Top Hedge Funds and Quant Shops: AI Agents, Workflow Signals, and Failure Modes

See also (wiki): wiki/quant-asset-management-ai.md · asset-management-quant-ai-gap-map-2026.md · buy-side-quant-ai-practitioner-signals-2026.md

Source credibility: LOW-MEDIUM to MEDIUM. TIER 2-3. Public hedge-fund AI disclosures are asymmetric. Balyasny, Man AHL, Two Sigma, Bridgewater, Schonfeld, Millennium, D. E. Shaw, Tower Research Capital, WorldQuant, XTX, CFM, Squarepoint, PDT, Jump Trading, Citadel/Citadel Securities, Jane Street, Point72/Cubist, HRT, Voleon, G-Research, QRT, Winton, and Renaissance Technologies have useful public AI/ML, quant-platform, or research-infrastructure signals. Only a smaller subset discloses agent-specific workflows. Do not infer hidden deployments from hiring, rumor, generic AI commentary, or raw compute spend.

Executive Summary

• The clearest public pattern is AI as research acceleration infrastructure, not autonomous portfolio management.
• The most concrete agent disclosures are Balyasny’s OpenAI research engine, Man AHL’s Alpha Assistant / AlphaTrend distinction, Two Sigma’s LLM feature-forecasting, 2026 outlook, and May 2026 ACM CAIS agentic-systems material, Bridgewater’s AIA Labs, Jump Trading’s LLM-agent tooling page, and Schonfeld’s FE AI Lab.
• The industrial-ML shops disclose less about chat/agent UX and more about infrastructure: XTX price-forecasting at 50,000+ instruments, CFM ML Lab and 6+ PB data, Squarepoint automated strategy implementation, PDT’s research-test-live loop, Citadel’s DSG/EQR research platforms, Jane Street’s CoreWeave AI cloud commitment, HRT’s advanced research environment, Voleon’s ML-first investment process, G-Research’s quant-ML platform, QRT Labs’ foundation-AI / agentic-systems partnership, and Winton’s idea-to-live systematic research infrastructure.
• Jump Trading now has one of the stronger public non-vendor signals for production AI: its official AI/ML page says ML is integrated across trading, research, and infrastructure, with deep learning, RL, LLM, generative modeling, custom foundation models, and LLM agents integrated across tools and data.
• Bridgewater and Schonfeld move out of the quiet-firm bucket. Bridgewater’s official AIA Labs page connects AI tools to portfolio management, trading, risk, and investment processes, and says AIA systems now manage billions while generating alpha; treat that as a company claim, not audited proof. Schonfeld’s FE AI Lab page describes PM/analyst workflow training, proprietary tools, model partnerships, and structured pilot gates.
• Millennium, D. E. Shaw, and Tower also move out of the quiet bucket. Millennium has a named Global Head of AI and AI advisory group. D. E. Shaw has an official optimizer / human-machine investment-process signal. Tower’s May 2026 article gives a concrete agent-harness, governance, token-budget, and zero-trust-permission view for capital-markets AI.
• The failure modes are now consistent across sources: overfitting, temporal leakage, hallucinated research claims, formula/state-tracking errors, generic assistants with no firm context, governance gaps, and unproven live alpha.
• The new AQR OCR capture strengthens the negative-evidence lane. AQR’s “Can Machines Learn Finance?” is older than genAI, but it cleanly explains why finance return prediction is small-data and low-signal-to-noise, why flexible models overfit, and why economic theory plus human expertise remain necessary controls. It is now frozen as source-acquisition fixture sa-120.
• CFA and Aon add the external diligence lens: governance is not a soft appendix. Professional and allocator sources now expect manager AI programs to disclose workflow location, policy, data privacy, responsible sourcing, staff training, bias checks, human review, and monitoring.
• AIMA’s 2025 alternative-investment GenAI report adds the hedge-fund allocator lens with disclosed methodology: 150 fund managers representing about US$788B AUM plus 18 large institutional investors. It shows near-universal GenAI use, rising front-office expectations, investor DDQ pressure, and a governance gap among smaller managers. Treat this as adoption/governance evidence, not alpha proof.
• The U.S. Senate HSGAC June 2024 report adds a government/regulatory evidence lane. Staff received information from Citadel, Renaissance Technologies, Bridgewater, AI Capital Management, Numerai, and WorldQuant and found inconsistent AI/ML terminology, inconsistent human-review boundaries, high-level client disclosures, uneven testing/review practices, and a need for operational baselines, version control, risk assessments, standardized audits, and audit trails.
• The 2026 agentic-AI-in-finance survey adds the market-structure lens. Once agents coordinate, adapt, and act across trading, portfolio, risk, or compliance workflows, evaluation must include system-level stability, liquidity, market impact, shock recovery, accountability, and continuous validation. A single agent’s task accuracy is not enough.
• XTX and G-Research add the infrastructure diligence lens. XTX’s TernFS disclosure shows that research-scale ML depends on storage architecture, immutable artifacts, snapshots, and external permission systems. G-Research’s May 2026 LLM code-review article shows a production pattern for treating LLM output as untrusted input: structured JSON, source-of-truth validation, two-pass verification, provider abstraction, cost telemetry, and behavior-level tests.
• Official talent, hiring, hackathon, and university-partnership pages now have their own promoted source lane. They are strong enough to reveal workflows, tooling, governance vocabulary, and skill demand; they are not strong enough to prove alpha or production agent reliability.
• Renaissance now moves out of the unsupported watchlist, but only as an official quant/HPC research-infrastructure signal. Its public careers pages and brochure support mathematical/statistical investment process, research/data-processing software, CPU/GPU/HPC, C++/Rust, and scientist/programmer infrastructure claims. They do not disclose LLM agents or genAI alpha.
• Publicly quiet firms should stay in the ingestion watchlist, not in the findings column.

What Top Shops Are Doing

What Is Not Working

1. General Assistants Are Too Shallow Without Firm Context

Man AHL’s distinction is the cleanest public evidence: broad assistants help with coding and exploration, but narrow workflows like AlphaTrend create depth, reproducibility, and auditability. The useful system has a predefined workflow, modular stages, firm code/data context, and explicit evaluation gates.

Gary Collier’s Tech Talks Daily interview strengthens the same point from the platform side. ManGPT is described as a safe/audited access layer for LLMs, while Alpha/Rosa, data provenance, graph-composed strategies, and ArcticDB are the substrate around the assistant. The system is not just a model; it is the model inside a governed research and data platform.

2. More Hypotheses Can Make Overfitting Worse

Two Sigma’s 2026 outlook is unusually direct: AI widens the research funnel, so the bottleneck moves from idea generation to idea evaluation. If agents can generate more hypotheses and run more backtests, they can also increase false discoveries unless temporal validation, holdout discipline, and independent review improve at the same time.

Its May 2026 ACM CAIS article adds the agent-systems failure surface. The firm-authored signal is that durable value lives in the harness around the model: architecture/composition, evaluation, security, optimization, and engineering operations. For finance, that means an agent can fail through bad tool documentation, schema drift, unsafe runtime actions, policy violations hidden inside traces, artifact contamination across PDFs/spreadsheets/decks, or poor model routing even when the final memo reads well.

3. Temporal Leakage Is A New LLM-Specific Risk

Two Sigma explicitly flags pretrained-model knowledge cutoffs as a concern in forecasting pipelines. A model may “know” future facts when asked to simulate a historical research decision. This is the LLM version of look-ahead bias.

4. Financial Artifact Generation Is Still Not Fully Reliable

FrontierFinance shows why hedge-fund agent claims need skepticism. Frontier agents can be much faster than human analysts on long-horizon financial modeling, but they still fail on formula dependencies, cell-reference state tracking, and auditability. That failure pattern maps directly to investment research artifacts: spreadsheets, backtests, and investment memos can look plausible while being structurally wrong.

5. Governance Is Behind Deployment Velocity

Deloitte’s investment-management outlook reports that AI governance language in job postings remains generic and non-AI-specific even as firms scale AI. MSCI’s IndexAI guide shows the correct direction: tool-call provenance and entitlement checks should be explicit. Hedge-fund agents need the same controls for internal data, licensed data, and investment evidence.

CFA and Aon sharpen this failure mode. CFA’s 2025 materials require explainability, validation, auditability, accountability, monitoring, and human judgment across the investment AI lifecycle. Aon’s 2026 survey page reports that heavier AI users have stronger formal governance and that 90% of surveyed managers agree accountability currently requires strict human-in-the-loop review. A hedge-fund AI disclosure without these controls is incomplete diligence evidence.

AIMA’s 2025 report adds a sharper hedge-fund-specific control. It reports that 95% of surveyed fund managers used GenAI, but more than 60% had formal restrictions and half of smaller managers reported no restrictions. Investor scrutiny is also becoming explicit: 29% of surveyed institutional investors already asked GenAI DDQ questions and another 29% expected to add them. For StateBench, that means a credible finance agent should answer due-diligence questions about policy, access controls, sensitive-data handling, model oversight, explainability, and compliance before any workflow claim is trusted.

The U.S. Senate HSGAC staff report makes the same control surface regulatory rather than allocator-only. It found that hedge funds and regulators do not use uniform definitions for AI, ML, expert systems, algorithmic systems, and optimizers; all six funds reported human review, but no uniform requirement or clear intervention point exists; and investor disclosures were often too high-level to explain how systems are developed, tested, reviewed, or monitored. For StateBench, a finance agent should lose credit if it says “human in the loop” without specifying when, by whom, against what test, with what audit trail, and under what client/regulatory disclosure.

6. Multi-Agent Finance Can Create System-Level Failure

The agentic-finance survey sharpens the risk beyond individual hallucination. When multiple learning agents optimize local objectives, system-level behavior can change: market depth and resilience become endogenous, agents can withdraw liquidity together, similar signals can create correlated behavior, and individually rational policies can produce collective fragility. The benchmark implication is that finance agents need stress tests for liquidity, market impact, execution cost, regime shift, shock recovery, and coordinated behavior, not just a score on a single research or trading task.

7. Alpha Claims Remain Unproven In Public

No public source in this corpus independently proves genAI-driven live trading outperformance. Balyasny publishes workflow and efficiency. Man publishes research-workflow architecture. Two Sigma publishes feature-workflow and evaluation warnings. Bridgewater publishes the strongest company claim: AIA systems have real-capital deployment, manage billions, and generate alpha. That is important, but it is still company-disclosed rather than independently audited public attribution. Schonfeld publishes FE workflow transformation, not trading outcomes. XTX/CFM/Squarepoint/PDT publish infrastructure/process signals. None publish reproducible alpha attribution from genAI agents.

Man Group’s Collier transcript explicitly supports this caution: the guest says the industry has not visibly produced a “killer finance app” and frames most GenAI use cases as co-pilot / human augmentation. For benchmark design, the right question is whether an agent can produce lineaged, explainable, reviewable research artifacts, not whether it claims autonomous alpha.

AIMA’s investor evidence points in the same direction: investors may reward meaningful GenAI budget and research implementation, but AIMA also warns that credibility comes from measurable action rather than AI washing. Adoption statistics, front-office expectations, and investor optimism should therefore feed governance and evaluation tasks, not performance claims.

8. Compute And Industrial ML Are Not Agent Proof

Jane Street’s official ML and performance pages are stronger than a generic compute-spend signal. They describe neural-network trading models, mostly-noisy market data, ultra-low-latency processing, regime shifts, self-impact from the firm’s own trading, tens of thousands of GPUs, exabyte- scale storage, production trade study, and inference latencies far below human-timescale ML. The CoreWeave deal adds capacity evidence: large-scale ML, noisy financial data, continuous model refinement, and trading/research operations.

This still should not be read as evidence of a disclosed LLM-agent workflow. Treat compute commitments, GPU fleets, custom storage, low-latency runtime engineering, and AI roles as capacity and deployment-constraint signals until a source describes the workflow, tools, permissions, evaluation boundary, and failure modes.

9. Public Benchmarks Often Do Not Match Trading Reality

HRT’s ML-benchmark article is one of the better public failure-mode sources from a top trading firm. The core lesson is that finance teams care about simplicity, reproducibility, and generality under low signal-to-noise constraints. A model or agent can win a public benchmark and still fail on trading value, robustness, or implementation cost.

Two other official HRT articles make this directly useful for StateBench. HRT’s “Applying Artificial Intelligence to Trading” argues against a monolithic “trading AI” abstraction: markets are not clean board games, market state is partially hidden, order outcomes are uncertain, and there is only one slowly arriving reality rather than infinite self-play data. HRT decomposes the problem into prediction, optimization, and execution. Its data article adds the source-acquisition gates a finance agent should pass before research begins: relevance, uniqueness, lookahead avoidance, sample-size/noise checks, and data provenance.

For StateBench, HRT should become a negative/control check: penalize agents that collapse prediction, portfolio construction, and execution into one fluent recommendation; penalize agents that accept hindsight-tagged, widely-crowded, undersampled, or noisy alternative data without provenance and as-of-date checks.

The promoted Odd Lots transcript sharpens the operational boundary. Dunning describes HRT-style AI around short-horizon market data, low signal-to-noise prediction, and full-stack integration from data capture to storage, GPU training, serving, and market routing. The critical control point is that the neural network is not treated as directly sending market orders; audited, risk-checked layers act on model output. That turns HRT into a useful StateBench control for release checks, intraday sanity checks, numerical-stability checks, regulator-facing trust controls, and contamination checks when an LLM is used to backtest historical speeches or news.

10. Finance ML Is Small-Data And Low-Signal-To-Noise

AQR’s “Can Machines Learn Finance?” provides the Acadian-adjacent skeptical baseline that was still too thin in this corpus. The useful point is not just “avoid overfitting.” AQR argues that return prediction is constrained by the number of independent return observations available for the target, while much of the “big data” discussion in finance focuses on adding more predictors. That is exactly the trap an LLM research agent can worsen: it can generate thousands of features, explanations, and backtests without adding independent evidence about the target.

For StateBench, this should become a scored failure mode. A model should lose credit when it equates more alternative data, more generated hypotheses, or a bigger neural model with more reliable return prediction. It should gain credit when it asks about independent observations, turnover/trading costs, out-of-sample validation, economic structure, interpretability, and whether ML is being used for a better-suited task such as risk, transaction cost, or portfolio construction rather than raw alpha discovery.

The standalone sa-120 fixture preserves the evidence boundary. This is an official AQR source and useful negative control, but it is not GenAI-agent evidence, not a current model benchmark, not alpha proof, and not a backend portability claim for CUDA, MLX, GGUF, SageMaker, Bedrock, or AWS Neuron.

10. Infrastructure Failures Can Masquerade As Model Failures

XTX and G-Research make a separate point: many apparent model failures are actually platform failures. If a research artifact is mutable, partly written, unversioned, unauthenticated, or disconnected from the source of truth, the model cannot make it trustworthy. G-Research’s standalone sa-130 source card adds the sharper LLM failure mode: if a finding is not validated against an authoritative rule index, structured schema, and behavior-level test harness, it can be valid JSON and still be wrong.

StateBench should therefore score infrastructure behavior alongside model quality: artifact hashes, dataset snapshots, permissioned retrieval, rule-ID validation, bounded repair, provider switching, cost logging, and non-blocking human review.

11. Regulatory Definitions And Audit Trails Are Still Underspecified

The Senate report adds a gap that practitioner podcasts rarely cover: even where funds say they use AI/ML in research, pattern identification, portfolio construction, or trading-decision support, the terms are not consistent enough for clients or regulators to reason about risk. It also records a practical backtesting warning from AI Capital Management: financial markets have only one historical path, so overfitting historical data is a structural risk rather than a nuisance.

For benchmarks, this means “the model produced a good backtest” is too weak. The artifact should record the system category, use case, human-review stage, testing cadence, version lineage, audit trail, client-disclosure boundary, overfitting controls, and whether the test says anything about unprecedented market regimes.

Watchlist And Promotion Rules

No currently tracked top-fund source should be promoted to agent-specific conclusions unless it describes the workflow, tools, permissions, evaluation boundary, and failure modes. Renaissance is now source-backed for quant/HPC research infrastructure, but remains unproven for agent-specific conclusions.

These firms have now moved to verified AI/ML or infrastructure signal, but not agent-workflow proof:

• Bridgewater
• Citadel / Citadel Securities
• D. E. Shaw
• G-Research
• Millennium
• Point72 / Cubist
• Qube Research & Technologies
• Jane Street
• Hudson River Trading
• Renaissance Technologies
• Schonfeld
• Tower Research Capital
• Voleon
• Winton

Bridgewater is a verified investment-process AI signal: its official AIA Labs page says AI tools are used in portfolio management, trading, portfolio risk management, and other investment-management processes, including through AIA, and says AIA systems now manage billions while generating alpha. Schonfeld is a verified investment-team workflow signal: its FE AI Lab page describes PM/analyst training on earnings prep, idea generation, document analysis, inbox triage, Excel workflows, proprietary systems, model partnerships, and structured pilot evaluation. Neither should be promoted to public proof of live alpha.

Millennium has a verified AI leadership/advisory signal, D. E. Shaw has a verified human-machine optimizer signal, and Tower has a verified capital- markets AI / agent-harness governance signal. These are valuable architecture and operating-model evidence, but still below the threshold for agent-workflow performance claims.

Jump Trading should be tracked as agent-specific tooling signal because its official AI/ML page explicitly mentions LLM agents integrated across tools and data. It still does not prove autonomous trading authority or live alpha.

G-Research, QRT, and Winton should be tracked as verified adjacent quant-ML signals. G-Research is now a standalone production-LLM infrastructure signal: source-of-truth validation, two-pass recall/precision filtering, provider abstraction, bounded repair, cost telemetry, and non-blocking human review. QRT Labs is a 2026 academic frontier-AI / agentic-systems partnership signal; Winton is an older but useful failure-mode source for selection bias, interpretability, and the journey from idea generation to live trading. These are useful benchmark design inputs, not agent-performance claims.

Renaissance should be tracked as a verified quiet-firm quant/HPC infrastructure signal. Its official pages are useful for source-discovery and talent-leakage tasks because they expose research software, statistical-model infrastructure, CPU/GPU and distributed-computing skill demand, and official-site authority. They do not describe a current AI-agent workflow.

Promotion rule: move a firm into the agent-findings column only when an official page, named practitioner transcript, paper, report, or verified conference talk describes the AI workflow, tool boundary, permission model, evaluation loop, or failure mode.

LinkedIn and job-post rule: official company LinkedIn posts and official job listings can strengthen a workflow/platform claim, especially when they expose tooling, hiring, architecture, controls, or user workflow. Third-party LinkedIn commentary should remain source discovery unless it points to a primary source.

Talent/hiring leakage rule: use quant-ai-talent-hiring-leakage-2026.md as the promotion filter for official careers, hackathon, leadership, and university-partnership material. The benchmark may score whether an agent finds the official source, classifies the claim type, extracts controls, and converts the signal into an eval task. It should not reward models for amplifying uncorroborated social posts or inferring hidden deployments.

What To Benchmark Because Of This

StateBench finance tasks should include:

1. Firm-context research assistant: use repo conventions, internal source ledgers, and evidence tiers.
2. Feature discovery: turn messy source material into timestamp-safe candidate features.
3. Backtest scaffolding: generate code/configs that execute and preserve assumptions.
4. Temporal discipline: answer only with information known at the simulated date.
5. Spreadsheet/model audit: detect formula, linkage, and state-tracking errors.
6. Tool provenance: show whether an answer came from licensed connector, internal corpus, public web, or model memory.
7. Abstention: refuse unsupported alpha, Sharpe, or investment claims.

Sources

• OpenAI, “How Balyasny Asset Management built an AI research engine”: https://openai.com/index/balyasny-asset-management/
• Man Group, “A Trend Following Deep Dive: AlphaTrend and Agentic Research Workflows”: https://www.man.com/insights/alphatrend-agentic-research-workflows
• Man Group, “A Trend Following Deep Dive: AI, Agents and Trend”: https://www.man.com/insights/ai-agents-trend
• Man Group / Anthropic partnership: https://www.man.com/man-group-anthropic-partnership
• Tech Talks Daily, “3192: Inside the AI-Powered Hedge Fund: Man Group’s Data-Driven Future”: https://techblogwriter.libsyn.com/man-group
• Local promoted extract: …/13-multimodal-sources/tech-talks-daily/2025-02-26-man-group-data-driven-future.md
• Two Sigma, “AI in Investment Management: 2026 Outlook (Part I)”: https://www.twosigma.com/articles/ai-in-investment-management-2026-outlook-part-i/
• Two Sigma, “AI in Investment Management: 2026 Outlook (Part II)”: https://www.twosigma.com/articles/ai-in-investment-management-2026-outlook-part-ii/
• TWIML, “LLMs for Equities Feature Forecasting at Two Sigma with Ben Wellington”: https://twimlai.com/podcast/twimlai/llms-for-equities-feature-forecasting-at-two-sigma/
• Bridgewater, “AIA Labs: The Future of Investment Intelligence”: https://www.bridgewater.com/aia-labs
• Bridgewater, “Artificial Intelligence”: https://www.bridgewater.com/research-and-insights/artificial-intelligence
• Bridgewater LinkedIn, “Senior Product Engineer, AIA Labs”: https://www.linkedin.com/jobs/view/senior-product-engineer-aia-labs-at-bridgewater-associates-4400923727
• Schonfeld, “Inside Schonfeld’s FE AI Lab”: https://www.schonfeld.com/insights/inside-schonfelds-fe-ai-lab/
• Millennium, “Gideon Mann”: https://www.mlp.com/people/leadership/gideon-mann/
• Millennium, “Technology”: https://www.mlp.com/people/technology/
• D. E. Shaw, “Machine Teaching: What I Learned From My Optimizer”: https://www.deshaw.com/library/machine-teaching
• D. E. Shaw, “Investment Approach”: https://www.deshaw.com/what-we-do/investment-approach
• Tower Research Capital, “AI in Capital Markets: A Practical Perspective from Tower’s Ramit Sawhney”: https://tower-research.com/ai-in-capital-markets-a-practical-perspective-from-towers-ramit-sawhney/
• Tower Research Capital LinkedIn company page / AI panel post: https://fr.linkedin.com/company/tower-research-capital
• WorldQuant, “How AI Is Changing Who Gets to Compete in Quant Finance”: https://www.worldquant.com/ideas/how-ai-is-changing-who-gets-to-compete-in-quant-finance/
• XTX Markets homepage: https://www.xtxmarkets.com/
• XTX Markets, “TernFS”: https://www.xtxmarkets.com/tech/2025-ternfs/
• CFM, “Our approach”: https://www.cfm.com/our-approach/
• Squarepoint, “About Us”: https://www.squarepoint-capital.com/about
• PDT Partners, “Work”: https://pdtpartners.com/work
• Jump Trading, “AI/ML”: https://www.jumptrading.com/ai-ml
• Jump Trading, “Technology”: https://www.jumptrading.com/technology
• DDN, “DDN Selected by Jump Trading to Accelerate AI Quantitative Trading at Massive Scale”: https://www.ddn.com/press-releases/ddn-selected-by-jump-trading-to-accelerate-ai-quantitative-trading-at-massive-scale/
• Citadel, “Quantitative Research”: https://www.citadel.com/our-teams/quantitative-research-team/
• Citadel, “Inside EQR: Building the Future of Systematic Investing”: https://www.citadel.com/careers/career-perspectives/inside-eqr-building-the-future-of-systematic-investing/
• Citadel, “Inside EQR: What It Means to Be a Quantitative Researcher”: https://www.citadel.com/careers/career-perspectives/inside-eqr-what-it-means-quantitative-researcher/
• Citadel Securities, “Quantitative Research”: https://www.citadelsecurities.com/careers/quantitative-research/
• Citadel Securities, “Machine Learning Researcher - PhD Graduate”: https://www.citadelsecurities.com/careers/details/machine-learning-researcher-phd-graduate-us/
• CoreWeave, “Jane Street Signs $6 Billion AI Cloud Agreement with CoreWeave”: https://www.coreweave.com/news/jane-street-signs-6-billion-ai-cloud-agreement-with-coreweave
• Jane Street, “Machine Learning”: https://www.janestreet.com/join-jane-street/machine-learning/
• Jane Street, “Quantitative Research”: https://www.janestreet.com/quantitative-research/
• Jane Street, “Performance Engineering”: https://www.janestreet.com/performance-engineering/
• Hudson River Trading, “In Trading, Machine Learning Benchmarks Don’t Track What You Care About”: https://www.hudsonrivertrading.com/hrtbeat/trading-machine-learning/
• Acadian Asset Management Inc., “2026 Investor Forum”: https://ir-forum-2026.acadian-inc.com/home/default.aspx
• Acadian 2026 Investor Forum source card: …/…/sources/06-industry-verticals/acadian-2026-investor-forum-systematic-scale-raw.md
• AQR, “Can Machines Learn Finance?”: https://www.aqr.com/Insights/Research/Journal-Article/Can-Machines-Learn-Finance
• AQR “Can Machines Learn Finance?” source card: …/…/sources/06-industry-verticals/aqr-can-machines-learn-finance-raw.md
• Local AQR OCR: raw/aqr-can-machines-learn-finance-2019.chandra.txt
• Agentic AI in Finance survey source card: …/…/sources/06-industry-verticals/agentic-ai-finance-survey-2026-raw.md
• U.S. Senate HSGAC, “AI in the Real World: Hedge Funds’ Use of Artificial Intelligence in Trading”: …/…/sources/06-industry-verticals/us-senate-hedge-funds-ai-ml-2024-raw.md
• Point72, “Cubist Systematic”: https://point72.com/cubist/
• Point72, “Market Intelligence”: https://point72.com/market-intelligence/
• Point72, “Students & Early-Career”: https://point72.com/students-early-career/
• Point72, “Five Years of the Cubist Quant Academy”: https://point72.com/blog/five-years-of-the-cubist-quant-academy/
• Voleon: https://voleon.com/
• Voleon ML-first investment-process source card: …/…/sources/06-industry-verticals/voleon-ml-first-investment-process-raw.md
• G-Research, “About us”: https://www.gresearch.com/about/about-us/
• G-Research, “NeurIPS paper reviews 2025 #1”: https://www.gresearch.com/news/neurips-paper-reviews-2025-1/
• Qube Research & Technologies, “QRT Labs”: https://www.qube-rt.com/qrt-labs
• Winton, “Different approaches to quantitative investing”: https://www.winton.com/news/experiment-and-observation-in-quantitative-investment
• Winton: https://www.winton.com/
• Renaissance Technologies: https://www.rentec.com/
• Renaissance Technologies, “Research Engineer”: https://www.rentec.com/Careers.action?jobs=true&selectedPosition=researchEngineer
• Renaissance Technologies, “Research Infrastructure Programmer”: https://www.rentec.com/Careers.action?jobs=true&selectedPosition=researchInfraProgrammer
• Renaissance Technologies careers brochure: https://www.rentec.com/pdf/careers_brochure.pdf
• Renaissance careers brochure OCR: raw/renaissance-technologies-careers-brochure.chandra.txt
• FrontierFinance note: …/21-benchmarks/frontierfinance-2026.md
• Raw source ledger: …/…/sources/06-industry-verticals/buy-side-quant-ai-practitioner-signals-2026-raw.md
• CFA/Aon governance note: cfa-aon-asset-manager-ai-governance-2026.md

Brandon Sneider | brandon@brandonsneider.com May 2026