus adds deep hedging, scenario generation, digital-asset risk, and climate/transition risk. These are dated teaching artifacts without implementation or performance claims.

Michael Rockinger’s Lausanne profile describes financial econometrics and computational finance work using sustainability and bank-risk text, neural networks for time-series and multivariate forecasting, term structures, pension funds, and systemic risk; his CV was updated in May 2026. The profile does not identify production datasets or live portfolio use.

Tilburg’s Quantitative Finance programme lists empirical finance, risk regulation, valuation, data science, uncertainty, time series, and DeFi, with examples involving isolation forests and an ML climate- transition-risk index. Its SoFiE school archive records 2025 financial-ML teaching by Bryan Kelly, Dacheng Xiu, and Semyon Malamud, plus an ML-for-economics-and-finance school with Hui Chen, Simon Scheidegger, and Fabio Trojani. Programme and event pages do not establish fund deployment, public code, or reproducible trading results.

These routes add licensed datasets, thesis-level idea discovery, semantic finance, LLM/LOB project surfaces, and European finance curricula without comparing institutions, firms, or methods.

September 4, 2026 — regional professors, finance programmes, and inspectable artifacts

The regional-language pass adds several academic routes that sharpen the idea map without turning course or faculty descriptions into hedge-fund claims.

Wang Yuan’s SUSTech profile records a SUSTech–Zhongliang joint quantitative lab spanning fundamentals, options and volatility, behavioural finance, machine learning, macro cycles, information theory, credit, liquidity, stock returns, and AI. It also lists prior Penn State training and papers in Management Science, Journal of Banking and Finance, Decision Support Systems, and pattern-recognition venues. This is a public professor/lab and academic-lineage route; it does not disclose a fund model, live data, or production deployment.

Makoto Naito’s KUAS profile connects a Japanese finance-engineering lecturer to a public Head of Domestic Equity Strategy role at Rheos Capital Works. The listed papers cover deep-learning factor models for Japanese equities, firm-characteristic portfolios, ESG and human-capital signals, and asymptotic-expansion methods for ML portfolio problems. The profile does not establish code, point-in-time data, costs, or live authority.

The Tokyo Metropolitan University AI and Finance symposium programme names MUFG/Japan Digital Design CEO Renpei Iwata on GenAI transformation, Nissay Asset Management Data/AI and Quant Research Head Toruki Kanoko on GenAI in asset management, and Osaka Metropolitan University Professor Kei Nakagawa on inductive bias in empirical finance. It is event and personnel evidence, not a model or deployment record.

Chung-Ang’s finance–AI convergence guide connects finance/accounting with ML, big-data analysis, NLP, and AI coursework. Excelia’s Miia Chabot profile adds BERT/NLP, ConvLSTM, urban digital twins, and climate-risk teaching with Python, Bloomberg, and Refinitiv. Both are faculty/curriculum routes; neither identifies a production investment system.

Bogazici’s 2026 report is a more concrete research-design lead: it describes annual-report topic extraction, Qwen2.5-14B factual anchors, GPT-4o labels for a 20,000-document corpus, RoBERTa classification, and Chinese A-share data from 2010–2024. The report does not supply a reproducible corpus, prompts, split, code, or live-investment link. A September 2025 UPV thesis reports Ridge, XGBoost, and DNN factor stock selection using LSEG and Nasdaq data; its reported CAGR remains thesis/ backtest evidence requiring a point-in-time, cost, and leakage audit.

University of Coimbra’s 2026/27 course lists real financial datasets, Python, a broker simulator, AI trading, and backtesting. UGM’s dissertation record lists MLP, BiCuDNNLSTM, and Cascading-MARS/DNN for Indonesian stock-return direction. These expose implementation vocabulary, not student results or production access.

The OJK OSIDA-PMDK release adds a supervisory route: Indonesia’s capital-market big-data platform is exploring graph, ML, and GenAI for surveillance. The Arab Monetary Fund’s 2026 catalogue similarly maps preprocessing, time-series validation, supervised/unsupervised ML, and AI applications in banking, investment, and risk. Neither source establishes a hedge fund’s investment authority or model inventory.

The implementation search also found public artifacts that can anchor future replication work. DeepMarket/TRADES exposes diffusion/ CGAN limit-order-book simulation with PyTorch, Lightning, W&B, and ABIDES references. LOBFrame covers LOB processing, HLOB/DeepLOB replication, simulation, and evaluation. Microsoft MarS describes a generative foundation-model market simulator with order models, background agents, and stylised-fact tests. These are inspectable research surfaces; venue fidelity, calibration, and profitability remain unestablished.

QuantReplay provides an Apache-licensed multi-asset order-driven simulator with FIX, auctions, replay, and randomized orders. StockSim describes multi-agent LLM market simulation with order books, slippage, impact, news/fundamental analysts, and decision traces. Twelve Data’s World Model Dataset specifies temporal train/validation/test periods with OHLCV, indicators, macro variables, text prompts, and trajectories. Data rights, model access, and out-of-sample validity must be checked separately.

TAT-QA, FinQA, UniFinEval, Finam-FinBench, and FinanceBench provide distinct evaluation surfaces for financial-report arithmetic, multimodal finance questions, technical and derivatives tasks, and filing-grounded evidence. They test document or model behaviour, not investment returns. Notre Dame’s SRAF repository adds the Loughran–McDonald dictionary, SEC/EDGAR text, CRSP files, and finance text-analysis materials, subject to academic/noncommercial and commercial licensing boundaries.

Together, these additions connect faculty lineage, taught methods, concrete research designs, simulator code, document benchmarks, and data-access constraints. They remain idea and diligence routes; they do not establish that any named fund uses a particular method or that any academic result transfers to live capital. See the expanded academic source note.

September 4, 2026 — Hong Kong research conference and agent-evaluation routes

CityUHK’s report on its 5th Hong Kong Conference on FinTech and AI in Finance says the June 14–15 conference brought roughly 100 academic and industry participants. It names Allan Timmermann’s keynote on computer complexity and scaling laws of return predictability, Olivier Scaillet’s double-ML carbon-emissions work under sample-selection bias, and Markus Pelger’s multi-agent-RL study of AI traders and market instability. The university report provides speaker and research-title evidence, not the papers, recordings, code, or investment deployment.

FinToolBench, submitted in March 2026 and revised August 1, defines a runnable financial-tool benchmark coupling 760 executable tools with 295 tool-required queries. It evaluates timeliness, intent type, and regulatory domain alignment and proposes a finance-aware retrieval/reasoning baseline. The paper says the manifest, execution environment, and evaluation code will be open-sourced; that is a benchmark route, not evidence of live trading authority or returns.

Hui Gong’s AI Agents in Financial Markets paper, revised April 22, proposes four layers—data perception, reasoning, strategy generation, and execution with control—and an Agentic Financial Market Model linking autonomy, heterogeneity, execution coupling, infrastructure concentration, and supervisory observability to market outcomes. Its event-study application is explicitly exploratory, and its bounded-autonomy framing should remain a research hypothesis rather than a claim about any manager’s operating model.

• Fin-Analyst at FinMMEval 2026 describes an eight- specialist LLM pipeline combining news, SEC filings, fundamentals, analyst forecasts, technical indicators, and social sentiment, with a meta-agent for an equity task and a rule-based comparison for crypto. The paper reports a short live-evaluation window, ablations, and repeated-error analysis. Its leaderboard and return figures are author-reported and require independent replication, cost controls, and a longer point-in-time study.

University and finance-programme signal routes

The public academic trail adds model-idea and talent-sourcing context, while remaining separate from evidence about any tracked hedge fund.

UPES School of Business’ profile for Aniruddha Ghosh lists financial data analytics, financial econometrics, risk analytics, machine learning in finance, and generative AI in finance. The profile names bankruptcy prediction, green finance, financial risk modelling, and behavioural finance as application areas, and lists courses in financial econometrics, risk analytics, portfolio management, and AI applications in finance. This is evidence of an academic applied-research route, not a disclosed fund implementation.

Pontificia Universidad Católica del Perú’s Machine Learning para Finanzas syllabus covers the model-development sequence that matters in finance: sampling and preprocessing, feature engineering, train/test/validation design, cross-validation, resampling and bootstrapping, supervised and unsupervised learning, tree ensembles, gradient boosting, SVMs, neural networks, and dimensionality reduction. The page does not provide student results, code, data rights, or production evidence.

CEMLA and Deutsche Bundesbank’s VII Course on Machine Learning and Central Banking, held by videoconference on August 3–4, 2026, emphasizes the connection between ML and conventional statistics, practical R exercises, adoption challenges, and strategies for developing and implementing models. CEMLA identifies IT, statistics, and research professionals as the audience and names Gerardo Hernández del Valle as coordinator. This is a useful model-governance and implementation surface; it does not reveal participant strategies or investment performance.

Wits University’s February 2026 “Shaping the Future of Finance” programme connects AI, quantum computing, blockchain, digital currencies, and data analytics with finance, economics, actuarial science, quantum physics, and machine intelligence. Its named participants include Solomon Assefa, whose listed education includes MIT degrees and whose biography records prior IBM Research leadership; Chimwemwe Chipeta, director of the Wits Fintech Hub; and Rendani Mbuvha, whose listed research spans machine learning, climate risk, weather forecasting, and actuarial science. These are university and cross-disciplinary research routes, not evidence of a hedge fund’s deployed models.

Manipal’s profile for Rakshith Bhandary identifies AI/ML in finance, risk management, digital banking, time-series and panel-data regression, neural networks, asset pricing, and fintech among his expertise and research areas. The profile describes a 2024 PhD on loan-default prediction using AI/ML and structural-equation modelling. It is a banking and credit-risk research lead; the page does not establish hedge-fund employment, proprietary data, production authority, or performance.

Taken together, these university routes expose research directions that can be followed into papers and reproducible artefacts: default and bankruptcy prediction, risk and econometric validation, ensemble and neural methods, central-bank adoption controls, fintech and climate-risk analytics, and cross-disciplinary quantum/data-science work. They should be used as discovery and lineage leads, not as a basis for ranking firms or attributing academic work to an investment manager.

Monash University’s BFF5555 Financial Machine Learning unit describes a Python-based progression from statistical learning and supervised methods to unsupervised learning, deep learning, model evaluation, regularisation, cross-validation, and NLP in finance. The handbook names Dr Hoa Briscoe-Tran as chief examiner. This is curriculum evidence and a talent-sourcing route, not a disclosed trading system or performance record.

Xi’an Jiaotong–Liverpool University’s FIN424 module follows financial and business data from raw inputs to features, signals, trading algorithms, model evaluation, economic interpretation, and risk-management interaction. The 2026/27 catalogue includes Python labs and says sessions may use the IBSS trading floor. It does not disclose student strategies, data rights, live capital, or results.

National Chengchi University’s Chinese-language Machine Learning and Financial Econometrics syllabus identifies a 2026 graduate finance elective taught by Vincent Kendro and describes ML applications to asset allocation, with calculus and linear-algebra prerequisites. This adds a Taiwan and non-English programme route; the syllabus does not establish a deployable strategy, dataset, or performance.

Johns Hopkins Carey’s Yinan Su provides a public professor-and-papers route connecting a University of Chicago Booth–Economics PhD and Tsinghua undergraduate training to empirical asset pricing, financial econometrics, ML, and AI. The page identifies research inputs including firm characteristics, financial news, investor holdings, and trading activity. Listed papers cover narrative factors from Wall Street Journal text, predictive cycles and volatility, trading-volume prediction, quantity-aware factor pricing, structural breaks, and an early AI-compute asset-pricing framework. These are research hypotheses and lineage evidence; the page does not establish hedge-fund deployment, proprietary data, or live performance.

Newly discovered external firm disclosure

The Finance Lab’s About page describes the organisation as an AI-native research and investment firm building generative models and autonomous agents for financial markets. Its Bloodhound Model 1 page describes a compound architecture in which specialist quantitative models provide market-regime, trend, volatility, historical-similarity, cross-asset, distribution, portfolio-rule, and momentum outputs to a financial-reasoning layer. The page says that layer uses a fine-tuned Gemma 4 26B A4B variant, GRPO, long-context memory, and feedback from realised market outcomes. A separate research page lists reinforcement learning from market feedback, latent market representations, a 9,000-agent architecture, and work on extrapolation beyond interpolation.

This is unusually detailed public architecture language, but it remains a first-party marketing disclosure. The pages do not provide independently verified corporate identity, named personnel, code, model weights, data rights, audited results, or proof of live trading. The claims therefore belong in the discovery ledger and monitoring queue, not in a capability ranking or performance comparison.

Title-blind engineering disclosure: Hudson River Trading

HRT’s October 2025 engineering post describes a firm-wide operating topology without presenting itself as an AI announcement. It separates Algo teams—which research, deploy, and monitor trading models—from Trading Tech, which handles exchange data, order placement, acceleration, processing, clearing, and reconciliation, and Research & Development, which handles storage, data centres, CPU/GPU clusters, scheduling, ETL, and research tools. The post also says platform engineers support multiple trading teams rather than being confined to a single silo, and identifies internal development tooling such as distributed testing and cached builds.

This is useful evidence about how a quantitative trading organisation publicly describes the path from historical research to live execution and the infrastructure around it. It does not identify particular AI models, datasets, performance, investment authority, or personnel beyond the post’s author and role context.

September 4, 2026 — professor-led idea supply and finance-programme implementation routes

The academic layer is useful not only for lineage. Current faculty pages, course descriptions, and thesis repositories expose specific research problems, evaluation choices, and implementation boundaries that can be followed into papers or student artifacts. They remain academic evidence; they do not establish that a tracked manager uses a method.

The Chicago Booth Center for Applied AI paper by Ralph Koijen and Bradford Levy describes a real-time, out-of-sample benchmark for earnings-announcement analysis. The stated design restricts the system to information available at announcement time, then evaluates whether agentic systems can extract structured signals from transcripts and explain contemporaneous returns. The paper’s proposed research object is therefore not simply “an LLM that predicts stocks”: it is an information-timing, signal-selection, economic-interpretability, and reflexivity problem. The authors report an exploratory comparison with standard benchmarks and release an SDK according to the page. Those results and the SDK still require independent inspection of code, data access, time splits, costs, and replication before being used as investment evidence.

Leland Bybee’s Chicago Booth profile adds the professor route behind the adjacent curriculum. Bybee is listed as an Assistant Professor of Finance and Fama Faculty Fellow, with research using machine learning and natural-language processing to study beliefs, asset pricing, and behavioral economics; his 2026–27 schedule includes Machine Learning in Finance. The course catalogue describes penalised forecasting, clustering, factor models, unsupervised learning, nonlinear prediction, financial text, and large language models using real financial datasets. This gives a clean way to split the queue into belief measurement, text representation, cross-sectional prediction, and factor construction rather than treating every finance-ML project as one category.

Chicago’s FINM 33500 Systematic Trading Technologies provides an engineering counterpart. The course description names asynchronous market-data and execution pipelines, event-driven backtesting, machine-learning signals, walk-forward validation, CI, Docker, tests, dashboards, and reproducible end-to-end artifacts, with Sebastien Donadio listed as instructor. This is a useful implementation checklist for the research queue: a promising model is not the same thing as a replayable data path, a cost-aware evaluator, or an operationally testable system. The page does not disclose student projects, proprietary data, or live capital.

MIT Sloan’s IAP 2026 “AI and Money” syllabus is a different kind of professor-led route. It names Gary Gensler as professor and frames the two-day course around traditional predictive models, generative AI, asset management, trading, underwriting, compliance, AI supply chains, data centres, technology stacks, performance drivers, economics, and regulation. The syllabus also requires a paper and places substantial weight on discussion and human reasoning. It is useful for expanding the research map beyond alpha generation into infrastructure, model-risk, cost, regulatory, and organisational questions. It is a course design, not a disclosure of MIT’s or a manager’s production system.

Georgia Tech’s Fall 2026 Practice of Quantitative and Computational Finance syllabus exposes a practical bridge between programmes and firms. It allows industry-partner projects under NDAs, separately permits public-data projects, and explicitly lists quantitative trading, ML prediction, NLP/LLMs in finance, Bayesian and ML credit-risk models, crypto, and applications across asset classes and institutional functions. The document also warns that WRDS and other purchased data carry usage restrictions and says projects may involve data annotation. This is valuable competitive-intelligence context because the public syllabus reveals where applied work can occur while also showing why project details, datasets, and results may not be public. It does not identify the participating firms or imply adoption of any student output.

Luiss adds an unusually inspectable student-to-research trail. The 2026/27 Data-Driven Models for Investment course names Antonio Simeone and Villy Edoardo de Luca and lists alternative data, LLMs, agents, time-series foundation models, fuzzy and neural methods, systematic trading, and multi-agent portfolio construction. Its LuissThesis supervisor index then exposes project titles around RAG for time-series prediction, alternative-data trading systems, fuzzy logic plus reinforcement learning, adaptive neuro-fuzzy forecasting, and RL for optimal control. These titles are discovery leads rather than validated strategies, but they show how model ideas move from a syllabus into concrete student artifacts and which papers, PDFs, and code should be recovered next.

The ETH Zurich Machine Learning for Finance and Complex Systems course adds another distinct research vocabulary: graph cuts, matrix factorisation, RL, MCMC, LSTMs, attention, neural ODEs, physics-informed neural networks, transformers, and Black–Litterman, taught through a coding project by Nino Antulov-Fantulin and Patrick Cheridito. The combination points to structured representation, dynamic systems, and allocation constraints as separate investigation lanes. The course page does not disclose student results, proprietary data, or production deployment.

These sources make the academic map actionable without turning it into a ranking: earnings-event agents and belief measurement; text and multimodal representation; time-series foundation models; alternative-data and geospatial collection; execution and walk-forward validation; credit, climate, and cyber risk; market interaction and reflexivity; and governance around data rights, cost, and human review. The next research step is to retrieve the linked papers, thesis PDFs, code, datasets, recordings, and named industry links, then keep academic hypotheses separate from firm-specific evidence.

September 4, 2026 — Northern European and Canadian finance-research routes

The academic search also surfaced programmes and labs that expose how research questions are turned into finance projects, without making claims about any manager’s internal systems.

NHH’s Autumn 2026 FIN545A Asset Pricing II is a restricted PhD course responsible to Adjunct Associate Professor Lars A. Løchstøer. It combines time-series and cross-sectional asset pricing across equities, currencies, commodities, bonds, and derivatives with recent machine-learning research. Students must produce an original empirical project specifying the research question, data, and results, then present it. The explicit deliverable makes this a useful discovery route for papers and replication artifacts; the course page does not reveal student projects or firm use.

Walter Pohl’s NHH profile adds a named faculty and lineage route. NHH identifies Pohl as a Professor of Finance whose fields include asset pricing and the application of machine learning to finance, with prior postdoctoral work at Zurich and a PhD in finance from the University of Texas at Dallas. His listed publications include commodity-price variation and heterogeneous-agent/long-run-risk asset pricing, and he teaches Big Data with Applications to Finance. This points to commodity, long-horizon-risk, and big-data asset-pricing research lanes; it does not establish a hedge-fund relationship or production model.

The University of Copenhagen’s 2026–27 Advanced Empirical Finance: Topics and Data Science course provides a particularly concrete data and validation route. Its syllabus covers historical company and price data, high-frequency transaction and order-book data, factor selection with Lasso and neural networks, realised-volatility estimation, market-microstructure noise, transaction costs, liquidity risk, parallel/cloud computing, and reproducible R reports. The course requires assignments with peer feedback and allows AI tools subject to disclosure. This separates factor discovery, intraday risk, data engineering, and reproducibility into distinct queues; it does not disclose a student strategy, proprietary data, or live capital.

RiskLab Toronto supplies a Canadian lab and industry-collaboration route. Its public history connects the University of Toronto lab to an earlier Algorithmics sponsorship and describes current pillars of machine learning, mathematical modelling, and investment and trading strategies. The page names Prof. Luis Seco and Prof. Hamid Arian in leadership, Alan Peng as managing director, and members across the University of Toronto, York University, the Fields Institute, and Technical University of Munich. This is evidence of an institutional research and talent network; it does not identify a particular manager’s dataset, model, permission boundary, or performance.

These additions extend the academic queue toward original empirical projects, commodity and long-horizon risk, high-frequency volatility and order-book data, factor selection, cloud-scale reproducibility, and industry-linked quantitative finance labs. They are idea and personnel sources, not a basis for ranking funds or attributing academic work to them.

September 4, 2026 — additional professor and finance-programme routes

Several further university pages add distinct idea channels and personnel bridges. They are useful for finding papers, theses, public code, and recruiting relationships; none by itself establishes a tracked manager’s use of a method.

• Kristian Bondo Hansen, Copenhagen Business School studies the adoption of machine learning and AI in asset management and quantitative trading using interviews and archival work. His profile also covers the political economy of cloud provision and big-data infrastructure in finance. This is a route into organisational adoption, compute procurement, and workflow-governance questions rather than a quantitative-signal claim; the page does not identify a fund, proprietary data, or production authority.
• Maxime Bonelli, London Business School links investment and asset-management research to the way Big Data and AI reshape financial-sector organisation and capital allocation. LBS records a mathematics PhD from Inria, a finance PhD from HEC Paris, and more than four years as a quantitative researcher in asset management before academia. This is an academic-to-industry lineage lead for studying labour substitution, research-team design, and investor decision-making; it does not disclose the prior employer, a current manager relationship, or a deployable model.
• Nanyang Business School’s finance faculty page exposes several Asia-Pacific routes in one official directory. It lists Will Cong with joint finance and AI/Data Science appointments and expertise in AI for economics and finance; Nelson Lau with systematic and high-frequency trading, machine learning in finance, market microstructure, and alternative data; and practice faculty connected to trading and technology roles. The same page lists Paras Kharbanda as Head of Crypto Engineering at Tower Research Capital and instructor for BA2215, “Vibe Coding for Algorithmic Trading.” These are public academic and teaching links, not evidence of Tower’s internal models or of student work being used in production.
• Fuwei Jiang’s Xiamen University research page provides a non-English-to-English paper trail. The official page lists work on investor and manager sentiment, narrative-based inflation forecasts, financial-news sentiment, machine-learning stock-crash risk, model uncertainty, and dynamic CAPM; its Chinese-language list includes large-language-model text sentiment and financial markets, GAN-based deep learning for Chinese factor investing, media sentiment and stock-return prediction, and machine-learning bond-default warnings. The route points to sentiment, narrative, uncertainty, factor construction, and credit-risk replication tasks, especially for Chinese equities; it does not establish a fund mandate, proprietary data, or live performance.
• Hongik University’s Fintech Lab is a cross-school lab spanning finance, computer engineering, and industrial engineering/data science. Its public research page names ML/deep learning for finance, NLP, massive financial data, explainable and fair AI, AutoML, and quantitative algorithm development. The member page records Hyuncheol Hwang’s prior Wall Street and Korean financial-industry experience and research interest in AI quant/HFT, while the project page lists regulator AI-guideline work, an Ernst Young asset-valuation project, a Koscom factor-model stock-rating and portfolio-recommendation project, anomaly detection, and reinforcement-learning dynamic pricing. These are concrete research and industry-project leads, but the public pages do not identify current sponsors, unrestricted datasets, model weights, or trading results.
• Gianluca De Nard, University of Liechtenstein combines a practice-professor role, supervision of master’s theses, machine-learning-for-asset-pricing research, and Head of Quantitative Research at OLZ AG. The profile records a University of Zurich finance PhD, ETH Zurich/University of Zurich quantitative-finance master’s training, a visiting PhD period at NYU with Robert Engle, a Yale postdoctoral period with Bryan Kelly, and recent publications on data-driven shrinkage, portfolio construction, climate-risk factor-mimicking portfolios, and tracking-error management. This is a lineage and portfolio-risk research route; it is not evidence that any tracked hedge fund uses those methods.

The resulting idea queue is broader than LLM-based text extraction: organisational adoption and compute economics; labour and research-team design; high-frequency and alternative data; crypto-engineering education; Chinese sentiment and narrative signals; crash and default risk; explainability, fairness, and AutoML; reinforcement-learning control; and shrinkage, covariance, climate-risk, and portfolio-construction methods. These routes should be followed into papers, code, thesis PDFs, and event recordings while keeping academic hypotheses separate from firm-specific disclosures.

September 4, 2026 — additional U.S. professor and programme routes

Another pass produced four distinct academic surfaces with more specific model ideas, data routes, and training signals. They are research and talent leads, not evidence of any tracked manager’s production system.

• Tengfei Zhang, Rutgers–Camden is listed as an Assistant Professor of Finance whose research interests include generative AI in finance, machine learning, and textual analysis. Rutgers records a prior postdoctoral role at Cambridge Judge’s Centre for Endowment Asset Management, a Louisiana State finance PhD, and a course titled “Generative AI and Textual Analysis in Finance.” The profile lists a 2026 Review of Financial Studies paper, “Dissecting Corporate Culture Using Generative AI,” suggesting a research route from text measurement to corporate culture and analyst/investor interpretation. The page does not disclose fund use, proprietary corpora, or live performance.
• Terrence O’Brien, University of Maryland Smith is Faculty Director of the Private Credit Program and teaches quantitative finance across undergraduate, graduate, and doctoral programmes. The profile connects empirical asset pricing to fintech, cryptocurrency, and machine learning, including research on exchange volume manipulation and false market-quality signals. It also records prior executive work at a company buying and selling Chinese macroeconomic and industrial government data. This opens private-credit, market-integrity, and China alternative-data research lanes; it does not establish a hedge-fund relationship or a deployable model.
• Georgia Tech’s MGT 8803/4803 AI in Finance syllabus, taught by Sudheer Chava, explicitly separates ML/DL foundations from an applied NLP, LLM, domain-specific LLM, small-language-model, and GenAI finance course. The syllabus describes lectures plus implementation labs, finance datasets via WRDS or the course system, possible annotation work, Python assignments, guest lectures, and a multi-deliverable group project. This is a concrete talent and workflow route for testing finance-specific language models, but it does not reveal student outputs, sponsor identities, proprietary data rights, or production deployment.
• Shengyu Huang’s Stevens dissertation announcement and Winthrop faculty profile expose a 2026 model-complexity route. The dissertation studies when nonlinear ML is essential in asset pricing, uses XAI to measure firm complexity, and reports state-dependent relations with skewness, growth characteristics, uncertainty, trading activity, distress, funding costs, and bank-stock returns. The official announcement identifies the dissertation committee and the theoretical, empirical, and XAI essays; the Winthrop page identifies Huang as an Assistant Professor of Finance. These are hypotheses and methods to replicate, not evidence of a manager’s factor or performance.

This pass sharpens the queue around corporate-culture text measurement, analyst and investor interpretation, private-credit and market-integrity signals, Chinese macro-data provenance, domain-specific and small language models, annotation and reproducible project design, and state-dependent nonlinear complexity. The next step is to recover the underlying papers, course artifacts, data terms, and public talks, while keeping academic evidence separate from firm-specific disclosures.

September 4, 2026 — title-blind conference and allocator-media routes

The media sweep also found conference and allocator pages where the relevant signal is in the programme or speaker biography rather than in a hedge-fund or AI keyword.

• SIFMA’s “Heard at Ops 2026” series packages short interviews filmed at the 2026 Operations Conference and links each conversation to a fuller session replay. The page frames AI, data, digital assets, tokenisation, resilience, and always-on markets as operating-model topics, and names speakers including Fredric Crosnier of J.P. Morgan Asset Management, Bob Santella of BetaNXT, and leaders from RBC, FINRA, Jefferies, DTCC, and Altruist. This is a useful path into market-operations, infrastructure, and implementation language that may not appear in investment-team media; the page does not disclose a hedge fund’s model, data permissions, or performance.
• QUANT 2026 in Varese is an Italian-language asset-management and quantitative-investing workshop held March 5, 2026. Its programme names Cesare Orsini of Allianz Global Investors for “When Machine Meets Human: Integrating AI Into Systematic Equity Investing,” Gabriele Susinno for an econophysics session, and academic/industry speakers including Paul Wilmott and Andrea Ferrero. The event explicitly invites asset managers, consultants, and family offices, making it a regional allocator and personnel-discovery route; it does not establish what any speaker’s employer deploys.
• CFA UK’s Technology & Innovation Community Summit is scheduled for September 21, 2026 in London and offers a route to the keynote recording. Its published speaker list includes Jeremy Leung, Investment AI Solutions Manager at T. Rowe Price; Rob Otter, Head of Global Technology Applied Research at J.P. Morgan; Shane Jocelyn at Citi; Justin Xu, Chief Quant & AI Officer at Milltech; Kara Byun, Head of Fintech at HSBC Asset Management; and Pablo Riveroll, Global Head of Equities Research at Schroders. The keynote description focuses on workflow embedding, human overrides, judgement, and the shift from insight to decision. The event page is a speaker and recording lead, not evidence of any employer’s production system or results.

These routes expand the media queue toward operations and resilience, AI-enabled systematic equity investing, econophysics, allocator/family-office networks, research-platform design, human override boundaries, and the organisational step from generated insight to investment decision. The next recovery step is to capture the linked replays, session pages, and recordings, then inspect transcripts for dated, attributable statements.

September 4, 2026 — regional allocator, regulator, and quant-media routes

The regional search adds several high-value capture paths that sit outside a conventional hedge-fund keyword search.

• CFA Society Netherlands’ “AI, Machine Learning and Agentic Investing” event lists Kristina Usaite, Senior Quantitative Researcher at Robeco, on AI-plus-human investing; Michael van Baren, Senior Quantitative Researcher at Northern Trust Asset Management, on ML applications in quant equity; and Reza Kahali, CIO of Argan Technologies, on agentic AI in macro investing. The page also records Kahali’s prior work at a European family office, NN Investment Partners, Ortec-Finance, and GSK. This is a named personnel and lineage route with explicit research topics, not evidence of any employer’s internal model.
• ESMA’s webinar on AI adoption in EU securities markets provides a regulatory and recording route often missed by investment-media searches. ESMA dates the webinar March 4, 2026, links presentation materials and a YouTube recording, and says the underlying analysis uses recent survey data to assess AI adoption, use cases, benefits, and challenges in EU securities markets. It can anchor claims about market-wide adoption context; it does not identify a tracked fund’s system or performance.
• UBS’s 2026 Asian Investment Conference release exposes three distinct discovery lanes in one event: an AI-focused wealth forum, an APAC Family Office Summit for principals and decision-makers, and a Trading and Execution Track covering quantitative strategies, data, and execution. UBS says the May 25–29 Singapore/Hong Kong programme connected asset-management, investment-bank, wealth-management, institutional, and family-office audiences. This is a network and session-recovery route, not evidence of UBS or a participant’s proprietary workflow.
• India’s Investing Accelerator Summit 2026 advertises full recordings from its August 7–8 event, with access through an Online Pass. The public page describes fund managers, independent investors, SEBI-registered advisers, listed-company founders, and an AI research companion called “fuzz” that answers from official filings and exchange data with linked sources. This is a valuable India-specific archive and title-blind discovery route; the public page does not expose the complete recording library, model internals, or investment results.
• Scientific Beta Days Europe 2026 publishes a detailed programme around real-world signals, ML-improved risk forecasts, commodity factor premia, and black-box AI challenges. Named participants include Scientific Beta researchers, Joëlle Miffre of Audencia, Maxime Ricomes of Mercer, and Robinson Rouchie, CIO for Systematic and Quantitative Investments at BNP Paribas Asset Management. The programme explicitly cautions that gains from complex return signals can be more modest than headline results and stresses implementation discipline; it is a conference-source route, not a verified performance comparison.
• HSBC Asset Management Singapore’s videos and webinars hub provides a firm-owned title-blind media archive. Its 2026 entries include “Quant equity strategies 101,” an Asia-investing video, India fixed-income coverage, and current named investment specialists and portfolio managers; the page also preserves older podcast and video categories. This is useful for extracting the firm’s public vocabulary around quant process, regional data, and portfolio roles, but the hub does not expose model code, training data, or performance attribution.

These sources expand the queue into regulator survey evidence, YouTube-linked event capture, family-office and allocator networks, India filing-linked research tooling, risk-forecast and black-box-AI governance, commodity factors, and firm-owned quant-media archives. The next step is to resolve each recording or session asset, preserve captions/transcripts, and separate programme claims from speaker-level employer disclosures.

September 4, 2026 — causal machine learning and manager-skill research routes

Two additional academic routes are worth preserving because they produce testable research objects rather than broad AI labels. MIT’s David Bruns-Smith faculty profile places him in a shared Finance and EECS appointment and describes machine-learning methods for causal inference applied to macroeconomics and household finance. His public research list includes deconfounding and representation learning under weak overlap, robust off-policy evaluation with sequentially unobserved confounding, two-stage ML for nonparametric instrumental variables, and synthetic-panel generation; it also records a Stanford Data Science postdoctoral route and a UC Berkeley computer-science PhD. For the investment-research map, this is a route into treatment-effect estimation, counterfactual policy evaluation, and robustness to hidden confounders—not a claim about a tradable alpha signal or a manager’s deployment.

Columbia’s February 2026 working paper, “Detecting Skilled Bond Fund Managers” provides a separate manager-selection idea. Ron Kaniel, Markus Pelger, Stijn Van Nieuwerburgh, and Luofeng Zhou state that they use machine learning on 3,021 unique U.S. bond funds from May 1995 through November 2024, combining fund and family characteristics, holdings, and macroeconomic variables. The abstract reports that a prediction-weighted long-best-10%/short-worst-10% portfolio generated 30 basis points of monthly abnormal return with a reported 24.6% information ratio, with persistence reported for up to 36 months; it also says holdings-based characteristics did not separate managers, reports signal concentration in corporate and municipal funds, and describes Treasury-fund differentiation as weak. These are authors’ reported working-paper results, not an independent validation or a recommendation. The replication queue should preserve the exact vintage, fund-selection rule, shorting and fee assumptions, survivorship treatment, and capacity before treating the result as usable evidence.

Together, these routes broaden the professor/programme map beyond language models: one tests whether predictive relationships survive causal and confounding scrutiny; the other tests whether observable manager histories contain information about future bond-fund outcomes. Neither route identifies a tracked hedge fund’s internal model, data rights, permissions, or live performance.

September 4, 2026 — computational-finance programmes and quant-engineering personnel routes

Carnegie Mellon’s MSCF curriculum exposes an education-to-implementation chain. Its machine-learning sequence covers regression, classification, boosting, ensembles, clustering, topic modelling, NLP, reinforcement learning, and deep learning; adjacent courses cover financial time series, factor portfolios, optimisation, market microstructure, and algorithmic trading. The market-microstructure course specifies Kdb+ and Python work on actual intraday quotes, trades, and order books, while the ML capstone pairs students with a financial firm and company mentors. This is a talent and research-surface signal, not proof of any named sponsor’s production system or student strategy.

The University of Washington Foster School’s MSF course catalogue adds a broader finance-ML route: Data Analytics in Finance, Fintech and Applications, Machine Learning in Business with predictive modelling, algorithmic-trading and risk-management exercises, and two-quarter capstones often based on local finance or corporate needs. The public catalogue exposes a potential recruiting and project channel; it does not identify a fund partner, disclose student outputs, or establish live-capital access.

Imperial’s Computing in C++ course blog, maintained by Paul Bilokon for MATH70112, joins C++, derivatives, electronic and high-frequency trading, software engineering, machine learning, and AI in one public teaching surface. An August 2026 post reports—through eFinancialCareers—that C++ creator Bjarne Stroustrup joined Susquehanna International Group as a part-time Technical Fellow and would contribute to codebase optimisation; it also says SIG became a CppCon 2026 Platinum sponsor. This is a dated academic-blog and secondary-reporting personnel lead, not a first-party SIG announcement and not evidence of an AI or GenAI deployment.

Carnegie Mellon’s Center for Computational Finance profile for Bryan Routledge adds a named professor route spanning asset pricing, NLP, commodities, macroeconomics, and corporate finance. It is older institutional profile material, so it is retained as a research-lineage lead rather than treated as a current project or firm relationship.

These sources separate four queues that are often conflated: finance-specific ML training, market-data and execution engineering, professor-led research ideas, and personnel movement reported outside an employer’s own channels. The next checks are the named capstone sponsors, course project artefacts, CppCon programme, SIG first-party confirmation, and public papers or code from the faculty routes.

September 4, 2026 — European finance-ML programmes and open research artefacts

The UZH–Peking University PhD summer school on Machine Learning for Macroeconomics and Finance is a particularly concrete academic-to-code route. The organisers say the July 6–10, 2026 programme received more than 300 applications and selected about 120 faculty and PhD participants. It covered deep learning, heterogeneous-agent models, structural reinforcement learning, continuous-time macro-finance, asset pricing, and numerical methods; the page names Yucheng Yang, Felix Kübler, and Bo Li as organisers and says slides, tutorials, and code were made openly available. It also names DeepHAM and JAX as implementation tools. This is public research-network and training evidence, not a fund system or a performance claim. Yang’s UZH faculty profile records his finance, macroeconomics, and ML focus and Princeton, Peking University, and Wisconsin lineage.

TUM’s Digital Finance teaching page announces a 2026/27 programme with Asset Management, Financial Markets, and “AI in Investments,” plus doctoral seminars and thesis supervision under Professor Florian Weigert. It creates a current German recruiting and project-surface lead; the page does not disclose student projects, partner firms, models, or investment results. TUM’s Christoph Knochenhauer profile adds a professor route spanning stochastic control for institutional and private investment, ML in finance, and probabilistic PDE methods, with TU Kaiserslautern/Dublin City University, University of Trier, and TU Berlin lineage. The profile is a research and personnel signal, not evidence of a manager relationship or production deployment.

The University of Bologna’s 2026/27 Advanced Machine Learning course in its Quantitative Finance master’s names Giovanni Della Lunga and focuses on neural networks for pricing and market-risk problems. Its description acknowledges that the required large-scale compute is beyond the student provision and therefore uses smaller models to teach the methodology. This exposes a useful compute-and-validation boundary: a research method may be teachable on toy problems without demonstrating scalability to production. It is programme evidence, not proof of student or manager deployment.

These sources add four specific recovery tasks: obtain the UZH summer-school code and tutorials; inventory DeepHAM’s economic assumptions and numerical evaluation; resolve TUM thesis and seminar topics; and inspect Bologna’s reading list and toy-model assignments.

September 4, 2026 — Boston University finance-ML pipeline and personnel lineage

Boston University adds a more specific Boston talent and curriculum route. The official MS in Financial Management Artificial Intelligence Applications concentration names financial technology, algorithmic trading, quantitative analysis, and risk management as target contexts, and lists natural-language processing, predictive modelling, deep learning, asset allocation, automated strategies, risk assessment, and fraud detection. BU’s MS in Mathematical Finance & Financial Technology adds a 39-unit quantitative route with Python, R, C++, MATLAB, and SQL; high-frequency and large-data analysis; Bayesian econometrics, volatility and factor models; GPU and parallel computation; tree methods, deep learning, text analysis, algorithmic and high-frequency trading, credit risk, and portfolio construction. These are official programme descriptions and recruiting surfaces. They do not identify a fund sponsor, student output, production model, or investment result.

The BU Hariri Institute profile for Alexander Becker identifies him as an assistant professor of finance at BU MET, with a Boston University physics PhD concentrated in econophysics. The profile connects complex networks, financial stability, corporate finance, and machine learning in finance, and says his current work examines private debt, covenants, and bank-firm lending relationships. The route is relevant to network representation and financial-stability research; it does not disclose a manager relationship or a live model.

The BU profile for Eugene Pinsky provides a distinct practitioner-to-classroom lineage. BU records a Harvard mathematics degree, a Columbia computer-science PhD, a 1993–94 visiting-scientist period at MIT’s Laboratory for Information and Decision Sciences, and self-reported computational market-risk work at Bright Trading, Letra Group, F-Squared Investments, Harvard Management Company, and Trading CrossConnects. It also lists quantitative-finance applications of data mining and ML, portfolio construction, forecasting, and trading algorithms, alongside recent public work on deep-learning stock-movement prediction, algorithmic-trading comparisons, and crude-oil ETF/futures strategies. These employment and publication details are profile-level evidence; they do not establish that any named employer adopted the methods or that the reported research results survive independent, point-in-time evaluation.

This Boston route adds a concrete search path through course syllabi, capstone projects, faculty papers, and alumni destinations, while keeping academic training, self-reported prior employment, and investment-manager disclosure as separate evidence classes. See the academic finance source note for the capture details and recovery queue.

September 4, 2026 — HEC Paris model-validation and AI-research controls

HEC Paris’s Christophe Pérignon profile exposes a useful control layer around financial AI. HEC identifies Pérignon as Professor of Finance, Deloitte Chair in AI for Business Innovation, Associate Dean for Research, and Scientific Co-Director of Hi!PARIS; it also records Swiss Finance Institute doctoral training, a UCLA postdoctoral fellowship, and earlier faculty work at Simon Fraser. The profile says he co-founded CASCaD, a reproducibility-certification agency, and founded RAMP-UP, a science-job matching platform. These are institutional and personnel signals, not evidence of a hedge-fund relationship.

The linked paper list adds several testable safeguards. The July 2026 working paper “Randomness in Large Language Models: What Researchers Need to Know (and Report)” describes variation in repeated filing-sentiment classifications from deliberate sampling, silent model updates, numerical rounding, and expert routing. Its public abstract says temperature zero does not remove every source of variation, while local open-weight execution offers more control but still depends on the complete hardware and software stack. The paper’s proposed reporting discipline treats LLM outputs as draws from a distribution and records model and execution conditions. This is directly relevant to any fund using an LLM as a measurement layer, but it is not a fund disclosure.

Pérignon’s “The Challenger” working paper frames replacement of an incumbent model with one using new or alternative data as an economic decision. It links learning-curve behaviour, data-acquisition and retraining costs, and discounting, then describes one-shot, greedy-sequential, and look-ahead procedures evaluated on a credit-scoring dataset with gradually arriving alternative data. The “Safe-Tail Paradox” paper adds a stress-testing route: its abstract reports that AI-exposure risk can be concentrated in borrowers classified as safest and demonstrates the proposed test on a synthetic French mortgage portfolio. Both are research designs and author-reported results, not evidence of a manager’s data, model, capital rules, or performance.

The “Persistent Anomalies and Nonstandard Sharpe Ratios” paper extends the same discipline to asset-pricing research by modelling variability across analysis paths rather than reporting only a conventional standard error. HEC’s abstract describes a 107-anomaly study and a nonstandard-error decomposition; those results still require versioned data, exact selection rules, and independent replication. Together, these papers create a practical validation queue for finance AI: freeze prompts and model versions, log repeated draws, price data acquisition and retraining, test challenger switching prospectively, and audit multi-design inference. They do not establish that any tracked firm follows these controls.

September 4, 2026 — family-office principal and Stanford AI network route

The family-office search also recovered a named allocator with an academic AI-network connection. Stanford HAI’s current Advisory Council roster names Liqian Ma among its members. A St. Paul’s School trustee profile, which says she joined the school’s board in 2026, describes nearly two decades of investment experience at Temasek International across technology, financial services, consumer, and healthcare. It says she now runs her own family office focused on asset management and philanthropic initiatives, and records a Stanford master’s in management science and engineering and a Cornell electrical-engineering degree.

This is a useful family-office-to-university network route for finding allocator conversations, research sponsorship, and AI-governance participation that would not necessarily appear in hedge-fund media. The reviewed pages do not disclose the family office’s portfolio, internal AI tools, data sources, model permissions, or investment results. The HAI relationship establishes an institutional network connection, not a family-office trading or GenAI deployment claim.

September 4, 2026 — allocator research on AI value capture and deployment economics

Cambridge Associates’ “Artificial Intelligence Investing After the First Wave”, published June 26, 2026 by Celia Dallas, Katharine Campbell, and Theresa Sorrentino Hajer, adds a first-party allocator lens. The report says the investment question is moving beyond access to frontier models and toward the bottlenecks and control points around deployment: power and infrastructure, workflow ownership, customer relationships, domain expertise, permissions, governance, and the ability to turn model output into completed business action. It also says improving open and lower-cost models can reduce the defensibility of raw model access where customers do not need tightly integrated, mission-critical systems.

For the hedge-fund map, this is useful context rather than a manager disclosure. It gives the research team a neutral checklist for reading public AI claims: identify the workflow, the data and permission boundary, the model and inference costs, the control or approval layer, and the path from generated output to an auditable decision. The report’s portfolio views are the authors’ allocator opinions and are not used here to rank firms or predict returns. It does not identify a named hedge fund’s internal model, training data, or live deployment.

September 4, 2026 — Rotman workshop adds professor-led finance-AI artifacts

The University of Toronto’s First Annual Asset Pricing and Investments Workshop was scheduled for May 9, 2026 and published a full single-track programme. Its AI in Finance session paired “Mimicking Finance,” by Lauren Cohen, Yiwen Lu, and Quoc Nguyen, with “A Financial Brain Scan of the LLM,” by Hui Chen, Antoine Didisheim, Mohammad Pourmohammadi, Luciana Somoza, and Hanqing Tian. Leland Bybee and Alejandro Lopez-Lira were listed as discussants. The wider programme added a global-markets paper by Yu An and Amy Huber and a subjective-belief-factor paper by Tingyue Cui, Ricardo Delao, and Sean Myers. This is a dated academic research and personnel route. It does not establish that a tracked manager uses one of these papers, has access to its data, or obtained a trading result from it.

The Rotman FinHub page lists “Agents Are Not Algorithms: The Tradeoffs of Decision-Time Reasoning in AI Trading,” by Ing-Haw Cheng and Justin Shi of Toronto with Maurice Granger and Vasily Strela of Royal Bank of Canada. Its public abstract describes a real-time market simulator for tender selection and execution. The same page lists RSM 338 Machine Learning, RSM 2328 Machine Learning and Financial Innovation, and RSM 8341 Analytical Methods in Finance, as well as Python, data-science, portfolio, risk, and trading case studies. FinHub’s public alumni list includes OMERS, Columbia, Berkeley’s MFE, JPMorgan Quantitative Strategies, Q Wealth Partners, CIBC World Markets, and RBC Capital Markets. These records show a finance-AI research and talent channel; they do not show an employer’s production model, permissions, proprietary dataset, or AI-attributed performance.

The practical follow-up is paper-level recovery: obtain the workshop PDFs, inspect the “Financial Brain Scan” methods and data, and document the simulator assumptions, transaction-cost treatment, and decision-time budget in “Agents Are Not Algorithms.” Academic papers, course catalogues, and alumni destinations remain separate evidence classes from manager disclosures.

September 4, 2026 — finance programmes expose execution, RL, and model-validation ideas

The college-programme sweep adds several concrete teaching surfaces. HEC Montréal’s MATH 60610A course is taught in the same faculty route as David Ardia’s public profile, which lists quantitative asset management, quantitative risk, and sentometrics. The course covers structured and unstructured financial data, textual analysis, filters for unobservable variables, and high-frequency data. Ardia’s profile also exposes supervised student routes around earnings-reaction analysis, text-based sentiment, volatility, option-return prediction, and hedge-fund model misspecification. These are teaching and research signals; no manager’s data, production model, or performance is shown.

NTNU’s new BBAN4010 course, coordinated by Denis M. Becker, starts in Spring 2027 and requires a group project applying reinforcement learning to a self-selected finance or economics problem. The published outline names MDPs, Q-learning, policy gradients, actor-critic methods, PPO, multi-agent RL, portfolio management, trading, risk, dynamic pricing, interpretability, and Python libraries including Gym, Stable Baselines 3, TensorFlow, and PyTorch. Becker’s profile separately records RL-for-finance research and public lectures on portfolio optimisation and limit-order-book crypto trading. The course is a future educational route; projects and outcomes are not yet public.

EPFL’s FIN-407, taught by Semyon Malamud, is listed as a mandatory 2026–27 Financial Engineering course. Its public coursebook moves from ridge regression and penalisation through neural nets, embeddings, sentiment, LLMs, transformers, and deep/shallow ML portfolios to transformer portfolio construction. It requires substantial Python coding, notebooks, projects, and model-performance evaluation. This is a detailed curriculum and talent-pipeline record, not evidence that any fund uses the course’s methods.

ENSAE Paris’s MScT AI for Markets and Quantitative Investment, directed by Charles-Albert Lehalle and Vianney Perchet, links ML to concrete financial market and systematic-investment problems. The programme advertises projects, internships, hackathons, and alternative inputs such as satellite imagery and maritime traffic for investment and risk-management applications; it also describes possible paths into financial AI/data-science labs or a finance-linked doctorate. Those statements describe programme design and career intent. They do not identify a fund partner, project result, data licence, or live deployment.

Across these programmes, the actionable idea inventory is wider than news sentiment: unobservable-state extraction, high-frequency modelling, RL under sequential decisions, multi-agent interaction, transformer portfolio construction, and non-price alternative data. The appropriate next step is to recover syllabi, assignments, project artefacts, and point-in-time evaluation rules, keeping student work and faculty research separate from manager disclosure.

September 4, 2026 — Stanford and Chicago professor routes for public-information alpha

Stanford GSB’s “The Shadow Price of Public Information” names Ed deHaan, Chanseok Lee, Miao Liu, and Suzie Noh and describes an “AI analyst” that uses only public information to make selective quarterly changes to the portfolios of about 3,300 actively managed U.S. equity mutual funds from 1990–2020. The authors’ working-paper abstract reports $17.1 million of incremental quarterly gains, compared with $3.6 million of average fees and $2.8 million of average alpha, and reports lower risk and outperformance of 93% of managers over the sample. The Stanford research account says the model was trained on public market and accounting variables, including sentiment analyses of earnings calls and regulatory filings, and was constrained to make small portfolio adjustments while preserving characteristics such as risk and number of holdings. These are author-reported research results, not an independent investment audit or evidence of a hedge fund’s system. They require checks of the paper’s point-in-time data, feature construction, model revisions, turnover, costs, capacity, portfolio constraints, and multiple-testing treatment before any replication or investment interpretation.

The experiment is useful because it defines a testable automation boundary: the model need not replace a manager or choose an unconstrained portfolio; it can be evaluated as a constrained information-processing layer that proposes marginal changes. The public record does not show the model weights, a durable data licence, a live deployment, an agent runtime, or whether the promised model/code release is available. Miao Liu’s faculty materials also place the paper alongside work on AI disclosure and narrative risk, but a co-authorship or conference appearance does not establish a manager relationship.

Chicago Booth’s AI-and-finance research overview provides a complementary professor/programme map. It names Stefan Nagel, Serhiy Kozak, Shrihari Santosh, Dacheng Xiu, Bryan Kelly, Shihao Gu, Raghuram Rajan, Luigi Zingales, Pietro Ramella, Ralph Koijen, Xavier Gabaix, Motohiro Yogo, and related collaborators. The page describes augmented elastic-net selection across many characteristics, nonlinear asset-pricing models using neural nets and regression trees, BERT analysis of more than 8,000 shareholder letters, asset embeddings, and text models that forecast or backcast economic variables from newspapers. Dacheng Xiu’s research archive adds public material on autoencoder asset-pricing models, image-based price trends, weak-factor and multiple-testing controls, high-frequency covariance, and recent work on expected returns and LLMs; it also links supplemental material, code, data, and a finance-analytics course. These sources give the research programme a concrete implementation vocabulary—representation learning, text/image inputs, high-dimensional selection, chronological prediction, and statistical limits—without establishing that any tracked manager uses a particular method.

The combined professor/programme route suggests several replication questions: whether an AI layer adds value after preserving a manager’s style and constraints; whether earnings-call and filing features remain available at the decision timestamp; whether embeddings add information beyond characteristics; and how much of a reported result is selection, turnover, or capacity. Those are research questions and validation gates, not comparative claims about schools, model families, or firms.

September 4, 2026 — Harvard finance-AI routes: local information, financial statements, and coverage economics

Harvard’s finance research pages add three distinct idea routes beyond generic sentiment. Charles C.Y. Wang’s HBS Working Knowledge feature describes a study with Sean Cao and Yi Xiang that prompted ChatGPT 4.1 and DeepSeek R1 to analyse 4,978 Shanghai and Shenzhen companies using common accounting inputs and six-month price-direction forecasts. The report says ChatGPT was more optimistic and less accurate in that setting, and that adding Chinese-language news removed the relative optimism. The HBS page also says the experiment used a cutoff around the models’ training data and compared predictions with year-end prices. This is a model-comparison and data-coverage result—not a trading strategy, current-model audit, or evidence of a manager’s deployment. The relevant replication questions are local language coverage, model cutoff, prompt stability, forecast horizon, and whether the added news is timestamped and legally usable.

The HBS AI Institute’s core-earnings summary names Matthew Shaffer of USC Marshall and Charles C.Y. Wang of HBS. Their public description compares a minimally guided prompt with a sequential prompt that first identifies unusual losses, then unusual gains, and then aggregates persistent earnings from 10-K filings. The page reports that the sequential measure performed better than traditional measures on the authors’ longer-horizon tests and that an average API call cost less than one dollar and about one minute per firm in the reported setup. These are HBS’s summary of a working paper, not an independent audit. The model version, filing vintage, prompt transcript, extraction error rate, corporate-action handling, and out-of-sample investment translation still need paper-level recovery.

The HBS AI Institute’s Seeking Alpha study names Mark T. Bradshaw, Chenyang Ma, Benjamin Yost, and Yuan Zou and links their working paper on generative-AI use by capital-market information intermediaries. The reported natural experiment follows AI-assisted articles before and after Seeking Alpha’s 2023 no-AI policy. The page reports a rise in AI-assisted writing, more output from adopters, weaker immediate market and reader reactions to AI articles on average, and broader coverage of smaller firms. It also reports that experienced writers’ AI-assisted work differed less from human work than newcomers’ work. The underlying paper is explicitly a working paper, and the public summary relies on AI-detection classifiers and platform-specific rules; it does not show that AI-generated analysis improves returns or that a hedge fund uses the method.

For enterprise quant research, these Harvard routes suggest a concrete division of labour: local-language retrieval and missing-information checks for global coverage; structured, multi-step extraction for filings; and low-cost coverage expansion with human review where domain knowledge affects information quality. They are research hypotheses and workflow designs, not a ranking of models, schools, or firms.

Harvard Data Science’s Ron Papka seminar page adds a practitioner bridge: Papka is identified as Voya Investment Management’s SVP, Head of Data Engineering and Governance, with prior capital-markets technology roles and a stated research interest in financial-news modelling and NLP prediction. The seminar is titled “Monetizing Data Science on Wall Street” and covers ML, NLP, analytics, change-management skills, and the business value of data. This is a dated seminar and biography route; the public page does not disclose Voya model weights, training data, permissions, production endpoints, or AI-attributed performance.

September 4, 2026 — academic tests of agent accountability and failure modes

Columbia’s Center for Digital Finance and Technologies Year 4 grants add two unusually specific finance-agent research designs. Zhuo Zhang’s proposed TrailFI system would connect an agent’s decision context to tool, wallet, system, and blockchain evidence so that a later incident can be reconstructed causally. Xiaodong Wang’s proposed Fin-GPT prototype would record prompts, retrieved evidence, tool calls, computations, and candidate outputs, then use smart contracts and selected zero-knowledge proofs to enforce evidence, computation, and authorization checks. The page describes proposed grant deliverables, not completed systems, hedge-fund use, production permissions, or investment results.

Stevens’ multi-agent spoofing project names Sidharth Koduru and advisor Steve Yang. The student simulator combines fundamentalist, chartist, zero-intelligence, and spoofing agents in a continuous double-auction order book. It compares a Q-learning detector using price change, volume imbalance, volatility, and order-cancellation rate with a GPT-4 detector. The project reports approximately 99% accuracy but F1 of zero for both detectors, showing why class imbalance and temporal recall can invalidate an apparently strong accuracy number. This is a simulation and student-project result; no public code, live surveillance deployment, or market result is established.

Dartmouth’s 2026 thesis by Rohan Ray, advised by Nikhil Singh, reports 150 backtests across Kimi K2 and Qwen3-235B-Instruct, 12 augmentation strategies, five seeds, and 82 trading days on the top 20 Dow constituents. The experiment injected Kronos time-series forecasts and FinGPT sentiment into StockBench agents. The thesis reports exploratory differences associated with signal format, model identity, gating, and narrative context, while stating that no full-period comparison survives multiple-comparison correction. It is an undergraduate research artifact, not a deployable strategy or a manager disclosure.

Together, these academic routes add three validation requirements to the research map: preserve causal traces around agent actions; evaluate rare-event detection with recall, F1, and temporal context; and test the receiver’s processing of an injected signal, not just the standalone accuracy of that signal. These are research and governance questions, not comparative claims about firms or models.

September 4, 2026 — university finance programmes expose the idea-to-system supply chain

The Columbia MSFE curriculum is a particularly dense public map of the finance-ML stack. Its electives include machine learning for financial engineering, deep learning for operations research and financial engineering, reinforcement learning, causal inference, big data in finance, AI applications in finance, algorithmic trading, model-based trading, programming for financial data and risk systems, and high-performance financial systems. The programme also publishes concentrations in asset management, computational finance/trading systems, financial technology, and machine learning for financial engineering. This is curriculum evidence, not proof that a student, faculty member, or employer has deployed any of these methods.

The associated Columbia IEOR ML and Analytics roster names a broad faculty network spanning interactive learning, bandits, reinforcement learning, online learning, interpretability, fairness, optimization, stochastic simulation, and finance. Three profiles make the finance connection unusually concrete: Agostino Capponi lists market microstructure, systemic risk, robo-advising, cryptocurrency analysis, and deep neural-network performance, and records research funding or collaborations involving public agencies and named technology and financial organizations; Wenpin Tang connects stochastic analysis and machine learning to blockchain-protocol queues, dynamic portfolio selection using repulsive point processes, and robust fintech methodology; and Ali Hirsa provides a practitioner-academic route through algorithmic trading, ML, deep learning, data mining, computational finance, and prior quantitative-strategy leadership. These pages establish faculty interests and affiliations, not a tracked manager’s model inventory, data permissions, or live authority.

The MIT Sloan MFin curriculum adds a different implementation signal. Alongside financial econometrics and analytics, the programme describes an Action Learning requirement in which teams work with finance practitioners on real-world problems and present findings to decision-makers. Its public course list includes Advanced Analytics of Finance, AI and Money, AI and Machine Learning Research in Finance, Advanced Data Analytics and Machine Learning in Finance, and Modeling with Machine Learning: Financial Technology. The latter explicitly spans valuation, credit, proprietary trading and hedge-fund strategies, portfolio management, market structure, risk, stress testing, NLP, and personal finance. The page also says practitioner participants can include investment managers, hedge funds, private equity, venture capital, risk, and consulting. It does not identify confidential project sponsors, datasets, model ownership, or production adoption.

This creates a more useful research map than a single “AI” label: (1) information extraction and representation, (2) cross-sectional and time-series prediction, (3) interactive decisions and execution, (4) portfolio and risk constraints, and (5) governance, cost, and human review. The next recovery step is to match named faculty papers, syllabi, capstones, guest speakers, and alumni records to explicit employer evidence, keeping course descriptions and manager disclosures as separate evidence classes.

September 4, 2026 — MIT professor route: patents, technology shocks, and automated markets

Leonid Kogan’s MIT Sloan profile adds a professor-led idea surface that is not reducible to text sentiment or generic forecasting. Kogan’s current research list includes using large language models to measure the information content of patents, studying technology’s effect on labor income risk, stock prices, and inflation, valuing crypto assets and automated market-making, and examining managerial compensation and agency problems in mutual funds. The profile also records his MIT faculty-leadership role for the MFin programme and his prior Wharton and Lehman Brothers research affiliations.

For the research queue, these are distinct lanes: patent and innovation information as a corporate-intangible signal; technology exposure as a cross-sectional or macro state variable; automated-market-making as a market-structure and inventory problem; and manager-compensation data as a governance or incentive feature. The public profile does not disclose a hedge-fund implementation, model weights, patent corpus licence, crypto venue, trading authority, or investment result. The MIT route should therefore be followed through papers, appendices, data/code statements, and programme-linked student work before any manager connection is made.

September 4, 2026 — professor and programme routes add non-sentiment signals

Cornell’s Bowers College research report describes a finance-AI design that does not reduce text to positive or negative sentiment. Martin Wells, Liao Zhu, Robert Jarrow, and Peter (Haoxuan) Wu are named in the report on “News-Based Sparse Machine Learning Models for Adaptive Asset Pricing.” The researchers scraped online financial news from 2013–2019, converted text and market information into numeric representations, and built asset embeddings that connect particular words and tradable assets to individual stocks or industries. The two named models separate the tasks: NEUSS selects stock-level information, while INSER selects industry-specific words before predicting industry returns. Cornell reports relative differences of 50% and 10% against the cited Fama–French five-factor benchmarks. Those are author- and institution- reported research results; they do not establish a live strategy, downloadable corpus, current model, data licence, or fund adoption. The recovery target is the paper’s exact rolling split, embedding construction, sparse-selection path, turnover and cost treatment, and any code or news archive.

Carnegie Mellon’s MSCF faculty roster adds a different professor route. Zeigham Khokher is listed as an Associate Teaching Professor of Finance; his profile describes energy-finance work on predicting potential oil mispricing and a research path from modeling and estimation toward machine learning in asset-management computation and statistics. Javier Peña’s profile lists financial optimization, machine learning, and convex optimization, with a Cornell Applied Mathematics PhD and prior consultation with Axioma on algorithmic portfolio- management tools. The pages expose methods, lineage, and an industry implementation connection for portfolio tooling. They do not identify a tracked manager’s model, data, decision rights, or investment result; those claims require paper- and project-level recovery.

The Smith AI Initiative for Capital Market Research at Maryland is a useful institutional route because it publishes both the supply chain and the people. The page names Sean Cao as director and co-founder, says the initiative is designed to move academic AI into finance and accounting practice, and lists unstructured inputs including conference-call transcripts, press releases, annual reports, ESG disclosures, social media, product or operational images, and fund-manager disclosures. It also describes an open textbook, companion videos, a cross-school AI literacy course, and planned agentic-AI projects that reduce verification costs. The 2026 donor update on the same page says a course has used real business problems supplied by GRF CPAs and Advisors and reports a global Coursera reach of approximately 3,000 learners. This is unusually explicit about education, partner-supported workflow design, and multimodal source classes; it is not evidence that a hedge fund receives the student work, owns a model, or grants an agent portfolio authority. The next pass should archive the textbook, course notebooks, public video library, initiative faculty roster, and GRF project boundaries.

Rice’s Bruce Carlin profile and the NBER working paper “AI Managed Household Portfolios” add a prospective LLM-portfolio experiment. Bruce Carlin, Ryan Israelsen, and Christopher Wazzan collect a daily time series of stock recommendations from several LLMs and examine style, media attention, diversification, momentum, size, and book-to-market exposure. The paper’s abstract reports that the generated portfolios are undiversified, load on several familiar characteristics, and do not show statistically significant abnormal returns under the cited Daniel et al. methodology. This is valuable as a reproducible evaluation design and a public negative result, not as a manager disclosure. The research file to recover is the dated prompt archive, model/version history, recommendation timestamps, trading rules, corporate-action handling, and cost and capacity assumptions.

UBC Sauder supplies a multimodal and programme-level route. Allen Hu’s faculty page lists Big Data and AI in Finance as a research interest and records work on video-based investor persuasion, financial-news production, asset complexity, and information acquisition. Sauder’s COMM_V 486I Applications of AI in Finance describes student projects using alternative data, factor and fundamental analysis, economic forecasting, portfolio optimization, and trading strategy development. Its public methods list spans decision trees, random forests, CNNs, RNNs, Vision Transformers, and NLP, with real market data and end-to-end applications. Sauder’s related account of the video-persuasion study says the researchers processed more than 1,000 real-world pitch videos using facial, vocal, and verbal features and made the code available. The page and course expose a research and talent pipeline around visual and vocal signals; they do not show a tracked fund’s use of those signals, licensing, or portfolio authority.

McGill’s Chengyu Zhang thesis-defense record is a compact map of three finance-ML problems. The 2024 event names Ruslan Goyenko as committee chair and describes essays on trading-cost-aware nonparametric dynamic portfolio optimization, a stock-versus-option-characteristics prediction exercise, and earnings- surprise prediction from 10-Q and 10-K text. The abstract specifically separates classic ML and bag-of-words or sentiment methods from finance-objective-trained LLMs, and states that the latter were the models that captured contextual filing information in the reported experiment. The public event record does not provide the thesis PDF, prompts, training corpus, code, or a manager connection. It should be treated as a research hypothesis until those artifacts and point-in-time tests are recovered.

The University of Queensland route adds risk-aware numerical learning rather than a new sentiment label. Duy-Minh Dang’s profile identifies him as Senior Lecturer and Director of the Master of Financial Mathematics, with Toronto PhD training, a Waterloo postdoctoral route with Peter Forsyth, and industry collaborations across superannuation, investment, banking, and finance. The page lists 2026 work on Fourier-mixture neural density estimation, neural-network policies for risk-reward optimization, and Fourier-trained transition kernels for Mean-CVaR optimization, alongside a machine-learning project for defined-contribution superannuation. The research direction is density and policy learning under financial constraints, not simply text classification. The public page does not establish a fund partner, production model, or investment result; the papers, data provenance, and project briefs are the recovery targets.

Finally, Technion’s CRML project page provides a useful failure-mode record. A student team built a crypto trading agent using price and volume history, Reddit posts and comments, NLP, deep learning, crawlers, and a trading algorithm, then reported that it did not find the expected significant correlation between the textual inputs and prices. The same project tested risk-factor constraints in portfolio construction. This negative result is relevant to any alternative-data pipeline: the public page supports the existence of the experiment and its stated limitation, but not its full report, sample construction, leakage controls, or generalisation. It should be indexed as a falsification and evaluation route, not as evidence of a deployable system.

Taken together, these university routes expand the idea inventory into interpretable asset embeddings, local-energy or commodity mispricing, multimodal persuasion, filing-context understanding, risk-aware policy and density learning, and explicit no-signal outcomes. The crosswalk remains evidence-class based: professor or course pages show research and talent routes; papers show stated methods and results; employer pages or dated media are needed to establish a manager’s implementation, permissions, and decision authority.

September 4, 2026 — Shanghai and Tilburg routes expose factor-search and finance-data curricula

Yuan Zhang’s public research page identifies him as an Associate Professor of Finance at Shanghai University of Finance and Economics since 2025, with a PhD in Finance from the Swiss Finance Institute at EPFL. His current project list names consumer credit and asset prices, large and deep factor models, evolutionary factor search, and language-model methods for alpha mining and executable option strategies. The page links the 2026 ICLR paper AlphaBench: Benchmarking Large Language Models in Formulaic Alpha Factor Mining and an in-progress project on evolutionary factor searching for sparse portfolio optimization using LLMs. It also lists collaborations with Semyon Malamud, Bryan Kelly, Boris Kuznetsov, and Teng Andrea Xu. This is a concrete public route into LLM-assisted factor generation and option-strategy expression; it does not establish model access, data rights, trading authority, or a fund’s use of the work. The recovery target is the AlphaBench evaluation protocol, factor grammar, selection constraints, out-of-sample design, and any public code or benchmark data.

The Tilburg MSc Finance: AI and Data Science track provides a separate European programme route from Tilburg’s quantitative-finance track. Its 2026–27 curriculum combines advanced corporate finance, investment analysis, empirical methods, AI and data-science methods in finance, AI and data-science applications in finance, and a master’s thesis. The page explicitly describes structuring large volumes of information for M&A, investment, trading, and banking analysis, while the public course catalogue and thesis route leave instructors, data partners, student models, and results unspecified. This is evidence of a formal talent pipeline and intended finance applications, not evidence that an employer receives a student system or grants it production authority.

These two routes add evolutionary factor discovery, formulaic-alpha benchmarking, executable option expressions, and information-structuring curricula to the idea map. They should be cross-checked against the existing SUFE AI Finance Laboratory and Tilburg faculty records; shared authorship or programme affiliation alone should not be converted into a manager link.

September 4, 2026 — Nanjing University route: finance-economic foundation-model claims and data quality

Nanjing University’s Chinese-language FinEconLLM laboratory page is another title-blind source for finance-AI research. The lab says it combines financial, economic, and technical expertise to build large models and tools for asset pricing, market forecasting, economic policy, decision support, sentiment and public-opinion monitoring, text mining, wealth management, investment, and risk management. It describes a project called FinEconBrainGPT that would use financial-market data, economic indicators, news, and financial reports for analysis, forecasting, and decision support. These are laboratory descriptions and intended applications, not a model card or evidence that a fund uses the system.

The page names Xiaowei Ding as a finance-information and intelligence faculty member and describes a Stanford financial-engineering and statistics training route. A separate Nanjing University faculty profile publicly claims prior big-data and AI quantitative-finance experience at JPMorgan, Morgan Stanley, and Citadel, plus current work on large-data/large-model/large-compute methods, financial information, blockchain, and financial security. Because those employment and adoption statements are displayed by the institution but were not independently corroborated in this pass, they are recorded as first-party biography claims. The profile also advertises courses on AI large models and next-generation digital finance and a 2026 research recruitment route covering foundation models, agents, world modeling, pre-training and post-training, alignment, and low-cost model construction.

The useful extraction is the lab’s proposed chain from data collection and quality to model construction, decision support, and trusted infrastructure, plus the explicit recruitment of finance, statistics, computer-science, and AI researchers. The recovery queue is to obtain FinEconBrainGPT technical artifacts, papers, code, model/data statements, course materials, and independent employment records. Nothing in the reviewed pages establishes a tracked manager’s reporting line, proprietary data access, trading authority, or investment result.

September 4, 2026 — Stevens LLM-agent route: memory, agent hierarchy, and benchmark design

Stevens’ university account of LLMs in financial decision-making names doctoral researchers Yangyang Yu, Haohang Li, and Yupeng Cao and identifies Jordan H. Suchow as the Stevens adviser. The article describes three related public projects. FinMem uses agent profiling, layered memory, and a decision module to retain and prioritize market information. FinCon uses a manager–analyst communication hierarchy: specialist agents process sources such as news, industry reports, and earnings calls, while a central agent makes decisions and sends feedback. Its paper describes selective verbal reinforcement and a risk-control component that periodically self-critiques and updates investment beliefs. FinBen supplies an open evaluation route spanning 36 datasets and 24 financial tasks across information extraction, textual analysis, question answering, generation, risk management, forecasting, and decision-making; the paper also describes stock-trading and retrieval-augmented-generation evaluation.

The Stevens article says unnamed firms contacted the team about applying the single-agent system to live investment strategies. No firm names, contract, data rights, model version, production endpoint, capital authority, or return audit are disclosed, so this remains a partnership lead rather than hedge-fund evidence. The useful research implication is architectural: keep memory and decision state distinct, route evidence through specialist roles, make belief changes reviewable, and benchmark the whole finance task surface rather than relying on sentiment or document-QA scores. The relevant recovery targets are the repositories, dataset licenses, temporal splits, trading assumptions, and independent confirmation of any industry contact.

September 4, 2026 — University of Florida AI-in-finance professor route

Yuehua Tang’s University of Florida faculty profile identifies him as the Scott J. Friedman Professor of AI in Finance and chair of the Finance, Insurance and Real Estate Department. The page lists AI in finance, fintech, mutual funds, hedge funds, investments, and climate finance as expertise areas. It also records a Georgia State University finance PhD, prior Singapore Management University and Emory appointments, public research-workshop and doctoral-teaching roles, and a linked media route to the ChatGPT-and-stock-movements study with Alejandro Lopez-Lira. The displayed publication list includes fintech disruption, mutual-fund incentives and skill, climate risk, and investment research.

This is a professor, publication, and talent-network route. The profile does not identify a manager’s model, data licence, production system, agent permissions, investment authority, or AI-attributed result. The research recovery queue is Tang’s current CV and papers, the Lopez-Lira study materials, seminar recordings, doctoral-student graph, and any explicit industry sponsor or project statement.

September 4, 2026 — Queen’s Belfast finance programme and real-money student-fund route

Queen’s Business School’s 2026/27 Finance PhD page describes strengths in empirical asset pricing, machine learning and AI in finance, and engagement with the Finance & Artificial Intelligence Research Lab (FAIR). It also lists WRDS, CRSP/Compustat, the London Share Price Database, Bloomberg, and S&P Capital IQ, alongside high-frequency trading simulations and the FinTrU Trading Room.

The FinTrU Trading Room page describes 26 Bloomberg terminals, trading simulations, and financial software embedded in taught programmes and the Queen’s Student Managed Fund. A 2026 university account says the student fund oversees £100,000 in partnership with Davy Group and that students make investment decisions with faculty oversight. The 2023–24 student-fund report lists speakers and contacts from BlackRock, Clarendon Fund Managers, Citadel, and Mediolanum. These are dated academic, sponsor, and speaker-network records; they do not identify a manager’s AI model, data licence, production system, or use of student work.

Queen’s research profile for Alan Hanna adds a named professor and practitioner lineage. It lists current teaching in Python for Finance and earlier teaching in Computational Methods in Finance, with research interests in computational finance, derivative pricing, machine learning, and sentiment and text analysis. The profile records a PhD and BSc in pure mathematics from Queen’s, prior financial-engineer work at Microgen, and an Equity Risk Project Lead role at Citi. It also identifies Hanna as a member of the Queen’s Student Managed Fund oversight committee and links research outputs on news media and investor sentiment and bull/bear market identification. These are first-party academic and career records; they do not establish that Hanna’s research, course code, or student-fund activity is used by a named manager.

The research recovery queue is FAIR personnel and projects, current PhD supervisors, student-fund presentations and recordings, and any explicit research or recruiting relationship with a named manager. The Hanna profile adds a specific personnel and artefact queue: recover the Python-for-Finance and Computational Methods syllabi, linked papers and open files, the student-fund oversight materials, and any dated external teaching or conference recordings. Keep Hanna’s academic role, prior Citi employment, student-fund oversight, research output, and any later employer or manager relationship as separate evidence states.

September 4, 2026 — Wharton funding, finance-AI projects, and quant-talent route

Wharton’s AI & Analytics Initiative funded-research index adds a new institutional source surface. The page reports 179 funded projects across 10 departments and more than $2.57 million invested since launch; it says awards can support data acquisition, computing resources, research assistance, and in some cases doctoral or postdoctoral support. This is evidence of an academic funding mechanism, not evidence that a hedge fund sponsors, owns, or deploys the resulting work.

• Winston Dou and Yuecheng Su’s “In Search of Distress Risk with Textual Data and Machine Learning” builds a sentence-level database from analyst reports, 10-K filings, and earnings-call transcripts, then uses an LLM API to classify economically motivated distress channels such as liquidity, leverage, competition, supply-chain pressure, governance, and demand shocks. The scores are combined with accounting and market variables. This creates a concrete filing-and-call route for default-risk and early-warning research; the page does not disclose prompts, model versions, corpus vintages, point-in-time controls, licences, or manager use.
• Marius Guenzel and Shimon Kogan’s projects use computer-vision embeddings from LinkedIn profile images. The initiative describes a nearly 1.5-million-person dataset linking images, labour histories, demographics, and credit-bureau records, and reports facial features adding predictive information for delinquency beyond conventional credit inputs, especially for thin-file borrowers. This is a multimodal soft-information and governance research lead, not evidence of an investable signal or permission to use sensitive personal data in a portfolio process.
• A separate Dou–Itay Goldstein project studies whether reinforcement-learning trading algorithms can coordinate, create bubbles, execute coordinated exits, and alter liquidity or market stability. The page frames these as theoretical, numerical, and empirical questions with regulatory implications; it should not be read as a finding about any named manager or as proof that such behaviour occurs in live markets.
• Alina Song and Joao Gomes’s “Data Gravity” links the near-universe of data-centre locations to online job postings and reports that firms shift AI human-capital investment toward establishments near greater data-centre capacity. This supplies a public alternative-data idea for mapping compute concentration and AI labour demand; it does not reveal the raw job-posting data, identifiers, update cadence, or a trading implementation.

The related Wharton Jacobs MSQF programme adds the talent and practitioner layer: its public description combines quantitative methods, machine learning, AI, financial markets, data labs, and an Applied Research Practicum with industry partners. The page identifies David Musto as faculty director and records prior quantitative-asset-management work at Roll and Ross Asset Management and Trout Trading. That is a documented academic-to-industry lineage, not evidence that a current student, faculty member, or partner fund uses a particular model.

These Wharton pages extend the idea map beyond generic sentiment: narrative distress channels, multimodal credit information, AI-agent market interaction, compute-geography and labour signals, and an applied quant-finance talent pipeline. The recovery queue is the project papers, model cards, prompts, data statements, privacy/consent boundaries, course projects, practicum sponsors, and any explicit employer-authority disclosures.

September 4, 2026 — Michigan Ross finance-ML curriculum and PEAK6 ecosystem route

Michigan Ross’s finance area page and course catalogue expose a distinct academic-to-practice route. The catalogue lists FIN 427, “Artificial Intelligence and Machine Learning in Investment Strategies,” which describes quantitative approaches using firm information to develop trading strategies; it also lists FIN 342, “Big Data in Finance,” as a quantitative-asset-management course. The finance PhD page adds two weekly seminars, cross-registration across economics, engineering, and mathematics, and research papers under faculty guidance. These pages describe training and research infrastructure, not a live manager system.

The Michigan Ross FinTech Initiative adds the people and implementation surfaces. It identifies Andrew Wu as faculty leader, describes machine learning and automated text analysis on large unstructured data, and records a Wharton finance PhD and Yale mathematics/economics undergraduate background. The initiative says it offers undergraduate and graduate fintech courses, industry/alumni outreach, and sponsor-facing multidisciplinary action projects. Its page identifies Chicago-based PEAK6 as the founding partner. The public record supports a university–industry network and a route for recovering projects and personnel; it does not establish that PEAK6 funded a particular model, received student output, or granted an AI system trading authority.

The Ross FinTech academic-offerings page also lists courses in Python, big-data manipulation, AI, quantum technologies, and global FinTech projects, plus an online “Financial Technology Innovation” series and the MAP project format. This creates a concrete capture queue for syllabi, project sponsors, student artefacts, public lectures, and alumni destinations. It should remain separate from evidence about any tracked hedge fund’s internal research or GenAI stack.

September 4, 2026 — PEAK6 Applied AI, options engineering, and academic-to-trading route

PEAK6’s first-party history is a direct AI-strategy signal. In a page dated January 11, 2024, PEAK6 says AI tools were being evaluated across the business and identifies a new “Applied AI” department whose stated purpose was to automate everyday tasks so teams could spend more time on new possibilities. This is a firm-reported organizational and workflow claim. The page does not name the department’s leader, disclose a model or vendor, describe training data, or establish that an AI system made investment decisions.

The firm’s PEAK6 Capital Management page describes the investment entity as a proprietary options-trading firm trading U.S.-listed equities and supplying liquidity in the U.S. options market. It says the business manages long and short options inventory over periods from days to years, and combines technologists, engineers, and traders with proprietary technology and a data-first approach. A separate 2020 acquisition release says Hardcastle Trading added proprietary technology and engineers, and was intended to extend electronic-options liquidity in the U.S. and Europe. These disclosures locate the public technology strategy in options-market infrastructure and trading workflow; they do not reveal the feature set, model class, data rights, or performance contribution.

The Trading Associate 2027 posting adds an unusually clear authority and talent signal. PEAK6 says associates move through basic training, basic trading, rotations, and individual portfolios; learn proprietary tools; collaborate with trading teams on new tools; explore both automated and discretionary trading; and test and analyze their own strategies in the market. The campus page describes an 18-month graduate route, Series 57 licensing within two months, supervised rotations, and individual-portfolio strategy testing, as well as a nine-week options and volatility internship. This describes a human-led training and experimentation pipeline, not an autonomous-agent permission map.

The academic connection is explicit at the network level but not at the project level. Michigan Ross identifies Jenny Just as a BBA ’90 alumna and PEAK6 co-founder/managing partner, and describes her discussion of financial technology and adapting to artificial intelligence on the school’s Down to Business podcast. Ross’s FinTech Initiative names PEAK6 as its founding partner and Andrew Wu as faculty leader, while the course catalogue lists FIN 427 (“Artificial Intelligence and Machine Learning in Investment Strategies”) and FIN 342 (“Big Data in Finance”). The reviewed public pages do not identify a PEAK6-sponsored FIN 427 project, student deliverable, transferred model, or trading-authority arrangement.

The PEAK6 news archive is now a required first-party discovery surface for this firm. It exposes dated announcements across PEAK6 Capital Management, Apex, Bruce Markets, and other portfolio companies, including Jenny Just’s public media appearances, the Standing Table podcast, and a Nasdaq data partnership for Bruce Markets. Those entities must remain separated: a PEAK6 group-company data or media announcement is not evidence of PEAK6 Capital Management using that data or technology in investment research. No named Applied AI personnel, model inventory, GenAI vendor, training corpus, or AI-attributed return was found in the reviewed first-party pages. The recovery queue is the Applied AI team’s personnel history, job snapshots, Michigan Ross MAP and course artefacts, podcast audio and transcript, and any explicit project-level disclosure.

September 4, 2026 — Chicago quantitative-developer certificate creates an explicit industry advisory channel

The University of Chicago Data Science Institute’s July 28, 2026 programme announcement adds a new, concrete academic-to-industry route. The DSI and Financial Mathematics Program are launching a Quantitative Developer Certificate for Autumn 2026, combining advanced software engineering with financial mathematics for systems used in trading and risk management. The stated Industry Advisory Board includes Chicago Trading Company (CTC), DRW, and Millennium; the announcement says the firms contribute industry expertise and case-study topics based on challenges they face. This is direct evidence of curriculum influence and talent engagement, not evidence that a firm supplied confidential data, adopted student code, or transferred an investment model.

The certificate page specifies one capstone with deliverables presented to advisory-board members and UChicago faculty. It says the programme is open to graduate students from data science, financial mathematics, computer science, computational and applied mathematics, finance, statistics, physics, and related fields, and is anticipated to launch in Autumn 2026. The announcement describes the applied work as including real-world systems, backtesting engines, production data pipelines, live-trading integrations, and AI tools. The public pages do not provide the case-study prompts, sponsor-specific requirements, data permissions, student identities, or post-capstone employment outcomes.

The same university page identifies Mark Hendricks as Director of Financial Mathematics and an Associate Senior Instructional Professor, with prior experience spanning hedge funds, trading firms, asset managers, and private equity, including quantitative research, systematic trading, and risk management. It identifies David Uminsky as DSI Executive Director and Research Professor with machine learning, signal-processing, pattern-formation, and dynamical-systems interests, and Arnab Bose as Senior Instructional Professor and Faculty Director of the MS in Applied AI programme. These are public faculty and programme lineages; they do not establish a current investment remit, firm reporting line, model ownership, or production authority.

This route should be tracked as a talent-and-method supply chain: quant-development curriculum → advisory-board problem framing → capstone artefact → employment or adoption. Only the first two links are public in the reviewed material. The recovery queue is the certificate course list, capstone deliverables, later advisory-board updates, faculty and student profiles, and any authorized recordings or project write-ups.

September 4, 2026 — Wharton hedge-fund coursework adds a portfolio-reality control surface

Wharton’s Ronen Israel faculty page lists FNCE2450 and FNCE7450 Hedge Funds in the Fall 2026 schedule. The course description specifies real-data exercises and projects, mathematical modelling, and backtesting of hedge fund and proprietary-trading strategies. It also covers fund structuring, trading, liquidity, funding, risk management, and performance measurement. This is valuable as a finance-programme control surface: an AI or alternative-data idea can be tested against the same implementation, risk, and measurement vocabulary used to study hedge-fund strategies. It is not evidence that any named manager supplied data, adopted course material, or uses a particular model.

The page also lists an INSP4997 senior capstone for Huntsman students. The public description does not expose current student names, project topics, datasets, employer sponsors, or results. The recovery queue is the two course syllabi, project and guest-speaker material, public capstone outputs, and any explicit connection between course participants and tracked firms.

September 4, 2026 — Regional finance-AI programmes expose additional research and talent routes

The academic map now includes several first-party routes outside the existing North American, European, and East Asian cluster. They are useful for idea discovery and personnel lineage, but none is evidence that a tracked hedge fund uses the named curriculum, data, or models.

The University of Dundee’s BU52057 module adds a named UK module lead and a practical assessment route. Dr Murat Mazibas leads a 20-credit course covering financial-data preparation, feature engineering, supervised, unsupervised, and reinforcement learning, algorithmic trading, portfolio optimisation, risk measurement, Python and Matlab, ethical/regulatory constraints, and a Financial Machine Learning Application Project. The page says it is available within the Business Analytics and Financial Technology master’s routes and assesses both formal knowledge and group application. This is an education and talent-pipeline signal; it does not disclose employers, datasets, project results, or any fund’s adoption.

PUC–São Paulo’s “Artificial Intelligence in Actuarial Science: Methods and Applications” is a live, synchronous 32-hour extension course scheduled from September 12 to November 7, 2026. Its public outline is unusually concrete about the applied stack: regression, decision trees, random forests, XGBoost, neural networks, Python, TensorFlow, Keras, ChatGPT, and the Whisper API. The stated projects cover risk classification, proposal approval, pricing, churn, fraud, document OCR, and claims prediction. Elizabeth Borelli is identified as coordinator, with Bruno Pereira Cunha and Gabriel Cassagni as instructors. This is a Portuguese-language training and practitioner route spanning insurance, pensions, and finance; it does not disclose student identities, datasets, model results, employer adoption, or live portfolio authority.

King Fahd University of Petroleum & Minerals’ graduate finance bulletin adds a stronger research-method signal from Saudi Arabia. FIN 654, “AI in Finance and Big Data Analytics,” combines finance data architecture, cloud and scalable analytics, event studies, causal inference, quantitative investing with pandas, scikit-learn, TensorFlow, LSTM/Transformer forecasting, leakage controls, model risk, reproducibility, and MLOps. The same bulletin lists portfolio management, hedge funds, and an industrial computational-finance project whose report and presentation are reviewed by supervisors and industry experts. The course outline is evidence of an explicit educational stack and evaluation vocabulary, not of which firms provide the projects, what data are licensed, or whether any strategy is traded.

The University of Sharjah’s MSc in Financial Technology is a new 33-credit graduate programme with its first intake in Fall 2026/27. It combines financial econometrics, big-data management in finance, AI, blockchain, cybersecurity, a research thesis, and computational finance in C++. The public course descriptions include real-time analytics, predictive modelling, portfolio optimisation, Monte Carlo and VaR, algorithmic trading, and a final financial-application project. The programme is explicitly oriented to UAE, Gulf, and global financial ecosystems. It is a talent and research pipeline; the page does not identify thesis topics, supervisors, datasets, industry sponsors, or production deployment.

The Islamic University of Madinah profile for Dr. Asaad bin Mahdar Sindi provides a personnel-and-lineage route rather than a programme-only signal. The profile lists a 2024 PhD in Financial Economics and MSc in Computer Science (AI and ML) from the University of New Orleans, a finance MSc, current academic leadership in Madinah, and research topics including AI and stock-market forecasting and comparisons of GRU, LSTM, and Transformer models in technology equities. It also lists ESG, oil-price volatility, market efficiency, and portfolio-resilience work. These are public biography and publication claims; the page does not establish peer-reviewed status for every listed title, code, data provenance, hedge-fund employment, or live investment authority. The profile should therefore be mined next for DOIs, working-paper versions, coauthors, and university research records before any lineage claim is strengthened.

These routes broaden the idea map in three directions: actuarial and insurance decisioning, research controls for leakage and reproducibility, and Gulf-region finance/FinTech talent formation. They should be treated as hypothesis generators and recruitment-network surfaces, not as evidence that the associated methods have been adopted by any particular tracked firm. The recovery queue is the Dundee module materials and Mazibas research/profile records, PUC–SP Lattes records and project artefacts, KFUPM syllabus and industrial-project sponsors, Sharjah thesis supervisors and student outputs, and DOI-level verification of Sindi’s listed research.

The University of Glasgow profile for Professor Charalampos Stasinakis adds a particularly useful professor-to-idea route. His public training is in computer and electrical engineering at the National Technical University of Athens and quantitative finance at Glasgow; his stated interests include AI, neural networks, heuristics, machine learning, big data, forecasting, risk, portfolio optimisation, and financial technology. The profile lists 2026 work on machine learning and big data in finance, alongside research on ESG portfolio diversification and decentralised lending, and earlier papers on hybrid volatility/sentiment Bitcoin trading, XGBoost–MLP credit risk, Bayesian FX trading, and neural/Kalman forecasting. It also identifies current doctoral supervision on ESG investment with ML, narrative-disclosure quality, and FinTech/shadow banking, plus current postgraduate teaching in ML in finance and financial technology. This exposes a research agenda and training network—not a hedge-fund relationship, proprietary dataset, or deployable alpha claim. It is a strong recovery target for paper-level validation, code, doctoral coauthors, and conference recordings.

The University of Otago profile for Dr Muhammad A. Cheema adds a New Zealand investment-research route with a named market-data facility: he is Senior Lecturer in Finance and Director of the BNZ Bloomberg Markets Lab. The profile connects his teaching and research to investments, behavioural finance, machine learning, and energy finance, and lists current Applied Investments and Behavioural Finance courses. Its public publication list includes investor sentiment and stock-market anomalies in Islamic countries, safe-haven assets across market downturns, ESG decoupling and crash risk, and environmental shocks to markets. It also identifies active PhD supervision in investments, behavioural finance, ML, and energy finance. This is a research and talent surface; the Bloomberg lab, papers, and courses do not establish a hedge-fund deployment, proprietary data right, or live strategy.

Otago’s Dr Olena Onishchenko profile adds a second personnel and supervision route. It lists empirical capital markets, short selling, securities lending, market microstructure, and ML in finance, and records current PhD projects on ML applications for financial risk management and geopolitical risk using ML. The profile also connects her to investment and portfolio-management teaching and Bloomberg Trading Challenge advising. These entries expose research questions and a student pipeline, not a manager relationship, portfolio authority, or evidence that student work is used in production.

Xiamen University Malaysia’s profile for Dr Bochuan Dai is a cross-border personnel route. Dai is an Assistant Professor of Finance whose stated research includes empirical asset pricing, corporate innovation, trading strategies, and LLMs in financial analysis. The profile records a PhD and undergraduate finance degrees from Massey University in New Zealand, a visiting-scholar period at Edinburgh, current adjunct research at Tsinghua, and prior positions at Tsinghua and Central South University. It lists research on half a million Chinese firms, stop-loss rules, international equity allocation, and corporate innovation. The LLM reference is a research-interest disclosure; no model, corpus, benchmark, vendor, code, or investment deployment is identified.

Universidad del CEMA’s quantitative-finance programme adds a Spanish-language Southern Cone curriculum with implementation detail. The programme is jointly designed by finance and computer-science departments and covers portfolio theory, derivatives, credit, risk, algorithmic-trading systems, strategy testing, Python, databases, data mining, big-data modelling, and quantitative strategy generation. Its ML-for-finance module names bias–variance trade-offs, train/test design, cross-validation, probabilistic and neural models, automatic differentiation, scikit-learn, and TensorFlow; related modules cover scraping, asynchronous Selenium/Playwright workflows, exchange APIs, and on-chain data. The programme requires a final project or integrative exam. It is a public training and talent route, not evidence of a fund’s data access, production stack, or performance.

Together these routes add Australasia and Argentina to the academic coverage, and make three idea clusters more explicit: sentiment and behavioral effects, risk/microstructure/geopolitical signals, and LLM-assisted financial analysis plus the data plumbing needed to test them. The next recovery targets are Otago syllabi and thesis artefacts, Dai’s LLM papers and code, UCEMA faculty biographies and project outputs, and any named finance-industry partners.

September 4, 2026 — Practitioner-linked professors and finance research routes in Asia and Africa

The City University of Hong Kong profile for Professor Li Wei connects an academic appointment to a current institutional-investment role. It identifies Li as an adjunct professor, Head of Multi-Asset Investments at BNP Paribas Securities (China), and a former Citi quantitative-investment practitioner in London and Hong Kong. His listed research areas are quantitative investment, machine learning, big data, operations research, reliability engineering, and computational finance; the publication list includes macro-news effects on FX implied volatility, secrecy-preserving verification for dark pools, high-frequency FX news, and ESG funds in China. The page also states that he has managed multi-billion-dollar funds, but that is a self-reported profile claim rather than an independently audited AUM or performance record. The public page does not disclose models, data rights, internal BNP systems, or investment authority boundaries.

The CUHK Business School profile for Professor Gang Li adds a research-and-practitioner lineage with unusually specific public work. It identifies a Renmin statistics → Fordham MBA → University of Toronto PhD path, a former global-macro hedge fund role in the United States, and prior asset-management work in China. His interests include asset pricing, derivatives, machine learning, investment, and FinTech. The profile lists a 2026 forthcoming paper on option trading and market risk premia, a 2024 working paper on forecasting option returns with news, a 2025 paper on stock-return autocorrelations and expected option returns, and a 2021 deep-learning/option-return-prediction grant. These records expose research hypotheses around news, options, volatility, and risk premia; they do not establish that the former employer or any tracked fund adopted them.

The HKUST profile for Andrew Chiu connects academic training to a named proprietary-trading firm. It lists a PhD and MBA from HKUST, a computer-engineering degree from Waterloo, research interests in option pricing, empirical asset pricing, machine learning, and quantitative trading, and identifies Chiu as founder of Algomeric Trading, described as a systematic stock-and-derivatives firm trading U.S. exchanges. This is a public founder and research-interest disclosure; it does not provide strategy descriptions, model versions, data sources, execution infrastructure, performance, or portfolio permissions.

Stellenbosch’s CV for Professor Evan Gilbert adds an African investment-management and academic route. It identifies Gilbert as a professor teaching and supervising investment-management research, with a Cambridge Judge Business School PhD and senior practitioner roles in South African investment management. The CV lists work on machine-learning segmentation and prediction of investor behavior, unsupervised ML risk archetypes, low-volatility effects in African frontier markets, factor investing, and portfolio risk measures. It also records substantial supervision across UCT, Stellenbosch, and Reading. The CV is a useful personnel and paper index, but it does not establish a specific fund’s use of the models, private data, or live trading authority.

These routes add practitioner-to-academic links in Hong Kong, mainland-China-connected finance, U.S.-exchange proprietary trading, and South African investment management. The recovery queue is independent employment verification, paper-level code and datasets, firm media and conference appearances, and any public evidence of model or research transfer.

September 4, 2026 — UT Austin AIM Investment Conference adds an academic allocator route

The University of Texas at Austin McCombs AIM Investment Conference page dates the 2026 conference for October 9–10 in Austin. It names Wei Jiang, Charles Howard Candler Professor of Finance at Emory and President of the American Finance Association, as keynote speaker, and identifies the AIM Investment Center and Dimensional Fund Advisors as joint sponsors. The call explicitly includes mutual funds, hedge funds, pension funds, and investment management among its submission areas. This is a professor, programme, and allocator-network route; it is not evidence of a Dimensional model, proprietary dataset, or deployment.

As of the September 4 retrieval, the page still said the 2026 programme would be published in August but did not expose the paper, discussant, or recording list. That unresolved publication gap is now tracked for recovery. Older programme entries provide follow-up research routes around hedge-fund performance fees, public-information acquisition, institutional trading around corporate news, and talent allocation in asset management. They identify questions and researchers, not current AI use, live strategy authority, or investment results.

September 4, 2026 — professor rosters and finance programmes as an idea-triage layer

The University of the Witwatersrand School of Economics and Finance people page adds faculty detail to the existing Wits Financial Technology programme route. It lists Moinak Maiti as Associate Professor and FirstRand Chair in Financial Data Science, with research interests in artificial intelligence, data science, digital-asset pricing, financial econometrics, and FinTech. It separately lists Lwazi Mhlambi with interests in machine learning, climate finance, and project finance, and Daniel Page with interests in asset pricing, asset management, investment styles, and factor modelling. These public roster fields identify potential research and recruiting paths in African and emerging-market finance; they do not identify a fund, proprietary dataset, model, or live mandate.

The USP e-Disciplinas page for EAD0830 confirms a 2026 course taught by Leandro dos Santos Maciel. The stated objectives cover regression/forecasting, optimisation, clustering, quantitative financial-data analysis, and applied projects. The official USP course syllabus provides the finer-grained historical teaching record: neural networks for economic and financial time series, K-nearest neighbours for company evaluation and credit rating, genetic algorithms for portfolio optimisation, and deep learning, decision trees, random forests, and support-vector machines. The 2026 page establishes current course identity; the 2023 syllabus is used only for the documented module content and should not be read as a 2026 syllabus or evidence of student deployment.

The UBA Institute of Calculus teaching page lists second-semester 2026 electives in “Quantitative finance, science and markets,” mathematical foundations of causality, and machine learning/predictive modelling. The UBA Faculty of Economic Sciences sustainable-finance seminar programme adds a dated Spanish-language speaker route from November 11, 2025: Rita Morrone and Darío Bacchini on ML beyond greenwashing, Rodrigo del Rosso on neural networks for sustainable finance, and industry-linked sessions with Banco Galicia and NEORIS. This connects causal, ESG, greenwashing, and neural-model research questions to named instructors and practitioners, but the programme does not expose code, datasets, validation, or investment outcomes.

Interpretation: these pages are useful for constructing an academic-to-manager watchlist. A candidate idea should be recorded with its professor, course, method, target variable, market, data vintage, and reproducibility artefact; it should then be tested independently before being linked to a manager. The present evidence supports idea and talent discovery only. Recovery targets are Wits faculty publication lists and research-group pages, USP Maciel/Jakel syllabi and project outputs, and UBA lecture recordings, papers, and programme-linked datasets.

September 4, 2026 — Strathclyde and Oxford expose model-governance and training routes

The University of Strathclyde MSc Quantitative Finance page adds a detailed UK programme route. For September 2026 entry, the cross-faculty programme lists Statistical Machine Learning, Machine Learning for Data Analytics, and Evolutionary Computation for Finance. The latter covers forecasting, portfolio optimisation, and algorithmic trading; the 40-credit Quantitative Finance Research Project can be university-based or jointly shaped with an industrial partner, with written-report and presentation assessment. The course page also says it was designed with finance-industry input and that industry projects may be possible. This is a talent and implementation pipeline, not evidence of a named employer, dataset, model, or live strategy.

Strathclyde also advertises a funded PhD project on the financial judgement of AI with one place, a 36-month duration, an August 24 opening date, and a September 18, 2026 deadline. The proposed research tests whether LLM political orientations affect financially grounded judgements, then compares baseline behaviour with analyst/adviser personas on stock forecasts, investment recommendations, ESG advice, portfolio construction, and justifications. It proposes a survey instrument and benchmark for cross-model auditing. This is a particularly useful disqualification and model-risk idea: it concerns whether an apparently neutral research assistant changes recommendations through latent preferences. The page does not yet identify the supervisor, model roster, prompts, data, results, or any fund relationship.

The University of Oxford’s Generative AI for Finance course is scheduled for September 7–17, 2026 and lists prompt engineering, portfolio management, risk assessment, predictive analytics, fine-tuning for financial tasks, fraud detection, compliance, privacy, and regulation. Oxford identifies Konrad Kleinfeld as tutor, with prior BlackRock, Nomura, and State Street Investment Management experience, and Claudia Otto as a legal guest lecturer on GenAI and regulation. The sessions are live and explicitly not recorded, so the course page is a programme/personnel signal and an archival recovery target rather than a transcript source. It does not establish the practices of those prior employers or any fund.

The University of South Carolina 2026–27 finance bulletin adds a concise US curriculum route: FINA 589 is a three-credit course applying ML and AI to investment, corporate finance, and banking. The same bulletin places it near courses in climate finance, financial-statement analysis, investment management, and risk. No instructor, project, dataset, or employer partner is named. Together these routes show why academic discovery should track not only papers but also course requirements, project supervision, evaluation methods, model-governance topics, and private/recorded status.

Recovery targets: identify Strathclyde supervisors and programme industry partners; capture the PhD project’s supervisor, model list, benchmark design, and later outputs; search Oxford tutor pages and any authorized post-course material; and locate South Carolina syllabi or faculty assignments. Preserve all four as academic or professional-training evidence, separate from manager deployment.

September 4, 2026 — open paper/code routes for asset-pricing structure and belief measurement

The Chunjie Wang research page identifies him as an assistant professor of finance at KU Leuven, with a Stockholm School of Economics finance PhD and stated interests in empirical asset pricing and machine learning. His job-market paper, “Asset pricing, not equity pricing,” links a Dropbox paper and a code repository. The page reports a comparison between asset-return and equity-return characteristic factors, including an out-of-sample MVE portfolio comparison and a reduction in the number of factors needed in the paper’s setup. Those numbers are author-reported and require recovery of the paper, code, sample, portfolio rules, costs, and point-in-time design; the page does not establish a manager relationship or deployment.

The Do Lee research page identifies a 2026 NYU economics PhD and an economist role at the IMF beginning in September 2026. His job-market paper uses machine learning forecasts as a benchmark for objective corporate-cash-flow beliefs, compares them with survey expectations, and links belief distortions to hiring and subsequent stock-price dynamics. The page links the paper’s PDF, online appendix, and slides. This is a macro/asset-pricing route into belief measurement, expectation formation, and event-window research; the page’s empirical claims remain author-reported until the appendix, data construction, forecast vintage, and code are checked. It does not identify a fund, proprietary dataset, or live strategy.

Together these pages add a useful distinction to the idea map: one route changes the object being priced and the factor representation; the other uses ML as a benchmark for measuring human belief error. Both have inspectable paper artifacts, but neither should be translated into a manager claim without temporal, cost, and replication checks.

September 4, 2026 — professor-led model-uncertainty and finance-programme routes

George Mason’s Bo Hu profile adds a professor and research-lab route with unusually clear academic lineage. Hu is an Associate Professor of Finance who teaches managerial finance, fixed income, and international finance in MBA and MSF programmes. The profile records a University of Maryland PhD in Finance, a UC San Diego PhD in Physics, and research interests in asset pricing, market microstructure, machine learning, and AI. It also links the Future of Finance Lab.

Hu’s paper, “Whence LASSO? A Rational Interpretation”, co-authored with Wen Chen and Liyan Yang and published online in 2025, frames LASSO as an equilibrium response to model uncertainty among competing traders. The abstract says the paper’s robust-trading setting can produce LASSO-type strategies endogenously, while also highlighting a trade-off: sparse estimation can reduce competition but introduce bias into trading decisions. This is a useful idea for testing feature selection under uncertainty, competition, and turnover—not simply a claim that sparse models generate alpha. The paper discloses Bank of Canada and SSHRC support and supplemental-data access, but does not establish any hedge-fund deployment, proprietary data arrangement, or live investment result.

The Vienna–Copenhagen Conference on Financial Econometrics adds an untracked conference surface. Its August 13–15, 2026 programme in Vienna is organised by the University of Vienna and supported by the University of Copenhagen. The published sessions include machine learning for high-dimensional realised covariances, information leakage and opportunistic trading around the FX fix, the nonstationarity–complexity trade-off in return prediction, debiased tests for Sharpe ratios of ML portfolios, Nelson–Siegel autoencoders for global yield-curve forecasting, and missing-data substitution for state-space forecasting. Those titles expose testable objects—covariance, market leakage, model complexity, inference, yield curves, and missingness—rather than treating “AI in finance” as one task. The event is a route to papers, presenters, and recordings; its programme does not establish a speaker’s employer system, model, or investment performance.

The academic-to-manager watchlist should therefore preserve three separate edges: professor and PI lineage; programme or conference method vocabulary; and any later firm or project corroboration. A paper’s theory, a course’s curriculum, and a manager’s production disclosure are different evidence classes.

September 4, 2026 — France’s MAQi programme and Louvain’s finance-ML research route

The new AI for Markets and Quantitative Investment (MAQi) programme from École Polytechnique and ENSAE Paris is a two-year programme beginning in September 2026. Its public description combines market finance with machine learning, requires mathematics, statistics, and Python, and includes projects, internships, and hackathons. It specifically mentions satellite imagery and maritime-traffic data for investment and risk-management work. The January 2026 institutional announcement names Charles-Albert Lehalle and Vianney Perchet as programme directors and says courses are co-taught by an AI/ML professor and a quantitative-finance professor. It lists BNP Paribas, Qube Research & Technologies, S&P Global, and the French AMF as supporting the programme. Those partnerships establish an education and industry-network surface, not a partner’s model, dataset licence, student work product, or investment deployment.

Louvain Finance adds a European research-centre route. The centre says it includes about 11 professors and 15 researchers, supports a Bloomberg learning centre, and receives support from TreeTop Asset Management and Candriam. Its 2026 news identifies PhD researcher Arnaud Germain’s work on clustering for credit risk, loan selection, and default prediction, while the centre’s discussion-paper list includes European sovereign-spread forecasting with machine learning and covariance shrinkage for portfolio selection. These are academic, sponsor, and talent-network signals; they do not show that a named manager uses the methods or that the public research produced an investable result.

September 4, 2026 — University of Bern workshop: ML asset-pricing ideas and limits

The University of Bern Institute for Financial Management adds a professor-led route that was not present in the coverage ledger. Its Machine Learning in Asset Pricing workshop ran May 18–21, 2026 and hosted Dacheng Xiu of the University of Chicago Booth School of Business. The official page frames Xiu’s work around when and why modern ML tools work, and where their limits lie. It lists return prediction, factor models, portfolio construction, textual and statistical learning, prediction, estimation, inference, and economic interpretation as workshop topics.

This is useful for the idea map because it ties model choice to an explicit validation agenda: forecasting and portfolio construction are paired with inference, economic interpretation, and algorithm limits. A practical research queue would therefore test whether a proposed signal is stable across vintages and regimes, whether its factor representation survives costs and turnover, and whether its apparent improvement remains after appropriate inference controls. Those are research-design implications, not claims about Xiu’s or any manager’s live strategy. The Bern page does not identify a hedge-fund participant, model implementation, dataset licence, recording, production permission, or investment performance.

September 4, 2026 — Academic routes for research agents, network text, and model uncertainty

Three additional public research artifacts extend the professor-and-programme map into underexplored methodological territory.

The September 1 Agentic Empirical Asset Pricing: Methodological Foundations preprint lists Yingjian Pan of Stanford’s Advanced Financial Technologies Laboratory and Management Science & Engineering, Xiaowei Ding of Nanjing University, and Kay Giesecke of the Hasso Plattner Institute. It defines “agentic empirical asset pricing” as an LLM-based system that can autonomously complete a hypothesis–formalisation–evaluation cycle. The paper proposes separate evaluation of the discovery process, not only the factor or trade it emits; describes the SEADS reference architecture; evaluates it against five re-implemented baselines on two US equity panels; and adds rolling re-execution across historical decision dates. The authors report negative findings and limitations, including that no single metric consistently distinguishes the systems. This is a research-agent evaluation protocol, not evidence of a hedge fund using SEADS or any other named system.

The Supply Chain Propagation of Textual Signals preprint by Asef Yılkı proposes augmenting LLM embeddings of annual-report disclosures with a supply-chain knowledge graph. The abstract describes FinBERT embeddings of 10-K MD&A sections for 255 S&P 500 firms over 2011–2025, then compares direct firm-level representations with representations whose signals propagate through inter-firm linkages. This creates a concrete alternative-data and network-design route: information can be tested both at the issuer and connected-firm level. The public abstract does not establish an employer, dataset licence, trading deployment, or independently verified investment result.

The DisclosureBeta theory preprint by Ping Kuen Wong treats an LLM as a noisy measurement channel for risk disclosures and puts channel noise and regime misclassification into an explicit beta error budget. It proposes a regime-conditional Fama–French five-factor setting, an adaptive blend of text-based and rolling-window estimators, and a frozen pre-registered panel of IPOs and recent listings; its empirical outcome is stated as forthcoming. The operational implication is a traceable measurement protocol: retain source quotes, model rationale, repeated scores, dispersion, and the resulting uncertainty rather than storing only a point estimate. This is a theory and pre-registration route, not evidence of a manager’s model, data rights, or live risk process.

September 4, 2026 — professor-led ideas beyond sentiment classification

The academic layer adds several specific research objects that can be followed without turning a professor, course, or paper into evidence about GMO, Acadian, Arrowstreet, or another manager. The useful split is by data object and decision point:

Research object	Public academic route	Idea exposed	Boundary for interpretation
Manager-behavior imitation	NBER Working Paper 34849, Mimicking Finance by Lauren Cohen, Yiwen Lu, and Quoc H. Nguyen	AI/ML prediction of mutual-fund trade direction from prior behavior; the abstract reports 71% predictability overall and higher predictability for some managers.	Reproduce with point-in-time holdings, disclosure lags, survivorship controls, and costs. Mutual-fund evidence is not automatically hedge-fund evidence, and a 13F-following use case has reporting delay.
Holdings-network stress mapping	NBER Working Paper 35227, The Optimal Use of AI in Financial Regulation by Christopher Clayton and Antonio Coppola	Graph-based deep learning over intermediary holdings to forecast trading behavior and stress-period asset-return variation; the authors describe nearly $40 trillion of non-bank-intermediary wealth in the study universe.	Recover the full paper and code, validate graph and timestamp construction, and keep systemic-risk forecasting separate from a tradable portfolio claim.
State-dependent covariance shrinkage	Learning the Shrinkage Intensity by Gianluca De Nard and Damjan Kostovic	An offline contextual-bandit policy chooses market-state-dependent linear or nonlinear covariance shrinkage for constrained risk portfolios.	Test across markets and regimes, turnover, constraints, and net-of-cost portfolio stability.
Learning market-maker pricing	Algorithmic Pricing and Liquidity in Securities Markets by Jean-Édouard Colliard, Thierry Foucault, and Stefano Lovo	A stylised Q-learning market-maker experiment studies adverse selection, limited exploration, noisy feedback, spreads, and noncompetitive pricing.	Treat it as a market-design and monitoring benchmark. It is not a disclosure of any proprietary market maker or proof that the stylised result transfers to live markets.

This professor-and-programme layer therefore supplies four additional research lanes: behavioral replication from public holdings, network structure during stress, adaptive risk estimation, and execution or pricing under learning. The linked papers report their own research designs and findings; they do not establish manager deployment, proprietary data rights, portfolio permissions, or investment performance.

September 4, 2026 — academic finance routes add compute markets and agentic curricula

The academic search found several additive routes beyond sentiment models. The inaugural Liechtenstein Workshop on AI in Finance was held May 6–7, 2026 and lists eleven invited papers with discussants, including machine learning in asset pricing and LLMs in financial analysis. Its later session was open to academically trained asset managers, financial-institution researchers, and quantitative analysts. This is a paper and speaker recovery route, not evidence of a named manager’s attendance, model, or deployment.

The University of Chicago Financial Mathematics 2026–27 ML/AI concentration lists Machine Learning for Finance, high-frequency-data analysis, Generative and Agentic AI for Finance, reinforcement and deep learning, multivariate statistics, and applied optimisation. It supplies a clear programme vocabulary for splitting future work into signal modelling, high-frequency data, research agents, risk, and optimisation. This is curriculum evidence, not evidence of any manager’s production architecture.

The Illinois 2026–27 Financial Engineering catalogue combines machine learning in finance, statistical methods, financial computing, optimisation, stochastic calculus, and a practicum, with applications across equities, bonds, derivatives, OTC markets, and digital assets. The catalogue identifies the programme director and practicum structure but does not disclose sponsor projects, datasets, or live trading use.

The Early AI Compute Asset Pricing preprint by Federico M. Bandi and Yinan Su adds a different research object: GPU compute rentals and proposed compute futures. The authors argue that non-storable compute does not inherit ordinary storage-based no-arbitrage links, construct preliminary synthetic return panels by GPU generation and maturity from term-rental data, and report preliminary evidence consistent with a positive compute risk premium and provider hedging pressure. This is an early framework using provider-supplied data, not a verified futures-market history or a manager strategy. The relevant follow-up is to test index construction, rental-contract selection, basis risk, GPU obsolescence, and contract-launch timing.

Together these routes widen the professor-and-programme map into agentic finance training, high-frequency implementation, and AI-infrastructure risk. They remain academic or educational evidence, separate from firm disclosures and performance.

September 4, 2026 — professor and programme routes for market design and institutional AI

The MIT IDE 2026 Annual Conference adds a Boston-based research-group surface beyond a single professor presentation. Its April 1 agenda names Eric So as lead of the AI in Financial Markets and Decision-Making group, with adjacent sessions on general social agents, workflow models for generative AI, AI-risk mapping, partial automation, human learning about AI, and distortions from AI search. So’s MIT profile identifies him as a Sloan Distinguished Professor, records Stanford GSB doctoral and Cornell economics training, and describes research on AI, human behavior, and market incentives. This is a personnel and research-paper recovery route. The agenda does not establish a hedge-fund attendee, commercial model, data rights, or investment deployment.

The ESSEC–Amundi Chair on Asset & Risk Management is a direct academic–asset-manager bridge. The chair says it was created in 2016 to support joint ESSEC–Amundi research in asset and risk management, including AI in finance; it publishes research letters and working papers, runs workshops and webinars, offers seminars to Amundi collaborators and institutional clients, and supports ESSEC PhD finance students working on related topics. Its February 17, 2026 webinar, “AI in Asset Management: From Gen-AI Signals to Portfolio Embeddings”, put generative-AI adoption measurement and asset embeddings on the same programme. A separate May 20 webinar was titled “AI and Machine Learning in Practice: Measuring Risk and Extracting Returns”. These pages establish a recurring research and talent channel, not Amundi’s production model, dataset permissions, portfolio authority, or independently verified performance.

The Princeton Bendheim Center for Finance paper by Markus Brunnermeier, dated August 2026, supplies a market-design idea rather than another signal-extraction proposal. Its abstract argues that agentic AI may create asymmetric understanding: agents can learn how humans respond while people cannot reliably anticipate the agents’ behavior. The paper links that possibility to less readable prices and a strategic disadvantage for central banks, then proposes preserving a human fallback, simpler and more robust rules, and less strategic ambiguity. This is a testable governance and market-structure hypothesis. It is not evidence of a hedge-fund system or a tradable edge; the working paper and its assumptions require separate validation.

These routes add three research lanes to the queue: human–AI decision experiments in financial settings, a sponsor-backed academic channel for GenAI signals and portfolio representations, and market-design tests for agentic systems whose behavior may be hard for people to anticipate. Keep professor, programme, sponsor, paper, and manager claims as separate evidence classes.

September 4, 2026 — Canadian workshop and practitioner-seminar routes

The inaugural University of Toronto Asset Pricing and Investments Workshop was scheduled for May 9, 2026 by Rotman and FinHub. Its programme has a dedicated AI-in-Finance session chaired by Goutham Gopalakrishna, pairing Mimicking Finance with A Financial Brain Scan of the LLM, and lists the authors, discussants, and institutions. The keynote speaker, Kent Daniel, is identified as a Columbia finance professor with prior faculty appointments at Northwestern, Chicago, and British Columbia and prior service in Goldman Sachs Asset Management’s Quantitative Investment Strategies group, where the page says he became head of the QIS equity-research effort. This is a Canadian academic and personnel bridge with an explicit academic-to-practitioner lineage. The page does not establish a fund’s attendance, a Goldman system, proprietary data, or deployment of either paper.

The Columbia Mathematics of Finance Practitioners’ Seminar Fall 2026 is another title-blind recovery surface. Its schedule includes Irene Aldridge of Risk AI Center, Thomas Yang of North Rock Capital, formerly Head of the Central Liquidity Book at Balyasny and a senior quantitative researcher and product head at Citadel, Karen Pham Van and Esen Ersoy of Davidson Kempner, Antoine Savine of Barclays, and Justin Hott of Hudson River Trading with Patrick Nichols of Old Mission. Yang’s October 14 abstract frames portfolio value through market impact, liquidation horizon, concentration, cross-asset risk, pre-trade cost estimation, and scenario analysis. The schedule supplies dated personnel, employer, and programme evidence; it does not establish any firm’s internal model, current reporting line beyond the published biography, data permissions, trading authority, or performance.

These routes expand the recovery queue in two directions: Canadian academic networks where AI papers sit inside broader asset-pricing programmes, and practitioner seminars where execution, liquidity, and model-governance details can appear without “AI” in the event title. Treat the programme, speaker biography, abstract, and any eventual recording as separate evidence objects.

September 4, 2026 — agentic financial engineering and executive-programme routes

Columbia’s Fall 2026 Agentic AI for Operations Research and Financial Engineering course is a more specific programme signal than a generic AI-in-finance label. The official course record names Agostino Capponi as instructor and describes hands-on design and deployment of end-to-end agentic pipelines using large language models, retrieval- augmented generation, conversational agents, semantic workflows, and streaming architectures. The stated applications include forecasting, optimisation, reasoning, supply-chain risk, energy systems, inventory management, and portfolio management. This exposes an implementation vocabulary for research and decision agents; a course description does not establish a firm’s production stack, data permissions, or autonomous investment authority.

MIT Sloan Executive Education’s Artificial Intelligence for Financial Services course is led by Andrew W. Lo and publishes a two-day sample schedule. It includes the evolution of AI in finance and quantamental investing, measuring AI exposure across firms and the labour market, interpreting and trusting LLMs in high-stakes applications, technology and data in modern investing, governance and regulation, deployment in financial institutions, complex-organisation implementation, computing constraints and the economics of AI deployment, and models-to-decisions exercises. The course is explicitly aimed at investment-management, broker/dealer, risk, and insurance decision-makers. That is a useful executive-programme and professor route for tracking how implementation questions are framed publicly; it is not evidence of a participant’s internal system or investment result.

Together these routes add two concrete research questions: what permissions, observability, and evaluation are needed when an agent connects forecasting to optimisation; and how should a financial institution price compute, data, governance, and human review when moving from a model demonstration to a decision workflow? Both remain programme-level evidence until course artefacts, speakers, or firm disclosures establish more.

September 4, 2026 — UK governance, M&A, and student-finance innovation routes

The University of Sheffield’s CRAFiC workshop reflection records a June 2026 workshop on AI in finance and accounting. Professor Fangming Xu’s keynote used machine learning to predict takeover premiums from pre-announcement accounting fundamentals, deal and adviser characteristics, and macro-financial indicators, with explainability treated as part of decision usefulness. The same programme covered mutual-fund ESG disclosure specificity, LLM classification of media-reported greenwashing accusations, and AI agents in speculative bubbles. These are distinct research objects—M&A pricing, disclosure quality, selective media signals, and agentic market behaviour—not one generic sentiment model. The page does not provide full data, code, out-of-sample results, or a manager connection.

Cambridge Judge’s AI in Financial Services for Public Authorities course creates a regulatory-side personnel and programme route. Its October 14 to November 11, 2026 cohort is a five-week course for regulators, supervisors, central banks, and ministries. The public faculty list names CCAF Academic Director Raghu Rau, former Barclays Global Investors Principal; Gary Ang, a former Monetary Authority of Singapore AI-risk and investment-risk lead; and Nico Lauridsen of the Florence School of Banking and Finance, among others. This helps mine supervisory language, AI-risk controls, and public-authority participants that may not appear in hedge-fund searches. It is not evidence of any manager’s system or a participant’s implementation.

Warwick’s CADE 2026 conference page adds a European conference archive with AI in Finance as one of its named themes. The June 15–17, 2026 blended event in Venice is structured for academics and practitioners, with a stated emphasis on early-career work, projects, and discussion, and lists Carsten Maple as general chair. Its prior-publication links create a paper-recovery route, but the page does not identify a specific finance paper, fund participant, recording, or production system; it should remain a discovery and archive candidate.

Duke Kunshan’s Digital Innovation Challenge Finance Track recap exposes a student and cross-campus implementation surface. The recap describes 75 teams from six countries, a finance track co-organised with NYU Shanghai, AWS-supported activities, Devpost-based first-stage evaluation, and a finance finalist project called “Causality” designed to verify ESG compliance and reduce greenwashing. The page names Professor Luyao Zhang as general adviser and lists academic and industry judges. This is useful for finding public prototypes, student talent, and applied problem definitions; it does not establish production quality, proprietary data rights, or investment use.

Together these routes extend the academic search into pre-announcement M&A modelling, ESG and greenwashing verification, agentic-market governance, supervisory risk, and student-built causal-finance systems. Each needs its own paper, data, split, licence, and reproducibility record before it can inform a manager-specific hypothesis.

September 4, 2026 — professor-led agentic screening and market-behaviour experiments

The academic layer now includes several concrete research objects that are useful for idea discovery while remaining separate from evidence about GMO, Acadian, Arrowstreet, or any other manager’s internal system.

Research object	Public professor or programme route	What is exposed	What remains unproven
Agentic portfolio screening	Columbia professor Agostino Capponi’s AI and Agentics in Finance page and the revised arXiv paper by Mehmet Caner, Agostino Capponi, Nathan Sun, and Jonathan Y. Tan	Separate LLM agents screen fundamentals and news sentiment; agents deliberate over buy/sell signals; high-dimensional precision-matrix estimation sets portfolio weights. The paper reports S&P 500 tests over short and medium horizons and a theory of “sensible screening.”	The abstract and paper do not establish a live manager implementation, proprietary data rights, durable out-of-sample performance, or autonomous portfolio authority. Reproduction requires point-in-time text and recommendations, prompt/model versions, universe construction, turnover, costs, and leakage controls.
LLM market-bubble behaviour	Oxford-Man professor Álvaro Cartea’s AI Bubbles with Large Language Models, also listed in the Frontiers of Factor Investing programme	A sequential bubble game studies speculative trades by AI agents, reasoning capacity, simplified beliefs, and conditioning on counterparty labels.	This is a market-design experiment, not evidence of a trading system or an investable signal. Model names, prompts, settings, replications, and transfer beyond the game require recovery.
M&A and governance prediction	The University of Bristol profile for Fangming Xu plus the Sheffield CRAFiC workshop reflection	Xu is listed as Professor of Finance and Director of Postgraduate Studies, with M&A, market anomalies, governance, and supply-chain interests. The Sheffield account describes ML takeover-premium prediction from pre-announcement accounting, deal/adviser, and macro-financial inputs.	The public pages do not provide the complete paper, code, data licence, point-in-time split, or manager connection. They do not establish live deployment or performance after costs.

The resulting research lanes are deliberately non-ranked: multi-agent screening followed by statistical portfolio construction; tests of whether LLM agents create predictable market-design distortions; and pre-announcement M&A and governance prediction. The papers, programme, faculty profile, and workshop report remain separate evidence classes. Academic results can generate replication candidates; they are not disclosures of a tracked firm’s model, data permissions, portfolio authority, or investment outcome.

September 4, 2026 — Asia-Pacific finance programmes and faculty routes

The regional academic search adds four non-ranked routes for idea and personnel discovery. They are not evidence that any tracked manager uses the methods described.

Route	Public evidence	Research or talent signal	Boundary
Singapore Management University quantitative finance	Quantitative Finance Research Workshop archive	The July 13–14, 2026 workshop is framed around portfolio management, asset pricing, risk management, and academic–industry collaboration. The archive also preserves older, dated links to Maybank’s machine-learning/AI development leadership and Salmon Global Fund portfolio management.	The older industry entries do not establish 2026 attendance, a shared project, or a production system.
University of Auckland Financial Machine Learning	FINANCE 710 course record	The 2026 course covers supervised and unsupervised ML for finance, financial-domain differences, dataset critique, computational requirements, and performance limitations.	A course record does not disclose student work, employer sponsorship, datasets, or investment deployment.
University of Technology Sydney applied AI for finance	Graduate Certificate in Applied Artificial Intelligence for Finance	Four subjects cover sustainable finance, investment and risk, compliance/anomaly/fraud detection, and digital/decentralised markets, with stated hands-on cases and guest lectures.	These are programme descriptions, not evidence of a named employer’s stack, model ownership, data access, or outcomes.
Hang Seng University of Hong Kong data-science/finance symposium	Economics and Finance Session	The July 2026 session names Boston College’s Zhijie Xiao, Monash’s Dan Zhu, and Melbourne’s Ping Chen, with topics spanning financial time-series dynamics, mixed-frequency VARs, and neural-network catastrophe-bond/pandemic-risk design.	The event page does not provide full papers, code, licences, recordings, or hedge-fund connections.

These routes widen the academic watchlist beyond text classification into ML model validation, time-series and macro-finance, insurance-linked risk, sustainable finance, and the academic–industry personnel network. The next recovery targets are dated programmes, recordings, assignments, faculty papers, code, and any authorized guest speaker or employer disclosures.

September 4, 2026 — European financial-NLP, personnel, and applied-course routes

Four additional routes are now tracked. They expose datasets, tasks, personnel, or employer-defined projects, while remaining separate from evidence about any manager’s internal systems.

Route	Public evidence	Useful signal	Boundary
Lancaster Financial Narrative Processing	FNP workshop series	The current FNP route advertises a Greek financial-report RAG task over ten years of Athens Stock Exchange annual reports, a Financial Causality Detection task, multilingual analysis, social-media discourse, misreporting, and negative-result reporting.	Shared tasks and datasets are research infrastructure, not evidence of a fund’s deployment, licence, or performance.
Manchester and Alan Turing personnel route	Eghbal Rahimikia profile	The profile connects financial ML/NLP research, Manchester teaching, Alan Turing fellowship, a public Tehran–Iran University of Science and Technology–Alliance Manchester academic path, and a dated Hull Tactical Asset Allocation consultancy role.	The profile does not disclose client work, data, current reporting lines, or a production model.
Birmingham employer-defined student projects	AI Inside FinTech event report	An April 2026 event involved Lloyds, Nationwide, FinTech West, and others; the report says twelve industry-proposed MSc projects covered ESG compliance monitoring, model risk, and AI workforce strategy.	Event participation and project topics do not establish deployment, data access, or validated results.
Southampton Banking and AI module	2026–27 module record	The syllabus covers predictive and generative AI in banking and central banking; the teaching record says students receive Python scripts and datasets for practice and can develop dissertation topics.	The public page does not identify the datasets, licences, student outputs, or employer implementations.

The resulting recovery queue is more operational: capture shared-task datasets and evaluation rules; follow named researchers through papers and public affiliations; recover employer-defined dissertation artefacts; and request or locate course scripts, datasets, and recordings. None of these routes supports a ranking or a manager-specific deployment claim without additional evidence.

September 4, 2026 — Latin American finance-AI faculty and applied-programme routes

Two additional Latin American routes add dated sponsor, professor, language, and implementation evidence without implying manager deployment.

Route	Public evidence	Signal	Boundary
Insper / Bradesco Asset historical event	Portuguese-language event page	The 2023 CeFiM event was sponsored by Bradesco Asset and named finance/data-science faculty from Princeton, Illinois, London Business School, and Insper, plus Fernando Caio Galdi, then described as Bradesco Asset’s Strategy and Innovation head.	This is a historical event-network record, not evidence of a Bradesco production model, proprietary data, or performance.
Universidad Panamericana applied finance programme	Mexico City programme page	The October 2026 course publishes a stack of Excel, Google Colab, ChatGPT, Python-assisted scripting, financial-news NLP, and real financial/market datasets.	The page does not identify dataset licences, instructors, student outputs, or employer deployment.

These routes add Portuguese- and Spanish-language recovery paths: obtain the Insper recording/slides and dated biographies, then recover the Panamericana syllabus, instructors, data provenance, exercises, and any authorized classroom artefacts.

September 4, 2026 — finance professors and programmes as an idea-discovery layer

The academic search should remain part of the coverage system because professors and finance programmes often publish the method, benchmark, syllabus, or research question before an investment firm describes an internal implementation. The University of Connecticut 11th Annual Finance Conference was scheduled for May 12, 2026 and identifies Xavier Giroud as keynote speaker. Its official description places alpha generation and persistence, investor cognition, and the impact of AI in finance in the same discussant-driven research programme. It also links current finance-department faculty, including Yiming Qian, Hang Bai, and Yao Deng. This is a dated academic route, not evidence of a manager’s model or a conference recording that has been reviewed.

The resulting idea map is deliberately non-ranked. Current public academic routes cover agentic screening with explicit portfolio construction; structured earnings-announcement representations; multilingual financial retrieval and causal extraction; nonlinear and high-dimensional asset pricing; pre-announcement M&A and governance prediction; market-impact and liquidation-horizon estimation; mixed-frequency macro-finance; insurance-linked risk; and experiments on how AI agents may alter market behaviour. The associated programmes also expose practical implementation surfaces: Python and Jupyter assignments, WRDS or other finance datasets, annotation tasks, model evaluation, RAG and reranking, agent workflows, governance, and industry-defined dissertation projects. These are research leads and talent-network routes. They do not show that a tracked fund uses a method, has rights to a dataset, or grants an agent portfolio authority.

For this project, each professor or programme should be tracked as its own evidence object: dated faculty profile, paper or preprint, course or event page, code/data licence, recording, and any separately verified employer connection. Before a research lead becomes an investable hypothesis, require point-in-time data, leakage controls, baseline comparisons, turnover and cost assumptions, model/version logs, and an independent out-of-sample test.

September 4, 2026 — title-blind university media and research routes

The coverage queue now includes academic seminars and ordinary finance media because relevant AI evidence does not always carry a hedge-fund label. A Sasin School of Management seminar featured University at Buffalo finance professor Sahn-Wook Huh on a study of ChatGPT’s ability to interpret firm disclosures. The public abstract describes a disclosure- interpretation metric and reports market outcomes involving informed trading, adverse-selection costs, volatility, order imbalances, and industry differences. This is an academic replication lead, not evidence of an investment firm’s use of the metric.

The MIT Industrial Liaison Program’s Money, Markets, and Machine Intelligence symposium adds an access-gap route. Its June 2026 overview combines market design, digital currencies, payment infrastructure, continuous trading, AI-enabled finance, automation, coordination, autonomy, and machine-to-machine economic networks. The public page does not expose the full roster or recording, so it should be tracked as a recovery target rather than mined for claims beyond the published overview.

The Knowledge at Wharton episode on AI stocks, oil prices, and the Fed shows why generic finance-outlook programmes belong in the search set. The May 2026 episode identifies Wharton emeritus finance professor Jeremy Siegel and links the same episode across Acast, YouTube, Spotify, Apple Podcasts, and RSS. It is useful for capturing how academic and practitioner audiences discuss AI and markets, but it does not disclose a hedge-fund system, proprietary model, or manager relationship.

The practical rule is now explicit: use professors and finance programmes as seed nodes for papers, recordings, guest appearances, assignments, and personnel trails; use generic finance podcasts as title-blind media sources; and preserve the audio or authorized transcript before extracting claims. Academic research, classroom material, media commentary, and investment-firm evidence remain separate evidence classes.

September 4, 2026 — additional university finance-media routes

The latest academic-media sweep adds a new Wharton macro-finance episode, “The Hidden Financial Risks of the AI Boom”. The September 4, 2026 page identifies Joao Gomes, Wharton finance professor and senior vice dean, and frames the discussion around AI financing, private credit, financial stability, and the need for better models. The page links Acast, Apple, Spotify, YouTube, and RSS. This is a useful macro-risk source, but it does not disclose a hedge-fund system or proprietary signal.

The AACSB Pulse episode “Why Finance Leaders Must Lean Into AI” adds a timestamped transcript and a business-school curriculum route. The March 2026 conversation features Meta finance leader Jason Pikoos and AACSB host Eileen McAuliffe on changing finance roles and what schools should teach. It is organizational and talent evidence, not evidence of Meta’s model stack or any investment-firm deployment.

The University of Iowa Tippie Leads episode “Can AI Beat the Market?” adds a separate professor-led media record for Ashish Tiwari, whose faculty research route covers flexible hedge-fund benchmarks and nonlinear, time-varying factor exposure. The October 2025 page links Apple, Spotify, and YouTube. It should be preserved as media evidence rather than merged with the faculty publication record.

September 4, 2026 — transcript-bearing professor media with implementation boundaries

The Wharton Future of Finance episode with Joao Gomes and Itay Goldstein includes a public transcript and cross-platform links. The February 2026 discussion describes reinforcement-learning agents optimizing trading strategies, interactions among agents, possible reductions in competition, financial fragility, prediction-market manipulation, and regulatory response. This is a professor discussion of possible uses and risks, not evidence of a Wharton or manager-operated trading system.

The Suffolk University On-Ramp transcript provides timestamped academic commentary on information gathering, investment decision support, model variation, annual-report extraction difficulty, opacity, bias, and transparency. It is valuable because it records operational limits as well as possible uses. The interview date and recording remain to be recovered; it does not establish hedge-fund deployment.

This reinforces the media rule for the project: preserve transcripts and timecodes, separate what a professor says AI could do from what a firm says it does, and keep academic, curriculum, media, and investment-firm evidence in separate classes.

September 4, 2026 — topic-indexed conference artefacts and Asian investment research

The Stanford Computational Market Design Center’s “Market Design in the Age of AI” was scheduled for February 27, 2026. Its public description connects algorithm design, economics, machine learning, operations research, market platforms, AI analysis, and regulation. The page now redirects to login, so the roster, agenda, papers, and recordings remain recovery targets rather than evidence about any participant.

The Morgan State National HBCU Blockchain, FinTech, and AI Conference call records a November 8–10, 2026 event in Nashville under the theme “Engineering Trust: Scalable, Intelligent, and Secure Financial Systems with Blockchain and AI.” Its scope includes deterministic and auditable AI, smart-contract analysis, ML for trading, credit, fraud and AML, AI-enabled compliance, robustness, privacy, distributed systems, and workforce/curriculum development. This is a regional academic and talent route; a call for papers does not establish accepted work, recordings, production systems, or fund adoption.

The UMD/SMU/UBS Quant Investment Forum programme and materials provide a deeper research-artifact route. The June 2026 Singapore forum links papers and slides on LLM-guided hypothesis discovery, AI errors and disagreement, knowledge-informed deep learning, fund-firm semantic alignment, active ML trading, text-managed portfolios, foreign bias in AI predictions, and 30 years of daily hedge-fund trades. The linked paper describes a human-designed symbolic environment using 66 Compustat variables and 24 operators, seven proposal–test–revision generations, and a staged validation process. These are paper-reported results, not independent performance verification or evidence of UBS deployment. The relevant design clue is the separation of the fixed evaluation laboratory from the agent’s iterative search loop.

The research queue should therefore hash and preserve the linked PDFs and slides, find code and data releases, and keep each paper separate from the conference host and from any manager-specific claim.

September 4, 2026 — university finance programmes as model and personnel routes

The NBER Summer Institute Asset Pricing programme adds a separate academic route for “Assessing the Benefits of Optimized Agentic AI Systems for Asset Pricing.” The programme names Andrea L. Eisfeldt and Sydney C. Ludvigson as organizers. The related Stanford programme description presents a real-time, out-of-sample earnings-announcement benchmark that extracts structured signals from announcement text while addressing look-ahead bias and market reflexivity. That exposes a research design to follow—point-in-time inputs, event-clock discipline, and evaluation under possible adoption effects—but it is not evidence of a fund’s implementation or independently verified trading performance.

The University of Reading ICMA Centre’s Data Science meets Finance conference provides a United Kingdom professor and programme route. Its March 3, 2026 page names James W. Taylor, Blanka N. Horvath, Ilias Chronopoulos, Marcin T. Kacperczyk, Shixuan Wang, Michael Clements, Emese Lazar, and Xiaochun Meng, with published topics including high-dimensional penalised least squares, machine-learning market timing, extreme-heat effects on stock performance, and functional-time-series change-point detection. The page is evidence of academic topics and named affiliations only; it does not establish a hedge-fund model, investable rules, code, data rights, recording, or post-cost results.

The practical addition is a clearer professor-to-programme-to-artifact route: use finance departments and quantitative-finance programmes to locate papers, code, recordings, students, and later industry affiliations, while retaining academic work as its own evidence class rather than treating it as a manager disclosure.

September 4, 2026 — academic AI leadership and student-built quant artefacts

The GW Investment Institute episode page on Patrick Hall identifies Hall as a George Washington University School of Business professor and Chief AI Officer. The associated public recording adds operational detail that is useful as an academic-enterprise comparison point: Hall describes building visibility into existing AI activity, an inventory and awards process, an AI forum, clearer rules for permitted use, and work with coding agents including Claude Code and Codex to prototype solutions. The captions are automatic, so these are bounded paraphrases rather than quotations. This is evidence about a university AI-governance and enablement programme, not about a hedge fund, trading model, or investment performance.

GW’s Spring 2026 Quant Fund Pitch Day archive provides a second kind of professor-and-programme signal: visible model-development artefacts. The page says student teams designed, tested, and refined predictive models in Python. Public descriptions include momentum and volatility rules; value, quality, and size factors; a data-center-infrastructure overlay; R&D productivity and cash-flow measures; ridge regression with engineered macro features and regime classification; LightGBM with trend, volatility-targeting, and regime features; and a two-stage workflow in which Claude.ai evaluates the significance and quality of news before a second model considers growth, valuation, and market performance. The page also names Rodney Lake, Ethan Baron, Oscar Pulido, and Greg Wong in the surrounding programme. The associated wrap-up recording, team reflection, and Investment Advice reflection make the student workflow discoverable, but their automatic captions should not be treated as a reliable personnel register. These are educational model exercises and public programme descriptions, not evidence of a manager’s live implementation, data rights, or post-cost results.

Together, these sources justify a separate professor-to-programme-to-artifact lane in the research system. It can reveal model ideas, governance patterns, tool names, researchers, and later alumni links. It must remain separate from firm disclosure until a tracked manager or an independently documented paper supplies that connection.

September 4, 2026 — finance curricula as industry and model-discovery routes

The Columbia Business School Quantitative Investing course is a useful bridge between faculty, curriculum, and a named manager. The Fall 2026 course is listed under Columbia’s Finance division and is co-taught by Kent Daniel and Giuseppe Paleologo of Balyasny Asset Management. Its published scope includes factor construction, portfolio optimization, risk modelling, backtesting, transaction costs, attribution, dynamic portfolio optimization, and work with CRSP, Compustat, and TAQ using Python and SQL. It also says that Claude Code, Cursor, or Copilot will be used in assignments and the final project. This establishes a public teaching and personnel connection, not Balyasny’s production implementation or the contents of any guest discussion.

The NUS Risk Management Institute’s HS2301 course is listed for AY 2026/27 Semester 1. It combines financial markets, accounting, statistics, data analytics, risk, fintech, and AI—including generative AI—with model, governance, and ethical risks. The course is part of NUS’s banking-and-finance track, has no prerequisite, and uses a group project alongside quizzes. This is evidence of curriculum design in Singapore, not a claim about a manager’s AI stack or strategy.

The City University of Hong Kong EF5560 catalogue entry describes a 2026/27 course in Fintech and AI in Finance covering machine-learning and AI-powered quantitative investment strategies, with a stated practical orientation. The public search result exposed the course description, but the direct page returned a 403 response during this pass; it is therefore a partial-recovery record pending an authorized page or PDF capture. The catalogue result also says generative-AI tools are not allowed in the final examination, a useful governance clue but not evidence of institution-wide policy.

These curriculum routes broaden discovery beyond firm-name searches. Course pages can expose the methods, datasets, software, instructors, guest practitioners, and student artifacts that later connect to a firm. Each connection still needs independent identity and implementation evidence before being promoted as a manager disclosure.

Korea: computational finance and a dedicated Finance & AI degree sequence

Seoul National University’s 2026 AI and Computational Finance course listing connects deep learning, reinforcement learning, and natural-language processing with high-frequency trading, algorithmic trading, market microstructure, high-frequency data, limit-order books, and optimal execution. It is listed under the College of Natural Sciences’ Department of Statistics. This is a particularly useful curriculum route for finding professors, lecture materials, and student projects around execution and market structure; the listing does not identify a manager, dataset, or live system.

The Hankuk University of Foreign Studies Finance & AI curriculum publishes a more complete degree architecture: Python, data structures and SQL, data mining, investment, machine-learning practice, financial time series, deep learning, NLP, prompt engineering, a finance-and-AI capstone, Transformer-based advanced financial time-series analysis, reinforcement learning, LLMs, portfolio optimization, and a model for financial-product recommendation. The page also includes AI ethics and philosophy. This is evidence of curriculum design and a talent pipeline in Korea, not evidence of any fund’s internal model inventory, permissions, or performance.

September 4, 2026 — seminar archives that expose research ideas before deployment

The Tsinghua SEM finance seminar notice records a June 4, 2026 talk by Washington University professor Zhou Guofu titled “Generative AI, Episodic Factors, and Causal Inference in Asset Pricing.” The notice confirms the speaker, host department, language, date, and research topic, but does not provide the paper, slides, recording, data, or a manager connection. It is therefore a research-idea lead, not evidence of an implemented strategy.

The AMLEDS past-seminar archive adds a useful title-blind index for academic finance and economics media. Its 2026 list names Asaf Manela on measuring and mitigating LLM look-ahead bias, Jing Cynthia Wu on an LLM survey framework, Tatevik Sekhposyan on ChatMacro inflation forecasts, Robert Richmond on asset embeddings, Pietro Bini on LLM biases and corrections, and Scott Cunningham on AI agents for research workers. The archive links some slides and papers, but the recent entries do not expose a complete set of replay URLs on the page. Those items are discovery leads until the individual artifacts are captured. The archive also provides a route to earlier work on text analysis, machine-learning portfolio performance, high-frequency trading, and macroeconomic forecasting.

These seminar indexes are valuable because they surface methods and researchers without requiring a hedge-fund keyword in the title. They remain academic evidence: a named seminar or paper does not establish fund adoption, proprietary data, trading authority, or post-cost results.

September 4, 2026 — recoverable research artefacts behind the seminar trail

The Federal Reserve Bank of San Francisco’s ChatMacro paper turns an AMLEDS title into a reproducible methodological warning. Alam, Boyle, Li, and Sekhposyan compare real-time ChatGPT inflation forecasts with pseudo-out-of-sample tests and report that the real-time forecasts are largely inaccurate and stale even when the pseudo-out-of-sample results look comparable to conventional benchmarks. The paper is macroeconomic rather than a hedge-fund study, but the point generalizes directly to financial research: a backtest using a model with post-period information can materially misstate the value of the signal. This is paper evidence, not a claim about any manager’s forecasting process.

The revised Asset and Investor Embeddings paper by Xavier Gabaix, Ralph S. J. Koijen, Robert J. Richmond, and Motohiro Yogo provides a different research direction. It learns representations from institutional portfolio holdings, then uses asset embeddings to study firm characteristics and investor embeddings to study investor similarity, performance measurement, and crowded trades. The authors also use language models to interpret firm- and investor-level text. This creates a public bridge between holdings data, representation learning, and portfolio research. It remains an academic paper; it does not disclose a fund’s private holdings, production model, or performance.

Public reporting on Scott Cunningham’s Federal Reserve demonstration and the more detailed account of the talk describes an agent orchestrating OpenAI’s batch API to classify 305,000 congressional speeches in about 2.6 hours at a reported cost of $11, broadly reproducing a prior immigration-sentiment study. The reports also describe a systematic disagreement pattern: the model compressed many disagreements toward “neutral.” That is a public research workflow and a warning about measurement error, not evidence of a trading agent or a verified investment result.

Two public AMLEDS LinkedIn announcements and the January look-ahead-bias announcement also expose the speaker, moderator, date, paper, and method vocabulary behind the archive. They are social discovery evidence, not substitute citations for the underlying papers or recordings.

The practical rule is now stronger: follow a seminar title to the paper, code, slides, recording, and social announcement; preserve the artifact’s clock and evaluation design; then search the named researchers and datasets for verified industry links. Do not merge an academic method into a firm profile without that separate link.

September 4, 2026 — professor-led ideas and finance-programme routes

The academic layer adds useful research and talent routes to the firm disclosures. It does not rank managers or establish that a hedge fund uses any of these methods. The relevant distinction is between a paper, a public artifact, a course implementation, a person’s verified employment, and a firm’s own production disclosure.

Selected idea families

Temporal integrity for financial language models. The Chronologically Consistent Large Language Models paper by Songrun He, Linying Lv, Asaf Manela, and Jimmy Wu describes ChronoBERT and ChronoGPT models whose training data is restricted to information available before each historical date. Its asset-pricing application uses financial news to predict next-day stock returns. The public model card and slides make the design more inspectable, including fixed knowledge cutoffs and released instruction-tuning data. The reported portfolio results remain paper claims requiring independent replication, point-in-time data, costs, and leakage checks.

Holdings embeddings. The Asset and Investor Embeddings paper by Xavier Gabaix, Ralph S. J. Koijen, Robert J. Richmond, and Motohiro Yogo turns portfolio holdings into learned representations of securities and investors, with applications to similarity, performance measurement, crowded trades, and text interpretation. This is a different modality from news sentiment: the central input is ownership structure. A reproduction would need point-in-time holdings, a survivorship-aware security map, portfolio rules, turnover, and a test against conventional characteristics.

Theory-guided ML and uncertainty. Hui Chen’s MIT research page describes economics-informed ML, uncertainty quantification, interpretability, and model misspecification as a connected agenda. It includes conditional-probability “inner confidence” for LLM news predictions, interpretation of LLM-generated economic forecasts, and synthetic-structural-model pretraining followed by empirical fine-tuning for option pricing. These mechanisms suggest research questions around confidence-conditioned signals, model drift, and structural regularisation; they do not establish a fund implementation.

Voice in earnings calls. Boston College’s research summary describes a deep-learning measure of vocal delivery quality in earnings-call audio and reports associations with news, market reaction, analysts, and media. A University of Chicago thesis describes speech-emotion recognition, transfer learning, and sentence-level aggregation. Together with the Duke summary of earlier vocal-cue work, these are academic audio routes—not evidence of a manager’s live signal, permissions, or post-cost returns.

Clinical-trial outcomes. The Nature Health CTO benchmark from University of Illinois Urbana-Champaign and Keiji AI covers approximately 125,000 drug and biologics trials, combining publication interpretation, phase tracking, news, sponsor prices, and trial metrics. The authors describe LLM-assisted outcome inference, cross-phase linking, FDA Orange Book matching, and manual review, and disclose the Keiji AI relationship and its non-involvement in the study’s design and analysis. The earlier HINT benchmark adds drug, disease, eligibility, pharmacokinetic, and historical-trial inputs. The investable question is whether point-in-time phase or approval forecasts can be mapped to sponsor exposure after event timing, licensing, leakage, and price-reaction controls—not whether a benchmark automatically produces a trade.

Finance programmes as implementation and talent maps

The University of Chicago Financial Mathematics 2026–27 concentration lists ML for finance, high-frequency data, generative and agentic AI, reinforcement and deep learning, multivariate statistics, and optimisation. Stanford CME241 places reinforcement learning inside stochastic control for portfolios, derivatives, hedging, and order-book trading. Illinois’ finance catalogue connects ML to options, portfolios, algorithmic trading, earnings responses, SEC filings, sentiment, discount rates, and risk. Queen Mary’s Finance and Machine Learning MSc adds LLMs and textual analysis to asset pricing, trading, portfolio construction, and a supervised dissertation or research project. Yale’s SoFiE programme exposes a broader research vocabulary spanning high-dimensional prediction, NLP/LLMs, factor pricing, alternative data, CNNs, transformer asset-pricing models, and limits of financial ML. The pages reveal what methods and project structures are being taught; they do not establish employer partnerships, student deployment, or investment performance.

This route improves the discovery process: course titles, syllabi, assignments, guest speakers, theses, supervisors, and alumni can reveal methods before a firm publishes a role or technical disclosure. Each resulting firm connection still requires its own public evidence.

Additional university and benchmark routes

EPFL FIN-407, taught by Semyon Malamud, explicitly links portfolio optimisation, return prediction, text embeddings, sentiment, LLMs, transformers, factor portfolios, coding projects, and model evaluation in the 2026–27 Financial Engineering curriculum. Malamud’s EPFL profile also lists doctoral FIN-622, covering high-dimensional regressions, neural networks, complexity corrections, factor models, and double-debiased ML. SUSS FIN525 adds a Singapore route covering annual reports, news, white papers, topic modelling, neural networks, risk prediction, and ML portfolio analysis; it names lecturer Wang Zhiyuan and records NUS and NTU training. These pages are talent and method maps, not evidence of employer use.

Rotman RSM2328H names Jun Yuan as instructor for a Fall 2026 course spanning supervised, unsupervised, reinforcement learning, NLP, GenAI, large datasets, Python, and hands-on finance use cases. PolyU’s Financial Technology and Artificial Intelligence programme combines software engineering, systems security, AI, ML, crypto, e-finance, and big data, with required work-integrated education and named programme leaders. Neither page identifies a production system, a specific employer’s model, or investment results.

The Portfolio Optimization Benchmark Framework by Hanyong Cho and Jang Ho Kim proposes mathematically explicit optimisation questions with unique solutions, varied objectives, assets, and constraints for testing LLM quantitative reasoning. It is a benchmark rather than a portfolio strategy. A related Stanford macro-finance session describes RAG over filings, patents, earnings calls, peer disclosures, and news to estimate firm-level marginal projects, alongside ML for UK price-change data under structural breaks. These are research designs that broaden the search beyond sentiment; they do not establish a hedge-fund implementation.

September 4, 2026 — student investment programmes and a reproducible GPT-4 study

The GW Investment Institute episode page identifies Rodney Lake as Institute director and describes a February 2026 discussion of AI in markets, investment-analysis tools, and AI as a productivity multiplier. The same page says GWII students manage more than $10 million of university endowment assets across four student investment funds and use FactSet, Bloomberg, and PitchBook. This is useful evidence of a university investment-programme and talent route. It is not evidence of a production AI trading system, model ownership, performance, or hedge-fund deployment; the transcript and technical details remain recovery targets.

The Canisius University report and linked Modern Finance paper add Marc LoGrasso’s GPT-4 stock-selection study to the academic artifact graph. The journal reports a 1985–2021 retrospective window, ten annual selections, two-year holding periods, and approximately 1% average monthly alpha, while also reporting that year-level portfolio alphas were positive and significant in only about one out of four years. Those are published paper results, not independent replication, live deployment, or durable post-cost alpha. The relevant research question is whether prompts, model versions, point-in-time controls, portfolio construction, and costs change the result.

GWII’s public investment-programme operating model

The GWII BMPB framework page publishes a repeatable equity-analysis rubric: business, management, price/valuation, and balance sheet, each scored from 1 to 10 and weighted equally at 25 percent. The page says analysts refine the composite repeatedly and that the framework has been foundational to GWII since 2005. This is a human scoring methodology, not evidence of an AI model or automated trading rule.

The Q2 2026 quarterly report reports six guest speakers across three finance classes, 21 stock pitches, 12 startup presentations, and eight quantitative-investing models. It also reports more than $45,000 raised for FactSet access and a 12.3 percent quarterly student-fund return, 290 basis points below the S&P 500. These are first-party programme figures, not an independent performance audit or an AI-attributed result.

The student endowment case study describes a $30,000 semiconductor-position proposal requiring 80 percent committee approval and receiving 89 percent, alongside a reported $11.7 million under student management across four real student-run funds. The page does not identify the student or security and does not expose the model, data, execution, or trade outcome. These pages add a useful education-to-practice route for finding decision rubrics, model-development exercises, committee controls, finance-course guests, and later alumni affiliations.

Recovered GW video evidence

The GW page embeds the official YouTube recording. Its automatic English captions provide timestamped, bounded evidence that Rodney Lake described GWII’s use of Gemini, ChatGPT, Grok, and Claude alongside FactSet and Bloomberg for student investment analysis (02:49–03:21); applying a GWII business-management, price/valuation, and balance-sheet framework as a weighted 1-to-10 score and iterating with a model to compare companies (04:04–04:52); checking model work and avoiding uncredited claims (05:09–05:26); and using Google’s ecosystem, Colab, and NotebookLM for coding and content creation (06:03–06:23). The captions are automatic, so these are paraphrases rather than quotations. They establish an educational workflow description, not a proprietary model, production trading permission, performance result, or hedge-fund deployment.

September 4, 2026 — additional professor and finance-program routes

The academic search also recovered several implementation-oriented routes that are useful for idea generation and talent mapping. They should not be read as evidence that any tracked manager uses the methods.

The University of Sydney’s FINC6028 Artificial Intelligence in Finance unit is a postgraduate finance course whose stated learning outcomes include comparing AI models for financial applications, applying them to portfolio optimisation, default prediction, factor strategies, option pricing, fraud detection, and risk management, and evaluating the resulting systems. The page specifies assignments and group projects, but does not expose the syllabus, student code, employer partners, or investment results. It is therefore a curriculum and Australian talent-pipeline signal only.

Temple’s 2026–27 Financial and Quantitative Analysis MS lists separate courses in machine learning in finance, generative AI for finance, and AI in portfolio management alongside quantitative portfolios, econometrics, volatility modelling, and risk. This is more specific than a generic “AI” label because it places GenAI and portfolio-management coursework inside a quantitative finance degree. The catalogue does not identify course projects, model providers, students, or live investment use.

Penn State’s FIN 465 Data Science and Artificial Intelligence in Finance describes Python, SQL, financial datasets, data handling, visualisation, statistical analysis, machine learning, LLMs, and prompt engineering. The same catalogue records the Intrieri Family Student Managed Fund as an actual student fund and separately describes the Nittany Lion Fund lead-manager practicum. That adjacency is a useful route for checking whether student projects, fund governance, and later alumni roles produce public artifacts; it does not establish that the AI course feeds either fund or that any model has trading authority.

Finally, Deep Learning to Trade: An Experimental Analysis of AI Trading and Market Outcomes by I. Gufler, F. Sangiorgi, and E. Tarantino models deep-reinforcement-learning traders inside a calibrated market with endogenous demand and price impact. The experiment uses continuous-state/action DDPG agents, a rational full-information benchmark, and ten U.S. equities; the authors report that interactions among many AI traders create learning noise and can reduce performance relative to the benchmark. This is a particularly useful caution for any backtest that treats agents as isolated, price-taking learners. It remains a dated academic simulation, not evidence of a fund’s live system, model choice, permissions, or realised P&L.

Saint Louis University’s 2026–27 finance catalogue adds another implementation-oriented programme route. Its graduate sequence lists financial analytics with alternative data, markets and algorithmic trading, and Artificial Intelligence and Machine Learning in Finance. The AI course specifies tree regression, neural-network methods, textual analysis, Python, web scraping, and financial-data visualisation; the catalogue also describes an applied portfolio management course in which students manage an allocation from the university’s endowment. This is a curriculum and student-investment governance signal. The catalogue does not connect the AI course to the endowment allocation, identify a faculty-led model, or establish live performance.

September 4, 2026 — professor lineages and research hypotheses

The professor search adds several public research surfaces with more specific ideas than a generic “AI in finance” label. These remain academic or talent signals and are not evidence of a tracked firm’s production system.

Jimmy Wu’s Washington University profile says he is joining the University of Iowa as an assistant professor of finance in Fall 2026 and works on empirical asset pricing, behavioral finance, LLMs, and ML in finance. His job-market paper uses an LLM-derived news-perplexity measure to study whether cognitively demanding news prompts additional information acquisition and reduces market underreaction. His page also links the Chronologically Consistent Language Models work. This is a useful route for testing language-model features against EDGAR-search activity and event-time returns, while keeping the academic result separate from any employer claim.

Xin He’s USTC profile combines AI-for-finance, asset pricing, quantitative investment, and China-market research with unusually inspectable public artifacts. It lists a PhD in Management Sciences from City University of Hong Kong, a Shanghai Jiao Tong industrial-engineering degree, research grants, China and US equity data/code, a volatility dictionary, and an August 2026 cross-market corporate-bond/equity working paper. The page also advertises research-student hiring and links multiple 2026 finance-AI conferences. These are professor-authored research, data, and recruiting disclosures; they do not establish a fund connection or live use of the released materials.

Manuel Nunes’s University of Southampton profile is both a programme and research route. He is director of the MSc Financial Technology and AI, deputy director of the Centre for Digital Finance, and lists explainable AI, LLMs, and reinforcement learning for portfolio management. His publications include actor-critic reinforcement learning for bond portfolios, LSTM-based bond-yield forecasting, FinBERT/ARMAX-GARCH work, and a thesis on ML in fixed-income markets. The profile records a Southampton computer-science PhD in a multidisciplinary finance/electronics programme, an engineering PhD and MBA from Nottingham, and earlier degrees from Porto and Nova. It also says he previously worked as an equity-research analyst, portfolio manager, and head of fixed income. This is an academic-to-practitioner lineage, but not proof that a named fund uses the methods or that the reported research produces live returns.

Yan Ji’s HKUST profile adds a market-structure route. Ji lists an MIT economics PhD and current work on AI in finance, including “Financial Market Fragility in the Era of AI Planning” and “AI-Powered Trading, Algorithmic Collusion, and Price Efficiency,” both co-authored with Winston Dou and Itay Goldstein. The research questions concern how autonomous agents learn, coordinate, alter price efficiency, and create fragility—not simply whether an isolated model forecasts returns. The profile and CV expose the paper trail and lineage; they do not establish a commercial system, fund deployment, or performance.

September 4, 2026 — additional professor and finance-programme routes

The title-blind academic pass found three additional routes that are useful for idea discovery and talent mapping. They are academic, programme, or public-IP signals; none establishes a hedge fund’s production system, data rights, model permissions, trading authority, or performance.

Bryan Seegmiller’s Kellogg profile connects an MIT-trained finance professor’s research on technological change, AI-driven labour demand, firm profitability, and financial-sector frictions to an “AI Foundations for Managers — Finance” course. The course covers valuation, forecasting, automated trading, credit risk, fraud detection, generative-AI analysis of financial documents, governance, and organisational adoption, with hands-on exercises. This creates a route for testing AI exposure through labour and corporate information as well as the organisational controls around finance workflows. It is not evidence of a hedge-fund relationship or live investment use.

Sungkyunkwan University’s Hyung Jin Ko profile lists Financial AI, generative AI, and finance/business agents; a 2026 paper on sector-aware LLM reasoning for investment decisions; privacy-preserving optimal portfolio work; a generative-diffusion approach to price-chart images; and a Principal Investigator role on a knowledge-graph-and-LLM personalised-agent programme. It also lists a homomorphic-encryption portfolio patent and earlier RAG investment-decision work. These are concrete research and public-IP surfaces for testing sector conditioning, retrieval graphs, chart representations, and privacy constraints, not evidence of a manager’s deployment.

The 10th PKU–NUS Annual International Conference on Quantitative Finance and Economics supplies a China–Singapore conference route. The May 16–17, 2026 Shenzhen event is jointly organised by PKU HSBC Business School, NUS’s Risk Management Institute, and PKU’s Key Laboratory of Mathematical Economics and Quantitative Finance. Its programme spans algorithmic trading and market microstructure, computational finance, machine learning in economics and finance, portfolio selection, quantitative investment, and risk management, and explicitly invites academic and industry participation. Its organiser list provides named faculty and cross-institution search handles. The organiser page does not provide recordings, proprietary data, or evidence of live firm deployment.

The academic search also found two executive-education routes that expose finance agent workflows more directly than a generic programme title. The University of Zurich course announcement announces a November 2026 course led by Benjamin Wilding, Roland Schläfli, and Linus Graf. Participants are expected to build and present an agent for a finance use case, configure a personal assistant, and study limitations and human oversight. It is a future programme route, not evidence of participant projects, vendor use, or production deployment.

The Johns Hopkins University AI and Agentic AI in Finance programme publicly describes projects around earnings-call sentiment, PII-redacted bank-loan credit memos, portfolio-risk monitoring agents, KYC/AML workflows, and multi-agent “bank-in-a-box” systems. Its curriculum names RAG, embeddings, prompt chains, hybrid ML/GenAI, SHAP, tool use, agent memory, multi-agent delegation, model-risk governance, cost analysis, and build-versus-buy decisions. The page lists Ian McCulloh and Jim Kyung-Soo Liew as faculty figures and indicative mentors including Benito Lopez, Ashutosh Pandey, Mirata Hosseini, Omid Badretale, and Anuj Saini, with roles listed at Citi Global Wealth, BMO Capital Markets, and RBC Capital Markets. These are public programme and discovery claims; they do not establish that the named organisations provided data, endorsed a project, or deployed the described systems.

This adds an important discovery route: executive-education pages often reveal the documents, controls, tools, and human-review boundaries that degree titles omit. They should be mined for dated syllabi, project rubrics, guest recordings, mentor profiles, tool lists, and alumni outcomes, while marketing claims and indicative rosters remain leads until independently verified.

A second check of official personnel pages sharpens the evidence boundary. Johns Hopkins’ Ian McCulloh profile identifies him as Director of Partnerships for Engineering Executive and Professional Education, with oversight of AI and related portfolios, Carnegie Mellon computer-science training, prior Accenture federal-AI leadership, and courses in applied GenAI, human-in-the-loop AI, and evaluation. SoKat’s official team page identifies Jim Kyung-Soo Liew as founder and president, while Johns Hopkins’ JHBIC profile identifies him as an Associate Professor of Finance. His public LinkedIn profile describes work with boutique banks, RIAs, asset managers, and hedge funds. These pages corroborate public academic and AI-consulting roles; they do not name a fund client, disclose a model, or establish deployment.

Nanyang Business School’s July 29, 2026 professor video adds a Singapore academic and media route. The NTU page identifies Byoung-Hyoun Hwang as President’s Chair Professor in Finance and Head of the Division of Finance, and embeds a discussion of fiduciary accountability, human oversight, agentic investing, and the limits of fully automated decisions. It describes course-level use of code-generating agents under student supervision and reports collaborations with hedge funds without naming them or disclosing their systems. The article retains those as source-attributed claims and omits the page’s promotional comparisons and adoption language; the route does not establish any named fund deployment.

September 4, 2026 — family-office allocator and academic-network routes

The Citi Institute report AI in the Family Office adds a sector-level operating-model source. Based on interviews with family-office CIOs and principals, it separates current use cases into document and manager due diligence, meeting transcription, reporting, custodian/API data extraction, tax and alternative-asset intelligence, manager research, market research, and risk identification. Its maturity map moves from data extraction and repetitive-task automation through research assistants and investment copilots toward predictive analytics, agentic workflows, and complex analytical processes. The report also records privacy, accuracy, limited internal IT, and headcount constraints; it says human legal review and final investment responsibility remain important. It mentions general-purpose and enterprise LLM choices, but does not identify any family office, model, dataset, deployment, investment authority, or performance result.

The London Business School recap of its 2026 Family Office Conference adds a university-hosted allocator-network route. The Institute of Entrepreneurship and Private Capital says the February 24–25 event was private and non-commercial, with SFO Collaborative peer discussions covering AI, healthcare, and digital assets alongside governance, succession, and thematic investing. The public recap does not name participating families or expose technical sessions, so it is a discovery route for later public recordings, faculty pages, speaker posts, and follow-up material—not evidence of a named family office’s AI system.

These sources extend the professor-and-programme layer into a distinct allocator context. They support research questions about privacy-preserving manager due diligence, auditable document and meeting workflows, and the point at which human approval remains necessary. They do not support cross-firm capability comparisons. The capture note preserves the evidence classes and recovery queue.

September 4, 2026 — additional academic model routes

Two additional preprints add concrete research designs to the professor-and-programme map. Voronina, Romanko, Cao, Kwon, and Mendoza-Arriaga study LLM-based selection and weighting of 20 stocks per S&P 500 sector, followed by classical portfolio optimisation, across stable and volatile out-of-sample periods. The abstract reports temporal dependence and author-reported improvements from the hybrid workflow. The research question worth carrying forward is whether sector conditioning, explicit constraints, and point-in-time prompts reduce instability; the preprint does not establish persistence, costs, live use, or a firm connection.

Hounwanou and Gaba examine TimeGAN and variational autoencoders for synthetic S&P 500 daily-return data, evaluating statistical similarity, temporal structure, and downstream portfolio and risk tasks. The abstract reports different trade-offs between volatility-pattern fidelity and smoothing of extremes. This is a useful privacy-preserving benchmark design, but it is a preprint experiment rather than evidence that synthetic data can substitute for live data or that a manager uses either model.

These papers belong in the research-hypothesis layer: replicate with versioned prompts, point-in-time controls, transaction costs, rolling evaluation, tail-risk tests, and independent data splits. The academic capture note keeps paper claims separate from firm disclosures and personnel evidence.

Two further academic artifacts sharpen the research-agent and disqualification loop. The SSRN version of Can AI Do Financial Research? places an LLM inside a human-designed symbolic accounting language, automated validation layer, and fixed empirical pipeline. The agent proposes and revises asset-pricing formulas over successive generations using standardized feedback. That is a concrete architecture for auditable hypothesis discovery: constrain the grammar, freeze the test harness, retain proposal lineage, and apply multiple-testing and robustness checks. The paper’s reported results remain author claims; it does not establish an investment-manager deployment, and one public co-author identity is withheld here.

The American Economic Association / Journal of Economic Literature article Artificial Intelligence–Powered (Finance) Scholarship by Robert Novy-Marx and Mihail Velikov describes LLM-assisted generation of finance papers after mining more than 30,000 potential predictors and preparing templates for 95 signals that passed the authors’ criteria. Its central methodological warning is as important as its efficiency claim: an agent that sees results before generating explanations can industrialize HARKing. For any finance research agent, discovery, evaluation, and explanation should be separated with frozen information dates, registered selection rules, retained rejected hypotheses, and no outcome leakage. This is peer-reviewed academic evidence about research process design, not evidence of a named firm’s model, data rights, or investment performance.

September 4, 2026 — additional academic and practitioner conference routes

Two further title-blind conference surfaces expand the personnel and research map. The CFA Society New York “Beyond the Quant” panel is advertised for October 6, 2026 as a hybrid event hosted by the Data Science & Quantitative Investing Group. Its public description covers AI in research, portfolio construction, risk, trading, implementation lessons, and human judgment, and lists Serge Levin, Director of Volatility Strategies at Massar Capital. The page does not expose a recording, transcript, model, dataset, or firm process, so it remains a dated capture target rather than evidence about Massar Capital’s systems.

Shanghai Jiao Tong University’s “New Trends in Quantitative Finance” page dates an April 20–24, 2026 conference hosted by its Quantitative Finance Research Center and the China Academy of Financial Research. The advertised scope includes AI in financial modelling, stochastic control, systemic risk, mean-field games, and carbon-certificate markets. The organiser list spans NYU Shanghai, Shanghai Jiao Tong, ShanghaiTech, and University of Evry researchers, with one listed organiser’s identity omitted here. The public page does not provide a complete programme, recording, paper list, or firm-specific implementation; it is a regional academic and personnel discovery route only.

The capture note records event dates, speaker roles, recovery targets, and evidence boundaries.

September 4, 2026 — finance-programme signals and professor-led research routes

The professor and programme layer now adds several distinct signals that are useful for idea discovery without implying manager adoption. SDA Bocconi’s Artificial Intelligence applications in finance programme, scheduled for November 25–27, 2026 in Milan and directed by Marta Zava, separates generative, predictive, and agentic AI across six pillars: strategic clarity; productivity, risk, and governance; forecasting and risk decisions; model risk and explainability; autonomous systems; and institutional integration. Its target audience includes boards, CFOs, CROs, and digital- transformation leaders. This is a public governance and curriculum signal, not a named-fund system or result.

NUS Business School’s Enhancing Banking & Finance with AI course, scheduled for December 14–16, 2026 in Singapore, names Teo Ted as programme director and Robin Lee of HAWK and Xperientia as co-course instructor. The outline covers supervised and unsupervised learning, reinforcement learning, deep learning, NLP, LLMs, practical demonstrations, and a dedicated risk day covering adversarial AI, deepfakes, synthetic identities, privacy, bias, black-box risk, and regulation. The course is aimed at non-STEM executives; its contents do not establish a production system at NUS, HAWK, Xperientia, or any manager.

The University of Bergamo’s XXVII Workshop on Quantitative Finance was held March 30–April 1, 2026 and explicitly solicited work in financial applications of AI and ML alongside portfolio optimisation, computational finance, risk, econometrics, and practitioner research. The public page lists FactSet, Amases, and REVO Insurance among sponsors. That sponsor graph is a route for finding data and practitioner material; it is not evidence of a particular dataset, customer relationship, model, or investment result.

The NUS Institute for Mathematical Sciences Quantitative Finance 2026 programme, held June 1–19, 2026, combines stochastic control, systemic risk, and market frictions with machine learning for pricing, hedging, forecasting, and portfolio optimisation. Its summer-school roster includes Christa Cuchiero’s “Generative AI in Finance,” alongside optimal transport, carbon-emissions regulation, and mean-field games. The organising committee includes Daniel Bartl, Agostino Capponi, Ying Chen, Min Dai, Julian Sester, Nizar Touzi, and Marko Weber. This is a professor-and-method discovery route, not evidence of a covered firm’s deployment, data rights, agent permissions, or performance.

These programmes suggest bounded follow-up questions: whether agentic finance is being taught with explicit approval and model-risk gates; whether forecasts, portfolio construction, execution, and risk are evaluated separately; and whether vendor sponsors and course speakers expose papers, code, datasets, or hiring signals. The programme capture note keeps those questions separate from firm disclosures and academic performance claims.

September 4, 2026 — proprietary financial-data and research-media route

The title-blind media sweep also recovered a new adjacent finance-AI route. New Constructs’ company article on The Principal Podcast identifies CEO David Trainer as a guest discussing proprietary data, corporate-filing footnotes, financial-analysis technology, and the firm’s proprietary AI systems. The Spotify listing names the episode and lists a roughly 30-minute runtime; directory metadata dates the episode July 22, 2026, while the New Constructs article is dated September 3, 2026. That date discrepancy is preserved in the capture note.

New Constructs’ partnerships page also states that detailed ratings data is available through Quiver Quantitative’s standardized API, and the company’s media archive lists a July 30, 2026 partnership item. This creates a concrete data-provider follow-up: inspect field definitions, point-in-time availability, revision history, coverage, and access terms before any research use. The public pages do not disclose model families, weights, training corpus, data licences, customer identities, agent permissions, live portfolio authority, or independently audited performance. Product and partnership statements remain company or publisher claims, not evidence of a tracked hedge fund’s deployment.

The title-blind pass also recovered an Enterprising Investor episode with Irina Bevza, published September 1, 2026. The listing identifies Bevza, PhD, CFA, as Head of Quantitative Solutions at Fineco Asset Management in Dublin and frames agentic AI as a possible move from portfolio-manager assistance toward more autonomous portfolio decision-making. This is publisher metadata and role evidence only: the transcript, model identity, data rights, permissions, and any live system were not recovered. The capture note keeps it separate from stronger transcript-level claims.

September 4, 2026 — Apollo allocator media and professor-led idea routes

The title-blind media sweep recovered Apollo Academy’s The Allocation episode, “Can AI Earn It Back? The $2 Trillion Test and Other Forces Reshaping Markets and Portfolios.” The publisher names Tal Barak Harif, Managing Editor of Apollo Thematic Investing, as host and Rob Bittencourt, who leads Apollo Thematic Investing, as guest; the Apple Podcasts listing provides a separate distribution record and approximately 39-minute runtime. The page is dated September 2, 2026, while the audio says it was recorded July 15, 2026. That discrepancy is preserved rather than silently normalised.

The audio adds a useful allocator and financing lens, but not a hedge-fund model disclosure. Bittencourt describes a thematic-investing group that follows technology and geopolitical forces and collaborates across origination, underwriting, and risk management. At roughly 08:04–10:51, he discusses the shift from experimentation toward enterprise deployment, the need to show ROI, and possible cost controls including lower-cost or open-source models, per-employee token budgets, and prioritised access. Those are speaker-reported operating ideas, not evidence that Apollo uses each one internally.

At roughly 13:42–14:13, the discussion cites Apollo involvement in infrastructure financing involving Broadcom and Anthropic, and in financing involving SpaceX and xAI. At roughly 16:15–19:36, it describes “AI canaries” as software and business-services companies whose unit economics may show early effects from AI. These are financing and thematic-research statements. They do not establish a technical partnership, training-data arrangement, model family, automated portfolio authority, or validated forecasting signal. The permanent audio and machine-generated timestamped sidecars are catalogued in the capture note.

The professor and programme search also adds idea and talent surfaces that are useful precisely because they are more specific than a generic “AI in finance” label. Yinan Su’s Johns Hopkins page connects a University of Chicago Booth–Economics PhD and Tsinghua undergraduate degree to financial econometrics, ML, and AI research using firm characteristics, financial news, investor holdings, and trading activity. His listed work includes narrative factors from Wall Street Journal text, scale-specific predictive cycles, trading-volume prediction, and structural-break methods. These are research hypotheses and lineage evidence; the page does not establish a fund’s deployment, proprietary data rights, or live performance.

Nabi Omidvar’s University of Leeds profile identifies him as Associate Professor of Artificial Intelligence in Finance and a lead researcher in the Leeds Centre for Financial Technology and Innovation. It connects large-scale optimisation and variable-interaction analysis to high-dimensional/deep-learning training efficiency and financial-services uses such as credit scoring, and lists RMIT University training. This provides a concrete personnel and algorithm search route, not evidence of a manager relationship or production credit model.

IE Business School’s AI-Powered Finance programme is scheduled for November 16–19, 2026 in Madrid. Its public overview describes hands-on work with Datarails, Copilot, and ChatGPT across forecasting, modelling, FP&A, treasury, and financial planning. The Oxford Generative AI for Finance course similarly lists prompt engineering, portfolio management, risk assessment, fine-tuning, fraud detection, privacy, and regulation, with Konrad Kleinfeld and Claudia Otto named as tutors/guest lecturer. These programme pages reveal the vocabulary and tool surfaces to search in hiring and practitioner media; they do not show participant deployment or a covered fund’s internal stack.

The Open Source Quantitative Finance conference, scheduled for October 23–24, 2026 at the University of Illinois Chicago, adds an open-source discovery route covering AI, risk, econometrics, high-performance computing, market microstructure, portfolio management, and time series. Its committee links may yield code and personnel leads, but the event page does not establish attendance by any tracked manager or adoption of a named library.

The capture note keeps the Apollo speaker claims, academic research, and executive-education curricula in separate evidence classes. The next checks are to recover Apollo’s cited thematic research, professor-linked papers and code, programme syllabi, and any employer-controlled source before connecting an idea or person to a manager.

The MIT–IBM Computing Research Lab profile for Yada Zhu adds a separate finance-lab route. The page identifies Zhu as an IBM Research staff member in the Future of Computing – Finance Research team and says she leads researchers developing foundational AI and ML for investment decisions and risk management in dynamic markets. It names high-dimensional time series, heterogeneous data, large-scale graph learning, and statistical modelling, and identifies an MIT–IBM AI Lab project funded by Refinitiv and an unnamed “Mega bank.” The page does not date the project or identify the bank, so neither is resolved here. The listed publication routes around outlier impact, explainability, complex-heterogeneity contrastive learning, and sample-based NLP explanations are useful idea and talent leads, but they do not establish a bank’s production system, model weights, data rights, portfolio authority, or performance.

The capture note records this lab route alongside the Apollo episode and the other professor and programme sources, with academic, allocator, and manager evidence kept as separate classes.

Two further academic routes make the idea surface more concrete. The Masaryk University project record for Sentiment-Driven Volatility Forecasting describes a 2026 student-supported project by Yingke Zhu under Štefan Lyócsa. It combines market-microstructure features and sentiment indicators for short-horizon stock-index-futures volatility and return dynamics, using regularised regression, ensembles, and SHAP in a multi-sentiment fusion design. The stated extension is an interpretable volatility-prediction system for commodity and equity futures. The record supplies a research design, not results, data rights, execution assumptions, or a manager connection.

The Korea University Financial Technology Lab research page is a particularly dense paper-index route. Its stated interests include automated investment management, portfolio optimisation, asset allocation, financial data analysis, and LLMs in finance. The 2025–26 list includes LLM embeddings for firm similarity; LLM-derived volatility and correlation estimates for portfolio construction; deep financial planning; random-forest feature selection; and GAN-based anomaly detection for mean-variance optimisation. It also lists work on LLM investor-risk profiles, portfolio-optimisation benchmarks, and LLM time-series forecasting. These are academic research artefacts and candidate replication ideas, not evidence of a fund’s model, data, permissions, live authority, or performance.

The capture note records the source pages and the follow-up requirement to recover papers, code, data statements, and point-in-time evaluation details.

September 4, 2026 — finance-agent benchmarks and professor-led research routes

The academic layer adds a useful distinction between a finance model and a finance research system. FinanceHarness, submitted July 30, 2026, describes a harness that constructs the research environment and data, runs an agent loop over finance tools, and applies reward modelling. Its public repository exposes search, source reading, reference chaining, calculations, valuation, and risk tools. The linked FinanceGym benchmark uses 400 point-in-time questions and 2,464 expert-rubric items, grading what was knowable before a cutoff separately from what happened afterward. The site also states that the benchmark, data, and code are CC BY-NC 4.0. This is a reproducible research-agent route, not evidence of any tracked firm’s internal system; its license must be checked before any commercial use.

Two other public benchmarks sharpen the evaluation surface. Deep FinResearch Bench separates qualitative rigor, quantitative forecasting/valuation, and claim credibility/verifiability when comparing agent reports with professional reports. Finance Agent Benchmark uses nine SEC-filing task categories and 537 expert-authored questions, with an agent harness that can use web search and EDGAR; its Zenodo record and repository are additional recovery routes, although the dataset files are restricted. These benchmarks are useful for testing research workflow quality, not for ranking firms or inferring live trading capability.

The professor route also becomes more concrete when the current ICLR record is followed. Yuan Zhang’s page identifies an SUFE finance professor with Swiss Finance Institute at EPFL PhD training and current work on LLM-assisted alpha mining, evolutionary factor search, large/deep factor models, and executable option strategies. The current AlphaBench poster tests LLM factor generation, factor evaluation, and factor search. The linked EFS preprint adds evolutionary feedback, redundancy-aware allocation, random-matrix denoising, and regularised quadratic programming for sparse portfolio construction across US, Hong Kong, and mainland-China datasets. These are public academic methods and reported experiments; they do not establish a manager’s deployment, data rights, model permissions, execution authority, or performance.

Yale SOM’s 2026 Behavioral Finance Summer School report names Nicholas Barberis, James Choi, Kelly Shue, and Suproteem Sarkar among the faculty for a June 8–12 PhD programme with new AI material. The report describes AI photo analysis of personality and career outcomes, and translation of news into mathematical representations for studying how companies are perceived. Yale’s linked research explainer describes a 96,000-person MBA-photo/LinkedIn study and records an explicit warning against using facial analysis in hiring. It is a multimodal research and governance route, not a recommended personnel-screening technique or investment signal.

CEMFI’s AI for Economic Measurement, Forecasting, and Simulation course lists Stanford Digital Economy Lab research scientist Sophia Kazinnik as instructor for August 17–21, 2026. Its syllabus covers extracting stance, uncertainty, risk, guidance, and narrative shifts from text or audio/video; generating structured forecast objects and labels; and simulating forecasters, committees, depositors, and other institutions under controlled information sets. It explicitly includes model and prompt robustness, bias correction, leakage safeguards, RAG, fine-tuning, versioning, and reproducibility. Kazinnik’s biography also records prior Federal Reserve stress-test and natural-language-tooling work. This course exposes a research workflow and personnel route; it does not establish a hedge-fund deployment or private data access.

The academic capture note keeps these routes separate from manager disclosures. The follow-up is to recover benchmark schemas, code and license terms, paper supplements, notebooks, recordings, and data statements, then apply the same point-in-time and disqualification checks before connecting a person or method to a firm.

September 4, 2026 — macro, volatility, and agentic-market research routes

MACROCAST, a 2026 Brandeis International Business School preprint by Andrea Carriero, Davide Pettenuzzo, and Shubhranshu Shekhar, makes data vintage part of the forecasting problem. It pretrains on synthetic series and fine-tunes on vintage-specific ALFRED data to target both temporal contamination and revision bias. The record describes a genuine real-time FRED-MD evaluation and reports approximate compute and fine-tuning time. The design is a useful control for macro or rates research agents; the reported results still require the code, vintage snapshots, and evaluation scripts before independent use.

Duke’s record for Alessio Brini describes a comparison of nine zero-shot time-series foundation models with eight econometric specifications on VOLARE: 50 assets spanning equities, FX, and futures, three forecast horizons, and formal pairwise and multi-model tests. The record reports non-uniform gains, a narrow TTM-versus-Log-HAR result, and a recalibration finding that much of the short-horizon advantage reflects level and scale rather than volatility dynamics. It also reports an equal-weight TTM/Log-HAR combination entering the Model Confidence Set for 98–100% of assets. These are author-reported study results, not a general superiority claim, firm deployment evidence, or a trading recommendation.

The academic capture note records the recovery queue: Brunnermeier’s paper PDF, MACROCAST code and vintage files, and the VOLARE dataset and full paper. Any replication should report asset- level results, data revisions, forecast vintages, model versions, costs, and confidence procedures rather than relying on pooled averages.

September 4, 2026 — Manchester foundation models and earnings-call audio routes

The Manchester route adds a concrete model-and-person surface. Eghbal Rahimikia’s University of Manchester profile identifies him as an Assistant Professor in Financial Technology and Senior Research Fellow at the Alan Turing Institute, with University of Tehran, Iran University of Science and Technology, and Alliance Manchester Business School PhD training. It lists work on machine-learning volatility forecasting, limit-order-book and news alternatives, and finance-AI teaching. His personal research page states that FinText released a suite of more than 600 finance time-series foundation models and maintains finance-specific LLM and benchmark artefacts. These are public research and repository claims. The exact model inventory, training corpus, licenses, and independent performance remain recovery items, and no manager deployment is established.

A Cornell CFEM/UBS seminar listing describes Rahimikia’s “Re(Visiting) Time Series Foundation Models in Finance” as a comparison of zero-shot forecasting, fine-tuning, and training from scratch on daily excess returns across 94 countries over several decades, against statistical and ML baselines. Cornell’s page was rate-limited during direct retrieval, so the event extract remains a lead until the recording, slides, and paper are archived. The cross-country design is a useful test surface, not evidence of a tradable signal or fund adoption.

The voice layer gains another academic route from Missouri’s Trulaske College of Business report. It describes Nargess Golshan’s Journal of Accounting Research paper, “Silent Suffering,” using a machine-learning measure of depression-related acoustic patterns over more than 14,500 S&P 500 CEO earnings-call recordings from 2010–2021. The report relates the measure to compensation, turnover, and firm-level risk indicators, while explicitly noting that the associations are not causal and that higher depression scores did not systematically imply worse observed performance. This is governance and accounting research, not a validated trading signal. Consent, construct validity, speaker-level leakage, temporal controls, and ethical review are prerequisites for any financial application.

The academic capture note records the next recovery targets: FinText checkpoints and model cards, the Cornell seminar materials, and Golshan’s paper and audio-feature construction. These routes remain separate from hedge-fund personnel or deployment evidence.

September 4, 2026 — multimodal finance agents, long documents, and professor-led research

The professor and programme search recovered a second set of research artifacts that make the idea surface more specific. They are not evidence that a tracked hedge fund uses any of these systems.

F²Agent: Financial Fusion of Agentic Intelligence for Multimodal Trading, submitted August 6, 2026, names Changshuo Liu, Yanzheng Jin, Shangfeng Cai, Peng Fang, Xiaokui Xiao, and Beng Chin Ooi. Its proposed architecture separates modality-specific signal extraction into a hierarchy of specialised agents, then uses modality-aware adaptive fusion and a noise-robust consistency objective. The public abstract reports experiments on six stock and cryptocurrency assets and author-reported comparisons with baselines; those claims remain preprint results and do not establish robustness after costs, turnover, unseen assets, or point-in-time controls.

Frontiers in FinTech: Multimodal Foundation Models for Financial Reporting and Decision Science describes FinVision as a model that reads text, tables, and images together, converts them into structured data, and checks cross-modal agreement. Its abstract claims two-stage training—public financial corpora followed by institution-specific investment data—and describes natural- language workflows for valuation, portfolio optimisation, and risk monitoring. The abstract reports a 200-company evaluation and a 48-professional user study, but the public record does not independently verify the data rights, institution, model weights, study protocol, or production use. Treat the numerical results as author claims pending paper, code, data statement, and replication review.

FinAcumen: Financial Multimodal Reasoning via Self-Evolving Experience Memory Harness, revised August 23, 2026, names Pianran Guo, Pengcheng Zhou, Yucheng Jian, Shuhua Chen, Zhongliang Yang, and Linna Zhou. The framework stores successful reasoning strategies and failure-derived caution rules in a persistent memory bank, activates memory only above a semantic-relevance threshold, and falls back when retrieved experience is irrelevant. It also describes a deterministic tool environment for numerical calculation, retrieval, visual decoding, and answer verification across four multimodal finance benchmarks. The paper says code is available through its arXiv record; the exact repository, license, benchmark splits, and independent results still need preservation.

IPO Finance Agent, by Mostapha Benhenda and revised June 30, 2026, extends an SEC-filing finance-agent benchmark to IPO due diligence. The paper identifies long S-1 filings as a distinct retrieval problem and describes contextual retrieval after a naive chunk-retrieval harness failed to return usable SpaceX S-1 output. It releases 70 public questions from a 1,000-question IPO dataset and uses an evaluator–optimizer pipeline that extracts candidate facts, audits draft criteria for omissions and hallucinations, and iteratively repairs and deduplicates the rubric before human review. The benchmark’s model comparisons are paper-reported; they are not a ranking of commercial models or evidence of a fund workflow.

Professors and finance-programme routes

The WashU Research Profiles record for AI Democratization and Trading Inequality names Anne Yanru Chang, Xi Dong, Xiumin Martin, and Changyun Zhou and identifies the 2026 peer-reviewed Journal of Accounting Research article. The authors use an AI-sentiment measure extracted from earnings-call transcripts and study how trading alignment changes around the broad deployment of ChatGPT, including retail traders, short sellers, information asymmetry, and AI-service outages. The record’s abstract reports the authors’ causal interpretation from outage variation; it does not disclose the exact model, prompt, transcript vendor, outage construction, or a deployable trading strategy. This is a useful professor-led route for studying AI-mediated information diffusion, with the paper’s identification and data construction requiring independent review.

The Johns Hopkins publication page for Nicholas Andrews exposes a connected faculty-and-paper route across finance, language models, speech, privacy, and uncertainty. In finance, Ross Koval, Andrews, and Xifeng Yan describe modality-specific experts for interleaved financial text and time series in Multimodal Language Models with Modality-Specific Experts for Financial Forecasting, including cross-modal alignment and an interpretability analysis. Their Context-Aware Language Models for Forecasting Market Impact from Sequences of Financial News uses a large language model for the current article and a smaller model to encode historical context into summary embeddings. Andrews’s page also lists work on semantic uncertainty and long-form audio privacy. Together these papers expose three testable controls for a financial research agent—modality-specific encoders, historical context retrieval, and calibrated uncertainty—without establishing a hedge-fund relationship or live investment use.

These additions change the academic search queue in a practical way. Search should now follow professors and courses into model repositories, benchmark schemas, syllabi, student projects, conference talks, and research-assistant lineages—not just search for “AI hedge fund.” For each route, preserve the publication date, information cutoff, modality, source and license, model revision, data rights, evaluation denominator, transaction-cost assumptions, and whether a result is a paper claim, a university description, or independently reproduced. None of these routes supports a cross-firm ranking or an inference that an academic method is used by a particular manager.

The academic capture note and coverage ledger keep the paper, professor, programme, and firm evidence classes separate. Recovery targets are the FinVision data statement and code, the FinAcumen repository and benchmark splits, the IPO benchmark’s public question set and rubric schema, the full WashU paper and outage design, and Andrews’s financial-paper code and data statements.

September 4, 2026 — Cambridge global baseline for finance AI adoption

The Cambridge Centre for Alternative Finance’s 2026 Global AI in Financial Services Report provides a useful context layer for the firm-specific evidence. The University of Cambridge Judge Business School study reports responses from 628 organisations across 151 jurisdictions: 203 fintechs, 149 traditional financial institutions, 146 AI vendors, and 130 central banks or other regulators. It measures adoption stage, use cases, investment, value, risks, cloud and model-provider choices; it is a survey-based sector baseline, not a census of hedge funds and not evidence about any named manager.

Among surveyed industry respondents, the report says 81% were adopting AI at some level and 52% had active adoption of agentic AI. It places current use most visibly in process automation, data visualisation, software engineering, and data/knowledge management, with risk-oriented examples including fraud detection and credit-risk modelling. The report also says 63% of industry respondents used external foundation models in internal workflows; its provider figures—OpenAI 76%, Google 57%, Anthropic 35%, and DeepSeek 15%—are survey percentages and should not be read as market share or as evidence of a particular firm’s vendor relationship.

The operationally relevant finding is measurement difficulty. The report says 55% of industry respondents had difficulty measuring AI value, rising to 76% among large financial institutions, while data quality, talent, legacy architecture, privacy, hallucinations, and loss of human oversight recur as constraints or risks. This gives the research programme a neutral cross-check: when a firm discloses an AI project, look for the workflow, model boundary, data rights, cost instrumentation, human review, and measurable outcome rather than treating a tool name as evidence of transformation.

The report also opens a new academic and institutional-people route. Its research team and acknowledgements name Cambridge CCAF/Fii staff and contributors from the BIS, IMF, World Bank, IDB, CGAP, Arab Monetary Fund, and WEF. Those names should be searched into their own papers, webinars, conference panels, regulator publications, and institutional disclosures; the report itself does not establish a hedge-fund connection or a production investment system.

The academic capture note records this as a global context and personnel-discovery route. The next pass should recover the report’s cross-tabulations and respondent definitions, then test whether the same workflow and control fields appear in public materials from tracked managers and adjacent family offices, banks, vendors, and regulators.

September 4, 2026 — Canada–Australia professor, allocator, and family-office routes

The regional search adds a different kind of evidence: professors and allocator events that expose model-risk, market-structure, and institutional-investor language without claiming a hedge-fund deployment.

The University of Guelph profile for Nikola Gradojevic identifies him as Professor and Fidelity Chair in Finance, with a University of British Columbia PhD and prior roles or affiliations involving the Bank of Canada, the Federal Reserve Bank of St. Louis, IÉSEG, and CARE-AI. The profile lists research interests in empirical asset pricing, market microstructure, high-frequency finance, forecasting, AI/ML, energy, and crypto markets. It also lists two public patents on AI/ML model validation and fuzzy-logic model-risk management, plus 2026 work on natural-gas sentiment and order flow/cryptocurrency returns. The profile’s media list includes Canadian Family Offices; a public LinkedIn post points to a March 2026 Canadian Family Offices discussion of geopolitical risk in investment models. These are academic, patent, and public-media routes; they do not establish a family office’s system, a fund relationship, or live performance.

The University of Sydney event “AI, Opinion Ecosystems, and Finance” records a May 5, 2026 presentation by Tsinghua’s Xiaoyan Zhang. Its abstract reports different associations for AI-assisted content on Seeking Alpha and WallStreetBets, including changes in information frictions, retail order flow, spreads, trading volume, volatility, and lottery-like returns. The event page is a research-seminar record, not a causal validation of a social-media strategy or evidence of a manager’s content pipeline. The public claims require the underlying paper, sample construction, platform data rights, bot/selection controls, and point-in-time tests.

The Sydney Business School Asset Management Forum provides an allocator and talent-pipeline route dated May 28, 2026. The listed speakers include Aware Super’s Head of Fundamental Equities and Senior Portfolio Manager Alvin Chan and Senior Portfolio Manager Dean Montgomery, alongside CoreData researchers and Sydney finance faculty Henry Leung and Vycke Wu. The page frames the conversation around investment decision-making, AI in finance, sustainability, and career pathways. It does not provide a transcript, model, data source, or evidence that Aware Super used AI in a portfolio process, so the speakers remain discovery handles rather than deployment evidence.

The 2026 Sydney Time Series & Forecasting Symposium is scheduled for November 26–27, 2026 and lists time-series econometrics, volatility and risk forecasting, high-dimensional modelling, robust inference, machine and deep learning, and investment applications. This should be monitored for recordings, papers, code, and speakers whose work bridges academic forecasting and asset management. The event page itself does not establish attendance or adoption by any tracked manager.

Together, these routes widen the search beyond hedge-fund names: model-validation patents, allocator panels, family-office publications, and time-series conferences can expose people and methods before a firm makes a direct disclosure. The evidence class remains explicit—profile assertion, patent record, event metadata, and public media are not interchangeable with a verified production system.

The academic capture note and ledger preserve the Canada–Australia routes separately. Next checks are to recover the Guelph papers and patents, the Canadian Family Offices article, the Sydney seminar paper and any recording, the Aware Super panel materials, and the symposium programme once speakers and papers are posted.

September 4, 2026 — university finance programmes and allocator-side family-office routes

Professor-led programmes that expose concrete research surfaces

• Columbia IEOR E4723: AI Investment Management lists Cyril Shmatov as instructor for Fall 2026. The course description covers prompting, retrieval-augmented generation, agentic frameworks, risk-factor extraction from equity data and unstructured text, real-time news monitoring, sentiment and portfolio-event flags, stress-test generation, multi-agent research and trading, guardrails, human oversight, and classical portfolio construction methods including risk parity, mean-variance optimisation, and Black–Litterman. The listing also says the course includes industry guest lectures. This is a curriculum-derived workflow map, but it does not identify guest firms, student systems, datasets, model providers, or production deployment.
• University of Mannheim FIN 6080 — Artificial Intelligence in Finance names Erik Theissen and PhD student Paul Seidel and describes a 13-session course spanning classical ML, transformers, LLMs, RAG, agents, asset pricing, trading, earnings calls, 10-K filings, credit, risk, RegTech, and GenAI in capital markets. The final assessment requires either a 15–20 page empirical paper with replication code and AI-workflow documentation or a functioning finance MVP with technical architecture. The programme creates a route into student projects, code, model choices, and industry guests; the course page does not establish that any project became a fund system.
• Singapore Management University SMU-X Financial Services in AI Immersion Program describes a six-week industry project after a boot camp and four weeks of blended sessions. The stated project areas include risk management, sales and marketing, technology and operations, regulatory compliance, and investment research. The public outline explicitly names prompt engineering, explainability, AI prototyping, data governance, security, auditability, responsible AI, and implementation roadmaps. It is a route to identify participating financial institutions, project artefacts, and student or faculty collaborators; the page does not name the host institutions or reveal their data and model boundaries.
• Warwick Business School IB9TF-15 Agentic AI for Finance lists Michael Mortenson as module leader and a two-week postgraduate module. Its outline covers process analysis and re-engineering, training and fine-tuning, tools and Model Context Protocol, retrieval and memory, supervision, human-in-the-loop governance, ethics, strategy, and finance case studies. The assessment includes a 3,000-word analysis with code and a group video presentation. This is a compact route for tracing what future finance practitioners are being taught to build and govern, not evidence of a named manager’s implementation.
• Luiss Data-Driven Models for Investment lists Antonio Simeone and Villy Edoardo de Luca for 2026/27. Its public syllabus combines satellite, geospatial, web, and news data with LLMs and GenAI; qualitative and quantitative AI; time-series foundation models including TimeGPT, Chronos, and TabPFN; multi-agent systems; fuzzy logic; genetic algorithms; and systematic-trading projects. It also describes an industry session involving a UK hedge fund’s Head of Capital Markets, but the page does not name the fund, disclose the speaker’s identity, or establish access to a proprietary system. Treat the named model families and alternative-data modalities as curriculum leads, not as proof of manager usage.

These programmes expose the implementation boundary that a job title or marketing page often hides: document and earnings-call ingestion, point-in-time news monitoring, stress-test generation, retrieval and memory, typed tools, model evaluation, data governance, auditability, human review, and final portfolio or research workflows. The next recovery pass should seek syllabi, project repositories, guest-speaker recordings, student theses, and named partner institutions. None of the pages supports a firm ranking or an inference that a hedge fund uses the methods taught.

Family-office reports and public allocator conversations

• UBS Global Family Office Report 2026 says its online survey covered 307 UBS family-office clients across more than 30 markets between January 22 and March 30, 2026. The participating families had average net worth of $2.7 billion and their family offices managed an average of $1.3 billion in assets, according to the publisher. UBS reports that 65% were invested across the AI value chain, with reported allocation interests in power and resources, infrastructure, and AI-enabled healthcare. The report is an allocator-demand and thematic-allocation signal; it does not reveal which family offices, their internal AI workflows, their manager selections, or their model providers.
• J.P. Morgan Private Bank’s 2026 Global Family Office Report describes 333 single-family-office participants from more than 30 countries and an average participant net worth of about $1.65 billion. Its public report page says 65% plan to prioritise AI investments, while the associated visual gives 43% with any venture-capital or growth-equity exposure and 21% with any infrastructure exposure, with small average portfolio exposures among those surveyed. It also highlights increasing outsourcing of investment functions. This creates a route into family-office operating models, external-manager demand, and the gap between thematic interest and direct exposure; it does not identify a family office’s AI stack or establish investment outcomes.
• PwC Family Office Stories episode 26 identifies Lisa Cornwell and Danielle Valkner, PwC’s US family-office leader, in a podcast about professionalisation, AI, governance, next-generation engagement, talent, processes, controls, and technology. The page has no public transcript in the captured HTML, so it should be recovered through the podcast feed or audio page and transcribed with timestamps. At present this is a first-party episode and named-practitioner route, not evidence of a specific family office’s implementation.
• BMO Beyond the Portfolio — “AI: What Matters for Investors” provides a publisher-hosted seven-page transcript dated June 1, 2026. Mike Miranda identifies himself as president of BMO Family Office and head of investments for wealth management; Dan Phillips is identified as BMO Wealth Management’s US chief investment officer. The conversation divides the AI investment stack into energy and materials, data centres, semiconductors, platforms, and applications, and discusses build-out risk, concentration, capital intensity, and a long investment horizon. This is an allocator-side investment framework with unusually good text provenance, not evidence of an internal BMO model, a hedge-fund strategy, or AI-attributed returns.
• AWM Insights #252 — Smart Investing in the Age of AI is a May 14, 2026 family-office resource page naming Chief Investment Officer Justin Dyer and Portfolio Manager Mena Hanna. Its publisher-generated, human-edited summary discusses data-centre valuations, technology bubbles, diversification across venture stages and public/private markets, infrastructure suppliers, and concentration risk. The page links to YouTube, Apple Podcasts, and Spotify and labels its text as generated from the episode and edited by a human. It is a useful family-office media route, but the public page does not expose a full transcript, model, data source, or portfolio attribution.
• PwC Italy’s Family Office Survey 2026 surveyed 48 family-office structures legally based in Italy, including 41 single-family offices and seven multi-family offices. The report says 22 respondents selected artificial intelligence among the themes they were asked to assess for the following 12 months, and its chart records 88% for AI. It also assigns technology/AI a 3.6 priority score on a five-point scale for investment over the next three years, while discussing wealth aggregation, investment monitoring, and reporting as major digitalisation activities and market-ready solutions as the preferred path over internal development. These are small-sample survey results about family-office themes and operating preferences, not evidence of named systems or performance.

The allocator routes add a useful negative finding: public family-office material currently reveals more about AI as an investment theme, infrastructure exposure, governance, outsourcing, and workflow digitisation than about internal model training or autonomous investment authority. That distinction should remain explicit in future family-office coverage. The podcast and transcript routes should be searched by speaker, episode, and linked media rather than by “hedge fund” in the title, because the relevant language is usually family office, wealth management, investment committee, or portfolio construction.

A professor-led architecture route for fundamental research

FundaPod, submitted in May 2026 and revised in June, is by Di Zhu, Lei Nico Zheng, and Zihan Chen, with affiliations shown on the public record as Stevens Institute of Technology and UMass Boston. The paper proposes independent persona agents—such as value and macro roles—working under a shared provenance contract, with disagreements surfaced for human portfolio-manager adjudication through a knowledge graph. Its described mechanisms include persona distillation from public investor materials, a declarative skill registry, claim-to-source evidence links, and a knowledge-graph “second brain” connecting tickers, memos, analysts, and themes. This is directly relevant to the design of a research-agent evaluation harness because it separates evidence gathering and viewpoint generation from final human adjudication. It is an academic design paper and case study, not evidence that any named manager has deployed the architecture or that it produces investable alpha.

Recovery queue from this pass

Recover the Columbia guest roster and syllabus, Mannheim student projects and guest recordings, SMU partner institutions, Warwick assessment artefacts, and Luiss’s unnamed hedge-fund session. Recover the PwC episode audio or transcript, BMO’s linked AI-investing paper, AWM’s full episode transcript, and the UBS/J.P. Morgan report downloads. For each item, preserve source date, speaker role, information cutoff, model and data claims, rights, evaluation denominator, and the boundary between thematic allocation, workflow adoption, academic design, and verified production use.

September 4, 2026 — MIT faculty and practice-linked finance curricula

• MIT Sloan Executive Education — Artificial Intelligence for Financial Services lists Andrew W. Lo as course lead and Hui Chen and Haoxiang Zhu among the faculty. The 2026 brochure frames the two-day executive programme around applications across the buy side, sell side, banking, insurance, risk management, quantitative trading, retail investing, and wealth management, alongside LLMs, responsible adoption, failure modes, regulation, and case studies with practitioners and researchers. It names an executive audience responsible for AI or data strategy, but does not publish the case-company roster, participant list, private examples, or any production system.
• Hui Chen’s MIT teaching page describes 15.S06, “AI and Machine Learning Research in Finance,” as a project-based course with research-paper presentations, industry guest lectures, faculty meetings, and team research papers. MIT’s finance course catalogue adds practice-linked routes: 15.439 Quantitative Investment Management covers strategy construction, portfolio and trading costs, risk management, and guest lecturers; 15.451 Proseminar in Capital Markets/Investment Management assigns practitioner-posed problems to student teams; and 15.453 Finance Lab uses real-world finance questions from practitioners including investment managers, hedge funds, private equity, venture capital, risk, and consulting. These pages expose a route into sponsor names, project briefs, student reports, and guest appearances that may be more revealing than course titles. The catalogue does not identify current sponsors or disclose confidential project content.

The MIT route is useful for two separate searches. The executive brochure may lead to current practitioners and case studies around adoption and governance. The project-based courses may lead to research questions commissioned by market participants, but a sponsored project is not automatically a firm disclosure or evidence of model deployment. Search each named faculty member, guest, sponsor, student paper, and course archive independently, and preserve the course year and public/private boundary.

September 4, 2026 — Taiwan and Germany academic routes

• National Sun Yat-sen University’s Application of Artificial Intelligence in Finance syllabus lists instructor 王尚文 for the 2026 graduate course. The published syllabus explicitly names Hugging Face, LangChain, LangGraph, CrewAI, and AutoGen; vector stores and RAG; LoRA; synthetic data with GANs, diffusion models, and VAEs; reinforcement learning for trading, hedging, execution, and portfolio optimisation; federated and decentralised AI; multimodal data; and agentic workflows with memory, planning, tools, guardrails, audit trails, and recovery. It also assigns application plans and project presentations. This is a rare public tooling-and-governance syllabus, but it does not identify industry partners, student projects, data rights, or production systems.
• Sebastian Weibels’s 2026 PhD thesis at the University of Cologne is a public, refereed dissertation with three empirical-finance studies. The thesis covers direct ML optimisation of stock portfolio weights conditional on firm characteristics, corporate-earnings forecasting from financial-statement variables, and an autoencoder-based measure of firm atypicality. Its abstract reports that the importance of income-statement variables is greater for shorter earnings horizons while balance-sheet information becomes more relevant at longer horizons; it also reports an atypicality/forward-return relationship concentrated among firms with higher limits to arbitrage and lower investor attention. These are thesis-level, author-reported findings requiring review of the downloadable chapters, sample construction, timing, costs, and replication code. The dissertation does not establish a fund relationship or live deployment.

The Taiwan syllabus supplies implementation vocabulary for title-blind job and conference searches: vector stores, LoRA, synthetic stress data, federated learning, agent memory, and recovery controls. The Cologne thesis supplies a cross-sectional research route that is not limited to sentiment: portfolio weights, earnings horizons, information-processing frictions, and representation learning. Search both routes into repositories, seminars, student placements, and subsequent employer pages without treating curriculum or dissertation publication as proof of adoption.

September 4, 2026 — professor and finance-programme routes

• Imperial College London — AI in Finance: Strategy, Applications and Impact lists 2026 cohorts beginning October 15 and December 10 and names Marco Di Maggio in the programme description. Its outline separates AI in trading and investment management from modules on implementation, risk, and return on investment. The page connects Di Maggio to prior Harvard Business School leadership of a FinTech, Crypto, and Web3 Lab and to research on emerging technologies in financial markets and organisations. The page was publicly indexed but returned a 403 to the research fetch, so the dates and module outline should be rechecked against the downloadable brochure or registration materials. This is an executive-education and professor-lineage route, not evidence of a named manager’s system.

Across the academic routes now tracked, the public signal is distributed across distinct research and implementation surfaces:

1. Forecasting and portfolio construction: Southampton, Durham, Duke, Cologne, and SKEMA expose return forecasting, cross-sectional prediction, portfolio optimisation, factor discovery, time-series modelling, and comparisons between AI and classical baselines.
2. Unstructured and multimodal information: Illinois, Mannheim, CEMFI, Yale, and MIT connect filings, earnings calls, news, images, audio/video, and structured forecast objects to finance questions. These pages expose possible research inputs, not evidence that a manager has lawful access to the same data or uses the same models.
3. Agentic research and controls: Columbia, Warwick, NSYSU, SMU, and FundaPod expose RAG, memory, tools, multi-agent decomposition, audit trails, human review, prompt/model robustness, and recovery. The useful follow-up is to recover assignments, code, guest sessions, and evaluation rubrics while keeping student work separate from firm deployment.
4. Market structure and strategic behaviour: Wharton’s AI in Finance Lab route, NUS’s quantitative-finance programme, Monash’s AI-in-investment colloquium, and the MIT finance catalogue connect AI to information production, market power, liquidity, pricing, risk, and execution. These are research agendas and practitioner discussion routes, not disclosures of trading authority.

This academic layer improves title-blind personnel and media discovery: search professors with finance-and-ML papers, industry guests, course sponsors, student repositories, course-specific recordings, and alumni transitions into quant research. A public curriculum or publication establishes a research or talent route; it does not establish a hedge-fund relationship, proprietary data, production model, live permissions, or investment performance.

Academic recovery queue

Recover Imperial’s brochure and faculty biographies; retrieve syllabi, project briefs, guest rosters, recordings, and public student code for Duke, Southampton, Durham, Illinois, Mannheim, Warwick, Columbia, SMU, NSYSU, and MIT; and map papers, supervisors, and subsequent employment only from direct public evidence. Preserve whether each artifact is a course description, working paper, peer-reviewed result, conference talk, or firm disclosure.

September 4, 2026 — new professor, allocator, and index-provider audio routes

• George Washington University Investment Institute — Market News with Rodney Lake, episode 79 is a 21:58 episode dated February 12, 2026. Rodney Lake identifies himself as director of the GW Investment Institute and says the Institute’s students manage just over $11 million across student funds. The timestamped page describes a 2:44 segment on AI at the Institute and a 5:26 segment on Gemini 3.0; in the transcript Lake says students can use Gemini, ChatGPT, Grok, and Claude alongside FactSet and Bloomberg for fundamental analysis, then iterate on a business/management, price/valuation, and balance-sheet scoring framework. This is unusually direct evidence of an academic investing workflow and its human-review boundary, but it is a professor’s description of student practice, not evidence of a hedge fund’s stack or of AI-attributed returns.
• FTSE Russell Convenes — “AI, energy innovation and the UK stock market” is a June 15, 2026 Scholar & Investor episode with Professor Constantine Yannelis, Janeway Professor of Financial Economics at Cambridge, and Thomas Moore, Senior Investment Director at Aberdeen Investments. The publisher supplies both a transcript and chapter timestamps. Moore describes an active-equity team, a benchmark-relative process, and company meetings; Yannelis connects AI investment to compute energy, infrastructure, innovation financing, and the Oxford–Cambridge corridor. The episode includes a reference to an unnamed hedge-fund manager’s data-centre energy consumption, but that anecdote is not independently identified and should not be turned into firm evidence. The episode is a bridge between academic research and traditional active-management decision context, not a disclosure of a model.
• MSCI Perspectives — “AI Is Rewriting Risk — Are You Ready?” transcript is dated March 12, 2026 and identifies Hitendra Varsani, head of equity-index investment research and development, and Andy DeMond, head of analytics research platform and governance. Varsani describes a three-layer AI investment map—physical infrastructure, digital platforms/model providers, and applications. DeMond describes agents that pull data, calculate, generate code or visualisations, and iterate, with possible effects on portfolio construction, risk analysis, reporting, and governance; the transcript also stresses model risk, concentration, and human judgment. This is an index-provider/analytics-platform perspective with named roles and a preserved transcript, not proof that a particular hedge fund uses MSCI’s tools or the described workflow.

These routes extend the academic search beyond course catalogues. Student-managed funds expose how general-purpose models are actually framed in an educational investment workflow; scholar–investor media exposes how AI themes are translated into active-equity and macro narratives; and index-provider transcripts expose the operational vocabulary used around agentic risk and portfolio work. Preserve the transcript, timestamps, speaker role, publication date, and whether a statement is a personal view, an educational practice, a vendor perspective, or a firm disclosure.

Audio recovery queue

Download and retain the GW episode audio and chapter transcript, recover the FTSE Russell audio file and any associated video, and preserve the MSCI PDF plus adjacent episodes. Search each speaker and programme for earlier/later appearances, faculty pages, research papers, guest institutions, and public model or data references. Do not infer firm adoption from a guest’s presence or from a vendor’s description of what investors could do.

September 4, 2026 — further title-blind academic and allocator routes

• S&P Global Masters of Risk, Season 4 episode 7 is a video/transcript episode recorded at the 2026 Definitive Risk Conference and published August 27, 2026. It identifies Matthew Tuttle as CEO and CIO of Tuttle Capital Management. Tuttle describes a thematic and ETF process that disaggregates the AI opportunity into memory, photonics, energy, utilities, materials, and space infrastructure, and discusses valuation, capital-spending slowdown signals, leverage, and private-credit ETF risk. The transcript also contains a self-reported scale statement of about $5 billion and roughly 70 ETFs. This is a public investment-manager interview and strategy-language route, not evidence of a hedge fund, proprietary AI model, or AI-attributed performance.
• Cambridge Associates — “The Rearview Mirror Problem” is a May 26, 2026 Cambridge Conversations episode with Chief Investment Strategist Celia Dallas and Head of Global Capital Markets Research and Investment Communications Kevin Rosenbaum. The page frames AI enthusiasm, private-market growth, concentration, credit signals, and diversification as allocator issues and links to a transcript. Cambridge Associates identifies its work across endowments, foundations, private clients, family offices, hedge funds, and other institutional portfolios. This is an allocator and outsourced-investment-research route; the page does not disclose a client’s models, vendor stack, or manager-level process.
• Morgan Stanley Institute Roundtable — “AI Disruption: What’s Fact, What’s Fiction?” is dated February 27, 2026 and provides an audio transcript. It identifies Andrew Sheets, global head of fixed-income research; Lisa Shalett, CIO of Morgan Stanley Wealth Management and chair of its Global Investment Committee; and Rui de Figueiredo, global head of investment and client solutions and CIO of the solutions and multi-asset group at Morgan Stanley Investment Management. The discussion separates AI-capex expectations, sector disruption, credit and valuation transmission, and implementation across macro, sectors, and companies. It is a senior allocator and investment-management perspective, not evidence of a named model or hedge-fund deployment.
• University of Arizona Eller Student Managed Investment Program describes a faculty-supervised, multi-tier student investment structure. The May 2026 page identifies Daniel Kinnear as director, reports an $8.8 million senior-fund value, a $3.8 million individual-stock portfolio, a $5 million semi-active enhanced-index strategy, and a separate master’s fund focused on factor returns through ETFs. The programme describes a pipeline from 230-plus senior valuation projects into a filtered set of portfolio ideas. This is a talent and applied-portfolio route; it does not disclose AI use, hedge-fund employment, proprietary data, or live-manager adoption.

Together these sources widen the search surface from explicit AI-labelled episodes to institutional risk programmes, allocator conversations, investment-manager interviews, and student portfolio operations. The next step is to recover linked transcripts and videos, inspect adjacent episodes and speaker profiles, and trace faculty, alumni, sponsors, and programme guests only where the public evidence supports the connection.

Further media and academic recovery queue

Recover the S&P Global video/transcript asset, the Cambridge Associates linked transcript, and the Morgan Stanley audio file; capture Arizona’s programme reports and faculty profiles; then search each series and speaker by episode, role, employer, paper, course, and conference rather than by “AI” alone. Keep self-reported AUM, educational fund values, allocator commentary, and firm disclosures in separate evidence classes.

September 4, 2026 — new academic–industry and finance-programme routes

• Rutgers Business School’s Managing Scarlet Student Fund is a concrete academic-to-practitioner route. The April 8, 2026 first-party account identifies finance professor Rose Liao and Daniel Davidowitz, a Polen Capital portfolio manager and Rutgers alumnus, as the people who developed the two-semester course and its $1 million student-managed fund. The account describes students presenting company and industry research, voting on initial holdings, and appointing five student portfolio managers. It exposes a personnel, curriculum, and real-money workflow surface; it does not disclose AI use, proprietary data, Polen Capital’s internal systems, or performance attributable to the course.
• Florida Gulf Coast University’s Eagle Fund describes more than $1 million of real university-endowment money managed by students under faculty and industry guidance. The page says the programme uses Bloomberg, practitioner meetings, CFA events, portfolio reviews, and an “AI-Assisted Research Report” example. The public page does not identify the model, prompt, training data, validation design, or decision authority behind that example. It is therefore a useful lead for recovering the report and student workflow, not evidence that a hedge fund or the university has deployed a validated AI investment system.
• Stevens Institute of Technology’s Dragos Bozdog profile identifies him as Teaching Associate Professor of Financial Engineering and Deputy Director of the Hanlon Financial Systems Laboratories. His listed research includes high-frequency financial data, rare events, early-warning systems, machine learning in finance, and liquidity. The profile also records prior quantitative-analyst work, CFTC research-analyst consulting, Cohen Capital Group and Foochee Trading consulting, and student theses on an AI-integrated FPGA for market making, analyst-EPS forecasting, financial text mining, LSTM stock prediction, and corporate-bond ML. These are valuable personnel, lab, and idea-lineage clues; the page does not establish a current employer deployment or production trading authority.
• King’s College London’s AI-in-finance PhD project brief names Carmine Ventre as supervisor and frames finance AI around safety and trustworthiness. Its proposed directions include latent-space explainability and formal assurance mechanisms, symbolic methods versus model-based deep learning for computational efficiency, intrinsic-time labelling, and transformer-based ESG analysis where data can be manipulated through greenwashing. This is a research agenda rather than a fund disclosure, but it adds a non-sentiment route for model ideas: robust labels, adversarial data quality, explainability, and compute-aware validation.

These sources add four distinct idea surfaces to the academic layer: practitioner-mentored real-money training, AI-assisted fundamental research, hardware-aware market making and liquidity research, and safety-first financial AI. The cross-source implication is limited but useful: search the associated course projects, student reports, lab theses, guest speakers, and alumni destinations for implementation clues. A student fund, thesis, or research proposal remains separate from a verified manager system, lawful data access, live trading permissions, or investment performance.

Recovery queue for this tranche

Recover the Rutgers course materials and fund policy; request or locate the Eagle Fund AI-assisted report and its methodology; inspect Stevens theses and Hanlon lab artefacts; and follow the King’s College project into Ventre’s publications and any doctoral appointment. Preserve the exact academic year, document status, model/data claims, and public/private boundary for each follow-up.

September 4, 2026 — international labs, finance-AI conferences, and one partial capture

• Florida International University’s Environmental Risk & Finance Lab adds a non-text and non-sentiment route. The lab page says Dan Li joined FIU in August 2026 as an assistant professor in the Environmental Finance & Risk Management Program and describes models of how hurricanes, floods, droughts, and wildfires propagate through engineered and financial systems. It names stochastic modelling, machine learning, catastrophe-risk modelling, environmental-systems analysis, parametric insurance, and derivatives. This is a university lab and decision-support research route; it does not establish a manager’s data feed, insurance book, or production model.
• Chicago Booth’s AI and Economics Summer Conference took place on August 11, 2026 and included finance-adjacent work on returns to intelligence, generative-AI empirical likelihood, pretrained embeddings for unstructured data, agentic monoculture under firm competition, and an asset-pricing session listed on the programme. The event names Suproteem Sarkar as an assistant professor of finance and applied AI and identifies the Center for Applied Artificial Intelligence as organiser. These are research and personnel discovery routes, not evidence of a firm adopting any presented method.
• Sungkyunkwan University’s Global Finance Research Center seminar page is a Korean-language title-blind route dated July 30 for an August 1, 2026 hybrid seminar, “Factor Zoo and Machine Learning,” presented by Yeonchan Kang. The page lists domestic and overseas academic participants, including researchers from Virginia, RMIT, EDHEC, Singapore Management University, Southern Denmark, Wollongong, and Zhejiang. It exposes a regional seminar and personnel network that would be missed by English-only or hedge-fund-keyword searches; the page does not expose the paper, code, data, or industry deployment.
• NUS Edge Research’s case study on AI tools is a partial capture. The official PDF’s indexed text describes using Gemini, ChatGPT, and DeepSeek for sector research and prototyping a fact-checking process based on link reachability, recency, and source-credibility checks. Direct retrieval currently returns an HTML/error response because of the site’s delivery layer, so those details remain medium-confidence until the PDF is recovered and its authors, course, prompts, outputs, and review process are captured. Do not treat the indexed summary as evidence of a fund system.

This tranche widens the idea map beyond ordinary news and earnings sentiment: physical-hazard propagation into financial risk, embeddings and agentic behaviour in economic research, multilingual factor-discovery seminars, and source-validation infrastructure for AI-generated investment research. The useful next links are lab publications, conference papers and recordings, student repositories, named participants, and the NUS PDF itself. Academic, conference, and case-study evidence remain separate from verified investment-manager deployment.

Recovery queue for international academic routes

Recover the FIU lab’s publications and project partners; obtain the Chicago conference papers, recordings, and speaker pages; translate and inspect the SKKU seminar archive and linked researcher profiles; and retrieve the NUS PDF through an alternate first-party path. Preserve language, event date, paper status, and whether a claim is a proposal, reported experiment, educational exercise, or production disclosure.

September 5, 2026 — academic labs with investment partners and reproducible student workflows

• Carnegie Mellon’s Accounting AI Research Lab annual report is a new lab-and-partner route. The FY25–26 report, described by the lab’s founder and director Pierre Jinghong Liang, records five working papers, twelve research presentations, four course innovations, and more than $250,000 in new funding, including Quantinno Capital Management LP. It also names the AGE/AGER panel dataset, a controlled data-sharing framework, Business Language Analytics courses, and a dedicated Nvidia Blackwell server planned for Fall 2026. The public account connects a quant-capital partner to accounting structure, graph-based data, and language analytics, but it does not disclose Quantinno’s internal use, data access, model deployment, or investment performance.
• Stanford’s NEA-sponsored “Automating market research for venture capital” project exposes a complete public workflow rather than only a course title. Students used CB Insights, PitchBook, Gartner Research, and eight expert interviews to create a manual baseline for AI-in-financial-services research. They then built a tool that uses STORM to retrieve sources, assign agent roles, surface perspectives, and feed four codified scores—automation potential, competition, customer readiness, and regulatory landscape—into a PDF report. Stanford says the tool was designed for NEA and could be extended into investment sourcing. This is a sponsored student prototype and a valuable architecture lead; it does not establish NEA production use, proprietary data rights, or investment results.
• Rutgers’ Scarlet Student Managed Fund 2026 Annual Conference is a personnel and media-discovery surface that the fund article alone did not expose. The April 24–25 conference lists Rose Liao and Dan Davidowitz as fund advisors and names Andreas Strzodka of Annaly Capital, Rick Romano of PGIM Real Estate, Michael Gonnella of Pershing Square, Melanie Vangopoulos of Gaia Global Capital Management, and Raunak Kasera of Polen Capital. Its “Moving Beyond the AI Hype” session explicitly covers AI-driven workflow efficiency, suitable investment-process applications, bottlenecks, and talent. The agenda identifies speakers and topics but does not provide their internal tools, a transcript, or evidence that any listed firm deployed a particular system.
• Quinnipiac’s Spring 2026 Student-Managed Portfolio account describes 28 finance students managing an endowment-linked portfolio reported at approximately $6.1 million, with value and growth mandates, sector assignments, two-thirds voting for changes, and a 48-page macroeconomic outlook. The page identifies finance professors Matthew O’Connor and Ted Koly and says the course used supervised student analysis rather than AI answering the investment questions. It is useful negative evidence for automation boundaries: the public workflow emphasizes human research, explanation, voting, and accountability. The publisher’s reported portfolio change is not evidence of skill attribution, AI use, or a hedge-fund process.

These routes add a sharper academic-to-industry map: sponsor-funded research infrastructure, source-grounded investment-report automation, public practitioner rosters, and supervised human decision processes around real capital. The most valuable recoveries are now the CMU annual report and controlled-dataset documentation, Stanford’s project presentation and NEA-facing deliverables, the Rutgers conference brochure or recordings, and Quinnipiac’s 48-page outlook. Each should be kept in its own evidence class rather than collapsed into a claim about manager deployment.

Recovery queue for the September 5 academic tranche

Recover CMU’s full annual report and Quantinno lecture materials; capture Stanford’s linked presentation video and project outputs; locate the Rutgers conference brochure, recordings, and speaker biographies; and obtain Quinnipiac’s macro outlook and course materials. Search the named practitioners and faculty independently for later employment, papers, public code, or AI disclosures, preserving the distinction between sponsorship, education, prototype, and production use.

September 5, 2026 — student-investor guardrails, bank-sponsored AI projects, and finance job signals

• University of Dayton’s Davis Center for Portfolio Management provides unusually concrete student-investor workflow evidence. The April 28, 2026 account says students manage more than $80 million through the Center. Student investor Will Otterbein describes a Python script that gathered five years of 10-K and 10-Q filings plus 20 recent 8-Ks, loaded them into NotebookLM, and became a 40-page workflow for research with trusted financial data. The account says the workflow was adopted by most Center students, while first-year interns were kept from using the AI tools until they had completed foundational training. This is a first-party educational implementation report and a clear human-in-the-loop boundary; it does not establish a fund’s model evaluation, data licence, autonomous authority, or performance attribution.
• SMU-X’s AY2026/27 financial-services course listing names UBS as an industry partner for a course in which multidisciplinary student teams work with UBS professionals on financial-services challenges. The public outline covers LLM-powered applications, project scoping, data analysis, integration, workflow automation, and final client presentations; it also describes related sponsor projects involving risk, operations, compliance, and investment research. The page exposes a sponsor, faculty, and project-delivery route, but not the selected UBS problem statements, data, model configuration, or delivered prototypes. It should not be treated as evidence of UBS production deployment.
• Yale’s public listing for a Larson International Group Investment Research and ETF Analyst is a title-blind recruiting signal. The May–June 2026 listing asks analysts to research companies, industries, markets, policy, and geopolitics; contribute to ETF research using quantitative methods and AI-powered analysis tools; and mentor students. It identifies an Investment Challenge and potential internship or career consideration, but does not establish the organization’s assets, specific AI tools, data sources, investment authority, or relationship to any hedge fund. Treat the employer page and any resulting personnel profiles as separate follow-up evidence.

This pass adds implementation details that ordinary “AI in finance” searches miss: a filing-ingestion workflow with staged access, a bank-sponsored prototype lifecycle, and a recruiting title that embeds AI in investment research without using “quant” or “GenAI” in the job title. The recovery targets are the Dayton 40-page workflow, SMU project outputs and sponsor rosters, and the Larson employer and alumni trail. Each remains distinct from verified hedge-fund deployment.

Recovery queue for the September 5 student and recruiting routes

Capture the Dayton workflow artefact and Center policy; recover SMU’s selected UBS project briefs, faculty supervisors, presentations, and permitted data descriptions; and inspect Larson’s employer site, Investment Challenge materials, and alumni or employee profiles. Preserve the difference between a public job listing, an educational prototype, an adopted classroom workflow, and a production investment system.

September 5, 2026 — professor-to-allocator and finance-programme bridges

• NYU Stern’s Ian D’Souza profile is a useful bridge into family-office and allocator research. The university profile identifies him as a Clinical Professor of Finance, a founding member of the NYU Stern Family Office Council, and an investment-committee member for several family offices, managed funds, and endowments. It also describes earlier CIO roles at TMT funds and research interests spanning satellite imagery, weather patterns, sentiment data, distributed-ledger forensics, labor-automation trends, and the interaction of human cognition with AI. This is a faculty biography and advisory disclosure, not evidence that any named allocator uses a particular model, data vendor, or live strategy.
• Stockholm School of Economics’ AI in Finance course adds a programme-level map of prediction tasks. The Fall 2026 course description covers investment management, portfolio construction, credit, insurance, private equity, and implementation on real data; group homework may include earnings prediction, deal selection, and mutual-fund-flow prediction. It lists tree models, neural networks, and NLP as core tools and allows practitioner guest lectures. The public course page does not name the instructor, guest firms, datasets, or student results, so these are curriculum and discovery leads rather than evidence of a manager system.

The academic layer now exposes two complementary routes: faculty with explicit allocator and family-office interfaces, and courses that name testable prediction targets beyond generic “sentiment.” The useful follow-up is to recover the Stern Family Office Council roster and recordings, D’Souza’s papers and patents, SSE’s slides and homework, and the destinations or public projects of students and guests. These routes should remain separate from firm-specific disclosures and from any claim that a taught method is deployed.

Recovery queue for the professor and allocator tranche

Find the public NYU Stern Family Office Council member and event pages; map D’Souza’s cited work, patents, and prior-firm affiliations; capture SSE course materials and any practitioner sessions; and trace named students or guests only through public, attributable sources. Record institution, role, date, model/data surface, and evidence boundary for every link.

September 5, 2026 — live family-office coursework and Swedish finance-AI conference routes

• NYU Stern’s Impact Investing in Family Offices syllabus documents a concrete academic-to-family-office interface. The Spring 2018 course required MBA teams to spend 12 weeks on a quantitative analysis of a live family-office opportunity or challenge; the syllabus says 4–6 family offices would provide projects, interact with teams, and receive a customized work product and final presentation. It also describes client feedback, faculty clinics, expert speakers, and participation by Family Office Council members. This reveals a research-and-consulting access path and deliverable structure, not the identities of the clients, their data, any AI tooling, or investment performance.
• The Swedish House of Finance AI & Machine Learning in Finance conference page gives a historical academic/practitioner map. The August 22–23, 2022 programme was designed around an industry panel and academic papers covering human–machine interaction, asset-price forecasting, algorithmic and high-frequency trading, portfolio construction, new structured and unstructured data, crowdsourcing, credit risk, and fraud. The page lists Bryan Kelly with Yale University and AQR Capital Management among the keynote affiliations, alongside Stefan Nagel, Tarun Ramadorai, and Thierry Foucault. The call does not establish what any firm implemented, which data were used in production, or whether any presented method generated returns.

These sources sharpen the academic-to-market map in two directions: family offices can expose live problem statements through supervised consulting engagements, while finance-AI conferences expose the research vocabulary and practitioner affiliations around forecasting, HFT, alternative data, and human–machine workflows. The next step is to recover client/project names where lawfully public, conference papers and recordings, and the named researchers’ current affiliations without treating any of those links as deployment evidence.

Recovery queue for the live-client and conference routes

Search the NYU syllabus authors, Stern Solutions archives, Family Office Council materials, and public student outputs for client/project identifiers; retrieve the Swedish conference programme, papers, video archive, and keynote biographies; then trace each named academic or practitioner to current profiles and publications. Record the year and whether the item is a proposal, syllabus, talk, paper, or verified firm disclosure.

September 5, 2026 — new professor, family-office, and reproducible-model routes

• University of Georgia’s Terry College account describes “Expectations Matter: When (not) to Use Machine Learning Earnings Forecasts,” by Zhongjin (Gene) Lu, John L. Campbell, Harrison Ham, and Katherine Wood. The researchers evaluated 3,024 model configurations against historical analyst forecasts, using five years of computing time and code released with the paper. The university account says the design varied loss functions, validation, forecasting approach, and other choices; its reported result was that 20% of configurations cleared the analyst comparison, with the correction-of-analyst-forecast setup and time-series validation central to the reported result. The paper’s 40-page online appendix makes the specification surface more concrete: OLS, LASSO, ridge, elastic net, random forest, and gradient-boosted regression trees; temporal versus k-fold validation; MAE/MSE choices; rolling versus expanding windows; and explicit tuning of tree depth, learning rate, feature fractions, and refit frequency. The authors state that replication files are available through the publisher. This is a reproducible academic benchmark and model-selection lesson, not evidence that a tracked manager uses the code or that the result transfers to live trading.
• Duke’s profile for John A. Forlines III provides a family-office, university, and financial-data-company bridge. Duke identifies Forlines as Chairman and CIO of JAForlines Family Office, an Executive in Residence teaching behavioral finance, decision-making, and private investing, and Executive Chairman of ARTD.AI, which supplies fine-art data, analytics, and custom indexes to financial institutions, family offices, wealth managers, and insurers. The profile also names Tappanwood Finance Labs and a Boston-based investment-management board role. It exposes a public personnel and data-vendor route; it does not establish a named client’s model, data licence, or investment outcome.
• HKU ICB’s profile for Patrick Mui identifies a part-time lecturer in financial markets and portfolio management who is also CIO of a family office and head of a venture-capital fund focused on AI and Web3 investments. The profile records prior senior roles at OCBC, National Australia Bank, Nomura, and Citi, and an Oxford engineering-science master’s degree. This is a direct academic-to-family-office personnel signal, but it does not identify the office, portfolio, AI vendors, models, or investment results.
• HKUST Business School’s “Machine Learning: The Best Investment Adviser?” summarizes research by Yang Ha (Tony) Cho, Xi Chen, Yiwei Dou, and Baruch Lev on earnings-change prediction. The public account names random forest and stochastic gradient boosting, detailed financial data, and comparisons with regression and analyst forecasts; it links to the full paper. The page is useful for a model/data/author trail and a possible student or research-assistant lineage, but the public summary does not establish a manager connection, production use, or a point-in-time investable backtest.

This tranche adds four different evidence classes: a model-selection benchmark with released code, a family-office principal linked to a financial-data company, a Hong Kong family-office CIO teaching in a portfolio-management programme, and an HKUST earnings-prediction research route. They should not be collapsed into a league table: the evidence concerns research design, public affiliations, or educational access, with deployment and performance boundaries left open.

Recovery queue for the reproducible-model and personnel routes

Retrieve the UGA paper, released code, appendices, and author profiles; inspect ARTD.AI’s product pages and public technical material; recover HKU ICB family-office programme events and Mui’s public work; and follow the HKUST paper to its full text, data description, code, citations, and student/researcher network. Record only direct affiliations and clearly labeled paper results.

September 5, 2026 — market-information leakage and alternative-investment systems

• Columbia Law School’s Joshua Mitts profile adds a law-and-finance research route that is relevant to information leakage. The profile identifies a PhD in Finance and Economics from Columbia Business School and says Mitts uses statistical analysis and machine learning to study short selling, securities lending, informed trading around cybersecurity breaches, information leakage, hedge-fund activism, insider trading on corporate disclosures, and information transmission. It also identifies his Columbia Data Science Institute affiliation and a course called Data and Predictive Coding for Lawyers. This is academic and regulatory-market research, not evidence of a fund’s proprietary surveillance system or deployment.
• Princeton’s John Mulvey profile connects alternative-investment research to operational systems. Princeton identifies him as a professor of Operations Research and Financial Engineering and founding member of the Bendheim Center for Finance; the page describes work on hedge funds, private equity, venture capital, commodities, dynamic investment strategies, large-scale optimization, integrated risk-management systems, and a graphical network approach using unsupervised learning to model stock returns. It also links a public video titled “Python and machine learning for asset management.” The page does not identify a current fund, proprietary dataset, model evaluation, or live trading result.

Together these routes widen the research map beyond alpha prediction: information leakage and market-transmission measurement on one side, and portfolio/risk-system architecture plus unsupervised return modeling on the other. The links are useful for finding papers, course material, public video, graduate lineages, and practitioner contacts; they do not support a ranking or an inference about any firm’s current system.

Recovery queue for market-information and optimization routes

Recover Mitts’s linked papers, CV, course materials, Columbia conference appearances, and Data Science Institute work; retrieve Mulvey’s current-research and publication pages plus the linked video transcript; then trace named coauthors, students, and practitioner collaborators through direct public sources.

September 5, 2026 — hedge-fund benchmark research and quantitative-finance training

• Tengjia Shu and Ashish Tiwari’s SSRN record for “Evaluating Hedge Funds with Machine Learning-Based Benchmarks” is a paper-level extension of the University of Iowa faculty route. SSRN records the paper as posted September 28, 2022 and last revised April 3, 2025. Its abstract says machine-learning algorithms improve the tracking of individual hedge-fund performance, particularly for funds with near-zero explanatory power under conventional multi-factor models; it connects the improved tracking to more precise alpha estimates, fund selection, and failure prediction. The abstract identifies Bayesian ensembles of trees as an important source of improvement and points to nonlinearities, interactions, and time-varying factor exposures. This is a research claim from an academic working paper, not evidence of a tracked firm’s adoption, a production data feed, or live trading performance.
• Tippie’s Master of Finance course listing names “Quantitative Finance and Machine Learning (FIN:9160)” and describes regression and machine-learning applications to derivative, equity, and interest-rate valuation. The programme and investment-management certificate page place that course inside formal finance training routes. The catalogue does not expose the instructor, syllabus, assignments, data, guests, or graduate destinations; it is therefore a programme-discovery lead rather than evidence of a manager pipeline or production use.
• The Data Mining Iowa Group seminar page lists Tengjia Shu presenting “From Stock Return Predictability to Mutual Fund Performance: A Machine-learning Approach.” The page also exposes an interdisciplinary research-methods network spanning finance, business analytics, economics, statistics, and computer science. It currently provides a named speaker and topic but not slides, a transcript, code, data, practitioner attendance, or a firm connection.

This route adds a useful three-layer distinction: a paper that states a hedge-fund measurement method, a formal programme that teaches quantitative finance and ML, and a seminar network that can expose adjacent researchers even when their titles do not contain “AI.” The analytical lead is a testable question about whether flexible, time-varying benchmarks change conclusions about alpha, selection, and failure once point-in-time data and multiple-testing controls are enforced.

Recovery queue for the Iowa research-and-training route

Recover the latest SSRN manuscript, appendices, code and data references; inspect FIN:9160’s syllabus, instructor, assignments, guests, and public student outputs; and capture the DMIG talk materials and related seminar titles. Preserve the difference between an abstract-level academic claim, a taught method, a seminar presentation, and verified manager deployment.

September 5, 2026 — Asia-Pacific finance-research network with alternative-data and hedge-fund tags

• Monash Centre for Financial Studies’ affiliated-faculty page is a useful Australia/Asia-Pacific discovery surface. MCFS says affiliated faculty collaborate with the centre on financial-industry challenges. The roster tags Klaus Ackermann, Simon Angus, and Paul Raschky with alternative data, big data, and machine learning; Andreas Deppeler with financial technology, AI, and blockchain; and Charlie Nave with institutional finance, research and innovation, start-ups, and hedge funds. The page links individual research profiles and, in several cases, LinkedIn. It exposes a personnel and paper-finding network, but not a specific manager partnership, dataset, model, production system, or performance result.

This is a useful regional branch because the page’s tags combine alternative data, ML, institutional finance, and hedge funds without requiring an “AI in investing” title. The next analytical step is paper- and profile-level verification: identify which topics are empirical finance research, which are teaching or practice roles, and which—if any—have a named industry partner or deployable artifact.

Recovery queue for the MCFS route

Open the linked faculty profiles, collect current publications and working papers, inspect MCFS partnership and event pages, and verify current practitioner affiliations. Preserve faculty interest tags, practice appointments, sponsored work, and verified manager deployment as separate evidence states.

September 5, 2026 — FIU graduate finance curriculum with ML and big-data tracks

• Florida International University’s Online Master of Science in Finance curriculum documents a current applied-analytics pathway. The 36-credit, 12-class programme includes core portfolio management and financial risk management, then offers “Machine Learning Applications in Finance,” covering Python, finance theory, and machine-learning applications, alongside “Big Data in Finance,” covering financial-data analytics and turning data into intelligence. The page does not identify instructors, assignments, datasets, student projects, employer outcomes, or a manager connection. It is a finance-programme and talent-discovery route, not evidence of production use.

The useful signal is curricular adjacency: ML and big-data training are placed beside portfolio construction and risk rather than presented as an isolated coding elective. That suggests a route for finding student work and faculty links around implementation, but it does not support a conclusion about the quality or adoption of any method.

Recovery queue for the FIU programme route

Recover the course syllabi, instructors, assignments, project showcases, faculty publications, and alumni/employer destinations. Search for public repositories or talks, and keep course descriptions, student prototypes, and verified firm systems in separate evidence classes.

September 5, 2026 — Stanford investment lab infrastructure and investor-summit bridge

• Stanford GSB’s Real-Time Analysis and Investment Lab (RAIL) is a concrete finance-lab route. Stanford describes 24 workstations designed to reproduce an investment-management environment and identifies Kevin Mak as faculty director. The lab-features page names BarraOne for factor-risk analysis, Portfolio123 for backtesting with frictions such as commissions and market impact, Rotman Interactive Trader for scripted market simulations, and Rotman Portfolio Manager for portfolio simulation using real quotes and benchmarking. Stanford says Mak oversees the student-managed fund; the RAIL faculty page names affiliated faculty using the facility, including Kostas Bimpikis, Peter DeMarzo, Dan Iancu, Charles Lee, Sridhar Narayanan, Joshua Rauh, and Kathryn Shaw. These pages expose infrastructure, tools, and personnel, but not student strategy files, live performance, proprietary model configuration, current vendor contracts, or hedge-fund adoption.
• Stanford GSB’s 2026 Investor Summit adds a current event bridge. The January 24, 2026 summit description says sessions covered AI’s impact on investing, asset-class boundaries, leadership and talent, and digital assets. The public speaker list includes Edwin Jager of D. E. Shaw, Phillip Lee of Citadel, Jackson J. Garton of Makena Capital Management, Raj Agrawal of KKR, Jody Jonsson of Capital Group, and Stanford faculty. This is a public convening and personnel-discovery surface; the page does not provide the speakers’ remarks, recordings, internal tools, datasets, or firm-wide conclusions.

RAIL is valuable because it exposes the operational middle layer between a finance syllabus and a manager: risk analysis, backtesting with explicit frictions, simulated market interaction, portfolio administration, and a named faculty network. The summit is a separate event layer. Neither should be read as proof that an academic workflow or a speaker’s firm uses a particular AI system.

Recovery queue for Stanford lab and event routes

Recover RAIL’s “Building Your Own Lab” guide, current course materials, student-managed-fund outputs, software versions, and public alumni/employer links. Recover the Investor Summit agenda, recordings, transcripts, and follow-up posts, then search each named participant independently while preserving event participation, teaching, research, and verified deployment as separate evidence states.

September 5, 2026 — South Dakota and Boston routes: e-trading labs, alternative data, and finance talent

• South Dakota State’s profile for Zhiguang (Gerald) Wang exposes a combined faculty, lab, and student-fund route. SDSU identifies Wang as DuBois Professor of Business Finance and Investments, coordinator of the First Dakota National Bank eTrading Education Lab, and teacher of the student-managed investment fund. His public research interests include machine learning, market microstructure, volatility, and derivatives; the profile lists a 2025 paper, “Machine Learning for Stock Return Prediction: Transformers or Simple Neural Networks,” and a 2022 paper with a graduate student on multistep deep-learning forecasts of the implied-volatility surface. The institutional symposium record names an Autoformer transformer and simple-neural-network comparison for U.S. stock returns by Wang and two students. SDSU’s research account reports the study’s result as transformer performance exceeding simple neural networks at one-, three-, and twelve-month horizons. These are academic and educational signals; they do not establish the fund’s live use, code, data rights, or performance.
• Faraz Moghimi’s UMass Boston dissertation adds an explicit research lineage and feature-design route. The December 2023 dissertation, “Contemporary Empirical Asset Pricing: Alternative Big Data and Machine Learning Models,” names Chi C. W. Wan as first advisor, Rui R. L. Li as second advisor, and Farid F. K. Khosravikia and Safer Y. Yuksel as additional committee members. Its abstract describes three essays: loss-function design for finance ML, an interpretable LSTM for temporal dependencies, and a dynamic measure of technology-skill demand in firm hiring profiles for future-return prediction. UMass’s public placement document lists Moghimi’s initial 2023 placement as Machine Learning Scientist at Zillow. This establishes academic training, a concrete feature family, and a public career path; it does not establish a hedge-fund affiliation, production model, code release, or investable performance.
• UMass Boston’s MS Finance curriculum places Machine Learning in Finance (MBA AF 642) alongside the Asset Management Practicum capstone, portfolio analysis and investment management, applied econometrics, and an industry research practicum. The College of Management’s finance page also describes a student-managed fund, Bloomberg certification, industry speakers, and data-science courses in NLP, machine learning, and textual analysis. This is a Boston talent and workflow route. The public pages do not expose syllabi, models, data, student outputs, or employer outcomes, so they are not evidence of manager deployment.

The new regional signal is a continuum from market simulation and student capital, through alternative-data research and interpretable sequence models, to graduate practicum training. The relevant follow-up is artifact recovery: exact papers, code, data boundaries, student-fund mandates, supervisors, and employment trails. Academic prototypes and curricula remain separate from live-manager evidence.

Recovery queue for the SDSU and UMass Boston routes

Recover Wang’s 2025 paper, code, data, and appendices; the implied-volatility and student-managed-fund papers; and the e-trading lab’s tools and events. Download Moghimi’s dissertation, map the supervisor/coauthor network, and recover UMass ML/practicum syllabi, student-fund material, Bloomberg challenge artifacts, and public graduate destinations. Test any predictive claim only with point-in-time inputs, explicit horizons, transaction costs, and held-out evaluation.

September 5, 2026 — Monash and Columbia research-lab routes for portfolio learning

• Monash’s Centre for Quantitative Finance and Investment Strategies describes work spanning stochastic control, asset pricing, financial econometrics, machine learning, and data science. Its public 2026 archive names a Q-Group Colloquium and new PhD students with supervisors; the archive also records BNP Paribas affiliates and a ClimateWorks Australia / BNP Paribas / ISS ESG / Monash index project. It identifies a Head of Quantitative Solutions at Cbus Super Fund as an adjunct associate professor. The related 2019 colloquium page lists practitioners from VFMC, CSIRO Data61, NAB Asset Management, Omega Global Investors, and Research Affiliates alongside academic and PhD sessions. This is a dated collaboration and talent network, not evidence of any firm’s current model or deployment.
• Columbia FDT Center’s Haoran Wang profile gives the portfolio-learning route a named researcher and lineage. Columbia identifies Wang as a postdoctoral research scientist in IEOR and the FDT Center; the profile describes continuous-time reinforcement learning, stochastic control, optimization, and a risk-sensitive asset-management agenda. It says Wang and Xunyu Zhou presented an interpretable RL algorithm for continuous-time mean-variance portfolio allocation, and records Wang’s 2018 mathematics PhD from the University of Texas at Austin under Thaleia Zariphopoulou. The profile does not expose code, data, independent benchmarks, fund adoption, or investment results.
• The FDT Center’s About page says the centre was established in 2017 with support from financial-investment entrepreneur Bill Nie and is a sister centre of Oxford’s Oxford-Nie Financial Big Data Lab. It describes research using reinforcement learning, big-data analytics, portfolio theory, behavioral finance, machine learning, and data science for asset allocation and risk management. It says the centre has developed continuous-time RL theory and interpretable algorithms, and has explored diffusion-model extensions for potential data and strategy generation. These are centre-described research directions and claims; independent validation, code, data rights, and firm deployment remain unresolved.

The important cross-cohort clue is the handoff from academic methods to industry-facing networks: Monash’s public record exposes bank, superannuation, ESG-index, and quant-colloquium interfaces; Columbia exposes a named RL researcher, supervisor lineage, and a centre whose stated agenda reaches from stochastic control to generative-model research. The correct next step is artifact-level recovery and reproduction, not a ranking of methods or firms.

Recovery queue for Monash and FDT routes

Recover Monash’s current people, papers, PhD theses, colloquium programmes, recordings, and BNP/VFMC/Cbus project artefacts. Recover Wang/Zhou papers, appendices, code, data, and doctoral lineage; map FDT/Oxford-Nie people and publications; and test any portfolio claim with point-in-time data, costs, risk constraints, and held-out periods.

September 5, 2026 — CMU computational-finance lab and industry-data pathway

• Carnegie Mellon’s 2026–27 MSCF announcement describes an applied-research lab built around industry datasets, competitions, and practitioner engagement. The page says students and faculty will work on capstone and applied-research projects tied to real market questions, while industry partners may contribute data, project themes, and practitioner expertise. It also describes a second year of partnership with CMU’s Computer Science Department and an expansion in deep learning, algorithms, and data-intensive finance, alongside additional real-world datasets and new coursework in blockchain, cryptocurrency, and digital assets. This creates a concrete academic-to-industry discovery surface, but the announcement does not name the participating firms, expose data rights, publish project outputs, identify model configurations, or establish production deployment.

The useful distinction is between a programme’s engagement contract and its research output. The former is public here; the latter remains a recovery task. Partner briefs, capstone reports, competition notebooks, and alumni profiles may expose model and data clues, but they must be labeled as student, sponsored, or production evidence separately.

Recovery queue for the CMU MSCF route

Recover the applied-research lab page, partner roster, project briefs, capstone outputs, course syllabi, CS partnership details, competition artefacts, and public alumni/researcher profiles. Preserve intended programme structure separately from completed projects and verified manager deployment.

September 5, 2026 — professor-led finance-AI routes and programme talent signals

• Papa Momar Ndiaye’s Stevens profile is a direct, named academic-to-practice route. Stevens identifies Ndiaye as a Teaching Associate Professor and describes research spanning robust optimization, dynamic systems, machine learning, risk modelling, signal generation, and portfolio allocation. The profile specifically names reinforcement learning for robust covariance matrices incorporating alternative data, LLM extraction from earnings calls paired with deep-learning volatility prediction, and semidefinite optimization plus clustering for factor portfolios and covariance matrices. It also lists his CEO role at Aleph1 Portfolio, a 2024 ACM AI-in-Finance paper on earnings-call signal extraction, a 2024 paper on reinforcement learning and text-based networks for correlation estimation, and courses in portfolio theory, asset allocation, stochastic calculus, and market microstructure. These are disclosed research and professional affiliations; they do not establish that a tracked hedge fund uses those methods, or that the published signals survive independent point-in-time testing.
• Zhiyuan Yao’s Stevens doctoral profile exposes a more operational research agenda. The dissertation summary describes deep learning and RL for trading and execution, with communication latency and realistic market simulation as explicit problems. It proposes model-based RL for delayed feedback, an agent-based simulator with heterogeneous agent utilities and flash-crash shocks, and a hierarchical system joining portfolio selection with order execution. The profile names Ionut Florescu and Chihoon Lee as academic advisors. This is useful because it identifies simulator fidelity, feedback delay, and the portfolio-to-execution interface as research objects; it is not evidence of a production simulator, a fund affiliation, or live trading permission.
• Rutgers’ MQF Industry Board is a programme-level personnel map rather than a hedge-fund disclosure. It names practitioners including Ratul Ahmed, Gary Ang, Justin Xu, Alexander Fleiss, Lukasz Szpruch, Giovanni Beliossi, and others. The page describes Justin Xu as Chief Quant and AI Officer at MillTech, with stated responsibility for AI strategy, quantitative research, FX hedging, risk management, and production AI/ML systems across portfolio construction and enterprise decisions. It identifies Szpruch’s remit around machine learning, RL, GenAI, agentic systems, model-risk governance, testing, and validation, and describes Beliossi’s systematic-investment path through FGS Capital, First Quadrant, Auriel, Astarte, and Axyon AI. The page is a public advisory-board description; titles and biographies are evidence of roles and networks, not evidence of specific models, data, or deployment at any tracked hedge fund.
• Adam Tashman’s University of Virginia profile joins a finance programme, research, and industry lineage. UVA lists him as Associate Professor of Data Science and Director of the Capstone Program, teaching reinforcement learning and mathematical finance while researching quantitative finance, computer vision, and LLMs. His prior roles include quantitative research at Citi and the Royal Bank of Scotland, senior risk consulting at Citizens Financial Group, and data-science leadership at FinMason, Carpe Data, CliniComp, and AWS. His public publications include stressed-beta option pricing and portfolio optimization, alongside recent data-science and LLM-evaluation work. This is a candidate network for further personnel and student-project discovery; it does not show a current hedge-fund role or a deployed finance model.
• S. Jaimungal’s University of Toronto research page provides a Canadian academic route with unusually explicit algorithmic-trading coverage. The page lists high-frequency algorithmic trading, stochastic control, mean-field games, and machine learning in finance; current student topics include FX markets, algorithmic trading, stochastic portfolio theory, and partial information; and it names postdoctoral researcher Omid Namvar Gharehshiran’s focus as reinforcement learning in algorithmic trading. The page also links a public code site for Fourier-based option methods and papers on ambiguity-aware execution, latent-alpha learning, order-flow signals, market impact, and robust market making. The source exposes methods, student topics, and research artefacts, not manager adoption or live returns.
• Will Cong’s Princeton finance-AI seminar page records a 2023 talk on “Building AI Models for Finance: Goal-Oriented Search.” The abstract links Transformer-based RL to portfolio management and managerial decision-making, and describes Panel Trees as economically guided, interpretable search for asset sorting, basis portfolios, pricing kernels, grouped heterogeneity, regime switching, and structural breaks. It also identifies Cong’s Cornell FinTech initiative and DEFT Lab, and his Harvard/Stanford/Cornell lineage. The talk’s reported empirical figures are claims in an event abstract and require recovery of the underlying paper, code, sample construction, and out-of-sample protocol before they can be used quantitatively. Nothing on this page establishes deployment by a tracked fund.
• Columbia IEOR’s financial-engineering page adds programme-level breadth around the previously tracked FDT/Nie Center. Columbia explicitly lists RL in stochastic control, mean-field games, systemic risk, quantitative behavioral finance, and principal-agent problems, and names affiliated faculty across machine learning, optimization, and stochastic modelling. The page also says the department uses New York practitioners for seminars, collaborations, and internships, and links the FDT Center as an industry-supported centre. It establishes a research and talent surface; it does not identify a hedge-fund partner, dataset, model owner, or production system.

The cross-programme pattern is specific enough to guide discovery but not to support a league table: finance-AI research is publicly legible around signal extraction, covariance and factor construction, execution under latency and impact, market simulation, RL, interpretable asset sorting, model risk, and governance. The next recovery step is to retrieve the underlying papers, code, syllabi, student projects, seminar recordings, and named alumni/employer trails, then label each artifact as academic, educational, advisory, or verified manager evidence.

Recovery queue for professor and finance-programme routes

Recover Ndiaye’s ECC Analyzer and correlation-estimation papers, appendices, code, and data definitions; Yao’s dissertation and simulator artefacts; Rutgers board members’ primary biographies and any named employer evidence; Tashman’s finance publications and capstone outputs; Jaimungal’s linked datasets/code and student theses; Cong’s P-Tree and Transformer-RL papers; and Columbia IEOR/FDT people, seminars, and project outputs. Search each named person independently across conference agendas, public repositories, employer biographies, and university placement records. Do not convert academic method descriptions into claims about hedge-fund deployment.

September 5, 2026 — finance-programme placement and production-training interfaces

• MIT Sloan’s 2025 Master of Finance Employment Report is a dated talent-flow source relevant to the tracked manager universe. The report says 100% of the class was covered by its reporting and lists quantitative research/data science and quantitative trading/research-sales among the finance functions. Its employer list includes Acadian Asset Management, Balyasny Asset Management, Citadel, Qube Research & Technologies, Squarepoint, Two Sigma, AQR, Man Group, Point72, Verition, and other asset managers, banks, and trading firms; it separately lists internship employers including Acadian, Qube, Squarepoint, and Millennium Advisors. The report establishes recruiting and placement visibility, not AI specialization, team assignment, model ownership, or deployment at any named firm.
• The University of Chicago’s Quantitative Development Certificate shows what a production-oriented finance curriculum makes explicit. The programme says its advisory board includes Millennium, DRW, and Chicago Trading Company; its capstone is presented to industry professionals; and its required choices include machine-learning operations and inference engineering, data pipelines for quantitative research, systematic-trading technologies, low-latency trading systems, and high-performance numerical computing. The page also says submitted resumes may be shared with the board for recruiting. This is a strong implementation and talent-pipeline surface, but the public page does not identify capstone datasets, project results, student placements, or any firm’s internal technology.
• Wharton’s Jacobs MSQF careers and recruiting page describes a practitioner-mediated research loop. It says students work in groups of four to six on research questions overseen by finance faculty, with approximately 16 practitioners posing questions, providing real-world data, and attending final presentations. The page displays logos for firms and market participants including Two Sigma, Millennium, Citadel, Bridgewater, Point72, Optiver, Susquehanna, Vanguard, Capital Group, and others. Those logos and programme claims expose a recruiting and project interface; they do not prove that every displayed firm supplied data, that a project reached production, or that a model generated investment performance.

This programme layer adds a distinct observable: not just what professors research, but how finance employers can shape student projects, access resumes, supply data, and observe presentations. Those interfaces are useful for finding people, project titles, and model/tool vocabulary; they remain separate from verified internal hedge-fund systems.

Recovery queue for placement and production-training routes

Recover MIT’s underlying employment tables and class-year reports; identify which named employers recur across years without inferring team placement; recover Chicago capstone titles, project sponsors, syllabi, and public alumni paths; and recover Wharton’s advisory-board names, project abstracts, seminar recordings, and data-use boundaries. Cross-link each person to public papers, code, conference appearances, and employer biographies, but do not treat a school-employer connection as evidence of deployment.

September 5, 2026 — student funds, academic labs, and executable research exercises

• MIT’s Rakhlin Lab roster is a compact academic-lineage source. It lists current work in statistical learning, reinforcement learning, and theoretical statistics, and names former mentee destinations including Two Sigma, Jump Trading, NYU Courant, the University of Chicago, Boston University, Cornell, NUS, and the University of Hong Kong. The page records co-advising relationships for some mentees. This is useful for mapping researcher lineage and recruiting surfaces; it does not establish that any destination firm uses a particular research method or that the lab’s work was transferred into a trading system.
• Emory’s Blue Eagle Capital gives a detailed student-fund workflow. The programme describes a student-run fund integrated with practicum courses, faculty and professional-portfolio-manager oversight, monthly investment-committee meetings, sector and theme research, a proprietary trade-booking and blotter system, Bloomberg portfolio tracking, exposure analysis, market/sector limits, stop losses, weekly commentary, and monthly attribution reporting. It also publicly identifies student officer roles and responsibilities, including CIO, research, operations, technology, and risk. This is an unusually concrete educational workflow and personnel surface, but it is not evidence of a hedge fund’s internal system or of independently verified returns.
• Florida’s Caimanes Fund account describes a student-managed equity hedge fund created in 2024 with support from Point72 founder Steven Cohen. The 2025 account says students undergo early-morning training, review earnings and industry developments, build models and pitches, use Bloomberg and other news sources, and cover sectors with long/short research. It also records links to a Bridgewater/Metaculus forecasting tournament and student internship outcomes. The public article documents training, capital, and recruiting interfaces; it does not disclose an AI model, proprietary dataset, execution stack, or performance attribution.
• Tulane’s account of the Aaron Selber Jr. Hedge Fund Course documents a concrete student research exercise. The winning Xtract Capital Partners team proposed using a proprietary LLM to analyze changes in 10-K, 10-Q, and proxy-statement language, using similarity/sentiment measures to inform a systematic equity long/short strategy. The university reports a ten-year S&P 100 backtest and a comparison with the Fama–French five-factor model, but the public account does not provide code, feature definitions, timestamp rules, transaction costs, delisting treatment, or the full evaluation notebook. Treat the numerical result as a student/team-reported claim requiring reconstruction, not as validated evidence.

The student-fund layer reveals a different set of implementation clues than professor biographies: role segregation, committee cadence, research-to-blotter handoff, risk limits, publication routines, recruiting signals, and the boundary between an LLM research proposal and a reproducible backtest. These are discovery surfaces for project files and alumni paths, not a ranking of schools or firms.

Recovery queue for student-fund and lab routes

Recover Rakhlin’s publications, co-advisor network, and public placements; Blue Eagle’s officer rosters, weekly commentaries, annual outlooks, and system documentation; Caimanes’ fund charter, student deliverables, mentor identities, and forecasting records; and Tulane’s Xtract proposal, source paper, code, feature timestamps, and full backtest assumptions. Preserve student, academic, and manager evidence as separate classes.

September 5, 2026 — Asia-Pacific practitioner professors and allocator-data programmes

• HKUST’s Keith Law profile is a current Hong Kong practitioner–academic bridge. HKUST identifies Law as an Adjunct Associate Professor and Managing Director of Quantitative Research at Polymer Capital Management, with prior quant roles at Credit Suisse, JPMorgan, Morgan Stanley, and UBS. The page records a UC Berkeley Master of Financial Engineering, a University of Hong Kong PhD in Statistics, interests in portfolio analytics, derivatives, strategy design, and quantitative trading, and publications on pairs trading, tail risk, covariance models, and portfolio-management evaluation. It provides a named current role and research lineage; it does not disclose Polymer’s models, data, AI systems, permissions, or performance.
• The University of Sydney’s Applied Portfolio Management programme describes a two-semester student-managed investment fund using real money and real-time Australian equities. The 2026 programme limits participation to 20 finance undergraduates, assigns sector and functional-team responsibilities, and requires research reports, presentations, portfolio monitoring, quantitative analysis, portfolio construction, and risk management. This is a useful Australian student-capital and talent route; the page does not expose the fund’s code, data vendors, AI use, holdings history, or independently audited performance.
• UNSW’s AI-driven managed-account research project is an allocator-data and financial-infrastructure route rather than a hedge-fund disclosure. The project proposes AI tools to maintain Separately Managed Account Standards, normalize portfolio names, fees, asset allocations, and fiduciary roles, classify holdings with machine learning, and use cleaned SMA data to study future outcomes and post-fee attribution. UNSW names Jerry Parwada as Professor of Finance and Director of Academic Strategy and identifies support from a Fintech AI Innovation Consortium. The page describes a proposed PhD/project structure; it does not disclose the consortium’s members, data-sharing contracts, model architecture, completed results, or deployment.

The regional signal is not a claim about relative sophistication. It is a set of distinct research surfaces: a named quant-research executive teaching at HKUST, real-money student portfolio operations in Sydney, and an Australian project aimed at making managed-account data comparable enough for machine-learning analysis and post-fee evaluation. Those surfaces point to different recovery targets—personnel papers, student-fund records, standards documents, data dictionaries, and project outputs.

Recovery queue for Asia-Pacific routes

Recover Law’s current employer biography, papers, seminars, and public code; Sydney’s 2026 SMIF charter, holdings history, student reports, guest speakers, and risk policy; and UNSW’s SMARS documentation, consortium members, PhD proposal, data schema, classification labels, and empirical-study outputs. Search the HKUST faculty network and Sydney/UNSW alumni independently for public practitioner, conference, and employer records while maintaining academic-versus-deployment boundaries.

September 5, 2026 — finance curricula as model-idea and personnel discovery surfaces

• Fordham’s Finance and Business Economics catalogue describes FNBU 4430, Computational Finance, as a Python-based course using alternative financial data, textual analysis, tone and sentiment analysis, topic modelling, and neural networks to examine investment decisions. The same catalogue lists FNBU 4431, a Student Managed Investment Fund: ESG-Impact, in which students analyse and select a real set of securities and opportunities. Fordham’s Fintech concentration also places computational finance beside algorithmic trading, advanced Python with AI, and machine learning for business. These pages expose curriculum and student-fund routes, not a named hedge-fund system, dataset licence, model owner, or live performance.
• Penn State’s Finance catalogue provides an unusually direct adjacency between finance-AI training and a real student portfolio. FIN 465, Data Science and Artificial Intelligence in Finance, specifies Python, SQL, financial-data handling, visualisation, statistical analysis, machine learning, LLMs, and prompt engineering. The catalogue separately describes the Intrieri Family Student Managed Fund as an actual fund with an approximately $250,000 stock portfolio, and FIN 462 as its officer/lead-analyst practicum. The catalogue does not say that FIN 465 supplies that fund, nor does it expose the course’s assignments, fund code, model permissions, or performance attribution.
• Cornell’s Milstein student profiles expose a dated student-talent route that a hedge-fund-name search would miss. The page’s Class of 2022 cohort includes a computer-science student describing algorithmic-trading-bot programming for Cayuga Capital, a student-managed hedge fund, and another profile describing interests in machine learning and data analytics. This is self-described student activity in a university profile, not evidence of a current employee, a production trading system, a fund mandate, or a particular AI method. The route is still useful for recovering student-fund rosters, project artefacts, alumni paths, and faculty or club connections.
• Trinity University’s Finance and Business Analytics department page combines finance, business analytics, machine learning, and data-driven decision-making in its public programme description. It says students can manage a real portfolio through a Student Managed Fund, work on live data projects with companies, and pursue internships with firms including JPMorgan Chase, Dell, USAA, and PwC Advisory. This is a school-described training and employer-interface surface; it does not identify the fund’s holdings, model stack, data suppliers, named student projects, or any employer’s internal AI use.
• Singapore’s Wealth Management Institute faculty page identifies Dr Anser Aftab as Principal Program Director and lead faculty for the Dalio Market Principles programme, with a PhD and MPhil from Cambridge, a prior senior leadership role at Bridgewater Associates, and stated expertise that includes machine-learning applications in investing. WMI’s programme-and-research-leadership page adds that Aftab’s former Bridgewater remit included political-economy and policy research, currency and emerging-market signals, AI-driven research initiatives, client-facing publications, and hiring/training responsibilities. Separately, the Dalio Market Principles course page publicly describes a proprietary portfolio simulator, case studies, and a generative-AI tutor. These pages establish a teaching, personnel, and product surface. They do not establish current Bridgewater systems, model architecture, data rights, trading authority, or performance; WMI’s promotional size/ranking language is not used here.
• San Francisco State’s finance bulletin adds a lower-specificity but useful West Coast curriculum route. The 2026–27 bulletin lists FinTech II covering DeFi, business blockchain, and machine learning; Financial Analytics; Mathematics and Programming for Financial Analytics; and a Student Managed Fund in ESG Investments, alongside alternative investment and financial-risk courses. It also links separate certificates in financial analytics, fintech, and AI for business technologies. The bulletin is evidence of course availability and talent discovery only; it does not disclose instructors, assignments, fund holdings, model use, or manager adoption.

The new programme evidence adds a practical discovery rule: search for courses and student funds where the technical vocabulary is embedded in finance, analytics, fintech, risk, or portfolio-management pages rather than in an “AI” programme title. The public clues are the interfaces—Python/SQL and LLM coursework beside a real fund, computational finance beside textual signals, practitioner-linked teaching, portfolio simulation, and student-managed capital. They remain academic or educational evidence classes and must not be converted into claims about any hedge fund’s internal strategy.

Recovery queue for curriculum and professor-led routes

Recover Fordham’s current instructors, syllabus, projects, and ESG-fund outputs; Penn State FIN 465 materials, Intrieri/Nittany Lion fund records, and faculty affiliations; Cornell Cayuga Capital rosters, alumni destinations, student repositories, and faculty sponsors; Trinity’s Student Managed Fund materials and finance/business-analytics faculty publications; Aftab’s public papers, talks, Bridgewater-authored material, programme recordings, simulator documentation, and WMI family-office programme links; and SFSU course instructors and student-fund outputs. Search the associated professors and student-fund officers across papers, GitHub, conferences, podcasts, and employer biographies. Keep course availability, student activity, faculty research, advisory participation, and verified manager deployment as separate states.

• Bryant University’s 2026–27 Finance catalogue adds a named-instructor, portfolio, and AI-course route. FIN 414, AI Applications in Finance, is listed with H. Kuang and covers coding basics, model APIs, robo-advisors, textual analysis, sentiment analysis, and content classification. The same catalogue describes the Archway Equity Portfolio Management sequence as a student-managed fund with presentations to investment professionals and a Financial Markets Center, and lists a fintech portfolio course funded by Bryant alumni plus an Archway fixed-income portfolio course governed by an investment policy statement. These are explicit educational and student-capital workflows; they do not establish the instructor’s research agenda, the fund’s model use, external data rights, or a hedge-fund deployment.
• Lancaster University’s Investments and Asset Pricing research area adds a UK faculty-and-industry route. The page lists systematic and factor investing, risk, derivatives, mutual-fund and hedge-fund research, and a recent focus on machine learning and AI for information reduction and investment-signal extraction. It names faculty attached to the Financial Econometrics, Asset Markets and Macroeconomic Policy centre and says the group maintains industry networks including Invesco, Quoniam, and Robeco. This is an institutional research and partnership surface; it does not identify a partner’s model, data rights, project results, or deployment.

September 5, 2026 — Canadian and UK academic routes with data, model, and family-office clues

• UBC Sauder’s COMM 486I: Applications of AI in Finance is a dated 2026 course-and-industry interface. The May 7 information-session deck names Jan Bena, Jose Pizarro, and James Shou; requires teams to source live market data and select AI models; and lists decision trees, CNNs, vision transformers, NLP, investment strategies, sentiment analysis, and forecasting as project areas. It specifies interim research workshops, final reports/slides/code, and a capstone presentation with Connor, Clark & Lunn Investment Management, whose quant team is described as participating in the summer camp. This is unusually concrete curriculum and sponsor evidence, but it does not expose student project results, CCL systems, data rights, or deployment.
• CIRANO’s David Ardia profile adds a Montreal research and reproducibility route. CIRANO identifies Ardia as a researcher since 2025 and full professor at HEC Montréal, with interests in quantitative risk management, quantitative asset management, sentometrics, and ML/NLP methods for finance and economic forecasting. The page records a 2025 research-impact award, a reproducible-finance award, and an April 2026 event on asset pricing and machine learning. This is evidence of an academic and reproducibility network, not of a named manager’s model, data access, or deployment.
• King’s College London’s Marie Grunert profile adds a non-model family-office route that is relevant to automation questions. Her PhD project studies how financial professionals collaborate, generate ideas, and make investment decisions; her teaching includes “AI at Work,” focused on AI/ML across professions and expertise; and her prior experience includes political-risk analysis in a family office. The page names Alex Preda and Crawford Spence as supervisors. This is organizational and personnel evidence for studying human–AI workflow boundaries in asset management and family offices; it does not establish a family office’s AI system or investment process.

September 5, 2026 — New York and London programmes connecting AI coursework to investment practice

• NYU’s Finance catalogue exposes a particularly useful finance-programme stack. FINC-UB 54, Data Driven Investing with AI and Python, is described as hands-on work with real financial-market data, portfolio construction, and strategy performance. FINC-UB 55, AI in Finance, covers generative AI in financial services, investment management, financial technology, financial analysis, risk modelling, and ethical governance. The same catalogue includes Alternative Investments I, which treats hedge-fund selection from an asset-manager perspective including pensions, endowments, family offices, and funds of funds; Hedge Fund Strategies, which uses real data and case studies across event-driven, equity, debt, FX, cross-market, global-macro, and activist strategies; and Research on Wall Street, which covers alternative data and the challenge posed by AI. These are course descriptions, not evidence of a named manager’s models, assignments, data, or deployment.
• Bayes Business School’s Finance (Investment Management) MSc adds a 2026/27 London talent and sponsor surface. The programme makes Data Analytics with AI for Finance compulsory, places it beside asset management, alternative investments, derivatives, and risk management, and offers Machine Learning and AI for Finance as an option. It describes industry projects on live challenges, examples involving Deutsche Bank, guest speakers from BlackRock, Sterling Investments, and UBS, simulations for sales and trading and asset management, an asset-allocation game, a research-project option, and company visits including Tikehau Capital and Allianz Global Investors. The page says specific providers and companies may change; it does not disclose project data, model implementations, sponsor deliverables, or employer deployment.

Together these programmes expose a useful academic-to-industry search surface: course catalogues reveal which investment domains are being paired with AI, while capstones, simulations, industry projects, and family-office-oriented coursework identify where student artifacts, instructors, guest speakers, and employer links may be recoverable. They remain curriculum and recruiting evidence, not a comparison of firms or proof of production use.

Recovery queue for NYU and Bayes programme routes

Recover NYU FINC-UB 54/55, Alternative Investments, Hedge Fund Strategies, and Research on Wall Street instructors, syllabi, project briefs, guest speakers, and student outputs. Recover Bayes module leads, project sponsors, simulation providers, research projects, guest-lecture recordings, company-visit records, and public alumni paths. Keep course descriptions, sponsor participation, student work, and verified manager systems as separate evidence states.

These routes strengthen the academic layer in three different ways: a finance course with live market-data projects and an identified quant-firm participant, a reproducible-finance research network, and a family-office/work-practice researcher whose topic is collaboration rather than model construction. They should be used to discover syllabi, capstone code, sponsor briefs, papers, supervisors, and alumni—not to infer manager deployment.

Recovery queue for Canadian, UK, and family-office academic routes

Recover UBC COMM 486I capstone reports, code, project topics, faculty biographies, CCL participation, and data boundaries; Ardia’s papers, appendices, code, event recording, and HEC/CIRANO collaborators; and Grunert’s dissertation, interviews or publications, supervisor network, AI-at-Work teaching materials, and family-office practitioner links. Track live-data coursework, sponsor participation, reproducibility evidence, and family-office employment as separate evidence states.

September 5, 2026 — professor-led research labs and finance-programme pipelines

• MIT’s 15.S06 AI and Machine Learning Research in Finance, maintained by Hui Chen, is a particularly concrete professor-to-artifact route. MIT describes a project-based semester with paper presentations, industry guest lectures, faculty meetings, team research, and a final paper. The public project page lists an embedding-based political-risk measure built from S&P 500 earnings calls, a simulated limit-order-book market with reinforcement-learning agents, and interpretability work for financial ML. The page reports research-level evaluation details for the displayed projects, including out-of-sample volatility-forecast comparisons for the political-risk measure. This is student research and a faculty-run course; it does not establish a manager’s use of the methods, proprietary data access, production permissions, or live returns.
• Rensselaer Polytechnic Institute’s Finance, Markets, and Emerging Technologies B.S. exposes a talent pipeline where finance, technology, and applied capital management are intentionally joined. The programme page lists financial modelling, portfolio management, emerging financial technologies, fintech and algorithmic trading, data analytics, Bloomberg access, simulation labs, AI-in-markets research, and the James Student-Managed Investment Fund. It also names internships and industry-facing presentations as programme features. These are institutional programme claims and student-capital infrastructure, not evidence that the fund uses AI or that an employer adopted a student method.
• Stevens Institute of Technology’s CRAFT Fall 2025 Industry Advisory Board meeting provides a rare public agenda joining professors, financial-services practitioners, and applied research proposals. CRAFT identifies Steve Yang as director and Mohammed Zaki as site director, with Stevens and RPI faculty in the research network. The public agenda lists projects on financial multimodal LLMs using semantic GraphRAG, privacy-preserving synthetic data for financial ML, causal safety testing for financial foundation models, algorithmic market making for tokenized securities, alternative-data risk, and agentic AI in accounting; it also lists public-industry use-case sessions on FinAgents/FinGPT and deployment of AI agents in data science. Some sessions are labelled closed, and the page does not publish their underlying artefacts. The open agenda establishes research themes and named participants, not a hedge-fund system, model ownership, data rights, or investment performance.

The combined signal is a repeatable academic discovery pattern: course pages identify the modelling vocabulary; student funds and simulations reveal where methods may be exercised; and advisory-board agendas reveal named researchers, practitioners, vendors, and project titles. The evidence classes must remain separate. A professor’s paper, a student fund, a sponsored proposal, and a practitioner use case are four different observations, not a chain proving deployment by an investment firm.

Recovery queue for professor and programme routes

Recover the MIT 15.S06 Fall 2026 project PDFs, authors, code, data descriptions, guest speakers, and faculty lineages; RPI’s current course schedule, instructors, student-fund mandate, project outputs, and alumni destinations; and CRAFT’s linked bios, project abstracts, papers, recordings, partner disclosures, and public follow-up. Preserve the distinction between a named professor, a course, a student-managed portfolio, a sponsored research proposal, and a verified manager system.

September 5, 2026 — Boston and Harvard finance faculty routes

• Harvard Division of Continuing Education’s Machine Learning Techniques in Economics and Finance lists Sudhakar Raju, Professor of Finance and Analytics & Technology at Rockhurst University, as the 2026 instructor. The course description covers regression, regularization, classification, portfolio optimisation, and other machine-learning techniques for economics and finance. Harvard’s instructor biography adds a public route through quantitative-methods teaching at Harvard Kennedy School and consulting work involving the World Bank, Chicago Board of Trade, exchanges, and central banks. This is course and biography evidence; it does not establish a fund relationship, a proprietary dataset, or a live investment system.
• Harvard DCE’s Financial Statement Analysis course page lists Surjit Tinaikar, Chair and Associate Professor of Accounting and Finance at UMass Boston, as an instructor for a 2027 section. Tinaikar’s public biography identifies a University of Toronto doctorate, Boston College finance training, CFA qualification, financial-markets-regulation experience, and research in fundamentals valuation, quant-based investment strategies, and financial reporting. The biography specifically says his current research uses machine-learning algorithms to process financial reports for future-stock-return prediction. This is a self-described academic research direction and Boston personnel lineage; the page does not name a production model, data vendor, fund, or independently verified result.

These pages add an important Boston search rule: faculty and programme pages can expose model-relevant research vocabulary—regularisation, portfolio optimisation, financial-report processing, and return prediction—even when neither the course title nor the professor’s title says “AI strategy.” The next step is paper-level recovery and student-artifact discovery, not an inference that the associated institutions or any manager use these methods.

Recovery queue for Harvard/Boston faculty routes

Recover Raju’s current course syllabus, datasets, assignments, papers, and public project outputs; Tinaikar’s papers, CV, coauthors, UMass Boston courses, students, and conference appearances; and any public employer or practitioner links. Keep course teaching, faculty research, consulting history, and verified investment-firm deployment as separate evidence states.

September 5, 2026 — sponsored academic research and quant-talent programmes

• IIT Bombay’s S. Baskar profile provides a rare first-party research-sponsorship disclosure. The profile lists “Deep Learning for American Puts: A Neural Approach to Optimal Stopping and Option Pricing” as a project running from March 2026, with Baskar as principal investigator and professors Suresh Kumar and Harsha Hutridurga as co-PIs; it identifies Citadel Securities Quantitative Research Lab as the sponsor. The same page connects the project to mathematical finance, deep learning for hedging and portfolio optimisation, and a July 2026 workshop on quantitative finance from stochastic models to deep learning. This is a self-described sponsored academic project and a named research interface. It does not disclose the research contract, data, code, model results, transfer rights, or Citadel Securities production use.
• Boston University Questrom’s MS in Mathematical Finance and Financial Technology curriculum describes a 33-credit STEM programme with Python for quantitative finance, portfolio theory, option pricing, machine learning and AI in investing, NLP, alternative data, advanced machine learning for finance, deep learning, statistical learning, high-frequency trading, and a required internship or corporate/innovation project. The page also places AI electives alongside portfolio construction and market-risk management. It is a programme and recruiting interface; the public page does not identify project sponsors, datasets, student outputs, or employer deployments.
• NC State’s Financial Mathematics career-services page publishes unusually specific industry-project examples. It describes 10–12-week industry-originated projects supervised by industry mentors and gives examples involving mortgage models, hazard-rate and LGD models, NLP sentiment analysis sponsored by Wells Fargo, algorithmic trading and durability of published financial modelling sponsored by QMS Capital Management, and model development and validation. It separately lists student projects on stock-price forecasting with ML/time series/deep learning, neural networks for financial data, algorithmic trading, climate risk, and financial contagion. These are programme-described training and sponsor interfaces; they do not establish that a sponsor adopted a student project or that any listed method produced investment returns.

This tranche adds a useful evidence distinction to the academic map: a named corporate sponsor is stronger than a generic course label for establishing an external research interface, but it still falls short of evidence about ownership, access, transfer, deployment, or performance. Programme project pages are also a route to model families and data domains—options, mortgage credit, sentiment, execution, climate risk, and contagion—without assuming that the student or academic implementation crossed into a live portfolio.

Recovery queue for sponsored research and quant programmes

Recover Baskar’s project proposal, grant or sponsor documentation, co-PI profiles, workshop recording, papers, code, and any public results; BU’s module leads, syllabi, project sponsors, student outputs, and alumni paths; and NC State’s current project presentations, sponsor identities, mentor biographies, and public artefacts. Preserve sponsorship, academic output, student project, employer adoption, and production deployment as separate evidence states.

September 5, 2026 — institutional research studios and practitioner faculty

• Princeton ORFE’s undergraduate academic guide exposes a dense faculty-and-studio route rather than a single “AI in finance” course. The department describes financial engineering alongside machine learning, optimisation, statistics, and probability; its faculty list includes Jianqing Fan’s high-dimensional statistics and financial-ML interests, John Mulvey’s asset-management and hedge-fund optimisation work, and Bartolomeo Stellato’s data-driven optimisation and ML interests. The Financial Econometrics Studio lists derivative valuation, portfolio allocation, risk modelling, volatility estimation, financial-data analysis, and financial-system simulation. This is a public research and talent map; it does not identify a fund sponsor, proprietary dataset, production model, or live performance.
• Strathclyde’s PhD/MPhil Accounting & Finance research page provides a UK doctoral-supervisor map with unusually specific topics. It lists hedge-fund and mutual-fund performance research, applied machine learning in finance, sentiment analysis, investor behaviour on social media, activist short sellers, and Mark Cummins’s stated interests in multimodal generative AI, explainable AI, agentic AI, Earth-observation data, model-risk management, and multiple-comparison correction. The page is useful for finding supervisors, students, papers, and project proposals; it does not establish a manager’s use of any method or access to the listed data domains.
• Sasin School of Management’s profile for Chonawee Supatgiat adds a practitioner-to-academia lineage. Sasin identifies Supatgiat as an Assistant Professor in finance and describes prior work as a quantitative analyst developing equity and commodity strategies at Thales Fund Management, research staff at IBM Zurich Research Lab, and quantitative roles at Tractebel/Suez, RWE Trading, Enron, and Esso across asset optimisation, valuation, hedging, risk management, portfolio optimisation, and energy-commodity trading. The profile is high-value personnel and career evidence, but it does not disclose any employer’s current model stack, data rights, permissions, or performance.

The cross-institution signal is a set of different research surfaces: Princeton exposes studios and faculty clusters; Strathclyde exposes doctoral supervision topics and model-risk vocabulary; Sasin exposes a former practitioner’s path across hedge-fund, research-lab, and commodity-trading contexts. These routes should drive paper, syllabus, student, and personnel recovery—not claims that an academic topic is deployed by a manager.

Recovery queue for studio and practitioner-faculty routes

Recover Princeton’s current faculty pages, studio rosters, senior theses, seminars, and public student artefacts; Strathclyde supervisors, student projects, papers, datasets, and recordings; and Supatgiat’s publications, CV, research collaborators, former-employer affiliations, and teaching materials. Preserve faculty interest, doctoral project, practitioner history, sponsored research, and verified deployment as separate evidence states.

September 5, 2026 — Singapore finance programmes and professor-led portfolio research

• SMU Academy’s Practical Machine Learning for Portfolio and Investment Management is a short, employer-sponsorable route aimed explicitly at portfolio managers, finance data scientists, investment professionals, and data analysts. The page describes hands-on work with regression, clustering, neural networks, model comparison, model interpretation, and financial predictive modelling. It names Liu Peng as the trainer and identifies him as an SMU Assistant Professor of Quantitative Finance (Practice), with an NUS PhD in Statistics and Data Science, an NUS M.S. in Business Analytics, and stated research interests in deep learning, sparse estimation, Bayesian optimisation, portfolio optimisation, and risk management. This is a public executive-education and personnel signal; the page does not identify participating firms, datasets, student projects, or production use.
• NTU Nanyang Business School’s MSc Finance programme overview places “Machine Learning in Finance” in a Fintech elective track alongside blockchain systems. NTU says the course covers artificial neural networks, decision trees, and support-vector machines, with a group project using realistic data-analysis problems; the wider programme also includes Python preparation and data-science training. This is a formal curriculum and recruiting route for Singapore and Asia-Pacific finance talent. It does not show the identity of the course instructor, project data, employer sponsor, student outputs, or a manager’s adoption of any method.
• NTU’s FlexiMasters in Business and Financial Analytics adds a continuing-education route aimed at the bridge between traditional finance and data science. Its public course descriptions cover machine-learning capability building, AI and analytics in enterprise, credit-risk assessment, regulatory compliance, financial modelling, portfolio management, predictive-risk analysis, simulation, and the design of AI/data-analytics solutions with ethical, regulatory, and operational considerations. The programme is evidence of an applied professional-training surface, not evidence that a named hedge fund provided data or deployed these systems.
• SUSS’s FIN525 course page exposes a six-session applied curriculum for mid-career finance professionals. It covers k-means, logistic regression, SVMs, ANNs, CNNs, time-series/cross-sectional/text data, risk forecasting, multilayer-neural-network stock prediction, annual-report/news/white-paper text, sentiment and topic modelling, ML portfolio analysis, and Python ML/DL implementation. The page names Dr Wang Zhiyuan as trainer and lists NUS and NTU degrees. This is a course and talent signal; it does not identify a fund, data licence, student output, or production model.
• SUSS’s Dr Ren Jing profile adds a professor-led research route that is easy to miss in a hedge-fund-name search. SUSS lists Ren’s 2026 paper on reinforcement-learning-guided NSGA-II for NASDAQ portfolio optimisation, a 2025 paper on GameFi text mining, and a 2025 book chapter using LSTM analysis of music-streaming chart content for stock-price prediction. The profile also records a 2018 PhD from Singapore Management University, a 2015–2016 exchange at Carnegie Mellon’s Heinz College, research interests in fintech and recommendation, and teaching in machine learning, predictive modelling, applied analytics, and FIN525. These are academic publications and training-lineage signals; they do not establish live trading, hedge-fund sponsorship, proprietary data, or independently verified performance.

The Singapore additions make the academic layer more operationally legible without turning it into a league table. One route is short employer-sponsored training for portfolio and risk practitioners; one is a formal finance degree with a realistic-data ML project; one is continuing education around AI, credit risk, modelling, and operational controls; and one is a named faculty publication connecting reinforcement learning to multi-objective portfolio optimisation. The next evidence step is artifact recovery—syllabi, instructors, project briefs, code, student work, papers, and alumni paths—while keeping course exposure, academic research, external sponsorship, and verified manager deployment separate.

Recovery queue for Singapore professor and programme routes

Recover Liu Peng’s faculty profile, publications, course exercises, industry background, and any public practitioner events; NTU’s course instructor, syllabus, project outputs, employment paths, and fintech-faculty roster; SUSS FIN525 assignments, datasets, student artefacts, and Wang’s publications; and Ren’s NASDAQ-portfolio paper, code, data, coauthors, doctoral supervisor trail, and public talks. Search these names across finance conferences, university placement pages, public repositories, and employer biographies without treating an academic or classroom method as evidence of hedge-fund deployment.

September 5, 2026 — Australian professor and research-lab routes

• RMIT’s profile for Xiaolu Hu exposes a finance-researcher route with unusually clear practitioner history. RMIT identifies Hu as an Associate Professor in Finance whose research areas include empirical asset pricing, bonds, machine learning, and Chinese capital markets. The profile records prior work as a fixed-income portfolio manager at Guosen Securities and credit-rating analyst at China Chengxin International, plus Tsinghua PBC School of Finance and Central University of Economics and Finance degrees. Her listed projects include network-based statistical learning for portfolio risk management, machine learning in mutual-fund evaluation, corporate-credit assessment using internal and external text, and ESG-disclosure text mining. The page says she is partnered with several financial institutions but does not name them or disclose their data, models, permissions, or deployment.
• UTS’s profile for Mohammad Hadi Sehatpour gives a current doctoral route from market practice into financial ML. UTS lists Sehatpour as a Finance PhD candidate supervised by Christina Nikitopoulos Sklibosios, Kylie-Anne Richards, and Gareth W. Peters. His topics include financial econometrics, time series, interest-rate modelling, financial ML, sustainable finance, and quantitative trading; the profile says he has nearly ten years of financial-market analyst and portfolio-manager experience and is applying statistical ML to green-bond dynamics. It records a master’s in industrial management and operations research from the University of Tehran and a mechanical-engineering bachelor’s from Shiraz University. This is a personnel and research-lineage signal, not evidence of a fund’s model or live strategy.
• Macquarie’s profile for Pavel Shevchenko identifies him as Professor in Actuarial Studies and Business Analytics, Director of the Risk Analytics Lab, and Co-Director of the Centre for Financial Risk. The profile describes long-running financial-risk and industry-commercial work across operational and credit risk, derivatives, commodities, FX, retirement products, extreme events, Monte Carlo methods, stochastic control, and machine learning. It records an M.S. from the Moscow Institute of Physics and Technology, a 1999 PhD from UNSW, and more than 80 journal papers and technical reports as stated on the page. The route is important for risk, portfolio allocation, and model-governance discovery; it does not identify current hedge-fund partners, proprietary data, model configurations, or investment performance.
• Bond University’s profile for Rand Low connects quantitative-finance teaching to disclosed large-institution model work. Bond describes Low’s interests in portfolio optimisation, risk management, systematic trading, multi-asset and commodities strategies, and statistical/ML automation. The profile says he previously worked at Bank of America Merrill Lynch and BlackRock in New York, leading quantitative teams building models for market, credit, and operational risk, securities lending, structured products, and portfolio management, and working on stress testing and model-risk governance. It also lists current research areas including corporate credit ratings, robo-advisers, merger arbitrage, and digital assets. These are biography and research-interest claims; they do not reveal former-employer code, data rights, current deployment, or performance.
• UWA’s profile for Kam Chan is a valuable title-blind route for GenAI and portfolio-data discovery. UWA says Chan designed a Data Analytics in Portfolio Investments micro-credential using Python and real fund-manager data, embedded ML and Bloomberg analytics into finance units, and previously served as Assistant Vice-President in Risk Analytics at United Overseas Bank. His listed research interests include generative AI in finance, textual analysis, information acquisition, political uncertainty, and empirical asset pricing; his qualifications include a PhD in Finance from the University of Queensland. This exposes a named educator, former risk-analytics practitioner, and data-informed curriculum route, but not the identity of the fund-manager data, model weights, permissions, or production use.
• Adelaide University’s profile for Reza Bradrania adds a professor-led AI, asset-pricing, and trading-lab route. Bradrania’s public profile lists AI and ML in finance, empirical asset pricing, behavioural finance, investments, liquidity, and capital markets; a Sydney PhD and Durham finance master’s; and funded collaborations involving CIMA, AFAANZ, and CIFR. It also says he co-founded and co-led the IRESS Trading Lab at UniSA Business and integrated IRESS trading technology into selected finance courses. His publication list includes machine-learning work on IPO pricing and portfolio-policy work on ESG index tracking. These are faculty, funding, publication, and educational-infrastructure signals; they do not establish a manager’s live system or independent performance.

The Australian route adds several implementation layers that a firm-name search misses: fixed-income and credit-rating experience feeding empirical asset-pricing research; doctoral work connecting green bonds and quantitative trading; a named risk-analytics laboratory; former bank and asset-manager model-governance experience; curriculum built on real fund-manager data; and a trading lab embedded in finance education. These layers are useful for finding papers, supervisors, student projects, data descriptions, and practitioner collaborations. They remain distinct from verified hedge-fund deployment and do not support a ranking of firms or methods.

Recovery queue for Australian professor and lab routes

Recover Hu’s listed papers, partner identities, student projects, and Chinese-market datasets; Sehatpour’s dissertation, supervisor publications, prior employers, and green-bond research artefacts; Shevchenko’s Risk Analytics Lab projects, code, industry partners, and conference talks; Low’s publications, former-employer references, RBA/Australian Bond Exchange materials, and model-governance work; Chan’s micro-credential materials, fund-manager data provenance, GenAI papers, and UOB lineage; and Bradrania’s AI/ML papers, IRESS Lab materials, funded-project documents, and student outputs. Preserve faculty research, teaching infrastructure, industry history, funded collaboration, and verified manager deployment as separate evidence states.

September 5, 2026 — European and Chinese finance-AI research networks

• The University of St.Gallen’s Artificial Intelligence in Finance page identifies Lukas Gonon as Assistant Professor in Artificial Intelligence in Finance at FSI-HSG and the School of Computer Science since November 2024. The page links deep learning and quantitative finance to time-series modelling, derivatives pricing and hedging, systemic-risk measures with graph neural networks, financial-time-series generation, deep hedging, stochastic-volatility calibration, and quantum-neural-network research. His listed lineage includes ETH Zurich, LMU Munich, Imperial College London, Sony, and J.P. Morgan. This is a named academic lab/personnel route with a dense publication surface; it does not establish a fund’s model use, proprietary data, or investment performance.
• Oxford Computer Science’s “Machine learning for finance” student-project page is an unusually explicit talent and method signal. Supervised by Mihaela van der Schaar and Edith Elkind, the project proposes ensemble learning and CNNs for spot-price prediction, evaluation on free and proprietary datasets using both profit and risk, and extraction of new “lucky factors” to improve investment strategies. It is an MSc/Part C project description, not a completed result: the page does not identify the student, data owners, code, out-of-sample design, or any manager adoption.
• The Alan Turing Institute’s Machine Learning in Finance programme states an explicit institutional aim of bringing researchers, practitioners, and regulators together around responsible financial AI. It names Blanka Horvath, Antoine Jacquier, and Lukasz Szpruch as organisers and describes work on transparent, reliable, reproducible research, explainability, privacy, accuracy, buy- and sell-side use, risk, and collaborative projects with industry partners. The page embeds an Oxford/Turing uncertainty-and-risk workshop recording. This is a public research-and-convening architecture; it does not reveal partner-specific systems, private datasets, or deployment outcomes.
• Shanghai Jiao Tong’s Quantitative Finance Research Center profile for Yiqing Lin identifies Lin as a tenured Associate Professor, Head of Statistics, and QFRC Executive Director. The profile connects nonlinear expectations, machine learning, financial mathematics, derivative-pricing model uncertainty, optimal transport, and quantitative risk models for high-frequency trading. It also says Lin mentors student quantitative-trading teams, promotes the open-source Fastbox fintech platform, and collaborates with domestic private-equity firms on quantitative-finance talent. These are first-party research, open-source, and collaboration claims; the page does not identify partner firms, data rights, code contents, production systems, or returns.
• Shanghai Advanced Institute of Finance’s profile for Dacheng Xiu provides a high-value China/U.S. academic lineage. SAIF identifies Xiu as a Visiting Professor and Chicago Booth Professor whose research spans financial econometrics, empirical asset pricing, ML in finance, high-dimensional statistics, high-frequency data, and derivatives. His listed publications include “Financial Machine Learning,” “Autoencoder asset pricing models,” “Empirical Asset Pricing via Machine Learning,” “Factor Models, Machine Learning, and Asset Pricing,” “Predicting Returns with Text Data,” and a working paper titled “Expected Returns and Large Language Models.” The profile records a Princeton applied-mathematics PhD/MA and Bendheim Center for Finance connection. These papers map methods and lineage; they do not establish deployment by a named manager.
• Xiamen University’s profile for Fuwei Jiang identifies him as a Nanqiang Distinguished Professor of Finance at the Department of Finance and WISE. The first-party page lists AI for finance, FinTech, machine learning, textual analytics, asset pricing, behavioural finance, and Chinese capital markets as his main research areas, and links his research-paper page and academic citation profiles. This is a Chinese-language/English institutional discovery route and a named professor surface; the page does not expose specific datasets, live systems, employer relationships, or performance.
• EPFL’s Swiss Finance Institute faculty roster exposes a European faculty cluster around quantitative risk and portfolio research. It lists Damir Filipovic’s machine learning in finance and quantitative-risk expertise and Semyon Malamud’s portfolio selection, nonlinear filtering, liquidity, asset prices, and ML-in-finance expertise alongside related finance researchers. This roster is useful for mapping supervisors, seminars, papers, and talent channels; it does not identify a manager’s data, model, deployment, or performance.

This tranche adds a different kind of academic signal: named labs and centres make people, papers, code, seminars, and industry interfaces discoverable even when the investment firm is absent from the title. The visible methods span deep hedging, graph neural networks for systemic risk, synthetic financial time series, text-based return prediction, high-frequency risk, model uncertainty, high-dimensional asset pricing, and classroom factor discovery. Those are research hypotheses and talent routes, not evidence of which manager uses them or whether they work after point-in-time and cost controls.

Recovery queue for European and Chinese academic routes

Recover Gonon’s full CV, code, datasets, student lineage, deep-hedging and financial-time-series artefacts; identify the Oxford project’s student, data sources, final paper, and evaluation design; recover Turing partner and event rosters, recordings, transcripts, and publications; inspect QFRC’s Fastbox repository, student teams, private-equity collaboration disclosures, and papers; recover Xiu’s LLM/text/ML papers, coauthors, code, and seminar recordings; map Jiang’s paper page, collaborators, and Chinese-market datasets; and follow the EPFL faculty roster into current courses, supervisors, and finance-AI publications. Preserve academic, open-source, sponsored, practitioner, and verified manager evidence as separate states.

September 5, 2026 — Canadian finance-AI sponsorship, student labs, and research lineage

• Rotman’s September 3, 2026 announcement of the iA Financial Group Professorship in Quantitative Finance records a new five-year, $600,000 philanthropic commitment from iA Financial Group. The announcement says the professorship will support asset pricing, risk management, portfolio construction, and the integration of AI and ML, and names Redouane Elkamhi as the holder. It also says Elkamhi has worked with pension plans, insurance companies, and global asset managers on investment strategy and total-fund management. This is a named academic sponsorship and practitioner-interface disclosure; it does not identify project data, model code, deliverables, or a live investment system.
• Waterloo’s Computational Finance Project describes a long-running, cross-department research project funded by NSERC and industry participants. The page names past and present sponsors including Scotiabank, Royal Bank of Canada, Bell University Labs, Sun Life, ITO33, Credit Suisse, and Morgan Stanley, and describes collaboration across computer science, finance, and statistics/actuarial science. This creates a concrete Canadian academic-to-industry archive for papers, technical reports, personnel, and data-method clues. The page does not establish which sponsor used which method, the terms of access, or any trading result.
• The University of Toronto’s Quantitative Finance Lab constitution documents a student organization with a stated mandate to train quants through a programme taught by PhD students and industry professionals. It lists empirical asset pricing, algorithmic trading, financial ML, and behavioural finance as research areas and sets out project-manager, mentor, workshop, and showcase processes. This is a student talent and project-discovery route; it does not identify current members, completed research, data rights, or employer adoption.
• University of Toronto PhD student Asic Q. Chen’s public profile records a two-year quantitative-researcher role at TD Asset Management before graduate study at the University of Toronto and the Vector Institute. The profile says Chen’s current work concerns probability-density estimation and generative modelling with probabilistic graphical models, with broader interests in trustworthy AI and ML for quantitative finance; it also links a NeurIPS 2023 paper on structured neural networks for density estimation and causal inference. This is a named personnel, academic-lineage, and former-asset-manager route, not evidence of TD’s current model stack or a live fund system.
• Waterloo Professor Tony Wirjanto’s indexed faculty profile links finance, statistics/actuarial science, and computer science. The public search result describes research interests in AI and deep learning, climate finance and risk, computational and quantitative finance, portfolio optimisation, and financial time series, and lists work on machine-learning proxies for high-dimensional nested simulation. The live page returned a retrieval error in this pass, so this is retained as a recovery lead rather than a fully verified profile claim.

The Canadian evidence adds three distinct layers: a fresh corporate-funded professorship, a multi-sponsor computational-finance project, and student/researcher routes exposing quant training, generative modelling, causal inference, portfolio optimisation, and climate-risk methods. The sponsorship and former-employer links are stronger than a generic course label for proving an external interface, but none establishes ownership, data access, production use, or performance. The appropriate next step is artifact recovery and lineage mapping rather than a firm comparison.

Recovery queue for Canadian academic routes

Recover Elkamhi’s research profile, professorship scope, project outputs, practitioner collaborators, and Rotman AI/FinHub event materials; Waterloo’s current sponsor roster, technical reports, faculty and student pages, and data/code disclosures; QuantLab’s current leadership, workshops, repositories, and alumni paths; Chen’s full CV, TD role details, generative-modelling papers, supervisor and coauthor lineage; and Wirjanto’s faculty page, current publications, nested-simulation work, student theses, and industry interfaces. Preserve philanthropy, sponsorship, student activity, former employment, academic research, and verified manager deployment as separate evidence states.

September 5, 2026 — Florida finance-AI lab, industry summit, and applied training routes

• UCF’s Cuneyt Akcora profile identifies a joint Finance/Computer Science professor who leads UCF’s Financial AI Lab. The profile names temporal graph learning, blockchain systems, and financial networks as research areas; it also says lab students intern at and collaborate with Fairwinds Credit Union and Equifax. UCF’s 2026 Institute of Artificial Intelligence research brochure, whose indexed first-party text was available in this pass but whose PDF fetch timed out, describes a related disclosure-integrity direction: checking filings, earnings releases, MD&A, risk factors, footnotes, and sustainability disclosures for invented facts, numerical inconsistencies, unsupported citations, and abrupt AI-like style shifts. The research idea is concrete—disclosure quality and AI-authorship detection as structured, longitudinal signals—but the brochure does not expose code, labels, precision, data rights, or an investment use. The lab’s academic and industry-interface evidence should not be converted into a claim about any tracked manager.
• UCF’s 2026 FinTech Summit page and its official agenda create a new speaker route. The April 10 programme put “AI in Academic Research and Industry Adoption” beside a “Man & Machine” finance panel and listed Sean Cao of Maryland, UCF’s Cuneyt Akcora, Alejandro Lopez Lira of Florida, David Yermack of NYU, BNY’s Chief Data & AI Officer Sarthak Pattanaik, Fairwinds’ VP of Enterprise AI Matt Cannon, and Voloridge Investment Management CTO Howard Haney. The summit also covered stablecoin fragility and tokenization. This is useful personnel, conference, and vocabulary evidence—especially the direct Voloridge CTO route—but neither the agenda nor the summit page provides a transcript, system design, model ownership, data permissions, or trading results.
• UC San Diego Extended Studies’ Artificial Intelligence for Finance course is a title-specific professional-training route with a future 2026 session listed for October 5–December 5. The course description names time-series forecasting, deep neural networks, NLP, ChatGPT, LLaMA, trading, portfolio management, credit risk, fraud, sentiment analysis, Python, TensorFlow, and real-world financial data. UCSD’s instructor page lists Biljana Aleksic as teaching the course alongside linear algebra, deep-neural-network practicum, probability/statistics, and TensorFlow courses. The same UCSD catalogue exposes a neighboring “Building Agentic AI Systems” course using CrewAI and LangGraph. This is a practical talent and toolchain signal; it is not evidence of a university research result, a hedge-fund system, or production performance.

These routes add three distinct academic signals to the research map. The first is graph-native finance and disclosure integrity: network structure, blockchain activity, and machine-checkable consistency in financial reporting. The second is a public bridge between faculty, enterprise AI officers, and a named quantitative-investment technology leader. The third is a workforce pipeline that pairs classic forecasting and NLP with open-weight language models and agent frameworks. None is a league table. The evidence states remain separate: faculty research, lab activity, course design, conference participation, former or current job title, and verified manager deployment are different claims.

Recovery queue for Florida and West Coast academic routes

Recover Akcora’s full publication and code trail, the Financial AI Lab’s project pages, UCF’s disclosure-hallucination brochure as a stable local mirror, and any public labels or evaluation design. Recover the UCF summit recording, speaker biographies, transcripts, and Voloridge/BNY/Fairwinds follow-up without treating presence on the agenda as a shared system. Recover UCSD’s full syllabus, instructor biography, project requirements, and any public student artefacts; map the adjacent agentic-systems course to instructors and repositories. Keep academic ideas, teaching content, conference appearances, industry interfaces, and manager deployment as separate evidence states.

September 5, 2026 — NYU–Barclays research-data and podcast route

• NYU Stern’s Jeffrey Meli profile identifies Meli as a Clinical Professor of Finance who joined Stern in January 2025 after nearly 25 years in financial-sector research. The profile says he spent ten years as Barclays’ global head of research and launched a Barclays research data-science platform that integrated alternative data and modern data techniques into investment research and content monetization. It also records Princeton mathematics and Chicago Booth finance training. NYU’s Research on Wall Street syllabus supplies a second source for the teaching route: it covers sell-side and buy-side research, fixed income, market structure, technology, regulation, and AI. The profile and syllabus expose a concrete research-operations lineage; they do not expose Barclays’ code, data licenses, model specifications, client access, or investment performance.
• Barclays’ official Flip Side page exposes a previously untracked research-media route with Apple, Spotify, YouTube, RSS, and other distribution links. The show is a recurring debate between Barclays Research analysts; its public episode surfaces include macro, credit, equity, ESG, private-credit, and market-structure topics. A public LinkedIn post by Meli says Barclays had launched a Data Hub to help clients integrate data and analytics into investment processes, with datasets described as relevant to portfolio optimisation, risk models, and pre- and post-trade analytics. The podcast page and post are public media/product evidence, not proof that any tracked hedge fund uses the platform or any specific dataset.

This route changes the search map in two ways. First, a senior research leader’s move into a finance faculty role creates a reliable bridge from a named investment-bank data platform to course content and future student/alumni searches. Second, the podcast’s platform distribution and long episode history make it a useful title-blind source for research vocabulary, guests, and market themes; it should be searched by episode transcript and guest rather than by “AI” in the title. The proper evidence states remain separate: a faculty biography, a course syllabus, a company product statement, a social post, and an episode transcript each support different claims.

Recovery queue for NYU–Barclays route

Recover Meli’s current CV and publications, the Barclays Data Hub product pages and archived client materials, the Flip Side RSS feed and platform episode inventory, transcripts for episodes with data/AI/alternative-investment themes, and Meli’s public talks or guest appearances. Preserve the distinction between Barclays Research media, Barclays client-product claims, faculty teaching, and verified manager deployment.

September 5, 2026 — Stanford trading-model curriculum and Wharton family-office research

• Stanford’s MS&E242 bulletin entry provides a new course-level route with unusually explicit model-family sequencing. “Machine Learning for Algorithmic Trading” is organized around regression for pairs trading and statistical arbitrage, reinforcement learning with a high-frequency-trading emphasis, and diffusion models plus generative AI for financial scenario generation and trading-strategy stress testing. Stanford lists the course within computational and mathematical engineering, data science, and management-science degree paths. The bulletin does not identify the instructor, assignments, datasets, code, student results, or any employer sponsor; it is curriculum evidence, not evidence of a live trading system.
• Wharton’s profile of Raffi Amit identifies him as founder and leader of the Wharton Global Family Alliance, an academic–family-business partnership whose work includes family-office governance, wealth management, succession, and a long-running biannual family-office survey. Knowledge at Wharton’s 2025 report summary describes the survey’s findings around family-office operating scope, next-generation involvement, succession, and comparatively slow AI adoption. This is a useful allocator and family-enterprise research route: it can lead to survey instruments, participating-office interviews, programme speakers, and governance practices. It does not identify a family office’s models, vendors, data rights, investment process, or AI-attributed results.

The Stanford route gives the research map a compact sequence from conventional statistical-arbitrage features to sequential decision-making and synthetic/scenario generation. The Wharton route supplies a separate family-office lens in which AI adoption is an organizational and governance question, not only a forecasting question. These are public education and research interfaces; they should not be converted into manager rankings or deployment claims.

Recovery queue for Stanford and Wharton routes

Recover the Stanford syllabus, instructor, project briefs, student repositories, guest speakers, and any public talks for MS&E242. Recover the underlying Wharton Global Family Alliance survey reports, methodology, fieldwork dates, public participants, family-office programme speakers, and any AI-specific follow-up. Keep course content, survey evidence, family-office practice, and verified manager deployment as separate states.

September 5, 2026 — student-managed capital and industry-connected finance programmes

• Stevens’ Student Managed Investment Fund is a two-semester undergraduate course in which students work as analysts, risk advisers, quants, and portfolio leaders responsible for part of the university endowment. The page says students build factor models, use portfolio-analysis tools, and work in Stevens’ financial-analytics labs; it also describes regular interaction with finance professionals and Goldman Sachs as the fund’s brokerage firm. The selection process, student leadership, risk/modeling/portfolio-allocation roles, and research-to-pitch workflow are public. The page does not disclose holdings history, code, data licences, AI use, or independently audited performance.
• Duke’s Quantitative Finance Concentration provides a public bridge from finance education to named practitioner personnel. The programme combines statistics, econometrics, stochastic calculus, machine learning, AI, and decision optimisation; its algorithmic-trading course describes research with financial datasets and model development for trading and investment strategies. The same page identifies David Ye, Massimo Cutuli, Hengzhong Liu, Nick Alonso, and Kai Cui in teaching or industry-facing roles. Duke describes Cui as Head of Data Science at Neuberger Berman, with prior data-driven long/short work at Point72 and a Duke statistics PhD; it describes Alonso as a quantitative-model leader at PanAgora; and it lists former risk leadership at Citadel Securities and Optiver among the practitioner backgrounds. These are programme-published biographies and teaching interfaces, not evidence that a named employer supplied a course project, transferred a model, or granted trading authority.
• Maryland Smith’s Master of Quantitative Finance creates a second student-capital and curriculum adjacency. The programme says students can manage part of Maryland’s Global Equity Fund with peers, learn to use AI for risk analysis, financial-data analysis, and process automation, and study financial data analytics, machine learning in finance, Python, textual analysis, portfolio management, and quantitative investment. Maryland’s public programme page also describes an experiential project that used housing and macroeconomic data to build a metropolitan housing-risk index. The page does not say that the student fund uses the listed methods, identify fund holdings or model owners, or provide a performance audit.

These routes expose three different kinds of evidence. Stevens makes team roles, endowment responsibility, factor construction, brokerage, and analytics infrastructure visible. Duke makes the practitioner-to-classroom bridge visible, including data science, long/short research, portfolio modelling, and risk leadership. Maryland puts student-managed capital next to explicit AI/ML and textual-analysis coursework. The useful next step is artifact recovery—fund policies, project briefs, instructor materials, student reports, alumni paths, and any public code—while keeping classroom exposure, student activity, practitioner biography, and verified manager deployment separate.

Recovery queue for student-capital routes

Recover Stevens’ current SMIF officers, fund policy, annual reports, holdings disclosures, project presentations, and alumni destinations; Duke’s course syllabi, industry-speaker recordings, project list, speaker biographies, and public student repositories; and Maryland’s Global Equity Fund records, BUFN640/BUFN650 syllabi, experiential-project materials, faculty profiles, and alumni/employer paths. Search each named practitioner independently across papers, podcasts, conference programmes, LinkedIn, and employer pages, without treating a university connection as evidence of a firm’s internal AI system.

September 5, 2026 — French quant education, an Alaska–asset-manager AI interface, and a student fund

• IPAG’s AI-in-finance lab page identifies Hans-Jörg von Mettenheim as head of its Chair of Quantitative Finance and Risk Management and as the person behind QuantIPAG, the school’s student investment fund. The page embeds a 2020 video in which he discusses applications of AI to finance. IPAG’s current profile PDF adds a more detailed academic and publication trail: neural networks, time-series forecasting, algorithmic trading systems, high-performance computing, ESG-news sentiment, cryptocurrency-market deep learning, high-frequency volatility, and neural-network pricing and hedging. It also records his PhD dissertation on advanced neural networks for finance and forecasting. The pages establish a professor, student-fund, video, and publication route; they do not expose QuantIPAG’s holdings, code, data rights, current AI workflow, or independently tested performance.
• The University of Alaska Anchorage’s AI-in-finance webinar page names Anureet Saxena as McKinley Capital Management’s director of quantitative research and describes a UAA–McKinley partnership around AI and data science in finance. The university page attributes to McKinley a workflow of identification, forecasting, and optimisation; it also states that the firm used AI to analyse very large cross-country stock universes. Saxena’s listed lineage includes Carnegie Mellon management science, Purdue economics, and IIT Bombay computer science, with prior roles at Lazard, Allianz Global Investors, Assiduous Investment, and Qontigo. A later UAA profile of Helena Wisniewski says McKinley supported the Alaska Data Science & AI Lab and describes planned portfolio-management competitions in which student teams would be evaluated against AI. These are first-party university accounts and attributed company statements; the webinar recording, partnership agreement, technical artifacts, data rights, current personnel, and independent validation remain unresolved.
• UC Riverside’s Hylander Student Investment Fund adds a student-capital route with explicit governance and recruiting interfaces. UCR describes a real-money portfolio founded with a $200,000 gift, a benchmark split between equities and bonds, weekly student recommendation meetings in a finance lab, supervision by Professor Greg Richey, and an advisory board of financial professionals and UCR Foundation officers. The page also links the fund to the Hylander Financial Group and records alumni media about finance careers. This is a useful control-group workflow for comparing student research, committee cadence, risk framing, and finance-talent pathways; it does not disclose AI use, holdings history, model ownership, data vendors, or independently audited performance.

These routes widen the academic search beyond named “AI in finance” degrees. IPAG connects a professor, student capital, a public video, and a publication trail. UAA connects an asset manager’s stated AI workflow to a named quant-research leader, academic lab support, and student experimentation. UCR provides a non-AI student-fund control surface with explicit governance and advisory structure. The distinction matters: attributed corporate claims, academic publications, teaching artifacts, and student-fund operations are different evidence states.

Recovery queue for IPAG, McKinley, and UCR

Recover the QuantIPAG video and transcript, von Mettenheim’s complete publication and conference trail, current fund materials, and student outputs; the UAA webinar recording, ADSAIL project pages, McKinley partnership documents, Saxena’s current profile, and any technical presentation; and UCR’s current HSIF roster, by-laws, holdings/performance reports, adviser biographies, student presentations, and alumni destinations. Do not upgrade the McKinley statements from attributed university/company claims to verified system evidence without a primary technical artifact or independent corroboration.

September 5, 2026 — Jacksonville University, MarketCipher, and a title-blind podcast route

• Jacksonville University’s Abdelwahab Missa profile identifies Missa as a Resource Professor of Finance and founder and CEO of MarketCipher Partners. The profile records a prior Deutsche Bank director role in which he built and managed an equity trading desk covering delta one, listed options, ETFs, and equity sales trading, plus an MBA from Harvard Business School. MarketCipher’s public team page separately describes Missa as Founder and CIO, Chrif Youssfi as Head of Quant Research responsible for quantitative/mathematical modelling, statistical analysis, and strategy back-testing, and SmartData asset-model work. These pages establish personnel and company roles; they do not establish current model architecture, data rights, production permissions, or performance.
• Jacksonville’s FinTech programme page links the Dolphin Student Investment Fund, the FinTech Lab, AI software, financial databases, GIS, and a public graduate story about MarketCipher technology combining AI with human insights to analyse market data for client decisions. The page currently says the undergraduate FinTech major is not accepting new applications, so its programme status should be treated as temporally important. A 2024 university account of the Dolphin fund adds two-semester student portfolio management, direct student decisions, Capital IQ and Eikon access, and Professor Missa as adviser. The AI-and-human-insights account is a university marketing/story surface; it does not disclose code, training data, model validation, or a causal link between the student fund and MarketCipher systems.
• The Managed Futures Podcast episode with Abdel Missa is a newly recovered 32-minute media route dated December 12, 2019. The public episode description says the conversation covers MarketCipher’s SmartData Investment Strategy and feeding proprietary investment drivers and relationships into AI for decision support. The episode is useful because the title contains neither “hedge fund” nor “quant”; however, the captured page does not provide a transcript or technical specification. Treat the product description as publisher/episode metadata until the audio is recovered and transcribed.

This route exposes a complete public discovery chain without collapsing its evidence levels: a professor and former bank trader; a named asset-management firm and quant-research title; a student fund and finance laboratory; a university story about AI plus human judgement; and a podcast description of proprietary investment drivers. The chain is valuable for recovery, but it is not proof that the student fund, MarketCipher, or any other manager uses a particular model in production.

Recovery queue for Jacksonville and MarketCipher

Recover the full Managed Futures Podcast audio, transcript, timestamps, RSS entry, and any alternate Spotify/Apple/YouTube copies; MarketCipher’s current product pages, SmartData documentation, team histories, papers, patents, public code, and model/data claims; Missa’s CV and faculty research; Chrif Youssfi’s public profile and publications; and Jacksonville’s current FinTech Lab, Dolphin fund, course, and alumni records. Preserve the distinctions among university marketing, podcast metadata, employee/company statements, student-fund operations, and independently verified deployment.

September 5, 2026 — AQR/LBS research interfaces and Princeton finance-ML lineage

• London Business School’s AQR Asset Management Institute pre-doctoral researcher specification describes a two-year research-assistant route intended to prepare recent graduates for finance, economics, or accounting PhDs. The specification names faculty-supervised data collection and analysis, one PhD course per term, workshops and seminars, and useful programming experience in Python, R, Matlab, SAS, or Stata. This is unusually direct evidence of a firm-supported academic talent interface and of the skills the programme sought; it is a 2020 hiring document, not evidence of a current cohort, current AQR staffing, proprietary data access, or production deployment.
• A dated LBS/AQR Institute discussion of machine learning and behavioral finance names Anna Pavlova as the Institute’s Academic Director and Marcos López de Prado as AQR’s Principal and Head of Machine Learning at the time. The article links public research ideas to de-biasing, textual sentiment analysis of market participants’ overconfidence, meta-labelling for position sizing, and pattern detection around greed and fear, while also emphasizing human judgement and organizational review. These are attributed practitioner and academic statements from January 2019; they are strategy vocabulary and historical personnel evidence, not a current model inventory, data contract, permission map, or return attribution. The Institute’s research archive also exposes a paper-and-author surface spanning alpha decay, institutional demand, asset prices, hedge-fund firms, portfolio-manager competition, and factor-related topics, but its displayed archive is not a current AQR system specification.
• Jianqing Fan’s Princeton teaching page currently lists Financial Econometrics, Statistical Machine Learning, and a 2026 course titled “Deep Learning and Generative AI.” His research page describes work on high-dimensional statistics, generative AI, deep learning, reinforcement learning, transfer learning, financial econometrics, asset pricing, portfolio choices, high-frequency trading, and risk management; it also records his Fudan, Academia Sinica, and UC Berkeley training. Princeton’s 2021 ORFE announcement says Two Sigma gave Fan a Faculty Research Award for high-dimensional statistics, machine learning, and doctoral mentorship. Together these pages establish a faculty research agenda, a current course signal, and a named historical research-support relationship. They do not establish that Two Sigma uses Fan’s methods, that the course reflects a manager’s production stack, or that any published result transferred into a live strategy.

This pass adds a useful academic-to-industry distinction. The LBS/AQR material exposes a formal predoctoral talent and research interface plus historically attributed method vocabulary. Princeton exposes a faculty research programme whose subjects overlap with systematic-investment needs and a separate historical sponsorship record. The overlap is a discovery lead only: academic subject matter, sponsor support, course content, and manager deployment remain separate evidence states.

Recovery queue for LBS/AQR and Princeton faculty routes

Recover the AQR Institute’s current advisory council, active fellows or predoctoral cohorts, awards, event recordings, and paper PDFs; identify which archived research items have reproducible code or data descriptions; and search the named historical personnel independently for later employer roles and public talks. For Princeton, recover Fan’s finance and AI/ML publication lists, software, seminars, current course materials, and public student artifacts without exposing restricted names. Preserve historical sponsorship, faculty research, curriculum, personnel placement, and verified manager deployment as separate fields.

September 5, 2026 — university investment-office prototypes and allocator-side data roles

• Duke’s 2020 “AI in the Investment Office” project deck is a concrete academic-to-allocator artifact. The Data+ team describes work for Duke University Management Company (DUMAC) in two tracks: a Python cost-optimization workflow for prime-broker cash and margin transfers, and venture-investment analysis using Burgiss, PitchBook, web extraction, Tableau, Python, and Excel. The deck exposes the operating logic rather than only naming “AI”: daily spreadsheet ingestion, broker-rate comparisons, margin constraints, proposed transfers, savings thresholds, weighted investment-duration calculations, and visualizations for analysts. It also describes a proposed Intralinks API integration to automate document retrieval and organization. The deck names DUMAC project lead Robert McGrail, student project manager Yi Wang, and faculty contact Paul Bendich. This is a student project presentation and a stated continuation plan; it does not prove that the prototype entered DUMAC production, that the data licences remain current, or that it affected investment performance.
• Duke’s first-person account by Priya Parkash independently corroborates the internship interface: the author says the team explored AI support for the investment office, built a financing-charge cost-optimization tool, and worked with the Private Investments team on DUMAC’s historical venture investments. The two Duke surfaces therefore establish an academic project, a named allocator-side project lead, a data-and-automation workflow, and a student talent route. They remain separate from verified DUMAC deployment or any claim about a manager hired from the project.
• A current Harvard career-board posting for HOF Capital’s Data Scientist / Researcher adds an allocator-side, title-specific recruiting signal outside traditional hedge-fund vocabulary. The employer description calls for data pipelines and ML models to evaluate and monitor millions of technology companies, products, open-source frameworks, and technology talent; raw-data analysis for diligence; a centralized portfolio-company metrics database; benchmarking for portfolio companies; and advice on data-stack and data-driven-culture buildout. Harvard dates the posting’s recruitment start to August 8, 2026 and expiration to September 7, 2026. This is an employer-provided job description hosted by a university career service, not evidence that a candidate was hired, that a model is in production, or that the firm’s performance improved.

The Duke artifact adds a model-idea category that the hedge-fund search can miss: allocator operations can produce measurable AI/automation opportunities before any predictive alpha claim—cash financing optimization, document-ingestion automation, private-market data normalization, investment-duration analysis, and analyst-facing visualisation. The HOF role adds a second allocator pattern: ML applied to sourcing, technical diligence, portfolio-company monitoring, and value-add rather than directly to public-market execution. These are academic-project and recruiting observations, not rankings or deployment claims.

Recovery queue for allocator and university investment-office routes

Recover Duke’s final summary, source code, data dictionary, project handoff, DUMAC continuation records, and any later team or employer pages; identify whether the Intralinks API plan was implemented. Search the DUMAC project leads and faculty independently across papers, talks, and current roles. For HOF Capital, recover the original employer posting, current team page, role outcome, and any public technical or investment-research material. Preserve project prototype, planned continuation, employer request, hire confirmation, and production evidence as separate states.

Student portfolio models and finance-programme operating surfaces

• WPI’s Investing Association provides a current student-capital and faculty-advisory route. Its first-party site describes equity research, portfolio management, sector analysts, fund managers, a structured investment committee, and five named faculty advisers: Eugene Okyere-Yeboah, A.J. Edwards, Kwamie Dunbar, Xin “Shane” Gao, and Marcel Blais. It says machine-learning engineers joined the executive board during the association’s FinTech expansion and that students built tools used by the club. The site describes a planned $300,000 initial investment and a target to reach $1 million, but also labels performance reporting “coming soon.” Those are institution-reported programme claims; the page does not establish holdings, model specifications, student employment outcomes, or independently audited performance. Current student names are not reproduced here.
• UC Berkeley’s Very Intelligent Portfolio capstone page, its project site, and presentation PDF expose a reusable academic model artifact. The team describes an open-source portfolio tool for concentrated employer-stock risk, using WRDS-provided CRSP daily data and Compustat fundamentals, with stock representations learned from historical relationships and portfolio weights produced through optimization. The project site identifies Ray Cao as a subject-matter expert and Head of Quantitative Portfolio Management at RPIA LP, with quantitative-strategy and portfolio-process responsibilities, and lists graduate training at Wilfrid Laurier, Birmingham, and Nankai. The university page and project site describe self-attention/embedding baselines, transformer variants, dynamic versus static embeddings, a PyTorch/AWS prototype, and a Colab/API interface. These are student and project-team descriptions; the production claim on the project page is not independently validated, the data licences are not examined here, and no RPIA adoption or performance attribution is established.
• Berkeley’s DeepLearning Sentiment Portfolio capstone page and final presentation add a separate news-plus-portfolio route. The project describes roughly 2.5 million finance and economics articles from 2001–2022, including FNSPID, New York Times, and scraped material; fine-tuning the last layers of FinBERT; attention pooling of daily news embeddings; and a two-transformer design that models both temporal sequence and cross-security relationships. It then uses expected returns to generate portfolio weights and experiments with sentiment both as an input and as an exposure scaler. The presentation reports student backtests, a London Stock Exchange senior quant as a subject-matter reviewer, and future work on turnover, transaction costs, implementation, and reinforcement learning. Those results and the reviewer statement remain project-reported; the public artifact does not provide a complete point-in-time audit, executable repository, cost model, or institutional deployment evidence.

The WPI, VIP, and DSP routes show three different research surfaces: an operating student fund with faculty oversight, an open-source personalized-risk portfolio prototype, and a multimodal text-plus-return regime model. They are useful for discovering models, data vendors, training methods, and personnel bridges. They must not be collapsed into evidence that an investment manager uses the methods or that the reported backtests survive independent replication.

Recovery queue for WPI and Berkeley student-model routes

Recover WPI’s fund policy, current faculty pages, board history, endowment records, tool repositories, and any public performance reports. Recover VIP’s archived presentation, linked notebooks, GitHub repository, model-architecture and data-pipeline pages, and Ray Cao’s independent RPIA profile. Recover DSP’s linked repository, presentation assets, dataset licenses, news-timestamp joins, model checkpoints, transaction-cost assumptions, and code needed to reproduce the reported comparisons. Preserve project claims, independent code, data rights, replication, and employer adoption as separate states.

Additional student-fund and quantitative-finance talent surfaces

• Cornell Quant Fund adds a student quantitative-finance organisation rather than a traditional classroom or endowment fund. Its current first-party site describes training in market microstructure, probability, statistical learning, game theory, and options theory; an annual trading competition with 150-plus participants and more than $10,000 in prizes; and engineering projects including a live exchange, strategy backtester, and trading games. The site also links a Trading Lecture Series, sponsor relationships, and a public team/application path. It does not name a faculty adviser, disclose a live fund or holdings, identify sponsor-specific projects, or establish that its software is used by an investment manager.
• ANU’s Student Managed Fund provides an Australia-Pacific operating model. The first-party page reports $911,378 under student responsibility as of July 23, 2026, with investment recommendations reviewed by an Investment Advisory Committee containing industry practitioners. It identifies Dean Katselas as fund convenor, Hua Deng as course convenor, and a two-semester progression from junior to senior roles through FINM3009/6009 and FINM3010/6010. The page links reports, publications, a team page, and social channels. It does not disclose AI use, model code, holdings history, data vendors, or audited investment attribution; the capital figure and role descriptions are institution-reported and date-sensitive.
• SMU’s Student-Managed Investment Fund adds a Singapore talent and allocator-adjacency route. The university’s 2026 student-activities page says the long-only public-equity fund was founded in 2005, held its first Investment Committee meeting in October 2025, and reported more than SGD 200,000 in AUM. It describes a three-stage selection process, weekly sessions, alumni links to investment banking, private equity, hedge funds, and asset management, and public participation in competitions associated with Citadel, Point72, Temasek, Fidelity, and other organisations. Those are programme and self-reported achievement claims; the page does not establish sponsor funding, sponsor model transfer, current holdings, AI use, or performance.
• Purdue’s Student Managed Investment Fund exposes a more inspectable real-money workflow. Its first-party site says continuous audited monthly statements begin in October 2013, holdings and performance are published, and the fund uses sector teams, written investment theses, a committee vote, quarterly reviews, and semester reporting. It names Lulu Zeng and Alexander Boquist as faculty advisers and describes an analyst-to-portfolio-manager-to-executive-board progression. The page also reports alumni destinations at named financial and professional-services employers. This is valuable for recovering positions, attribution, governance, and talent-path artifacts, but the public page does not show AI use, model ownership, data rights, or any manager adoption.

These four routes add geographic and operating coverage to the academic map. Cornell exposes engineering and trading-simulation infrastructure; ANU exposes a dated Australian student-capital process with named convenors; SMU exposes an Asia-based live-fund and alumni interface; Purdue exposes a public holdings/performance and governance surface. They are discovery routes for research artifacts and personnel lineage, not a league table and not evidence that any hedge fund uses their methods.

Recovery queue for Cornell, ANU, SMU, and Purdue

Recover Cornell’s current team, sponsor roster, lecture recordings, competition data, and public repositories; ANU’s reports, publications, Investment Advisory Committee, current convenors, linked video, and alumni destinations; SMU’s fund reports, committee materials, competition records, alumni profiles, and sponsor disclosures; and Purdue’s holdings, performance, research, team, Substack, policy, and faculty pages. Search named faculty, convenors, alumni, speakers, and sponsors independently across papers, podcasts, conference programmes, and employer pages. Keep student activity, university reporting, sponsor association, and verified manager deployment separate.

Recovered ANU fund report and Cornell event/personnel surfaces

• ANU’s approved Semester 1 2026 Student Managed Fund report is materially richer than the programme landing page. It describes a live portfolio with Australian active-equity and asset-allocation teams, a risk-and-compliance team, a relationship team, an Investment Advisory Committee, and a formal handoff between senior and junior cohorts. The report records portfolio stewardship and thesis revaluation, a three-year Australian macro outlook with 11 scenarios, collaboration between asset-allocation and equity teams on interest-rate exposure, a quantitative-risk-measures initiative, and an overhaul of the Investment Book of Records to improve accessibility for future analysts. It also records industry sessions with Significant Ventures, Australian Ethical Investment, Nanuk Asset Management, Team Super, and a former SMF CIO. These are institution-reported operating and educational artifacts; the report does not disclose AI, model code, data licences, or manager adoption. Its return, AUM, holdings, and attribution figures remain dated fund self-reports and are not independently audited here.
• The same ANU report shows a useful non-LLM automation target: durable handoffs, traceable portfolio records, thesis revaluation, exposure measures, and cross-team macro/equity collaboration. That is an inference about research workflow design from the report, not a claim that ANU uses an AI system or that the process produces investable alpha. The report’s explicit disclaimer says the students and staff are not licensed to provide financial-product advice.
• Cornell Quant Fund’s current team page exposes role design and talent movement without requiring a hedge-fund keyword. The executive board includes heads of systematic equities, derivatives, prediction markets, and software engineering; the biographies describe computer-science, mathematics, physics, statistics, and operations-research training, plus internships or prior roles at Millennium, Citadel, Citi, Peak6, Aristeia, IMC, JPMorgan, Amazon, and Xantium. The 2026 Trading Competition page schedules the fifth annual event for October 24, 2026 at Cornell Tech, welcomes undergraduate and master’s students from the United States and Canada, and advertises original trading games, cases, networking, 13 sponsors, and $12,000 in prizes. The lecture-series page adds an IMC guest lecture and a syllabus spanning market microstructure, systematic equities, algorithmic trading, options, and derivatives. These pages establish public talent and event surfaces; they do not identify confidential sponsor work, faculty supervision, model ownership, live capital, or investment performance. Current student names are not reproduced here.

These recovered artifacts improve the academic route in two directions. ANU exposes the operational controls around a live educational portfolio—records, handoffs, risk measures, scenario work, and thesis maintenance. Cornell exposes a student-built trading-infrastructure and recruiting surface with systematic, derivatives, prediction-market, and software roles, plus an upcoming event likely to yield speakers and employer vocabulary. Neither should be treated as evidence of hedge-fund deployment or ranked against another programme.

Recovery queue for ANU report and Cornell event routes

Recover ANU’s underlying report PDFs, Investment Advisory Committee roster, portfolio files, recommendation reports, risk-measure definitions, IBOR change history, speaker spotlights, and video/transcript assets. Recover Cornell’s 2026 competition sponsor list, case/game descriptions, lecture recordings, GitHub organisation, team history, and any public event recordings. Search named employers, speakers, faculty, alumni, and programme leads separately; preserve dates, public claims, student work, and verified employer evidence as distinct fields.

Family-office research, curriculum, and practitioner-faculty routes

• Chicago Booth’s Family Office Initiative is a new institutional route for family-office discovery. Booth describes a research, education, and network initiative; names Robert (Bobby) Stover Jr. as inaugural executive director; lists Emanuele Colonnelli, John C. Heaton, Steven Kaplan, Pascal Noel, and Eric Zwick among the faculty in this area; and identifies a Family Office MBA course, a practitioner council, and a 2025 summit attended by representatives of 147 family offices from the United States, Canada, EMEA, and Latin America. The page also describes research on family-office investment behaviour, performance, leadership, and relationships. This is a university programme and network surface; it does not disclose member-office identities, models, vendors, portfolio data, or AI deployment.
• Wharton Global Family Alliance’s 2026 survey page identifies Raphael Amit and a July 2026 executive summary for the eighth detailed family-office survey. The page says the research is designed to understand current practices and performance drivers while preserving participating families’ anonymity and confidentiality, and that detailed reports are distributed to participating offices. This is a useful route for recovering the public summary, survey methodology, programme leadership, case studies, and future speakers; it is not a dataset of identifiable offices and does not reveal their AI systems, managers, vendors, or results.
• Adelaide University’s profile for Zenki Kwan provides a rare public academic-to-single-family-office personnel bridge in Hong Kong. The profile identifies Kwan as an Adelaide adjunct professor and as CIO of an unnamed Hong Kong single-family office, with responsibility—according to the profile—for investment strategy and the investment committee across equities, fixed income, funds, and structured products. It also records adjunct and academic-advisor roles for HKU and HKBU family-office programmes, prior McKinsey, Samsung Securities, UBS, J.P. Morgan, and Finsoft experience, and Oxford, Harvard, HKU, HKUST, UCL, and Swiss Business School education. These are profile claims about an unnamed office and a public teaching/research interface; the office identity, portfolio, AI vendors, models, permissions, and returns remain unresolved.

The family-office layer now has three separate evidence surfaces: an academic institution building a practitioner council and curriculum; an anonymized longitudinal survey; and a named professor/practitioner whose employer remains undisclosed. That separation matters because family-office privacy is itself a discovery constraint. The appropriate next step is to recover public programme speakers, survey instruments and summaries, faculty papers, and disclosed technology or vendor relationships without attempting to identify anonymous offices.

Recovery queue for family-office academic routes

Recover Booth’s Family Office MBA syllabus, faculty papers, council biographies, summit programmes and recordings, and any public research outputs. Recover Wharton’s 2026 executive-summary PDF, survey methodology, case-study archive, programme events, and later survey updates. Recover Kwan’s public talks, papers, teaching materials, HKU/HKBU programme pages, and any independently disclosed family-office technology relationship. Keep anonymous survey evidence, named personnel biographies, academic teaching, and verified office deployment separate.

Recovered Wharton 2026 family-office survey PDF

• The Wharton Global Family Alliance 2026 executive summary adds measurable allocator operating context. The survey instrument was distributed in Q1 2026 to family offices and selected firms serving family-office clients, with respondents across 21 countries in the Americas, Europe, the Middle East, Asia, and Australia. The public summary says 61% of respondents manage more than $500 million, 56% employ seven or fewer professionals, and most investment-management work in the sample is handled in-house rather than outsourced. It also reports that the technology platform is mission-critical for financial management, custody, information consolidation, aggregation, and client reporting, while the average office employs fewer than one IT professional and 0.2 cybersecurity specialists. These figures are survey-summary claims with an uneven regional sample; they do not identify offices, vendors, model architectures, AI usage, or investment performance.
• The PDF also makes a useful workflow distinction for family-office research: the scarce technical layer sits alongside broad investment, legal, accounting, and client-service responsibilities. That suggests a concrete discovery queue around aggregation, reporting, custody-data reconciliation, document controls, cybersecurity, and investment-management handoffs before assuming a predictive-AI use case. This is an inference from the survey’s operating categories, not evidence that any surveyed office has implemented those automations.

The recovered PDF upgrades the Wharton route from “survey exists” to a dated, inspectable instrument summary with explicit technology, staffing, governance, and outsourcing fields. The detailed report remains confidential, so the research process should use the public summary to identify programme speakers, vendors, and public case studies—not to infer the identity or systems of participating offices.

Recovery queue for the Wharton 2026 survey artifact

Preserve the executive-summary PDF, recover its survey instrument and methodology if publicly released, compare the 2026 categories with the 2024 and 2022 summaries, and search WGFA events, case studies, faculty papers, and public partner material for technology and workflow examples. Keep all respondent-level claims anonymous and distinguish survey-reported operating patterns from named-office evidence.

WorldQuant University: finance-ML education and partner-network surface

• WorldQuant University’s current partners page says the university works with more than 40 government, industry, and social-sector organisations in 13 countries, with the list current through August 11, 2026. The page presents recruitment, employee development, dual-degree, blended-learning, and applied-lab routes, and names relationships including Sazience Technology, Kuramo Capital, Zindi, AIMS Rwanda, the National Bank of Rwanda, and the Tanzania Institute of Bankers. Its partnership news describes a 2025 Sazience co-op in which selected learners could demonstrate skills in real time while studying. This is public education and talent-network evidence; it does not disclose proprietary WorldQuant data, models, signals, or trading permissions.
• WorldQuant University’s January 2026 fact sheet lists an MSc in Financial Engineering, Applied Data Science, Deep Learning Fundamentals, and Applied AI: Computer Vision labs. It identifies Igor Tulchinsky as founder, chairman, and CEO of WorldQuant University and describes him as the founder of WorldQuant, a quantitative investment firm with more than 20 offices in 16 countries. That creates a public institutional-lineage and talent-pipeline clue, not evidence that university projects flow into WorldQuant’s investment process. Learner counts and outcomes in the fact sheet remain institution-published claims.

The useful discovery surface here is the intersection of finance-engineering education, applied ML labs, regional partner development, and public quantitative-investment lineage. Follow the partner list, programme artefacts, instructor profiles, learner projects, alumni destinations, and co-op materials. Keep the university, its partners, and the investment firm as separate entities unless a source explicitly links a model, dataset, or workflow.

WorldQuant University recovery queue

Recover the partner-list download, MScFE and applied-lab syllabi, instructor biographies, public learner projects, competitions, alumni profiles, Sazience co-op materials, and any independently disclosed Kuramo or WorldQuant University technology relationship. Search each named partner and programme lead separately across talks, papers, repositories, and job descriptions; do not infer proprietary-data access or investment deployment from the relationship list.

University-linked investment research and quant curricula

• The University of Chicago Data Science Institute’s Millennium announcement records a 2026 industry-affiliate relationship. UChicago says Millennium will participate in curriculum and applied-learning initiatives, help shape a quantitative-developer certificate being jointly developed by DSI and Financial Mathematics, and contribute industry expertise alongside other firms. The announcement names Pranat Pathak as Millennium’s International CIO and Global Head of Fixed Income, Commodities and Core Technology. It describes student engagement, project collaborations, seminars, and talent acquisition as part of the affiliate channel. This is a named academic-industry and personnel signal; it does not disclose a Millennium dataset, model, capstone, research result, or production system.
• MIT Sloan’s Finance Lab project archive lists public examples including machine-learning research on social-media alpha signals such as Glassdoor data, probability models for emerging-market currency crises, rates-and-currencies strategies using volatility risk premia, transaction-level intraday bond-liquidity measures, factor analysis of illiquid private-equity and distressed-credit investments, fixed-income ETF liquidity and factor crowding, and target identification for small family-owned manufacturers. The page establishes project topics and the sponsor/faculty/student structure, not sponsor identity, proprietary access, investable performance, or manager deployment.
• The University of Edinburgh’s 2026/27 Financial Markets Applications of Deep and Reinforcement Learning course specifies implementation and evaluation of ML systems on financial time series, including high-frequency trading and limit-order-book data. Its outline includes feed-forward networks, classification, autoencoders, CNNs, RNNs, reinforcement learning, and transformers, and lists Dr Adam Ntakaris as course organiser. This is explicit curriculum evidence about model families and the HFT research object, but it does not identify a fund partner, dataset owner, or live trading use.
• Adam Ntakaris’s University of Edinburgh research profile adds a professor-level recovery route behind that curriculum. The public profile links work on feature engineering for financial time series, a minimal-batch adaptive-learning policy engine for real-time mid-price forecasting in high-frequency trading, and a PhD viva on multi-agent reinforcement learning for market making. These links identify research objects and a named academic, not a hedge-fund affiliation, proprietary data source, or live deployment; the papers and thesis materials still need independent review.
• The profile’s 2026 paper on event-triggered LOB forecasting adds a concrete, inspectable model/data route. The paper proposes an RL-Arbitrator that switches between 1-millisecond data and adaptive subsampling, uses an adaptive Huber loss during structural price jumps, and reports a benchmark using millisecond-aggregated Databento feeds with ten-level depth for four S&P 500 stocks over three months against eleven baseline models, including deep-learning and unified GARCH–Itô approaches. The authors report higher out-of-sample accuracy during toxic-order-flow periods and computational savings; those are paper-reported results requiring independent reproduction, and the source does not establish a fund relationship or production deployment.
• Durham’s 2026/27 Quantitative Finance with Artificial Intelligence module combines professional finance data, programming, trading-strategy analysis, ML framing for prediction/classification/risk/regime tasks, robust validation, interpretability, slippage, and critical assessment of when ML adds value over classical methods. It also explicitly allows documented AI coding assistants for scoping, debugging, and refactoring. This is a current teaching artifact that exposes an evaluation and governance vocabulary; it does not establish student or employer deployment.
• Fordham’s MS in Quantitative Finance industry-project page names past employer relationships including AQR, Goldman Sachs, Guggenheim Partners, HSBC, J.P. Morgan, Morgan Stanley, Neuberger Berman, Numerix, and RBC, and lists project themes including asset management with ML, FOMC signals and asset prices, hedge-fund strategies, interest-rate derivatives, and FRTB. The page attributes the programme route to Professor Qing Sheng. These are school-reported employer and project claims; they do not reveal assignments, models, data rights, or results.

These sources add a useful distinction to the academic route. The Chicago announcement is a named hedge-fund-to-university affiliate relationship with a public technology leader; MIT exposes the kinds of unresolved investment and private-market questions sponsors can pose; Edinburgh and Durham expose model, validation, slippage, interpretability, and AI-assistant language in current curricula; and Fordham exposes a broader employer/project vocabulary. None should be converted into a claim about a firm’s internal AI strategy without a separate firm source.

Recovery queue for the new university-linked routes

Recover the Chicago quantitative-developer certificate’s current syllabus, DSI affiliate events, Millennium participant pages, and any public capstone or seminar artifacts. Recover MIT Finance Lab host/faculty pages, current and historical project briefs, sponsor names where publicly disclosed, and student outputs. Recover Edinburgh and Durham reading lists, project prompts, lecturers, recordings, repositories, and graduate destinations. Recover Fordham’s advisory board, project archive, faculty biographies, and public alumni/employer traces. Preserve the distinction between a programme’s stated curriculum, a sponsor relationship, a named employee, and verified firm deployment.

Academic venture, financial-institution, and formal-verification bridges

• Carnegie Mellon’s VentureBridge ’26 cohort announcement names Stephen Wu, a School of Computer Science fellow, and describes his project as an AI-driven hedge fund using machine learning for equity-signal generation and portfolio optimisation. This is a direct university-published founder/project statement, not evidence of a registered fund, assets, data rights, model architecture, live capital, or performance. The announcement places the project inside a broader CMU founder pipeline; that institutional label is not diligence on the venture.
• Columbia Business School’s Program for Financial Studies describes a research and education centre for financial economics, quantitative finance, and financial analytics. It cites computational resources, traditional and nontraditional datasets, industry engagement, research sharing, student recruitment, and the Society of Quantitative Analysts as an industry affiliate for MSFE students. This is an academic-industry interface and personnel-discovery route; it does not disclose a particular firm’s dataset, model, project, or deployment.
• The University of Oxford–UBS Centre for Applied AI announcement describes independent research and joint initiatives between Oxford and UBS, a dedicated team of 20 researchers, and research areas covering AI and society, AI for business and the economy, and future AI paradigms. UBS says the partnership is intended to produce practical tools that can be implemented at scale across the firm. This is a named financial-institution research centre, not a hedge-fund disclosure; the announcement does not identify specific models, datasets, researchers beyond the leadership announcement, or deployed systems.
• Imperial College London’s Logos Research announcement describes a spinout led by Dr Cristopher Salvi, an Imperial associate professor, that uses a Lean formal-logic library and an AI-agent feedback loop to verify AI-generated code. The announcement identifies algorithmic trading and other high-stakes applications as target use cases and names Khosla Ventures, XTX Ventures, and SOSV in the financing history. This exposes a technically specific verification workflow and an academic-to-finance venture bridge; it does not establish adoption by any tracked fund or a production trading integration.

These routes add different kinds of evidence. CMU exposes a student/fellow’s explicit fund concept; Columbia exposes a finance-research and industry-affiliate structure; Oxford–UBS exposes a resourced financial-institution AI centre; and Logos exposes a formal-verification mechanism aimed at reducing errors in AI-generated quantitative code. They should remain separate from manager-specific AI strategy evidence.

Recovery queue for academic venture and institutional AI bridges

Recover Stephen Wu’s public venture identity, research or product pages, legal entity and regulatory records where available, and any public model or data description. Recover Columbia PFS faculty, SQA events, MSFE projects, public seminars, and student work. Recover the Oxford–UBS centre’s appointed professor, researcher roster, publications, events, and implementation updates. Recover Logos’s technical papers, repositories, founder biographies, financing disclosures, and any separately named finance customer or pilot. Treat university announcements, founder claims, vendor claims, and independently verified deployment as different evidence classes.

Finance professors, programmes, and student-fund routes

• The SMU Finance Research repository paper by Xiaowen Hu, Maximilian Rohrer, and Hanjiang Zhang (published March 17, 2025) provides a particularly useful academic measurement route. It defines an Active ML-Based Trading (AMLT) measure from mutual-fund holdings and forward-looking deep-neural-network signals built from quantitative and textual information, then uses employee AI-talent analysis as a validation channel. The abstract reports a rising AMLT trend and a 2.4–3.0 percentage-point annual risk-adjusted difference between its top and bottom groups; it also reports that the full information/model specification produces a larger result than linear or reduced-information variants. These are paper-reported mutual-fund findings, not evidence about any named hedge fund, and they require independent replication, cost treatment, and point-in-time review before use in an investment process.
• The University of Chicago Financial Markets Program is a three-year undergraduate talent and vocabulary surface for quantitative finance. Its public curriculum includes coding, machine-learning training, trading simulations, price-prediction modelling, and practitioner workshops. The current programme page names site visits or presentations involving Two Sigma, Walleye Capital, Millennium, DRW, Citadel, Jane Street, Optiver, and other market firms. This establishes a recruiting and skills pipeline that can be mined for speakers, student projects, event recordings, and employer terminology; it does not establish that a named firm supplied a model, dataset, or production workflow.
• NYU’s profile of Jerzy Pawlowski identifies him as an adjunct professor in Finance and Risk Engineering and as Chief Investment Officer at MachineTrader. NYU’s public biography also describes prior hedge-fund portfolio-management work in credit derivatives and CDO tranches and research interests in machine learning for systematic investing and trading. The page links a public GitHub route. This is a named academic/practitioner and former-hedge-fund personnel bridge; it does not reveal a tracked-fund affiliation, proprietary model, dataset, or live strategy.
• University of Florida’s faculty profile for David Mascio lists him as Executive Director of the Fintech Research Center and as teaching Natural Language Processing & Generative AI for Finance. The profile also states that he is a founding partner of Della Parola Legacy Fund and previously served as chief investment officer of a billion-dollar trust bank. Those are first-party profile claims requiring normal source verification; they create a useful professor–fund–curriculum route, but do not establish the fund’s model, data, permissions, or deployment.
• The Claremont McKenna College Student Quantitative Fund describes a student-led research group focused on quantitative investment strategies, algorithm development, strategy testing, and industry-professional partnerships. It is a student-fund discovery surface for research artifacts, team lineage, talks, and sponsor vocabulary. The public description does not identify proprietary datasets, models, live capital, performance, or a manager deployment.

These routes add academic measurement, curriculum, recruiting, and professor/practitioner evidence to the map. The SMU paper can inform how public holdings and personnel signals might be tested; the programme and faculty pages help locate people, talks, papers, and source vocabulary. None should be converted into a claim about the internal AI strategy of GMO, Acadian, Arrowstreet, or any other tracked manager without a separate firm-level source.

Recovery queue for professor and programme routes

Recover the SMU paper’s full text, code, input construction, fund universe, point-in-time restrictions, transaction-cost assumptions, and author seminars. Recover UChicago programme speakers, employer event recordings, student projects, trading-competition material, and current programme staff. Recover Pawlowski’s linked GitHub and papers, and Mascio’s CV, course materials, fund disclosures, and published work. Recover the CMC fund’s team, faculty advisers, sponsor list, research outputs, and competition or event recordings. Preserve paper-reported results, profile claims, programme marketing, and independently verified firm deployment as separate evidence classes.

Asian financial-ML labs and international-model routes

• City University of Hong Kong’s FinTech and Digital Assets Centre describes a research centre spanning AI/ML asset pricing, quantitative investment, financial econometrics, portfolio choice, digital assets, blockchain, generative AI for financial-market intelligence, data-driven financial decisions, and industry collaboration. This is a named academic-lab route with direct relevance to Hong Kong and regional finance research. The centre page does not identify a particular hedge fund’s data, model, vendor, or production deployment.
• CityU’s profile of Min Dai records his July 2026 move to CityU as Hung Hing Ying Chair Professor of International Economics and Finance, his PhD from Fudan, and prior faculty positions at Peking University, NUS, and Hong Kong Polytechnic University. His publication page lists work on reinforcement learning for portfolio selection and stock-index-futures arbitrage, policy randomisation for data-driven Merton strategies, data-driven option pricing, and investment under transaction costs. This is a professor and model-lineage route; the papers do not establish adoption by a tracked fund.
• Sungkyunkwan University’s Financial Mathematics & Machine Learning lab identifies Jeonggyu Huh as PI and describes PG-DPO, dynamic portfolio choice, deep hedging, asset pricing, and stochastic control. The lab’s people page names a postdoctoral researcher and graduate topics including financial-time-series augmentation, dynamic allocation, bankruptcy anomaly detection with a residual deep compressive autoencoder, VIX-term-structure forecasting, flow matching for corporate-default prediction, and decision-focused learning. Its public paper list includes a 2026 ICML Pontryagin-guided framework and a 2026 reliability-screened equity-premium forecasting paper. The open-access PLOS paper specifies a two-stage pipeline: one-step predictor forecasts and uncertainty proxies are screened for reliability, then admitted signals are mapped to next-period excess returns with random forests, optional SHAP screening, and PCA/PLS representations; the study uses monthly U.S. data from 1952–2024 and reports net-of-cost and downside-state diagnostics. The authors publish reconstruction code and documentation through GitHub and Zenodo. These are Korean academic research and personnel surfaces, not evidence of manager deployment.
• The ABFER paper copy of Alpha Go Everywhere names Darwin Choi and Wenxi Jiang of CUHK and Chao Zhang of Oxford. It tests market-specific neural networks for international stock returns, adds variables constructed from U.S. firm characteristics, and records conference and seminar exposure across Asian and international finance programmes. The published Review of Asset Pricing Studies record confirms the paper’s authors and abstract while noting that publisher full text is restricted. This is a concrete international-equity model and academic network route; it is not evidence of any fund’s implementation.
• The recovered University of Florida CV for David Mascio adds lineage and temporal detail to the faculty profile. It lists a finance PhD at EDHEC, dissertation work on machine learning and applied-finance forecasting, Frank Fabozzi as internal adviser, Turin Bali as external adviser, papers on sentiment-index forecasting and combined-forecast/ML market timing, and current UF teaching in financial modelling and NLP/Generative AI in Finance. It also lists practitioner events including QuantMinds, Neudata, Quant Strats, and the New York Data Summit. These are CV claims and public teaching/speaking records requiring ordinary external reconciliation; they do not reveal a fund’s data, model permissions, or deployment.

The new Asian routes expose three different research objects: institution-level AI/quantitative-finance infrastructure at CityU; continuous-time, RL, and transaction-cost methods in named faculty research; and market-specific international-equity modelling with cross-market information. The personnel and programme links are useful for discovering papers, code, seminars, and employer interfaces, but academic proximity remains separate from evidence about a manager’s internal system.

Recovery queue for Asian labs and faculty lineage

Recover CityU FDAC’s current staff, projects, seminars, partner disclosures, and graduate outputs; Min Dai’s latest working-paper files and code; the SKKU lab’s complete paper/code links, adviser histories, and conference recordings; the Alpha Go Everywhere replication materials and market/data definitions; and Mascio’s external publication records and conference media. Do not reproduce personal contact details from CVs, and keep academic claims, self-reported CV claims, and independently verified firm deployment in separate fields.

Applied capstones, practitioner networks, and finance-AI coursework

• The University of Maryland Smith School’s March 2026 experiential-learning report describes an MQF team working with senior executives from the World Bank and T. Rowe Price on homeowners-insurance data as a natural-hazard signal for mortgage default, prepayment, MBS valuation, and credit analysis. The report says the team merged more than 500,000 loan records with ZIP-code insurance data and compared logit, Lasso, Random Forest, XGBoost, and neural-network models. It also describes a prior Google-sponsored project that used Gemini, AI Studio, and NotebookLM to analyze more than 25,000 pages of public financial documents, plus a current private-credit project using bank call reports and generative-AI tools. The MQF programme page adds a student-managed portion of Smith’s Global Equity Fund, an Agentic AI Challenge and Datathon, named Bloomberg/Capital One/Google experiential partners, and faculty including Michael Cichello, Albert Kyle, Steve Heston, and Russ Wermers. These are university-reported student projects and programme claims; they do not establish sponsor ownership, fund deployment, or production performance.
• Rice University’s Center for Computational Finance and Economic Systems partnership page describes two decades of academic/industry collaboration, simulated training environments, internships, real-world projects, workshops, and the Eubank Conference on Real World Markets. Its public invited-lecturer list includes NYU data-science professor Vasant Dhar, QuantRoll Capital’s managing partner, Vaughan Nelson’s chief risk officer, Fidelity’s quantitative/data-science lead, Bank of America executives, and an Indian Statistical Institute associate professor. This is a useful speaker and personnel graph for finding papers, talks, employer vocabulary, and project provenance; the page does not identify proprietary datasets or any specific manager’s deployment.
• UC Berkeley’s IEOR 198 Introduction to Quantitative Finance is a current student-led course with guest lectures, applied exercises, Jupyter assignments, market making, game theory, risk, options, statistical models, machine learning, and a coding project intended to build systematic-trading software. Professor Thibaut Mastrolia is listed as the sponsoring professor, while the student group states that it operates independently of the university. This exposes an unusually direct student-to-practitioner vocabulary and code-project route, not evidence of a firm’s research system.
• Stevens Institute’s Revenue Revolution 2026 report documents a conference on AI and digital transformation in asset-management distribution with keynotes, panels, demonstrations, and an evening reception. It names George Calhoun, the quantitative-finance programme director and Hanlon Financial Systems Center executive director; CRAFT, an NSF-backed academic/industry centre; Google Cloud strategic-alliances leader Paul Magnone; and Narrative Alpha founder Rick Lake. Sessions covered compliance in the AI era, tokenization, digital assets, ETF strategy, wealth management, and alternatives distribution. This is a dated conference and speaker route; it does not show a particular hedge fund’s implementation.
• Stevens’ Big Data/AI for Quantitative Finance capstone page exposes a concrete project specification from a 2022 industry engagement: scrape/API acquisition of multi-asset metadata and time series; AI and quantitative models for risk signals and algorithmic strategies; early-warning and volatility indicators across equities, crypto, and NFTs; backtests and simulations; documented GitHub functions; paper-trading automation; and monitoring dashboards. The host company is anonymized and the engagement is dated, so this is a workflow and deliverable artifact—not evidence of a named firm, proprietary data, live capital, or production performance.

These routes expose how finance-AI work is made legible outside a manager: sponsor-defined questions, permitted data, model comparisons, deliverables, code repositories, practitioner events, and recruiting vocabulary. They are valuable for discovering people and methods while preserving the boundary between a classroom or sponsored project and an operating investment process.

Recovery queue for applied academic finance routes

Recover UMD’s full project reports, sponsor presentations, data-use terms, faculty advisers, and student outputs; Rice’s Eubank agendas, recordings, research archive, and named practitioner biographies; Berkeley’s current syllabus, EdStem materials, sponsor list, competition recordings, and faculty/student history; and Stevens’ Revenue Revolution agenda, recordings, CRAFT project archive, capstone sponsor identity where publicly releasable, GitHub artifacts, and faculty supervision. Keep sponsor-reported figures, student work, public code, and verified investment deployment as separate evidence classes.

September 5, 2026 — Oxford–Man replay corpus and paper-level model routes

• The Oxford–Man Machine Learning and Finance Conference 2026 archive is more than an event listing: it provides direct MP4 replays and linked papers for talks by David Hirshleifer, Siew Hong Teoh, Stefan Nagel, Ansgar Walther, Shuang Chen, Carol Alexander, Jesús Gorrín, and Patrick Chang. The archive therefore creates a recoverable paper-to-talk corpus for local transcription, model extraction, and timestamped research indexing. The page does not connect any talk to a manager’s production system.
• Detecting Deception in CEO Interviews: An AI-Based Measure of Text-Audio-Video Modal Incongruence, by Minjae Koo, Siew Hong Teoh, Il Sun Yoo, and Meiling Zhao, describes affect extraction from text, audio, and video across 2,439 CEO interviews from 2010–2023. The paper constructs a cross-modal divergence measure and tests it against active non-public SEC investigations, the F-score, residual short interest, earnings surprises, restatements, SEC comment-letter amendments, insider selling, and analyst disagreement. It reports that single-channel measures are weaker and less stable than the cross-modal measure. These are preliminary paper claims requiring replication and careful label-timing review; they do not establish that an investment firm uses facial, vocal, or deception analysis.
• GIFfluence: A Visual Approach to Investor Sentiment and the Stock Market, by Ming Gu, David Hirshleifer, Siew Hong Teoh, and Shijia Wu, constructs a daily GIFsentiment index from millions of GIF appearances in Stocktwits posts from September 2020 through October 2024. The paper combines GIF-level valence inferred from bullish/bearish declarations with appearance weighting, then compares it with text, emoji, self-declared, and Seeking Alpha sentiment measures. It reports positive contemporaneous association and negative return predictability for up to a month, with volume and volatility relationships; the paper reports a one-standard-deviation association of 18.7 basis points with the contemporaneous S&P 500 return and a 191.4-basis-point lower return in the first month. These are paper-reported estimates, not a validated live signal or evidence of a firm’s social-media pipeline.
• The Growth and Performance of Artificial Intelligence in Asset Management, by Shuang Chen, Clemens Sialm, and David X. Xu, combines Form ADV and other regulatory disclosures, labor-market data, archived strategy descriptions, and fund-level performance. The authors define AI-driven investing around predictive modelling and trading-signal generation rather than generic workflow automation, report that 60% of their 2024 AI-fund sample is systematic diversified macro, and study 7,896 U.S. hedge funds from 2006–2024. Their abstract and introduction report earlier benchmark-adjusted differences that decline after 2017, a residual sibling-fund association, lower return comovement, and no extra inflows attributable to mentioning AI alone. These are preliminary academic estimates with classification and survivorship questions to audit; they are not a firm-by-firm ranking or evidence about any particular manager.
• Beyond Patent Ownership: Learning About Technological Usefulness, by Jesus Gorrín and Rory Mullen, uses NLP and positive-unlabeled learning to match descriptions of public firms with technology categories inferred from patent descriptions. The proposed dataset covers all U.S. public firms, a sample of 50,000 utility patents per year, hundreds of millions of firm–patent pairs, and nearly three decades of technology change; its output is a continuous usefulness probability rather than a binary patent-ownership flag. The paper argues that investors may inefficiently process technological information. This is a potentially useful route for technology-exposure and non-patenting-firm research, but it is a preliminary working paper and does not reveal an investment firm’s data, model, or deployment.
• AI Bubbles with Large Language Models, by Álvaro Cartea, Patrick Chang, Nan Chen, and Mingyue Zhong, uses a sequential bubble game in which an LLM decides whether to buy an asset with zero fundamental value and attempt resale at a ten-times price. The paper reports that greater reasoning effort reduces—but does not eliminate—irrational speculative trades, while agents can coordinate on a speculative equilibrium when multiple equilibria exist; it also identifies simplified market beliefs and framing sensitivity as mechanisms. This is a controlled market-design experiment funded by Oxford–Man, not evidence that a trading desk uses the prompt, model, or behavior.

The replay corpus changes the academic route from a list of people to a set of inspectable artifacts: multimodal CEO communication, visual social-media behavior, regulatory-and-labor disclosure classification, technology-use inference, and controlled LLM market experiments. The next step is to download permitted replays, preserve hashes, recover audio or captions, and align private timestamped transcripts to the linked papers. Academic results, conference attendance, and manager deployment remain separate evidence states.

Recovery queue for Oxford–Man 2026 replays and papers

Recover the direct MP4 files through browser or alternate publisher delivery, compute hashes, extract captions or audio, and retain timestamped transcripts privately. Review the full methods, appendices, code, data licenses, sample construction, model prompts, and point-in-time timing for each linked paper. Search the named authors and speakers across faculty profiles, seminars, practitioner events, and employer biographies; do not infer a fund system from Oxford–Man or Man Group adjacency.

September 5, 2026 — Canadian finance-lab and quant-programme routes

• The University of Waterloo Master of Quantitative Finance describes a 16-month programme spanning mathematics, statistics, econometrics, machine learning, computer science, and finance, with research-paper and thesis options. Its employer-facing page adds a four-month internship, electives in deep learning, machine learning, portfolio optimisation, numerical computation, and risk management, and a stated applicant pool of more than 150 qualified candidates for approximately 15–20 places. This is a concrete talent and curriculum surface; Waterloo does not publish a named hedge-fund project, proprietary dataset, model ownership, or deployment result on these pages.
• The BMO Financial Group Finance Research and Trading Lab at Rotman describes a lab founded in 1999, renamed after BMO support in 2013, with 71 workstations, real-time market-data and information feeds, research portals, industry applications, and custom simulation-based learning. Its stated scope includes portfolio management, financial engineering, risk, market microstructure, and trading. The facility is an education and research infrastructure signal; the page does not expose student strategies, vendor contracts, private data rights, or investment performance.
• Rotman FinHub provides a complementary project and personnel route. Its public research list includes press-release return prediction using more than 138,000 releases and BERT-based embeddings, online learning, and interpretability; its course and case pages describe Python, web/API data collection, portfolio formation, trading-strategy evaluation, supervised/unsupervised/reinforcement learning, and finance applications. The page’s research-catalyst alumni list includes paths to OMERS, JPMorgan Quantitative Strategies, Q Wealth Partners, CIBC World Markets, RBC Capital Markets, CDPQ, CPP Investments, and Visa Consulting & Analytics. These are public programme and alumni fields, not evidence that an employer received a particular model or data asset.
• The Imperial Centre of Excellence in Quantitative Finance joins finance, mathematics, computing, and industry collaboration. Its public events page lists the completed 18th Annual Hedge Fund Conference on Quantitative Investing hosted with Goldman Sachs on July 1, 2026, the November 10–11, 2026 CFM–Imperial Workshop on Quantitative Finance hosted with Deutsche Bank, and a prior market-microstructure workshop hosted with UBS. The leadership page identifies Robert Kosowski and Johannes Muhle-Karbe as co-directors; it states that Muhle-Karbe collaborates with hedge funds and banks. These are institutional and event-network signals, not evidence of shared models, data access, or deployment.
• Oxford–Man’s 2027 Finance and AI event page is a future monitoring point for March 18–19, 2027. The page confirms the date and venue but says the speaker list and registration will follow. It should be tracked as an upcoming academic-media route rather than treated as a speaker or content discovery until the programme is published.

These Canadian and UK routes expose three distinct layers: a selective quantitative-finance training pipeline, physical market-data and simulation infrastructure, and recurring academic–industry convenings with direct replay or paper surfaces. They improve personnel and model discovery without supporting a league table or any inference that academic content is used by a named investment firm.

Recovery queue for Waterloo, Rotman, and Imperial routes

Recover Waterloo’s current faculty, advisory board, internship employers, student research papers, and placement records; Rotman’s current FinHub staff, alumni publications, case-study files, data provenance, and BMO lab resources; and Imperial’s 2026 conference flyers, speaker rosters, recordings, papers, advisory board, and CFM workshop materials. Preserve programme marketing, alumni claims, event attendance, research artifacts, and verified employer deployment as separate fields.

September 5, 2026 — firm-funded doctoral routes and Asia-Pacific finance curricula

• Cambridge’s G-Research Trinity PhD Studentship is a direct firm-to-university talent interface that was not visible in a manager-name or AI-lab search. The 2026/27 award supports one PhD student in Engineering, Computer Science and Technology, or Pure Mathematics and Mathematical Statistics, with a stated preference for machine-learning and AI-modelling work. The page specifies 3.5 years of funding, a £1,500 annual conference/travel grant, biannual mentorship with a G-Research quantitative researcher, participation in G-Research’s Spring into Quant Finance programme, seminar invitations, and an annual dinner. This establishes a recruiting, mentoring, and research-exposure mechanism; it does not identify the selected student, supervisor, research topic, proprietary data, model, or transfer into a G-Research production system.
• Curtin University’s Finance and Investment Analytics major describes an applied postgraduate route linking programming, statistics, econometrics, financial-market operations, and predictive modelling. The current page says students develop predictive models for investment-strategy formulation, study machine learning alongside modern econometrics, and complete an industry-inspired project. Curtin dates the page’s update to September 3, 2026. This is a current curriculum and project surface, not evidence of a named sponsor, dataset, model implementation, or investment result.
• The University of Queensland Digital Finance Research Hub names Associate Professors Min Zhu and Sergeja Slapnicar as leads and states that the hub works with government, regulators, financial institutions, and technology firms. Its research themes include cyber-risk stress testing, algorithmic decision-making and AI governance, data assets and data markets, privacy-preserving data sharing, digital assets, and AI in capital markets. The page also announces the Digital Transactions in Asia VII conference for December 11–12, 2026 and describes a large-scale market-and-user-data project across India, Indonesia, China, Japan, Malaysia, Singapore, Vietnam, and the Philippines. This exposes a regional data-governance and market-research route; it does not disclose a hedge fund’s data rights, model, or deployment.
• Oxford’s MSc in Mathematical and Computational Finance makes the academic-to-industry model vocabulary unusually explicit. The current page describes a ten-month programme covering stochastic control, C++, Python, PyTorch, deep learning, Monte Carlo, optimisation, derivatives, fixed income, decentralised finance, and market microstructure. It says the fourteen-faculty group has industry ties with banks, hedge funds, central banks, and exchanges, and that students complete a supervised dissertation that may run alongside an industry internship. These are programme and faculty-group claims; the page does not identify which projects use which partner data or whether any dissertation enters a production investment process.

This tranche adds a firm-funded doctoral relationship, a project-based predictive-investment curriculum, a regional hub combining AI with data governance and market infrastructure, and a mathematically intensive programme with explicit industry interfaces. The G-Research relationship is stronger evidence of a talent channel than of a deployed model; all four routes remain separate from verified production evidence.

Recovery queue for G-Research, Curtin, UQ, and Oxford programme routes

Recover the selected G-Research student and supervisor when publicly disclosed, research topic, seminar and Spring into Quant Finance material, and later employment or publication trail. Recover Curtin’s unit handbook, project descriptions, teaching staff, and industry-project outputs; UQ’s current team, projects, conference programme, panel recordings, data-governance work, and student researchers; and Oxford’s faculty, dissertation topics, internship interfaces, alumni destinations, and industry-linked research artifacts. Preserve scholarship funding, curriculum exposure, academic research, recruiting, and verified firm deployment as separate fields.

September 5, 2026 — professor-to-programme cross-check and student trading infrastructure

• Stanford’s report on faculty use of AI for large-scale text analysis names finance professors Antonio Coppola and Matteo Maggiori in a Global Capital Allocation Project that processed more than 780,000 earnings-call transcripts and analyst reports from more than 21,000 companies over more than a decade on Stanford high-performance-computing clusters. The report describes semantic LLM analysis of tariff-related business responses and a public dashboard. This is a professor-led research and data-engineering route that can inform our earnings-call corpus and model-audit vocabulary; it does not disclose a hedge fund’s system, data rights, or investment deployment.
• The University of Chicago Financial Markets Program is a three-year student pipeline with coding, machine-learning training, trading simulations, price-prediction modelling, practitioner workshops, and employer site visits or presentations. Its current page names programme activity involving Two Sigma, Walleye Capital, Millennium, DRW, Citadel, Jane Street, Optiver, and others. This is useful for discovering faculty, speakers, student projects, and recruiting language; the page does not establish that any named firm supplied a model, dataset, or production workflow.
• Stevens’ 2026 High Frequency Trading Competition exposes an unusually concrete training environment. Students design and run intraday strategies on the SHIFT simulation platform, with realistic market microstructure, low-latency conditions, fees and rebates, daily capital resets, and Python/GitHub/FIX requirements. The public rules and examples include market making, momentum, statistical arbitrage, microstructure-aware execution, backtesting, optimisation, risk controls, documentation, and reproducibility. The competition is a student simulator and talent signal; it is not evidence of a tracked manager’s execution stack, live capital, or performance.
• The University of Iowa profile for Ashish Tiwari adds a professor-led model-evaluation route that should stay linked to the existing hedge-fund-benchmark paper. Tiwari describes current projects on flexible ML-based benchmarks for nonlinear, time-varying hedge-fund exposures, BART-based stochastic-discount-factor sparsity, and Bayesian safeguards against false positives in alpha search. His page also lists prior work on hedge-fund replication and model combination. These are academic methods for evaluating exposures and research claims, not evidence that a named manager uses them.

This cross-check reinforces a useful evidence split. Professors and finance programmes can expose model classes, data construction, evaluation discipline, student-to-employer pathways, and the language used to describe research. They can also identify speakers and potential personnel lineage that manager-name searches miss. They cannot, by themselves, establish a firm’s internal AI strategy, proprietary data access, live authority, or performance. The article therefore records these as academic, programme, recruiting, or practitioner-interface evidence until a separate firm-level source closes the gap.

Recovery queue for professor and programme cross-check routes

Recover the Stanford dashboard, paper, prompts, code, and timestamped earnings-call annotations; UChicago programme staff, faculty advisers, employer events, student projects, and trading-competition artifacts; Stevens’ competition handbook, simulator documentation, winning strategies where releasable, faculty supervision, and alumni paths; and Tiwari’s current working-paper files, code, seminars, and coauthor lineage. Keep curriculum exposure, academic claims, employer contact, former employment, and verified manager deployment in separate evidence states.

September 5, 2026 — Japan and South Africa academic finance-AI routes

• The University of Tokyo-linked Matsuo–Iwasawa Lab financial-market machine-learning course is already captured in the article as a Japanese-language course route. Its 2026 schedule explicitly covers market prediction, dataset creation, labeling, backtesting, CTA operations, stochastic optimisation, and post-LLM strategy investment, with named university and practitioner instructors. It is a curriculum and talent-network signal, not evidence of a live fund system.
• Hiroshima University’s 2026 “Data Science of Algorithmic Finance” syllabus names Ting Hian Ann as instructor for a graduate course taught in English. The official record covers alternative ETF construction, market microstructure, algorithmic trading, mathematical modelling, computational implementation, and statistical tests; its plan includes stock indexes and ETFs, financial statistics, random-walk and variance-ratio tests, asset-pricing models, options, volatility indexes, static replication, and GARCH. The page also identifies homework, data analysis, and quizzes as detailed teaching methods. This exposes a concrete academic model and validation vocabulary, but not a firm sponsor, proprietary dataset, production system, or investment result.
• The University of Johannesburg repository’s open MCom thesis, “Machine learning applications in quantitative finance: a South African perspective” identifies Ryno du Plooy and supervisors P. J. Venter and M. Booysens. The record links related work on neural networks for vanilla-option pricing in South Africa, comparisons of artificial-neural-network and bootstrap-aggregating ensembles for derivative pricing, and ANN approximation of option Greeks in classical and multi-curve settings. The repository record is useful for paper, supervisor, and regional-lineage discovery; it does not by itself establish a fund connection, tradable signal, data rights, or deployment.
• Wits University’s statistics staff directory identifies Dr Eden Gross as a lecturer whose research interests include machine-learning applications in quantitative finance, financial risk management, and actuarial science. The same directory identifies Professor Charles Chimedza as Head of School and lists related interests in machine learning, deep learning, robust regression, clustering, and time series. This is a faculty-discovery route for papers, students, and research collaborators; it is not evidence of a hedge-fund laboratory or live strategy.

This pass adds four distinct public research surfaces: a Japanese-language course already in the ledger but worth retaining in the cross-check, an English graduate syllabus that makes the statistical-testing workflow explicit, an open South African thesis and supervisor graph, and a Wits faculty route linking ML to quantitative finance and risk. These sources help map model families, validation habits, academic lineages, and regional talent. They do not support a ranking of firms or an inference that a tracked manager uses any method.

Recovery queue for Japan and South Africa academic routes

Recover Hiroshima’s full reading list, assignments, instructor publications, and any permitted course materials; retrieve the Johannesburg thesis PDF and cited papers with their methods, data windows, and code; and map Gross and Chimedza to current papers, doctoral supervisors, students, and seminars. Search Japanese, English, and South African institutional sources separately, and keep academic exposure, industry contact, current employment, and verified investment deployment as separate evidence states.

September 5, 2026 — Israel, Poland, and Vietnam professor/programme routes

• Reichman University’s profile for Dr Yael Eisenthal identifies her as co-head of the MA in Financial Economics and lists teaching in investment theory, fixed income, and derivatives. The profile records a Columbia finance PhD, computer-science degrees from Tel Aviv University focused on machine learning, prior quantitative-investing research and portfolio-management roles at Goldman Sachs Asset Management and Barclays’ Quantitative Portfolio Strategy group, later research at Eagle Trading Systems, and an investor-risk/product role at Pagaya. Its publication list includes credit-spread, credit-risk, hedge-fund peer-evaluation, and machine-learning work. This is a dated academic/practitioner lineage route; it does not disclose a current fund’s models, data rights, permissions, or performance.
• The University of Warsaw’s faculty competition notice announces a full-time Assistant Professor position in a teaching group for quantitative finance and machine learning, with an expected October 1, 2026 start. A separate University profile for Piotr Wójcik describes quantitative-finance research on constructing and testing algorithmic investment strategies using machine learning, deep learning, high-frequency financial quotations, and high-resolution satellite imagery. It also records coordination of a Horizon 2020 financial-supervision and technology-compliance consortium involving universities and financial-sector firms. These are faculty-recruitment and academic-project signals; they do not establish a hedge-fund relationship or live strategy.
• Kozminski University’s profile for Piotr Kotlarz identifies him as an Assistant Professor in Empirical Economic Analysis and a postdoctoral data-science researcher at Oxford. It records a Liechtenstein business-economics PhD, a Vienna University of Economics and Business quantitative-finance master’s degree, prior European Central Bank work, and research interests in ML and AI for financial markets, FX, Generative AI, and experimental methods. The page links to scientific publications but does not expose a manager sponsor, proprietary dataset, or production system.
• The University of Economics Ho Chi Minh City’s English-language graduate course record names “Machine Learning and AI in Finance” as course BAN606028. Its published scope includes regression, probabilistic modelling, Gaussian processes, neural and deep networks, interpretability, AR/GARCH and Box–Jenkins models, HMMs, state-space and particle filters, RNN/LSTM/GRU/CNN/autoencoders, and RL applied to derivative pricing, risk, trading, volatility, cryptocurrency prediction, portfolio optimisation, and wealth management. The course uses case studies and hands-on projects and explicitly compares ML tools with traditional econometrics. This is a detailed curriculum route, not evidence of student outputs, employer sponsorship, or investment deployment.

This geographic expansion adds four distinct signals: a finance-programme leader with directly documented quantitative-investing and ML training, an EU faculty and recruiting route linking algorithmic strategies to high-frequency and satellite data, a current academic bridge spanning Oxford and Kozminski with explicit GenAI/FX interests, and a Vietnamese course that exposes the full model ladder from econometrics through sequence models and reinforcement learning. These sources are useful for personnel lineage, paper recovery, programme scanning, and talent discovery; they do not support a firm ranking or deployment inference.

Recovery queue for Israel, Poland, and Vietnam academic routes

Recover Eisenthal’s publications, CV, courses, coauthors, and dated employer transitions; identify the Warsaw position’s eventual appointee and recover Wójcik’s papers, code, satellite-data terms, and Horizon 2020 outputs; retrieve Kotlarz’s linked publications, working papers, supervisors, and Oxford research; and obtain UEH’s assignments, instructor identity, readings, datasets, and student projects. Preserve academic, recruiting, former-employer, and verified manager evidence separately.

September 5, 2026 — Columbia financial-data-science course route

• Columbia’s Fall 2026 Financial Data Science and Machine Learning course names William G. Ritter as instructor and describes Python examples with real financial data, statistical learning methods intended to scale to large datasets, and a final project in which students build a trading strategy. The published topics include lasso, elastic net, cross-validation, Bayesian models, EM, support-vector machines, kernel methods, Gaussian processes, hidden Markov models, and neural networks. The registrar page shows 44 enrolled students against a 45-student cap as of September 4, 2026. This is a current professional-education and talent route with explicit model and project vocabulary; it does not identify student outputs, employer sponsors, proprietary data rights, or a live manager system.

The Columbia course adds a useful bridge between professor/programme discovery and implementation-oriented screening: the public curriculum exposes model families, scaling constraints, data modality, validation language, and a strategy-building deliverable. Those details can guide searches for student projects, faculty papers, public repositories, and employer event recordings, while remaining separate from evidence about any named investment firm.

Recovery queue for the Columbia course route

Recover the full syllabus, assignments, project rubrics, instructor profile and papers, public student artefacts, guest speakers, and any recorded course or seminar material. Preserve course description, student work, employer contact, and verified deployment as separate evidence states.

September 5, 2026 — China and Hong Kong finance-AI faculty and programme routes

• Tsinghua’s AI + Quantitative Finance course page, taught by Jian Li with named teaching assistants, exposes a graduate course that combines classical quantitative trading, factor models, alternative-data factor mining, genetic programming, machine/deep learning, LLMs, and agents. It requires coding projects and allocates most of the grade to project reports, presentations, and code. The page is headed “AIQUANT 2026 spring,” but an internal line says “2025 Spring,” so the course date needs reconciliation before being treated as a clean temporal observation. It is curriculum evidence, not proof of a fund’s system or use of any listed method.
• Tsinghua’s Algorithm–Data–Learning group page adds a personnel and lineage surface. The public group roster names Jian Li, doctoral and master’s researchers working on large language models, machine learning, quantitative finance, and learning theory, and alumni destinations labelled as private funds, Qianxiang Asset Management, Zhongguancun AI Institute, Qwen, and other technology organisations. These are self-published group-page research-interest and first-job labels; they require individual CV and employer reconciliation and do not establish a manager’s model ownership or deployment.
• Peking University’s Financial Engineering Laboratory teaching page publishes Chinese-language course and lecture material. Its quantitative-investment course covers market and financial data, factor models, index enhancement, hedging, statistical arbitrage, portfolio allocation, high-frequency data/trading, and machine learning with program implementation and simulated investment analysis. A separate 2025/26 machine-learning-and-asset-pricing course is attributed to Li Xinping, whose profile on the same page describes prior public-fund investment leadership, U.S. hedge-fund quantitative-strategy and investment-manager work, Stanford economics training, and research in quantitative trading, asset pricing, ML, and financial text. The page also advertises a March 2026 lecture by Guangfa Securities’ financial-engineering chief on OpenClaw deployment and investment-research applications. These are Chinese first-party teaching and biography claims; they do not identify a current fund system, data permissions, or performance.
• Shanghai Jiao Tong University’s May 2026 “Large and Deep Factor Models” seminar page names Yuan Zhang of Shanghai University of Finance and Economics and describes a deep-neural-network stochastic-discount-factor decomposition through a Portfolio Tangent Kernel. The abstract connects the learned representation to nonlinear characteristic discovery, factor pricing, portfolio optimisation, and finite-sample complexity; the biography identifies Swiss Finance Institute/EPFL doctoral training, an MIT visiting period, and work on benchmarking LLMs in quantitative finance. These are speaker-page and abstract claims requiring recovery of the underlying paper, code, and data definitions; they do not establish a fund relationship or production use.
• CUHK Shenzhen’s bilingual SAI Spring 2026 course page names Tsang Ka Wai for a course on using AI tools for quantitative financial-market analysis and systematic-strategy development. The syllabus combines portfolio optimisation, derivatives pricing, volatility forecasting, statistical arbitrage, empirical validation, rigorous backtesting, and practical deployment challenges with multi-agent systems and “Data Spaces” intended to constrain agent access and preserve data/process traceability. This is a rare public route joining finance modelling, agent architecture, data governance, and backtesting in one course; it remains curriculum evidence rather than a manager-deployment claim.
• City University of Hong Kong’s profile for Li Wei identifies him as an adjunct professor whose research areas include quantitative investment, machine learning, big data, operations research, reliability engineering, and computational finance. The first-party profile also describes him as Head of Multi-Asset Investments at BNP Paribas Securities (China), with prior Citigroup and BNP Paribas experience in London and Hong Kong and responsibility for multi-asset quantitative investment. These are public biography claims and a practitioner–academic bridge; they do not disclose the employer’s models, data rights, trading authority, or results.

This regional tranche adds three different research signals: a public model-and-agent curriculum, an academic-to-practitioner personnel graph, and a Chinese-language route into finance-lab lectures and student projects. The most actionable follow-up is artifact recovery—Chinese/English syllabi, lecture recordings, project code, paper versions, group alumni CVs, and employer biographies—while keeping course content, self-reported lineage, current employment, and verified deployment distinct.

Recovery queue for China and Hong Kong academic routes

Reconcile Tsinghua’s AIQUANT date and recover its full syllabus, projects, Jian Li’s publications, group repositories, and alumni CVs. Recover Peking University’s Chinese course assets, OpenClaw lecture recording, lecturer biographies, and any public simulation code. Retrieve Yuan Zhang’s full paper, code, data, and academic lineage; CUHK Shenzhen’s assignments, Data Spaces materials, projects, and recordings; and Li Wei’s CV, publications, seminars, and independently corroborated BNP Paribas role. Preserve Chinese-language originals with translations and do not infer production deployment from an academic or teaching association.

September 5, 2026 — geospatial finance and alternative-data research route

• Boston University’s profile for Chishan Zhang identifies him as a Global China Post-doctoral Research Fellow with a co-appointment at the BU Center for Remote Sensing. The profile describes a framework combining the Global Development Policy Center’s finance databases with satellite imagery and machine learning to assess how China’s overseas development finance affects land use, carbon emissions, and biodiversity. It records doctoral training in geography and environmental studies at the University of Illinois Urbana-Champaign and prior computational-science fellowship work. This is a concrete geospatial-finance data and personnel route for discovering alternative-data methods; it does not disclose a tradable signal, hedge-fund relationship, data licence, or investment deployment.

The route expands the academic search beyond market-price and text data: finance databases joined to remote sensing can support event timing, infrastructure, environmental exposure, supply-chain, and country-risk research. The article treats those as research hypotheses until source-level data rights, labels, point-in-time availability, and out-of-sample testing are recovered.

Recovery queue for geospatial-finance research

Recover Zhang’s publications, project datasets, satellite sources, geospatial methods, code, supervisors, collaborators, and conference presentations. Check whether any public finance database or imagery provider imposes licensing or temporal restrictions, and keep academic environmental-impact modelling separate from a portfolio-signal or manager-deployment claim.

September 5, 2026 — European professor and finance-syllabus routes

• ESCP Business School’s profile for Mathis Mörke identifies him as an Assistant Professor of Finance in Paris whose research covers asset pricing, investments, derivatives, machine learning, and big data. The profile lists “Option Return Predictability with Machine Learning and Big Data,” a forthcoming “Bayesian Stochastic Discount Factor for the Cross-Section of Individual Equity Options,” “Option Factor Momentum,” and the 2026 forthcoming “Machine Forecast Disagreement,” coauthored with Turin Bali, Bryan Kelly, and J. Rahman. It records a University of St Gallen finance PhD, a Kellogg visiting-scholar period, and quantitative-finance training at ETH Zurich and the University of Zurich. This is a named professor, paper, and academic-lineage route; it does not establish a fund affiliation, proprietary data, or deployment.
• The Barcelona School of Economics 2026 Machine Learning for Finance syllabus names Professor Argimiro Arratia and specifies a three-ECTS course using R/Python, financial time series, feature selection, stationary bootstrap, neural networks, RNN/LSTM, Gaussian processes, sentiment analysis, algorithmic trading, deep-learning portfolio management, alternative-data factor models, robust optimisation, portfolio replication, and ML option valuation. The syllabus requires programming experiments, simulations, and team take-home work rather than a written exam. This is a particularly useful public curriculum map for model families and validation language; it does not disclose student projects, datasets, employer sponsors, or a live investment system.

Together these sources extend the academic map from finance-AI survey language into two inspectable artifacts: a professor’s current paper and lineage surface, and a course syllabus that makes alternative-data, uncertainty, sentiment, portfolio, and option-model workflows explicit. The next step is to recover the linked papers, code, data definitions, assignments, and student outputs while keeping academic research separate from manager deployment.

Recovery queue for European professor and syllabus routes

Recover Mörke’s personal site, working papers, code, coauthor and supervisor lineage, seminars, and any independently verified industry interface. Recover the BSE course’s complete reading list, assignments, data terms, project artefacts, instructor materials, and recordings. Preserve paper claims, syllabus content, student work, employer contact, and verified deployment as separate evidence states.

September 5, 2026 — academic implementation artifacts and family-office media

• UCLA’s MATH 279 Data Science and Machine Learning for Finance syllabus, taught by Mihai Cucuringu, is a project-based course that spans dimensionality reduction, networks, clustering, ranking, high-dimensional noisy data, limit-order-book modelling, order-flow imbalance, lead–lag detection, price impact, optimal execution, realized volatility, covariance forecasting, factor models, change-point detection, and systematic-strategy backtesting. The syllabus moves from minute and daily data to high-frequency limit-order-book data and requires a reproducible project report. This is a concrete public map of research questions and validation requirements; it does not identify a manager sponsor, proprietary dataset, live exchange access, or production deployment.
• Georgia Tech’s Practice of QCF Fall 2026 syllabus, taught by Sudheer Chava, exposes a different bridge between university projects and the buy side. It describes industry-mentored projects using proprietary systems or databases under NDA, alongside public or Georgia Tech-accessible data through WRDS or course servers. The listed project families include quantitative trading, ML prediction, NLP/LLMs in finance, Bayesian and ML credit-risk models, alternative-data acquisition, NLP feature generation, and ML prediction. The syllabus also warns that licensed data must remain within the permitted class use and notes that students may perform data annotation. This establishes a project and data-governance route, not the identity of a sponsor, access to a named fund’s systems, or deployment of a student model.
• BlackRock’s 2025 Global Family Office Survey release reports interviews and surveys with 175 single-family offices collectively overseeing more than $320 billion. BlackRock says 45% were more likely to invest in companies building AI solutions and 51% in opportunities expected to benefit from AI, while 33% reported deploying AI internally to improve the investing process. It also reports gaps in reporting, deal-sourcing, and private-market analytics. This is an allocator-demand and operating-capability baseline, not evidence about any named office’s model, vendor, data rights, or investment results; the figures are publisher-reported survey results and should retain their denominator.
• The UHNW Institute Podcast episode “Integrating AI into the Family Office”, published December 15, 2025, names host Kristen Oliveri and guests Tania Neild of InfoGrate Wealth and Bill Wyman and Dan Gregerson of Summitas. The public episode notes point to reporting, data analysis, investment research and monitoring, governance, privacy, team literacy, human judgement, and technology-partner selection. This is a title-blind family-office media route with named practitioners; the captured page does not provide the full spoken transcript, model inventory, data sources, or a specific office’s implementation.

The academic artifacts add a sharper implementation vocabulary to the professor route: high-frequency microstructure, execution, reproducible project work, NDA-controlled data, annotation, WRDS access, credit risk, and NLP feature generation. The family-office sources add an allocator-side baseline: public material currently exposes adoption rates, capability gaps, governance, and workflow themes more readily than internal model training or autonomous investment authority. Neither source family supports a cross-firm ranking. The next recovery work is to obtain course projects and recordings, identify only publicly disclosed Georgia Tech sponsors, recover the UHNW audio/transcript, and preserve survey wording and denominators.

Recovery queue for these artifacts

Recover the UCLA MATH 279 reading list, project milestones, code, and data terms; Georgia Tech’s public project and guest roster while preserving NDA boundaries; Monash assignments, instructor research, and student outputs; and the BlackRock survey instrument or cross-tabs. Download and transcribe the UHNW Institute episode with timestamps, then search each named guest and linked technology provider independently. Keep curriculum, survey, practitioner commentary, and verified investment deployment as separate evidence states.

September 5, 2026 — student investment funds and allocator-network surfaces

• Imperial Business School’s Student Investment Fund says its 2025–26 student-led fund deploys live capital and is organized into fundamental equity research, quantitative strategy research, quantitative engineering, portfolio management, and marketing departments. This is a rare public organizational map for a student-managed fund: it exposes the separation between research, engineering, portfolio decisions, and capital deployment, and links to leadership, supervisors, departments, and industry-partner pages. The public home page does not identify capital size, models, data vendors, decision logs, or performance, so it should be treated as a talent and workflow route rather than a proxy for institutional-fund practice.
• George Washington University’s Spring 2026 GW Quant Fund Pitch Day report provides a more concrete student-model artifact. The report says the fund had more than $160,000 as of May 5, 2026 and describes QuantFlow Trading’s LightGBM momentum approach using trend indicators, volatility targeting, regime detection, Fama–French factors, sector adjustment, and portfolio-risk controls. The same page embeds three reflection videos and describes the broader GW Investment Institute as managing university endowment money across four student funds. These are university-published student-project and programme claims; they do not establish independent performance, production-grade controls, data licensing, or use by a tracked hedge fund.
• The Chicago Booth Family Office Summit 2026 hub reports a May 2026 event with 200 attendees from 150 family offices across the United States, Canada, EMEA, and Latin America. It describes faculty-led research, practitioner sessions, student engagement, and an account-gated attendee directory, bio book, videos, presentations, and recap material. This is a valuable route for public follow-up: names and artifacts may become discoverable through speakers’ own pages, university announcements, or later media. The hub itself does not expose the gated directory, confidential discussions, individual office systems, or AI implementation, and no account-gated material is treated as public evidence here.

These routes make the academic-to-industry boundary more tangible. Imperial exposes organizational decomposition around live student capital; GW exposes a named model recipe and embedded media; Chicago Booth exposes a high-density family-office network and a clear public/private boundary. None supports a firm ranking or an inference that a professional manager uses the student-fund methods.

Recovery queue for student funds and allocator networks

Recover Imperial’s leadership, supervisors, department pages, public videos, and disclosed industry partners; GW’s embedded videos, full project presentations, student-team identities, and public data/model documentation; and Chicago Booth’s non-gated announcements, public speakers, faculty research, and post-event materials. Do not attempt to bypass the Booth account gate or recover confidential attendee information. Preserve student work, programme claims, allocator-network metadata, and verified professional deployment separately.

September 5, 2026 — additional academic faculty and quantitative-finance curriculum routes

• Illinois Tech’s profile for Benjamin E. Van Vliet identifies him as an Associate Professor of Finance, Director of the Center for Strategic Finance, and co-editor-in-chief of Algorithmic Finance since 2024. The profile lists teaching in machine learning for finance, Python financial modelling, high-frequency finance, C/C++, and automated trading-system design and development. It also records research interests in high-frequency trading, machine learning in finance, and fintech/innovation, plus more than 40 listed research articles and books on automated trading. This is a practitioner-facing academic and programme route; it does not identify a fund sponsor, proprietary data, or a live strategy.
• Sewanee’s profile for Huarui Jing describes an Assistant Professor of Finance whose work combines asset pricing, financial econometrics, macro finance, and machine learning. The profile names current projects on distributionally robust optimisation in asset-pricing models, robustness of nonparametric recursive-utility models, and high-dimensional asset-pricing DRO problems. It records a 2021 economics PhD from the University of Connecticut, an MS from Illinois, and earlier statistics and public-economics degrees from Shanghai University of Finance and Economics. This is a clear robustness-and-lineage route for paper recovery; it does not establish a manager connection or investment deployment.
• Rutgers–Camden’s profile for Wei Jiao links an Assistant Professor of Finance to international investments, investor behaviour, machine learning, and an “Investment Management and Machine Learning” course. The page lists work on global mutual funds and multinational-firm returns, country rotation and international mutual-fund performance, post-event returns in global markets, and manager home-country culture. It records a Binghamton University finance PhD and presentations at the AEA and AFA. This route broadens the academic search from model architecture into international allocation, manager behaviour, and cross-border outcome variables; it does not establish any fund’s use of the methods.
• Singapore Management University’s quantitative-finance curriculum exposes a sequence from quantitative trading and limit-order-book practice to QF209 Machine Learning in Quantitative Finance and QF210 Reinforcement Learning in Quantitative Finance. The page describes deep neural networks and reinforcement learning applied to stock-price prediction and portfolio management, and Python work for dynamically constructing and adapting portfolios. Its QF206 description also refers to an online financial-markets simulation and an MSCI Singapore Free Index futures limit-order-book case. This is curriculum and simulator evidence, not a named employer’s execution stack, data licence, live authority, or performance.

These routes add four distinct research lenses: trading-system engineering, distributional robustness, international manager and market behaviour, and reinforcement-learning/portfolio curriculum. They are useful for expanding professor, student, seminar, and alumni searches, but the source boundary remains unchanged: academic profiles and curricula reveal research vocabulary and talent pathways, not a tracked firm’s production system or results.

Recovery queue for additional academic routes

Recover Van Vliet’s paper list, course syllabi, journal archive, and public talks; Jing’s working papers, code, coauthors, and doctoral lineage; Jiao’s papers, course materials, international datasets, and seminar recordings; and SMU’s project briefs, simulator documentation, assignments, and faculty/student outputs. Keep academic research, simulator exposure, employer contact, and verified deployment separate.

September 5, 2026 — India and Indonesia professor-led finance-AI routes

• Universitas Indonesia’s profile for Zaäfri Ananto Husodo identifies him as Professor of Financial and Market Risk Management and leader of the Computational Intelligence for Finance, Business and Risk research group. The profile lists asset pricing, market microstructure, AI-based risk modelling, liquidity, volatility, systemic risk, digital assets, news sentiment, and high-frequency financial modelling. It also exposes 2025–26 projects on price discovery and asymmetric risk premiums in fiat-backed stablecoins, high-frequency crypto/US-stock volatility spillovers, and sentiment analysis of news, narratives, and social media for US and Indonesian firms. This is a concrete ASEAN research-group and data-modality route; it does not establish a fund relationship, proprietary data, or live deployment.
• IIM Indore’s faculty profile for Devika Arumugam records IIT Madras finance PhD and integrated economics training, a Fulbright award, an Emory Goizueta visiting-researcher role, and specialization in algorithmic trading. Her listed papers cover intraday tail risk, intraday profitability and trading behaviour, volatility exploitation, commonality and contrarian trading among algorithmic traders, and causal links between proprietary/buy-side algorithmic traders and market quality. This is a named paper and academic-lineage route; it does not establish any manager’s implementation, data rights, or performance.
• IIT Ropar’s profile for Dr Puneet Pasricha identifies research areas in mathematical finance and ML in finance and records an IIT Delhi PhD and master’s degree, a scientific-collaborator role at the Swiss Finance Institute/EPFL, and a University of Wollongong fellowship. Its publication list includes a 2025 paper on European option pricing under regime switching using physics-informed residual learning, alongside work on stochastic liquidity, Hawkes jump processes, credit risk, and portfolio optimisation. This adds a model-family and lineage route connecting physics-informed learning to derivatives and credit; it does not disclose a fund sponsor, code, data, production use, or investment results.

This pass extends the regional map into Indonesia and India with three different research surfaces: a professor-led computational-intelligence group using high-frequency, sentiment, and digital-asset data; an algorithmic-trading research programme focused on intraday risk and market quality; and a mathematical-finance lineage using physics-informed learning, stochastic liquidity, and Hawkes processes. These are discovery routes for papers, code, students, and seminars, not evidence for ranking firms or inferring deployment.

Recovery queue for India and Indonesia academic routes

Recover Husodo’s current papers, project methods, datasets, research-group members, code, and seminars; retrieve Arumugam’s full papers, data windows, identification designs, coauthors, and IIT Madras lineage; and recover Pasricha’s 2025 paper, code, calibration/data definitions, doctoral supervisors, EPFL collaborators, and related student work.

September 5, 2026 — Malaysian finance programmes and Indian student-research pipelines

• The University of Malaya Department of Finance staff directory, last updated July 31, 2026, publicly lists Dr Asyraf Abdul Halim’s expertise in asset-pricing anomalies, Islamic and ESG asset pricing, financial econometrics, portfolio optimisation, and AI/machine learning. The same directory lists fintech, digital finance, investment, derivatives, and capital-markets coverage elsewhere in the department. This is a faculty and programme-discovery surface for Islamic-finance, ESG, portfolio, and ML intersections; the directory does not disclose a fund relationship, proprietary data, live system, or performance.
• The University of Reading Malaysia / Henley Business School academic page identifies Anthony Yap as Lecturer in Fintech and Machine Learning. It records a PhD in Economics from Universiti Putra Malaysia, an MSc in Finance, prior banking experience at Citibank, HSBC, and OCBC, and teaching in fintech and cryptocurrency, programming for finance, and machine learning in finance. The same institutional page describes BSc Finance and Business Management, a financial-dealing room, and faculty with banking, stockbroking, and investment-research experience. This links a named educator, applied financial-risk/ML teaching, and a practical training environment; it does not establish any hedge-fund deployment or access to bank data.
• IIT Guwahati Professor Natesan Srinivasan’s public student and thesis page, updated June 25, 2026, exposes a long-running project pipeline. Named student topics include computational finance and option pricing, portfolio management using ML, portfolio optimisation, Bitcoin algorithmic trading, financial-news sentiment and stock-price prediction, Forex reinforcement learning, deep-RL portfolio optimisation, tree-based market-regime detection, and agent-based market simulation. The page also identifies 2025–26 students working on Forex-RL, deep-RL portfolio optimisation, and market-regime detection. This is unusually specific evidence of student research themes and a talent-discovery route; it does not establish validated returns, production controls, licensed data, or adoption by a professional manager.

The Malaysian and Indian sources add three observable layers to the academic map: faculty expertise that combines portfolio and ML language, a finance programme with a dealing-room and programming/ML coursework, and a dated student pipeline that makes model families and project titles searchable. These are useful for finding papers, students, supervisors, seminars, and alumni; they are not evidence that a tracked hedge fund uses any listed approach.

Recovery queue for Malaysian and Indian academic routes

Recover the University of Malaya faculty profiles, papers, student theses, seminars, and any public industry projects; retrieve Anthony Yap’s publications, teaching materials, programme artefacts, and professional history; and follow the IIT Guwahati student names into repositories, papers, code, presentations, and current employers. Preserve curriculum, student projects, academic publications, employer contact, and verified production use as separate evidence states.

September 5, 2026 — Saudi and Southern African finance-AI faculty and lab routes

• The University of South Africa profile for Dr MM Mpanda identifies him as a Senior Lecturer in Decision Sciences and lists teaching in financial-risk modelling. His stated interests include stochastic-volatility modelling, financial econometrics, robust portfolio optimisation, financial machine learning, and interest-rate modelling; listed projects include calibrating fractional-volatility models with ML. The page also records a quantitative-finance programme-development role and 2025–26 work on BRICS+ market connectedness, volatility, and Heston option pricing. This is a named professor, programme, and model-family route; it does not establish a fund relationship, proprietary data, live use, or performance.
• The University of South Africa profile for Professor Ebenezer Esenogho says he joined UNISA in June 2025 and is establishing the Centre for Artificial Intelligence and Multidisciplinary Innovation Studies (CAIMIS). The university describes finance and business work involving fraud prediction/detection, AI auditing, volatile-stock forecasting using ARIMA and LSTM, and a feature-engineered neural-network ensemble for credit-card fraud detection, drawing on his earlier Institute for Intelligent Systems work at the University of Johannesburg. This is a first-party academic-centre and research-history signal; it does not establish a hedge-fund sponsor, investment model, data rights, or production deployment.
• The KFUPM Business School profile for Dr Muhammad Tahir Suleman records a PhD in Finance from Victoria University of Wellington, an MSc in Quantitative Finance from Hanken, and an MS in Financial Economics from the University of Skövde. His specialization includes quantitative finance, financial markets, climate and energy finance, and ML in finance. The listed research includes cross-market sentiment and gold-volatility predictability, ML models for industry-based stock-volatility prediction, high-frequency return dispersion, and Bitcoin fear-and-greed connectedness. These are paper and lineage routes; the profile does not establish professional-manager use or investable performance.
• KFUPM’s Master of Quantitative Finance programme and Interdisciplinary Research Center for Finance and Digital Economy expose the surrounding talent and institutional architecture. The master’s degree combines artificial intelligence for business, financial markets, stochastic processes, financial econometrics, derivatives, fixed income, and a project; the centre describes research themes in AI applications, fintech, quantitative economics and policy, and industry engagement. This is a programme-and-lab pipeline in Saudi Arabia, not evidence that a particular fund supplies projects, data, or trading authority.

This cluster adds model and workflow vocabulary that the firm pages rarely expose publicly: ML calibration of fractional stochastic-volatility models, robust portfolio optimisation, market connectedness, AI auditing, fraud detection, LSTM forecasting, cross-market sentiment, high-frequency dispersion, and an academic project pipeline. These are research surfaces for paper, code, student, and personnel recovery; they do not support ranking managers or inferring deployment.

Recovery queue for Saudi and Southern African academic routes

Recover Mpanda’s full papers, calibration methods, data definitions, student theses, and programme materials; retrieve Esenogho’s CAIMIS charter, researchers, finance papers, grants, and public outputs; and recover Suleman’s papers, code, student lineage, seminars, and KFUPM centre projects. Keep university claims, paper-reported results, industry contacts, and verified investment deployment as separate evidence states.

September 5, 2026 — Latin American finance-ML curricula and faculty interfaces

• Universidad Torcuato Di Tella’s Spanish-language “Análisis de Data en Finanzas” course page publishes a concrete implementation syllabus: Python/Jupyter notebooks, fractional differentiation, information-driven bars, entropy and market-microstructure features, time-series model validation, XGBoost, covariance denoising and detoning, Hierarchical Risk Parity, mixed-frequency data, and post-Markowitz allocation. A separate “IA en Finanzas” page describes intensive Python work, finance-specific models, and short weekly assignments. The pages name Pablo Roccatagliata, who teaches quantitative finance, visualization, and data strategy and previously led data at Digital House and taught algorithmic trading through ROFEX; they also name Lionel Modi, with IOV Labs and prior investment-business experience, and Matías Macazaga, a Crisil Argentina quantitative analyst responsible for ML-model validation. This is an unusually detailed Spanish-language talent and workflow route; it does not establish a fund’s use of the curriculum, proprietary data, or live performance.
• Tecnológico de Monterrey’s official profile for Luis Arturo Bernal Ponce describes a professor working across finance, AI, and educational innovation. It records leadership of the 2026 Financial Econometrics curriculum design for the Bachelor’s Degree in Finance, an approved “Market Anomalies with AI and Python Applied to Finance” project, generative-AI use as an applied consulting tool in Treasury and Accounting courses, faculty training in programming/ML, and business cases focused on algorithms and ML financial modelling. The profile also lists executive education in AI applied to finance and participation in the MIT Digital Currency Initiative. These are university profile claims about education and professional networks; they do not disclose a hedge-fund sponsor, production model, data rights, or investment results.

The Latin American route contributes a Spanish-language implementation vocabulary that is easy to miss in English-only searches: feature construction tied to market microstructure, information bars, entropy, fractional differentiation, validation, covariance cleaning, HRP, mixed-frequency data, and model validation as a named professional responsibility. Mexico adds a separate curriculum-innovation and applied-GenAI teaching route. These sources are useful for recovering instructors, students, papers, notebooks, cases, and alumni; they remain academic or educational evidence rather than manager-deployment evidence.

Recovery queue for Latin American academic routes

Recover Di Tella’s current syllabi, notebooks, assignments, instructors’ papers, ROFEX programme artifacts, and student/alumni outcomes; verify the named practitioners’ current roles independently; and retrieve Bernal Ponce’s cases, publications, course materials, innovation-project documentation, and public finance/AI seminars. Preserve Spanish-language curriculum, faculty biography, industry contact, and verified investment use as separate evidence states.

September 5, 2026 — Brazilian finance professors, theses, and supervisor pipelines

• University of São Paulo’s Lattes profile for Flavio Almeida de Magalhães Cipparrone, updated in August 2026, identifies an Associate Professor whose areas include operations research, optimisation, simulation, and quantitative finance. His public supervision list exposes a current student pipeline: 2026 projects on NLP and machine learning applied to media for financial-index prediction, HMM–LSTM economic-regime and inflation forecasting, ML analysis of ADR/ordinary-share premia, and quantitative investment decision support; earlier work includes ML applied to financial markets, optimal equity-order execution, and high-dimensional covariance estimation. This is a rare supervisor-level route into concrete student research topics and model families; it does not establish validated returns, a fund sponsor, licensed data, or production use.
• Insper’s 2025 professional-master’s thesis by Emerson Sousa Vieira studies Brazilian stocks’ monthly excess returns using Brazilian equity factors plus a broad macroeconomic feature set. The thesis names Elastic Net, PCR, PLS, random forests, and gradient-boosted regression-tree methods in its model discussion and reports paper-level out-of-sample comparisons with OLS and variable-importance analysis. Its abstract highlights Brazil country risk (EMBI), economic-expectations measures, a commodities composite, and credit-to-GDP among the macro variables identified as important. The thesis is a reproducible research route with a named supervisor, Gustavo Barbosa Soares; its reported result is not an independently audited live strategy or evidence of manager deployment.
• FGV EPGE’s doctoral-defence record for Gabriel Brum Cardoso documents a finance dissertation on machine learning in the Brazilian cross-section of returns. The public record names Felipe Saraiva Iachan as supervisor and Marcelo Fernandes and Gustavo Bulhões Carvalho da Paz Freire as examiners, linking FGV EPGE, FGV EESP, and Erasmus School of Economics. The page exposes the title, date, finance research line, and academic lineage, but not the full thesis methods, data, code, or results; it should be followed as a paper-and-supervisor recovery route rather than treated as a performance claim.
• FGV EAESP’s profile for Alan De Genaro identifies an Associate Professor of Finance with USP economics/statistics training and a Courant Institute postdoctoral appointment. His listed quantitative-finance areas include stress testing, market, credit, liquidity and model risk, asset pricing, ML, and blockchain. The publication list includes Brazilian-market spoofing detection, securities lending and short selling, OTC equity-lending price transparency, interest-rate derivatives, Monte Carlo option pricing, and central-counterparty liquidation costs. This adds a market-infrastructure and model-risk research route around ML, rather than a directional-prediction-only route; it does not establish a hedge-fund relationship, proprietary data, deployment, or performance.

The Brazilian pass adds a useful student-to-professor-to-paper chain. Public supervision records expose emerging project choices before they become polished publications; the Insper thesis provides a fully named asset-pricing design with macro features and model comparisons; FGV adds doctoral lineage and a market-infrastructure/risk faculty surface. These are routes for recovering papers, code, data definitions, students, supervisors, and industry interfaces, not evidence for ranking firms or inferring their internal systems.

Recovery queue for Brazilian academic routes

Recover Cipparrone’s linked student outputs, theses, repositories, code, data definitions, and current employers; retrieve Vieira’s complete thesis tables, splits, feature vintages, code, and replication materials; obtain Cardoso’s dissertation and supervisor publications; and follow De Genaro’s papers, coauthors, seminars, and public industry interfaces. Preserve student work, paper-reported results, faculty biography, employer contact, and verified investment deployment separately.

September 5, 2026 — Andean finance-ML labs, faculty, and recent-paper routes

• Universidad Externado de Colombia’s active “Machine Learning y Finanzas Computacionales” project record names Diego Ismael León Nieto as principal investigator and lists Javier Hernando Sandoval Archila, Carlos Armando Mejía Vega, and John Freddy Moreno Trujillo as co-investigators. The project’s stated objectives cover portfolio selection, time-series prediction, and asset valuation through ML, organized around data volume, velocity, and variety. Its public outputs include a deep-multilayer-perceptron one-minute HFT price-prediction presentation, deep learning and wavelets for high-frequency forecasting, and an interpretable automated-ML credit-risk model. This is a named university project and publication route; it does not establish a fund sponsor, proprietary data, live deployment, or performance.
• Universidad de Chile’s profile for Juan Díaz identifies an Assistant Professor with PhDs in Statistics from Harvard and Economics from Universidad de Chile. His research areas include econometrics, causal inference, statistics, and ML; the profile lists recent work on ML stock-market volatility, gold-risk-premium estimation, oil-and-gas volatility, real-estate returns using big data, and price effects of forced asset sales. This creates a Harvard–Chile faculty and paper lineage across volatility, commodities, real estate, and causal market events; the profile does not establish manager adoption, proprietary data, or live strategy use.
• Universidad EAFIT’s profile for Andrés Ramírez Hassan records a PhD in Statistical Sciences from Universidad Nacional de Colombia, an EAFIT master’s in finance, postdoctoral work at Monash, and a visiting role at Melbourne. It lists Bayesian econometrics, causal effects, and ML as research interests, a minimum-expected-loss portfolio-choice paper, and consulting or evaluation work for the Inter-American Development Bank, Colombia’s energy regulator, Medellín, EPM, and Grupo Nutresa. This is a Bayesian/causal finance and applied-industry interface; it does not establish a hedge-fund relationship, proprietary data, or investment deployment.
• Universidad del Valle’s finance faculty profile for Javier Humberto Ospina Holguín identifies a Full Professor in Finance associated with the Economic Value Generation and Financial Solvency/Risk groups. His listed interests include financial econometrics and forecasting, ML applications to finance, empirical asset pricing, portfolio theory, and econophysics; his education combines physics and finance/economics training. This is a Colombian professor and research-group discovery route; the profile does not expose current code, datasets, fund relationships, live systems, or performance.

The Andean pass adds a project-level bridge between high-frequency prediction, interpretable credit risk, and portfolio/valuation research, alongside faculty lineages in volatility, commodities, real estate, Bayesian inference, causal methods, and econophysics. These routes should feed searches for papers, student theses, code, seminars, and alumni. They remain academic and applied-research evidence, not evidence about a tracked manager’s internal system.

Recovery queue for Andean academic routes

Recover the Externado project’s papers, code, data definitions, HFT presentation, student outputs, and current affiliations; retrieve Díaz’s volatility and commodities papers, code, Harvard lineage, and seminars; follow Ramírez Hassan’s finance papers, industry project artifacts, students, and causal/ML methods; and recover Ospina’s publications, group members, student work, and seminars. Keep project statements, paper-reported results, industry consulting, and verified investment use separate.

September 5, 2026 — Turkish algorithmic-trading labs and finance-research lineages

• Middle East Technical University’s Algorithmic Trading Research Group page describes a multidisciplinary group focused on financial markets, quantitative modelling, and financial risk management. Its listed methods include machine-learning algorithms, algorithmic and high-frequency trading, risk management, and Monte Carlo methods; the page describes the intended integration of “mini-modules” into a real-time algorithmic-trading and financial-tools prototype. The page also says the group is being reactivated under a Portfolio Optimisation and Management name. This is a university research-group description and an explicit prototype objective, not evidence of production trading, a hedge-fund sponsor, proprietary data, or performance.
• Marmara University’s AVESIS publication profile for Mahmut Bağcı lists a 2026 Computational Economics paper, “The Optimal Threshold Selection for High-Frequency Pairs Trading via Supervised Machine Learning Algorithms,” coauthored with P. Kaya Soylu. The same profile lists teaching in computational finance and algorithmic trading. This is a current paper-level route into threshold selection, supervised learning, and high-frequency pairs-trading methodology; the profile does not establish live use, a manager connection, code, data rights, or investable performance.
• İstinye University’s profile for Nadi Serhan Aydın records BSc training at Istanbul Technical University, MSc and PhD work in mathematical finance at METU, doctoral research in Imperial College London’s Financial Signal Processing Lab, and research fellowships at Heidelberg and Ulm. The profile lists quantitative finance, computational optimisation, financial signal processing, and machine/reinforcement learning, and links work on reinforcement-learning-based optimal trading in simulated futures markets, financial signal processing with Lévy information, and stochastic optimisation. This is a named academic lineage and model-research route; it does not disclose a fund system or deployment.
• Kadir Has University’s profile for Associate Professor Oğuz Ersan lists market microstructure, behavioural finance, informed trading, high-frequency and algorithmic trading, and big-data analysis. It records a Yeditepe financial-economics doctorate and publications on probability-of-informed-trading estimation, a computationally adjusted PIN model, HFT detection from order and trade data, Borsa Istanbul market quality, and an R package for PIN estimation. This is a concrete market-microstructure and software-publication route; it does not establish an investment manager relationship, proprietary data, live use, or performance.

The Turkish pass adds a research-group architecture, a current supervised-learning paper, a cross-country academic lineage through METU and Imperial’s financial-signal-processing lab, and a reproducible market-microstructure software route. These details are useful for paper, code, student, and conference recovery. They remain academic or publication evidence and do not support a firm ranking or deployment inference.

Recovery queue for Turkish academic routes

Recover METU’s successor Portfolio Optimisation and Management page, member roster, prototype documentation, papers, code, and seminar recordings; retrieve Bağcı’s 2026 paper, data definition, code, coauthors, and citations; follow Aydın’s thesis, FSP Lab supervisors, papers, simulations, and current affiliations; and recover Ersan’s PIN package, datasets, HFT papers, coauthors, and conference appearances.

September 5, 2026 — Abu Dhabi allocator lineage and extreme-risk ML research

• The Hebrew University Business School’s profile for Professor Alexander Lipton describes him as Global Head of Research & Development at Abu Dhabi Investment Authority, Professor of Practice at Khalifa University, a visiting professor and Dean’s Fellow at Hebrew University, a Connection Science Fellow at MIT, and an advisory-board member at ADIA Lab. The same profile records earlier quantitative roles at Citadel, Credit Suisse, Deutsche Bank, and Bankers Trust; leadership of Bank of America’s Global Quantitative Group and Quantitative Solutions; visiting appointments at EPFL, NYU, Oxford, Imperial, and Illinois; and more than 100 scientific papers and 12 books as stated on the page. This is an academic-to-allocator personnel and lineage route. It does not disclose ADIA’s internal models, datasets, research handoff, investment authority, or performance, and ADIA Lab remains a separate entity from ADIA’s investment business.
• KAUST’s applications-to-finance research page identifies Raphaël Huser as an Associate Professor of Statistics working across extreme-value theory, spatio-temporal statistics, data science, machine learning, copulas, and applications to finance. The Extreme Statistics publication list includes 2026 work on the “efficient tail hypothesis” and market efficiency, alongside neural Bayes estimation, graph neural networks, variational autoencoders, and extreme-quantile regression. This provides a route into tail risk, dependence, and representation-learning research; it does not establish a fund relationship, a tradable signal, proprietary data, or deployment.

These routes connect allocator-side research leadership, university appointments, and extreme-risk methodology while keeping entities and evidence states separate. The follow-up paths are Lipton’s papers and current research interfaces, ADIA/ADIA Lab boundary checks, Huser’s finance paper and code, and student/coauthor lineages. No ranking or deployment inference is warranted from the profiles.

Recovery queue for Abu Dhabi and extreme-risk routes

Recover Lipton’s current publication list, seminars, university and ADIA Lab affiliations, and dated employer history from first-party sources. Retrieve Huser’s 2026 market-efficiency paper, code, data definitions, doctoral lineage, finance collaborators, and related seminars. Keep allocator employment, independent lab activity, university research, and verified investment use as separate evidence states.

September 5, 2026 — Southeast European academic finance-AI routes

• The Bucharest University of Economic Studies’ Ciprian Necula profile identifies him as a professor in Money and Banking, director of the DOFIN MSc programme and the Center for Advanced Research in Finance and Banking (CARFIB). The page lists quantitative-finance teaching in financial engineering, stochastic calculus, risk management, econometrics, and quantitative methods, alongside option pricing, stochastic volatility, macroeconomic modelling, and World Bank or government-funded project experience. This is a programme-and-research-centre route into mathematical finance and risk modelling; it does not establish a manager relationship, model deployment, proprietary data, or investment performance.
• ASE’s Dan Gabriel Anghel profile adds a finance faculty route explicitly spanning asset pricing, market efficiency, financial econometrics, data-snooping, technical analysis, and machine-learning applications. It also records PhD supervision in Finance, an Applied Mathematics degree, a 2023–24 Fulbright visit to the University of Pennsylvania, and current interests in climate and financial tail risk, sentiment-driven market dynamics, and high-frequency financial networks. This identifies paper, student, and seminar recovery paths; the profile does not establish a live trading system or hedge-fund use.
• ASE’s Florian Neagu profile records research interests in financial stability, systemic risk, banking, green finance, and AI/ML in finance, and says he founded the Innovative Finance Laboratory in 2025. His concurrent role as Director of the Financial Stability Department at the National Bank of Romania creates a central-bank/academic route for model-governance, systemic-risk, and supervisory research. It is not evidence of an investment manager’s internal AI platform.
• Belgrade’s Master in Computational Finance and teaching roster provide a concentrated programme route combining finance, data science, AI, IT, quantitative methods, and ML. The roster names Branko Urošević, who lists PhDs in Physics from Brown and Economics from UC Berkeley and says he created the MCF; Drago Inđić, whose profile describes hedge-fund experience, 500-plus automated trading systems, and interests in algorithmic trading, AI/ML, NLP, speech recognition, and visual systems; Mladen Stanojević, whose research interests include asset pricing, allocation, financial engineering, and ML; and Nikola Vasiljević, who teaches Advanced Quantitative Finance at the University of Zurich and previously worked in market-risk and quantitative-analyst roles. The programme page also reports a strategic partnership with ARPM and encourages trading algorithms, bots, robo-advisors, and blockchain applications. These are programme and instructor disclosures, not proof of production systems, fund sponsorship, or returns.
• The Belgrade Banking Academy profile for Vladimir Vasić adds a statistics/econometrics-to-industry route: it lists teaching in advanced business analytics, econometrics, and quantitative finance, a recent focus on ML, and participation in commercial regional projects using statistics, econometrics, and ML. The page does not identify the commercial clients or establish investment deployment, so those claims remain an unresolved lead.
• Athens University of Economics and Business’ Apostolos Katsafados profile records an Assistant Professor in Accounting and Finance who also serves as a Financial Advisor in the Bank of Greece Risk Management Department. His current research uses AI, ML, and NLP for credit-risk prediction, corporate-text analytics, central-bank communication, social-media and online-review data, corporate behaviour, and financial markets. This is a concrete credit/text/supervisory research route with a named public-sector interface; it does not establish a hedge-fund model or live trading use.
• AUEB’s George Chalamandaris profile records an Imperial College quantitative-finance PhD, prior risk-analyst/trader/structurer work at NatWest Markets and Eurobank, and research on derivatives, fixed income, and ML applications for assessing investment strategies. The page’s combination of market-practice history and academic ML/quantitative-finance work is useful for recovering papers and seminars, but it does not disclose current employer systems, proprietary data, or investment results. The AUEB study guide separately exposes a Machine Learning in Finance teaching route.
• The University of Piraeus profile for Georgios Papayiannis identifies research in robust decision-making for finance, learning from multiple information sources, model uncertainty, risk quantification, and statistical learning for complex data. The profile for Charalampos Agiropoulos adds computational finance, econometrics, and ML for financial time series, alongside prior work as a risk-and-investment officer and cofounder of a licensed Greek microfinance institution. These profiles expose robust-learning, aggregation, risk, and financial-time-series routes; they do not establish a hedge-fund relationship or live investment deployment.
• Sofia University’s CV for Deyan Radev lists principal-investigator and research-supervisor roles on AI-based credit-risk assessment, ML-based credit scoring for traded and non-traded companies, alternative-data-based sustainable credit scoring, and financial health/P2P lending. The CV gives public project budgets and dates. The projects are especially useful for tracing alternative data, fairness, credit labels, students, and co-investigators; they are not evidence of hedge-fund deployment or performance.
• The AUBG profile/interview for Andrey Gurov and its quantitative-finance event page expose a Bulgarian teaching and practitioner network around risk management, data analytics, ML in finance, quantitative finance, automated trading, and advanced trading strategies. The event page also names Genko Vasilev, described there as Head of Data Science at KBC Group Bulgaria and an Assistant Professor at Sofia University, and Veselin Filev, described as Head of AI at Ablera and an Associate Professor at the Bulgarian Academy of Sciences. These are public personnel and event leads requiring current-employer corroboration; the pages do not establish any hedge-fund system or performance.

This Southeast European pass adds four distinct discovery routes: finance faculties with named ML and risk researchers; formal computational-finance programmes; academic-to-bank, regulator, and practitioner interfaces; and funded projects where credit labels, alternative data, model governance, and fairness may be recoverable. The useful next step is to recover papers, code, project datasets, student theses, and conference recordings, then separately check whether any alumni or collaborators moved into tracked managers, vendors, banks, family offices, or allocator research teams. Nothing in this pass supports ranking firms or inferring deployment from academic affiliation.

Recovery queue for Southeast European academic routes

Recover ASE’s DOFIN, CARFIB, and Innovative Finance Laboratory rosters; retrieve Anghel’s ML, sentiment, high-frequency, and tail-risk papers plus PhD students; map RAF MCF’s ARPM partnership, instructors, student projects, alumni, and event recordings; verify Vasić’s commercial project interfaces; retrieve Katsafados, Chalamandaris, Papayiannis, and Agiropoulos papers and code; recover Radev’s funded-project documents, datasets, fairness definitions, students, and co-investigators; and verify the current KBC/Ablera roles named by the AUBG event page. Keep academic profiles, project statements, current employment, and verified investment use as separate evidence states.

September 5, 2026 — Student capital, applied programmes, and non-traditional finance data

• The University of Luxembourg’s FinSat project page is a concrete academic route into satellite and alternative data. It names Symeon Chatzinotas and Michael Halling as principal investigators, lists Eva Lagunas and doctoral researcher Khushboo Gehi on the project team, and describes a 24-month FNR-funded project using satellite imagery, maritime and aeronautical data, and socioeconomic indicators to build deep-learning models for market-behaviour prediction, portfolio optimisation, and automated investment advice. Halling’s faculty profile adds a computer-science and finance PhD lineage through Vienna, prior appointments at Utah and Stockholm, and a sustainable-finance chair. The project is an explicit research objective and data-modality disclosure; it does not establish a hedge-fund sponsor, investable performance, proprietary data rights, or production deployment.
• The University of Auckland’s FinTech applied-project route says ten-week student teams work with organisations on confidential datasets and produce joint analysis plus individual reports and presentations. Its examples include financial projections, bankruptcy and credit-card-default prediction, fraud detection, machine valuation, and portfolio optimisation; the programme teaches Python, R, SQL, ML, predictive analytics, and optimisation. A related finance analytics project page explicitly invites investment firms implementing quantitative strategies to propose projects, while describing outputs as exploratory rather than professional consulting advice. This exposes a structured academic-to-industry intake channel and potential data/workflow leads, not evidence that a particular fund participated.
• The University of Maryland’s Vertex Fund is a newly visible 2026 student-led fund page. It says members manage a real portfolio of equities and digital assets, meet weekly for research and pitches, build financial models, and publish equity research; the page shows zero active positions and a placeholder portfolio-value field at the time checked. This is a talent and student-capital discovery route with an important completeness caveat: the public page does not expose faculty advisers, holdings, model methods, AI use, data rights, or verified performance.
• Brian Silverstein’s 2025 SSRN paper, “Reinventing the SMIF”, hosted through the University of South Carolina’s Darla Moore School of Business, describes an AI-assisted student-managed-fund portal. The proposed workflow runs from idea generation through valuation and risk review to a documented decision; AI notes and summaries support analysis, while live portfolio views, market and stock notes, buy/sell limits, DCF, factor analysis, P&L attribution, and mean-variance optimisation remain visible. The paper explicitly says the portal does not replace professional terminals or execute trades. This is a useful control case for where AI can improve research traceability and teaching without granting execution authority; it is a paper and design proposal, not evidence of a live fund system.

This pass adds four research routes that are easy to miss in firm-centric searching: satellite, maritime, and aviation data; confidential university-to-organisation project intake; newly formed student capital pools; and AI-assisted investment-research pedagogy with explicit non-execution boundaries. The most actionable follow-up is to recover FinSat papers and data definitions, Auckland project hosts and anonymised outputs, Vertex’s team and faculty links, and SMIF portal implementations. None of these sources warrants ranking a firm, programme, or model, or inferring that an academic design is deployed by an investment manager.

Recovery queue for student-capital and applied-programme routes

Recover FinSat’s project publications, doctoral researchers, data suppliers, satellite/maritime feature definitions, code, and finance collaborators; retrieve Auckland’s FinTech project examples, host organisations, faculty supervisors, anonymised reports, and alumni destinations; inspect Vertex’s team, faculty sponsor, filings, portfolio disclosures, and campus channels; and trace Silverstein’s portal design into course materials, software, student-fund implementations, and subsequent papers. Keep student-managed capital, academic project proposals, sponsor access, and verified manager deployment as separate evidence states.

September 5, 2026 — Academic data labs and student portfolios with inspectable workflows

• The University of Colorado Boulder’s NLP::FIN::LAB describes a research centre developing tools and datasets at the intersection of NLP, ML, OCR, big data, and financial economics. Its current-projects page names EDGAR, newspapers, earnings announcements, analyst reports, academic CVs, patents, accounting information, and scraping algorithms as textual or visual corpora. Projects include extracting historical bond yields from New York Times images, mining Moody’s manuals for firms and manager/director networks, and studying academic output with NLP. The Asaf Bernstein faculty profile adds a finance faculty and information-economics route. These are concrete corpus and OCR disclosures; they do not establish a hedge-fund data licence, investment use, or model deployment.
• The University of Queensland’s Student Managed Investment Fund page reports a live equity portfolio, monthly fund updates, and more than AUD 550,000 as of August 2026, while naming convenors Eric Tan, Saphira Rekker, and Alexander Cameron and publishing current portfolio managers and analysts. The 2026 research-challenge brief gives the operating loop: teams select an Australian-listed company, write an equity-research report, build a valuation model, present to academic and industry judges, and—if selected—serve as equity research analysts with recommendations subject to Investment Committee approval. The February 2026 update publishes holdings, weights, contributions, and a cash position. This is a detailed student-capital and talent-flow surface; it does not show AI use, proprietary model ownership, or hedge-fund deployment.

This pass adds two different kinds of academic evidence: a lab whose public remit exposes document, image, and archival corpora, and a student fund that exposes a dated research-to-committee workflow with named personnel and recurring portfolio updates. The follow-up is to recover the Colorado lab’s datasets, papers, code, and personnel lineage, and to parse UQ’s monthly reports, challenge submissions, judges, alumni, and sponsor links. These sources remain evidence about academic research and student capital, not a basis for ranking firms or inferring live manager systems.

Recovery queue for academic data labs and student portfolios

Recover NLP::FIN::LAB research outputs, datasets, OCR pipelines, archival corpora, student researchers, and Bernstein’s finance/NLP papers; download UQ SMIF monthly updates, holdings histories, challenge briefs, judging panels, Investment Committee materials, current students, alumni, and faculty affiliations; and cross-check any employer or sponsor links independently. Keep public student-fund records, academic datasets, employer relationships, and verified investment deployment separate.

September 5, 2026 — Family-office allocator network and university-trained personnel route

• TABOR Capital Conferences provides a public allocator-network route that is largely absent from firm-name searches. The site lists an invitation-only Family Office Conference in Dallas for October 11–13, 2026, says the upcoming event is intended for approximately 50 family offices and 30 managers, and describes a two-way diligence process with attendance capped. It also reports, as company-provided network statistics, more than 500 managers reviewed each year and 27 conferences held to date. TABOR’s founder profile says Troy Johns graduated from the University of Colorado Boulder, worked across audit, technology consulting, risk-assessment software, hedge funds, and capital raising, and uses information gathered from actively investing family offices to curate manager invitations. These claims expose a useful LP/GP discovery channel but are not independently audited and do not reveal invitee identities or investment decisions.
• TABOR’s public analyst profile says Gavin Johns studied real-estate finance at Southern Methodist University, interned with a single-family office and alternative-investment managers, joined TABOR in early 2026, and brings stated AI expertise to the conference business. This is a university-to-allocator personnel lead and a route to SMU alumni, family-office internships, conference attendees, and manager-selection vocabulary. The page does not specify the AI methods, data, systems, clients, or investment authority involved.

This route broadens the family-office search beyond academic councils and fund websites: conference organisers can sit between family offices, managers, and university-trained analysts, and can expose selection criteria without publishing a portfolio. Treat all TABOR scale and conversion figures as first-party claims; use the public conference date, biographies, and role descriptions as discovery leads only.

Recovery queue for allocator-network routes

Recover TABOR’s 2026 agenda, speaker and manager list where publicly released, event recordings, founder and analyst profiles, SMU alumni links, family-office attendee references, and any public material on AI-enabled manager due diligence. Keep event-organiser claims, family-office identity, manager marketing, and verified investment decisions separate.

September 5, 2026 — Princeton financial-engineering faculty and practitioner-interface route

• Princeton’s ORF 311 course page identifies John M. Mulvey as the Spring 2026 instructor for “Stochastic Optimization and Machine Learning in Finance.” The description links decision trees, Monte Carlo simulation, stochastic programmes, forecasting and planning systems, and machine-learning methods in financial applications. This is a public curriculum signal for connecting prediction to constrained decision-making; it does not identify enrolled students, a sponsor, a proprietary dataset, or a live portfolio system.
• Mulvey’s Princeton publication list exposes a long-running research programme rather than a single LLM claim. The list includes work on large language models for financial and investment management, regime-switching signals and factor allocation, neural-network risk-budgeting and portfolio optimisation, regime-switching execution with nonlinear impact costs, multi-period planning with Monte Carlo tree search and neural networks, NLP for investment decisions, and competitive multi-agent reinforcement learning. Princeton’s research-output record separately lists 2026 outputs on allocation-focused regimes, deep generative models for financial-regime identification, regime-aware asset allocation, and a modular ML trading rule with transaction costs. These publications expose model families, objective functions, and implementation constraints to investigate; titles and abstracts do not establish profitable use or professional deployment.
• Mulvey’s consulting page says he has designed or implemented asset-liability management systems for PIMCO, Towers Perrin/Tillinghast, AXA, American Express, Siemens, Munich Re-Insurance, Ant Group/Alibaba, First Republic Bank, and other organisations, and says current projects include regime identification and factor approaches for long-term investors such as family offices and pension plans. The university’s faculty research guide additionally says his work has involved large insurance companies, hedge funds, global FinTech firms, banks, Ant Financial, and an unnamed multi-manager hedge fund in Austin. These are first-party biography and consulting disclosures; “Renaissance Re-Insurance” on the consulting page should not be conflated with Renaissance Technologies, and no named hedge-fund system, client mandate, data right, or performance result is disclosed.
• Mulvey’s public profile also links an older statistical-learning paper on equity return forecasting whose abstract describes market-neutral weekly stock allocation, transaction-cost sensitivity, and statistical-arbitrage terminology. That is a published academic result from 2005, not evidence of a current manager’s strategy. The historical-to-current sequence is useful for tracing how academic work moved from market-neutral prediction to regime-aware allocation, neural optimisation, transaction-cost modelling, and LLM-oriented investment-management research.

This route adds a faculty member whose public materials span programme teaching, papers, consulting, family-office planning, and named financial-organisation interfaces. It is useful for building a university-to-industry lineage graph and a model-family timeline, but it does not support ranking firms, attributing any paper to a tracked manager, or inferring that a client adopted the published methods.

Recovery queue for Princeton faculty and practitioner-interface routes

Recover Mulvey’s 2024–2026 papers, working-paper versions, code, data definitions, student theses, course materials, lecture recordings, coauthor and supervisor lineage, and public descriptions of client work. Verify every client name and date independently; distinguish PIMCO, insurance, banking, reinsurance, family-office, pension, and unnamed hedge-fund references; and preserve academic method claims, consulting claims, and verified investment deployment as separate evidence states.

September 5, 2026 — Essex computational-finance centre, doctoral lineage, and market-agent route

• The University of Essex’s Centre for Computational Finance and Economic Agents (CCFEA) is a dedicated academic route combining computer science, machine learning, economics, and finance. Its public description says teaching and research use high-frequency data and cover algorithmic trading, derivatives pricing, algorithmic game theory, and multi-agent systems motivated by real markets; it also describes industry and UK-government links and participation in the IEEE Computational Finance and Economics Technical Committee. This is a centre-level research and talent surface, not evidence of a particular fund’s sponsorship or deployment.
• The centre’s postgraduate study page names training in machine learning, market-microstructure design and testing, risk management, financial engineering, adaptive and reinforcement learning, heuristic optimisation, and evolutionary computation. The 2026/27 Machine Learning for Finance module specifies structured and unstructured financial data, data-mining algorithms, optimisation for asset allocation, risk management, option pricing and calibration, lectures plus labs, and a 40% code-and-report component. It names Dr Ana Matran-Fernandez as supervisor/teaching staff and records UCL and University of York external examiners. The page does not expose student code, datasets, employer sponsors, or professional deployment.
• Michael Kampouridis’s faculty profile provides a professor-and-supervisor graph. His listed interests include genetic programming, financial forecasting, fundamental/technical/sentiment analysis, directional changes, weather derivatives, and evolutionary algorithms. The page names eight current doctoral supervisees across computational finance, computer science, and AI, and prior theses on deep reinforcement learning for high-frequency currency trading, ML/statistical earnings and free-cash-flow forecasting, genetic-programming time representations, mixed-asset real-estate allocation, and sentiment/technical/fundamental algorithmic trading. His listed publications include multi-agent systems for computational economics and finance, algorithmic trading with directional changes, and work on genetic programming in finance. This exposes a concrete student-to-model lineage; it does not establish where graduates went or whether any method entered a fund.
• Two public Essex repository records make the model questions more inspectable. Ivan Evdokimov’s doctoral work on fundamental stock valuation frames next-quarter EPS and free cash flow as sparse-data regression targets, compares eight ML and five statistical estimators over 100 U.S. companies, and reports interpolation, quantile transformation, transfer-learning, overfitting, and fixed-period backtesting issues. Xinpeng Long’s doctoral work on financial forecasting combines physical and event-based time with genetic programming and multi-objective optimisation. These are thesis abstracts and repository claims; recover the full methods, splits, feature vintages, costs, code, and out-of-sample definitions before treating any reported portfolio result as portable.
• The CCFEA people page also exposes current student names and research labels, including a finance/Python/stock-trader/fintech profile and a published financial-image-recognition workshop paper. Preserve those public labels as discovery leads only, and independently verify identities, dates, publications, employers, and any later hedge-fund or vendor connection.

This route is valuable because the academic evidence is operationally specific: a named course, code-and-report assessment, high-frequency data, market microstructure, multi-agent systems, directional-change sampling, genetic programming, sentiment, sparse fundamental targets, and named doctoral supervision. It remains a university research and talent network; no firm ranking, manager attribution, or professional deployment inference is warranted.

Recovery queue for Essex computational-finance routes

Recover CCFEA’s current and historical rosters, doctoral theses, supervisors, code, datasets, feature and label definitions, high-frequency-data terms, paper appendices, workshops, recordings, industry-partner pages, IEEE committee events, alumni destinations, and public student projects. Preserve original repository files and dates; independently verify employer links and separate course exercises, thesis results, centre-level industry claims, and live investment deployment.

September 5, 2026 — Leicester FinTech/RegTech programme and finance-faculty lineage

• The University of Leicester’s 2026 MSc FinTech and RegTech is a new finance-programme route combining financial markets with data analytics, AI, blockchain, DeFi, cryptoassets, trading, payments, and market structure. The programme says students use Python or R for hands-on analytical work, interpret model outputs, and examine transparency, auditability, fairness, responsible data use, and compliance-by-design for automated decision systems. Its final module offers a dissertation, professional-practice placement, or AI business project. This exposes a curriculum and applied-project route; it does not identify a fund sponsor, dataset, model owner, or live trading system.
• The programme specification adds the formal structure: a 30-credit “Data Analytics and Machine Learning for Finance” module, a 30-credit “RegTech, Ethics & Governance in AI” module, and a 60-credit choice among dissertation, placement, or AI-agent business project. These are programme-design signals, not evidence that a student project reached production or that a named employer supplied data.
• Hadi Movaghari’s faculty profile links the programme to a quantitative finance lineage. It records a PhD and postdoctoral route through the University of Glasgow’s Adam Smith Business School, statistics degrees from Tarbiat Modares University and Razi University, and a visiting position at Radboud University. His public publications include double machine learning for corporate cash policy, time-varying coefficients, transfer entropy, robust variable selection, and LASSO; his teaching includes FinTech, AI and Blockchain, Investment Management, Financial Econometrics, and Data Science and Machine Learning in Finance. This provides paper, supervisor, and course recovery targets without establishing a hedge-fund connection.
• Professor Daniel Ladley’s profile adds a computational-finance and market-microstructure route. It records undergraduate and PhD training at Leeds, finance-firm collaborations, publications on high-frequency speed versus sophistication, noise trading, systemic risk, contagion networks, and agent-based models, and supervision in market microstructure, behavioural finance, regulation, or banking using computational or numerical methods. The profile does not name collaborators, data terms, employer systems, or investment deployment.
• Professor Ania Zalewska’s profile adds a cross-border faculty and conference network. It records London Business School doctoral training, prior appointments at Bristol, Maastricht, Bath, and Leicester, research on financial-market development, intermediaries, fintech, regulation, and investor behaviour, and a current role as UK representative and academia-government coordinator for the EU-COST Action on Fintech and Artificial Intelligence in Finance. Her profile also links the Finance Research Consortium, International Virtual Research Seminars in Finance, and policy/industry events. This is a route to speakers, papers, and programme participants; it does not identify a professional manager’s AI stack.
• Bogdan Grechuk’s profile extends the mathematical lineage through MIPT, Stevens Institute of Technology, and Edinburgh. It combines financial mathematics, ML, capital allocation, benchmark-based deviation and drawdown measures, portfolio optimisation, and risk analysis, and records prior work on high-gain investment strategies and error-correction mechanisms in AI systems. These are public academic and technology-transfer claims; they require paper, patent, and implementation verification before any investment relevance is inferred.

This route links a current finance/AI programme to faculty with different research emphases: data analytics and governance, double ML and variable selection, high-frequency market structure, agent-based systemic risk, cross-border fintech regulation, and portfolio-risk mathematics. It is a personnel, syllabus, and paper-discovery surface; it does not support ranking institutions or attributing any academic work to a tracked fund.

Recovery queue for Leicester finance-AI routes

Recover the programme’s AF7001–AF7004 materials, instructors, assignments, AI-agent project examples, placement hosts, dissertation topics, data terms, and student/alumni destinations. Retrieve Movaghari’s ML-finance papers and supervisors; Ladley’s current funded projects, code, students and finance-firm collaborators; Zalewska’s EU-COST, consortium and seminar rosters; and Grechuk’s portfolio-risk papers, patents and implementation records. Keep programme claims, academic publications, policy interfaces, employer links, and verified investment deployment separate.

September 5, 2026 — Canadian student-capital and academic-industry routes

• The Ted Rogers School Student Managed Investment Fund guidelines describe a real-money fund founded in 2020 with $500,000 of seed capital and a reported value above $900,000 in January 2026. The redesigned FIN650/FIN750 structure separates analysts from portfolio managers: analysts produce sector reports, update holdings, and conduct research; portfolio managers mentor analysts, make recommendations, prepare term reports, and assist faculty advisers with monitoring, compliance, and reporting. This is a concrete research-to-oversight workflow; it does not disclose AI use, proprietary data, or professional-manager adoption.
• The University of Toronto’s Rotman Commerce Student Fidelity Fund announcement records a $250,000 Fidelity Investments Canada contribution, a 12-member board of faculty, alumni, and industry professionals, and a first trade planned for March 2026. The student team covers seven sectors and operates under a dual mandate involving risk-adjusted returns and liquidity for an annual charitable distribution. Named faculty board members include Catherine Barrette, Claire Célérier, David Goldreich, and Anwar Husain. This is a named asset-manager sponsorship and talent route, not evidence of Fidelity’s internal models or AI systems.
• The University of New Brunswick Student Investment Fund reports five Bloomberg terminals, a two-year analyst-to-fund-manager progression, professional portfolio-manager interaction, and growth from $1 million at launch in 1998 to more than $15 million. Current participant biographies expose unusually specific role and employer pathways: an Investment Data Analyst building portfolio dashboards and visualisations, a fund trader managing trades and cash positions, portfolio managers covering TSX sectors, and links to Vestcor, RBC Dominion Securities, PwC, and Irving Oil. The page is a public talent and workflow surface; it does not establish AI use, production-grade model controls, or professional-fund deployment.
• Simon Fraser’s Quantitative Finance Lab provides the research-lab layer behind the Canadian pipeline. Its public remit covers asset pricing and risk premia, digital assets and derivatives, market microstructure and trading, and ML/data-driven finance. It says most student projects involve Python and large financial datasets and that the lab meets weekly. This supports searches for faculty, student code, papers, and data definitions; it does not disclose a named manager, data licence, or live strategy.

This Canadian cluster exposes four distinct interface types: an analyst/PM/compliance teaching structure, a named Fidelity-sponsored student portfolio, a large student fund with public data and trading roles, and a research lab organized around market microstructure and digital assets. These should be mined for papers, supervisors, public code, event recordings, alumni destinations, and sponsor materials while keeping student capital, academic research, employer history, and verified professional deployment separate.

Recovery queue for Canadian student-capital routes

Recover TMU’s faculty advisers, sector reports, FIN650/FIN750 assignments and holdings history; Rotman’s post-launch reports, board biographies, student research, and Fidelity event material; UNB’s faculty oversight, annual reports, historical holdings, MQIM course material, student outputs, and public alumni paths; and Simon Fraser’s lab roster, publications, projects, code, seminars, and industry interfaces. Do not infer AI adoption from dashboards, Bloomberg access, sponsor support, or quantitative-finance course language.

September 5, 2026 — Italian finance programmes, AI coursework, and faculty research routes

• The Catholic University of the Sacred Heart’s Artificial Intelligence for Banking and Finance MSc is a two-year, English-language programme scheduled for a September 2026 intake. Its public curriculum combines programming, statistical methods, machine learning on large-scale data, financial modelling, asset management, risk, pricing, regulation, simulations, and collaborative industry projects. This is a direct syllabus and project-interface signal; the page does not identify a tracked fund, a proprietary dataset, or production deployment.
• Cattolica’s Financial Engineering for Portfolio and Risk Management master describes a one-year, English-language programme starting November 2, 2026, with 30 places. Its stated methods span classical quantitative models, AI and quantum-computing approaches for pricing, statistical arbitrage, trading, portfolio, risk and wealth management. The programme lists AI-powered quantitative investment, systematic trading, probabilistic-ML engineering and computational finance among its stated career pathways. Those are programme claims about intended training outcomes, not evidence that an employer uses a particular method.
• The University of Pavia’s English-taught Finance master combines quantitative finance, portfolio management, risk, derivatives, econometrics, probability, stochastic processes, statistics, Python/Matlab and machine learning. The programme draws faculty from economics and management, physics and mathematics, and its 2026–27 study plan includes financial econometrics, statistics for finance, portfolio management, advanced quantitative finance and computational methods. This exposes a cross-disciplinary recruiting and research route; it does not establish a manager relationship or model deployment.
• The University of Milano-Bicocca ECOFIN programme names Machine Learning for Finance as an 8-credit course inside a broader quantitative structure covering financial mathematics, statistics, econometrics, derivatives, portfolio theory, risk, asset management and FinTech. The programme also describes a 12-terminal Bloomberg laboratory, trading competitions, professional reports, internships/theses and a second-year Birkbeck University of London route. Its March 2026 presentation identifies Fabio Bellini as programme coordinator and names alumni destinations including BNP Paribas Cardif, KPMG and Arca Fondi SGR. These are public education and alumni-interface signals, not evidence of employer model use.
• Milano-Bicocca’s Andrea Maurino profile lists a 2025 ICAIF paper, “Optimizing Large Language Models for ESG Activity Detection in Financial Texts”. The paper introduces a 1,325-segment ESG-Activities benchmark and reports experiments in which fine-tuning on original and synthetic data improved domain classification, with open 7B models outperforming proprietary systems in particular configurations. The result is an inspectable financial-text classification experiment; it is not a trading signal, a fund disclosure, or proof of production use.
• The Fabio Stella faculty page provides a separate model lineage: Bayesian networks and continuous-time Bayesian networks applied to finance, feed-forward neural networks applied to computational finance, and online portfolio-selection algorithms with and without transaction costs. Stella’s profile records a PhD in Computational Mathematics and Operations Research from the University of Milan, a former risk-management consulting role at Banca Intesa, and teaching in causal networks, machine learning and unsupervised learning. The Fabio Bellini profile and Bicocca archive expose a mathematical-risk route through expectiles, elicitable risk measures, risk parity and backtesting. These pages support paper and faculty-lineage recovery; they do not link either researcher to a tracked hedge fund.
• Marco Guerzoni’s profile adds an alternative-data and energy-finance route. His 2026 paper uses machine-learning-based variable selection inside a GARCH-MIDAS approach to study Italian electricity-price volatility; his biography describes advanced analytics and big data applied to economics, business intelligence and policy. This is a concrete research design for testing nontraditional covariates in a financial-volatility setting, not evidence of a manager’s live factor.

The Italian pass adds programme-level talent surfaces and paper-level model clues: large-scale financial-text classification, synthetic-data fine-tuning, Bayesian and neural finance, online portfolio optimisation, risk measures, and weather-linked electricity volatility. The useful next step is to recover syllabi, assignments, code, datasets, theses, supervisors, event recordings, alumni paths and independently verifiable employer interfaces. Academic method, programme marketing, alumni destination and professional deployment remain separate evidence states.

Recovery queue for Italian finance-AI routes

Recover Cattolica’s MSc and master syllabi, instructors, industry-project examples, placements and data terms; Pavia’s faculty, thesis topics, supervisors, code and quantitative-finance projects; and Milano-Bicocca’s ECOFIN teaching materials, Bloomberg-lab outputs, Birkbeck route, faculty papers, students and alumni. Retrieve the Maurino ESG benchmark and fine-tuning details, Stella’s finance papers and code, Bellini’s risk-measure lineage, and Guerzoni’s data and feature definitions. Do not infer hedge-fund adoption from a syllabus, alumni employer, Bloomberg access or a published academic model.

September 5, 2026 — European finance courses, model classes, and academic–industry project interfaces

• The Vrije Universiteit Amsterdam Machine Learning for Finance course is listed for the 2026–27 academic year and is restricted to the Finance and Technology and Quantitative Finance honours programmes. Its objectives require students to code predictive models in Python and apply them to real datasets; the course covers regularised linear models, logistic classification, trees, neural networks, support-vector machines and ensembles, with assignments and a written exam. The Anne Opschoor profile links the course to an associate professor and Tinbergen Institute fellow whose research includes financial econometrics, volatility, risk management, copulas and high-dimensional extreme risk. This is a concrete training and faculty route; it does not identify a fund sponsor, dataset licence, or live strategy.
• Leipzig University’s Artificial Intelligence & Machine Learning in Finance module is scheduled for Summer Semester 2026 and names Gregor Weiß as module owner. The German-language module description requires students to structure and analyse large, unstructured financial datasets with Python, TensorFlow and R and apply ML/deep learning to risk management, asset allocation and derivatives valuation. Weiß’s official profile records a finance and sustainable-banking chair, a PhD in economics from Ruhr University Bochum, mathematics and management training, and teaching in computational finance and ML/AI in finance since 2020. The source exposes a dated curriculum and faculty lineage, not evidence of investment-manager adoption.
• The University of Basel Computational Finance & AI course ran in Spring 2026 under Dietmar Maringer. The official catalogue describes hands-on Python implementation of pricing trees, Monte Carlo, artificial markets, agent-based models, portfolio rebalancing, backtesting, algorithmic trading, ML-based asset pricing, learning and self-adaptation. Assessment combines participation, assignments and a written exam. Maringer’s faculty profile records computational finance, algorithmic trading, high-frequency markets, portfolio optimisation, financial networks and ML, plus an earlier director-of-research role at Essex’s CCFEA. This is a model-and-course route with an unusually explicit implementation surface; it does not disclose student code, proprietary data or professional deployment.
• ETH Zurich’s Fall 2026 Machine Learning in Finance & Insurance course lists regression and regularisation, dimension reduction, kernels, tree ensembles, neural networks, autoencoders, graph neural networks and transformers, with coding projects in pricing, credit analytics, deep hedging and insurance claims. The ETH RiskLab course catalogue provides the academic context under Patrick Cheridito. These are teaching and implementation clues, not evidence that students or lecturers operate a fund strategy.
• ETH RiskLab’s 2026 public project register adds a deeper academic–industry interface. Bálint Binkert’s PhD combines numerical guarantees for conditional expectations, neural-network low-latency stochastic filtering and latent-factor dynamics for robust investor growth. Mehdi Hachimi’s MSc, partnered by Qube Research and Technologies, studies transformations that conceal sensitive features while preserving model classes and predictive performance, including RFAnon and quantile discretisation tested with random forests and neural networks. Other listed projects connect ADIA to diffusion models for S&P 500 implied-volatility surfaces trained on 2010–2020 data with 2021–2023 validation, AXPO to regime-aware energy-spread execution using HMMs, XGBoost and a temporal-convolution attention model, and Zürcher Kantonalbank to deep time-series stock-return forecasting. These are public thesis abstracts and partner listings; they expose research questions, model families, data windows and confidentiality concerns, but not client mandates, code, production ownership, permissions or independently validated performance.

This European pass adds several concrete model and workflow surfaces: Python implementation, real financial datasets, unstructured-data preparation, derivatives valuation, agent-based markets, backtesting, deep hedging, privacy-preserving transformations, implied-volatility diffusion, regime-conditioned execution and time-series asset pricing. The academic–industry links are useful discovery routes for papers, supervisors, data terms, code, personnel and event recordings; they are not evidence that a named tracked fund adopted a method, nor a basis for ranking institutions.

Recovery queue for European academic implementation routes

Recover VU Amsterdam assignments, faculty publications, data definitions and thesis projects; Leipzig module materials, instructors, student work and doctoral postings; Basel notebooks, assignments, papers and seminar recordings; and ETH RiskLab thesis files, appendices, code, data dictionaries, partner agreements, supervisor lineages and presentations. Independently verify all industry-partner relationships and distinguish public thesis abstracts, partner association, confidential data access, production deployment and performance.

September 5, 2026 — Additional U.S. student-capital and finance-lab routes

• The University of San Francisco’s Silk Family Investment Institute Student Managed Fund reports a $3.0 million portfolio and a workflow combining fundamental and technical analysis, financial modelling, valuation, risk analysis, macro research and investment theses. The page identifies Bloomberg and FactSet as lab data sources and publishes a 2024 fund-versus-index figure, while directing readers to request the underlying performance report. This is a first-party student-capital and data-access disclosure; the page does not identify a model, faculty investment authority, dataset licence, or AI use, and the performance figure is not independently audited here.
• The University of Texas at Arlington’s Student Managed Investment Fund describes up to 20 interviewed undergraduate or graduate participants managing real money donated in 2020. Students are organized as sector security analysts and senior analysts, can join a fund/portfolio-manager team, and collectively vote on purchases, sales, liquidations and rebalancing; the chief adviser/professor executes the decisions. The programme exposes a 10-terminal Bloomberg lab, DCF and relative-valuation models, technical analysis, live portfolio updates, and a separate quantitative-finance MS route. It does not disclose AI use, proprietary data, model ownership, or professional-manager adoption.

These two routes add different public surfaces: San Francisco exposes an investment lab with named commercial data vendors and a requestable performance report; Arlington exposes a vote-to-execution boundary, role progression, a Bloomberg lab, and a professor-controlled order-execution step. Both are educational investment programmes. Neither is evidence of a tracked fund’s internal AI strategy or a basis for ranking student funds.

Recovery queue for additional U.S. student-capital routes

Recover USF’s faculty, fund charter, current team, 2024 performance report, portfolio history, data terms, Investment Lab materials and alumni destinations. Recover Arlington’s faculty advisers, live-update data, course assignments, committee records, student projects, quantitative-finance curriculum, public portfolio history and alumni paths. Keep vendor access, student decisions, faculty execution, reported performance, academic work and professional deployment separate.

September 5, 2026 — Wofford and Houston student-capital operating surfaces

• Wofford’s James–Atkins Student-Managed Investment Fund describes a real-time student fund whose research and decisions determine the portfolio. Its public operating model separates Managing Partners, Portfolio Managers, and Research Associates, with four research groups: domestic equity/international, domestic equity/alternative, domestic equity/real estate, and domestic equity/fixed income. The page also publishes Spring 2026 meeting dates, links to a current portfolio and annual report, and describes an application gate for decision-making membership. This is a governance and artifact-recovery route; it does not disclose an AI system, data licence, model permissions, or professional-fund deployment.
• Wofford’s member-profile page exposes a dated personnel surface rather than an anonymous student-fund description. The listed Spring 2026 cohort includes Robbie Joseph as Managing Partner; Leah Mitchell, Bridger Jones, Matt Johnson, and Mason Mitchener as Portfolio Managers; named assistant portfolio managers and research associates across the four groups; and Dr. Patrick M. Stanton as Associate Professor of Finance. Stanton’s biography lists finance teaching, a Mississippi College undergraduate degree, Louisiana Tech MBA/DBA training, and research interests in microfinance, behavioral finance, and international corporate finance. These names and titles are public discovery leads; they do not establish where students later worked or that their research used machine learning.
• The University of Houston Bauer College’s Cougar Investment Fund is a multi-million-dollar private equity fund managed by MBA and MS Finance students in the Graduate Certificate in Financial Services Management. The page names Thomas George as Bauer Professor of Finance, AIM Center director, and fund Chief Investment Officer, and Michael J. Murray as Instructional Professor and Chief Operating Officer. It describes fundamental company analysis, financial valuation, real-time and locally collected financial information, and Bauer proprietary analysis and valuation software. The page says the fund receives capital from individual investors rather than only a university endowment. None of these statements identifies an AI model, model-training dataset, or production system.
• Bauer’s graduate experiential-learning page gives more detail on the fund-to-workflow boundary: students identify U.S. equity securities, present recommendations, execute trades, prepare reports for private investors and prospective clients, and perform accounting and reporting for the private fund. A separate Bauer 2024 alumni account follows Jeffrey Detwiler from the Cougar Fund and MS Finance programme to a partner/fixed-income portfolio-manager role at Garcia Hamilton & Associates, and records a $50,000 fellowship directed toward finance students, especially those participating in the fund. This is a concrete education-to-manager lineage and philanthropic interface; it is not evidence that Garcia Hamilton or Bauer uses a particular AI technology.

These routes add two different kinds of evidence to the academic map. Wofford makes the current research-team hierarchy, sector grouping, faculty biography, and portfolio/report links inspectable. Houston makes the student-to-execution workflow, outside-investor structure, named operating officers, data/software environment, and an alumnus’s subsequent portfolio-management path inspectable. Both are training and capital-governance surfaces. Neither supports a ranking of programmes, an inference about hedge-fund deployment, or an AI claim beyond what the pages explicitly state.

Recovery queue for Wofford and Houston routes

Recover Wofford’s current and historical member resumes, annual reports, portfolio snapshots, faculty supervision, student research outputs, recordings, alumni destinations, and any public code. Recover Houston’s fund charter, current student roster, course syllabi, investor/reporting materials, AIM Center facilities and data descriptions, Thomas George and Michael Murray’s papers and students, the TIPS programme archive, and alumni pathways including Detwiler’s. Keep student research, faculty authority, outside-investor capital, vendor/software access, academic publication, employment history, and verified professional deployment as separate evidence states.

September 5, 2026 — Barry, Stirling, and Purdue: model-policy, allocator, and AI-talent routes

• Barry University’s Student Managed Investment Fund page says the fund is student-founded and student-run, backed by university-endowment capital, and governed by an advisory board of investment practitioners. It displays a $1 million portfolio, Bloomberg, FactSet, proprietary software, and an industry-resource library. Its published investment policy explicitly separates quantitative models and techniques, qualitative judgement, and external research, and places oversight of investment policy, strategy, and performance with the advisory board. Dr. Stephen Morrell is listed as the application contact. This is a rare explicit public model-policy disclosure in a student-fund setting; it does not identify the quantitative models, training data, code, or any professional manager’s system.
• The University of Stirling’s SMIF “Who we are” page publishes a 2025/26 committee with titles that resemble a small investment-organisation operating chart: Michael Kehoe as Chief Investment Officer, Meet Thaker as Head of Portfolio Management, Ignacio Sanchez as Chief Risk & Compliance Officer, Cole Connolly as Chief Research Officer, Alasdair Cross as Chief Economics Officer, and Agnes Wiborg as Chief ESG Officer, alongside two co-presidents. The same page names academic and practitioner advisers, including Kevin Campbell (MSc Investment Analysis course director), Patrick Herbst (finance lecturer with prior equity-portfolio-management experience at Allianz Dresdner Asset Management), and Isaac Tabner (MSc Finance course director). This is a dated public personnel and curriculum-lineage surface, not evidence of AI use.
• Stirling’s advisory page also identifies Jon (JB) Beckett as a 1996 alumnus who had recently been an investment gatekeeper for £180bn of Scottish Widows funds and who serves on Royal London’s Investment Advisory Committee. The page describes that as adviser biography, not as sponsorship or access to confidential allocator data. It is a useful route for tracing allocator governance, professional fund-investing careers, committee language, reports, and student-to-industry interfaces; it does not establish any hedge fund or family-office relationship.
• Purdue’s current SMIF team page says 33 students and two faculty advisers work across eight sector teams, Fixed Income & Macro, and Portfolio + Risk Management. Its executive-board structure assigns students responsibility for strategy, research, risk, recruiting, education, and operations. The public bios expose several AI/quantitative-talent routes: Hunter Specht combines computer science, data science, applied statistics, and finance and lists an MFS Investment Management software-engineering internship; Parth Dama lists IBM Data & AI work and an incoming Barclays Markets Quantitative Analyst role; and Abhipsa Prajapati lists current AI-and-finance research through the Wharton AI Research Fellowship and prior private-credit analysis. These are student biographies and future/employment claims published by the fund; they do not show that Purdue SMIF uses those employers’ systems or that its portfolio process uses AI.
• Purdue’s page also names Lulu Zeng and Alexander Boquist as faculty advisers and describes a separate Portfolio + Risk Management team responsible for allocation, position sizing, risk monitoring, trading and rebalancing, and performance attribution. That combination makes the page a useful recruiting and process-discovery surface: search the named students and faculty across papers, GitHub, conferences, finance clubs, internships, and later employer biographies. It remains a public educational structure, not evidence of a production trading model or of a particular employer’s internal strategy.

These three routes add different evidence states: Barry publishes a model-versus-judgement policy boundary and named data/software categories; Stirling publishes an unusually explicit investment-organisation role taxonomy and an allocator-adviser biography; Purdue publishes a current operating chart with identifiable AI, data, software, and markets-quant talent paths. None supports ranking programmes or attributing professional AI deployment to a student fund, university, employer, or adviser.

Recovery queue for Barry, Stirling, and Purdue

Recover Barry’s current advisory-board roster, faculty and fund-manager names, investment-policy history, quantitative-model descriptions, holdings/performance records, course materials, service-learning partners and alumni destinations. Recover Stirling’s reports, events, committee history, adviser biographies, MSc syllabi, student research, public code and allocator/fund-investing pathways. Recover Purdue’s research posts, holdings and performance methodology, team history, faculty papers, student LinkedIn/paper/GitHub trails, AI-and-finance fellowship outputs, and internship-to-employer transitions. Keep published bios, educational assignments, employer affiliation, adviser status, confidential access, model ownership and verified deployment separate.

September 5, 2026 — Family-office curriculum and Canadian allocator/data interfaces

• The Hong Kong Polytechnic University’s MSc in Asset and Wealth Management lists a September 2026 entry, one-year full-time or two-year part-time study, and a Family Office Wealth Management specialization alongside Digital Asset Management. The programme description explicitly includes AI, blockchain, cloud computing, data science and entrepreneurship as technology elements; its structure includes a six-credit Asset and Wealth Management Project using real-world cases through industry collaboration and an elective, Decision Analytics by Machine Learning. The Family Office track also includes Wealth Planning and Family Office, while the broader programme describes data-driven case studies, advanced analytics, visualisation, and immersive AI/blockchain training. These are programme design and intended-learning claims, not evidence of a family office’s production model, investment mandate, or data access.
• PolyU’s official programme-brochure page and the current programme page name Prof. Jimmy Jin as programme director. The route is useful for finding family-office practitioners, project briefs, guest webinars, student outputs, faculty papers, and Hong Kong wealth-management networks. The public materials do not identify the industry partners, their supplied datasets, the AI vendors used in class, or any investment authority.
• York University’s YUSIF “About” page describes a current student endowment fund with 24 students across York faculties, an equity portfolio, an academic-year operating cycle, and alumni/advisory-board surfaces carrying employer links across banks, pensions, asset managers and alternative-investment firms. The page is a current talent and allocator-interface lead, but its employer imagery does not by itself identify individual alumni, dates, job functions, or employer systems.
• The official 2016 York/Schulich launch release makes the historical data and custody interface more concrete: S&P Global Market Intelligence supplied access to Capital IQ, CIBC Mellon supplied custody and its Workbench information-delivery platform, and the student structure included a Chief Investment Strategist, Portfolio Manager, senior analysts and junior analysts. The release names founding donor George Engman, a retired Alberta Investment Management Corporation senior vice president, and records a former student leader’s path to Scotiabank Global Banking. These are dated sponsorship, infrastructure, and education-to-industry records; they do not disclose current YUSIF model use, S&P or CIBC Mellon internal AI, or the practices of any employer shown on the current alumni page.

These routes extend the allocator and family-office map in two directions. PolyU exposes a current curriculum where family-office governance, emerging technology, machine learning, and industry projects are intentionally placed in the same programme. York exposes the less visible operating layer around a student portfolio: data-provider access, custody/information delivery, student role hierarchy, allocator alumni, and current employer-linked discovery surfaces. Neither establishes live AI use by a family office, fund, university, vendor, sponsor, or alumnus.

Recovery queue for PolyU and York routes

Recover PolyU’s 2026–27 syllabi, faculty, project briefs, named industry partners, guest webinars, student work, family-office case materials, ML assignments, data terms and alumni destinations. Recover YUSIF’s current roster, historical teams, holdings/performance reports, Capital IQ and Workbench documentation, board/adviser biographies, sponsor records, student research, alumni job histories and employer media. Preserve programme intent, vendor access, custody, alumni association, current employment, confidential data access and verified deployment as separate evidence states.

September 5, 2026 — Systematic-student engineering, academic–industry labs, and finance-AI research surfaces

• The UConn Husky Quant Group describes itself as a student-run systematic fund researching financial markets and deploying live strategies. Its public description names academic finance, econometrics and machine learning in research and backtesting, and engineering work on research tools and execution systems. It reports $26,000 AUM and 19 members and alumni. Its named data sponsors are Carbon Arc, which supplies alternative-data categories such as credit-card activity, foot traffic and TikTok Shop metrics, and DataBento, which supplies direct exchange feeds and level-3 order-book data. The team page names Shium Mashud (Portfolio Manager), Ethan Carty (Quant Research Lead), Josef Karpinski (Engineering Lead), and Brendan Barnett (Founder, Advisor), alongside current researchers and developers. This is a student-fund description and vendor relationship, not evidence of a professional firm’s systems or of the sponsors’ internal data practices.
• Monash’s Centre for Quantitative Finance connects stochastic control, asset pricing, financial econometrics, machine learning and data science with practitioner-linked research. Its public timeline records a 2020 climate-transition-index project with ClimateWorks Australia, BNP Paribas and ISS ESG; Henry Wong joining as an adjunct associate professor while Head of Quantitative Solutions at Cbus Super Fund; Mark Aarons joining from VFMC’s portfolio-risk and solutions function; and a research agreement with Vigeo Eiris. The people page identifies Ivan Guo as Centre Director, Oscar Tian as Deputy Director and a senior consultant in pricing and risk analysis at NAB/MLC Wealth, and Henry Wong as the Cbus-linked adjunct. These records establish research and practitioner interfaces; they do not establish which models, data rights or production systems any employer uses.
• Imperial’s AIDA-Finance lab says its work spans AI, signal processing and optimization, with topics including generative AI for financial decision-making, synthetic or surrogate datasets, NLP/LLMs for financial text, graph methods, high-frequency data and limit-order-book modelling. Its linked FinLlama project describes a Llama 2 7B fine-tune using four labelled public financial-text datasets, PEFT/LoRA, 8-bit training and a single A100, followed by a 2015–2021 S&P 500 article/return experiment. The page reports the authors’ experimental portfolio construction and metrics, but the public material does not establish independent replication, production ownership, transaction-cost treatment or live deployment. The lab says it works with investment banks, hedge funds and asset managers, without naming them on this page.
• Columbia’s FABULYS centre publicly lists AI/deep learning/ML, data science, financial engineering and algorithmic trading, asset and wealth management, optimization and real-estate finance among its research and education areas. It advertises an Artificial Intelligence in Real Estate course and describes capabilities such as data collection, feature engineering and AI/ML/deep learning, with an intended industry and government partnership surface. The page mentions former quantitative traders from Wall Street firms and hedge funds but does not name them. This is evidence of a centre-level research and partnership interface, not a named fund relationship or proof that a particular capability is deployed.
• Rice’s Center for Operations and Financial Engineering partnerships page describes academic–industry research pipelines, simulated training, internships, real-world projects, workshops and the Eubank Conference on Real World Markets. Its undated invited-lecturer roster includes academics and practitioners from NYU, Bank of America, Prota Financial, Westwood Holdings, QuantRoll Capital, Fidelity Investments, Liberty Mutual, Vaughan Nelson, Gunvor and Troop Capital Management. The roster is a discovery surface for speakers, papers, employers and project histories; because the page does not date each appearance or state each person’s project scope, it does not establish a current firm relationship, data source or AI deployment.

These routes add evidence about academic curricula, student-fund operations, vendor interfaces, research labs, practitioner appointments and named research themes. They do not support ranking institutions or inferring that an employer, sponsor, alumnus or academic lab has deployed a model merely because a public page mentions machine learning, generative AI, financial data or a related job title.

Recovery queue for UConn, Monash, AIDA-Finance, FABULYS, and CoFES

Recover UConn’s dashboard, GitHub, blog, sponsor terms, current and historical rosters, research notes and execution disclosures. Recover Monash project papers, PhD supervisors, practitioner biographies, ClimateWorks/BNP/ISS ESG/Cbus/VFMC/Vigeo Eiris project artefacts and current events. Recover AIDA-Finance paper code, datasets, model cards, author affiliations and named collaboration records, while separating research experiments from live trading. Recover FABULYS faculty, course materials, project briefs, named partners and former-trader biographies. Recover CoFES event dates, recordings, slides, speaker affiliations and project references. Preserve public capability, coursework, sponsorship, employment, confidential access, model ownership and verified deployment as separate evidence states.

September 5, 2026 — Professors, finance curricula, and explicit research objects

• Emory Goizueta’s faculty profile for Lakshmi Shankar Ramachandran identifies him as an Associate Professor in the Practice of Finance whose areas include FinTech and asset pricing. The profile records a PhD in Finance from EdHEC, an MS in Computational Finance from Carnegie Mellon, and a B.Tech. from IIT Madras. It says he has taught FinTech, AI, ML, deep learning and blockchain alongside investments, derivatives and financial-risk analytics, and records executive training for named banks and market institutions. This is a faculty biography and teaching route; it does not establish a firm’s use of his course material, any model ownership, or investment deployment.
• NJIT’s Ajim Uddin research page exposes a more specific asset-pricing research surface. Uddin describes nonlinear tensor factorization, network representations of financial markets, dynamic graph structures, graph neural networks, explainable ML and fairness-aware credit models. The page links code for earnings-forecast missing-data work and attention-based dynamic graph learning, and lists papers on signed graph Laplacian factors, dynamic graph neural networks, explainable stock-return prediction, nonperforming-loan forecasting and earnings forecasts. It also names a PhD student working on LLM financial sentiment analysis. These are public academic and code links, not evidence of a manager’s production stack or data rights.
• MIT Sloan’s Taha Choukhmane profile identifies a current research line on whether AI can provide personal financial advice. Its June 2026 entry names co-authors Weidong Lin, Matthew Akuzawa and Tim DeSilva for “AI Financial Advice: Supply, Demand, and Life Cycle Implications,” and records a 2026 Swiss Finance Institute paper award. The page connects the work to household saving, investing, prompts and differences in user characteristics. This is relevant to allocator and wealth-management workflow design, but it is not evidence about hedge-fund alpha, institutional deployment, or model performance beyond the cited research claims.
• Lehigh’s Donald Bowen III profile names Data Science for Finance and FinTech Capstone as courses and describes NLP, LLMs and coding used to build datasets about corporate investment, innovation, patent markets, venture capital and IPOs. His public links include a data-science course, SSRN work, patent data and a CV; the biography records Arizona State finance/economics training and a University of Maryland finance PhD. The route is useful for alternative-data and document-to-dataset design, but the page does not establish a fund partnership, proprietary corpus, live system or investment result.
• Florida State’s Alec Kercheval profile identifies him as Director of the Financial Mathematics MS and PhD programmes and an affiliated researcher with UC Berkeley’s Consortium for Data Analytics in Risk. The page lists portfolio-risk covariance shrinkage, statistical and ML models of intraday limit-order-book behaviour, derivative pricing, credit-risk models, and agent-based asset-pricing dynamics. Its publication list includes a 2015 SVM study of high-frequency limit-order-book dynamics and 2025–26 work on high-dimensional eigenvector shrinkage and long-only minimum-variance portfolios. This is a concrete programme, model-class and paper route; it does not disclose a trading firm’s implementation or deployment.

Together these routes broaden the academic search beyond an “AI lab” label: faculty biographies can expose training lineages and practitioner teaching; course pages can expose the intended workflow; code links can expose reproducible model objects; and mathematical-finance programmes can expose microstructure, risk and portfolio-construction targets. None should be converted into a ranking or a claim about a firm’s internal practice without separate corroboration.

Recovery queue for the professor and programme routes

Recover Emory course syllabi, executive-training dates and public research outputs; NJIT repositories, student papers, code licenses and dataset provenance; the MIT AI-advice paper, data, prompts, experimental design and replication materials; Lehigh course repositories, patent-data documentation and paper code; and FSU programme syllabi, student supervisors, seminars, code and current industry links. Track education, supervision, employment, collaboration, data access, model ownership and deployment as separate fields.

September 5, 2026 — Programme infrastructure and data-access signals

• UCL’s Computational Finance MSc lists financial engineering, numerical methods, data science and machine learning for finance as core areas, with options in algorithmic trading, market microstructure, networks and systemic risk, numerical optimisation and advanced ML. It requires a substantial research or engineering project; the page says projects are often conducted through placements organised with a bank, hedge fund, fintech, financial-services firm or regulator, but that statement does not identify a specific host or guarantee a placement. This is a programme design and talent-pipeline signal, not evidence of a host’s model or data access.
• Leeds’ 2026 Financial Mathematics MSc combines derivative pricing, stochastic models, risk management, statistical and machine learning, portfolio optimisation, R and Python. The page names two on-campus trading rooms and access to Bloomberg, WRDS, CSMAR, Refinitiv Eikon and Datastream, and describes guest input from finance and regulatory professionals. Those are published educational resources and intended access arrangements; they do not establish a student’s or employer’s use of any particular dataset, model or strategy.
• Sussex’s Advanced Financial Modelling and Machine Learning module specifies ARMA, GARCH, cointegration, LASSO, PCA, tree-based models and copula dependence analysis, applied in Python with Reuters data. The module is planned for 2026/27 and is assessed through a project and computer-based examination. This is a dated syllabus-level method and data route; it does not establish production use, licensing beyond the course context, or investable performance.

These programme pages reveal a different discovery layer from faculty biographies: named databases, trading rooms, project placements, coding languages, module assessments and model families. They are useful for tracing talent and possible industry interfaces, but programme access, student work, employer affiliation, data rights, model ownership and deployment must remain separate evidence states.

Recovery queue for programme infrastructure routes

Recover UCL project-host records, faculty and dissertation topics; Leeds module handbooks, database terms, trading-room rules, guest-speaker records and student projects; and Sussex staff, module materials, Reuters access terms and assessed-project outputs. Search each programme’s alumni, placement hosts, GitHub, papers, seminars and employer biographies without converting programme marketing into firm-level claims.

September 5, 2026 — Hong Kong and Singapore finance-AI programme infrastructure

• City University of Hong Kong’s profile for Houmin Yan identifies him as Chair Professor of Management Sciences, director of the MSc in Accounting and Finance with AI and FinTech Applications, and director of the Hong Kong Laboratory of AI-Powered Financial Technologies. The profile lists stochastic models, machine learning and algorithms, and risk modelling among his research areas; records a University of Toronto PhD and Tsinghua engineering degrees; and links a 2025 paper on conditional generative modelling for credit-risk management in supply-chain finance. This establishes a named programme/lab and model-method route, not a hedge-fund system or production deployment.
• CityU’s profile for Qi Wu identifies him as director and PI of the JD Finance–Hong Kong CityU FinTech and Engineering Joint Laboratory. The public profile lists machine learning and financial technology as research interests, a Columbia financial-engineering PhD, Peking University and Wuhan University training, and prior quantitative roles at DTCC, UBS and Lehman Brothers. It also lists interpretable ML for financial-risk modelling and a strategic collaboration with JD Finance. These are first-party biography, lab and grant-route claims; the page does not establish JD Finance data access, a fund relationship, a live model or model ownership.
• NUS Risk Management Institute’s MFE curriculum exposes a detailed educational workflow. It lists systematic equity strategy development, risk models, liquidity and trading costs, technical and sentiment factors, portfolio construction and optimisation; a Machine Learning and FinTech course covering deep neural networks, topic modelling, cryptocurrencies and sentiment analysis; an electronic-financial-market course covering front-, middle- and back-office architecture and algorithmic-trading solutions; and a C++ project building components of a risk-management system. These are published course objectives and exercises, not evidence that a bank, hedge fund or student project uses the same methods in production.

These Hong Kong/Singapore pages show why programme mining must include director biographies, joint-lab names, prior quant employment, research-grant pages, course objectives and capstone descriptions. They expose model vocabulary and talent pathways, but not proprietary data, permissions, employer adoption, model ownership or live deployment.

Recovery queue for CityU and NUS routes

Recover CityU’s AI-finance programme syllabus, Houmin Yan’s full publication and teaching record, Qi Wu’s joint-lab agreement and project outputs, named fund or bank collaborators, and any public lab seminars. Recover NUS MFE module handbooks, instructors, invited practitioner sessions, code/data exercises, student projects, internship destinations and risk-system assignments. Preserve curriculum intent, grant funding, prior employment, sponsorship, data access and verified deployment as separate evidence states.

September 5, 2026 — Canadian student investment firms and quantitative-investment training

• McGill Desautels’ Desautels Capital Management page describes a university-owned, student-run, licensed and regulated investment firm with $9 million in assets and four funds: global equity, fixed income, Alpha Squared equity, and socially responsible investing equity. The page says the first two funds are managed by Honours in Investment Management students and the latter two by Master of Management in Finance students, with student analysts led by strategists and guided by faculty and the Desautels Global Experts network. For the Alpha Squared fund, it explicitly names research, trade execution, portfolio and risk management, compliance, settlement, macro views, fundamental valuation and quantitative analysis, and identifies Vadim di Pietro as CIO. This is a public educational investment-firm workflow, not evidence of a hedge fund’s AI system or of any named sponsor’s model.
• McGill’s Martlet Fund describes a student-run fund with event-driven, long/short-equity and quantitative divisions. Its quantitative description says students build systematic models using data, statistics and machine learning; the site also exposes student finance/STEM recruiting, programmes, team and alumni routes. No capital figure, dataset, code, model specification, performance audit or professional-employer adoption is disclosed on the reviewed page, so the ML language remains a self-described educational activity.
• The University of New Brunswick’s MQIM6605 Quant Student Investment Fund course says students manage a multi-million-dollar portfolio under faculty guidance and the fund’s investment policies, using quantitative techniques to forecast return and risk and construct portfolios. The course requires a strategy pitch book and a group presentation with detailed backtesting information. This is a concrete academic deliverable and oversight route; it does not identify the models, data rights, live execution controls or any employer’s use of the work.

These Canadian routes add a useful spectrum of evidence: a regulated educational investment firm with explicit operational roles, a student fund that publicly names ML as one division’s method, and a graduate course with specified portfolio/backtest deliverables. They should be mined for faculty oversight, strategy documents, alumni, code, data terms and events without being treated as rankings or evidence of professional-firm deployment.

Recovery queue for Canadian student-capital routes

Recover McGill fund publications, current strategists and faculty approvals, DGE biographies, fund reports, data/vendor terms, student research and alumni employer paths. Recover Martlet’s current team, programme materials, model/code artefacts, events, competition records and alumni network. Recover UNB MQIM syllabi, Vestcor partner records, faculty advisers, portfolio reports, pitch books, backtests, student outputs and internships. Separate student activity, sponsor/partner affiliation, educational access, model ownership and verified deployment.

September 5, 2026 — Faculty research factories and finance-practitioner lineages

• Saint Louis University’s Norman Guo research page describes research on FinTech, machine learning, investments, hedge funds, mutual funds and financial analysts, with an emphasis on explainable models. It also says his empirical workflow is run by an “AI Research Orchestra” of 28 specialized agents and nine conductors, spanning literature retrieval, data extraction, analysis, drafting, review and validation, with the professor retaining approval gates. The linked public GitHub repository describes the same architecture as a finance/economics-oriented research factory with locked research designs, data-quality checks, adversarial review, reproducibility controls and human approval. This is concrete evidence of an academic research-workflow implementation; it is not evidence that a hedge fund uses the system or that its outputs are investable.
• Guo’s public research list links “Decoding Mutual Fund Performance: Dynamic Return Patterns via Deep Learning,” a sequential deep-learning study that the page says forms a long-short portfolio and reports a 2.8% annualized four-factor alpha persisting up to four years. It also lists “The Impact of AI Adoption on Hedge Fund Performance,” which the page says studies returns, risk and portfolio concentration, and “Can Machines Understand Human Skills?,” which uses machine learning to select financial analysts and aggregate their forecasts. These are author-page summaries and working-paper/publication claims; they require paper-level, data, split, transaction-cost and replication checks before being treated as financial evidence.
• The Norman Guo CV gives a temporal and training lineage: Assistant Professor of Finance at Saint Louis University since 2022; Georgia State finance PhD; Bentley finance MS; and South China Normal University finance BS. It lists the deep-learning mutual-fund publication, the AI-adoption hedge-fund paper as under submission, an analyst-selection paper at revise-and-resubmit stage, and a separate working paper on AI accessibility and institutional trades. The CV also records conference presentations and computational tools, but does not establish a firm sponsor, proprietary data licence or production trading permission.
• The University of Edinburgh profile for Ioannis Psaradellis identifies him as a Lecturer in Finance who joined Edinburgh in 2024 after St Andrews and researches empirical asset pricing, asset management, hedge- and mutual-fund performance, and machine-learning applications in finance. It records BSc/MSc training at Athens University of Economics and Business and a D.Phil. in Finance from the University of Liverpool, and links a personal website and LinkedIn route. The profile is useful for recovering papers and seminar networks around fund evaluation; it does not disclose a manager’s model, data, deployment or employer relationship.
• Brandeis’s Steve Xia profile records a professor-of-practice route with interests in quantitative investment, portfolio management, fixed income, asset allocation, derivatives, machine learning and AI in finance. The first-party profile lists current senior-management work at Guardian Life, prior active-asset-allocation research at Fidelity Investments in Boston, prior principal work at Vanguard, and an MIT PhD. This is a Boston/MIT practitioner and teaching lineage worth following through papers, course materials and alumni, but it does not establish that Fidelity, Vanguard or Guardian use a particular AI system, dataset or model.

The academic signal here is not a claim that these routes are equivalent to professional-firm practice. It is a set of observable research objects: an agentic research workflow with explicit human gates; academic tests of nonlinear fund evaluation and AI adoption; a finance faculty pipeline focused on hedge-fund measurement; and an MIT-linked practitioner biography. The next research step is to recover paper versions, code, data provenance, model revisions, conference recordings and any separately documented industry collaborations.

Recovery queue for faculty research-factory and practitioner routes

Recover Guo’s paper PDFs, repositories, agent prompts, guard hooks, data manifests, empirical splits and conference versions; Psaradellis’s publications, co-authors, seminars, datasets and fund-performance methods; and Xia’s papers, course materials, practitioner talks, Fidelity/Vanguard/Guardian dates and public research identifiers. Treat personal-page claims, CV claims, university affiliations, prior employment, academic results, confidential access and verified deployment as separate evidence states.

September 5, 2026 — New academic tests of AI adoption, ownership, and investment horizons

• CEIBS’s “Generative AI and Asset Management” research page proposes a measure of investment-company reliance on generative AI and says its hedge-fund analysis uses a difference-in-differences design plus ChatGPT-outage shocks. The page reports an annualized abnormal-return difference for adopting funds and attributes the result to AI talent and firm-specific information analysis, while noting no comparable result for non-hedge funds. The page does not name the paper’s authors, publish its data construction, adoption classifier, outage assumptions, costs, fund sample, code or independent replication. It is therefore a research lead and a claim to verify, not evidence of any tracked manager’s GenAI system or performance.
• The Tsinghua paper “AI Adoption, Mutual Fund Short-Termism, and Real Investment” by Xiang Huang and Haifeng You studies a different mechanism. Its abstract says AI adoption is associated with greater acquisition of short-term fundamental information and less long-term fundamental information, with implications for near-term earnings price informativeness and corporate investment. The paper says it uses manager biographies and public sources, applies LLMs to identify AI-related expertise, and classifies a fund from the quarter of its first AI-related manager hire; its sample period is primarily 2010–2023, before broad GenAI diffusion. This is a paper-level measurement and market-impact route, not proof that any identified fund used an LLM in production.
• Wharton’s Mack Institute page on “Ownership, Corporate Ambidexterity, and AI Adoption” identifies Elaine Pak as the PhD-candidate author and frames private-equity and hedge-fund ownership as a comparison group against long-term institutional owners. The public abstract asks whether short-horizon monitoring and performance metrics steer AI toward optimization at the expense of exploratory innovation. It is useful for the allocator/family-office question because it turns ownership structure and monitoring capacity into explanatory variables, but the page does not disclose the sample, firm identifiers, measures, estimates or production AI systems.

These sources add three distinct research objects to the academic map: a public GenAI-adoption measure for investment companies, an LLM-assisted manager classification tied to information horizons and real investment, and an ownership/governance framework for AI implementation. Their claims must be checked at paper and data level, and they must not be translated into rankings or deployment assertions about professional firms.

Recovery queue for AI-adoption and ownership studies

Recover the CEIBS paper and author metadata, fund universe, adoption measure, outage event windows, code and appendix; the Tsinghua paper’s manager-biography corpus, LLM prompts, labels, dates, instrument construction, robustness tables and replication materials; and Elaine Pak’s full paper, sample, ownership definitions, AI measures, conference versions and advisor/doctoral lineage. Track published result, author summary, model-assisted label, firm identification, employment, data access and verified deployment separately.

September 5, 2026 — Family-office data plumbing and academic–industry training surfaces

• The EDHEC CV for Victor Planas-Bielsa records an unusually concrete family-office data interface. It says that since 2020 he has designed and implemented a custom automated ETL for a family office using about 30 daily data sources. The same CV places him at EDHEC from 2024 as academic director for the Data Science and AI for Business track and professor of Introduction to Machine Learning, and records earlier work as research director of the Hedge Funds Research Institute, a centre sponsored by Alpstar Management. In that earlier role it lists backtesting improvement, derivatives-model validation, global hedge-fund trend and capital-flow monitoring, and board-level investment recommendations. These are CV claims; the family office is unnamed, the sources and pipeline are not disclosed, and the record does not establish current hedge-fund or family-office deployment beyond the stated consulting engagement.
• HKUST’s Center for Wealth Management explicitly combines quantitative investment and market efficiency, private wealth and family-office ecosystems, digital wealth management and fintech integration. Its description names data-, algorithm- and machine-learning-driven quantitative strategies as research objects and separately lists family-office governance, succession, cross-border regulation and investment preferences. The public team list includes Jianfeng Yu, Dong Lou, Abhiroop Mukherjee, Nicholas Barberis, Lauren Cohen, Haizhou Huang, Howard Kung and Christopher Polk. This is a research and personnel discovery surface for Asian allocators; it does not identify a family office’s data, model, vendor, investment authority or live system.
• Bond University’s Rand Low profile records an Associate Professor of Quantitative Finance whose research includes portfolio optimisation, risk management, systematic trading and multi-asset investing. The profile says he previously led quantitative teams at Bank of America Merrill Lynch and BlackRock building market, credit and operational-risk, securities-lending, structured-product, asset-backed-security and portfolio-management models, and describes model validation, stress testing and governance work with regulators. It also lists current interests in ML for corporate credit ratings, robo-advisors, digital assets, commodities, systematic active investment and business-process automation. These are profile-reported academic and employment claims, not evidence of current employer systems or a fund’s production use.
• The GAME XIII programme adds a title-blind conference route. Its March 2024 programme includes “Machine Learning Driven Investment: How to Use AI as an Investment Tool,” presented by Arezu Moghadam, Managing Director and Global Head of Data Science at J.P. Morgan Asset Management; a panel with Jessica Iorio, Divisional Operations Director at Rockefeller Global Family Office; and sessions on student-managed funds, hedge-fund industry careers, portfolio competitions and teaching/research assessment. The programme is dated and does not provide recordings, slides, employer implementation details or evidence that the speakers’ organizations used a shared system. It is a speaker and follow-up discovery route.

These sources expose a practical expansion of the academic map: family-office ETL design, hedge-fund research-center history, allocator ecosystem research, model-risk governance, and conference speakers whose titles reveal data-science or family-office operating roles without using “AI” in the event title. They do not justify comparing firms or attributing academic, conference, or prior-employer material to a current production system.

Recovery queue for family-office and academic–industry routes

Recover Planas-Bielsa’s ETL architecture, source categories, data contracts, family-office identity only where publicly documented, HFRI publications and Alpstar dates; HKUST center papers, events, team biographies, family-office research and project partners; Low’s publications, model-governance talks, dated employer roles and teaching materials; and GAME recordings, slides, session pages, speaker profiles and student-fund artifacts. Preserve consulting claims, prior employment, academic research, event participation, data access, model ownership and verified deployment as separate evidence states.

September 5, 2026 — Caltech student capital and quant-talent pathways

• The current Caltech Student Investment Fund site reports a portfolio valued at more than $1.9 million as of Q1 2026 and describes weekly board meetings covering market trends, portfolio performance and new opportunities. Its separate investing page says the fund had begun exploring quantitative finance but “do[es] not yet employ quantitative strategies for active portfolio management.” That explicit non-use statement is useful negative evidence: the presence of a technology-oriented university and a quantitative-finance discussion does not establish a live systematic process.
• Caltech’s board page publicly lists roles and next-step employment surfaces: Harsh Kooshal Gandhi as co-president and incoming fixed-income and macro-trading intern at Citadel; Dhruv Verma as co-president and incoming quantitative-research intern at Statar Capital; Kevin Cai as portfolio manager for commodities and incoming quantitative-trader intern at Jane Street; Ali Niazi as portfolio manager for healthcare and incoming Point72 intern; and Varun Gabbita as secretary and incoming Princeton research fellow. It also lists applied-math, computer-science, electrical-engineering and bioengineering training among officers. These are dated public student biographies and talent-pipeline signals, not evidence that the firms recruit from, sponsor, or use the fund’s work.
• The Caltech pages expose a useful temporal consistency check: the current home page reports more than $1.9 million in Q1 2026, while the older investing page reports $1,371,816.19 as of May 16, 2025 and approximately $1.4 million as of Q3 2024. The figures are not merged into a growth series because the pages use different dates and presentation contexts, and no audited statement was reviewed.

Caltech belongs in the discovery map as a student-managed capital and talent route with an explicit boundary between quantitative-finance exploration and active use. The board roster is useful for recovering public alumni transitions and conference or recruiting surfaces, but it must not be used to infer firm strategy, model ownership, data access or deployment.

Recovery queue for Caltech student-capital routes

Recover Caltech’s current portfolio history, board minutes or public proposals, quantitative-finance project materials, officer biographies and alumni transitions; confirm dates and titles from employer or student sources; and look for public seminars, code, competitions and faculty oversight. Keep student activity, future employment, recruiting, university affiliation, sponsor relationship and verified professional deployment separate.

September 5, 2026 — Oxford multi-agent learning and finance-research route

• The Oxford Martin AIGI profile for Jakob Foerster identifies him as an Associate Professor leading the Foerster Lab for AI Research at Oxford Engineering Science. The profile records prior research time at Google Brain, OpenAI, DeepMind and Facebook AI Research; work on multi-agent cooperation, coordination, communication and zero-shot coordination; and a J.P. Morgan AI Research Award for work on AI in finance. It says the current lab focus combines large-scale pretraining with reinforcement learning and search, alongside meta- and multi-agent learning. This is a named academic lab and research-method route, not evidence of a J.P. Morgan, hedge-fund or asset-manager production system.

The Oxford route adds a coordination and multi-agent layer to the academic map. The public profile exposes research lineage, lab ownership and award-topic evidence, but not finance datasets, model weights, market simulator design, decision permissions, code, partner scope or live deployment.

Recovery queue for the Oxford agentic-finance route

Recover Foerster Lab publications, J.P. Morgan award details, finance-specific papers, simulators, datasets, code, student and postdoctoral rosters, seminars and any explicitly documented industry collaboration. Keep award, academic affiliation, research method, collaboration, data access, model ownership and deployment as separate evidence states.

September 5, 2026 — Academic finance programmes and faculty as idea-discovery routes

The KIT Computational Risk and Asset Management Research Group (C-RAM) page adds a distinct applied-research surface to the existing KIT route. Led by Prof. Dr. Maxim Ulrich, it describes financial-economic modelling combined with machine learning, data science and high-performance computing; studies across equities, options and macro-financial variables; model-free dividend-risk-premium estimation from option and survey data; and NLP analysis of ECB communications across asset classes. The page stresses robustness, interpretability and computational scalability and links public YouTube, X, Instagram and Facebook accounts. These are academic methods and discovery surfaces, not evidence of a named manager’s system, partner data access, portfolio authority or deployment.

Two programme pages are useful because they expose research and talent interfaces rather than just course marketing. Johns Hopkins Carey’s MS in Finance lists Financial Data Analytics, Machine Learning for Finance, a Financial Data Science concentration, a $500k-plus student-managed fund, and course-level use of RAG and agentic AI. Wharton’s Jacobs MSQF FAQ describes four-to-six-person, faculty-guided applied research teams moving from messy data to hypotheses, risk analysis and communication, within a curriculum covering price/risk modelling, portfolio theory, ML, AI, regulation and business analytics. Duke’s Quantitative Finance Concentration adds practitioner-supported projects and a course sequence spanning time series, algorithmic trading, decision optimisation, risk and derivatives. None of these programme descriptions establishes a firm’s model, data rights, adoption, trading permission or performance.

The practical implication for the research map is to index professors and finance programmes as separate evidence routes: faculty papers for methods and hypotheses; labs for researchers, code and infrastructure; student funds for explicit non-use or test artifacts; and programmes for guest speakers, project briefs, employer interfaces and alumni transitions. Recover those artifacts before cross-linking a person or idea to any investment manager. Academic affiliation, prior employment, sponsorship, data access, model ownership, portfolio authority and verified deployment remain separate states.

Recovery queue for the academic-programme routes

Recover C-RAM publications, KABFI projects, named team members, linked media and European collaborations; Johns Hopkins syllabi, Frank Talk episodes, student-fund reports and guest identities; Jacobs practicum briefs, faculty/advisory roster, public project data and alumni destinations; and Duke syllabi, project artifacts, speaker recordings and student-fund or competition materials.

September 5, 2026 — Doctoral data access and title-blind finance-AI events

Reading’s ICMA Centre PhD in Finance is a useful research-infrastructure route. The programme says researchers can access WRDS, LSEG, Bloomberg, Datastream, SAS, Stata, Matlab, Python and Eventus, as well as three dealing rooms. Its published topics include fintech and machine learning, algorithmic trading and execution, market microstructure, institutional investors, model risk, textual analysis, volatility and correlation, and portfolio-performance assessment. The programme also describes weekly research seminars and a practice-centred connection to industry. This supports searches for supervisors, doctoral work, papers, code and seminars; it does not establish access to a fund’s proprietary data or live trading authority.

King’s College London’s “AI and Finance — Collaborative Frontiers” programme PDF provides a title-blind event route dated April 11, 2025. It names Bart De Keijzer of King’s Distributed AI Research Group on clearing financial networks with CDS; Silvia Bartolucci of UCL’s Financial Computing and Analytics Group on deep learning for limit-order-book forecasting; James Hamp, Director and Head of FX Data Strategy and Analytics at Citi London, on AI applications, organisational constraints, productivity and automation in large-bank markets; and Christian Julliard of LSE on Bayesian model averaging for a joint stock-and-bond factor zoo. The PDF supplies public roles and abstracts but no recordings, employer implementation details, data permissions, model ownership or investment results.

The discovery value is methodological: doctoral database lists can reveal feasible public-data replications and supervisor networks, while event PDFs expose practitioner titles that do not appear in episode headlines. Each name should be searched across institutional pages, papers, seminars, code repositories, LinkedIn and video archives, while academic research, employer role, event participation, data access, model ownership and deployment remain separate evidence states.

Recovery queue for doctoral and title-blind event routes

Recover ICMA supervisor and doctoral rosters, current theses, seminar recordings, database/licensing descriptions and alumni transitions; and King’s/LSE/UCL recordings, slides, paper versions, speaker biographies, Citi public material and any explicit code or data disclosures.

September 5, 2026 — Academic agent artifacts and model-behaviour controls

University of Illinois Gies Business identifies Clinical Assistant Professor Tony Zhang and describes FIN 580 students assigning separate agents to valuation, risk and alternative-data collection while building trading platforms. The account describes a five-agent simulated hedge-fund exercise and a student-reported short-horizon return claim after thousands of backtests. That performance claim remains explicitly unvalidated; the durable evidence is the classroom architecture—role separation, agent cross-examination, missing-data handling and human judgment—not a professional-firm system or investment result.

The University of Chicago Data Science Institute capstone names four student authors building a fine-tuned LLaMA agent for repeatable quantitative-research work. Its design combines external tools, agentic orchestration, RAG and domain-specific instructions, with paper replication as the representative task. The page links a presentation timestamp and identifies Justin Kurland as faculty advisor and a Goldman Sachs Engineering Division Tech Fellow. The university’s reliability and analyst-time claims remain project-page claims pending the paper, code, data and recording; nothing here establishes adoption by Goldman Sachs or a hedge fund.

Auburn’s Harbert College report names Stace Sirmans, Javad Keshavarz and Cayman Seagraves and describes 25 investment questions presented in biased and neutral forms to 48 AI models. The report says the experiment surfaced framing, anchoring, narrative-cue and loss-aversion effects, with newer model versions reducing some effects but showing stronger loss aversion. This is a governance and model-behaviour route, not a return forecast or deployment disclosure; exact prompts, model versions, the linked SSRN paper and replication remain open.

The academic layer now offers three distinct test surfaces beyond news sentiment: multi-agent research orchestration, paper-replication throughput and behavioural robustness under prompt perturbation. They are useful benchmark ideas, but any investment-facing test still needs fixed model versions, point-in-time data, reproducible traces, cost accounting, human approval gates and out-of-sample evaluation. No cross-firm conclusion is drawn from these university artifacts.

Recovery queue for academic agent artifacts

Retrieve Illinois page assets and student-project details; recover the Chicago Box recording, paper, repository and data; retrieve the Auburn SSRN paper and exact prompt/model matrix; and independently verify every performance or productivity statement.

September 5, 2026 — Academic–manager interfaces and finance-agent evaluation routes

The University of Chicago Data Science Institute’s April 2026 Millennium announcement is a direct institutional partnership disclosure. UChicago says Millennium joined the DSI Industry Affiliate Program, with joint initiatives involving student engagement, applied learning, project collaborations and seminars. It also says Millennium will lend strategic insight and industry expertise to a new quantitative-developer certificate being developed with UChicago Financial Mathematics, and names Pranat Pathak as Millennium’s International CIO and Global Head of Fixed Income, Commodities and Core Technology. This verifies an academic–manager interface and a named executive route; it does not disclose the project corpus, code, model ownership, investment use or portfolio authority.

Berkeley RDI’s AgentX–AgentBeats competition adds a public evaluation route with a dedicated Finance Agent track. The platform uses a “green” evaluator agent to define tasks, environments and scoring and a “purple” agent under test communicating through A2A. The page describes research and finance tracks involving document extraction, multi-document computation and statistical analysis. This is benchmark infrastructure, not evidence that a fund uses a participating agent or that a benchmark score predicts returns.

UChicago Career Advancement’s finance-competition page lists a September 18, 2026 AI-Enabled Investing Competition in which students use AI tools to build investment analyses and models with industry feedback. The same page says restructuring-competition participants receive access to Constellation Finance, an AI-based credit-analysis tool, and lists Marblegate Asset Management, Oaktree, Aegon, Lazard, Houlihan Lokey and other sponsors. These are competition and vendor-interface signals; they do not establish sponsor data access, firm adoption, model ownership, trading authority or performance.

Stevens’ 2026 High Frequency Trading Competition describes student teams coding and deploying intraday strategies on the in-house SHIFT simulator through Python and FIX Protocol interfaces. Its rules mention Dow Jones 30 historical assets, synthetic tickers for agent-based simulation days, market fees/rebates, multi-week scoring and reinforcement-learning-agent scenarios. This is an academic simulation and recruiting surface, not evidence of live-market access, employer use or production deployment.

The research map now separates affiliate status, curriculum influence, project collaboration, seminar access, data provision, tool access, student competition and employment. These public interfaces can expose personnel, vendors, evaluation protocols and recruiting paths without revealing proprietary model internals or live investment authority.

Recovery queue for academic–manager interfaces

Recover Millennium–UChicago certificate materials, project briefs, speakers and student outputs; AgentBeats Finance Agent definitions, datasets, leaderboards and traces; UChicago mentors, Constellation Finance terms and competition artifacts; and Stevens SHIFT rules, recordings, student repositories and employer or sponsor disclosures.

September 5, 2026 — Quant-developer curriculum and architecture-selection research

UChicago’s Quantitative Developer Certificate page deepens the Millennium partnership route. The Data Science Institute and Financial Mathematics Program describe a certificate combining advanced software engineering with financial mathematics, a capstone presented to the Industry Advisory Board and faculty, and an anticipated Autumn 2026 launch. The displayed advisory board includes Chicago Trading Company, DRW and Millennium. The page names Mark Hendricks, David Uminsky, Arnab Bose, Emily Backe and Anne Brown in programme leadership and says the certificate will draw graduate students from data science, financial mathematics, computer science, applied mathematics, finance, statistics and physics. This verifies a curriculum and employer-interface structure; it does not disclose project data, model ownership, production access or investment use.

MIT Media Lab’s “Towards a science of scaling agent systems” project reports a controlled evaluation of 180 agent configurations across single-agent, independent, centralized, decentralized and hybrid architectures. The project evaluates Finance-Agent alongside web navigation, planning and tool-use benchmarks. Its page reports gains for centralized coordination on a parallelizable financial-reasoning task, degradation for multi-agent systems on sequential planning, a tool-coordination trade-off, error-amplification measurements and a predictive model using task properties such as tool count and decomposability. These remain project-page claims pending inspection of the paper, benchmark definitions and code; they do not establish any fund’s architecture or deployment.

The research implication is an architecture-selection test rather than a universal multi-agent prescription. A finance research system should measure decomposability, sequential dependencies, tool density, communication overhead and error propagation before choosing a single agent, parallel workers, an orchestrator or a hybrid. Separately, the UChicago certificate shows an employer-advisory route into quant-development education, but the public page does not show what any advisory firm contributed or adopted.

Recovery queue for quant-development and agent-architecture routes

Retrieve the UChicago certificate course list, capstone briefs, advisory-board participation, faculty materials and student outputs; retrieve the MIT paper, benchmark definitions, code, task-level results and Finance-Agent data; and verify claims independently before using them in any investment-facing evaluation.

September 5, 2026 — CMU quant-finance talent map and Stanford applied-AI research agenda

Carnegie Mellon’s MSCF governance page describes a programme spanning computer science, statistics and data science, mathematical sciences and the Tepper School of Business. Its 2025–2026 advisory board lists Max G’Sell at PDT Partners, Roni Israelov and Yumi Oh at Citadel, Reha Tutuncu at Point72, Evans Xiang at Two Sigma, and Manuela Veloso as head of AI research at J.P. Morgan. The alumni board lists Matthew Lyberg as head of asset-management AI at Manulife, Braxton McKenzie at Citadel, Neel Purohit at ExodusPoint, and Xingyao Sun at Balyasny, among other finance and trading roles. This is a programme-published talent-network map for the stated academic year. It does not prove current employment beyond the page, an individual’s research remit, model ownership, or deployment.

Stanford Economics’ Session 3: Applied Artificial Intelligence in Macro-Finance provides several concrete research routes. Allen Hu and Song Ma describe RAG over filings, patents, earnings calls, peer disclosures and news to estimate firm-level marginal projects and construct a “q AI” measure. Bradford Levy and Ralph S.J. Koijen describe a real-time, out-of-sample benchmark using announcement-time information and agentic systems that extract structured signals from earnings calls; the event abstract reports explained variation moving from 8% to close to 20% for the tested systems, pending recovery of the paper and code. Antonio Coppola and Christopher Clayton describe graph-based deep learning over portfolio-holdings networks for intermediary trading behaviour and stress vulnerability. Winston Dou, Itay Goldstein and Yan Ji model market fragility when AI planning enables coordinated trading. These are academic abstracts and programme claims, not disclosures of any tracked fund’s systems or returns.

The practical research agenda is therefore testable by modality: document-grounded latent-project extraction, announcement-time structured signal generation, holdings-network representation learning, and strategic interaction under AI planning. Recover the papers, appendices, code, data definitions, seminars and author biographies before drawing any investment conclusion. A programme board, academic result, or alumni role is not evidence of hedge-fund deployment.

Recovery queue for CMU and Stanford academic routes

Capture CMU MSCF speaker-series recordings and board biographies; recover the Stanford papers, presentation materials, benchmark protocol and author pages; verify each listed employer role independently; and map papers to supervisors, labs, doctoral institutions and public code without inferring a firm’s investment process.

September 5, 2026 — Harvard recruiting surfaces and a CMU quant-talent funnel

Harvard FAS’s Federated Hermes MDT posting is a current Boston hiring artifact. The posting says the MDT investment team builds proprietary software and systems for quantitative strategies and assigns interns to alpha-factor research, model development, machine learning, risk management, portfolio optimisation, large-scale datasets, asset-price forecasting and investment-signal evaluation. It also describes collaboration between researchers and engineers and presentation of findings to senior researchers. This is a role description reproduced by a university career portal; it is evidence of stated hiring scope, not an audit of the system, a filled role, a model inventory, or AI/GenAI deployment.

Harvard FAS’s Trexquant PhD posting describes the employer as a systematic hedge fund using large datasets, statistical and machine-learning methods and scientific experimentation. The role separates alpha research, data science and strategy research, and names novel features, alternative datasets, simulation, academic-paper review, production-ready strategies and a stated path from idea generation to live trading. This is employer self-description in a job advertisement: a useful architecture and talent vocabulary, but not independent verification of live deployment, permissions, data rights, model versions or performance.

Harvard FAS’s Shanghai Sixie Capital posting adds a regional and language-specific route. The role requires Chinese as the working language and lists statistical learning, deep learning, reinforcement learning, feature extraction, market-price prediction, portfolio optimisation, representation learning, generative modelling, time series, NLP, image/video processing, PyTorch and TensorFlow. This is a hiring artifact for a Shanghai-based firm, not evidence that each listed modality is in production or that any model generates attributable returns.

CMU’s MSCF Trading Competition page describes a September 19, 2026 one-day simulated-market event for more than 150 students, organised with the Women in Quant Finance Club, Smart Woman Securities and the CMU Quant Club. Teams are encouraged to split financial-strategy and coding/data-analysis leadership; participants may place résumés in a sponsor-shared resume book, while sponsors receive recruiting, interview and event-access options. This is an explicit academic-to-employer screening funnel. The page does not name sponsors, disclose case data, or establish that student strategies transfer to professional systems.

University recruiting pages expose the vocabulary firms use to divide research work—alpha discovery, data acquisition, feature engineering, model development, simulation, risk, productionisation and senior review—without revealing proprietary implementation. The China posting also shows why language and modality fields belong in the coverage ledger. These signals require comparison with firm-controlled job boards, company media, papers, code and named personnel before being promoted beyond hiring intent.

Recovery queue for Harvard and CMU hiring routes

Archive the Harvard posting pages before their expiry dates; capture the full Trexquant and Sixie application text and employer pages; recover CMU’s 2025 competition video, 2026 case materials, sponsor acknowledgements and results; and independently verify all firm, role and modality claims.

September 5, 2026 — Cross-university AI venture and finance-application funnel

Columbia Business School’s AI Startup Challenge 2026 is a new academic-to-market discovery surface. The challenge accepts graduate-student founders from Columbia, Cornell Tech, NYU and Yale, explicitly includes finance among its target markets, requires AI to be core to the proposed product, and routes finalists toward investor, operator and industry feedback. Its judging criteria include customer validation, meaningful use of AI, differentiation and team execution; the page advertises more than $40,000 in prizes. The page does not disclose judge names, finalist projects, datasets, model choices, customer contracts or investment-firm adoption.

Student venture competitions may expose finance-specific AI ideas before they appear in fund media or academic papers, but they are an ideation and diligence funnel rather than evidence of a deployed investment system. The next recovery targets are final-pitch recordings, finalist pages, public pitch decks, disclosed judges, accelerator affiliations, and any transition from prototype to a named financial institution.

Recovery queue for the Columbia finance-AI venture route

Retrieve the final-pitch page and recordings, finalist or alumni project pages, public decks, sponsor and judge disclosures, and any subsequent company or finance-partner evidence; preserve the distinction between a proposed product, a pilot, and production use.

September 5, 2026 — Academic measurement of GenAI adoption and manager imitation

UC Irvine’s Paul Merage School article on “Generative AI and Asset Management” names Zheng Sun and Jinfei Sheng, with Baozhong Yang and Alan Zhang, and describes a research design that attempts to measure hedge-fund GenAI adoption from holdings data and an April 2025 survey. The school page reports that its analysis finds more than 60% of hedge funds using GenAI for investment decisions and that roughly 70% of surveyed hedge funds said they did so. It frames the proposed mechanism around interpretation of unstructured information such as earnings calls, geopolitical news and regulatory developments, alongside AI-talent measures. These are university communications describing an evolving working paper; the paper, sample construction, adoption classifier, survey instrument, data, code, costs and independent replication remain open.

Harvard Business School’s Working Knowledge explainer for “Mimicking Finance” supplies a more detailed research and data route for manager-behaviour imitation. It says the model predicts quarterly buy, sell or hold directions from U.S. equity mutual-fund holdings in Morningstar Direct from 1990–2023 plus Federal Reserve data, with funds required to be at least seven years old and hold at least ten securities. The article reports 71% overall predictability and describes differences by fund type, experience, size and concentration, while linking lower predictability to subsequent risk-adjusted outcomes. This is mutual-fund evidence and an HBS explanation of an NBER working paper, not hedge-fund evidence, a 13F execution strategy, or proof of live AI use by a manager.

These sources expose two distinct measurement problems: whether a fund appears to use GenAI, and whether a manager’s disclosed behaviour can be imitated. They require point-in-time holdings, disclosure-lag controls, survivorship rules, stable model versions, leakage tests and a clear separation of descriptive predictability from executable returns. The pages identify research designs and personnel; they do not establish a tracked firm’s deployment.

Recovery queue for academic adoption and imitation studies

Obtain the UC Irvine working paper, appendices, survey instrument, holdings-based classifier and code; retrieve the HBS-linked NBER paper, model inputs, sample filters, appendices and any replication; and search the authors’ faculty pages, seminars, datasets and employer links independently.

September 5, 2026 — Finance-programme curricula and Oxford network-learning papers

• LSE’s ST458 Financial Statistics II course guide is a current 2026/27 programme artifact convened by Dr Tengyao Wang. It lists tree ensembles, neural networks and deep learning, LSTMs, factor models, cointegration, Granger causality, high-frequency portfolio allocation, refresh-time choices and gross or maximum-exposure constraints for large portfolios. The course is compulsory on the MSc in Financial Statistics and available on the MSc in Data Science and MSc in Quantitative Methods for Risk Management; LSE reports a 40-student 2025/26 cohort. This exposes the methods and risk/portfolio vocabulary being taught in a London quantitative-finance pipeline, not a fund’s production stack, data licence or live performance.
• Oxford’s Research Archive record for “Text and network based processing for financial machine learning” identifies Dragos Gorduza’s 2025 DPhil thesis, supervised by Stefan Zohren and Xiaowen Dong and funded through the Grand Union Doctoral Training Partnership. The record’s keywords span text processing in finance, graph neural networks, asset pricing, financial regulation and network science. This is a useful faculty/supervisor and research-object route for tracing how textual and relational signals are combined; the archive record does not establish a hedge-fund sponsor, proprietary data access or deployment.
• The related Oxford graph-autoencoder paper studies market-wide stock-correlation networks and reports that edge-reconstruction errors on the S&P index over 2015–2022 correlate with volatility spikes and improve out-of-sample autoregressive volatility modelling. The 2024 analyst-network paper uses analyst co-coverage networks to construct a momentum signal and reports 29.44% annualised returns and a 4.06 Sharpe ratio in its own experiments. Those are paper-reported backtest results, not independently verified investable performance: the implementation, point-in-time controls, costs, turnover, capacity, multiple-testing treatment and replication must be checked before using them as evidence.

Research implication: the university layer is exposing finance-AI ideas that are not just news sentiment—high-frequency portfolio constraints, text-plus-network representations, analyst information-brokerage graphs and volatility monitoring. These are hypotheses and research artefacts to test under a common point-in-time protocol, not a basis for comparing firms or inferring that a named manager uses them.

Recovery: obtain LSE’s ST458 reading list, project briefs, lecturer materials and student outputs; download and inspect Gorduza’s thesis and papers; recover code, feature timestamps, network construction, trading costs, capacity assumptions and all robustness tables; and map Zohren, Dong and Gorduza to current lab, conference, student and employer surfaces without treating academic supervision as industry sponsorship.

September 5, 2026 — Financial-AI synthetic-data route

• University of Surrey’s profile for Dr Amit Kumar Jaiswal adds a concrete industry-to-academia data route. Surrey says Jaiswal worked on a Leeds industry-led Financial AI project involving Credit Suisse and Finastra as a research engineer building a production-grade synthetic-data generator for mortgage-lending and stock-portfolio use cases. The same profile records computer-science and information-retrieval training, multimodal search and recommendation research, and current financial applications involving information representation and modelling. The page supports a synthetic-data and representation-learning recovery path; it does not disclose the project’s data contract, generator architecture, validation results, client permissions, or live investment use.

Research implication: synthetic-data engineering is a distinct research route from forecasting or sentiment analysis. It should be traced through project documents, data contracts, privacy and fidelity tests, collaborators, code and dated employer disclosures before any financial-use claim is made.

Recovery: recover Jaiswal’s Leeds Financial AI project documentation, synthetic-data specification, datasets, evaluation protocol and collaborators; and follow Credit Suisse and Finastra through first-party media, jobs, papers, podcasts and conference pages without inferring firm-wide practice.

September 5, 2026 — Student-capital and sponsored financial-AI research routes

• UCL’s Institute of Finance & Technology research-student directory adds a personnel layer that a lab landing page does not show. It lists Raad Khraishi as a PhD student whose research is funded by NatWest while he works there as a Lead Data Scientist on financial crime, pricing, credit risk and portfolio-management problems; Renda Rundle as a senior portfolio-analytics specialist at ALUWANI Capital Partners researching macroeconomic regime shifts with ML; Philipp Wirth as a PhD student after an AI-investment role at Arabesque and work at an AI-focused German asset manager; Pin Ni in Financial AI, knowledge graphs, NLP and deep learning; and Pornpanit Rasivisuth researching NLP for due diligence and investment strategy in venture capital and private equity. These are university-published biographies and research leads, not proof that the named organisations share systems, data, models or investment authority.
• The University of Manchester NaCTeM NVIDIA project is a new sponsored academic route explicitly focused on “Multi-Agent Reinforcement Learning for Financial AI.” The page frames markets as adaptive multi-agent systems and names Prof. Sophia Ananiadou, Jimin Huang and Dr. Kailai Yang as the project team; it attributes support to the NVIDIA Academic Grant Program and links financial-NLP workshop activity at EMNLP 2026. This exposes a research team, sponsor, multi-agent-RL framing and conference-discovery path. The page does not identify a fund partner, financial dataset, model weights, simulator, live permissions or investment deployment.
• ANU’s extended Student Managed Fund description adds governance and allocator-lineage detail missing from the main fund page. It describes real-money student responsibility, an Investment Advisory Committee, Asset Allocation, Active Australian Equities, Risk and Compliance, and Relationships sub-teams, and says the fund was established with a donation from Russell Clark, an ANU alumnus and Partner at Horseman Capital Management. The page also describes overlapping junior and senior cohorts and rewards for durable process or model contributions. This is a university-managed student-fund and alumni-network route; it does not establish Horseman involvement in the fund’s research, an AI system, a data licence or a production strategy.

Research implication: faculty and programme discovery should capture not only course names but named student researchers, employer-funded doctorates, portfolio-analytics roles, sponsor grants, student-fund governance and alumni donations. These routes can reveal where methods, talent and data questions meet while preserving the distinction between an academic relationship, a prior or concurrent job, a sponsor, and verified deployment.

Recovery: archive UCL biographies and linked papers, supervisors, employer dates and AIRiskLab projects; recover NaCTeM project updates, code, datasets, simulator design, workshop papers and recordings; and obtain ANU SMF reports, committee membership, portfolio history, student models and Horseman-related public material. Do not infer a fund’s internal system from a student, sponsor or alumni connection.

September 5, 2026 — Clemson faculty, student capital, and family-office research

• Clemson’s finance faculty directory links a named ML/asset-pricing researcher to a student-capital pipeline. It identifies Harrison Ham as an Assistant Professor whose research focuses on machine learning, asset pricing, investor information processing and expectation formation, and says he guides the student investment fund while teaching in the MSF programme. The same page identifies Daniel Greene’s research interest in AI applications in corporate finance and records Ron Klotter’s prior institutional-investment roles, including senior portfolio management at Northern Trust, Wellington and Invesco. These are faculty and biography routes; they do not establish a student-fund model, employer data access, or live institutional deployment.
• London Business School’s research archive adds a family-office governance route. The “Family Office Culture and Value Creation” project names Alexander Hayward and Randall Peterson and says it draws on experience analysing the setup, structures and investment strategies of more than 100 family offices. The public description focuses on internal culture, governance and long-term value creation, but does not publish the family-office identities, data, AI tooling or investment decisions.
• The same LBS archive lists Maxime Bonelli’s “Data-driven Investors” project, which asks how financial intermediaries’ use of data technologies such as machine learning affects capital allocation in venture financing. The public summary describes more automated screening as concentrating investment in companies resembling historical data and reducing exposure to rare major outcomes. This is an academic mechanism about investor technology and allocation, not evidence about a hedge fund, family office or deployed model; the paper, sample, identification and replication remain recovery targets.

Research implication: the academic map now includes three separate discovery surfaces—faculty who advise student capital, family-office operating-structure research, and empirical work on how investor automation changes the opportunity set. The appropriate follow-up is to recover papers, datasets, student projects, family-office sample definitions and public personnel links, while keeping academic, practitioner, donor, employer and deployment states separate.

Recovery: retrieve Clemson syllabi, student-fund reports, Ham and Greene papers, Klotter’s dated practitioner materials and programme events; recover LBS project papers, author biographies, family-office sample construction, data sources, appendices and any public follow-up; and do not infer a named investor’s process from faculty or research-archive proximity.

September 5, 2026 — A single-family-office quant route and an AI-forward MFin syllabus

• Christoph Frey’s public academic/professional homepage says he is a quantitative researcher for a single-family office in Hamburg and a research fellow at Lancaster University’s Centre for Financial Econometrics, Asset Markets and Macroeconomic Policy. It lists portfolio and multi-asset allocation, Bayesian statistics, financial econometrics, forecasting, predictive models, machine learning, AI, neural networks and big data among his interests, and identifies Winfried Pohlmeier as thesis advisor and Gary Koop as thesis referee. The Tidy Finance project and publisher biography connect Frey to a reproducible Python finance text covering CRSP/Compustat/TRACE preparation, factor selection via ML, option pricing via ML, portfolio optimisation and backtesting. This is unusually direct family-office-to-academic and public-code evidence; the office is unnamed and the sources do not disclose its models, data rights, permissions, trading decisions or results.
• Carleton University’s FINA 5523 Financial Analytics outline, taught by Dr Yuriy Zabolotnyuk in the 2026 MFin summer term, states that students use Python and AI-assisted coding tools for equity, fixed-income and fundamental data, risk measurement, factor models, portfolio optimisation, backtesting and ML return prediction. The schedule names Cursor, Claude Code, GitHub Copilot, pandas, FRED API, scikit-learn and PyPortfolioOpt; its GenAI policy asks students to use tools as assistants, verify outputs, apply judgement and cite AI use when relevant. This is a concrete curriculum and toolchain signal in a finance programme, not evidence that students or the university authorise live trading or that any named employer uses the course stack.

Research implication: this adds two important academic discovery routes. Public code and teaching materials can reveal the reproducible data plumbing and validation vocabulary surrounding ML finance, while a named family-office practitioner can expose a talent and research lineage without identifying the office. The useful next step is to recover Frey’s papers/software and the Carleton project outputs, then separately verify any alumni or employer links.

Recovery: archive the Tidy Finance Python source, changelog, code licence and Frey research/software pages; recover the Carleton course projects, presentation materials and later course outlines; verify Pohlmeier, Koop and Zabolotnyuk through their institutional profiles and papers; and keep course exposure, public code, family-office employment and production deployment as separate evidence states.

September 5, 2026 — Canadian professor-led risk, portfolio, and AI programme routes

The University of Ottawa Telfer profile for Jonathan Yu-Meng Li adds a more specific academic-industry route than a generic AI-in-finance course. Li is an Associate Professor and RBC Financial Group Professor in Financial Risk Analytics. The profile lists a PhD in Operations Research with a Financial Engineering specialisation from the University of Toronto, and describes work spanning financial econometrics, optimisation, foundational models, large language models, and deep reinforcement learning for portfolio management, asset pricing, fraud detection, and regulatory compliance. It also lists papers on deep-RL option pricing and hedging under dynamic risk measures, and permutation-invariant deep-RL policy networks for portfolio management.

The same profile names two Mitacs-funded industry routes: a 2020–23 Brane Capital project on a deep risk-sensitive reinforcement-learning framework for portfolio management, and a 2019–20 EVOVEST project on portfolio management by reinforcement learning. These entries verify funded research relationships and research topics. They do not establish that either company adopted the models, that proprietary data were used, or that any result transferred to a live book. The current profile also lists a 2025–28 SSHRC project on foundational models for unified AI-driven financial regulation, which is a governance and risk-intelligence route rather than evidence of an investment system.

Tolga Cenesizoglu’s HEC Montréal profile is worth recording at greater depth. It identifies him as a full Professor of Finance, Director of the Canadian Derivatives Institute, and co-responsible for the BNI–HEC Montréal Fund, described as a student-run investment portfolio with more than $7 million in assets under management. The profile also lists CFO responsibilities at Delta Vega Financial, a structured-product risk and valuation firm, and a senior-director role in KPMG financial-risk management advising Canadian pension funds and financial institutions on portfolio and risk-model validation. His stated research fields include asset pricing, market microstructure, financial econometrics, AI, machine learning and big data; his training is a UCSD PhD/MA in Economics and MSc in Statistics after a Boğaziçi industrial-engineering degree.

This is a useful professor–programme–student-capital interface: portfolio management, model validation, derivatives, microstructure, student investing and AI/ML appear in one public biography. It is still not evidence that the BNI–HEC portfolio uses AI, that Delta Vega or KPMG clients share data with HEC, or that any listed method has trading authority. HEC’s David Ardia profile supplies a complementary reproducibility and text route, including a 2026 paper on optimal text-based time-series indices, a 2026 paper revisiting a published asset-pricing result, sentometrics, and supervised student projects on earnings reactions, NLP sentiment, volatility, option returns and hedge-fund model misspecification.

The idea queue should therefore separate three testable surfaces: risk-sensitive policy learning for portfolio decisions; text-derived time-series indices with timestamp and revision controls; and model-validation methods for nonlinear or misspecified exposures. The public pages support those research hypotheses and personnel lineages. They do not support a conclusion about any tracked manager’s relative capability, deployment stage, data access, or performance.

Recovery: retrieve Li’s thesis, code, appendices, Mitacs project reports and any public Brane/EVOVEST presentations; archive the BNI–HEC governance and portfolio materials; recover Cenesizoglu’s academic and professional CVs; and obtain Ardia’s 2026 papers, replication files, student theses and seminar recordings. Keep research funding, teaching, student capital, consulting, former employment, proprietary access, model ownership and live deployment as separate ledger states.

September 5, 2026 — Professors of practice as family-office and allocator discovery routes

• Marquette’s profile for David Bauer identifies him as an Adjunct Professor of Finance and, since 2005, a partner and Chief Investment Officer at Lubar & Co, a Milwaukee single-family office. The page says he teaches Corporate Finance in the Executive MBA and Private Equity to undergraduates, and describes direct investing, private equity, acquisitions, valuation, boards and portfolio-company work. This is a direct school-to-family-office operating link, but the page does not disclose Lubar’s AI tools, data, portfolio process or returns.
• Wharton Executive Education’s profile for Mike T. Kane identifies him as founder and president of Kestrel Investments, a Philadelphia-area single-family office investing in venture capital and private equity. It also says he teaches philosophy and economics at Penn and teaches Wharton’s Private Wealth Management programme, which has included high-net-worth investors from more than 50 countries. The profile exposes a family-office governance, private-market and executive-education route; it does not establish a model, data stack, AI use or investment authority beyond the stated biography.
• Rice Business’s profile for Marc Sharpe identifies him as a Lecturer in Finance, founder and chairman of The Family Office Association, and an adjunct professor teaching “The Entrepreneurial Family Office” at Rice and SMU. The page says the association has connected senior family-office executives and principals since 2007 and promotes education, shared practices and co-investment opportunities; it records Cambridge, Oxford and Harvard Business School training. This is a network and curriculum surface that can lead to speakers, programme materials and governance vocabulary, not evidence of any member office’s AI or investment system.
• Chicago Booth’s profile for Priya Parrish identifies her as Partner and CIO at Impact Engine, Adjunct Associate Professor of Strategy and Impact Investor in Residence. It records her prior CIO role at Schwartz Capital Group, described as a global-markets single-family office, and prior strategy work at Aurora Investment Management plus ESG product development at Northern Trust and KLD. Booth lists Impact Investing in her 2027 course schedule. This is a named allocator, family-office, hedge-fund and academic lineage route; it does not show AI tooling, data access, model ownership or performance.
• UVA Darden’s profile for Joe Andrasko identifies him as a Professor of Practice in Data Analytics and Decision Sciences and Finance, teaching Decision Analysis, Applied Security Analysis, a master seminar and Data Visualization and Analytics. The page also identifies him as managing partner and CIO of Fry’s Path Capital and records his prior role as director of investments at Murray Enterprises, a single-family office where he oversaw outside funds and direct investments in healthcare, technology and real estate. The profile is useful for manager-selection, asset-allocation and family-office practice discovery; it does not establish AI use or a production system.

Research implication: academic discovery should include professors of practice, executive-education faculty, adjuncts and course instructors—not only ML-labelled professors. These profiles expose how family offices teach governance, private markets, manager selection, investment judgement and operating experience. They are personnel and programme interfaces, not evidence that a family office or fund uses the methods taught by the person.

Recovery: retrieve course syllabi, guest-speaker lists, family-office association programmes, public talks, papers and dated employer records; map the people to the existing family-office and quantitative-finance ledger; and preserve the distinction between current role, prior role, teaching relationship, network membership, academic research and verified AI deployment.

September 5, 2026 — Queensland professor, student capital, and AI-portfolio lineage

QUT’s official profile for Professor Anup Basu adds a distinct Australia-Pacific academic route. QUT identifies Basu as a Professor of Finance and Behavioural/Experimental Economics, founder of the QUT Student Managed Investment Fund, and visiting faculty fellow at MIT Sloan. His teaching list includes Student Managed Investment Fund 1 and 2, Security Analysis and Portfolio Management, Behavioural Finance, and Data Analysis and Decision Making. The profile lists a completed doctoral supervision titled “Portfolio Selection Using Artificial Intelligence” from 2014, alongside work on investment strategies, performance evaluation, pension finance, behavioural decision-making and data analysis.

QUT’s School of Economics and Finance page says the student fund was established in 2018 to give students hands-on experience managing a real investment portfolio, and that the Bloomberg Lab houses the fund’s technology. A separate QUT fund page says the fund is financed by QUT Business School and external donations, and exposes students to investment advisers, professional fund managers and other practitioners. QUT’s 2025 annual report separately describes the fund trust as established in 2019; those dates are retained as different institutional statements rather than reconciled into one history.

This route is useful for finding AI-related portfolio-selection theses, student models, Bloomberg workflows, faculty-supervised fund artefacts and practitioner speakers. It does not establish that the QUT fund uses AI, that Basu’s older supervision is current research, or that QUT, MIT Sloan, Bloomberg or any external donor shares proprietary data or runs a live institutional strategy. No comparison or relative assessment is implied.

Recovery: obtain the “Portfolio Selection Using Artificial Intelligence” thesis, QUT SMIF investment-policy and governance documents, dated portfolio reports, Bloomberg Lab specifications, student presentations, adviser/speaker rosters and alumni transitions. Keep the 2018 profile statement and 2019 annual-report trust statement separately timestamped.

September 5, 2026 — Laurier MFin links ML coursework, student capital, and employer pathways

• Wilfrid Laurier’s Master of Finance programme page lists MF791 Machine Learning in Finance as an elective and describes it as an end-to-end applied course using Python for financial analysis. The same programme page says the MFin includes Bloomberg, financial-database, Excel/VBA and Python training, and has a faculty-supervised Laurier Graduate Student Investment Fund in which students can work as portfolio managers, equity analysts, risk analysts or macro analysts managing real money in live markets. Laurier also lists a methods field that includes MF791 and a co-op stream with named financial-sector employers, including BMO Capital Markets, CPP Investments, Fidelity, OMERS, Ontario Teachers’ and RBC. These are programme-level claims and employer examples, not evidence that the student fund uses ML, that any employer recruits from a particular project, or that any model has live authority or attributable performance.

Research implication: Laurier offers a compact route for recovering the handoff from finance education to investment artifacts: course exercises and projects, a supervised fund, co-op placements, and named employer pathways. It belongs in the academic/talent graph, not in the firm-deployment layer.

Recovery: retrieve MF791’s syllabus, assignments, project briefs and instructor profile; obtain LGSIF governance, holdings and student research where public; recover seminar and speaker pages; and compare named co-op employers with independently verified personnel and job records without inferring adoption from programme proximity.

September 5, 2026 — Boston-area student capital and quantitative-finance programme

Babson’s official College Fund page adds a Boston-area student-capital route with unusually visible operating detail. Babson says selected undergraduate and graduate students manage $9 million of the college endowment in a long/short equity portfolio, organised into sector teams and supported by a faculty director, executives in residence and industry professionals. The page describes fundamental and quantitative analysis, portfolio construction, tracking-error and risk-exposure monitoring, and a two-semester continuity requirement. Its Fall 2026 cohort section exposes current co-executive portfolio-manager and sector-team roles, creating a time-bounded personnel and recruiting surface.

The Babson MBA quantitative-finance concentration places the fund beside financial modelling, forecasting, machine-learning methods for business, derivatives, fixed income, portfolio management and institutional trading-strategy coursework. The page identifies Jasmina Hasanhodzic as a faculty member with prior AlphaSimplex Group and Credit Suisse experience, while the FIN7572 catalogue entry specifies that the fund course selects analysts and portfolio managers and uses Cutler Center analytical tools. These are programme and biography disclosures; they do not establish the fund’s AI use, the student portfolio’s production equivalence to a hedge fund, proprietary data access, or model ownership.

The page publishes calendar-year return and contributor tables. Those are retained as publisher-reported records for later reconciliation, not treated here as an evaluation or comparison. The research value is the observable workflow: sector-team idea generation, analyst-to-portfolio-manager progression, risk and tracking-error language, industry mentorship, and a public recruiting surface in the Boston finance ecosystem.

Recovery: archive the current student-manager roster, faculty director and executive-in-residence biographies, fund policy, holdings/transaction reports, Cutler Center tool descriptions, course syllabi and public presentations. Preserve the page’s $9 million description and catalogue/programme figures as separate dated claims, and verify any alumni transition independently.

September 5, 2026 — Bristol hedge-fund behaviour research and Strathclyde AI-finance capacity

• The University of Bristol profile for Dr Xinyu Cui identifies him as a Lecturer in Finance with a PhD from Alliance Manchester Business School. His listed expertise includes empirical asset pricing, hedge funds, institutional investors, market anomalies, toxic emissions and corporate decision-making. The profile says his recent research examines institutional-investor trading behaviour, including mutual funds and hedge funds, and market effects.
• Bristol lists Cui’s 2025 Journal of Corporate Finance article “Do Hedge Funds Still Manipulate Stock Prices?” with Olga Kolokolova and his 2024 Management Science article “On the Other Side of Hedge Fund Equity Trades.” The profile also records a common-ownership doctoral supervision with Kirak Kim and Neslihan Ozkan. These are academic personnel, publication and supervision links. They do not establish AI use, a fund affiliation, proprietary data access, deployment or performance.
• Strathclyde’s Professor of Financial Technology (AI) job brief describes a proposed professorial role funded through the UKRI Global Talent Fund. The brief places it in Strathclyde’s FinTech Cluster and links the remit to the Financial Regulation Innovation Lab, a partnership involving Strathclyde, FinTech Scotland and the University of Glasgow. It names multimodal generative AI, explainable AI, agentic AI and earth intelligence as areas in the surrounding research ecosystem, and asks the appointee to lead research, teaching, doctoral supervision, external funding and industry/regulator partnerships. This is a recruitment specification, not confirmation of an appointment or evidence of a working system.

Research implication: the academic layer now includes a title-blind route into institutional trading behaviour and a separate institutional-capacity route into multimodal, explainable and agentic financial technology. Recover the papers, code, datasets, seminars, appointment outcome and partner records before linking either route to a firm’s AI or GenAI activity.

Recovery: archive Cui’s papers, appendices, data and conference versions; capture his CV, seminars and doctoral/coauthor network; verify Strathclyde’s appointment and current FinTech/FRIL roster; and retrieve funded projects, doctoral work, code, regulator engagement and partner disclosures. Preserve the boundaries between academic research, proposed capacity, prototype, pilot and production deployment.

September 5, 2026 — Boston-to-Vanderbilt finance professor with an earlier ML research lineage

Adonis Antoniades’ current public site identifies him as an Associate Professor of the Practice of Finance at Vanderbilt’s Owen Graduate School of Business and records his move from Northeastern’s D’Amore-McKim School of Business. His public CV traces a mixed engineering, machine-learning, finance and policy path: a Berkeley MSc in Electrical Engineering and Computer Sciences, Columbia economics PhD training, earlier NICTA research in artificial intelligence, finance appointments at NUS and Northeastern, and economist roles at the ECB and BIS.

The CV lists early papers on machine learning for adversarial-agent microworlds and pursuit–evasion with multiple agents under incomplete information. It also records finance-sector student projects involving a discretionary strategy, portfolio selection and backtesting, a robo-adviser for a global insurer, and an investment competition. These artifacts expose a personnel lineage from engineering and agent research into finance education and policy analysis, but the CV is dated and does not establish a current employer’s model, data rights, investment authority, or live deployment. The current Vanderbilt page should be treated as the current role source; the Northeastern CV is historical evidence.

Recovery: archive the current Vanderbilt faculty page, updated CV, papers, teaching materials and any public finance/AI seminars; verify the Northeastern, NUS, ECB, BIS and NICTA dates independently; and keep historical teaching projects, academic research, policy work and verified manager deployment as separate evidence states.

September 5, 2026 — UMass Boston’s finance-data curriculum and historical Bloomberg pipeline

• UMass Boston’s current Accounting & Finance department page says the department trains students in financial and market-data analytics in the context of AI and Big Data. It also says the finance PhD curriculum has been updated with data-science classes in natural-language processing, machine learning and textual analysis, alongside access to a student-managed fund, Bloomberg certification and industry-speaker events. This is current programme-level evidence of methods and infrastructure; it does not establish a hedge-fund relationship, proprietary data access, model ownership or live investment deployment.
• The current MS in Finance page places financial-statement and market-data modelling, risk analysis, Bloomberg training and the student-managed fund in a Boston financial-industry learning pathway. The page’s first-person student account is useful as a recruiting and workflow lead, but it is not independent evidence of a fund system or employer practice.
• UMass Boston’s 2015 Chancellor’s report records more than $200,000 of assets under management in the College of Management Student Managed Fund, a public-private partnership, Bloomberg data certifications and interactions with Boston financial-services decision-makers. A 2021 College of Management report names Brian Walker, Valentino Palmieri and Laura Rodriguez as student organisers of a Bloomberg Terminal competition, with 19 teams from UMass Boston, Clark and Bryant; it describes remote Bloomberg Lab access and seven weeks of $1 million U.S.-equity competition allocations. These are dated educational records, not current AUM or proof of an AI strategy. The 2021 allocation should be read as a competition parameter, not as a claim about deployed capital.

Research implication: this route adds a Boston-area title-blind search surface where finance data infrastructure, textual-analysis training, student capital and local industry contact coexist. The relevant follow-up is to recover the current finance PhD course list, faculty and student-fund governance, Bloomberg Lab history, competition materials, student research and alumni employment records. The public pages support talent and method discovery, not an inference about any tracked manager’s capabilities or relative position.

Recovery: obtain UMass Boston finance faculty CVs, course syllabi, dissertation titles, student-fund reports, Bloomberg certifications, competition results, speaker rosters and alumni paths; verify which 2015–2021 facilities and activities remain current; and keep programme exposure, student competition, employer contact, proprietary access, model ownership and production deployment as separate evidence states.

September 5, 2026 — Boston College’s current quantitative-finance programme route

• Boston College’s current M.S. in Finance page describes a STEM-designated quantitative-finance track with Financial Econometrics, Quantitative Portfolio Management, Derivatives & Risk Analytics, Data Analytics in Finance, and FinTech and Cryptocurrencies among its 2027 course options. The same page includes applied-fundamental-analysis, fixed-income and cross-asset-valuation options, creating a current Boston-area curriculum surface that spans statistical modelling, portfolio construction, derivatives, risk, market data and digital assets.
• The programme page also exposes a practical talent route through a 12-month graduate format, quantitative course sequencing and admissions for a January 2027 quantitative-track start. These are programme and recruiting claims, not evidence that Boston College operates an AI lab, that student work reaches a named hedge fund, or that the listed methods have live trading authority.

Research implication: the Boston academic map now includes a complementary programme whose public language is more specific about the handoff from econometrics and portfolio construction to data analytics, risk and FinTech. Recover the course syllabi, instructors, project briefs, employer or corporate-project partners, student-fund links, alumni placements and any public research outputs before connecting it to a manager or model.

Recovery: archive the 2027 curriculum and course catalogue, verify MFIN8870 Data Analytics in Finance and the quantitative-track electives, recover faculty CVs and project sponsors, and keep curriculum exposure, student work, employer recruitment, proprietary data, model ownership and deployment as separate evidence states.

September 5, 2026 — Miami student capital, option ML, and Indian finance-AI research interfaces

• Paul Borochin’s public research and teaching site identifies him as a financial economist and data scientist with a Duke Fuqua finance PhD and Wharton finance/statistics training. It links work on public-market information, equity and option performance, corporate events, institutional ownership and financial applications of ML. His implied-volatility research describes tree ensembles, nonlinear interactions and delta-hedged option-return tests, while the site also exposes an ML/NLP-in-finance conference, Bloomberg portfolio tutorials and University of Miami Student Managed Investment Fund courses. These are first-party academic, data and teaching disclosures; they do not establish fund deployment or independently validate the displayed performance claims.
• Miami University’s Farmer School Student Managed Investment Fund, also called Red Brick Capital, describes more than $1 million of university-endowment capital, FIN 481 analyst and FIN 482 portfolio-manager courses, sector teams, alumni advisory bodies and reporting to University Treasury Services. Students conduct quantitative and qualitative research, vote on recommendations, perform attribution analysis, and submit orders through faculty and Treasury approval. The page names Dr. Xi Liu as a faculty-adviser contact and publishes sector reports. This is a detailed student-capital governance route; it does not disclose AI use, model code, data licences or professional-manager adoption.
• IIT Madras’s August 2026 announcement with Computer Age Management Services says the CAMS IIT Madras FinTech Innovation Lab, established in 2022, is entering a renewed phase focused on AI in financial services, payments, tokenisation, cybersecurity, RegTech, WealthTech, InsurTech, lending, analytics and financial inclusion. It names Anuj Kumar, V. Kamakoti and M. Thenmozhi, and describes research, proofs of concept, technology transfer, internships, hackathons, startup mentoring and professional certificates. The release also describes CAMS as an asset-management-industry infrastructure provider and notes its majority stake in Think360.ai. These are institutional and company-reported partnership claims, not evidence of a specific model, dataset permission, hedge-fund use or production outcome.
• IIT Bombay CFILT’s significant-projects page lists State Bank of India Foundation projects titled “Hallucination Detection and Mitigation on Finance LLMs” and “Table to Text Generation Maintaining Subjectivity LLMs.” Adjacent projects include time-series prediction and knowledge graphs, speech-to-speech translation for Indian languages, and explainability grants. The separate SBI Foundation/C-MInDS hub required JavaScript and was not directly readable in this pass, so the static CFILT page is the retained evidence. The titles expose concrete research and regional-language directions, but not model weights, datasets, results, ownership or deployment.

Research implication: add option-implied volatility forecasting, student-fund governance, asset-management infrastructure partnerships, finance-LLM hallucination evaluation, subjectivity-preserving table-to-text generation, time-series/knowledge-graph integration and Indian-language speech translation to the research queue. Keep academic, education, partnership and deployment evidence separate.

Recovery: capture Borochin’s linked papers, conference recordings, syllabi, datasets and code; obtain Miami’s current Red Brick Capital reports, sector analyses and faculty/adviser roster; recover CIFIL project lists, MoU terms, PI names, funded outputs and Think360.ai work; and use browser capture to recover the SBI hub’s project metadata without treating the JavaScript failure as absence of content.

September 5, 2026 — Darden faculty, CMU quant talent, and IIT Bombay–Optiver research scope

• UVA Darden’s profile for Shuaiyu Chen lists financial markets, asset pricing, hedge funds and mutual funds, retail investors and social media, and machine learning in finance as areas of expertise. It records a 2025 move from Purdue, a Rochester finance PhD, a Johns Hopkins applied-mathematics-and-statistics master’s degree, and MBA teaching in quantitative portfolio management. This is a faculty, paper and curriculum route; it does not establish a hedge-fund relationship, proprietary data access, model ownership or deployment.
• CMU MSCF’s profile for Saagar Shah says he led fixed-income research in a Stevens student-managed fund, built momentum and ETF-based bond signals, integrated them into an XGBoost RFQ-bidding algorithm at SumRidge Partners, and built what he describes as a fully automated intraday trading system using ML forecasts of return and volatility. The profile also names a Kalshi statistical-arbitrage project. These are self-reported talent-profile claims, not an audit of the system, strategy, permissions, performance or CMU involvement.
• The IIT Bombay–Optiver AI Innovation Lab page describes a 2,000-square-foot facility with 15 workstations, GPU/server capacity, Bloomberg and Refinitiv Eikon, real-time NSE Data & Analytics feeds, mutual-fund and ESG ratings, and statistical/computing software. Its stated research themes include sparse and irregular data, tabular/time-series foundation models, online learning and drift, graph-stream analytics, multimodal text/time-series/image fusion, metadata-driven feature discovery and automated data-quality checks. The page also describes PhD/postdoctoral routes and a ₹10–15 lakh average proposal budget. These are lab-scope and infrastructure disclosures, not project results, researcher data rights, Optiver employment for academic staff, production use or performance evidence.

Research implication: add bond/RFQ signal engineering, return-and-volatility forecasting, sparse-event learning, tabular/time-series foundation models, online drift, graph streams, multimodal fusion and data-quality intelligence to the academic research queue. Keep Darden faculty research, a student’s self-report and sponsored-lab scope as separate evidence classes.

Recovery: retrieve Chen’s research page, papers, CV and seminars; independently verify Shah’s SumRidge dates, code and competition artifacts; and archive the IIT proposal call, named project awards, principal investigators, fellowship outcomes, symposium recordings and publication rules.

September 5, 2026 — NYU professor and MIT recruiting route into applied hedge-fund replication

• MIT’s Career Advising & Professional Development board carries a September 2, 2026 research-analyst internship posting from Unlimited Funds. The posting describes the firm as founded by Bob Elliott, Bruce McNevin and Matt Salzberg, says it manages four listed ETF products with approximately $250 million in assets under management, and connects machine learning and data science to replication of alternative-investment return characteristics. This is a current university-hosted recruiting disclosure and a firm-provided description; the AUM and product claims are not independently audited here.
• Unlimited’s current team page identifies McNevin as co-founder and Chief Data Scientist, a professor of economics at NYU, and a former data-science leader at Bank of America, Midway Group, Clinton Group and BlackRock. Unlimited’s technology page says its models use index-return data to infer and replicate real-time positioning of alternative managers, with cyclical, tactical and structural components. This exposes a concrete model/data surface and a professor-to-industry route, but not the training sample, feature construction, permissions, model weights, replication error, or live decision authority.

Research implication: university job boards can surface current investment-firm personnel, capital scale, recruiting needs and model descriptions that are absent from academic course pages. The research object here is return-based inference and portfolio replication, not a claim about any underlying manager’s process. Keep firm marketing, university recruiting, professor employment, public product filings and independently verified performance as separate evidence classes.

Recovery: obtain the internship’s closing status, full role text and named team; verify McNevin’s NYU faculty page and employment dates; retrieve Unlimited prospectuses, regulatory filings, model-risk disclosures, product holdings, methodology notes and public talks; and compare the stated return-based approach with academic replication literature without ranking firms or inferring proprietary manager positions.

September 5, 2026 — Hanken, RiskLab Finland, and Peter Sarlin’s finance-AI bridge

• Hanken’s official appointment notice records Peter Sarlin’s 2017 appointment as Professor of Practice, with machine learning and AI applications in financial, economic and societal disciplines. The dated notice describes a bridge role between Hanken researchers and financial-sector organisations, and records prior consulting work for the IMF, ECB and several European central banks, plus his founder roles at RiskLab Finland and AI-focused ventures and his then-current executive/chief-scientist role at Silo AI. It is historical evidence of the appointment and remit, not proof of a current Hanken title or any investment-firm deployment.
• Hanken’s 2018 International Alumni Day programme records Sarlin’s “Machine and Human Intelligence in Finance” talk and identifies him as Professor of Practice at Hanken and Executive Chairman and Chief Scientist of Silo.AI. The Strategic Growth Investing syllabus adds a more concrete programme surface: a Sarlin AI guest lecture, a knowledge-graph-with-ESG-data session, and an applied hackathon. Those are dated educational and event artifacts; they do not disclose private participants, models, data permissions or production use.
• Hanken’s Digging into High Frequency Data project notice identifies Sarlin as the Finnish project lead for interpretable ML and exploratory analysis of large, high-frequency financial data, including visual dynamic clustering. It lists PIs at EUROFIDAI, Berkeley, Goethe, UMass Amherst, UCL and LSE. A related RiskLab paper describes CrisisModeler as an interactive financial-crisis prediction and model-evaluation framework applied to European-bank data. These sources expose research design, data modality, cross-school lineage and evaluation tooling, not a hedge fund’s live system or investable performance.

Research implication: this route adds a finance-specific academic bridge spanning systemic-risk modelling, high-frequency data, interpretable ML, knowledge graphs, ESG, AI venture building and practitioner-facing teaching. It also creates a conference and syllabus recovery path: follow the named PIs, guest lecturers, RiskLab projects, course outputs and event recordings while preserving appointment dates and separating research, consulting, vendor work and investment deployment.

Recovery: verify Sarlin’s current institutional and company roles; retrieve RiskLab Finland project archives, code, datasets, model documentation and papers; recover the named PIs’ profiles and cross-collaborations; obtain the Hanken course materials and event recordings; and keep dated academic, consulting, company, prototype, pilot and production evidence separate.

September 5, 2026 — Columbia practitioner-professors with quantitative, AI, and allocator links

• Columbia’s profile for Satyajit Bose identifies him as a Professor of Practice and Principal Investigator in the university’s sustainability-management programme. The profile says he teaches sustainable investing, cost-benefit analysis and mathematics; it also records prior investment-banking, asset-management and financial-restructuring work, including directing quantitative trading strategies at a convertible-arbitrage hedge fund that the page describes as managing $1.5 billion. The same biography says he developed machine-learning algorithms for automated weather-risk decision tools. The page does not name the hedge fund, disclose the trading models or data, or establish that the weather-risk work transferred to a live investment process.
• Columbia Business School’s profile for Michael Weinberg identifies him as an Adjunct Professor of Business in the Finance Division and says he teaches the Institutional Investing course he created. It records a current Managing Director, Head of Hedge Funds and Alternative Alpha, and Investment Committee role at APG; prior CIO roles at MOV37 and Protege Partners; prior portfolio-management roles at FRM, Soros and Credit Suisse; and a special-advisor role to The Tokyo University of Science’s Endowment. The profile also records his former co-founding role at The Artificial Intelligence in Finance Institute and links him to an NLP-for-SDGs paper with George Mussalli and Amir Amel-Zadeh. These are current/prior personnel and publication links, not evidence that any employer adopted a named model or that the academic paper describes a deployable trading signal.

Research implication: professor-of-practice pages can expose a different bridge than finance curricula: named investment mandates, prior hedge-fund strategy roles, allocator governance, AI-in-finance institution building and co-authored text-analysis research. Bose’s route is a quantitative-strategy and weather-risk/ML lead; Weinberg’s route is an allocator, hedge-fund-selection, AI-governance and corporate-text-analysis lead. The source pages support those research queues but do not support ranking firms or inferring deployment.

Recovery: retrieve Bose’s CV, papers, weather-risk project records and dated fund employment; retrieve Weinberg’s current APG disclosure, course materials, AI-in-finance institute archive, NLP-for-SDGs paper and code/data provenance; verify Mussalli and Amel-Zadeh independently; and keep academic role, prior employment, current employer, advisory activity, public research and production deployment separate.

September 5, 2026 — George Mason professor, student capital, and wealth-management curriculum

• George Mason’s profile for Derek Horstmeyer identifies him as an Instructional Professor of Finance, co-founder and director of the GMU Student Managed Investment Fund, and director of the Financial Planning & Wealth Management concentration. The profile describes research interests in hedge-fund activism, mutual-fund and ETF performance, boards, corporate governance, mergers and acquisitions, and executive compensation. It also links a public media trail including the Wall Street Journal, ABC News, Fortune, and the Zacks “ETF Spotlight” podcast. This is a professor, student-capital and media-discovery route; it does not disclose AI use or a live systematic investment process.
• The same first-party profile gives a three-stage quantitative lineage: a University of Chicago mathematics/economics bachelor’s degree, a Stanford financial-mathematics master’s degree, and a USC finance and business-economics PhD. It lists research and conference outputs on dissident directors, governance and investment transparency. The education and publication trail is useful for mapping training and topic formation, but it does not establish that the student fund or any employer uses the methods, data or software associated with those topics.

Research implication: this route adds a useful control surface to the academic map: an explicitly investment-oriented student fund and a wealth-management programme alongside hedge-fund-activism research, but without a public AI claim. The media clipping list is also a title-blind expansion queue for finance podcasts and appearances that may not mention AI or hedge funds in episode titles.

Recovery: capture the current GMU fund site, mandate, governance documents, holdings or performance reports, student-manager roster, course syllabi, software and adviser biographies; recover the Zacks episode and other media transcripts; and preserve professor role, student-fund activity, research topic, media appearance and verified manager deployment as separate states.

September 5, 2026 — Chicago Booth family-office programme as a confidential allocator discovery route

• Chicago Booth’s “The Future of Your Family Office” programme lists a November 2–6, 2026 session for senior leaders of single-family offices with more than $250 million in assets under management. The public page lists a $20,000 fee, a September 21 application deadline, selective admission and a confidential, solicitation-free environment. It identifies John C. Heaton as programme co-director and Family Office Initiative director, Bobby Stover as Executive Director of the initiative, and a wider faculty/practitioner group including Stuart Lucas, Stephan Roche, Sharon Schneider, Rick White and Quan Mac.
• The programme outline exposes operating questions that are relevant to the family-office search: which functions remain in-house, which are outsourced or handled through external partnerships, how ownership and governance are structured, and how accountability and performance management are designed. Participants complete a proprietary questionnaire across nine strategic domains and develop a plan through workshops and a capstone review. This is an executive-education and allocator-network signal, not evidence about any attendee’s AI stack, vendor contracts, data access, investment decisions or portfolio returns.

Research implication: a confidential programme page will not identify participants, but it gives a precise discovery route for family-office principals, operating executives, faculty, practitioners and future public speakers. Its in-house-versus-outsourced framing is a useful search vocabulary for AI automation, data operations, reporting and governance without assuming that any family office has automated those functions.

Recovery: archive the programme brochure, faculty biographies, public talks, alumni or participant material where lawfully available, and any later disclosures from named practitioners. Do not attempt to infer identities from the confidential cohort; keep programme attendance, public employment, network membership, vendor relationship and verified AI deployment separate.

September 5, 2026 — Simon Fraser’s SIAS and BEAM student-capital operating routes

• SFU Beedie’s SIAS page describes a graduate-finance programme that teaches asset pricing and data analytics alongside a student-managed portfolio with a published market value above CAD 30 million. SFU identifies HSBC Canada and the Lohn Foundation as donors, says performance is compared with benchmarks and reviewed quarterly by an industry-expert committee, and separates the operating teams into Canadian equity, global equity, fixed income, cash, economics, compliance and risk metrics. The page describes broker coordination and approved-transaction execution, but it does not disclose AI use, model code, data licences or professional-fund adoption.
• SFU’s BEAM page describes a separate undergraduate endowment programme with CAD 12.5 million invested as of December 31, 2025 across Canadian equity, global equity, fixed income and cash. It names Deniz Anginer and Geoffrey Poitras as faculty advisers, Manolo Pineda as Operations Portfolio Manager and Kai Nicholson-Karp as Investments Portfolio Manager, and says students handle research, asset allocation, security selection, trading and risk management under faculty/staff oversight. Bloomberg certification, quarterly reporting and daily compliance are public workflow signals, not evidence of a production AI system.

Research implication: SFU exposes two linked but distinct talent surfaces—graduate SIAS and undergraduate BEAM—with real endowment capital, explicit team decomposition, benchmark review, compliance, risk metrics, named advisers and current student leadership. That makes it useful for tracing how research and operational roles are taught before looking for alumni, employer or software links. The published portfolio values are institutional claims at stated dates and should not be treated as a performance comparison.

Recovery: archive the SIAS investment-team and reports pages, BEAM participant and reports pages, investment-policy documents, software/vendor references, student presentations, adviser biographies and alumni destinations. Preserve SIAS and BEAM as separate funds, and keep curriculum exposure, student capital, donor support, employer contact, proprietary access, model ownership and verified deployment separate.

September 5, 2026 — Peking Financial Engineering Lab’s dated practitioner lecture cluster

The Peking University Financial Engineering Lab teaching archive contains more than a generic course description. Its April 23, 2026 Optiver lecture identifies Yang Yaowei as Head of Technology and describes deep learning, time-series prediction, reinforcement learning, order-flow modelling and multi-asset engineering. The notice also says he was responsible for developing Optiver’s Shanghai AI Lab and places the lab’s Shanghai/New York build-out in 2025. These are Chinese-language event and personnel statements; they do not provide a recording, code, model weights, training corpus, permission map or trading results.

The same archive records a March–April 2026 sequence of practitioner lectures: BigQuant founder Liang Ju on an “Agentic Quant” workflow spanning data acquisition, factor mining, strategy development, backtest optimisation and risk control; Huatai Securities’ Shen Yang on LLMs for factor discovery, technical analysis, text-based stock-selection enhancement and thematic investing; Guosen Securities’ Wang Kai on multi-model, agent and multimodal research workflows; and Guosen’s Zhang Xinwei on extracting predictive information from high-frequency order and transaction data, including order size, duration, aggressive buy/sell behaviour and intraday timing. These lectures expose search terms, personnel routes and claimed workflow surfaces—not proof that any named organisation deployed the methods in a live fund.

Research implication: this is a useful title-blind conference/lecture source because conventional finance-event names carry concrete AI, agent, multimodal, high-frequency and data-engineering details that would be missed by searching only for “hedge fund,” “AI” or “GenAI” in episode titles. Recover each event’s detail page, slides, recording, transcript and speaker profile separately, and preserve employer claims, academic hosting, vendor marketing and verified production evidence as distinct states.

Recovery: capture the five lecture notices, linked detail pages and any video or social reposts; verify Yang Yaowei, Liang Ju, Shen Yang, Wang Kai and Zhang Xinwei through independent first-party profiles; and retrieve cited reports, datasets, model documentation and evaluation protocols without inferring deployment from a talk description.

September 5, 2026 — UMass Amherst’s MIT-to-alternative-investments research and data route

• Mila Getmansky Sherman’s UMass Amherst profile identifies her as Fuller and Meehan Endowed Professor of Finance and Director of the Center for International Securities and Derivatives Markets (CISDM). The profile records an MIT BS and MIT Sloan PhD, a post-doctoral fellowship at MIT’s Laboratory for Financial Engineering, quantitative research experience at Deutsche Asset Management, and a research portfolio spanning empirical asset pricing, hedge funds, investment-trading-strategy performance, financial institutions, systemic risk, ESG and system dynamics. It also lists current/recent work on government communication and presidential tweets, ETF-network dynamics, sustainable short selling and hedge-fund greenness. These are academic and career records; they do not establish a current investment mandate or an employer’s use of a specific model.
• CISDM’s first-party centre page exposes an unusually relevant data and network surface: the centre says its CISDM–Morningstar database tracks more than 12,000 hedge funds and CTAs, including graveyard records, with data tracking since 1994. It describes the centre as a bridge between academic and business communities and links to its indices, database, Journal of Alternative Investments and CAIA-related activity. This is evidence of a substantial alternative-investment research infrastructure and a potential historical-universe discovery route, not evidence that any tracked manager’s confidential positions, features or models are available.
• The UMass finance major page says students use Bloomberg terminals for real-time and historical data across securities and financial institutions, and participate in the Minuteman Equity, Fixed Income and Alternative Investments funds. The page describes fund administration, investment decisions, peer interviews, student-run security-selection courses, analyst reports and newsletters. It is a concrete finance-programme and talent-development surface, but it does not say that the funds use machine learning, generative AI or proprietary alternative-investment data.
• UMass’s 2021 CISDM conference programme is a valuable conference-recording recovery route. “The Future of Finance: From DeFi to AI” listed Joseph Simonian, founder and CIO of Autonomous Investment Technologies, for “Modular Machine Learning: The Best of Both Worlds?” and Sanjiv Das for “AI, FinTech: New Paradigms,” alongside Campbell Harvey and Fahad Saleh. The associated UMass conference notice links to the recording. The programme identifies public speakers and topics; it does not disclose attendee identities, proprietary client work or live trading results.
• Sherman’s UMass CV and the PRISM project description add a cross-domain modelling path. Her CV lists NSF support for “Digging into High Frequency Data” and PRISM, a project studying multi-layer dynamic interconnections and catastrophe risk with collaborators across finance, engineering, environmental science and data-intensive research. The public project description names power outages and natural disasters as target risks; a KDD workshop summary records the use of convergent data-intensive research and collaborators from Cornell, UMass Amherst, Tufts, Lincoln Park Zoo and industry. This exposes methods, collaborators and data modalities—not a hedge-fund alpha claim.

Research implication: this route joins three otherwise separate discovery surfaces: a finance professor with MIT financial-engineering lineage and prior buy-side quantitative research; an alternative-investment database with long historical coverage; and a finance programme with Bloomberg-enabled student funds. The most actionable follow-up is to recover CISDM database methodology and historical coverage, Sherman’s recent working-paper materials, the 2021 conference recording/transcript, Simonian’s modular-ML claims and the Minuteman fund’s public reports. Academic research topics such as hedge-fund liquidity, portfolio similarity, dynamic risk exposure, high-frequency returns, systemic connectedness, ETF networks and market impact of government communication should be treated as hypotheses for replication—not as evidence of any named firm’s live use.

Recovery: capture CISDM database/indices documentation, access conditions and historical schemas; preserve the conference recording, speaker biographies and transcripts; retrieve Minuteman fund reports, course syllabi, student rosters and Bloomberg/data-use documentation; map Sherman’s coauthors, doctoral committees and MIT/UMass lineage; and keep academic work, public database access, classroom activity, consulting, employer roles, vendor relationships and verified production deployment in separate evidence states.

September 5, 2026 — KAIST Financial Engineering Lab’s research-to-employer talent map

• KAIST Financial Engineering Lab’s home page identifies the lab as a research group in KAIST’s Department of Industrial and Systems Engineering and lists a 2024–2026 publication stream. The public list includes deep financial planning, decision-focused sparse tangent portfolio optimisation, limit-order-book representation, cross-asset order-flow prediction, transformer Hawkes processes, random-forest feature selection in asset management, GAN-based anomaly detection for portfolio optimisation, graph-based ETF forecasting and curriculum/imitation learning for financial time series. These are named papers and research topics, not evidence of a particular firm’s production system or performance.
• The lab’s team page identifies Professor Woo Chang Kim as a Princeton Operations Research and Financial Engineering PhD/M.A. with Seoul National University industrial-engineering degrees. It records his KAIST roles, a 2019–2023 directorship of the Shinhan–KAIST Artificial Intelligence Finance Research and Development Center, a 2016–2021 advisory-professor role at Samsung Asset Management, and earlier founder/executive-advisor work at DPT Capital Management. The page also records current students and postdoctoral researchers, including Haeun Jeon, Mingyu Yang, Seunghoon Choi, Gwanghyun Lee and Hwayong Choi. These dated affiliations should be kept separate from any claim that Samsung, Shinhan or DPT adopted a lab model.
• The lab’s AI-for-finance research page states that its deep-reinforcement-learning work targets sequential decisions under complex states. It explicitly describes training limit-order-book dynamics for high-frequency-trading policies and using deep reinforcement learning for constrained portfolio optimisation. Separate areas cover personalised life-cycle goal-based investing, uncertainty-aware investment management, robust optimisation, stochastic programming and dynamic programming. This gives a concrete research agenda and modality vocabulary, while leaving data access, simulator realism, execution authority, risk limits and out-of-sample results unresolved.
• The lab’s public alumni page exposes a longitudinal placement and thesis surface. It lists Guhyuk Chung’s 2024 PhD thesis on neural order-book modelling and market making followed by a quantitative-trader role at LINE Investment Technologies; Do-Gyun Kwon’s stochastic-programming thesis followed by a vice-president, quantitative-development role at WorldQuant; and Geum-il Bae’s regime-switching thesis followed by Korea Asset Management and later HuGraph CTO roles. These are page-reported destinations and dates, useful for lineage mapping but not proof that the destination firms used the thesis methods.
• The same alumni page lists a 2022 M.S. thesis on optimising high-frequency pairs trading with reinforcement learning and spread prediction; a 2021 M.S. thesis on enhancing hedge-fund-index tracking with deep reinforcement learning; a 2024 M.S. thesis on type-based limit-order-book events and deep neural-network mid-price prediction; and a 2023 M.S. thesis on an action-constrained deep Q-network for goal-based investment. It also lists work on politically themed stocks using text mining and entropy-based network dynamics, differentiable sorting, goal-based investing, index tracking and portfolio decision-focused learning. This is a useful map of research hypotheses and student specialisations, not a validated alpha catalogue.
• The lab’s current publication list shows the research moving across the full investment stack: feature selection and anomaly detection before portfolio construction; decision-focused learning that couples forecasts to portfolio objectives; order-flow and limit-order-book representations for market microstructure; graph and information-theoretic methods for ETF and cross-market relations; and stochastic or deep-learning methods for long-horizon financial planning. The source does not disclose training corpora, licensed feeds, simulator-to-live transition, model ownership, code availability or production monitoring.

Research implication: KAIST provides a particularly useful college-to-industry map because the public evidence joins a named professor, Princeton/Seoul training, a finance-AI centre, a stated research agenda, current graduate researchers and a multi-year alumni trail into quant development, quantitative trading, asset management and financial technology. The top follow-up questions are empirical: which thesis artefacts became public code or papers; which datasets and simulators were used; whether the alumni destinations involved the same methods; and how research outputs changed after industry placement. The evidence supports a research queue around order-flow representation, decision-focused optimisation, constrained reinforcement learning, financial planning, portfolio similarity, anomaly detection and financial-network information transfer. It does not support ranking KAIST or any destination firm.

Recovery: preserve the lab’s research, projects, courses, publication metadata and alumni pages as dated snapshots; retrieve linked papers, theses, code, datasets and conference presentations; verify current employment independently; map Woo Chang Kim’s students, coauthors and doctoral lineages; and distinguish university research, sponsored centre activity, advisory work, alumni employment, public code, pilot activity and verified production deployment.

September 5, 2026 — Morgan State and RIT: finance programmes as AI and talent surfaces

• Morgan State’s profile for George Micheni identifies him as Director of the Capital Markets Lab. The university says the lab’s U.S. capital-markets programmes use Bloomberg for data analysis and market insight, and that Micheni has built fintech and investment programmes, corporate partnerships, machine-learning research and a Student Managed Investment Fund. The profile also describes a summer 2023 project in which his team used AI to forecast prices for AAPL, MSFT and MU, and lists a research interest in combining machine learning with time-series analysis for credit-card-default prediction. This is first-party university evidence about a professor, lab, educational capital and described projects; it does not disclose code, data rights, validation design, model ownership, fund authority or attributable performance.
• The Morgan route is useful because the AI signal is embedded in a Capital Markets Lab and faculty biography rather than a hedge-fund or generative-AI page. The relevant follow-up is to recover the lab’s project briefs, student-fund mandate, faculty and student rosters, Bloomberg/data permissions, conference or Finance TV appearances, and any public research artifacts. The page does not establish that the Student Managed Investment Fund uses the described forecasting work or that any external manager adopted it.
• RIT’s profile for Jialin Qian identifies Qian as an Assistant Professor of Finance whose expertise includes AI and Finance and textual analysis in finance. The profile lists invited 2025 presentations of “Harnessing Generative AI for Economic Insights,” earlier work on ChatGPT and corporate policies, and a conference paper on short-selling hedge funds. Qian’s current FINC-475 course covers supervised and unsupervised learning, NLP, language models, financial-disclosure analysis, ML time-series forecasting, ethics and regulation; the adjacent Financial Analytics course covers portfolio optimisation, valuation and default modelling in R or Python. These are public faculty, paper and curriculum signals, not evidence of a proprietary fund model, live trading use, data licence or investment result.
• RIT’s page creates a particularly clear paper-to-course recovery queue: obtain the full generative-AI and ChatGPT papers, conference versions, code or data statements, course syllabus and assignments, and any student or employer project links. The public record supports research into disclosure analysis, corporate-policy effects, language models, return forecasting and short-selling-fund behaviour; it does not support claims about which firms use those methods.

Research implication: finance programmes and named professors should be indexed as separate evidence objects—faculty research, lab infrastructure, course methods, student-managed capital, and employer or practitioner interfaces. Morgan adds a lab-and-student-fund route around Bloomberg, forecasting and credit risk. RIT adds a current finance-AI curriculum and a public generative-AI research trail. Both are useful for finding people, papers, assignments and datasets while preserving the boundary between academic activity and verified investment deployment.

Recovery: capture Morgan’s Capital Markets Lab and Investment Club materials, project documentation, student-fund reports and speaker records; retrieve Qian’s paper versions, course materials, code/data statements, coauthors and conference recordings; map student and alumni transitions through public university or employer sources; and keep teaching, research, advisory work, fund activity and production deployment separate.

September 5, 2026 — Ohio State Fisher’s asset-pricing research and applied FinTech curriculum

• Ohio State Fisher’s profile for Aditya Chaudhry identifies him as an Assistant Professor of Finance whose primary fields are asset pricing and macro-finance. The profile says his work uses new empirical strategies and sophisticated statistical and machine-learning methods to identify structural parameters, and lists a University of Chicago Booth finance PhD and University of Virginia commerce/mathematics degrees. His public working-paper list includes a machine-learning approach to extracting high-frequency expectations from asset prices and work on factor proliferation, macroeconomic uncertainty and expected returns. These are faculty research records; they do not disclose a hedge-fund mandate, proprietary data or live deployment.
• The same profile lists a 2025 Journal of Financial Economics paper on prices and analyst cash-flow expectations and a 2017 ICML paper on uncertainty assessment and false-discovery-rate control in high-dimensional Granger causal inference. The combination is relevant to the research map because it connects asset-pricing questions with causal/time-series methodology and uncertainty control. It should be treated as a research-method route, not as evidence that a particular manager uses the methods or that the findings form a tradable signal.
• Fisher’s FinTech Micro-Credential page describes a 4.5-credit graduate-level programme covering AI, machine learning and blockchain in financial services. Its public curriculum includes supervised, unsupervised and reinforcement learning; ML for credit modelling, investment and insurance; Python implementation and deployment; and forecasting price and volume of blockchain assets. The page names Taner Pirim as Academic Director and Senior Lecturer in Finance and Nathan Craig as Associate Professor. Its detailed schedule is dated January–April 2024, while the page also contains a Fall 2025 application notice, so the programme’s current delivery status should be verified rather than assumed.
• The programme page describes two project components: one applies ML in finance, including credit, investment and insurance, and the other uses data-mining tools and Python to build and deploy models. It also frames the programme for both Ohio State graduate students and professionals with quantitative finance/statistics backgrounds. This exposes a practical training and talent-discovery surface, but it does not identify project datasets, student submissions, employer sponsors, model governance, evaluation splits or investment authority.

Research implication: Ohio State adds a useful bridge between academic asset-pricing identification and applied finance-technology training. The research queue is to retrieve Chaudhry’s high-frequency-expectations paper and code/data statements, identify the data-generating assumptions and uncertainty controls, and recover the FinTech programme’s project briefs, instructor materials, student outputs and employer links. The public record supports investigation of high-frequency expectations, factor discovery, causal inference, credit/investment/insurance modelling, blockchain-market forecasting and Python model deployment; it does not support claims about a named firm’s use, superiority or performance.

Recovery: archive Fisher faculty/CV and paper links, the FinTech credential page and dated schedules, programme faculty biographies, project rubrics, student or employer showcases and any public data/code disclosures. Keep current faculty roles, historical programme schedules, academic papers, classroom exercises, corporate training, external consulting and verified investment deployment separate.

September 5, 2026 — Rochester Simon’s manager-skill, investor-attention, and finance-AI curriculum route

• Ron Kaniel’s Simon Business School profile identifies him as the Jay S. and Jeanne P. Benet Professor of Finance and records prior faculty roles at Duke and the University of Texas at Austin, a Stanford visiting-scholar role, and a Wharton finance PhD preceded by computer-science and mathematics/computer-science degrees from the Hebrew University of Jerusalem. The profile explicitly lists AI among his research interests and includes “Machine-Learning the Skill of Mutual Fund Managers,” “Using Machine Learning to Predict Mutual Fund Performance,” “Filing Speed, Information Leakage, and Price Formation,” and “Unmasking Mutual Fund Derivative Use.” These are public academic records; they do not disclose a hedge fund’s implementation, proprietary data, or investment authority.
• Kaniel’s profile describes a research agenda around mutual-fund investment decisions, security-price effects, investor communities, trading volume, order flow and return predictability. That combination creates a concrete research queue: examine whether manager-skill labels are trained from holdings, flows, returns, text, trading behaviour or other observables; identify the point-in-time information set; and test whether results survive fund-selection, publication and multiple-testing controls. The public profile alone does not provide those methodological details or establish a tradable signal.
• Yukun Liu’s Simon profile identifies him as William H. Meckling Associate Professor of Business Administration, with Yale economics PhD and Cornell economics/mathematics training. It lists asset pricing, labor and finance, and FinTech as primary fields and records a 2026 Journal of Finance paper, “Institutional Investor Attention,” alongside work on shareholder governance, long-run risk and cryptocurrency risk factors. This adds an investor-attention and institutional-behaviour route that can be cross-referenced with public fund flows and filings, without inferring any employer’s use.
• Simon’s finance PhD curriculum describes formal modelling and empirical testing across asset pricing, financial institutions and corporate finance. Its public course list includes continuous-time trading and stochastic asset-price models, empirical asset pricing with time-series predictability and mutual-fund performance, causal empirical corporate finance, advanced empirical asset pricing and financial/economic networks covering OTC, production and payment networks. The programme is a lineage and methods surface; the page does not identify a particular hedge-fund sponsor or production system.
• Simon’s current MS in Finance curriculum lists AI & Deep Learning, Asset Management, Financial Technology and Intro to AI & Finance. The asset-management course explicitly covers institutional and private clients, public equity, fixed income, hedge and liquid portfolios; the FinTech concentration links FIN 446 Financial Technology and FIN 478 Intro to AI & Finance, with AI & Deep Learning as an elective. The page says the programme is accepting applications for January 2027 entry. Curriculum presence indicates training exposure, not that students or faculty operate live funds.

Research implication: Rochester adds a particularly clear academic idea stack around measuring manager skill, predicting mutual-fund performance, information leakage, derivatives use, investor attention, order flow, networks and asset pricing, alongside formal finance and AI coursework. The next evidence to recover is the full Kaniel and Liu paper set, code/data statements, replication files, seminar recordings, student theses and public programme projects. These can inform model and data hypotheses, but neither the university pages nor the paper titles establish current use by any named hedge fund, manager or allocator.

Recovery: archive Kaniel and Liu CVs, papers and working-paper versions; retrieve Simon course syllabi, datasets, assignments, student research and seminar recordings; map doctoral advisers and placements where publicly documented; and preserve separate states for academic method, classroom exercise, public data, employer relationship, consulting, live investment use and performance.

September 5, 2026 — Wisconsin’s finance-AI faculty, student capital, and M&A-ML route

• Zhongtian Chen’s Wisconsin School of Business profile records his July 2025 appointment as Assistant Professor of Finance, Investment, and Banking, with a Wharton finance PhD, Duke economics MA, and Renmin University MS in Finance and BS in Mathematics and Applied Mathematics. The profile describes a research intersection of asset pricing, machine learning, and behavioral finance, focused on belief formation and investor behavior. Wisconsin’s faculty feature makes the research object more specific: rather than only forecasting prices, Chen asks whether models can predict what investors do next, and describes a line of work that manipulates AI-generated video attributes such as speaker gender, race, and energy level to study how financial information changes beliefs. This is a public academic research description—not evidence that a fund uses synthetic video, owns such a dataset, or has granted a model investment authority.
• Mark Fedenia’s faculty profile adds a second finance-AI personnel node. Fedenia is Academic Executive Director of the Hawk Center, Baird Professor in Finance, and former director of Wisconsin’s Applied Security Analysis Program (ASAP), where he spent 21 years overseeing a programme in which students manage multimillion-dollar portfolios. His public publication list includes the 2024 paper “Machine Learning and Trade Direction Classification: Insights from the Corporate Bond Market.” The paper title establishes the research object; the profile does not establish a fund implementation, proprietary bond feed, production classifier, or investment result.
• The Hawk Center describes ASAP as a student-managed investment programme established in 1970, with students using real-world tools to manage more than $25 million across equity, investment-grade, high-yield, and Treasury portfolios. It also records more than 700 alumni serving as speakers, mentors, employers, advisory-board members, and hosts for firm visits and conferences. This is a substantial public talent and practitioner-contact surface. It is not evidence that the portfolios use machine learning or that any particular hedge fund sponsors the programme.
• Wisconsin’s MS in Finance, Investment, and Banking and curriculum page connect the finance degree to the Hawk Center, applied security analysis, an investment-firm research internship, portfolio-management responsibilities, practitioner presentations, and a student team managing more than $25 million. The current pages describe the programme structure and recruiting surface; they do not provide a current roster of employer placements, technical assignments, model code, or data entitlements.
• A 2019 Nicholas Center report, “From ML to M&A” is a useful public implementation artifact. A student data-analytics team, advised by Brad Chandler, combined random forest, three-layer neural-network, and ensemble models to predict acquisition announcements from publicly available data. The report says the model used more than one billion public data points, trained on information through June 30, 2019, and evaluated the following twelve months. It reports 10.8% correct ensemble predictions against a 1.71% base rate in its test period, while also flagging rarity, limited data, repeated predictions, prediction lag, and the need for more industry-specific fundamental analysis. Those figures are the report’s own backtest claims, not an independently reproduced trading result or evidence of an investable signal.
• Wisconsin’s AI Hub for Business also exposes a discovery route that would be missed by a hedge-fund-only search. Its public material lists finance, predictive modelling, financial time series, text and video analytics, and machine-human collaboration, and provides a transcript for a public “UNsupervised” podcast episode with Erik Mayer discussing robo-advice, real-time news and big-data processing in hedge-fund-style trading. The episode is faculty commentary, not a manager disclosure; its value here is the vocabulary and speaker graph it supplies for further searches.

Research implication: Wisconsin links three distinct surfaces that should remain analytically separate: Chen’s multimodal behavioural-finance research; Fedenia’s market-microstructure and institutional-investor research; and a long-running student-capital programme with practitioner access. The Nicholas Center report is especially useful for a model-reconstruction queue because it discloses model families, point-in-time cutoff, forecast horizon, target rarity and acknowledged failure modes. The public evidence still does not show which, if any, investment manager supplied data, recruited a named student, adopted a model, or ran it in production.

Recovery: retrieve Chen’s CV, working papers, AI-generated-video study materials and any public seminars; obtain the full Fedenia corporate-bond paper and code/data statements; archive Hawk Center conference programmes and public alumni/employer rosters; recover the Nicholas Center process manual, source-data definitions, feature list and out-of-sample history; and ingest the AI Hub podcast transcript with timestamps. Track academic method, classroom project, student-managed capital, employer contact, sponsorship, consulting, production deployment and performance as separate evidence states.

September 5, 2026 — Illinois Gies’ MIT finance lineage, agentic trading classroom, and quant curriculum

• Victor Duarte’s Gies profile records an MIT finance PhD and current Assistant Professor of Finance role. It lists forthcoming work titled “Machine Learning for Continuous Time Finance” and current courses that cover neural networks, regression trees, gradient boosting, clustering, principal-component analysis, Deep Q-Networks, Q-learning, SARSA, policy gradients, option pricing, portfolio selection, and credit-card fraud detection, with implementation in Python using PyTorch, scikit-learn, XGBoost and TensorFlow. This is unusually concrete academic and curriculum evidence about model families and finance objects; it does not establish a hedge-fund deployment or a live trading mandate.
• Dmitriy Muravyev’s Gies profile adds a market-microstructure and high-frequency-data node. The profile lists asset pricing, derivatives, high-frequency data and applied machine learning as research interests, and links work on informed-trading intensity, close auctions, retail/options trading costs, data-driven measures of high-frequency trading, price discovery and liquidity. The page also records current editorial roles at finance journals and derivatives/options teaching. These publication and personnel signals identify research objects for further paper and code recovery; they do not show that an investment manager uses the methods.
• The Gies MSF curriculum exposes a formal training map rather than a generic “fintech” label. Its specializations include asset management, quantitative finance, data analytics and fintech, and finance research. Named electives include FIN 553 Machine Learning in Finance, FIN 554 Algorithmic Trading Systems Design & Testing, FIN 556 Algorithmic Market Microstructure, FIN 550 Big Data Analytics in Finance, FIN 557 Financial Data Management and Analysis, and FIN 580 Quantamental Investment. The page describes an employer-facing research and career-preparation surface, but does not disclose student projects, employer sponsors, proprietary datasets or production controls.
• Gies’ July 1, 2026 account of FIN 580: Quantitative Investment is a new agentic-AI discovery surface. Clinical Assistant Professor Tony Zhang says students build institutional-style dashboards, use agents to translate designs into code, and scrape and verify interest and retail-sentiment data that conventional terminals may not contain. The article describes a student-built multi-agent platform for trading on news-derived narrative signals, with separate agents assigned to valuation, risk and alternative-data collection that challenge one another and troubleshoot errors. The article also quotes students describing human supervision, workflow design and extensive debugging. These are classroom projects and participant accounts, not audited performance, a firm partnership, or evidence that any student system trades external capital.
• The same Gies programme page identifies a faculty and talent route beyond Duarte and Muravyev. It names Tony Zhang as the FIN 580 instructor and describes the course as moving from passive market-data consumption to active AI engineering. The school’s FELP brochure describes Zhang’s public work around alternative data, AI techniques, quantitative finance and portfolio management, and identifies his development of an AI teaching assistant called AristAI. The brochure records an electrical-and-computer-engineering PhD from the University of Minnesota and an analytic-finance/entrepreneurship MBA from Chicago Booth. These are university-reported credentials and project descriptions; the brochure does not establish product adoption, external-fund use or investment results.
• The Gies MSBAi programme requirements provide a parallel implementation surface. They describe FIN 550 as a predictive-modelling pipeline that ends in a trading-signal system, BADM 590 as agentic AI covering LLMs, RAG, prompt engineering and agentic analytics workflows, and BADM 576 as advanced ML covering transformers, time series, MLOps and LLMOps. The named portfolio artifacts—trading-signal system, functioning AI analytics workflow and deployment-oriented advanced ML project—are educational deliverables, not proof of live investment use.

Research implication: Illinois adds an explicit bridge from finance theory and high-frequency market structure to model implementation, alternative data and multi-agent workflow design. The most actionable public artefacts are not claims about returns; they are the decomposition of research labor into data collection, verification, valuation, risk, narrative-signal construction, code generation, backtesting and human review. That decomposition supplies a search vocabulary for fund job descriptions, conference talks, student projects and vendor partnerships while keeping classroom activity separate from verified firm deployment.

Recovery: obtain Duarte’s continuous-time ML paper and code/data statements; retrieve Muravyev’s HFT and informed-trading papers, datasets and replication files; archive current Gies syllabi, project rubrics and FIN 580 student showcases; recover the Tony Zhang/AristAI materials and any public alternative-data book artefacts; and search Gies finance seminars, podcasts, YouTube and alumni/employer pages for named fund or data-vendor links. Preserve separate evidence states for faculty research, course assignment, student-managed or simulated capital, public data, employer contact, sponsorship, production deployment and performance.

September 5, 2026 — Minnesota Carlson’s finance-ML workflow, dissertation, and structural-estimation route

• Adam Zhang’s Carlson profile identifies him as an Assistant Professor of Finance with Stanford economics PhD and MA training, following UCLA mathematics/economics training with a computing specialization. The profile describes work on asset-price mechanisms, retirement saving, housing, income and wealth inequality, and machine-learning techniques for structural estimation. This is a macro-finance and economic-mechanism route, rather than a simple price-prediction label; the public profile does not establish a manager relationship or live investment use.
• Ruoxi (Hank) Tian’s profile identifies a finance PhD candidate with a Duke economics MA and Guangdong University of Foreign Studies BA in English and Finance. Her stated research interests combine causal inference, machine learning and natural-language processing in finance, sustainable finance and financial technology. The profile also links to LinkedIn, creating a personnel-discovery route, but it does not provide a current employer, paper-level method, dataset, or fund connection.
• Carlson’s MS Finance course sequence supplies an unusually operational curriculum surface. It names Bloomberg, FactSet, CRSP and Compustat in the computing-for-finance course; Python for financial-data access and visualisation; econometrics and computational methods; quantitative portfolio analysis; and an “Introduction to Machine Learning for Finance” course focused on asset pricing and credit assessment. The ML course explicitly breaks a real project into data collection, data management, exploratory analysis, learning/prediction and communication, while experiential-learning teams work with client companies and present findings. The programme page does not identify client identities, data permissions or production adoption.
• The Carlson finance dissertation page gives the academic lineage and idea queue more depth than a faculty keyword search. It lists “Essays on FinTech and Machine Learning in Finance” by Keer Yang, “Two Essays on Dynamic Learning Under Information Asymmetry” by Fangyuan Yu, and “Extrapolative Expectation, Financial Frictions, and Asset Prices” by Yao Deng, alongside placements and presentations. These titles expose research objects—FinTech adoption, learning under asymmetric information, extrapolative expectations and asset pricing—that can be followed to papers, supervisors and code. The page does not imply that any graduate work was adopted by a hedge fund.
• Murray Z. Frank’s Carlson profile adds a corporate-finance and seminar-network node. It lists machine-learning methods in finance as an expertise, a Queens University economics PhD, and co-founding the Virtual Corporate Finance Wednesdays seminar series. The seminar link is a useful route to speakers, paper titles and recordings, but the profile does not establish that seminar participants share proprietary data or that any method is used in a live portfolio.

Research implication: Minnesota exposes a complete public learning workflow—from named financial databases and Python, through empirical finance and ML for asset pricing/credit, to client-facing experiential work—alongside doctoral topics in FinTech, dynamic learning and investor expectations. It is a valuable bridge for searching the vocabulary of data management, causal NLP, structural estimation and information asymmetry across fund jobs and conference talks. The record still does not show which client, employer or investment manager received a model, supplied confidential data, or used a system in production.

Recovery: retrieve Zhang’s papers and code/data statements; recover Tian’s working papers, supervisors and conference appearances; obtain the Carlson ML course syllabus, project briefs and experiential-learning client records where public; follow Keer Yang, Fangyuan Yu and Yao Deng to dissertation repositories and current roles; and archive Virtual Corporate Finance Wednesdays schedules and recordings. Keep faculty research, doctoral work, classroom projects, client work, employer contact, production deployment and performance as separate evidence states.

September 5, 2026 — Delaware IFSA’s bank-linked financial-services analytics pipeline

• The University of Delaware’s Institute for Financial Services Analytics is a distinct academic-to-industry object, established in 2012 by Lerner, the College of Engineering and JPMorgan Chase. Its public research menu includes business and social-network analytics, causal learning, fairness in machine learning, graph-based data science including hypergraph neural networks, financial risk analytics, customer targeting and loan optimization. The institute says faculty span business, electrical engineering, computer science, finance, MIS and statistics, while students combine research collaboration, corporate internships and coursework in machine learning, data mining, process analysis and optimization. This is a finance-specific research and talent surface; it does not disclose JPMorgan’s internal models or a hedge fund’s process.
• Bintong Chen’s first-party profile identifies him as Chaplin Tyler Professor and Director of IFSA and the MS/PhD FSAN programmes. It records a University of Pennsylvania PhD in decision science and MS in systems engineering, after electrical-engineering and naval-architecture degrees from Shanghai Jiao Tong University. His public publications include work on credit-card fraud detection with consumer incentives, personalized pricing, frontier portfolios and optimization. The profile establishes faculty lineage and research objects, not a production financial model or investment result.
• Delaware’s M.S. in Financial Services Analytics describes a 30-credit interdisciplinary STEM programme jointly run by Lerner and Engineering. Its public tool list includes R, Python, SAS, SQL, Gurobi, CVX and PyTorch, with core areas in financial risk management, financial operations and financial management. The page names an intended path into fintech and financial services and identifies Bintong Chen as programme director; it does not publish current project datasets, employer-specific assignments or model-governance details.
• The FSAN PhD page makes the research-to-employer bridge more explicit. It lists dissertation topics including Bayesian causal inference, predictive analytics, bank systemic risk, credit-card fraud, NLP/chatbots, text-based industry classification, loan-process optimization, customer targeting and active learning on graphs. It also documents corporate-sponsored internships in wealth management, machine learning, data architecture, investment banking and global finance, with public internship destinations including JPMorgan Chase, Barclays, Sallie Mae, Amazon, LinkedIn and HP. A separate 2026 IFSA advisory-board announcement names senior personnel from JPMorgan Chase, Capital One, Bank of America, Barclays US Consumer Bank and Citi. These are institutional and personnel connections; they do not show which firm used which research, what data was shared, or whether any model reached production.
• The institute’s 2019 Machine Learning in Financial Services conference account supplies an additional media and conference recovery route. It records sessions involving industry and academic participants, a student research presentation, and themes around the opportunities and challenges of ML in financial services. The institute’s 2018 FSAN programme video/transcript describes JPMorgan Chase support, multidisciplinary engineering/business research and the role of seminars in exposing students to applied research. These are dated public programme artifacts, not disclosures of a particular bank or asset manager’s live system.
• Xiao Fang’s profile and CV add a methods-and-publication node. Fang is an IFSA Senior Fellow whose stated method areas are machine learning and optimization, with applications in financial technology and social-network analytics. His public record includes deep-learning imputation, expert-knowledge-assisted firm classification, link recommendation and social-network methods. The public pages provide paper and personnel leads for artifact recovery; they do not establish an external manager relationship or deployment.

Research implication: Delaware exposes an end-to-end finance-analytics pipeline—interdisciplinary faculty, a named bank-founded institute, explicit methods and data-science tools, doctoral topics, corporate internships, industry advisory governance and dated conference/video surfaces. The most useful search vocabulary is broader than price prediction: credit risk, fraud, causal learning, fairness, hypergraphs, process mining, graph-based active learning, text classification and loan optimization. Those clues can improve title-blind searches across fund, bank, allocator and vendor materials while keeping bank sponsorship, student work and verified production deployment distinct.

Recovery: archive IFSA faculty and advisory-board rosters; retrieve Chen, Fang and Laux CVs and papers; recover FSAN course syllabi, dissertation abstracts, code/data statements and public student projects; capture the 2019 conference programme, recordings and seminar archive; verify internship and graduate placements; and search the named banks, asset managers and financial-technology employers for attributable follow-on evidence. Preserve separate states for academic method, classroom work, sponsored research, internship, employer relationship, proprietary data, production deployment and performance.

September 5, 2026 — Clemson conference exposes a mutual-fund ML measurement route

The 2026 Clemson Finance Research Conference agenda places “Active Machine-Learning-Based Trading and Mutual Fund Performance” in its April 18 programme, with Xiaowen Hu of Southern Methodist University presenting alongside Maximilian Rohrer of the Norwegian School of Economics and Hanjiang Zhang of Washington State University. The same agenda lists “Dinner Table Alphas?” by Sean Cao, Huaizhi Chen, Lauren Cohen, and Tianchen Zhao, providing a second academic route into information and investment research. The conference programme establishes dated author and presentation metadata; it does not establish a fund partnership, production model, proprietary dataset, or investment authority.

The related SMU paper record describes AMLT as a holdings-based measure of how mutual-fund decisions align with forward-looking deep-neural-network signals built from quantitative and textual information, with employee AI-talent analysis used as a validation channel. Its reported results and methodology remain paper claims pending independent replication. The useful follow-up is artifact-level: recover the paper, slides, appendices, data and code statements; inspect the point-in-time information set, talent labels, turnover and transaction-cost treatment; and keep any academic measurement separate from claims about named hedge funds or allocators.

September 5, 2026 — NYU and Singapore routes connect academic finance to family-office personnel

• An NYU School of Law profile for Winston Wenyan Ma identifies him as an adjunct professor and Executive Director of the Global Public Investment Funds Forum. The profile states that he is a partner of Dragon Global, described there as an AI-focused family office and founding member of Dragon.AI, and records a prior ten-year role as Managing Director and Head of the North America Office at China Investment Corporation. His NYU seminar covers sovereign funds, alternative investments, digital and technology sectors, cross-border regulation, and strategic capital. This is public personnel, teaching and allocator-network evidence; it does not expose Dragon Global’s models, data, vendors, portfolio process or investment results.
• A CFA Society Singapore programme brochure provides a dated Southeast Asian talent route. It describes Phoebe Gao as an Assistant Professor at Singapore Institute of Technology with research interests in asset pricing, risk management, AI, machine learning and sustainability, including information flow between equity and options markets, digital transformation and analyst-forecast sentiment. The same brochure identifies Phuah Keng Keat as an equity analyst at Singapore multi-family office Kamet Capital, covering technology, banking, healthcare and consumer staples. The brochure also attributes consulting work on quantitative trading strategies to Gao; these are dated programme biographies and should not be treated as proof of Kamet’s AI use or of a shared strategy.

These routes widen the personnel graph beyond hedge-fund job titles: sovereign-investment teaching, an AI-focused family-office affiliation, finance-faculty research, and multi-family-office analyst biographies can reveal relevant people and vocabulary. They do not establish a model, dataset, partnership, production deployment, authority or performance outcome for any named allocator.

September 5, 2026 — Rice and GW expose finance-programme operating surfaces

• Rice Business’ Kerry Back profile identifies a finance professor who teaches machine learning in finance and quantitative investment strategies to MBA students, quantitative finance to data-science students, and asset pricing theory to PhD students. Rice’s 2025 teaching-award account says his Generative AI in Finance course trains students to apply AI and coding to financial analysis and describes his tutorial on integrating AI into teaching. This is curriculum and faculty evidence; it does not disclose student datasets, code, employer projects, fund authority or investment outcomes.
• Rice’s 2026–27 finance-seminar calendar adds a recovery queue: its August 28, 2026 programme listed Alejandro Lopez Lira for “Can LLMs Discover Novel Economic Theories?”, and the calendar lists later finance speakers from Columbia, MIT, Berkeley, NYU, Harvard, Stanford and other universities. Seminar listings establish dates, speakers and titles; they do not establish recordings, paper methods, fund partnerships or deployment.
• The GW Investment Institute’s student-fund page says students serve as analysts and portfolio managers across four university-endowment funds totaling more than $11 million. It lists a quantitative fund in which FINA 4103 students build and test predictive models, reports that the fund was established in 2021 with capital from donors and Aron Kershner, and identifies Kershner as portfolio adviser. The page also names FactSet, Bloomberg, BlackRock Aladdin and PitchBook as research tools. These are current programme and toolchain claims; they do not establish model quality, data entitlements, live-trading controls, external-fund adoption or attributable performance.

Together, these pages add two different research-to-talent interfaces: Rice exposes faculty-led GenAI finance pedagogy and a seminar graph, while GW exposes student-managed capital, quantitative-model coursework, an industry adviser and institutional research tools. The appropriate follow-up is to recover syllabi, assignments, student reports, seminar recordings and alumni paths while keeping classroom activity, endowment management, employer contact and professional deployment separate.

September 5, 2026 — Family-office programmes and allocator conference reports

• Wharton’s June 2026 Advanced Finance Program account profiles Warren Sheng, Principal of IVision Holdings, who built a family office and studied private equity, distressed investing and restructuring, venture capital, portfolio management, fixed income and credit, and private wealth management. It identifies Kevin Kaiser as a finance professor and senior director of the Harris Family Alternative Investments Program, and states that Sheng’s family office later became a limited partner in a fund run by a professor met through the programme. The account does not identify the fund, strategy, commitment size, AI use or investment result.
• London Business School’s Leading the Family Office programme is a five-day route aimed at family members and family-office professionals, including investment-committee members, trustees, board members, executives and next-generation members. Its curriculum names family-office models, governance, investment execution, risk evaluation, asset-class assumptions, portfolio allocation and dynamic portfolio strategy. This is a participant-discovery and allocator-vocabulary surface, not evidence of any attendee’s model or AI workflow.
• Yale’s Swensen Asset Management Institute report on the 2026 Sohn Investment Conference is a student-authored account dated May 28, 2026. It names public speakers from Philosophy Capital, Whale Rock, Maplelane Capital, Ananym Capital, Epicenter Capital, Atreides and Diameter Capital, and records the student’s summary of AI-infrastructure, software, private-credit and labor-market themes. The report is useful for speaker and thesis discovery, but it is not a transcript or independently verified position book and does not establish each firm’s AI use, holdings, sizing, model, data or performance.

These routes add allocator and family-office discovery surfaces that do not depend on “hedge fund” or “AI” appearing in the title. Follow participant biographies, faculty papers, event recordings, firm pages and regulatory records, while keeping education, networking, investment, firm disclosure and verified deployment separate.

September 5, 2026 — Harvard AI-investment policy, NYU sovereign-fund events, and Sentient lineage

• Harvard Berkman Klein Center’s Paul Fehlinger profile identifies him as an affiliate working on AI investment, entrepreneurship, policy, and ethics/governance. The profile states that he is Senior Director of Policy, Investment & Innovation at Project Liberty Institute within a family-office-backed USD 500 million effort, and that he works with institutional investors, venture firms, family offices, entrepreneurs and policymakers on systemic risk, demand signals and capital allocation across the AI stack. It also says he co-built a responsible-AI investment process involving LPs and VCs representing more than USD 6 trillion. The page links dated commentary on AI alpha/risk, investing with AI versus investing in AI, and LP/data incentives. These are profile and linked-publication claims, not evidence of a hedge-fund model, data entitlement, investment authority or audited result.
• NYU’s Global Public Investment Funds Forum describes a university forum focused on sovereign wealth funds, public pension funds and other public asset owners at the intersection of law, finance, politics and technology. The page states that public investment funds manage more than USD 30 trillion and lists events involving APG and German KfW on AI digital infrastructure, Norway NBIM on AI investing, Ireland’s ISIF on strategic development in the AI economy, Temasek on public investment funds, and an IFSWF/Jain Family Institute workshop on sovereign AI. It identifies Kevin Davis as Faculty Director and Winston Ma as Executive Director. This is an allocator and event-discovery graph; it does not establish adoption of a named AI system, proprietary-data sharing or investment performance.
• The University of Indianapolis AI Summit dates its event to April 9, 2026 and identifies Babak Hodjat as keynote speaker and Chief AI Officer at Cognizant. The organizer describes Hodjat as former co-founder and CEO of Sentient, founder of Sentient Investment Management, and a contributor to Artificial Life, Agent-Oriented Software Engineering and Distributed Artificial Intelligence; it lists a PhD in Machine Intelligence from Kyushu University and 40 issued US patents. The page’s “world’s first AI-driven hedge-fund” wording is retained as an organizer characterization, not an independently verified historical conclusion. The event page does not expose Sentient’s model architecture, investment data, authority, live performance or continuity with Hodjat’s current employer.

These routes add Harvard policy-investment, sovereign-fund, and historical AI-hedge-fund personnel surfaces to the academic map. Follow linked publications, recordings, papers, patents, event speakers and first-party employer material while preserving the boundaries between policy, historical firm claims, academic lineage and verified deployment.

September 5, 2026 — Miami’s computational-finance faculty and explicit AI/ML modelling course

• The University of Miami research portal’s Alok Kumar profile identifies Kumar as a Professor in Finance and lists behavioral finance, empirical asset pricing, corporate finance and computational economics as expertise. It links his faculty CV, research archive and SSRN author page, and records a lineage through a Cornell economics PhD and MA, Yale management MA, Dartmouth engineering-management degrees, and an IIT Kharagpur engineering degree. This is first-party personnel and publication-index evidence; it does not establish a hedge-fund affiliation, proprietary data access or production deployment.
• Kumar’s dated CV records University of Miami MSF teaching of “Advanced Financial Modeling using Artificial Intelligence and Machine Learning” in 2020 and 2022, alongside behavioral finance and quantitative/analytical finance teaching. His linked research archive lists computational approaches including neural-net modelling of historical market calls, network analysis of search dynamics, sentiment and attention, analyst forecasts, fund-performance prediction and a working paper on aggregating artificially intelligent earnings forecasts. These are academic research and teaching disclosures. The archive’s paper descriptions should be checked against the underlying papers, data statements and replication materials before any performance or tradability interpretation.
• This route adds a finance-programme surface more specific than a generic “fintech” label: model families and computational methods are named in an instructor’s course and research history, while the same public record exposes doctoral lineage, coauthor networks and a paper-recovery queue. It is still not evidence that Miami’s MSF students, Kumar or any named fund used a particular model with live capital. The relevant next checks are the current course syllabus, assignments, software and datasets, paper versions and code/data statements, student projects, and any employer or practitioner interface.

Research implication: add computational economics, investor-behaviour and attention signals, analyst-forecast aggregation, neural-network replication of historical decisions, and finance-programme teaching of AI/ML modelling to the academic idea map. Treat them as hypotheses and talent/research surfaces, not as proof of an investment manager’s strategy.

Recovery: preserve the dated CV and current university profile, retrieve Kumar’s AI/ML course materials and underlying papers, map coauthors and doctoral lineage where publicly documented, and search Miami student-fund, seminar and alumni surfaces. Keep academic research, classroom exercise, public data, employer contact, consultancy, model ownership, production deployment and performance in separate evidence states.

September 5, 2026 — Amsterdam’s quantitative-finance curriculum and public trading-strategy assignment

• Tomislav Ladika’s University of Amsterdam profile identifies him as an Associate Professor of Finance, Program Director of the MSc Finance, and a teacher of programming applications in finance. The profile lists current teaching in Investments and Algorithmic Trading, Data Analytics and Quantitative Trading, and an on-demand Executive Education AI in Finance module. It also records a Brown University PhD in Economics and prior research on financial-market attention, executive incentives and corporate investment. This is faculty and curriculum evidence; it does not establish a hedge-fund relationship or deployment.
• The Amsterdam Business School’s Quantitative Finance track describes a programme built around mathematical pricing models, computing, big-data analysis and Python. Its Computational Finance course includes data handling, risk-management calculations and machine-learning techniques; the track also includes applied financial econometrics, empirical methods, quantitative finance and algorithmic trading, advanced risk management, behavioural finance and a supervised thesis. The page identifies a programme architecture that combines prediction, causal/econometric discipline, market microstructure, risk and implementation rather than treating model output as a standalone strategy.
• The same page describes a “Real-life Case: Construct a Quantitative Trading Strategy.” Students select Dutch, Belgian and French stocks using price data extending back to 1990, define purchase and sale timing, pitch the strategy to an investment committee, and evaluate it with current prices out of sample. The university calls this a hedge-fund-style classroom exercise. It is a valuable artifact-recovery route for data provenance, leakage controls, feature construction, portfolio rules, turnover, costs and evaluation design, but it is not evidence of live capital, student performance, fund sponsorship or institutional adoption.
• The UvA PhD project for Oskari Veijalainen, already tracked as a separate research route, provides an adjacent method vocabulary: return predictability, predictive uncertainty, cross-sectional risk factors, momentum, under- and over-reaction, and signal generation. The project is supervised by Simon Rottke and Florian Peters. The programme and project pages do not disclose code, full data entitlement, model weights, a firm partner or production authority.

Research implication: Amsterdam adds a particularly concrete academic-to-practice surface: a public classroom assignment that specifies geography, historical sample depth, investment-committee presentation and out-of-sample evaluation, alongside a quantitative-finance curriculum that joins Python/ML with econometrics, microstructure, risk and algorithmic trading. Follow the course catalogue, assignment rubric, student presentations, thesis titles, supervisor graph and alumni paths. Keep classroom design, research method, public data, employer contact, proprietary data, production deployment and performance separate.

Recovery: capture the current UvA course catalogue and detailed assessment materials; preserve the quantitative-trading case description and its 1990–present data claim; recover Ladika’s CV, papers and AI-in-finance teaching materials; map Rottke/Peters and Veijalainen papers; inspect public student and alumni outputs; and search UvA corporate partnerships and recruiting pages for attributable follow-on evidence.

September 5, 2026 — St.Gallen’s quantitative-finance programme and public volatility-model artifacts

• The University of St.Gallen’s MiQE/F master’s page describes training in econometrics, quantitative methods, machine learning and big-data analysis, with a finance/economics foundation and a research thesis. Its curriculum page names Data Analytics I: Predictive Econometrics, Data Analytics II: Causal Econometrics, Financial Programming with Matlab, Data Handling: Databases, Quantitative Asset Management, Derivatives Modeling in Python, Applied Quantitative Asset Management, Quantitative Behavioral Finance, AI for Decision Making and Machine Learning. These are programme disclosures; they do not establish a fund sponsor, proprietary data arrangement or live investment authority.
• Lukas Gonon’s FSI-HSG profile identifies him as Assistant Professor in Artificial Intelligence in Finance, with research on deep learning and quantitative finance, particularly time series, derivatives pricing and hedging. The page lists papers on operator deep smoothing for implied volatility, deep hedging under distributional adversarial attack, graph neural networks for systemic-risk measures, generative financial time series, stochastic-volatility calibration, and deep hedging with second-order optimisation. It records prior research and teaching roles at ETH Zurich, LMU Munich and Imperial College London, plus work with Sony and J.P. Morgan. These paper titles and affiliations are research/personnel evidence, not evidence of employer adoption or trading performance.
• Francesco Audrino’s St.Gallen profile adds a long-running financial-econometrics route: computational statistics, machine learning, sentiment analysis, volatility forecasting and regime-switching models. The profile records projects on high-dimensional volatility matrices, implied-volatility surfaces, behavioural asset pricing, cross-asset dependence in high-frequency data, SentiVol sentiment analysis with Bayesian model averaging for volatility prediction, causal analysis with high-dimensional financial data, and a 2026–28 Financial Social Networks project. It links public FGD code and a SentiVol forecast site, although the latter was unavailable during this check. The profile establishes public code/project links, not a production system or investable result.
• Lyudmila Grigoryeva’s quantitative-methods page names reservoir computing, recurrent/deep learning, dynamical systems, time-series forecasting and financial econometrics as research areas. This creates a separate paper and talent route into sequence models and dynamic-process learning. It does not establish that any named allocator uses reservoir methods.

Research implication: St.Gallen broadens the academic idea map from generic “AI in finance” to specific artifact classes: deep hedging, implied-volatility smoothing and calibration, adversarial robustness, graph-based systemic-risk measures, generative financial time series, Bayesian model averaging for volatility, high-dimensional cross-asset dependence, sentiment-to-volatility modelling, reservoir computing and causal econometrics. The public FGD code and SentiVol link justify artifact-level recovery and reproducibility checks, not claims about alpha or professional deployment.

Recovery: archive the MiQE/F course catalogue and thesis topics; retrieve Gonon, Audrino and Grigoryeva papers, code, data statements and supervisor lineages; test the FGD code in its stated research-only context; recover the SentiVol forecast history if the service returns; map student projects, HSG finance-industry contacts and alumni; and preserve separate states for curriculum, research code, public forecast, employer contact, proprietary data, production deployment and performance.

September 5, 2026 — Cambridge legal-finance AI and an agent-generated asset-pricing paper

• The University of Cambridge 3CL seminar page dates “The Deep Learning of Hedge Funds” to March 17, 2026 and names William J. Magnuson, Professor of Law at Texas A&M University School of Law. Its abstract discusses machine learning and AI across exchanges, derivatives and currency markets, including flash crashes, insider-trading algorithms and adversarial attacks, and frames hedge funds as an important sector in the shift. This is legal and market-structure research, not evidence about a named fund’s system.
• Cambridge University Press’ Deep Learning in Quantitative Trading supplies a technical vocabulary for follow-on searches: financial time-series and supervised learning, time-series and cross-sectional momentum, predictive signals, portfolio optimization, high-frequency microstructure data, finance-specific cross-validation and public code examples. The publisher description is a teaching/reference surface and does not establish manager adoption or investment results.
• The University of Exeter event page lists Andrew Chen of the Federal Reserve Board for a May 15, 2026 seminar on “Hedging the Singularity.” The arXiv record identifies Andrew Y. Chen as author, records the April 18 submission and August 4 revision, and says the paper was generated by AI using a public repository. Its model treats AI stocks as a hedge against displacement risk under incomplete markets; these are the paper’s stated assumptions and results, not a recommendation or a live strategy disclosure.
• Chen’s public Ralph-loop repository describes a repeatable pipeline: a paper specification and economic-background file feed an author-plan agent; an author-improve agent applies the plan; tests check the result; and failed tests trigger another iteration. The README describes optional referee agents, branch-based iteration, and test families covering facts, specification, theory, visuals, writing and build integrity. It reports that the full 25-test example run consumed two $200/month Claude Code subscriptions; this is an author-reported project cost, not a general cost benchmark.

These sources add two academic discovery routes: legal analysis of AI-market structure and a public example of agent-orchestrated finance research. They do not establish that a tracked hedge fund, allocator or central bank approved, deployed or traded the described system. The follow-up queue is to recover the Cambridge recording and underlying article, archive public code and licences, preserve the arXiv revision and repository history, and search the resulting method vocabulary across quant hiring, finance labs, conference talks and firm engineering pages.

September 5, 2026 — BIMSA’s bilingual quantitative-investing and generative-finance course surface

• BIMSA’s Quantitative Investing and Machine Learning course page lists Xing Yan as an Associate Professor and dates the course from September 14 through December 28, 2026. The page describes factor construction, mean-variance optimisation, Barra-style risk modelling, rigorous backtesting, transaction-cost modelling, performance attribution and dynamic portfolio optimisation. It names CRSP, Compustat and TAQ as example financial databases and says students will design, implement and backtest an original quantitative strategy. The page exposes Chinese and English language metadata and public notes/video indicators. This is curriculum and research-training disclosure, not evidence that a professional manager supplied the data, adopted a student strategy or granted live trading authority.
• BIMSA’s Deep Generative Models for Quantitative Finance course records an earlier 2026 course taught by Yan. Its public description names generative adversarial networks, normalizing flows and diffusion/score-based models, and places them in return and scenario generation, portfolio construction, risk-neutral density estimation, derivative pricing and calibration, stress testing, risk management, and market-microstructure/order-book simulation. It explicitly frames financial consistency, probabilistic foundations, stochastic processes and no-arbitrage as constraints on generative modelling. These are model/application clues; the page does not disclose weights, training data, licensing, validation results, simulator-to-live transition or deployment.
• BIMSA’s AI-for-Digital-Economy archive records Yan’s seminar “Machine Learning in Finance: Uncertainty Quantification, Generative Learning, Model Stability, and Their Applications.” The public abstract describes finance-specific uncertainty, dependency structures and tail properties, cautions against generic end-to-end black-box assumptions, and names risk forecasting and portfolio construction. It identifies Yan’s prior Renmin University role, Chinese University of Hong Kong financial-engineering PhD, Chinese Academy of Sciences computer-science master’s degree and Nankai pure-mathematics bachelor’s degree. This is a dated seminar and lineage source; its performance language requires paper-level evaluation.
• BIMSA’s Digital Economy Lab seminar archive adds title-blind recovery vocabulary: robust conditional portfolio decisions via optimal transport, neural importance sampling for option pricing with normalizing flows, text narratives and volatility forecasting for Chinese banks, information flow between asset classes during extreme events, stablecoin-arbitrage fragility, and generative AI meeting data quality. It exposes speakers, dates and public video/PDF markers, creating a Chinese/English paper and transcript queue. It does not establish a shared model, investment-firm collaboration or production use.

Research implication: BIMSA adds a cross-border academic route spanning forecast, risk and portfolio construction rather than only sentiment or news NLP. The concrete model/data queue includes CRSP/Compustat/TAQ workflows, transaction-cost-aware validation, Barra-style risk models, tail-aware generative models, normalizing-flow option methods, diffusion scenario generation, no-arbitrage constraints, order-book simulation, uncertainty quantification, optimal transport and Chinese-bank narrative data. These remain academic hypotheses and talent signals until underlying papers, datasets, code and independent replications are recovered.

Recovery: capture the public BIMSA notes, videos and language metadata; retrieve Yan’s papers, CV and supervisor lineage; recover both course syllabi, assignments and project outputs; identify database versions and point-in-time controls; and search the BIMSA speaker graph, Renmin alumni and finance-industry pages for attributable follow-on evidence. Keep academic teaching, seminar claims, public artefacts, employer relationships, proprietary data, deployment and performance separate.

September 5, 2026 — Warwick’s mathematical-finance programme and Gillmore research network

• Warwick’s 2026 Mathematical Finance MSc page dates the one-year programme to a September 28, 2026 start and describes training jointly across Statistics and Warwick Business School. Its seven core modules span financial statistics, financial econometrics, asset pricing and risk, programming for quantitative finance, stochastic calculus, applications of stochastic calculus, and simulation and machine learning for finance. The page identifies C++, Python and R as programming languages and says the programme concludes with a dissertation. These are programme and talent-pipeline disclosures, not evidence of a named manager’s models or data.
• Warwick’s course-structure page gives a more operational view: the programme combines probability, stochastic processes, numerical methods, derivatives, asset pricing, financial time series, risk management and PDEs with Python, Matlab/R and C++ work. Its public recent-dissertation list includes change-point detection in financial time series, optimal strategies for high-frequency trading, systemic-risk modelling, option-gamma estimation, and spark-spread options for Bitcoin miners. These titles identify research hypotheses and student-artifact routes; they do not validate the methods or show transfer to a professional portfolio.
• Vicky Henderson’s Warwick profile records a current Statistics faculty role, prior work at the Oxford-Man Institute of Quantitative Finance, financial-derivatives teaching, and a connection to the joint Mathematical Finance MSc. This is a named professor and lineage route. The profile does not establish that any Oxford-Man or Warwick research method was adopted by a fund.
• The Gillmore Centre for Financial Technology provides a related finance-AI lab surface. Its public description spans artificial intelligence, machine learning, data science, finance, behavioural science and engineering, and states research pillars around consumers, firms, AI/ML investment and process use, and wider societal effects. Its people page names Head of Centre Ram Gopal, Lead Academics Moris Strub and Kalina Staykova, and other centre personnel; the page does not identify a hedge-fund model, proprietary data feed or live deployment.
• Warwick’s 2026 PhD Workshop on Quantitative Finance and FinTech adds a dated conference recovery route. Its parallel tracks cover machine learning, text and credit risk; AI, learning and financial decision-making; machine learning for volatility and derivatives; information aggregation and price discovery; factor pricing and momentum; high-frequency dependence; and intraday trading. The programme names Roman Kozhan, Moris Strub, Xinyi Zhang, Amir Hosseini, Bazil Sansom and Mohammad Nourbakhsh in organising roles. It exposes a paper, speaker and recording-recovery queue, not evidence that the listed topics were deployed by a manager.

Research implication: Warwick adds a UK academic pipeline linking stochastic finance, market simulation, machine learning, high-frequency trading, volatility/derivatives, change-point detection, systemic risk, and finance-technology governance. It is useful for finding dissertations, supervisors, paper code, student placements and practitioner interfaces. The programme and centre pages do not establish model ownership, licensed data, investment authority, production use or performance.

Recovery: capture the 2026/27 module descriptions and dissertation archive; retrieve Henderson’s papers and Oxford-Man lineage; recover the Gillmore people, seminar and workshop recordings; map student placements and supervisors from direct sources; and keep programme exposure, student work, centre research, employer contact, proprietary data, deployment and performance separate.

September 5, 2026 — Bath’s hedge-fund classification research and finance/data-science pipeline

• Emmanouil Platanakis’s University of Bath research-portal profile identifies him as an Associate Professor of Finance focused on portfolio theory, financial forecasting, asset pricing, machine learning, investments, fintech, cryptofinance and estimation-risk management. The profile records a PhD in Finance from the University of Reading’s ICMA Centre, an MSc in Mathematics from Southampton, and an MEng in Electrical and Computer Engineering from Aristotle University of Greece. It also links current working papers on commodity-inflation risk, equity-premium prediction and data-mined anomalies. This is first-party faculty, research and academic-lineage evidence; it does not establish a hedge-fund affiliation, proprietary data access or production use.
• Bath’s research record for “Hedge Fund Performance, Classification with Machine Learning, and Managerial Implications” identifies Platanakis, Dimitrios Stafylas, Charles Sutcliffe and Wenke Zhang as authors of a peer-reviewed 2025 British Journal of Management article. Its abstract says machine-learning techniques are used to test whether reported hedge-fund strategies align with realised performance and to examine how classification affects managerial decisions, abnormal returns, risk exposures and benchmark construction. The public record does not specify the full model comparison or expose the underlying database in enough detail to treat the result as a deployable signal; the paper itself and its data/method appendices require separate recovery.
• Bath’s 2026 taught-postgraduate catalogue lists Finance, Finance with Risk Management, Advanced Machine Learning, and Financial Mathematics with Data Science programmes. The university’s maths-with-data-science guide explicitly connects randomness, risk, machine learning and mathematical modelling to financial-market behaviour, and says Financial Mathematics with Data Science includes an industry-placement option. Bath’s Economics MSc page also lists programming and machine-learning applications in economics and finance, with dissertation or limited consultancy-project routes. These are current programme and talent-pipeline disclosures, not evidence that a fund recruited from, sponsored, or adopted work from the programmes.

Research implication: Bath adds a route for searching beyond sentiment and text models into manager/strategy classification, abnormal-return and risk-exposure measurement, benchmark design, data-mined anomaly research, portfolio forecasting, and the interaction between mathematical finance and machine learning. The useful cross-correlation queue is Platanakis’s coauthor and supervisor graph, Reading ICMA Centre and Southampton lineage, Bath dissertation and placement artefacts, and the paper’s actual model/data/validation design. Do not infer a professional manager’s strategy from this academic record.

Recovery: obtain the open-access paper and supplementary materials, record the precise features, labels, model families, sample period and point-in-time controls, and inspect code/data statements; retrieve Platanakis’s linked papers and doctoral work; capture Bath module pages, placement language, dissertations and seminar recordings; and search the named authors, programme alumni and employer pages for attributable follow-on evidence. Keep academic method, classroom exposure, placement, employer contact, proprietary data, production deployment and performance in separate evidence states.

September 5, 2026 — Columbia MAFN’s faculty directory exposes a university–industry quant graph

• Columbia’s Mathematics of Finance faculty directory is a concentrated personnel and curriculum surface. The programme says it combines mathematics, statistics, stochastic processes, numerical methods and financial applications, and that many applied courses are taught by finance professionals. Its faculty page currently lists Alberto Botter as an AQR Capital Management partner responsible for equity long-short and tax-aware products, portfolio construction, risk management and portfolio implementation; the page also identifies his Columbia course as quantitative methods in investment management. This is a first-party biography and teaching connection, not evidence about AQR’s internal models, data or deployment.
• The same directory lists Gordon Ritter as founder and CIO of Ritter Alpha, after portfolio-management roles at GSA Capital and Highbridge’s statistical-arbitrage group. It describes his Harvard mathematical-physics PhD, current research in portfolio optimisation and statistical machine learning, and public work on machine learning for trading and turnover/liquidity/autocorrelation. Columbia lists Eric Yeh, formerly of Morgan Stanley, Deutsche Bank, Tower Research Capital and AllianceBernstein, as President of Vermillion Leaf Capital and an adjunct teaching hedge-fund strategies and risk; the page describes his Harvard mathematics and computer-science degrees. These are personnel and research-lineage records, not independently verified fund results.
• Columbia also lists Harvey J. Stein as a senior vice president in Two Sigma’s Labs group and an adjunct professor, following a long Bloomberg path across quantitative risk analytics, counterparty and credit risk, interest-rate derivatives and quantitative-finance R&D. Laura Leal is listed as a Goldman Sachs vice president on the AI Strats team, previously in GSAM’s Quantitative Investment Strategies group, teaching “Mathematics of Generative AI” as a Columbia adjunct and holding a Princeton ORFE PhD. The page does not reveal the firms’ systems, models, permissions, vendor relationships or performance.
• Additional practitioner-academic bridges include Luca Capriotti, Global Head of Quantitative Strategies Credit at Credit Suisse, whose listed research interests include credit models, computational finance, machine learning and adjoint algorithmic differentiation; Jaehyuk Choi, MAFN programme director, with nine years as a fixed-income quant at Goldman Sachs and a co-founder/adviser role at quants.net; and Alexei Chekhlov, listed as Head of Research and Partner at Systematic Alpha Management. The directory also records course-level interfaces for numerical methods, quantitative methods, generative AI and hedge-fund strategies and risk. Course and biography evidence should not be converted into firm strategy claims.
• Columbia’s programme page says MAFN has more than 1,400 alumni in 42 countries and invites employers to engage with students and alumni. That is a recruiting-network claim made by the programme. It is useful for an alumni and job-history recovery queue, but it does not identify which alumni worked on AI, which employers supplied projects, or whether any academic method reached production.

Research implication: this directory changes the unit of analysis from an isolated professor to a repeatable university-to-industry graph. It supplies names, titles, prior employers, course labels, degree lineages and research vocabulary across systematic equity, statistical arbitrage, credit modelling, portfolio implementation, risk, generative AI and hedge-fund strategy. The page does not establish that AQR, Two Sigma, Goldman Sachs, Ritter Alpha, Systematic Alpha, Vermillion Leaf, Credit Suisse or any other listed organisation used a Columbia course, shared data or deployed a named model.

Recovery: archive the faculty page and MAFN programme pages as dated snapshots; recover current syllabi, practitioner-seminar recordings, faculty papers, alumni pages and public code; verify current titles through first-party employer pages where available; and keep teaching, adjunct affiliation, prior employment, advisory work, research, public code, proprietary data, production deployment and performance as separate evidence states.

September 5, 2026 — Oxford’s Jeremy Large links academic market microstructure to former Tudor portfolio management

• Oxford’s profile for Jeremy Large identifies him as a Senior Research Fellow at INET Oxford who teaches at the Saïd Business School. The profile lists interests in financial economics, machine-learning techniques, consumer demand and agent-based modelling, and says he has traded global macro, FX, listed equities and commodities. It records a fellowship at All Souls College, a move into the hedge-fund industry, and a 2013 appointment at Tudor Investment Corporation as a Quantitative Portfolio Manager. This is an Oxford employment-history statement; it does not show Tudor’s data, models, authority or results.
• The profile’s public publication list includes work on quadratic variation with price-tick and microstructure controls, limit-order-book market clearing and inefficiency, moving-average estimators of integrated variance, pro-rata matching in one-tick futures markets, and electronic limit-order-book resiliency. Its working-paper list includes a very large demand-system estimator using negative sampling from machine learning to handle a massive choice set. These papers expose research hypotheses around microstructure, execution, demand estimation and high-dimensional choice; they do not demonstrate that the methods were used at Tudor or remain in a current trading process.

Research implication: this route connects market-microstructure research, machine-learning estimation and agent-based modelling to a named former quantitative portfolio-management role, while preserving the distinction between published research and firm practice. Follow the Oxford publication records, coauthors, conference appearances, archived Tudor biographies and any public code without inferring continuity of employment or strategy.

Recovery: capture the Oxford profile, All Souls biography and linked paper records; retrieve papers, data descriptions, code and conference talks; verify the Tudor dates and later professional history from independent first-party sources; and keep publication, teaching, former employment and verified deployment separate.

September 5, 2026 — PKU–Baruch, HKU SPACE, and Imperial expose implementation-oriented quant curricula

• Peking University NSD’s 2026 NSD–Baruch MFE Summer Program notice is a bilingual China–New York talent and method route. The August 10–15 online programme lists machine learning for finance—tree methods, random forests, dimensionality reduction, support-vector machines and generative models—alongside options trading and arbitrage, volatility trading, and a technical seminar on relative-entropy-regularised optimal order execution. The seminar topics include stochastic differential games, HJI equations and linear-quadratic control. It names Dan Stefanica, Giulio Trigila, Tai-Ho Wang and Ken Abbott as Baruch MFE faculty and says participants receive buy-side/sell-side industry exposure and networking with Baruch alumni. These are programme claims; the notice does not identify participating firms, projects, datasets, models or deployment.
• HKU SPACE’s Postgraduate Diploma in Financial Analytics and Algo Trading lists a September 5, 2026 start and six modules: AI and Financial Computing; Financial Analysis and ESG Investing; Financial Risk Analysis and Portfolio Optimisation; Machine Learning for Financial Analytics; Web Scraping and Text Analytics in Quantitative Finance; and Algo Trading and Quantitative Investment Strategies. The page specifies Python and computational methods, text mining and web scraping, portfolio and risk analysis, performance evaluation, strategy optimisation, and group assignments/presentations. It also links public YouTube material, including an “AI and quantitative investment” event. This is a Hong Kong continuing-education and talent pipeline; the page does not expose instructor-level research, employer-sponsored projects, licensed feeds or live authority.
• Imperial’s 2026 Summer School on Machine Learning, Applied Statistics, and Quantitative Finance lists independent modules in modern ML, applied statistics and systematic trading. The page names Johannes Muhle-Karbe and Almut Veraart as course directors and lists Ed Cohen, Sarah Filippi, Johannes Muhle-Karbe, Almut Veraart and Yufei Zhang among the mathematics faculty. The Systematic Trading module (July 6–10, 2026) describes return predictions extracted from past prices, trades and quotes using ML, then a price-impact model to quantify trading costs and test whether paper profits survive actual trading. Its stated learning outcomes include designing and deploying statistical return models, calibrating price-impact models, and combining both for backtests. Imperial’s syllabus uses hedge funds and “smart money” as the application context; it does not establish that any named fund supplied data, hired a participant, adopted the syllabus or achieved a result.
• The Imperial page initially returned 403 to a bare request in this run, but returned 200 with a browser User-Agent. The recovered page is therefore retained as directly captured HTML with a request-header note, while the public synthesis relies only on the page’s visible programme and faculty content. This is a reproducibility detail for future media and university-page acquisition, not evidence about any investment firm.

Research implication: these programmes add implementation vocabulary that title-only hedge-fund searches miss: generative models, options arbitrage, volatility trading, relative-entropy execution, HJI control, web scraping, text analytics, point-in-time performance evaluation, return prediction, price impact, transaction costs and paper-profit survival. They are useful for finding syllabi, instructors, student projects, recordings, alumni and recruiting events. None establishes a firm’s internal system, data licence, model ownership, trading authority, production deployment or performance.

Recovery: archive the PKU–Baruch bilingual notice and Baruch faculty pages; recover its industry-session roster and any recordings; capture HKU SPACE syllabi, instructor biographies, linked YouTube videos and public project material; preserve the Imperial page and browser-header acquisition note, faculty pages, module pages and any recordings; and keep curriculum exposure, student work, employer contact, proprietary data, deployment and performance separate.

September 5, 2026 — UZH/ETH’s thesis archive is a supervisor-linked model map

The UZH/ETH quantitative-finance thesis archive warrants a deeper extraction than its headline topic list. In 2026 it records a multi-agent system hedge-fund project supervised by Markus Leippold; reinforcement learning for market making with informed and uninformed trader mixtures supervised by Patrick Cheridito; neural-operator deep calibration of rough-volatility models; financial-event forecasting on dynamic knowledge graphs with large language models; machine-learning prediction of German asset-swap spreads; and forward-looking equity-correlation forecasting. The archive supplies dates, presenters, supervisors and locations, but not thesis files, code, data rights, conclusions or live authority.

The 2025–2021 entries make the cross-modal route clearer: sectorial multi-transformer attention networks for volatility forecasting; realised-volatility forecasting using implied-volatility-surface data; open-source and proprietary company-level financial-news sentiment scores; deep RL for optimal execution; RL high-frequency market making; microstructure-driven crypto prediction with tick data; a 2021 earnings-call-transcript project for forecasting company fundamentals in factor-based quantitative investing; GAN scenario generation; ClimateBERT; image-based technical-pattern recognition; deep portfolio optimisation; meta-labelling; and deep no-arbitrage asset pricing. These titles identify public research hypotheses and personnel/supervisor links, not validated trading results.

The UZH/ETH track overview and Spring 2026 course schedule provide the programme context: machine learning for finance and complex systems, quantitative asset management and systematic investing, investments using pattern-recognition and ML tools, and optimisation methods. This creates a concrete recovery graph from programme → supervisor → thesis presentation → paper/code/data → public profile or later employer. It does not establish that a professional manager sponsored or adopted any project.

Research implication: UZH/ETH adds LLMs and knowledge graphs, earnings-call text, sentiment, images, volatility surfaces, order-book execution, market making, multi-agent systems, GAN scenarios, ClimateBERT and no-arbitrage constraints to the academic idea map. The archive is especially useful for locating supervisors and students connected to specific modalities. Keep thesis title, presentation and academic method separate from model ownership, employer contact, production deployment and performance.

Recovery: archive the full 2026 and historical tables; locate thesis PDFs, slide decks, recordings, code, datasets and data statements; map presenters and supervisors to papers, LinkedIn/X profiles and later employers; and test recovered artifacts for look-ahead, survivorship, corporate-action, transaction-cost, multiple-testing, model-version and licensing issues.

September 5, 2026 — Family-office AI workshops and student funds expose allocator workflows

• Simple’s London family-office AI workshop was dated July 2, 2026 and designed as a limited, 2.5-hour session for family-office principals and executives. Its public agenda names operations, data management, investment analysis, reporting, compliance and family communication as application areas, and identifies Francois Botha, Jimmy Otterdijks and Oliver Yorke as hosts. This is an allocator-side discovery route and a title-blind vocabulary source; it does not identify attendees, client systems, model providers, datasets, permissions or outcomes.
• The University of Arizona Eller College’s Student Managed Investment Program describes three tiers with fund values dated May 2026. It identifies Daniel Kinnear as programme director; lists Anne Anderson, David Zynda and Sheila Hanley in faculty roles; and describes an $8.8 million senior fund split between a $3.8 million individual-equity portfolio and a $5 million semi-active enhanced-index strategy. The page says the master’s equity fund uses a more quantitative factor-return focus, while the senior capstone routes 230-plus student valuation projects into a filtered list of possible small- and mid-cap overweights. This exposes a concrete research-to-portfolio governance pipeline, not evidence of AI use, a hedge-fund relationship, proprietary data, live model automation or audited performance.
• St. John’s Tobin College’s 2026–27 SMIF notice says the Student Managed Investment Fund requires registration in FIN 4327 Managing Investment Funds and reports more than $13 million in assets. The notice is useful as a New York talent and student-capital route, but it gives no model inventory, data vendor, faculty research agenda, holdings, code, AI use or performance attribution.

Research implication: allocator and academic surfaces should be searched for workflow nouns—not only “AI” or “hedge fund.” The most useful terms here are family-office operations, data management, investment analysis, reporting, compliance, enhanced indexing, factor returns, capstone valuation, portfolio governance and student-managed capital. Treat workshop agenda, classroom process, faculty supervision, employer contact, proprietary data, production deployment and performance as separate evidence states.

Recovery: retrieve Simple’s related workshops, reports, team biographies and any public recordings; obtain Arizona’s fund reports, investment-policy materials, faculty biographies and capstone artefacts; and follow St. John’s SMIF faculty, alumni, annual reports and employer events. Do not infer professional-fund practice from educational portfolio activity.

September 5, 2026 — Academic agent research, live model evaluation, and finance-programme infrastructure

• Princeton economist Markus Brunnermeier’s August 2026 working paper frames agentic AI as an “asymmetric understanding” problem: agents may learn how humans respond while humans cannot reliably anticipate agent behaviour. The paper connects that asymmetry to harder-to-read prices and central-bank communication, and proposes preserving human fallback channels and simpler policy rules. This is a macro-finance and market-structure hypothesis, not evidence about any fund’s implementation.
• “Can AI Do Financial Research?” by Huan Liu, Miao Liu, Zhizhe Liu and Danqing Mei describes a human-designed, auditable research environment in which LLM agents propose, test, critique and revise interpretable accounting-formula signals. The abstract reports 280 candidate signals across eight theme pairs and seven generations, 159 passing a conventional significance screen, 38 surviving multivariate horse races, and further tests against multiple-testing concerns and 209 published anomalies. These are author-reported working-paper results; the paper’s data, code, model versions, point-in-time controls, costs and independent replication remain open. SSRN returned 403 during this capture, so the abstract metadata and direct PDF route are retained separately.
• HKU Business School’s Artificial Intelligence Evaluation Lab describes an Agentic Trader experiment beginning in April 2026: ten LLMs received the same US$100,000 initial capital, tools, leverage and live market data across FX, the S&P Index and precious metals, with no prescribed trading strategy. The release reports roughly six weeks of trading and a range from approximately +9.9% for Qwen to -15.1% for DeepSeek, while also saying trade frequency and risk did not map directly to returns. The release is internally inconsistent: its key-findings section assigns the highest profit to Qwen, while a later detailed paragraph assigns the highest six-week return to Kimi. Preserve the full report and tables before using any model-performance claim; this is a university experiment, not evidence of a deployable investment system.
• The University of Stirling’s 2026 MSc Finance and Data Analytics combines finance theory with machine learning and AI applications, Python/R, a real-time trading boot camp with AmplifyMe, a student-managed investment fund, and access to Bloomberg, Capital IQ and Morningstar Direct. Its public page names course director Konstantinos Gavriilidis and faculty across data analytics, quantitative methods, stochastic processes, investments and risk; it also lists prior guest connections including Walter Scott, Aberdeen, UBS and SVM Asset Management. This is a programme, data-access and recruiting surface, not proof of those firms’ systems or student-fund AI use.
• Quinnipiac’s May 2026 student-portfolio account reports an approximate $6.1 million portfolio, up from approximately $4.8 million one year earlier, and identifies student fund managers who presented sector reviews and new-company cases to university leadership. The article mentions an AI-driven market backdrop but does not attribute portfolio decisions or performance to AI, name models or vendors, or disclose data and decision logs. Treat the asset and performance figures as university-reported and separate from any AI inference.

Research implication: the academic route now covers three distinct layers—economic theory about agentic markets, auditable agentic research loops, and live model-behaviour evaluation—plus the programme infrastructure that supplies data access and talent. Search faculty papers, thesis archives, course projects and conference recordings for terms such as symbolic accounting formulas, hypothesis-generation loops, point-in-time language models, live FX agents, human fallback, model-risk evaluation, market microstructure, Bloomberg/Capital IQ/Morningstar workflows and real-time trading labs. None of these sources establishes a named hedge fund’s production deployment, proprietary-data rights, investment authority or performance.

Recovery: obtain Brunnermeier’s full paper, discussion slides and Jackson Hole presentation; recover the SSRN PDF, appendices, code/data statements and model-version ledger for the LLM research loop; download HKU’s complete Agentic Trader report and tables and reconcile the Qwen/Kimi discrepancy; capture Stirling syllabi, boot-camp artefacts, SMIF reports and guest-session recordings; and archive Quinnipiac’s portfolio outlook, holdings methodology and performance report if made public.

September 5, 2026 — AI-authored finance research, Boston training pipelines, and agentic-trading security

• UCLA Anderson’s Human × AI Finance call for papers required a finance paper begun after February 18, 2026, encouraged extensive AI assistance from literature review through modelling and writing, and had AI agents select four papers for presentation at the Fink Center Conference on Financial Markets. It required machine-readable submissions and a description of the authors’ AI workflow. This is a controlled research-process experiment, not evidence that a fund delegated research or that the selected papers produced investable signals.
• Northeastern’s 2026–27 Boston MSF Quantitative Finance catalogue combines investment analytics, applied statistics/econometrics, programming, derivatives and risk with applied projects using real company data, Bloomberg training, co-op or organization projects, and participation in the student-managed 360 Huntington Fund. The public catalogue exposes a Boston talent and applied-workflow route but does not identify models, AI use, data licences, employer project sponsors, live authority or performance.
• FinRegLab is a nonprofit research and convening surface spanning AI in financial services, alternative data in credit underwriting, and small-business financial health. Its public site links projects, publications, testimony, podcasts, conferences and an AI Symposium scheduled for November 18, 2026. The organization’s stated work is policy and empirical analysis, not hedge-fund research; its sources are nevertheless useful for tracking model-risk, fairness, explainability, alternative-data and financial-agent vocabulary outside fund marketing.
• “Poisoning Agentic Alpha” studies adversarial vulnerabilities in multi-agent trading systems. The paper decomposes a pipeline into Analyst, Researcher, Trader and Risk Manager roles, evaluates role-specific attacks across four communication topologies, five assets, two model backbones and two target directions, and introduces an Adversarial Signal Preservation Score. Its abstract says no architecture is inherently robust. This is a security research result from an arXiv preprint, not evidence of an identified firm’s architecture or an actual compromise.

Research implication: academic discovery now needs separate routes for research generation, portfolio education, financial-AI governance and adversarial resilience. Add AI-assisted paper production, machine-readable research provenance, student-fund applied projects, alternative-data underwriting, model fairness, prompt/data poisoning, role-specific attack surfaces, communication topology and signal-preservation metrics to the cross-firm search vocabulary. Preserve paper claims, classroom activity, policy research, security experiments, firm disclosure and production deployment as distinct evidence states.

Recovery: obtain the UCLA submissions and conference recording; capture Northeastern’s 360 Huntington Fund page, syllabi, project sponsors and alumni paths; crawl FinRegLab’s AI Symposium, podcasts, projects and funding disclosures; and archive the arXiv PDF, code, attack prompts, benchmark assets, affiliations and revision history. Do not infer fund adoption from academic or policy research.

September 5, 2026 — Rotman and UBC add a measurable academic route into hidden data capacity

• Rotman’s 2025 finance and accounting grants page describes Zigang Li’s PhD project, “Data Acquisition in the Age of AI and Market Efficiency.” The abstract proposes measuring fund-level data-processing sophistication from staffing profiles, disclosed algorithms, NLP analyses of regulatory filings, vendor partnerships, surveys, and text-mined procurement records, then joining those measures to 13F holdings and market-microstructure variables. It proposes testing whether technologically advanced investors tilt toward data-rich firms and whether those flows affect price informativeness. This is a grant description of a research design, not a published result or evidence about a named fund.
• The Rotman route is useful because it turns vague “AI capability” into observable research variables: recruiting and staffing language, named methods, filing text, vendors, procurement signals, holdings disclosures and microstructure data. The design still requires point-in-time controls for 13F delays, survivorship, vendor-name ambiguity, staffing noise, multiple testing, data rights and sample construction. The source does not disclose the final sample, model, code, findings or production deployment.
• UBC Sauder’s Allen Hu profile identifies research interests in Big Data and AI in Finance and lists “Persuading Investors: A Video-Based Study,” “Banks’ Images: Evidence from Advertising Videos,” and “Financial News Production.” It also records Tsinghua engineering and Yale PhD training. This creates a professor and multimodal-research route into visual, vocal, verbal and information-production signals, but does not establish a manager relationship, proprietary data access or live investment use.
• UBC’s finance PhD programme describes faculty research ranging from machine-learning analysis of financial contracts to information and asset pricing, and lists empirical/theoretical asset-pricing training and recent placements including Columbia, UCLA, Berkeley and Toronto. This is a talent and research-pipeline signal, not evidence that a particular student or professor joined a tracked firm or that a method reached production.
• Rotman’s John Hull Financial Innovation Fund states that it supports teaching and research in derivatives, hedging, risk management, climate risk and evolving applications of ML and AI in finance. It names Ing-Haw Cheng and Andreas Park as Rotman FinHub academic directors and lists BMO Global Asset Management, Bourse de Montréal, CIBC, National Bank Capital Markets, RBC and TD among founding donors. Donor support and a university research centre do not establish a donor’s model, data access, deployment or performance.

Research implication: university finance programmes and professor grant pages should be treated as first-class discovery surfaces. They expose research hypotheses, student and faculty lineages, data vocabulary, project governance and potential public artifacts that ordinary firm searches may miss. The correct comparison unit is the evidence state—research design, curriculum, personnel, sponsor, dataset, code, deployment or result—not an assumed league table.

Recovery: obtain Li’s supervisor, paper, data dictionary, staffing labels, procurement corpus, 13F join logic, microstructure variables, code and presentations; recover Hu’s video-study paper and any data/code statements; capture UBC course projects and recordings; and map Rotman FinHub researchers and donor-linked public work separately. Preserve the distinction between academic design, education, sponsorship, former employment and verified firm deployment.

September 5, 2026 — Melbourne adds LLM context limits, trading infrastructure, and an Asia-Pacific research graph

• The University of Melbourne Finance Department’s “Supporting a more robust use of AI in finance” page says Antoine Didisheim and Attila Balogh tested LLMs in corporate finance and asset pricing and documented an “information overload” effect: beyond a threshold, extra context reduced accuracy in their tests. The university page describes two tests, not a universal law of all models or finance tasks, and was first published November 6, 2025. The linked paper, prompts, model versions, context construction, scoring rules and replication remain to be recovered.
• Melbourne’s Centre for Brain, Mind and Markets teaching overview describes a third-year Algorithmic Trading subject combining game theory, market microstructure, algorithm design, coding, statistical analysis and backtesting. It says two-thirds of the course is application-focused and students use a custom Python-based programming platform and a proprietary high-frequency-trading platform. The same page describes Foundations of Fintech using machine-learning models for security valuation and asset allocation, plus finance/ML/quantum-computing teaching. These are educational platform descriptions, not evidence of a professional firm’s systems.
• The 2026 FIRN Melbourne Asset Pricing Meeting is dated October 26, 2026. Its public programme lists Andrea Vedolin as keynote and speakers/discussants from Melbourne, Yale, UNSW, Nanyang, Imperial, Cambridge, Hong Kong, Arizona and other institutions. Topics include news pricing, macro-news price impact, option prices and FX expectations, SDF-based portfolio choice and forecast-agnostic portfolios. This is a conference-recovery and academic-network route; the programme does not identify fund sponsors, proprietary datasets or production use.
• Andrea Lu’s profile identifies her as Associate Professor, Master of Management (Finance) programme director and FIRN vice president, with a Northwestern Kellogg PhD and research interests in empirical asset pricing, market frictions and international finance. Melbourne’s Doctoral Program in Finance describes a five-year coursework-and-thesis route with supervision in asset pricing and portfolio/funds management, derivatives and risk management, financial institutions/regulation and market microstructure. This is a talent and faculty-lineage signal, not evidence of a named manager’s deployment.

Research implication: Melbourne adds three concrete search axes—finance-LLM context-budget testing, hands-on trading education with Python and high-frequency infrastructure, and a dated Asia-Pacific conference graph covering news, microstructure, options, macro information and portfolio construction. Use these to locate papers, code, course artefacts, student placements, recordings and practitioner links while keeping research results, classroom platforms, conference participation, faculty lineage, employer relationships, data rights, deployment and performance separate.

Recovery: retrieve the Didisheim–Balogh paper and appendices, Algorithmic Trading and Fintech handbooks, platform documentation, student projects, FIRN papers and recordings, and public CVs and placements for relevant faculty and PhD students.

September 5, 2026 — ESSEC and HEC expose French academic agent and talent routes

• ESSEC’s Metalab for Data, Technology and Society lists contributors across finance, information systems, statistics, AI, optimisation, data governance and human-machine collaboration. Its public focus areas include “AI Agents for Finance,” described as robust generative and predictive AI solutions for asset managers. The same page gives examples of “AI Acceptability with BNP Paribas” and “Robustness in AI, in partnership with BNP Paribas.” These are institute focus and example-project statements; the page does not identify a deployed BNP Paribas system, model, training set, vendor, evaluation result or investment mandate.
• Metalab’s named academic topics include sequence models, generative AI, reinforcement-learning bias, high-dimensional statistics, robustness, fairness, streaming data, option pricing, algorithmic decision-making, data ownership, privacy and data/AI strategy. Its student-facing “Student IDEAS” material adds a route into reward-model trade-offs, safety design and model bias. These pages are useful for finding faculty, students, project pages and recordings; they do not establish a finance-fund system or production deployment.
• HEC Paris’s February 2026 Master in International Finance account quotes Academic Director Evren Örs describing AI/ML integration, more than 30 mostly practitioner-taught electives, an AI-focused workshop, mandatory basic Python, empirical methods for capital markets or corporate finance, and at least one AI/ML-and-finance elective. It says paired student research papers must involve data collection and analysis under finance-faculty supervision. This creates a concrete finance-theory → coding/empirics → AI workshop/elective → supervised research route, but the page does not publish student papers, datasets, sponsors, code, model providers or professional adoption.
• HEC Paris and École Polytechnique’s Data & Finance announcement describes a joint programme covering statistics, probability, regression, classification, optimisation, machine learning, deep learning, Python, R, blockchain, algorithms, AI applied to finance and technology regulation. It names Olivier Bossard and Erwan Le Pennec in the launch account and describes outside-expert teaching. This is programme-design and talent-pipeline evidence, not evidence of a named fund’s model or data access.
• HEC’s alternative-data and finance-startup article summarizes Thierry Foucault’s work on forecasting horizons and discusses synthetic data, LLMs, GANs and reinforcement learning as possible responses to alternative-data and information-overload problems. It names Lemon AI, Synthera AI and Revaisor in the Creative Destruction Lab context and describes Claire Calmejane’s banking and fintech-investment background. This is HEC publication and practitioner-interview evidence; it is not independent validation or proof of HEC-affiliated production use.

Research implication: French academic pages add agent robustness, acceptability, reward-model and safety vocabulary to the finance search graph, alongside alternative-data horizon effects, synthetic market data, LLM/GAN/RL combinations, empirical student research, practitioner electives and financial-technology regulation. Search these terms across faculty papers, student repositories, employer pages, conference recordings and vendor partnerships while preserving separate evidence states for curriculum, institute focus, partnership mention, startup association, proprietary data, deployment and performance.

Recovery: capture Metalab project pages, contributors, student papers, BNP Paribas project descriptions and public recordings; recover HEC MIF syllabi, AI workshop materials, supervised papers, external-expert rosters and Data & Finance alumni/project artefacts; and verify each named person or organisation independently before making any firm-specific inference.

September 5, 2026 — Professors and finance programmes expose embeddings, graph learning, and production-facing AI validation

• The University of Chicago Becker Friedman Institute’s Asset Embeddings working paper by Xavier Gabaix, Ralph Koijen, Robert J. Richmond and Motohiro Yogo applies the embedding idea to portfolio holdings rather than only text, audio or images. The abstract says recommender-system methods, Word2Vec-style shallow neural networks and BERT-style transformers can represent firms and investors from holdings; it evaluates relative valuation, return comovement and institutional portfolio decisions, and discusses investor classification, performance evaluation, crowded-trade detection, generative portfolios, risk management, stress testing and LLM-generated firm narratives. This is a working-paper research design and reported benchmark programme, not evidence that any tracked fund uses it or has access to the underlying holdings data.
• Stanford GSB’s “Financial Regulation and AI: A Faustian Bargain?”, by Christopher Clayton and Antonio Coppola, describes a graph-based deep-learning model over security-level holdings of non-bank financial intermediaries. The public abstract says the model learns asset and investor representations, incorporates economic priors, covers nearly $40 trillion in wealth, and tests out-of-sample forecasts of intermediary trading behaviour including crisis periods. The abstract reports its own comparisons with traditional stress-event return measures and systemic-risk metrics; those are paper claims pending recovery of the full paper, data construction, code and independent replication. Stanford’s page returned 403 to a bare request during this pass, so the indexed first-party abstract is retained as the public evidence and the access failure is part of the record.
• UCL’s Institute of Finance & Technology industrial-team directory adds a personnel and curriculum bridge. It lists Raad Khraishi as Head of AI R&D at NatWest and an Industrial Professor teaching AI for banking and finance, with public interests in generative AI, LLM evaluation, deep learning, reinforcement learning and responsible AI; José Juan De León Guillamon as an Industrial Professor and senior quantitative researcher working across market microstructure, systematic strategy research, alternative data, econometrics, signals and risk controls; Budha Bhattacharya as Head of Systematic Research at Lombard Odier Asset Management, with public work on ML/AI, sustainability and quantifiable climate/nature exposures; and Luca Cocconcelli as an Industrial Professor with risk analytics experience at LCH/LSEG and M&G Investments. The directory is first-party personnel and teaching evidence. It does not disclose any employer’s models, data permissions, system architecture, investment authority or performance.
• UCL’s AI for Finance Lab page says the lab works on liquidity-risk estimation, risk profiling and risk management; offers AI validation, benchmarking and financial-data analytics design; and names collaborations involving the Saudi Central Bank, Santander UK, Consob, the UZH Blockchain Center and the DLT Science Foundation. UCL’s research-projects page separately describes a NatWest-sponsored offline-reinforcement-learning project led by Raad Khraishi and colleagues for contractual pricing, using historical data with online fine-tuning because live trial-and-error pricing is impractical. These pages establish a lab service line and named project description, not a hedge-fund deployment, model-provider choice, data licence, production authority or measured trading result.
• Andrew W. Lo’s MIT working-publications index provides a separate professor-led route into market dynamics and healthcare-finance forecasting. The index includes high-frequency country-ETF price-discovery networks and a “Debiasing Probability of Success Estimates for Clinical Trials” paper that adjusts recent clinical-trial probability-of-success estimates for a boundary-effect bias. These entries are publication metadata and abstracts; they are relevant research hypotheses for information diffusion, market networks and approval forecasting, but do not establish a fund relationship, investable signal, proprietary dataset or live deployment.

Research implication: the professor/programme layer now exposes several distinct model families and operational questions: holdings as latent asset/investor representations; graph inductive learning for unseen investors or assets; LLM narrative generation over economically structured embeddings; offline RL where online exploration is unsafe; model validation and benchmarking as a service; and market/network or clinical-trial forecasting. These are candidate research avenues, not a league table. The evidence does not support saying that any named hedge fund is using one, that one approach is superior, or that an academic affiliation transfers a model into production.

Recovery: obtain the Chicago and Stanford full papers, appendices, code and data statements; reconstruct holdings timing, missing-entity tests, investor labels, crisis windows and benchmark definitions; capture UCL lab project pages, named researchers, supervisors, employer dates, validation protocols and collaboration terms; recover Lo’s clinical-trial paper and data design; and search the named professors, students, alumni and industrial faculty across seminars, GitHub, job descriptions, conference recordings and employer disclosures. Keep professor research, course exposure, sponsor/project mention, personnel affiliation, proprietary data, production deployment and performance as separate evidence states.

September 5, 2026 — Singapore adds an interdisciplinary AI institute and an industry-linked finance talent pipeline

• NTU’s launch announcement for the Global Institute of Finance, Technology, and Society (GIFTS) is dated May 30, 2026. It says GIFTS is jointly founded by Nanyang Business School and the College of Computing and Data Science, and will work with local and global partners on AI, digital economics, fintech, digital assets, and emerging business and social systems. The announcement identifies Professor Lin William Cong as the institute’s leader and as President’s Chair Professor of Finance, Computing and Data Science. A separate NTU profile says Cong joined NTU from Cornell, where he led the FinTech Initiative, and frames Singapore as a laboratory for innovation and governance in an AI-enabled economy. These pages establish an interdisciplinary research surface; they do not identify a hedge fund, proprietary dataset, model, deployment permission, or trading result.
• NTU’s MSc Finance programme places Machine Learning in Finance in the Fintech elective route. It names artificial neural networks, decision trees, and support-vector machines and says students complete group projects using realistic data-analysis problems. The programme also includes Python preparation and data-science training. This is curriculum evidence, not a production-system or employer-adoption claim.
• SMU’s Young Talent Programme for AI in Finance announcement describes a 13-week Financial Services AI Immersion Programme with a two-week masterclass, term-time course and project work, experiential learning, and internships. It identifies collaboration with the Institute of Banking & Finance and Monetary Authority of Singapore, and names AWS and Bloomberg as partners. The announcement says more than 20 financial institutions are participating, including JPMorgan Chase, DBS and Prudential; it names Randall E. Duran and Emmaline Lim as course leads. This is a public talent-pipeline and partner statement, not evidence of a particular institution’s internal model or a hedge-fund workflow.
• SMU’s June 2026 programme account adds the mechanics: students from Singapore’s autonomous universities can apply; SMU and NTU can recognise the curriculum for academic credit; projects use Bloomberg data and deploy solutions on AWS; and the pilot targeted 100 students with 20 institutions. The page says the pilot began in August 2026. These details expose a university-to-industry data, cloud, project, and placement route; they do not establish that student work entered a firm’s production environment.
• NTU’s finance faculty directory provides additional search terms: Nelson Lau is listed with systematic and high-frequency trading, machine learning in finance, market microstructure, fintech and alternative data; Milind Goel with AI, climate finance, and empirical asset pricing; Eric Tham with AI/ML in finance and commodities finance; and Wang Xin with market microstructure, financial intermediation, macro-finance and fintech. The directory also lists Will Cong with joint finance and AI/Data Science appointments. These are faculty-interest and personnel signals, not claims about employer deployment.

This Singapore route adds a distinct academic-to-industry pipeline to the professor/programme map: interdisciplinary research on AI and markets; finance curriculum using ML and realistic projects; government- and industry-linked student placements; and faculty routes spanning systematic trading, market microstructure, climate finance, alternative data, digital assets, and AI governance. Search GIFTS working papers, partner announcements, project briefs, student submissions, Bloomberg/AWS usage terms, internship rosters, and later employer biographies. Keep course exposure, project participation, partner presence, employment, production access, and performance as separate evidence states. See the capture note.

Recovery: obtain GIFTS’s current research themes, people, partners, events, papers, and recordings; obtain the NTU course outline and project rubric; archive SMU’s pilot curriculum, institution roster, data/cloud governance, student projects, internships, and outcomes; and independently verify every named participant before linking them to a tracked firm. Do not infer professional-fund practice from academic curricula or student projects.

September 5, 2026 — professor and finance-programme routes for model ideas and talent

The academic search added several useful surfaces that do not use hedge-fund language. They should be read as idea, training, and personnel evidence—not as evidence of any tracked manager’s implementation.

• Johns Hopkins Carey’s Sudip Gupta biography names AI/ML, auctions, ESG, healthcare and fintech as current research and teaching areas. It lists alternative credit scoring with machine learning, alternative-data ESG ratings and portfolio formation, generative AI, nowcasting with alternative data, and current teaching in Machine Learning for Finance. The profile gives a Wisconsin economics PhD, prior Fordham MSQF programme-director work and faculty appointments at Indiana, the Indian School of Business, NYU Stern and Maryland. These are professor and publication routes; they do not establish a named manager’s model, data rights or deployment.
• Johns Hopkins Carey’s Frank Fabozzi biography records prior teaching at Yale, MIT, Princeton and EDHEC and 2025 publications/books on simulation, optimisation and machine learning for finance. The page lists work on ML-enhanced Markowitz selection, GAN-based anomaly detection, spread prediction and AI-based startup-investment evaluation. This is a useful paper and seminar-recovery route, not evidence that a student fund or professional manager uses those methods.
• Quinnipiac’s Fall 2027 finance faculty posting explicitly lists AI/ML in financial decision-making, alternative data, computational finance, GenAI/LLMs in finance, predictive modelling and AI governance among relevant areas. It also names an Analytics & Applied AI Lab, a Financial Technology Center, the GAME student-run finance forum, and a university-stated $6 million student-managed fund. The posting is recruiting evidence; it does not prove a hire, AI use by the fund, a partner firm, live authority or performance.
• Sacred Heart’s Financial Technology & Analytics programme says finance and analytics courses will incorporate discipline-relevant AI applications from Fall 2026. It describes model comparison, applied finance-analytics projects, generative-AI pilots, Google and IBM AI credentials, and a separately supervised Student Managed Investment Fund with sector analysis, equity research and investment-professional advisers. The page does not connect the fund’s decisions to the AI coursework or identify a production model.
• UTA’s finance catalogue separates FINA 5345, “Artificial Intelligence and Other Technologies in Finance,” from FINA 5328, Student Managed Investment Fund. The AI course emphasizes applications in asset management, credit and compliance, strategic evaluation, responsible innovation and adoption oversight; the fund course covers security research, trading and economic forecasting. This separation is a useful control for distinguishing technology governance from portfolio decision evidence. Neither course description identifies a fund model, dataset, employer project or performance.
• Winona State’s repository entry names Pat Paulson and Larry Schrenk and describes chatbot use in finance courses, including investments and the Student Managed Investment Fund. The abstract includes student engagement, feedback, challenges and ethics. It does not identify a provider, prompts, fund permissions or investment outputs, so it belongs in the instructional-AI branch rather than the manager-capability branch.
• Cal Poly Pomona’s MS in Financial Analytics page describes Python/R modelling, machine learning in finance, portfolio management, real-world financial datasets and a Bloomberg financial-markets room. Its related programme description names CRSP, Compustat, WRDS, a student-managed fund and an asset-management practicum. This is a concrete training and data-environment route; it does not show which project used which dataset or that student work entered a live portfolio.

The resulting idea queue is broader than news sentiment: alternative-data credit and ESG, nowcasting, fairness and bias testing, holdings and portfolio optimisation, anomaly detection, spread prediction, model comparison, responsible-AI adoption, and the boundary between chatbot assistance and human investment accountability. The dated academic capture note keeps papers, syllabi, student funds, faculty roles, data access, deployment and performance as separate evidence states.

Recovery: obtain the named professors’ papers and CVs; recover Quinnipiac GAME agendas and fund policy; retrieve Sacred Heart and UTA syllabi and project briefs; locate Winona’s slides and evaluation material; and capture Cal Poly Pomona’s instructors, practicum outputs, access terms and graduate destinations. Do not infer firm adoption or relative quality from academic exposure.

September 5, 2026 — Cornell and LSE connect professors, sovereign-capital research, and finance-data training

• Cornell Duffield Engineering’s profile for Marcos López de Prado identifies him as a Visiting Professor in Operations Research and Information Engineering and states that he is Global Head of Quantitative R&D at the Abu Dhabi Investment Authority, a founding board member of ADIA Lab, a research fellow at Lawrence Berkeley National Laboratory, and a Cornell Professor of Practice. The page describes his public work as financial machine learning and statistical inference. This is unusually direct academic-to-sovereign-capital personnel evidence, but it does not reveal ADIA’s models, data rights, decision authority, internal collaborators or performance; Cornell’s page is a source for the role descriptions rather than an independent audit of the employer claims.
• LSE’s Statistics doctoral programme describes a three-to-four-year MPhil/PhD route with research groups in Data Science, Probability in Finance and Insurance, and Time Series and Statistical Learning. LSE says these groups span AI, machine learning, statistical inference, quantitative finance and financial statistics; it also describes department seminars, reading groups, Digital Skills Lab support, restricted-data access through the library, and first and second supervisors. This is a research-talent and lineage surface: it supports recovery of supervisors, theses, code and industry transitions but does not establish a fund connection or proprietary-data access for any student.
• LSE’s 2026/27 MSc Financial Statistics regulations expose a broad finance/ML vocabulary in one formal programme: unsupervised ML, Bayesian ML, deep learning, reinforcement learning, graph representation learning, computational text analysis and LLMs, social-network analysis, market microstructure, computational methods in finance, stochastic simulation and calibration, causal inference, derivatives, fixed income and financial risk. The programme requires a dissertation and limits places on some options. This is a curriculum map and student-artifact discovery route, not evidence that the listed methods are used by a named manager or produce a live signal.
• LSE’s announcement of its MSc Accounting and Data Analytics says the first cohort begins in September 2026 and combines accounting, valuation and investing with AI, ML and predictive analytics. LSE describes hands-on projects using live financial data for identifying mispriced securities, detecting corporate fraud and evaluating M&A, and names Maria Correia as Head of the Department of Accounting. These are university programme claims and concrete search terms for accounting-data research, not a hedge-fund partnership, data licence, model specification, student output or production deployment.

Research implication: this cluster adds three linked but separate routes. Cornell supplies a named person whose public biography spans financial-ML teaching, quantitative R&D at a sovereign allocator and an external computational-science lab. LSE supplies a doctoral/supervisor graph and a formal course vocabulary that joins graph learning, LLM text analysis, RL, causal inference, microstructure, risk and finance. The new accounting programme adds a non-quant title-blind route into security mispricing, fraud and transaction analysis. These routes indicate where to search for papers, supervisors, student projects, job transitions and industry sessions; they do not support saying that any firm has adopted a method or that one method is better than another.

Recovery: capture López de Prado’s Cornell CV, papers, public talks, ADIA/ADIA Lab pages and dated role history; recover LSE doctoral supervisors, ST405/ST449/ST455/ST456/ST457/MY459/ST461/MA435 course guides, dissertations, seminars and student code; retrieve the accounting programme’s detailed structure, launch video, project briefs, faculty and employer connections; and keep academic affiliation, employer role, sponsor, data permission, model ownership, deployment and performance as separate evidence states. See the dated capture note.

September 5, 2026 — Chicago Booth adds a feedback-loop risk to earnings-call signals

• The University of Chicago Knowledge repository record identifies Pietro Ramella’s 2026 Booth accounting dissertation, “Better AI, Worse Disclosures? The Unintended Consequences of NLP on Financial Reporting.” The record describes a model in which investors use algorithms to process reports and managers can tailor language to misdirect those algorithms. It says the empirical work uses U.S. earnings conference calls and reports that after ChatGPT-3.5’s November 2022 release, managerial tailoring increased while textual informativeness declined, with the decline concentrated in prepared speeches rather than spontaneous Q&A. These are dissertation claims and require replication before use as an investable result.
• The repository exposes a downloadable dissertation PDF. Its conclusion describes a 2007–2024 panel of U.S. earnings calls, constructs standardized textual surprise by residualizing linguistic dissimilarity against observable information events, and says the measure is validated against abnormal returns. It reports a gradual post-ChatGPT-3.5 increase in managerial tailoring and a simultaneous decline in textual informativeness, concentrated in prepared remarks; it reports no corresponding decline in spontaneous Q&A. The PDF is the primary artifact for the sample, definitions, robustness checks and limitations.
• Chicago Booth’s recent PhD graduates page lists Ramella’s 2026 dissertation and placement at the University of Miami. A Booth AI-and-finance research account separately records Ramella’s earlier BERT-based work with Raghuram Rajan and Luigi Zingales on more than 8,000 shareholder letters from 1955–2020. These records establish academic lineage and prior text-model research, not a hedge-fund relationship or production deployment.

This route adds a feedback-loop risk to text-alpha research: as automated language processing improves, managers may change prepared language in response, while spontaneous Q&A can provide a different information channel. Earnings-call research should preserve speaker role, prepared-versus-Q&A segmentation, call clock, model revision, publication time, textual-surprise construction, and independent market joins, then test sector, issuer, time, model-adoption and multiple-testing controls. The dissertation does not identify a fund, vendor, model deployment, proprietary dataset, trading authority or performance net of costs. See the capture note.

Recovery: archive the PDF and repository metadata; extract the exact textual-surprise definition, event controls, model choices, sample exclusions, prepared/Q&A labels and robustness tables; locate code, data construction notes and seminar slides; and compare the mechanism against privately retained earnings-call datasets without inferring firm adoption.

September 5, 2026 — a Columbia-trained quant researcher links public finance-ML artifacts to a current Millennium role

• Jing “Jimmy” Guo’s personal academic page self-reports a Columbia Operations Research PhD, a University of Virginia Statistics MS, and a “Quantitative Researcher, Millennium Management, 2019 - now” role associated with “Alternative Data/ Alpha Research.” The same page lists prior quantitative and trading-research roles at Goldman Sachs, Guggenheim Partners, KCG, PNC and JPMorgan. Because this is a self-maintained biography, the employment history is a lead requiring independent confirmation, not a verified current headcount record.
• Guo’s public research page lists stochastic control, Bayesian learning and optimization applied to behavioral finance and financial engineering. It links work on order-book dynamics with hidden Markov models, electronic market making, high-dimensional financial PCA and Markov-regime-switching stochastic volatility, including public R code and plots for some projects. The data-science page and linked Machine Learning in Finance repository expose teaching notebooks and slides covering algorithmic trading, price prediction, limit-order-book price impact, deep-learning stock returns and trading strategies.
• This is a useful professor/programme-adjacent discovery route because it connects academic lineage, public educational artifacts, alternative-data vocabulary and a self-reported practitioner role. It does not disclose Millennium’s model inventory, team assignment, data suppliers, proprietary code, production permissions, investment authority or performance, and the public notebooks must not be treated as employer technology.

Research implication: personnel searches should join current or historical firm-role strings with public course repositories, code, research pages and supervisor networks. The candidate ideas here span alternative data, order-book state modelling, market making, regime switching, Bayesian learning and stochastic control; they are public research surfaces, not evidence of any firm’s adoption or of relative quality.

Recovery: verify the role and dates independently; recover the CV, paper versions, notebook commit history, data provenance, conference talks, public professional profiles and course materials; then map supervisors and collaborators to papers and other employer records while keeping self-reported employment, academic work, public code, proprietary data, deployment and performance separate. See the dated capture note.

HKU and MENAP: benchmark design and cost-aware financial agents

• HKU Business School’s Noisy Financial Machine Learning event page dates the seminar to September 3, 2026 and names Professor Federico Bandi, James Carey Endowed Professor at Johns Hopkins Carey Business School. Its abstract argues that a zero-predictability benchmark can be too conservative for operational cross-sectional prediction because apparent predictability can partly reflect level or market-timing effects. It proposes the panel historical mean as a separating benchmark and says richer ML models produce meaningful predictions mainly in smaller, value and less-liquid stocks in the described tests. This is a seminar abstract, not a published performance record or evidence of a firm’s implementation.
• Springer’s publisher record for MENAP identifies Yukai Su, Hui Chen and Lailong Luo and dates the open-access article to August 29, 2026. The paper describes an LLM-agent pipeline combining news, market/macro variables and metadata in a news-to-state-to-pricing-to-portfolio workflow. Its offline preference optimisation penalises token usage and refinement steps; the abstract specifies a two-year Wall Street Journal news window aligned to CRSP, Ken French rates and macro factors, a 9-month/3-month/1-year train/validation/test split, portfolio and pricing metrics, and token/step efficiency reporting. These are paper-reported design and outcome claims; full-text, code, exact model versions, point-in-time controls, costs and independent replication remain open.
• The Springer record places Su and Luo at the National University of Defense Technology and Hui Chen at Macquarie University; it reports no known competing financial interests or personal relationships. Hui Chen’s public profile describes Macquarie University and Shanghai University postdoctoral affiliations, a Macquarie PhD supervised by Longbing Cao and Jinyan Li, and research interests in trustworthy multimodal intelligence, uncertainty-aware learning, probabilistic ML, federated learning and temporal modelling. This is a self-maintained academic biography and does not establish a fund relationship or professional deployment.

Research implication: the benchmark route should separate market-timing or level effects from genuine cross-sectional ranking effects, while an agentic news route should report token and refinement-step costs alongside portfolio and pricing metrics. Neither academic route establishes a tracked manager’s research stack, data rights, production deployment, investment authority or net performance. See the capture note.

Recovery: obtain Bandi’s seminar slides or recording and the paper behind the abstract; retrieve MENAP’s full text, supplementary material, code/data statements, model checkpoints or prompts, preference data, transaction-cost assumptions and point-in-time joins; independently verify author CVs, supervisors, coauthors and subsequent employment; and reproduce the stated splits and metrics before drawing strategy conclusions.

LUISS, NTUT and HKUST: model and tooling vocabulary in finance programmes

• LUISS’s 2026/27 Data-Driven Models for Investment course page names Antonio Simeone and Villy Edoardo de Luca and describes a 6-credit, 48-hour English course. Its stated goals include satellite/geospatial data, web data and news/media with LLMs and GenAI; live-trading and systematic-strategy exercises; qualitative and quantitative AI; agentic AI and transformer architectures for financial time-series forecasting. The contents name FinBERT, AlphaAgents, TradingAgents, TimeGPT, Chronos and TabPFN, plus fuzzy logic, neural networks, multi-agent systems and quantum-annealing/Ising-model work. The page describes Python coding, Bloomberg use, a team project, and a week covering techniques said to be deployed at an unnamed UK hedge fund with a Head of Capital Markets guest. This is course and guest-session evidence; the fund, guest identity, datasets, code, permissions, production details and results are not public on the page.
• National Taipei University of Technology’s 2026 AI Applications in Finance syllabus names Chung Chien-Ping and describes a three-credit, 18-week course. The public syllabus covers digital finance, big data, cloud computing, insurance, payments, digital assets, sustainable finance, quantitative trading and generative-AI/AI-agent development. It explicitly mentions OpenAI cloud resources, Codex, Claude Code, chatbot development, project-based work, and a required Taipei fintech-expo case report. This is a Taiwan curriculum and tooling signal, not evidence of an employer’s software stack, student deployment, data rights or investment authority.
• HKUST’s 2025/26 postgraduate Generative AI in Finance course record names Juergen Harald Rahmel as instructor/coordinator and dates the course February 1 to June 30, 2026. It covers LLMs and related technologies in financial-institution operations, text and code generation, staff-augmentation tools, customer chatbots, guardrails, ethics and regulatory context in Hong Kong, Asia and globally. The record is a course description; it does not disclose firms, datasets, assignments, vendors, deployment or results.

Research implication: these routes expand the search vocabulary beyond news sentiment: alternative-data acquisition, geospatial and satellite signals, time-series foundation models, zero-shot inference, structured-data models, agent tool/memory loops, fuzzy and quantum optimisation, cloud/API governance, code-generation tools, chatbot controls and financial-regulatory guardrails. LUISS creates a recovery path from course materials to an unnamed UK-fund guest and practitioner disclosure, but the public course page does not identify the firm. See the capture note.

Recovery: recover LUISS slides, reading list, project brief, guest identity and any recording; inspect model licences, dataset provenance, Bloomberg and alternative-data permissions, backtest assumptions and course outputs; retrieve NTUT project reports and Taipei fintech-expo artefacts under permitted access; and obtain the HKUST syllabus, assignments, practitioner examples and instructor research profile.

September 5, 2026 — NYU independently verifies a quant-and-teaching route without naming the employer

• NYU Tandon’s official biography for Naftali Cohen identifies him as an Adjunct Professor and says he is a senior quant in the hedge-fund industry specializing in systematic, market-neutral equity trading at mid-frequency with data-driven approaches. It also records graduate-level teaching in data science and quantitative finance at NYU Tandon and Columbia IEOR, and a 2013 Applied Mathematics PhD from NYU’s Courant Institute.
• Cohen’s self-maintained professional page names Millennium Management as a current employer and adds a detailed role chronology, but NYU’s official page does not name Millennium. The sources therefore support an official academic/practitioner description plus a separate self-reported employer lead; they do not support treating the Millennium attribution as independently confirmed.
• This route is useful for recovering finance-programme syllabi, public teaching material, research papers, code and dated personnel evidence. Neither page discloses an employer’s models, data suppliers, production permissions, investment authority or performance.

Research implication: current personnel maps should distinguish an official institutional biography from a person-maintained employer history. The public method vocabulary here is systematic equity research, market neutrality, mid-frequency horizons and data-driven modelling; it is a discovery route, not evidence of a specific manager’s implementation.

Recovery: capture the NYU/Columbia course pages and syllabi, verify employer dates through independent first-party or professional records, and recover talks, papers, code and public profiles while keeping official role, self-reported employment, corroborated employment, proprietary data, deployment and performance separate. See the dated capture note.

September 5, 2026 — Dartmouth, Chicago, and CUNY add agentic, deployment, and international-finance research artifacts

• Dartmouth’s public thesis-proposal page for Junyan Cheng describes autonomous LLM-agent research from theory through systems engineering. The proposal names a Sense–Plan–Act world model trained on time series and text, soft propositional reasoning for forecasting, empirical asset pricing as a non-stationary testbed for long-term memory, distributed genetic programming for model discovery, and an open-source low-level language-model framework. Its committee includes Peter Chin, George Cybenko, Soroush Vosoughi, Kyle Richardson and Jay Stokes. This is a proposal and event abstract, not evidence of a validated trading system, fund connection or production deployment.
• Rong Bai’s University of Chicago thesis record describes a study comparing zero-shot GPT-4 API inference with supervised fine-tuned BERT models for stock-return prediction from more than 30,000 Dow Jones Newswire articles aligned to firm returns from 1989–2020. The record says it uses rolling-window evaluation and examines accuracy, operational feasibility, cost and scalability, and reports author-provided performance metrics. The thesis file is restricted, so the article/return alignment, model versions, code, costs, transaction assumptions and full robustness results remain unverified.
• Yanran Li’s 2026 CUNY Graduate Center finance dissertation lists Xi Dong as advisor and Lin Peng, Dexin Zhou and Guofu Zhou on the committee. Its public abstract covers 44 non-U.S. countries, a supranational anomaly-to-market framework, and LLM/ChatGPT-based measures of macro-uncertainty extracted from earnings-call transcripts. The dissertation is embargoed until June 2, 2028, so the abstract is the available evidence and its findings require later artifact-level review.

Research implication: the professor/programme route now adds three distinct questions: whether agentic memory and world models can be evaluated in non-stationary markets; how API versus locally tuned models trade off infrastructure and reproducibility; and whether earnings-call uncertainty can be studied across international and supranational markets. These are candidate research designs, not evidence that any manager has adopted them or that one approach is superior.

Recovery: recover Cheng’s thesis/code/recording; obtain Bai’s restricted thesis under permitted access and reconstruct its timing, data joins, model versions, costs and robustness; preserve Li’s embargo date and recover the dissertation when available; then map committees, advisors, students and project names to public papers, code, talks and employer biographies. See the dated capture note.

UZH and SFI: regulated-finance cloud training route

• The University of Zurich Department of Finance account dates the Swiss Finance Institute masterclass to June 3, 2026 and names Professor Markus Leippold, SFI Senior Chair at the University of Zurich, and Tobias Kaymak, a Google Cloud Customer Engineer, as co-leads. The page says the session covers GenAI and LLM fundamentals, financial analysis, risk assessment, client service, hands-on labs, and building and deploying GenAI solutions on Google Cloud.
• The page says the event is part of SFI–University of Zurich cooperation and that employees of SBA member institutions, FINMA and the Swiss National Bank could attend without charge. It describes a four-hour SAQ recertification and responsible-AI/ethical-implementation content. This exposes a regulated-finance talent and cloud-partner route, but not the identity of participating firms, their systems, data permissions, model versions, production use or performance.

Research implication: searches should combine professor, cloud-partner and regulated-institution vocabulary—financial-analysis assistants, risk-assessment workflows, client-service tools, build/deploy labs, responsible AI, SAQ accreditation, FINMA and SNB. The event should be joined to Leippold’s public research and thesis artifacts, but attendance or training exposure must not be treated as deployment evidence. See the capture note.

Recovery: obtain the SFI registration page, slides, labs, recordings, model/tool inventory, Google Cloud architecture, participant or employer disclosures and Leippold’s related research artifacts.

September 5, 2026 — CMU and Cornell expose practitioner-faculty routes into credit and execution research

• Carnegie Mellon’s MSCF profile for Roni Israelov identifies him as a current senior quant credit researcher at Citadel. The page also records prior Principal work at AQR Capital Management managing options portfolios and equity portfolio-implementation research, prior CIO/President responsibility at NDVR, finance-journal publications, and an MSCF advisory-board role. This is first-party programme biography evidence; it does not disclose Citadel or AQR models, data, permissions, authority or performance.
• Cornell’s CFEM & UBS seminar page for Michael Sotiropoulos dates an April 7, 2026 talk on smart-order-router allocation algorithms. The abstract discusses distributing passive orders across multiple limit-order books, constrained routing, online estimation and reallocation. Cornell identifies Sotiropoulos as Managing Director of Imperative Execution and records prior Citadel Securities, Deutsche Bank and Bank of America quantitative-research leadership, visiting professorship at Princeton, Fordham teaching and a Stony Brook theoretical-physics PhD.
• These routes add two distinct idea surfaces: credit-research personnel and execution research built around venue allocation, completion time, online estimation and market microstructure. The public pages do not establish that a former employer used the seminar’s method or that either named firm adopted any disclosed model.

Research implication: academic-to-industry mapping should preserve the difference between a current practitioner biography, a historical employer role, a teaching appointment and an event’s research topic. Follow papers, seminar recordings, course artifacts and dated affiliations before connecting the public method vocabulary to a firm’s internal system.

Recovery: recover Israelov’s publications and course materials; capture Sotiropoulos’s recording/slides, execution-paper versions, code/data statements and curriculum; and independently verify role dates while keeping academic work, employment, proprietary data, production deployment and performance separate. See the dated capture note.

September 5, 2026 — Stanford, Polytechnique and Bryant add academic-to-employer routes

• Stanford’s faculty CV for Markus Pelger lists research topics and selected first placements for doctoral students. Named examples include deep-learning asset pricing with a first placement at Two Sigma; deep-learning statistical arbitrage with a BlackRock placement; machine learning and institutional price impact with a Hudson River Trading placement; latent financial structure and option-price-surface work with a Citadel placement; and statistical or multi-agent-RL work with a Cubist placement. The same CV also lists topics involving causal inference, corporate default probabilities, TextGNN, robust ML discount curves, investment styles, stock-return prediction and multiple-testing/change-point methods. These are academic-placement and research-topic signals. They do not establish that any employer adopted a student’s work, or disclose employer data, permissions, production authority or performance.
• Polytechnique’s AI for Markets and Quantitative Investment X-ENSAE MaQI page describes a two-year English-language programme integrating finance and machine learning around practical market and investment problems. Its public partner surface displays QRT, BNP Paribas and Squarepoint, with S&P Global as a data partner and AMF as a supporting institution. This is programme and partner evidence, not proof of a partner-sponsored project, placement, dataset entitlement or production system.
• Bryant’s finance catalogue describes FIN 414, “AI Applications in Finance,” with coding, APIs, robo-advisors, textual analysis, sentiment/content classification and AI models. It also describes FIN 454, “Archway Equity Portfolio Management,” as the second course in a two-course student-managed-fund sequence with portfolio, risk and performance work and interaction with student analysts and investment professionals. The catalogue does not show that the student-managed fund uses FIN 414 methods, trades student models live or receives proprietary firm data.

Research implication: the academic map can now be joined at three separate edges—supervisor → research topic, programme → partner/talent surface, and applied course → student-managed investment environment. Keep each edge separate from employer system ownership, live deployment and measured performance. The useful model-research vocabulary to recover includes deep-learning asset pricing, statistical arbitrage, TextGNN, corporate-default prediction, institutional price impact, option-price surfaces, robust discount curves, multiple testing, change points and multi-agent market dynamics.

Recovery: retrieve Pelger’s linked papers, student code and seminar artifacts; obtain MaQI’s syllabus, project briefs, placement evidence and data-use terms; and obtain Bryant’s FIN 414 syllabus, assignments, Archway investment-policy materials, fund reports and practitioner-session artifacts. See the dated capture note.

September 5, 2026 — HEC, Bowdoin and Illinois add agent-market and research-workflow routes

• HEC Paris’s July 3, 2026 research account, linked to the Review of Financial Studies paper, describes simulations of AI market-making bots. The account reports prices remaining above the Glosten–Milgrom benchmark after about one million simulated trades, smaller tick sizes sometimes widening spreads, uncertainty slowing competitive learning, and more AI rivals gradually moving prices toward the competitive benchmark. This is a model and simulation result, not evidence about a named hedge fund or live deployment.
• David Byrd’s Bowdoin research page describes ABIDES and minABIDES multi-agent market simulations and lists research questions involving RL traders, LLM-enabled social-media pump-and-dump behaviour, spoofing, accidental cooperative price fixing, bubbles, interpretable execution and high-frequency historical order streams. Byrd’s public GitHub profile exposes stumpgrinder and minabides repositories, while the ABIDES paper documents the exchange/trader simulation design and market-impact use. This is an open academic/code route; it does not establish professional-firm adoption or trading performance.
• Illinois Gies’s June 11, 2026 AI Faculty Fellows announcement names a FIN 221 AI ecosystem with a custom tutor, CEO simulation, AI discussion engine and automated grading; a research-workflow guide covering coding agents, GitHub, version control and project architecture; a reusable AI-assistant architecture; and a governance/audit course. These are curriculum and institutional workflow artifacts, not hedge-fund systems or evidence of finance returns.

Research implication: the academic map now has a separate market-mechanism branch—agent objective → action space → exploration → interaction → emergent conduct → market-design response—and a research-operations branch—agent → repository/version control → domain assessment → human review → governance audit. The public sources expose concrete vocabulary for simulation, evaluation and controls, but not employer data, model ownership, production authority or performance.

Recovery: retrieve the HEC paper’s appendices and replication materials; run a permitted ABIDES/minABIDES smoke test and record versions; and obtain Gies FIN 221, research-workflow and governance artifacts. See the dated capture note.

September 5, 2026 — Santa Clara and Georgia Tech add family-office academic bridges

• Santa Clara University’s George Chacko profile identifies him as an Associate Professor of Finance whose research includes capital markets, market microstructure, financial security design and management of financial institutions including hedge funds and private equity funds. The page says he currently heads Confluentis Investments Pte Ltd, a family office focused on public and private investments globally, and records prior Harvard Business School appointments plus Harvard, Chicago and MIT degrees. This is a public academic/practitioner bridge; it does not disclose Confluentis’s AI, data, model, permissions or performance.
• Georgia Tech’s Mark Bell profile identifies Bell as Head of Private Capital and Family Office Services at Balentine, with prior CIO/Managing Director work at BlueArc Capital Management and prior Director of Strategy at D.E. Shaw. The page also records an Emory adjunct-professor and Center for Alternative Investments senior-research-fellow role, plus Stanford and Oxford education. This is a personnel and allocator-network route, not evidence of any firm’s AI system or deployment.

Research implication: the personnel graph should include a separate academic → allocator/family-office edge alongside the academic → hedge-fund and academic → programme edges. The public biographies expose market microstructure, private/public investment, alternative-asset management, private equity, reinsurance and alternative-investments vocabulary for follow-up. They do not support claims about model ownership, data access, production authority or returns.

Recovery: retrieve Chacko’s CV, papers, talks and permitted Confluentis disclosures; retrieve Bell’s Emory research artifacts, talks and dated Balentine/BlueArc/D.E. Shaw biographies; then search both across conference, podcast, LinkedIn/X and executive-education surfaces. See the dated capture note.

September 5, 2026 — Rockefeller, Duke, USC and Cornell add allocator-AI routes

• Rockefeller’s official June 10, 2026 release says the firm is building an AI-enabled wealth-management platform with Anthropic’s Claude model. It names client-meeting intelligence, operational workflows and internal support as initial use cases, but does not disclose architecture, prompts, retrieval sources, data permissions, evaluations, rollout status or investment authority.
• Rockefeller’s Global Family Office Private Advisor Team posting describes internal AI infrastructure, workflow best practices, tool evaluation, firm-sponsored pilots, prompts, templates, process guides, knowledge management and advisor-facing AI solutions. It lists Copilot, ChatGPT, Claude and Gemini as preferred experience. This is a job-description clue, not proof that each activity was completed or that the role controlled production systems.
• Duke’s March 20, 2026 alumni event page identifies Kai Cui as a Neuberger Managing Director leading data-science and AI investment-solutions efforts, with prior Point72 work integrating big data and advanced analytics into long/short equity research. The same page identifies Steven Nicklas as Cargill’s Director of Quantitative Trading, building a proprietary trading desk for a family office after managing a centralized futures/options execution platform; it records Duke, Tsinghua and NYU academic links. These are university-published biographies, not disclosures of models, data rights, deployment or performance.
• USC’s Joyce Shen profile describes an AI/ML and technology investor who has deployed institutional and family-office capital into technology companies, with prior Thomson Reuters emerging-technology investment leadership and IBM Cloud Platform experience. The profile is a personnel/capital-allocation bridge and does not identify a fund model or investment result.
• Cornell’s Rustom Desai profile identifies a Visiting Senior Lecturer who chairs his family-office companies and integrates AI into Strategic Alliances course materials and board-governance thinking. The profile exposes a family-office governance and executive-education route, not an AI investment-system disclosure.

Research implication: add separate evidence edges for provider partnership, job-defined workflow, named investment-solutions personnel, family-office quant-platform build, AI-capital allocator and family-office governance education. Do not collapse them into one firmwide AI strategy or infer production adoption beyond the source language.

Recovery: obtain Rockefeller architecture, privacy/data terms, evaluations and later postings; capture the Duke event recording and named personnel artifacts; retrieve Shen’s CV and research; and recover Cornell course and governance materials. See the dated capture note.

Georgia Tech: AI-in-finance capstone and student-fund bridge

• Georgia Tech Scheller’s 2025 Finance Group report identifies Satyajit (Jeet) Karnik as a finance lecturer whose specialties include risk management, derivatives and machine-learning methods in finance. It also describes the 2024 Center for Finance and Technology, supported by Mike and Jenny Messner, a former Seminole Management co-founder and chief investment officer.
• The report lists an “AI in Finance Capstone” applying machine learning and large language models to finance-industry problems, capstone projects with finance firms, and investment-management experience conducted with faculty and the Georgia Tech Student Foundation. It says selected students can participate in practical, real-world fund management.
• This is a useful academic and talent-network route: it joins ML instruction, an explicitly named AI/LLM capstone, finance-firm project language, and a student-managed investment environment. The report does not name capstone students, partner firms, datasets, code, model versions, live positions, deployment authority or performance. The student-fund description should not be treated as evidence that the fund uses AI methods.

Research implication: search Georgia Tech’s finance faculty, Center for Finance and Technology, QCF programme, Student Foundation and practitioner events together. Recover capstone briefs, project presentations, partner names, data-use terms, course materials and fund reports while preserving separate evidence states for teaching, sponsorship, employment, data access, deployment and performance. See the capture note.

Penn State Behrend: a dated student-fund and analytics baseline

• Penn State Behrend’s 2025–26 Finance Newsletter says FIN 497, “Use Cases in Financial Analytics,” was being offered to World Campus students for the first time in spring 2026. The page does not publish its syllabus or identify an AI model or vendor.
• The newsletter reports that the Intrieri Family Student-Managed Fund surpassed $1.6 million, names Samuel Greene, Sergej Stojanovic, Nicholas Kerner and Brayden Ervin as its 2025–26 officers, and says students serve as portfolio managers and analysts while executing trades of approximately $30,000 each. It lists Bloomberg, Value Line and Morningstar as valuation tools.
• The same page describes a CFA Research Challenge team led by Dr. Greg Filbeck, an industry adviser and a professional equity-report workflow. These are useful academic talent and workflow signals, but the newsletter does not connect FIN 497 to the fund, disclose AI use, or establish external-firm sponsorship, proprietary data, trading authority or performance attribution.

Research implication: recover FIN 497 materials, Intrieri fund reports, investment-policy documents, student presentations, practitioner events and officer/alumni histories. Reconcile this dated $1.6 million statement with older catalogue figures and preserve valuation dates, denominators and performance definitions. See the capture note.

Reading, Surrey and Stanford: programme and trading-lab routes

UK finance-AI practitioner and programme routes

• Middlesex University’s Alberto Pallotta profile describes research in portfolio optimisation, volatility prediction and quantitative trading, including machine-learning trading algorithms and graph theory in portfolio optimisation. It identifies an engineering degree, an MSc in Artificial Intelligence, Computational Finance teaching and a current Head of R&D role at an unnamed Swiss asset-management firm. The page does not name the firm, models, datasets or results.
• Coventry University London’s Shahzeb Mohammed biography identifies him as MSc Global Financial Trading course director and describes teaching in algorithmic trading, Python/R machine learning, quantitative risk and Bloomberg analytics. The university page also says he previously worked at London Business School’s Centre for Hedge Fund Research & Education and held hedge-fund quant-desk leadership involving quantitative-strategy development and deployment. These are university-hosted biography claims; no fund or strategy is named.
• Temple’s Quantitative Finance MS bulletin lists a 30-credit curriculum spanning Data Intelligence, Data Science in Finance, Machine Learning in Finance, Generative AI for Finance, AI in Portfolio Management, credit risk, volatility, asset pricing and quantitative portfolios. It also describes guest lectures and financial-firm visits. This is curriculum evidence, not production-system or employer-deployment evidence.
• The University of Georgia Terry College SMIF page describes a $4.5 million Athena Stock Fund across 41 holdings and a $2.25 million Arch Bond Fund in active commitments, with student sector allocation, stock pitches, macro/rate analysis, fixed-income security selection and Bloomberg Aggregate benchmarking. The page does not disclose AI use, model code, proprietary data or external manager deployment.

Research implication: follow the named researchers, publication repositories, course materials, instructors, guest rosters, student-fund reports and information-session artifacts. Keep biography claims, historical employment, curriculum, student-fund operation, current R&D, data access, deployment and performance as separate evidence states. See the capture note.

• The University of Reading’s BSc Finance (Investments) 2026/27 page lists an optional AI and machine-learning-in-finance module covering dataset intelligence extraction, Python, generative-AI tools such as ChatGPT and SQL for large datasets. It sits alongside derivatives, financial engineering, fintech and a research project. This is curriculum evidence, not a named firm’s live workflow.
• The University of Surrey’s 2026/27 Investment Management MSc specification names Vasileios Pappas as programme leader and explicitly includes ethical evaluation and use of AI in investment management, with textual analysis, portfolio management, valuation, behavioural finance and independent research. It does not disclose a model, dataset, employer project or deployment.
• Stanford GSB’s biography of Kevin Mak identifies him as director of the Real-time Analysis and Investment Lab and FIN362 instructor. It records Rotman training, prior management of Rotman’s Financial Research and Trading Lab, co-invention of two trading and portfolio-management simulation platforms, oversight of Stanford’s student-managed fund, and interests in microstructure, liquidity and network effects. This exposes educational infrastructure and talent-network vocabulary, not adoption by a hedge fund.

Research implication: recover the Reading and Surrey syllabi and Stanford RAIL/FIN362 materials, simulation-platform documentation, student-fund reports and public talks. Keep curriculum, educational infrastructure, practitioner consulting, external sponsorship, proprietary data, deployment and performance as separate evidence states. See the capture note.

UIC, Penn State and Rice add academic and personnel routes

• UIC’s profile for Tengjia Shu identifies an Assistant Professor of Finance whose research sits at the intersection of asset pricing, investment management and machine learning. The profile specifically names hedge-fund and mutual-fund performance evaluation, return-predictability signals and sparse stochastic discount factors, and records an Iowa finance PhD plus statistics training in Beijing. UIC also lists a 2022 FMA honor for a paper on hedge funds with ML benchmarks. This is an academic research and lineage route; it does not establish any fund’s adoption, data access, model authority or performance.
• Penn State Smeal’s Jingzhi Huang profile identifies a Professor of Finance and Mathematics and Faculty Chair in Finance. It lists credit risk, fixed income, derivatives, mutual funds, hedge funds and machine learning among his expertise, records a 1997 NYU finance PhD, and links a publication trail including ML-based return predictors, macro-finance, corporate bonds, liquidity, mutual-fund flows and hedge-fund performance analysis. The page also lists empirical research-methods and portfolio-management teaching. These are public academic outputs and course routes, not evidence of a named manager’s live strategy.
• Rice’s Michael D. Jackson profile identifies the director of Rice’s Professional Master of Statistics programme and a faculty lecturer connected to its Center for Computational Finance and Economic Systems. Rice says he is currently head of data science and machine learning for New Territory Advisors, LLC, a Houston hedge-fund company, and records Stanford, SMU and Rice education plus prior finance, analytics and venture-capital roles. This is a university-hosted personnel/title record. It does not disclose New Territory’s models, data suppliers, permissions, deployment, investment authority or results.

Research implication: add three separable graph edges—academic finance-ML research to fund-performance questions, a publication lineage spanning ML return predictors and credit/fixed-income markets, and a current named hedge-fund data-science leadership title connected to a computational-finance programme. Recover papers, code, supervisors, course materials, dated professional profiles and employer-side corroboration before making any firm-level inference. Keep academic work, university-hosted biography, independent employment confirmation, proprietary data, deployment and performance as separate evidence states. See the capture note.

Rutgers exposes an unusually specific public finance-ML and GenAI curriculum

• Rutgers–Camden FIN 582: Investment Management and Machine Learning names Dr. Wei Jiao and describes a Spring 2026 course using Python, financial data, portfolio construction, stock pricing and robo-advisors. The syllabus names lasso, elastic-net, random forests, gradient boosting and neural networks; it says the course uses S&P Global Compustat data through WRDS; and it schedules model-improvement, interpretation and multiple neural-network units across 13 weekly projects. This is a public training and data-environment signal, not evidence of a named manager’s live workflow.
• Rutgers–Camden FIN 583: Generative AI and Textual Analysis in Finance names Professor Tengfei Zhang and describes R/Python extraction from prices, filings, earnings calls, media, social posts and corporate websites using APIs and scraping. It names GPT, bag-of-words, sentiment, embeddings, BERT/FinBERT, OpenAI text embeddings, fine-tuning, RAG, multimodal learning, AI hypothesis generation and agent creation. The syllabus lists a course GPT, ChatGPT Plus, and instructor-provided earnings-call, Glassdoor, RavenPack, Reddit, Compustat and CRSP data. Its project menu includes culture/ESG, crypto and meme-stock returns, earnings-call risk, executive–analyst differences, 10-K risk factors and stock-return prediction. These are instructional claims and project prompts; they do not establish vendor contracts, production use, hedge-fund sponsorship or validated returns.

Research implication: Rutgers provides a concrete public baseline for comparing academic finance-AI routes: conventional investment ML with regularisation, ensembles, neural networks and interpretation; then unstructured-data collection, embeddings, fine-tuning, RAG, agents and multimodal research. The named data surfaces create a provenance checklist for any later backtest—timestamps, point-in-time joins, vendor entitlements, social and Glassdoor access, model/prompt revisions, held-out evaluation and costs. See the capture note. Nothing here establishes adoption by a named manager or relative quality among firms.

Rutgers and SDSU add finance-lab and student-fund infrastructure

• Rutgers–Camden’s Financial Markets Lab describes Bloomberg Professional access to market data, news, research and analytics, plus training technology for tactical trading decisions. The page names Ralph Giraud as lab manager and Teaching Instructor in Finance. It is an educational data/tooling surface adjacent to the Rutgers ML and GenAI syllabi, not evidence of a fund’s production stack.
• SDSU’s Financial Markets Lab says it operates twelve Bloomberg terminals and runs an eight-week stock-pitch challenge, a ten-week Bloomberg Trading Challenge with a mock $1 million portfolio, and an economics research project using Bloomberg, Excel, R and Python. The page says the Aztec Investment Fund uses Bloomberg to research and invest approximately $1 million of AUM, with students executing real trading and portfolio decisions. It also lists BlackRock, Bloomberg and Deloitte guest speakers and an MSCI donation of more than $500,000 in indices and databases in 2013. These are university-reported student-fund, data and talent signals; they do not disclose a machine-learning model, audited performance, current sponsor terms, or employer use of student work.

Research implication: these labs expose the operational substrate around finance education—terminal and data access, Python/R research, mock portfolios, real student capital, industry speakers and alumni routes. Keep adjacent curriculum, student-fund activity, model use, external sponsorship, employer contact, deployment and performance separate. See the capture note.

SUTD’s AIFi Lab exposes concrete multimodal and portfolio-model routes

• SUTD Associate Professor Dorien Herremans’s AIFi Lab page says she heads an AI for Finance lab working on deep-learning trading algorithms, portfolio management, generative AI for synthetic data and volatility prediction across markets including digital assets. The page links PreBit, a multimodal model using Twitter FinBERT embeddings for extreme Bitcoin-price-movement prediction; a volatility-spike paper using whale transactions, CryptoQuant data and Synthesizer Transformers; and DeepUnifiedMom, a multi-task, mixture-of-experts approach to time-series momentum portfolio construction. These are public academic model/data routes, not evidence of a hedge fund’s live system.
• The page lists generative models, LLMs and FinTech, deep learning for trading and portfolio optimisation, blockchain/digital assets and other AI-fintech topics as doctoral and postdoctoral research areas, alongside funded PhD recruiting. It does not identify a fund partner, commercial deployment, data licence, portfolio authority or audited result.

Research implication: SUTD expands the modality map into social-text/market fusion, on-chain or vendor crypto activity, volatility-event forecasting, synthetic data, time-series momentum, multi-task learning and mixture-of-experts portfolios. Recover the papers, code, data statements, timestamps, splits, costs and collaboration records; keep academic research, recruiting, external partnership, deployment and performance separate. See the capture note.

UCC adds a hedge-fund research-conference network and a historical SMIF baseline

• University College Cork’s Mark Hutchinson profile identifies him as Professor and Chair of Finance, Co-Director of UCC’s Centre for Investment Research, and founder/co-chair of the Annual FMA Conference on Hedge Funds, which the profile says has run in Europe since 2014. It describes research using novel data gathered through financial-services links and covering hedge funds, behavioural finance and futures-market trading strategies. The profile records conference venues including Imperial College London Business School, Aspect Capital and Cambridge Judge Business School, and a centre founded with researchers at CASS, DCU and UCC. This is an academic-network and conference-recovery route; it does not disclose a fund’s models, data rights, deployment or authority.
• The primary journal-volume PDF containing Jones and Swaleheen’s paper documents one Florida university’s student-managed equity fund from January 2005 through April 2013 using all holdings and transaction data for that period. It describes a $200,000 initial allocation, faculty and foundation oversight, and exchange, market-capitalisation, concentration and sector constraints. The authors report cumulative return, volatility, Sharpe and CAPM figures against the S&P 500 for the stated sample, including 53.60% versus 31.82% cumulative return and 13.41% versus 15.58% cumulative standard deviation. These are historical, single-fund paper statistics—not a general claim about student funds or professional managers—and the paper does not discuss AI or ML.

Research implication: recover the Annual FMA on Hedge Funds programmes, papers, discussants, sponsors, recordings and participant transitions; map UCC/CASS/DCU coauthors and supervisors; and use the SMIF paper as a governance-and-data baseline with explicit sample, constraint, transaction and benchmark fields. Preserve academic collaboration, student capital, employer contact, model use, deployment and performance separately. See the capture note.

September 5, 2026 — family-office governance, Princeton infrastructure, and university AI-finance routes

• Columbia’s Achilles Venetoulias profile describes an adjunct finance professor who, according to the university biography, founded and ran two hedge funds, supervised a European fund-of-hedge-funds process, founded a fintech company and most recently oversaw portfolio analytics for a large family office. It records Columbia mathematics and Stanford statistics/computer-science training, prior MIT Sloan teaching, and courses in quantitative finance and systematic investment strategies. The office and funds are unnamed, and the profile does not disclose models, data, permissions or results.
• UCF’s Eric Ambinder profile describes legal counsel for a private family office focused on strategic investments in cancer therapeutics, health technologies and AI, with prior Accenture work on AI contracting and responsible GenAI adoption. This is a governance and investment-domain signal, not a disclosed family-office portfolio, model or deployment.
• Princeton’s finance graduate-programme description explicitly joins Computer Science, Operations Research, the Center for Statistics and Machine Learning and the AI Lab with quantitative asset management, forecasting, quantitative trading and applied research. It describes tracks in quantitative asset management, data science and financial technologies, required internships, and analysis of hedge-fund portfolios, trading models, risk/return and macro relationships. This is programme evidence, not proof of a live Princeton or partner-firm system.
• Pace’s university-hosted Cornspring posting describes a family-office platform applying GenAI and LLMs to investment, accounting and operational data. The role exposes ingestion, validation/reconciliation, ML data integrity, Python, SQL, PostgreSQL, unstructured data, data-unit testing and Azure. The employer copy is not independent validation and does not name clients, models, vendors, data rights or investment authority.
• Sam Houston State’s College of Business account describes a Bloomberg-equipped financial centre, a university-stated student fund with up to $500,000 of initial proceeds under Jose “Joey” Gutierrez, and collaboration with Forward Edge-AI through COO George Stephenson. The page names Gutierrez’s interests in market microstructure, liquidity, price dynamics and international markets. It does not specify the R&D project, model, dataset, fund workflow or production use.

These routes add model-risk governance, family-office data reconciliation, AI-investment diligence, market microstructure, and the academic path from AI/ML infrastructure to hedge-fund analysis. They remain separate from evidence about GMO, Acadian, Arrowstreet or any other manager’s actual systems.

September 5, 2026 — practitioner professors and finance-programme infrastructure

• Queen Mary’s profile for Drago Indjic identifies him as a Visiting Lecturer of Fintech, an Industrial Professor at UCL’s Institute of Finance and Technology, and co-Head of Oxquant. The university biography says that, since 1993, he has held portfolio-management and quantitative-research roles at several hedge funds, a family office and a sovereign wealth fund, and co-founded five fintech and fund ventures. It records an Imperial College London engineering PhD and a University of Belgrade engineering degree. The page does not name those employers or disclose Oxquant’s models, data, clients, permissions or results.
• Columbia Business School’s profile for Robert Swigert identifies him as an Adjunct Assistant Professor whose research and writing cover macroeconomics, macro finance, portfolio choice, corporate finance and asset pricing. Columbia says his investment-management career began in Trading and Operations at Bridgewater, followed by operations-research and technology-management roles; that he later evaluated life-sciences investments at a family office with emphasis on computational drug discovery, infectious disease and digital health; and that he is General Partner of Tyr Partners LP, which designs passive, liquid portfolios for foundations, endowments and family offices. The profile does not disclose Tyr’s or Goodstead’s models, holdings, data suppliers, authority or performance.
• Stevens’ Hanlon Financial Systems Center describes two finance-company-style labs used for asset pricing, risk management, portfolio strategy and cybersecurity. It says students apply machine learning, data visualization and risk engineering; connects the center to quantitative-finance, financial-technology and financial-engineering degrees; lists certificates including algorithmic trading, financial computing, financial software engineering and financial statistics; and says the labs access real-time and historical data while partner companies send employees to training. The page does not identify those partners, disclose data licences or establish any manager’s deployment or performance.
• Fordham Gabelli’s M.S. Finance bulletin exposes a particularly concrete programme route: the financial-analytics track lists Python, cloud computing and finance uses, advanced machine learning, equity-factor investment strategy, fintech in portfolio management, algorithmic trading, computational finance and AI, automated trading systems, machine learning for finance and applied quantitative-investment strategy. The bulletin also lists a Value Investing Student Management Fund. Fordham’s M.S. Quantitative Finance bulletin describes a student prototype auto-trading system using live exchange data and discusses slippage, scale and asynchronous information processing. These are curriculum and educational-project claims, not evidence of a named hedge fund’s live system.

Together, these routes add a cross-sector quant/family-office lineage, computational life-sciences investing, finance-lab and employer-training infrastructure, and explicit training in cloud, ML, factor research and automated trading. They are useful search surfaces for papers, syllabi, supervisors, student projects, jobs and conference appearances; they do not support a ranking of firms or an inference that a programme or professor’s methods are deployed by a particular manager. The dated source capture and recovery boundaries are retained in the research ledger.

September 5, 2026 — academic model artifacts and current finance curricula

• Columbia’s 11th Annual Bloomberg-Columbia Machine Learning in Finance conference page adds a dense academic discovery surface. The programme describes Georgia Tech’s Sudheer Chava evaluating more than 197,000 questions about U.S. public companies to study temporal and cross-sectional knowledge bias in financial LLMs. It also describes UCL’s Silvia Bartolucci using deep learning and network representations for limit-order-book forecasting, with attention to tick size, information decay and trading feasibility. Its poster list includes satellite imagery for physical-risk and economic-activity modelling, graph portfolio optimisation, LLMs and graphs for financial texts, RAG-based synthetic scenarios, LLM-generated market colour, generative portfolio management and regime detection. These are academic topics and personnel routes, not manager disclosures.
• The Columbia-hosted poster “Deep Generative AI for Portfolio Management” by Stony Brook’s Jeongyeon Park and Young Shin Kim gives unusually concrete model and data detail: Student-t Glow, Flow Matching and score-based models learn heavy-tailed return distributions for CVaR optimisation; the experiment uses 6,367 daily log returns from nine sector ETFs and generates 200,000 synthetic return vectors for comparison with historical, Gaussian and Student-t baselines. The poster does not establish transaction-cost-adjusted performance, point-in-time validation, released code or institutional deployment.
• Fairfield’s M.S. Financial Technology catalogue describes a 30-credit STEM programme combining data design, generative-AI applications, algorithmic trading, AI systems, data engineering, portfolio management, risk management and fintech ethics/regulation. Its algorithmic-trading description includes strategy generation, backtesting, optimisation, infrastructure and performance evaluation; its AI systems and generative-AI descriptions cover lifecycle design, deployment, maintenance and tool chaining. The catalogue does not identify employer projects, licensed datasets, models, student outputs or fund deployment, and the direct page required a recovery path during capture.
• INSEAD’s current Master in Finance curriculum advertises an August 2027 France/Singapore intake with Data Science for Finance, AI & Deep Tech for Finance, an AI practicum and professional experience. Its Wealth Management track lists Hedge Funds and Family Offices; the page describes machine learning for asset-price forecasting, credit decisions and multi-asset portfolio optimisation, and generative/agentic AI for research, advisory work and workflow automation. Its statement that students may deploy live models in practicals is a programme claim, not evidence of a named employer’s system or investment result.

These additions turn the academic route into a more actionable research queue: distributional scenario generation for downside-risk optimisation; microstructure and information-decay modelling; financial-LLM temporal-bias evaluation; multimodal physical-risk signals; and curricula that connect data engineering to governance and trading-system design. They remain academic or programme evidence and should not be converted into a ranking or a claim about any manager’s internal deployment. The detailed capture note is retained in the research ledger.

September 5, 2026 — Asia-Pacific finance-AI programmes, labs and researcher lineages

• NUS AIDF’s Generative AI in Finance course exposes a Singapore-based executive-training route covering ML, deep learning, NLP, computer vision, reinforcement learning, LLMs, prompt design, investment analysis, risk, fraud and the Monetary Authority of Singapore FEAT framework. Its instructor list connects Ke-Wei Huang’s NYU Stern information-systems PhD and AI/data-mining/finance research with Yen Teik Lee’s SMU finance PhD, David Lee’s NUS/SUSS/Shanghai fintech network, MAS AI Development Office leadership and AI-governance practice. The page does not identify participant firms, datasets, prompts, model versions or deployment.
• SMU’s EngD curriculum explicitly pairs an “AI in Financial Markets Forecasting” course with foundation models, network and multimodal data, time-series forecasting and portfolio-management applications. The same programme lists generative models for NLP and multimodal reasoning, optimisation, intelligent-system planning, software mining and empirical research projects. This is a concrete academic stack for investigating multimodal forecasting and agentic decision support, not evidence of a fund sponsor or live trading system.
• NUS AIDF’s PhD roster publishes student–supervisor and research-topic links. Public examples include explainable AI and LLMs; ML, deep learning, reinforcement learning and portfolio optimisation; text mining and financial forecasting with generative language models; causality and probabilistic ML; and multimodal foundation models for financial narratives. Supervisors named include Stanley Kok, Julian Sester, Ying Chen, Johan Sulaeman, Ling Feng, Erik Cambria, Gianmarco Mengaldo and Paul Pu Liang. These are academic lineage signals, not manager-employment or deployment claims.
• SIT’s profile for Bullipe Chintha records an Indian School of Business accounting FPM, a Deakin finance PhD via the Deakin–ISB partnership, a Kellogg visiting-scholar period and earlier Edelweiss and Reliance roles. It describes current ML/GenAI research on intangible assets, firm value, labour markets and consistency between filings and earnings calls, plus teaching in data analytics and ML. The profile lists a former Harvard Business School AI Institute research-affiliate connection and a 2025 accounting/finance publication; it does not establish an investment manager’s model or data access.
• NUS Business School’s executive AI-and-finance course describes a three-day route from AI/ML foundations and LLMs to Python and practical work in portfolio optimisation, credit, property-price and time-series prediction. Yen Teik Lee and Ted Teo are listed as faculty, with an industry guest session. This exposes executive-facing implementation vocabulary, not a named institution’s production workflow.
• Lincoln University’s Master of Fintech and Investment Management links investment management and software development with big data, ML, AI, blockchain, cybersecurity and a research essay or placement. Its compulsory courses include Big Data and Machine Learning Tools and Techniques; programme contacts are Associate Professor Cuong Nguyen and Professor Christopher Gan. The programme advertises industry placement and agribusiness applications but does not identify a manager, dataset, model or deployment.

This regional pass adds a supervisor graph, finance-specific multimodal and foundation-model coursework, regulator-facing GenAI education, financial-disclosure research and a software/data-to-investment pipeline. These are academic and educational signals. They should guide searches for papers, code, syllabi, placements and personnel transitions, without being treated as evidence that any tracked hedge fund uses them. The detailed capture note is retained in the research ledger.

September 5, 2026 — European finance programmes expose model, market and control interfaces

• LSE’s MSc in Financial Mathematics regulations for 2026/27 list Computational Methods in Finance as compulsory and offer Machine Learning in Financial Mathematics and A Mathematical Approach to Generative AI. The same option set includes Portfolio Management, Advanced Time Series Analysis, Mathematics of Market Microstructure, and Stochastic Simulation, Training, and Calibration. This is a useful academic route for tracing how mathematical finance, ML/GenAI, microstructure and calibration are taught together. The page does not identify a manager, proprietary dataset, production model or investment result.
• IESE’s Master in Finance programme page gives unusually explicit finance-AI syllabus language. Its Deep Learning and Generative AI course covers deep learning, LLMs and reinforcement learning across financial text, time series and structured information; examples include earnings-call summarisation, news sentiment, fraud detection, regulatory-filing classification and trading-signal extraction. It also describes agentic LLMs interacting with data, tools and workflows, plus Q-learning and exploration–exploitation for trading, asset allocation and pricing. The course objectives mention training, fine-tuning, prompting, interpretability, bias and model risk. These are curriculum claims, not evidence of a named firm’s system or deployment.
• The University of Luxembourg’s Investment Management programme places AI strategies alongside factor investing, macro/style factors, smart beta and systematic strategies. Its accessible indexed course text also covers CTA/managed futures, alternative data, transaction costs, best execution, liquidity, leverage, collateral, fund restrictions and stress testing. The page was robot-gated on direct inspection, so this route remains a recovery item until a browser capture or official syllabus is retained; no fund, model, dataset or performance claim follows from it.
• ENSAE Paris’s Finance, Risks and Data specialisation names Jean-Michel Zakoïan and Peter Tankov as pathway heads and describes a 2026/27 course grid with Applied Statistical Learning, Machine Learning for Financial Time Series, Machine Learning for Portfolio Management and Trading, Mathematics of Machine Learning, AI for Better Risk Control, Algorithmic Trading, Credit Risk, Generative AI for insurance and actuarial studies, Portfolio Management and GPU programming. ENSAE describes destinations including quants, portfolio managers, risk managers, hedge funds, pensions, insurers and regulators; that is a training and career-path statement, not evidence of any employer’s live implementation.

These programmes add four different discovery routes: mathematical ML and microstructure; LLMs and tool-using agents for financial documents; systematic investing with alternative data and execution constraints; and statistical learning, GPU work, trading and risk control. The next evidence pass should recover course guides, instructors, theses, student code, employer events and paper/supervisor lineages. Academic material remains separate from verified manager disclosure.

September 5, 2026 — professors expose the research gap and the validation burden

• Goldstein, Spatt and Ye’s Review of Financial Studies introduction to “Big Data in Finance” defines finance big data through size, dimensionality and complex structure, explicitly including text, images, video, audio and voice. Its examples span FINRA order data, cross-asset market microstructure, clickstream information demand, earnings-call culture and machine-learning controls for multiple testing. It also identifies a historical gap between academic work using monthly returns or quarterly 13F holdings and practitioners operating at shorter horizons. That is a literature-gap observation, not evidence of any named manager’s current system.
• Mihail Velikov’s Oxford-Man “Assaying Anomalies” presentation provides a concrete research-quality route: it describes filtering 31,460 candidate signals to 17,074 after redundancy, breadth, history and data-availability screens, then varying portfolio construction, weighting and trading-cost assumptions. Its dual-hurdle protocol leaves roughly 2% of signals passing the stated screen. The useful takeaway is the validation design—point-in-time data, breadth, alternative construction, costs and independent hurdles—not a claim that any signal is investable or used by a fund.
• Imperial’s profile of Robert Kosowski records a professor of finance, CEPR research fellow and 2024 co-founder of the Centre for Excellence in Quantitative Finance. It links research interests in asset management, risk, ML, hedge funds, performance measurement, derivatives and forecasting to Cambridge and LSE training, prior Goldman Sachs, Boston Consulting Group and Deutsche Bank work, and part-time quantitative-research leadership at Unigestion. This is an academic-to-industry lineage signal; it does not disclose a manager’s model, data rights, investment authority or production transfer.

The academic route now contributes two kinds of evidence: hypotheses about under-mined modalities and trading horizons, and protocols for rejecting fragile signals. The next pass should recover the underlying papers, code, datasets, course artifacts and supervisor/employment links before making any stronger claim.

ANU exposes a finance-AI course and project-artifact route

• ANU’s 2026 FINM3015 course record identifies Foundations of Fintech as a six-unit course offered by the Research School of Finance, Actuarial Studies and Applied Statistics. It names Dr Lin Hu as convener and covers financial technology affecting trading, asset management, lending, insurance and capital-market infrastructure.
• The class summary lists robo-advisors and AI in investing, AI/ML, AI agents in finance and trading, APIs/data/open banking, and regulation/privacy. It requires a project proposal, final presentation and project report, and says lectures and workshops are recorded. This creates a concrete route to permitted student artifacts and personnel follow-up; it does not establish a professional manager’s deployment, data access or investment performance.

Research implication: recover public project titles, repositories, recordings and the convener’s research profile, then check any student-managed-fund connection separately. Keep course content, student work, employer relationships and live investment systems as distinct evidence states. See the capture note.

September 5, 2026 — regional academic routes expose personnel and implementation vocabulary

• IIT Bombay’s profile of Rohan Chinchwadkar identifies him as Associate Professor of Finance, associate faculty at the IITB–Citadel Securities Quantitative Research Lab, and former Professor In-Charge of the HDFC ERGO–IIT Bombay Innovation Lab. His public PhD-guidance statement specifically names large language models and agentic AI in finance. The profile records an IIM Calcutta finance PhD, prior McKinsey financial-services work, and the 2021–24 HDFC ERGO innovation-lab role. This is a personnel and academic-interface signal; it does not establish Citadel Securities or HDFC ERGO model use, data rights, production transfer or investment results.
• PUC-Rio IAG’s Portuguese “Machine Learning Applied to Investments” programme describes a 168-hour sequence covering quantitative finance, econometrics, time series, ML, financial ML and quantitative finance. Its published objectives include quantitative strategies, portfolio formation, long–short positions, automated trading robots and NLP for unstructured financial data; the module descriptions list factor models, Ridge/Lasso, random forests, neural networks, deep learning, ensembles, SVMs, clustering, central-bank minutes and news. Coordinators include Eduardo Marinho and Marcelo Cabús Klötzle. This is a Portuguese-language talent route, not evidence of a named fund’s deployment.
• FGV EMAp’s Quantitative Methods in Finance and Risk specialisation lists an 18-month, 12-course sequence spanning programming/data science, ML, time series/econometrics, derivatives, quantitative risk, deep learning, portfolio optimisation and algorithmic trading. It includes a final project on a real problem and names Rodrigo dos Santos Targino as coordinator, with a University College London statistics PhD and prior financial-industry experience. The page does not disclose project hosts, datasets, model versions, investment authority or results.

This regional pass adds a concrete India-based academic–industry personnel bridge and Portuguese-language curriculum evidence that is easy to miss in English-only searches. The recovery queue is the underlying IITB lab projects, course recordings, faculty CVs, student theses and permitted code—not an inference about live manager systems.

September 5, 2026 — Hong Kong universities expose family-office operating routes

• Hong Kong Baptist University’s 2026/27 handbook names Professor Monique Wan as programme director and lists Family Office Fundamentals & Governance, Family Offices Practicum, Digital Transformation in Family Offices, and Family Wealth Preservation & Growth. Electives include Digital Assets & Allocation, Behavioural Finance, Private Equity & Alternative Investments, and Global Macroeconomics & Cross-border Capital. This exposes a family-office operating and technology vocabulary; it does not identify participating offices, vendors, models, data or deployment.
• HKU’s Master of Family Wealth Management curriculum requires core work in asset allocation, investment management and current family-office issues, offers Hedge Funds and Private Market Investments as an elective, and culminates in a practical research project or industry internship. It names Professor Zhiwu Chen as institute director and Professor Bonnie Leung as programme director; the page records Leung’s former BlackRock COO roles, Asia-Pacific sustainable-investing work, McKinsey financial-services experience and Wharton MBA. This is a public academic/practitioner bridge and a route to capstones, placements and guest events, not evidence of a specific office’s AI system.
• HKUST’s MSc in Family Office and Family Business describes Asian family-office case studies, workshops, industry immersion, corporate projects, an FOFB Project and an FOFB Practicum. The programme site names Professor Roger King as founding director of the Roger King Center for Asian Family Business and Family Office and lists advisers from Shui On Group, Lee Kum Kee Group and Lee Heng Diamond Group. These are network and curriculum signals; no member-office identity, model, vendor, data entitlement, investment authority or result is disclosed.

The family-office academic route now has concrete recovery surfaces for governance, digital transformation, asset allocation, alternative investments, capstones, internships and practitioner events. The next step is to recover those artefacts and alumni transitions while keeping programme intent, adviser association, current employment and verified deployment separate.

ANU and Glasgow add model and validation lineages

• ANU’s Priya Dev profile describes long-memory asset-price research, machine learning for trust propagation in blockchain pricing, neural networks and extreme-value modelling for Australian electricity prices, and working papers on path-dependent options and long-range price dynamics. It also records consulting to ASX-listed companies and prior teaching at ANU and Columbia. This is academic and consulting evidence, not a disclosed hedge-fund system.
• Glasgow’s Georgios Sermpinis profile lists machine learning, financial trading, forecasting, risk management and operations research, along with bank consultancy and seminars. Its publication list includes interpretable deep learning for industry returns, LSTM-LagLasso bond-yield forecasting, ML portfolio construction, mutual-fund data-snooping controls, Bitcoin volatility/sentiment and neural-network copula ETF optimisation. These are academic publication and consulting routes; they do not establish manager deployment, data access or investment performance.

Research implication: recover the papers, samples, feature clocks, costs, code, supervisors, seminars and permitted collaboration records. Treat long-memory, energy, blockchain, fixed-income, portfolio and false-discovery methods as separate research branches, and keep academic affiliation, consulting, employer role and deployment distinct. See the capture note.

September 5, 2026 — JCU adds an Asia-Pacific liquidity and media-data route

• James Cook University’s profile for Huiping Zhang identifies her as an Associate Professor of Business at JCU Singapore Business School, with a 2011 PhD in Finance from the National University of Singapore and prior Associate Professor of Finance role at Shanghai University of Finance and Economics. The profile lists empirical asset pricing, market microstructure, international stock markets and machine learning, with particular attention to liquidity effects on returns, emerging-market liquidity measurement, and media coverage effects on liquidity and returns in China.
• JCU’s publication list includes 2025 work on credit-rating and stock-return comovement; 2023 work on liquidity shocks and the negative premium of liquidity volatility around the world; and earlier work on international illiquidity premia, information uncertainty and liquidity pricing, emerging-market liquidity, and limit-order-book commonality. This creates a concrete search path from the economic problem—liquidity and information—to possible ML and alternative-data research, without assuming that any named manager uses the work.

Boundary: the university profile supports the appointment, education, prior institution, research interests and listed publications. It does not identify a fund, family office, data licence, proprietary feature set, model architecture, investment authority, production deployment or independently verified investment result. See the capture note.

September 5, 2026 — CUHK adds a cost-aware factor-model lineage

• CUHK Business School’s profile for Sicong Li identifies him as an Assistant Professor of Finance with a BA from the Central University of Finance and Economics, an MS from Columbia, and a 2024 PhD from London Business School. The profile lists empirical asset pricing, applied financial econometrics and machine learning, with investment and portfolio management as teaching areas.
• Li’s listed work includes the 2024 Journal of Financial Economics paper “Comparing Factor Models with Price-Impact Costs,” co-authored with Victor DeMiguel and Alberto Martin-Utrera, as well as working papers on low-frequency risk factors and factors with economic targets. This is a useful academic route into cost-aware factor selection and portfolio construction: it foregrounds price impact and economic targets rather than treating model accuracy as sufficient. It remains academic evidence, not a disclosure of any manager’s model, data or deployment.

Boundary: CUHK’s official page supports the appointment, education, research interests, teaching areas, publication metadata and awards. It does not identify a hedge fund or family office, proprietary data, model architecture, live portfolio authority, production deployment or independently verified performance. See the capture note.

September 5, 2026 — classroom-scale document measurement, linguistic stress, and optimized narratives

• UT Austin McCombs’ “Keeping Humans in the Loop” identifies Michael Sury as an associate professor of practice in finance and describes his graduate course in analytic finance and machine learning. Sury says students process hundreds of thousands of earnings-call transcripts, collectively millions of sentences, and use LLMs to score tone, sentiment, and contradictions between prepared remarks and Q&A. The commentary also describes a classroom test in which an LLM return-prediction backtest reportedly collapsed after moving beyond the model’s training cutoff. This is a university commentary and classroom account, not a reproducible study or a hedge-fund disclosure; the transcript source, prompts, model versions, temporal split, code and results remain to be recovered.
• Claremont McKenna’s open-access senior thesis by Yiqi Sun is a distinct text-modality route. The 2026 economics thesis reports psychology-backed Claude Sonnet 4.6 prompts applied to 7,491 earnings-call transcripts from 470 S&P 500 companies over 2020–2024. Its abstract reports a 0.9% lower two-day cumulative abnormal return for a one-unit increase in the CEO linguistic-stress score, a stronger relationship in prepared remarks than Q&A, and no predictive power over the stated 2–7 or 2–180 trading-day horizons. These are thesis-reported results, not independently replicated performance evidence. The repository describes linguistic patterns in text; it does not establish an audio voice model, acoustic feature set, deception detector, or fund deployment. See the capture note.
• Chicago Booth’s Fama-Miller project page for “Optimal Narratives for Predicting Stock Returns” identifies accounting PhD student Alex G. Kim and describes Proximal Policy Optimization fine-tuning of LLaMA-2 to generate summaries optimized for stock-return prediction. The page reports a pilot in which optimized summaries contain more forward-looking information and strategies using them have higher Sharpe ratios than strategies using raw documents. It does not provide the document universe, point-in-time cutoff, reward construction, benchmark, costs, checkpoint, code or replication. This is a public model-objective hypothesis, not evidence of a named manager’s system or live performance.

These additions make the academic route more useful for diligence: classroom-scale measurement, text-only executive-stress extraction, and outcome-optimized narrative compression should be tracked as separate modalities. The next checks are source-data provenance, clock alignment, contamination/leakage tests, role and prepared/Q&A stratification, costs, and independent replication. Academic work, student projects and professor biographies should not be read as evidence of any tracked firm’s deployment, authority or results.

September 5, 2026 — HKUST exposes a hands-on ML equity-investing syllabus

• HKUST’s Spring 2023 ACCT 4720 syllabus describes Equity Investment with Machine Learning as a hands-on course covering quantitative-investment process, multifactor stock selection, portfolio construction, performance evaluation and AI/ML applications in equity investing. Students were required to evaluate, design, build and backtest equity strategies through assignments and a final group project; the stated learning outcomes include strategy design, backtesting and performance evaluation.
• The syllabus names Professor Haifeng You and says he is a co-director of HKUST’s Center for Securities Analysis with Financial Technology. It records prior quantitative-equity investment work at Barclays Global Investors and a Director of Quantitative Equity Research role at China Investment Corporation. These are dated syllabus and biography claims; they do not establish that the course, center, BGI, CIC or another manager used a specific model, dataset or production workflow.

Boundary: this is historical academic curriculum and biography evidence. The syllabus does not disclose student project results, model versions, datasets, live portfolio authority, current employer systems, production deployment or independently verified performance. See the capture note.