Executive Summary
- The leak is not a live portfolio. It is the research machine: what data a shop watches, how candidate features are created, which controls reject bad signals, and where humans still hold authority.
- The strongest public signals come from named practitioners at Two Sigma, Man Group, Acadian, Balyasny, Bridgewater, CFM, Jane Street, Hudson River Trading, and Versor. They disclose enough workflow detail to map the shape of serious quant AI without disclosing tradeable alpha.
- The recurring pattern is research-funnel expansion. LLMs and agents create more hypotheses, extract more features, monitor more sources, and write more code. The bottleneck moves to point-in-time provenance, leakage control, out-of-sample validation, execution checks, and governance.
- The useful output is a leak ledger, not an alternative-data catalog: fund, leaked clue, implied data source, implied control, and what not to infer.
The Leak Ledger
| Shop | What actually leaked | Implied data/source surface | Implied control | Do not infer |
|---|---|---|---|---|
| Two Sigma | LLMs sit upstream in equities feature forecasting. Ben Wellington gives the example of turning an observed Target hiring sign into a cross-company hiring feature, then warns that naive full-sample beta tests overfit history. | Observed real-world events, hiring indicators, text/tables/images, company-level feature stores. | Temporal validation: “what would I have known in 2011 using only 2000-2010 data?” | Do not infer LLMs pick stocks or that any disclosed feature works. |
| Man Group / AHL | ManGPT is a safe/audited internal LLM layer; Alpha/Rosa are the platform substrate; Man Numeric uses ML in roughly a quarter of signal models; Man emphasizes data provenance, ArcticDB, graph-composed strategies, and co-pilot before autopilot. | Earnings-call NLP, Reddit/meme-stock monitoring, market/order data, sector data, Python/open-source/proprietary code, ArcticDB/BQuant-like data frames. | Audited internal access, provenance, explainable trade rationale, human augmentation. | Do not infer autonomous trading authority or live alpha from ManGPT/Alpha Assistant/AlphaTrend. |
| Acadian | AI modules feed an existing systematic process: pattern detection across a broad stock universe, analyst-bias correction, earnings-call Q&A extraction, supplier/customer context, expected-return forecasts, portfolio construction, then order routing. | Earnings calls, analyst estimates, newsflow, supplier/customer and peer fundamentals, technical trading patterns, portfolio/risk/cost data. | Domain knowledge, train/validation/out-of-sample discipline, modular signal validation, transaction-cost and risk constraints. | Do not infer generic ChatGPT portfolio construction works. |
| Balyasny | The public Balyasny-style leak is the central quant services model around PM pods: factor models, hedging, PM advisory, performance diagnosis, risk/drawdown support, and execution research. LLM-generated factor definitions need point-in-time provenance. | PM/pod history, factor libraries, risk exposures, execution data, proprietary datasets, performance/drawdown logs. | Factor tests for pervasiveness, persistence, interpretability, diversified reproduction, and point-in-time definitions. | Do not infer an autonomous PM agent or disclosed Balyasny alpha. |
| Bridgewater | AIA Labs exposes a staged research-agent workflow: causal-map agent, data-finder agent, coder/chart agents, planner/supervisor/subagents, approved-plan retrieval, critique pane, editable code pane, and human-diagnosable oversight. | Bridgewater databases, causal maps, prior approved plans, code, charts, research reports, Textract/Bedrock/EKS-like infrastructure. | Causal reasoning over correlation, blueprint/chain-question planning, human critique, inspectable generated code. | Do not infer public demo behavior equals audited production trading. |
| CFM | CFM leaks research discipline: statistical significance, luck-vs-skill separation, Sharpe/risk framing, model clusters, decorrelation tests, rejection of correlated ideas, alternative-data investment, ML tools, shared research/data platform, cloud compute. | Alternative data, strategy/model libraries, portfolio correlations, cloud research platform, systematic strategy data. | Reject ideas that are not statistically significant, independent, or decorrelated from existing clusters. | Do not infer CFM disclosed LLM agents or strategy performance. |
| Jane Street | Jane Street leaks the research-to-production shape: exploration, data collection, modeling, productionization; low-data/high-noise markets; manual checks before scale; simple baselines before deep models; hardware and model shape co-design. | Market data, hand-collected datasets, ticker/split/error checks, research tools, production systems, hardware/serving constraints. | Out-of-sample belief discipline, data-quality checks, baseline ladder, productionization gate. | Do not infer public LLM-agent trading disclosure. |
| Hudson River Trading | HRT leaks the market-making control boundary: low-level market events are the raw ingredient; neural nets predict short-horizon behavior; audited risk-checked layers, not the neural net, act in markets; LLM historical news backtests are contaminated. | Order-book/market event data, petabyte-scale storage, GPU training, model serving, routing systems, release/intraday checks. | Release checks, intraday sanity checks, numerical stability, regulatory trust, separation of prediction from execution. | Do not transfer HFT evidence to long-horizon fundamental investing. |
| Versor | Versor leaks the junior-researcher agent pattern: agents read papers, implement ideas, structure conviction evals, use podcast/transcript sentiment as a research medium, and combine off-the-shelf with fine-tuned/open models. | Papers, podcasts, transcripts, factor ideas, model/code artifacts, structured eval logs. | Paper-to-implementation loop, structured conviction scoring, model/task fit. | Do not infer podcast sentiment is proven alpha. |
What The Leaks Say About Non-Standard Data
The non-standard sources that matter are the ones the research machines imply, not a generic vendor menu.
| Leaked source | Who points to it | Why it matters |
|---|---|---|
| Hiring and labor traces | Two Sigma’s Target hiring-sign example; talent leakage across official firm pages | Hiring becomes a measurable company-level feature only if timestamped and comparable across the traded universe. |
| Earnings-call and transcript NLP | Man Group, Acadian, FactSet, multiple finance-platform sources | Calls are not just summaries; they become Q&A features, sentiment/semantic features, contradiction checks, and management-consistency inputs. |
| Reddit / meme-stock monitoring | Man Group | Social data is treated as risk and market-pulse input, not a clean long-only alpha feed. |
| Expert-network and management-meeting transcripts | Shu Bai, FactSet | These feed thesis checks: demand, chips, token costs, management claims, and conference contradictions. Rights and MNPI posture matter. |
| Podcasts and YouTube | Shu Bai, FactSet, Versor | Public long-form media becomes a research source for market pulse, source discovery, and sentiment when transcripted and timestamped. |
| Causal maps and internal research plans | Bridgewater | Prior approved plans and causal relationships become retrievable research artifacts. |
| Order-book / market event exhaust | HRT, Jane Street | At short horizons, low-level market events matter more than generic alternative data. |
| Platform telemetry and code artifacts | Man Group, Bridgewater, Jane Street, HRT | Serious shops leak that the moat is not the model alone; it is data capture, code, serving, audit, and production controls. |
The Real Thesis
The public hedge-fund podcast circuit leaks four things:
- Where the data enters: hiring traces, transcripts, social/media, market events, expert calls, firm databases, and code artifacts.
- Where AI enters: feature ideation, transcript extraction, paper-to-code translation, source monitoring, thesis contradiction, causal-map drafting, and research-code generation.
- Where AI is stopped: point-in-time checks, out-of-sample tests, factor decorrelation, audit/provenance layers, execution gates, and human critique.
- Where the moat lives: proprietary data context, research platforms, reusable internal libraries, data lineage, productionization, and governance.
The funds are not leaking the trade. They are leaking the machine that decides whether a trade idea deserves to exist.
What To Build Next
The next product should be a weekly research-machine leakage tape with one row per leaked clue:
{
"fund": "Two Sigma",
"leaked_clue": "LLMs used for equities feature forecasting; hiring observations can become company-level features",
"source_artifact": "TWIML transcript",
"implied_data_source": ["hiring/labor traces", "text/table/image feature extraction"],
"implied_control": ["point-in-time validation", "overfitting checks"],
"confidence": "HIGH",
"what_not_to_infer": "No disclosed alpha or live strategy"
}
Rank by specificity, recency, source strength, named practitioner credibility, and non-consensus operating detail. Penalize generic AI enthusiasm, host framing, vendor marketing, and claims without a named workflow or control.
Sources
- Hedge Fund Research Signal Landscape
- Buy-Side Quant AI Practitioner Signals
- Finance and Quant Audio/Video Deep-Dive Roadmap
- Two Sigma LLMs for Equities Feature Forecasting
- Man Group’s Data-Driven Future
- Acadian Zhe Chen AI Investment Process
- Balyasny / Odd Lots Multi-Strat Quant Interview
- Bridgewater AIA Labs At AWS re:Invent
- CFM Systematic Multi-Strategy
- Jane Street Signals and Threads
- Hudson River Trading on Odd Lots
- Vinesh Jha on Mining Alternative Data
Brandon Sneider | brandon@brandonsneider.com June 2026