A research report for institutional allocators · Q3 2026
The three principles
Most AI-in-investing claims do not survive a basic empirical test. The harder problem is that the checks an allocator runs were not built to catch the way they fail.
The average AI-in-investing paper an allocator reads is methodologically unsound. There is no shortage of research. Too much of it is bad.
A review of 164 papers on language models in finance, by researchers at Oxford, BlackRock, Chicago Booth, and Florida, found that each of five core biases goes unaddressed in at least 72% of studies. Survivorship bias is handled in barely one paper in a hundred. The field went from 36 papers in 2023 to 250 in 2025. Volume is outrunning rigor.
Nine of ten frontier LLM trading agents have negative stock-selection alpha. The agent that tops the return leaderboard, up 85% over two years, picks stocks worse than chance.
A Tsinghua and Stepfun benchmark, KTD-Fin, ran ten agents over a 548-day window and split every return into market, style, and genuine stock selection. The best of the ten only breaks even on selection, at +0.2%; the other nine run negative, down to −77.8%. The return curve a manager shows you is almost entirely beta and factor exposure. And the sliver that looks like selection is often memory: move the test past the model's training cutoff and the edge collapses, with FinMem's total return falling 72% and QuantAgent's Sharpe 51%.
The best model scores 82.4% on financial spreadsheets. Translated, that is one number wrong in roughly every six. A junior analyst who misread that often would be let go.
And 82.4% is the ceiling, the single best model. Averaged across ten frontier models, accuracy runs from 86.2% on the simplest file to 48.6% on the largest, 152 companies across 8 funds. That is below a coin flip, because the models strip out the layout and visual structure that carry the meaning.
Specification error is the hidden fault line in quant investing. Add a control variable that is itself a consequence of your factor and your returns, and the factor's sign can reverse.
Across 85 Barra risk models, 26 cases show the wrong control flipping a factor's sign: the correct loading on a liquidity factor, +0.08, turns into −0.04. The statistics improve at the same time, so every standard diagnostic points the wrong way. This is not overfitting. The cause is a wrong assumption about how markets work, and no amount of backtesting catches it. It is why a clean fit and a strong track record can be selling you a broken model.
Put a frontier model in the seat of an investment advisor and its recommendations look tailored. They collapse onto a single input.
Across a thousand simulated client profiles, the allocation a leading model recommends is driven mostly by one variable, the client's self-reported risk tolerance, which carries 57 to 88% of the decision. The rest of the client's circumstances barely move it. The output is fluent and confident, and it would fail the suitability standard a human advisor is held to. Letting the model search the web softens the collapse without fixing it, and makes every client's advice look more alike. The failure hides because the answer reads like bespoke judgment.
Pull the failures together and the common thread is your diligence: your checks never see the failures, and quietly reward the part a machine already does for free.
A simple model predicts 71% of an active manager's trades, against 52% for a naive baseline. That predictable 71% is the mechanical part of the job, the part a machine reproduces at no cost, and it is exactly what a questionnaire scores well: long tenure, a clean process, size. The edge is the 29% the model cannot predict, and almost no diligence process is built to find it. The contaminated studies, the returns that are really market exposure, the tools that fail the basic task: none of it surfaces in a standard review, because the review rewards the commodity and never tests for the skill.
A few approaches in AI-driven investing do hold up. Here is where the edge is, and how long it lasts.
A new signal reads company news with a large language model and reports a 3.1 Sharpe, more than double the best factor anyone has catalogued. That number is a stack of best cases: a frontier model, equal weighting, the full small-cap universe, no trading costs, and the entire sample, all at once.
Strip it to what a desk can run, large-cap, value-weighted, point-in-time, after costs, and the Sharpe settles near 1.6, a number the authors themselves report once transaction costs and turnover smoothing are in. Most of that gain comes from lower risk: volatility falls about 70 percent while the average return barely moves. The conservative build also survives a clean look-ahead test, the one most AI-news strategies quietly fail; the flashier numbers run on models that may have memorized the news. The honest version is about half the brochure, and that gap is the diligence this report is built on.
| 2024 net return | 25.45% |
| Sharpe ratio | ~2.75 |
| AUM | ~$550M (2025), scaling toward $1B |
| Participants | thousands, globally |
| Core contributors | ~1%, power law |
Numerai turns prediction into a crowdsourced market: thousands of independent data scientists build models and stake real money on them. A stake-weighted metamodel aggregates the lot, and it consistently beats Numerai's own internal models.
Six thousand stocks allow six thousand factorial possible orderings, more than the atoms in the universe. Crowdsourcing explores an idea space no single team can cover. The edge is the crowd, the staking that forces skin in the game, and the machinery that turns thousands of submissions into one signal. The model is interchangeable. The infrastructure around it is the moat.
| Firm | What they have |
|---|---|
| Hudson River Trading | Its own data center and 100TB+ of proprietary market data. $12.3B in trading revenue in 2025, a record. Bought enough GPUs to strain the supply chain. |
| XTX Markets | A €1B+ data center in Finland (22.5 MW), 25,000+ GPUs, 650 petabytes of storage. Rivals a frontier AI lab's cluster. |
| High-Flyer | The quant fund behind DeepSeek, which repurposed the same GPUs from predicting prices to predicting tokens. |
The durable edge has slid down the stack. The model is the commodity layer now; what is scarce sits beneath it, in assets that took years and billions to build.
A rival can rent the same foundation model by the afternoon. It cannot rent a power contract or the years of proprietary data sitting behind it, because those accrue with time and capital, beyond what any purchase order can buy. That is why the advantage holds while models keep getting cheaper. An edge that lives entirely in a firm's models is renting what it calls a moat.
The most overlooked edge is telling correlation from causation, and it is moving from method to infrastructure.
In a global causal-discovery contest run by ADIA Lab, the research arm of one of the world's largest sovereign investors, entrants were handed 47,000 labeled synthetic datasets, each generated from a known causal graph, and asked to read cause from effect. The best of 1,904 competitors reached 76.7% against a 40% baseline, with an inventive graph-learning model. But that modeling was the commodity, solved by an open crowd. What the institution owned was the layer underneath, the simulation infrastructure that turns known causal structure into labeled data at scale, and the contestants could only borrow it. It is the purest case in this report of the edge living in the system rather than the model.
AI-driven US funds beat their peers by about six percent a year, risk-adjusted, through 2017. After that, by nothing.
Across nearly eight thousand funds, the advantage falls to statistically zero once AI investing stops being rare. This is the pattern to expect from anything that works: a real edge, competed away as it spreads. It is also quietly reassuring on one fear: AI funds move less in lockstep than ordinary funds.
The industry is buying AI by the cohort. Almost none of it becomes edge, because the layer being taught is the layer that gets commoditized.
| Firm | What changed |
|---|---|
| Lord Abbett$248B | Plain-English backtests went from a 70–80% failure rate to 80% first-try success. Work that took weeks now takes days. |
| IDX Advisors<$3M rev | $1M+ saved over three years. Legal tasks cut from 40 hours to 2, at $7 of compute. |
| Manulife$1.3T | Model-agnostic framework, 70%+ internal adoption, weekly office hours for new features. |
| NBIM (Norway)$2T | An AI-ambassador network, 171 projects identified, mandatory training for all staff. |
| Balyasny$32B | Central-bank speech analysis from ~2 days to ~30 minutes. Merger-arb monitoring automated. |
| BlackRock | AI drafts the first version of work across job families, from an email to a pitch to code. |
The clearest, least contested win is speed. These firms are in production, past the pilot stage, and the gains are specific.
The same method repeats across the table. Each firm turned ad hoc prompting into a shared prompt library and a defined workflow, then spread it through internal training and office hours. The gains land hardest on slow, repeatable work, a weeks-long backtest or a forty-hour legal review now done in a fraction of the time. This is what the AI-for-analysts training teaches, and done well it is real research leverage.
In barely two years, nearly every major bank rolled out an internal AI tool to its analysts, and asset managers did the same with research workflows and training. The chart shows how fast, and how uniformly, it happened.
A capability the whole industry acquires at once is table stakes. It is useful, and every desk needs it. It is also the same tool the desk across the street just bought.
When the entire industry rolls out the same kind of tool in the same eighteen months, the tool cannot be where the edge is.
Analysts reach for their favorite Claude cheat-sheets, and the help is real, but limited and not proprietary: the moment a cheat-sheet is shared, whatever edge it carried is gone.
What the tools reliably do is narrower. Give analysts AI and they get faster on the companies they already follow; they do not start covering new ones. A decade of sell-side data shows the same analyst growing far more accurate on names already on the desk while adding essentially none. The tool sharpens what an analyst already covers, and being faster at what everyone already does only keeps a manager on the floor, well short of an edge.
The dangerous failures are the ones that look right. Ask a frontier model to summarize a filing and it can hand back a clean, confident summary that points the opposite way from the document itself. Generation was never the hard part. Verification is.
In a study of S&P 100 filings co-authored by researchers at BlackRock, State Street and J.P. Morgan, an LLM summarizing a 10-Q reversed the call the document made, bullish to bearish or the reverse, on a third of them. The summaries read clean and plausible the whole way. The mechanism is quiet: it keeps the headline number and drops the caveat that frames it. A longer summary does not help, and a different model just tilts it a different way. What worked was structural: generate several summaries and audit them against the source. Inside a live institutional product, that step lifted the forecasting signal while the raw summary degraded it.
The same week a fund could not say what its AI spend returned, one engagement found a third of it was pure waste, producing no research and no signal.
Removing it expanded the research instead of shrinking it. When the cost per validated signal falls by more than half, the desk runs twice the experiments on the same money. The model is the commodity. The infrastructure that runs it cheaply and checks it honestly is the edge. The research points the same way: in an NBER study of news signals, holding the model fixed and changing only which confidence reading the pipeline trusts, its own inner probabilities rather than its stated confidence, raises the Sharpe by about a fifth. The edge sat in the configuration.
The tools to judge an AI claim yourself: the size of the gap, where your diligence stands today, and the test that closes it.
The last three parts argued that standard diligence cannot see an AI edge. This is that gap with a number on it. About 1,500 allocators and their teams have run the SPEC Process Diagnostic, the most rigorous structured assessment of AI-manager diligence available, and the average score is 31 out of 100.
That is the typical allocator near the bottom of the ladder below: able to name the tools but not yet to test whether the system behind them is real. The number flatters the field, since the people who seek out a diligence self-test are the ones already taking it seriously. The ladder shows where you stand and what it takes to climb.
From taking an AI claim at face value to testing it yourself. The average allocator sits at level two. Select a level to see what it takes.
You can independently run the tests: a blinded re-execution, a regime breakdown, an audit-trail reproduction, a base-model swap.
You run the four SPEC questions and can hear evasion from substance across specification, performance, explanation, and configuration.
Performance attribution, clean backtests, process documents, headcount. Rigorous by the traditional standard.
You separate a real AI system from a marketing label, and you ask which models and tools a manager runs.
A clean track record and a familiar model name are enough to say yes. The AI claim is taken as given.
SPEC is four axes: Specification, Performance, Explanation, Configuration. Each gives an allocator one question to put to a manager. Listen past the polish of the manager's answer for whether a real process sits behind the AI claim, or only a story.
"How did you decide which variables to include in your model, and which did you deliberately exclude? What was the reasoning?"
"Walk me through a time your model was wrong. What did you learn about its assumptions, and what did you change structurally, not just parametrically?"
"Walk me through a specific trade from last quarter. Which signals fired, what factors were present, and why did the model make that decision?"
"What happens to your model when you remove one variable? How sensitive are the results to specification changes?"
The same evidence, turned into action, whether you allocate or you manage.
Make the report your standard for judging an AI manager.
Answer the allocator's questions before they are asked.