This page separates public strategy disclosure, model-risk language, academic evidence, and research hypotheses. It does not reveal private models or establish performance.
Source status: prospectus and public-disclosure audit initiated 2026-08-15. The strongest new evidence comes from SEC filings for registered funds and advisers.
Source ledger: sources/06-industry-verticals/hedge-fund-ml-predictive-edge-prospectus-audit-2026-raw.md.
Executive answer
The public record is not empty. It is uneven.
Private hedge-fund filings rarely expose model mechanics. Registered-fund prospectuses do expose more: model family, forecast target, data modality, horizon, event universe, human override, or model-update language. Those documents create a useful bridge between broad AI claims and testable research questions.
The new prospectus evidence identifies several concrete predictive surfaces:
- four-week relative price appreciation and forecast uncertainty;
- predicted earnings growth, dividend growth, volatility, and valuation;
- event impact around earnings, filings, insider transactions, M&A, and restructurings;
- text and audio extraction for sentiment and return forecasting;
- risk-adjusted-return ranking using purchased and public alternative data;
- short interest and stock-borrow demand;
- alternative data from transactions, search, traffic, and social sentiment;
- short-term market regimes and multi-asset allocation;
- LLM-based financial-text classification, earnings-guidance analysis, and AI-opportunity assessment; and
- clinical-trial approval, delay, and market-reaction forecasts.
These disclosures do not prove that a particular model works, that the model is deployed across a firm, or that it generates returns after costs. They do give a more precise map of where to look and what to backtest.
Why prospectuses are a valuable new source
Private hedge funds usually do not publish a retail-style prospectus. A Form D may show that a fund exists. A Form ADV may identify an adviser, private fund, or business practice. Neither normally describes a live model.
Registered closed-end funds, interval funds, mutual funds, and ETFs have a different public surface. Their prospectuses describe principal strategies and risks. When a manager chooses to state that ML, NLP, LLMs, alternative data, or AI-driven forecasts are part of the process, the language can reveal the target without revealing the implementation.
The right interpretation is “public strategy disclosure,” not “independent model validation.”
Prospectus evidence map
| Manager or fund | Publicly disclosed ML/AI surface | What the document exposes | Boundary |
|---|---|---|---|
| AQR Funds / AQR Delphi | ML and NLP in certain strategies; alternative data; proprietary long/short quantitative models | Data categories, model/data testing language, quality/low-beta/value themes, and a 2026 Delphi filing | The AI paragraph is broad; the Delphi filing is preliminary and has no operating history in the document |
| Counterpoint Quantitative Equity ETF | ML ranking with gradient-boosted trees and neural networks | 30+ input variables, value/reversal/momentum/profitability/sentiment/stability, model diversity, backtesting | No current weights, live calibration, or independently audited performance |
| QRAFT AI-Enhanced U.S. Large Cap ETF | Licensed QRAFT AI; related LG-QRAFT system | 100–200-stock pool, four-week relative appreciation, Bayesian uncertainty, top-50 selection | Registered-fund strategy disclosure; not proof of a hedge-fund deployment |
| Intelligent Alpha ETFs | Human-initiated AI quantitative and qualitative selection | Human-defined universe, data, themes, philosophy, and periodic review | Foundation model and training details are not disclosed |
| Sparkline Capital | ML and computing on large unstructured datasets | Firm identity, CIO lineage, Boston quant-fund history, and broad ML purpose | Limited mechanics; historical filing context |
| Draco Evolution AI ETF | Draco Model: ML regime model plus macro model | Technical indicators, macro variables, short-term regime/volatility, and stated 70% asset-selection/weight contribution | Fund-of-funds ETF; model performance and current live behavior require separate testing |
| Bluerock AI200+ Future Leaders ETF | Ensemble of gradient boosting, neural nets, and adviser-fine-tuned LLM | Financial-text classification, sentiment, earnings guidance, AI-opportunity assessment, composite conviction, PM override | Fund-controlled disclosure; no weights, labels, or independent replication |
| FINQ AI-managed ETFs | Proprietary adaptive AI framework | Large-cap ranking/weighting and a dollar-neutral long/short variant | Adaptive is a stated design description, not a published validation protocol |
| Sterling / Guardian Capital | ML, deep learning, and LLMs | Predicted earnings growth, dividend growth, volatility, valuation, and alternative factors | Task-to-model mapping is not disclosed |
| Goldman Sachs Asset Management | NLP/ML for text and audio extraction within quantitative themes | Forecast-return score, sentiment, business quality, market themes, portfolio optimizer, execution costs | No specific audio model or earnings-call deployment is named |
| Rayliant NxtGen Multifactor ETFs | ML models estimate risk-adjusted returns and determine constituent weights | Purchased and public-web alternative data, proprietary data cleaning, automated and manual checks, human discretion around news and liquidity, U.S. and emerging-market variants | Registered ETF disclosure; the filing does not name model architecture, weights, or an independent evaluation |
| FS Event Driven / FS Market Neutral | ML and big data for event-impact prediction | Earnings, calls, filings, insider transactions, M&A, spinoffs, restructurings, management changes | Event labels, model architecture, and results are not public |
| Blackstone Alternative Multi-Strategy | Some managers’ model, data, execution, and risk systems may use ML | Proprietary/licensed technology and manager discretion over data subsets | Multi-manager risk disclosure, not a named strategy |
The most actionable public model targets
Relative price appreciation over short horizons
QRAFT’s disclosure is unusually explicit about a four-week relative price-appreciation target and forecast uncertainty. That target is different from generic “AI stock picking.” A research replication should ask whether uncertainty, model disagreement, and changes in the forecast distribution add information beyond momentum and traditional factors.
The necessary controls include point-in-time universe membership, delistings, corporate actions, transaction costs, turnover, borrow availability, and a clean separation between training features and revised financial data.
Earnings and dividend forecast revision
Sterling’s prospectus names predicted earnings growth and predicted dividend growth. GSAM’s filing describes forecast-return scores and text/audio extraction. Together, those disclosures point toward a layered test: extract facts and guidance from text/audio, compare them with the prior market expectation, and measure the revision surprise rather than raw sentiment.
The relevant unit is not “positive call.” It is the change in a forecast relative to what was available before the call, routed by speaker and section, and joined to the correct market clock.
Event impact across filings, calls, and corporate actions
FS’s event-driven filing describes a wide event universe: press releases, earnings calls, regulatory filings, insider transactions, M&A, spinoffs, restructurings, management changes, and other transformative events. It says quantitative models may use ML and big data to predict the impact of those events.
This is a promising taxonomy for a research corpus because it naturally produces event timestamps and outcome labels. The hard problem is expected-versus-surprise measurement. An event model can predict that a merger announcement matters while failing to predict the direction once the market’s prior probability is included.
Alternative data and demand nowcasting
AQR’s public document names consumer transactions or behavior, social-media sentiment, and internet-search or traffic data. These modalities can support demand, attention, or business-momentum estimates. They also carry unusual revision, access, privacy, licensing, and vendor-survival risks.
A credible test needs to preserve the source’s original timestamp and revision history. Re-running a backtest over today’s cleaned vendor database is not equivalent to reproducing what an investor knew then.
Market regimes and cross-asset allocation
Draco’s prospectus describes a regime/volatility model using technical indicators and a macroeconomic model using labor, housing, orders, money, rates, and expectations. The model’s output selects and weights ETFs across equities, bonds, gold, currencies, and other asset classes.
This makes the public target a state-classification and allocation problem rather than a stock-ranking problem. The right tests include regime persistence, allocation turnover, leverage, tail behavior, and performance under release-timing constraints.
Text, audio, and qualitative financial factors
Bluerock’s filing describes an LLM fine-tuned on financial text for classification, sentiment extraction, earnings guidance analysis, and AI-opportunity assessment. GSAM describes text/audio extraction within a quantitative return model. These documents provide public evidence that qualitative financial information can be routed into investment research, but they do not prove that the system uses voice deception detection or that language features are incremental after standard factors.
The productive next step is ablation: words only, metadata only, audio only, and multimodal combinations, all with point-in-time splits and multiplicity correction.
Risk-adjusted-return models with explicit data hygiene
Rayliant’s 2026 prospectus for its NxtGen Multifactor U.S. and Emerging Markets ETFs adds a useful implementation detail to the alternative-data category. The adviser says its ML models estimate expected risk-adjusted returns and determine selection and weighting through a rules-based optimization process. It describes purchased vendor data and public-web data, a proprietary cleaning process, and automated and manual checks. It also preserves human discretion to adjust trades based on news, liquidity, or portfolio-team insight.
That combination creates a more specific test surface than “AI stock selection”: data provenance, cleaning, model-to-weight mapping, and the boundary between model output and human adjustment. The prospectus does not disclose the architecture, training sample, feature list, or live model evaluation. (Rayliant prospectus)
Where private hedge-fund artifacts fit
The prospectus corpus complements the firm-specific artifacts already tracked in the hedge-fund research program:
| Public firm artifact | Research surface | What it adds to the prospectus map |
|---|---|---|
| Bridgewater PAT and Interrupt recording | Document/time-series retrieval, code generation, multi-agent investigation, validation | Architecture and control-loop evidence rather than a portfolio-factor disclosure |
| Man AlphaGPT and AlphaTrend | Hypothesis generation, code implementation, candidate-signal testing | Research-production workflow evidence rather than a fund prospectus target |
| Balyasny embeddings and merger forecaster | Retrieval embeddings and tool-using merger-probability forecasting | Model/data artifacts with paper-level evaluation, not public fund disclosure |
| Two Sigma feature forecasting | Timestamp-aware LLM feature discovery and forecast workflow | Practitioner evidence that connects model use to research process |
| Numerai Predictive LLM | Article-derived structured features and tournament research | A released firm model artifact and public feature surface |
| CFM financial NER | Narrow financial entity extraction | A compact domain model with a specific information-extraction task |
The prospectuses add target labels and disclosure language. The private-firm artifacts add system architecture, model training, agent control, and workflow details. A cross-reference should never turn one into evidence for the other.
New research ideas generated by the audit
- Forecast dispersion: measure disagreement across model families, forecast uncertainty, and portfolio-manager overrides around events.
- Signal handoff: test whether extracted text/audio features improve earnings or dividend forecasts after standard quantitative factors.
- Event-chain prediction: link clinical, regulatory, patent, financing, licensing, and M&A states instead of modeling each event independently.
- Borrow-aware event models: combine short interest, borrow demand, options skew, and event surprise.
- Alternative-data decay: measure signal half-life, vendor revisions, access outages, and crowding.
- Model-change detection: treat portfolio exposure shifts around disclosed model updates as observable events.
- Regime-to-risk budgeting: evaluate whether regime classification changes risk allocation without relying on leverage or hindsight.
- Commercial translation: connect clinical approval probability to reimbursement, adoption, claims utilization, and label breadth rather than treating approval as the final label.
- Research-agent economics: measure whether agents change idea throughput, validation time, or error rates while preserving a human approval gate.
- Cross-manager disclosure graph: link prospectus adviser, sub-adviser, portfolio manager, vendors, papers, jobs, and firm recordings into a dated entity graph.
Evidence boundaries
- The public record contains more evidence of ML use than of ML performance.
- A registered-fund prospectus is a disclosure artifact, not an independent audit.
- AI risk language may apply to an adviser or manager platform broadly and may not identify the exact fund using ML.
- A model family or forecast target does not establish live deployment, profitability, or causal alpha.
- A public ETF, interval fund, mutual fund, asset manager, and private hedge fund are distinct legal and operating categories.
- No ranking or “ahead/better” assessment is made.
- “No public evidence found in this pass” is not “does not exist.”