Ninety-five percent of surveyed hedge fund managers now use generative AI [1]. That stopped being an edge about the time it became a survey result. The edge is what you build around it - and how honestly you measure it.
Educational content, not investment advice. Nothing here is a recommendation to buy, sell or hold anything, or to hire any manager. Illustrations use historical or simulated data; past results do not predict future results.
How to read this article: claim versus corroborated
Fund marketing about AI is hard to check. So every external statement below carries one of three tags:
- [CORROBORATED] - The fact is confirmed by a primary or independent record: a regulator's order, a peer-reviewed or archived paper, a survey publisher's own release.
- [CLAIM] - The firm or a vendor says it, in its own press release or in an interview. We confirmed it was said. We did not and could not confirm it is true or that it makes money.
- [UNVERIFIED] - Circulating in secondary sources but we could not trace it to a primary source. We do not rely on it.
Notice the trap in the CLAIM tag. "Man Group announced a partnership" is corroborated. "The partnership generates alpha" is not - and nobody outside the firm can know. Keep that distinction in your head for the rest of this piece.
The headline everyone is quoting
The Alternative Investment Management Association (AIMA) surveyed 150 fund managers (about US$788 billion in assets) and 18 large institutional investors, and published in September 2025 [1]:
- 95% of managers use generative AI, up from 86% in 2023.
- 58% expect to use it more inside the investment process over the next year, up from 20% in 2023.
- 60% of institutional investors said they would be more likely to back a manager that devotes meaningful budget to it.
- 29% of those investors already ask about generative AI in due diligence questionnaires, and another 29% planned to add it that year.
- Among managers under US$1 billion, 50% reported no restrictions at all on using generative AI tools.
[CORROBORATED] as survey results published by AIMA; they describe stated usage, not performance. "Use" can mean anything from drafting emails to generating signals. The last bullet is the quiet alarm: adoption has run ahead of governance.
So "do you use AI?" now gets a yes from everyone. The allocator's next question is sharper: show me how an AI idea becomes capital, and how you know it is not noise.
WHAT ALLOCATORS ASK, IN ORDER OF DIFFICULTY
"Do you use AI?" -> everyone says yes (95%)
|
"Where, exactly?" -> ops? research? the portfolio?
|
"What is your baseline?" -> beats a simple model, net?
|
"How many ideas did you try?" -> the trial count (see below)
|
"Who can stop it, and how fast?" -> controls, logs, kill switch
What funds publicly say about their AI
Read each of these as a claim about intent and method, not as evidence of returns.
Bridgewater - [CLAIM] on method, [CORROBORATED] that it was said. Fortune reported in July 2024 that Bridgewater was launching a fund of nearly US$2 billion in which machine learning is the primary basis for decisions, combining proprietary technology with models from OpenAI, Anthropic and Perplexity, with humans helping on risk, data and execution. Co-CIO Greg Jensen was candid that language models have a hallucination problem [2]. Some aggregator sites quote specific returns and error-rate improvements for this fund; we could not trace them to a primary source, so they are [UNVERIFIED] and we leave them out.
Man Group - [CLAIM] on capability, [CORROBORATED] that the partnership exists. On 11 February 2026 Man Group announced it was deploying Anthropic's Claude across investment and operational processes, with the stated aim to "harness AI for alpha generation". It describes AlphaGPT, an in-house alpha idea generation tool, as part of that effort; its CTO said a key attraction was describing workflows "in plain English" [3]. Trade press describes AlphaGPT as proposing signals, writing code and backtesting them before a human committee decides; that description comes from press coverage and interviews, not an audited disclosure, so treat the mechanics as [CLAIM] [4].
Two Sigma - [CLAIM], thin evidence. Its public careers pages seek researchers who apply deep learning, large language models and reinforcement learning to noisy, complex data [5]. That corroborates hiring intent and nothing more. We found no first-party, quantified description of how language models feed live portfolios. Several secondary pages quote compute and data-source figures; we treat those as [UNVERIFIED].
What the regulators have corroborated about false claims. In March 2024 the SEC settled the first "AI washing" cases against two advisers. Delphia (US$225,000 penalty) had said it used AI and machine learning on client data to make its investment predictions "smarter"; the SEC found it did not have those capabilities. Global Predictions (US$175,000) had marketed itself as the "first regulated AI financial advisor"; the SEC found the representations false [6]. [CORROBORATED] - settled orders, neither admitted nor denied. Total penalties: US$400,000.
The pattern across the credible names: AI as a component inside a rigorous research pipeline, with humans and controls around it. The pattern in the enforcement cases: AI as a sentence in the brochure.
The stack that wins: five parts, one rule
- Old models (factors, risk, costs, optimisers): the backbone and the benchmark.
- LLMs: read filings, extract features, write code.
- Agents: do multi-step research in a sandbox.
- RAG: memory with citations, filtered by as-of date so it cannot see the future.
- MCP servers: narrow, read-only, logged doors between the AI and your data, positions and tools.
The rule: models propose, controls decide. No agent holds the keys to an order.
DATA (point-in-time) RESEARCH (sandbox) DECISION
+----------------+ +-------------------+ +------------------+
| prices, filings|--MCP--> | LLM + agents | | validation gate |
| macro, news | read- | propose features, |----->| baseline beaten? |
| as-of dated | only | code, backtests | | trials counted? |
+----------------+ +-------------------+ | holdout passed? |
^ | +--------+---------+
| every step logged |
| v v
+----------------+ +-------------------+ +------------------+
| RAG memory |<---------| audit log | | HUMAN + RISK |
| with citations | | prompts, versions | | limits, sizing, |
+----------------+ +-------------------+ | kill switch |
+--------+---------+
No arrow runs from the agent box to the order box. |
v
EXECUTION (rules)
The uncomfortable math: research is cheap, so luck is cheap
Agents make research nearly free. That is exactly the danger. In our own simulation (strategies with zero true edge, three years of daily data), the best of 1,000 random strategies shows a Sharpe ratio near 1.85; the best of 5,000 near 2.1. Nothing is there. An agent can try thousands before breakfast. (Author's simulation; the code is in the companion book.)
Best backtest Sharpe among N strategies with ZERO true edge (simulated)
N tried best-looking Sharpe
------- -------------------------------------------------
1,000 ##################################### ~1.85
5,000 ########################################## ~2.1
true edge (nothing) 0.00
Each # is about 0.05 Sharpe. More effort, better-looking luck.
This is old news in academia, which is why it should embarrass a team that ignores it. Harvey, Liu and Zhu argued that, after decades of data mining across hundreds of published factors, a new factor should clear a t-statistic of about 3.0, not the conventional 2.0 [7]. Bailey, Borwein, Lopez de Prado and Zhu formalised the probability that a strategy chosen as best in-sample underperforms out-of-sample, and the minimum backtest length needed to resist this [8]. Bailey and Lopez de Prado's Deflated Sharpe Ratio then corrects a Sharpe for the number of trials [9]. [CORROBORATED] - published research.
We also ran an honest end-to-end test on free Kenneth French factor data [10]: 36 timing rules, selected on 1990-2014, held out 2015-2026, costs on. The winner had an in-sample Sharpe of 1.07 and a held-out Sharpe of -0.44. The multiple-testing math said it was probably not luck. The holdout said it did not matter. (Author's analysis; replicate it.)
1990 ============================ 2014 | 2015 ======== 2026
SELECT the best of 36 rules | HOLDOUT, touched once
in-sample Sharpe +1.07 | held-out Sharpe -0.44
^
the only number you should believe
Language models add their own trap. Research by Gao, Jiang and Yan finds LLM forecasts look better on events the model has effectively seen in training, and the edge fades after its knowledge cutoff [11]. A 2024 paper reporting GPT-4 beating analysts at earnings prediction was withdrawn in February 2025 after replication problems [12]. Treat every LLM backtest as an upper bound. [CORROBORATED].
What goes wrong when teams go agentic on hype
Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls, and noted widespread "agent washing" - rebranding chatbots as agents [13]. That is a [CLAIM] in the sense of a forecast by an analyst firm, but it matches what is seen in practice. The failures are boringly repeatable. They are also testable.
THE EIGHT FAILURE MODES THE TEST THAT CATCHES IT
1 Multiple testing / overfitting ---> count trials; deflate Sharpe
2 Look-ahead leakage ---> as-of dates; shift-by-one test
3 No baseline ---> must beat a dumb model, net
4 No evals ---> fixed test set, scored each run
5 Hallucination ---> cite-or-reject; sample audits
6 Cost blow-ups ---> budget cap per task; log tokens
7 Weak controls ---> read-only tools; human sign-off
8 Demo-driven scope ---> pre-registered success metric
1. Overfitting and multiple testing. Every idea an agent tries is a lottery ticket. Fix: log every trial (including failures), set the success metric before the run, hold out data that is touched exactly once, and report a deflated Sharpe or a probability of backtest overfitting [8][9].
2. Leakage. Future information sneaks into the past: restated fundamentals, survivorship in the universe, normalising on the full sample, or a language model that memorised the outcome. Fix: point-in-time data, as-of filters on retrieval, and an LLM cutoff check (only score events after the model's training cutoff) [11].
3. No baseline. An "AI signal" that cannot beat a 12-month momentum rule, a simple factor tilt or last period's value is not a signal. Fix: always report the dumb model next to the clever one, after costs.
4. No evals. Without a fixed test set you cannot tell whether last week's prompt change helped. Fix: a frozen, versioned evaluation set with known answers, scored automatically on every change.
5. Hallucination. Bridgewater's own co-CIO named it [2]. In research, the damaging form is a confident, uncited number. Fix: cite-or-reject (no claim without a retrieved source), and random human audits with an error rate you track.
6. Cost. Agent loops multiply tokens quietly. Fix: a per-task budget cap, token logging, and a cost-per-validated-insight metric rather than cost per run.
7. Weak controls. An agent with write access to orders or data is an incident waiting for a trigger. Fix: read-only tool access, narrow scopes, full audit logs, human sign-off, and an out-of-band kill switch.
8. Hype-driven scope. The project is chosen because a competitor announced something. Fix: a one-page charter with a decision it will change and a metric agreed before the build.
The pre-flight checklist (print it)
[ ] Success metric and holdout period written down BEFORE the run
[ ] A simple baseline exists and is reported net of costs
[ ] Every trial is logged; the trial count appears in the result
[ ] Data is point-in-time; retrieval is filtered by as-of date
[ ] Test events fall AFTER the model's training cutoff
[ ] Frozen eval set, re-scored on every prompt/model/code change
[ ] Every factual claim carries a retrieved citation, or is rejected
[ ] Tools are read-only; no path from agent to order
[ ] Cost cap per task; spend and tokens are logged
[ ] A named human owns sign-off, and a kill switch exists
Five quick tests you can run this week
- Shuffle test. Randomly permute your signal's dates and rerun the backtest. If it still "works", you are fitting noise or leaking.
- Shift test. Lag the signal by one extra period. A real, tradeable edge degrades gently; a leak collapses.
- Placebo test. Feed the agent a randomised price series and count how many "discoveries" it reports. That is your false-positive rate.
- Cutoff test. Score only events after the model's stated training cutoff. Compare with before. A gap means memory, not skill.
- Cost test. Re-run with realistic spreads, fees and slippage. Many paper edges vanish here.
Twelve ideas to steal (each verifiable, none guaranteed)
- Diff consecutive filings for guidance, risk-factor and tone changes; test as features after costs.
- An agent that reproduces published anomaly papers on free data and reports which survive costs.
- A daily rates brief written from a Treasury-curve MCP tool.
- Positioning plus carry diagnostics from CFTC and FX data.
- Scenario narratives from an LLM, translated to factor shocks by your risk system.
- A red-team agent that argues against every proposal, using only its evidence.
- Automated post-mortems comparing predicted and realised outcomes.
- Statement reconciliation across custodians (family offices: start here).
- Capital-call and look-through exposure tracking from PDFs.
- Policy Q&A over your investment policy statement, with citations.
- A backtest auditor that checks your own code for look-ahead.
- A trade journal that tags each decision and tracks your hit rate by rationale.
Start free, today
You do not need a data budget. The US Treasury publishes its daily yield curve as a CSV [14]. Kenneth French's library gives a century of factor returns [10]. FRED, the ECB, the World Bank, EIA, CFTC and SEC EDGAR cover macro, FX, energy, positioning and filings. The free stack has real gaps (options chains, clean point-in-time fundamentals, bond prices).
In the free Treasury data alone, an agent would have flagged the 10-year minus 2-year spread turning negative on 1 April 2022, bottoming at -1.08 points on 3 July 2023, and last inverting on 5 September 2024 (author's calculation from the Treasury series [14]).
What to do this week
- Funds and desks: write the one-page AI policy and list of approved tools. Add a trial counter to every backtest. Expect due diligence questions about it [1].
- Family offices: put an enterprise-grade assistant on document work first, then RAG on your policy statement.
- Home traders: write your plan before the strategy, paper trade for a pre-set period, and never give a bot withdrawal rights.
- Everyone: describe only the AI you can demonstrate [6].
The bet
The firms that win will build the boring machinery of validation, point-in-time data and control faster than their competitors build flashy agents. Everything else is available to everyone for a few dollars a day. Let the models be fast. Let your controls be slow.
Want a second pair of eyes?
If you are looking for deep expertise or support, I have many write-ups and practical guides. I have worked with individuals leveraging AI/ML and systematic trading for years, and worked in hedge funds/asset management/investment banks.
Contact: alecordev | alecor [dot] dev [at] gmail [dot] com
The full guide, with architecture, tested Python code (an MCP server, a governed agent loop, a point-in-time retriever, and a validation case study), a catalogue of free multi-asset datasets, and a 12-month roadmap, is in the companion book, "Agentic Alpha".
Sources
Retrieved October-November 2026. Numbers are quoted from the sources; "author's" items are our own calculations and not independently corroborated.
- AIMA, press release on generative AI research, 16 Sep 2025: https://www.aima.org/article/press-release-front-office-gen-ai-adoption-shifts-from-if-to-when-for-leading-fund-managers-aima-research-finds.html
- Fortune, "Bridgewater launches $2 billion fund run by machine learning", 1 Jul 2024: https://fortune.com/2024/07/01/bridgewater-2-billion-fund-machine-learning-decision-making-openai-anthropic-perplexity
- Man Group, "Man Group teams up with Anthropic", 11 Feb 2026: https://www.man.com/man-group-anthropic-partnership
- Hedgeweek, "Man Group partners with Anthropic to embed AI across investment processes": https://www.hedgeweek.com/man-group-partners-with-anthropic-to-embed-ai-across-investment-processes/
- Two Sigma careers, AI Research Scientist postings: https://careers.twosigma.com/careers/JobDetail/New-York-New-York-United-States-AI-Research-Scientist-Internship-2026-Summer/13577
- SEC press release 2024-36, 18 Mar 2024: https://www.sec.gov/newsroom/press-releases/2024-36
- Harvey, Liu and Zhu, "... and the Cross-Section of Expected Returns", Review of Financial Studies 2016 (NBER w20592): https://nber.org/papers/w20592
- Bailey, Borwein, Lopez de Prado and Zhu, "The Probability of Backtest Overfitting": https://papers.ssrn.com/abstract=2326253
- Bailey and Lopez de Prado, "The Deflated Sharpe Ratio" (SSRN, 2014): https://papers.ssrn.com/abstract=2460551
- Kenneth French Data Library: https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html
- Gao, Jiang and Yan, arXiv:2512.23847: https://arxiv.org/abs/2512.23847
- Kim, Muhn and Nikolaev, arXiv:2407.17866 (withdrawn): https://arxiv.org/abs/2407.17866
- Reuters via The Star, "Over 40% of agentic AI projects will be scrapped by 2027, Gartner says", 25 Jun 2025: https://www.thestar.com.my/tech/tech-news/2025/06/25/over-40-of-agentic-ai-projects-will-be-scrapped-by-2027-gartner-says
- US Treasury, daily par yield curve rates: https://home.treasury.gov/policy-issues/financing-the-government/interest-rate-statistics
Note on verification: sources 1, 2, 3, 6 were read directly; 4, 5, 13 were identified via search results only; 7-9 were confirmed via search summaries; 9-12 and 14 are carried over from the earlier version of this article and the companion book's reference list and were not re-fetched in this revision.