AI can predict 71% of what your fund manager does
- Robin Powell
- 10 hours ago
- 9 min read
Harvard researchers have built an active fund manager AI — a machine-learning model trained on public data — and found that it can predict 71 per cent of what a human fund manager does next. The most predictable managers, it turns out, are the worst performers. So what exactly are you paying for when you buy an active fund?
Picture yourself settling into seat 14C for a long flight. The captain comes on the intercom to explain, with a reassuring baritone, that he'll be hand-flying the aircraft today. Good news, you think. A real pilot, earning his keep.
Then you learn what every pilot knows. On a typical flight, the autopilot does something close to 90 per cent of the actual flying. The captain's real value lies in the moments autopilot can't handle — the unexpected crosswind, the iced-up runway, the last-second diversion.
Now think about what your active fund manager actually does all day.
You're paying somewhere between 0.75 and 1.0 per cent of your portfolio a year because you believe a human expert is exercising skilled judgment on your behalf. A tracker would cost you a fraction of that — 0.07 to 0.20 per cent, depending on the index. The gap between the two is the price of the judgment. The premium for the pilot, if you like. It's also far more than any genuine alpha an active manager can plausibly justify.
A new paper from a Harvard-led team of researchers suggests that most of that judgment isn't judgment at all. It's routine. And a machine, armed with nothing more than publicly available data, can do most of it for you.
Your fund manager is on autopilot for seven trades in ten
An active fund manager AI trained only on past trading patterns can predict 71 per cent of what a human manager does next.
That's the central finding of Mimicking Finance, published in February 2026 by Lauren Cohen of Harvard Business School with Yiwen Lu and Quoc H. Nguyen. The method is disarmingly simple. Take a manager's publicly available trading history. Add the characteristics of the stocks they trade — size, sector, style, valuation. Feed it all into a machine-learning model. Then ask the model, for any stock in the manager's universe, whether the next move will be a buy, a hold or a sell.
Crucially, the model never sees a single current trade. It is guessing in advance, based on pattern alone.
Seven times out of ten, it gets the answer right.
A naïve benchmark — simply predicting each trade will follow the majority pattern — does little better than chance. The AI model gets you to 71 per cent. For some managers, in some quarters, nearly every trade is predictable.
It helps to be clear about what 'predictable' means here. The AI is forecasting trade direction — buy, hold or sell — not trade size or future returns. This is a classification problem, not a crystal ball. But it's an awkward finding for anyone selling active management as skilled judgment, because the judgment is being replicated by an algorithm that has never met the manager and knows nothing about their thought process. If beating the market is already closer to a game of chance than a game of skill, Cohen's work just narrowed the band of where any residual skill might live.
Think of it as the autopilot portion of the flight. Seven trades in ten are routine enough for a machine to anticipate. The remaining three in ten — the genuinely novel decisions — are where any actual judgment lives.
The trades the AI predicted were also the ones that lost money
Routine trading doesn't just fail to add value. It subtracts value.
Cohen, Lu and Nguyen sort fund managers into quintiles by how predictable their trading is. The pattern is almost monotonic: the least predictable managers outperform their peers, and the most predictable underperform. It's not a messy relationship. It runs in a straight line from skilled to scripted.
A long-short portfolio — buying the stocks traded by the least predictable managers and shorting the stocks traded by the most predictable — generates 424 basis points of annualised alpha.
More damning still is what happens inside a single manager's portfolio. The researchers split each fund's holdings into the positions the AI could predict and the positions it couldn't. The unpredictable positions — the ones the model was most uncertain about — outperform the predictable ones by a cumulative 0.79 per cent over four quarters.
The implication is hard to dodge. Whatever genuine skill active managers possess appears to live in the narrow slice of trades they make that don't fit their own historical pattern. The rest — the routine 71 per cent — is replicable at near-zero cost. And it is the part of the portfolio that drags performance down. It also lines up with Morningstar's recent finding that fees are now the most reliable predictor of future fund performance, period. What investors are paying for in the routine 71 per cent is, in effect, underperformance insurance.
If you believe in the value of active money management, the moments when the pilot takes the controls are where the real value lives. The autopilot portion isn't just neutral. In a world where you're paying for a pilot, it's the part that costs you.
How fund managers become creatures of habit
Three structural forces explain why so many active managers spend most of their time on autopilot. None of them is flattering to the industry.
The first is tenure. The longer a manager has been running money, the more predictable their trading becomes. Habits harden. Instincts calcify. An active fund manager AI trained on 15 years of trades learns a manager's patterns in a way the manager themselves might not recognise.
The second is competition. Managers in crowded, competitive categories — where differentiation is the price of survival — are harder to predict. Managers in quieter corners of the market, where peer pressure is gentler, slip into routine more easily. Competition forces novelty. Its absence permits habit.
The third is skin in the game. Managers who have significant personal ownership of their own fund — real money, their own money, at stake — trade in ways the AI finds harder to anticipate. They appear to exert more genuine effort. When it's not your money, it's easier to buy what you bought last quarter.
Underneath all three findings sits a single structural fact about the fund management business. The economics reward gathering assets, not generating insight. A manager can be highly predictable, modestly underperforming, and still retain billions under management for years, because the sales and distribution apparatus has already done the hard work of getting investors into the fund, and inertia does the rest. Active management still dominates the global asset pool, and still collects almost all the fees paid for fund management, precisely because that gathering machine is so effective.
Imagine an airline that paid its pilots the same hourly rate whether they hand-flew the aircraft or used the autopilot. You'd expect to find most pilots reaching for the autopilot button most of the time. That is roughly how the fund management industry works.
UK fund managers face the same verdict
The Cohen paper studies US funds. The UK story is the same story, and the Year-End 2025 SPIVA Europe Scorecard, published by S&P Dow Jones Indices, tells it with brutal clarity.
Start with the one-year numbers. Even in 2025 — a year of tariff turmoil, AI-driven dispersion and low correlations, precisely the conditions that should have favoured stock pickers — 89 per cent of UK large- and mid-cap equity funds underperformed their benchmark. In UK small-caps, where active managers claim the biggest informational edge, 97 per cent underperformed. The S&P Dow Jones Indices' long-running study of active versus passive performance found majority underperformance in 18 of 21 European equity categories.
Stretch the horizon and the picture worsens. Over ten years, between 93 and 95 per cent of UK equity funds — depending on the category — failed to beat their benchmarks. Euro-denominated Europe equity funds fared no better: 97 per cent underperformed over a decade. Globally, the pattern held across virtually every major market.
Those long-term numbers are flattered by survivorship. Only 41 per cent of UK large- and mid-cap funds that existed ten years ago survived to the end of the measurement period. More than half simply disappeared — merged away or quietly liquidated, usually after sustained poor performance. The true odds of picking a fund that both survives and outperforms over a decade are worse than the headline figures suggest.
Now add the fees. A typical UK active equity fund carries an ongoing charges figure of 0.75 to 1.0 per cent. A mainstream tracker can cost as little as 0.06 per cent on the FTSE All-Share. The average active manager starts every year roughly 0.8 percentage points in the hole before a single trade is made, purely because of what the fund costs to run. Most investors spend more time scrutinising an energy bill than they do their investment fees, even though the fund fees are usually the more expensive item over a lifetime.
If 71 per cent of what that manager does can be replicated by an active fund manager AI trained on public data, the value proposition on the other side of the fee becomes very thin indeed.
The regulator has noticed. Under the Financial Conduct Authority's Consumer Duty regime, 'Price and Value' is a live supervisory priority, and the assessment of value regime for authorised funds is set for further policy attention through 2025 and 2026. The Cohen paper doesn't change the rules. But it hands anyone who cares to use it a sharper tool for asking the central Consumer Duty question: is what this fund does actually worth what it charges?
How to stop paying pilot prices for autopilot
The practical response to all of this is not complicated.
If 71 per cent of active trading is routine, the rational move is to capture that 71 per cent at tracker prices, and to reserve any active fees — if you choose to pay them at all — for the narrow sliver of managers who consistently trade in ways a machine can't anticipate.
Consider the arithmetic. A £200,000 portfolio invested for 20 years at a 7 per cent gross annual return. In a tracker charging 0.15 per cent, it grows to roughly £753,000. In an active fund charging 0.85 per cent, it grows to roughly £660,000. The difference — £93,000 — is the compounded price of the autopilot portion, dressed up as judgment. That's not a rounding error. It's a house deposit, a decade of holidays, or three years of retirement income, handed to a fund manager for work that a machine can replicate.
The obvious objection is: fine, but what about the good managers — the ones in Cohen's least predictable quintile, whose novel trades genuinely outperform? The problem is identifying them before the fact, not after it. The latest evidence on fund performance persistence confirms what decades of research have shown: a fund's past outperformance tells you almost nothing about whether it will outperform in future. A manager who looks unpredictable today may simply be going through a noisy patch before reverting to routine. And even among the minority who do add value, most of that apparent skill turns out to be statistically indistinguishable from luck. You'd need to watch a manager for decades — longer than most funds survive — to be confident the outperformance was real.
For most investors, the conclusion is the conclusion it has been for a long time. A low-cost, globally diversified index fund captures market returns in full, without paying active fees for routine trading dressed up as expertise.
The autopilot and the pilot
The Cohen paper does not argue that active management is worthless. It argues something more precise. It shows, for the first time, which part of active management is worthless, and which part isn't.
Again, if you believe in active management, the pilot still matters, in the moments that matter. But most of the flight, and most of the trading, is autopilot — and the industry has been charging hand-flying prices for an autopilot-heavy service for decades.
You don't need to find the rare, genuinely skilled manager. You can capture the market's returns at a fraction of the cost, pay less in fees, and sleep better knowing that the 71 per cent you've declined to pay for was never worth paying for in the first place.
The best investment decision most people can make isn't finding a better pilot. It's noticing that the flight was on autopilot all along.
Resources
Cohen, L., Lu, Y. & Nguyen, Q.H. (2026). Mimicking Finance. NBER Working Paper No. 34849.
Cohen, L., Malloy, C.J. & Nguyen, Q.H. (2020). Lazy Prices. The Journal of Finance, 75(3), 1371–1415.
Cremers, K.J.M. & Petajisto, A. (2009). How active is your fund manager? A new measure that predicts performance. Review of Financial Studies, 22(9), 3329–3365.
Autor, D.H., Levy, F. & Murnane, R.J. (2003). The skill content of recent technological change: An empirical exploration. The Quarterly Journal of Economics, 118(4), 1279–1333.
S&P Dow Jones Indices. (2026). SPIVA Europe Scorecard: Year-End 2025. S&P Global.
One logical next step
If this has made you think about whether your current approach to investing is actually working, our Find an adviser directory is a good place to start. Everyone listed has publicly committed to low-cost, globally diversified investing — the kind of approach that reinforces what you've read here, rather than quietly undoing it.
Stay connected: YouTube | LinkedIn | X For advisers: Connect with Robin about content