AI Is Rewriting Alpha: Why Quant Trading Could Make Markets More Fragile

At 9:30 every night, a convenience-store owner notices the same pattern. Dog walkers flood the block, then stop for bottled water on the way home. He starts filling the fridge before they arrive and puts pet treats by the register.

No spreadsheet gave him the idea. He saw something his competitors had not bothered to see, and for a while he earns a better margin because of it.

Now give every shop in the city a cheap AI that spots the pattern. Within days, everyone is stocked with water and dog treats before 10 p.m. The extra margin disappears. If the stores keep ordering against the same signal, some will end up with too much stock and too few customers. What began as an edge becomes a crowded mistake.

Quant finance is moving toward the same trap.

The case for AI in finance starts with efficiency. It can read more filings, test more signals and react faster than human analysts. But the same technology that removes stale mispricings can also make investors more alike. Similar inputs, similar models and similar risk controls can produce similar decisions, especially under stress.

The autoimmune metaphor begins there. Markets correct errors by rewarding investors who find them first. If the correction mechanism becomes too synchronised, it creates a new source of instability: the crowd of investors trying to correct the same thing at once.

When alpha changes sign

Finance has seen the first half of this story many times. An anomaly is discovered, other investors copy it, capital arrives and the return shrinks.

David McLean and Jeffrey Pontiff studied 97 variables that academic research had linked to stock returns. Out of sample, the returns were 26% lower than the original findings. After publication, they were 58% lower. The authors estimated that trading prompted by publication explained 32 percentage points of that decline. Correlations among published predictors also rose after publication. The act of making an idea widely known changed the economics of exploiting it. (McLean and Pontiff, Journal of Finance)

That is ordinary alpha decay. Crowding can produce something worse.

Research by Dong Lou and Christopher Polk finds that common trading in momentum strategies changes the behaviour of the strategy itself. When common momentum trading is low, arbitrage tends to correct underreaction. When it is high, momentum stocks are more likely to reverse. At that point, investors are no longer just harvesting a mispricing. Their collective trading is helping create the next one. (Lou and Polk, Review of Financial Studies)

August 2007 showed how violent that process can become. Quantitative equity funds built around similar valuation, momentum and statistical-arbitrage ideas suffered losses at the same time. Amir Khandani and Andrew Lo later found evidence consistent with coordinated deleveraging: one portfolio's liquidation moved prices against other portfolios holding similar positions, which forced further selling. Goldman Sachs's Global Alpha fund lost 22.5% that August. (Khandani and Lo, Journal of Financial Markets)

The broad US equity market did not suffer anything comparable over the core of the quant sell-off. AQR's later account makes the key point: much of the damage came from relationships among positions held by similar investors, rather than from a sudden collapse in the businesses those positions represented. (AQR)

A factor can die twice. Competition first strips away the easy return. Crowding then makes what remains more fragile and, in some regimes, capable of reversing.

AI makes that lifecycle faster because it lowers the cost of discovery.

Consensus becomes the trade

An earnings report that once occupied an analyst for an afternoon can now be parsed in seconds. A model can compare management language across hundreds of companies, classify central-bank communication and test thousands of hypotheses before a human investment committee has finished its morning meeting.

The IMF reported that AI-related material accounted for 19% of patent applications connected to algorithmic trading in 2017 and more than half in every year from 2020 onward. Its market participants expected much deeper use of advanced AI in investment decisions over the following three to five years. (IMF)

The efficiency gain is obvious. The systemic cost is easier to miss.

In January 2026, the Bank for International Settlements warned that financial institutions using similar datasets, pretrained models and decision rules could respond to shocks in more correlated ways. Faster automation can compress those responses into a shorter window. A market move that once gave different investors time to interpret, disagree and adjust can become a race among systems trained on much of the same history. (BIS)

The models do not need to make identical forecasts. Their risk controls merely need to rhyme. Two funds can disagree about why a stock should rise and still sell it together when volatility jumps, margin requirements change or portfolio correlations break through the same limit.

Once that happens, the exit itself becomes the consensus trade.

Fully autonomous AI trading remains far from the dominant form of institutional investing. In its July 2026 Financial Stability Report, the Bank of England said firms were using more autonomous systems mainly in areas such as research, coding and surveillance. The Bank is working with the BIS Innovation Hub on simulated markets populated by LLM-based portfolio managers partly because the collective behaviour is still something regulators are trying to understand. (Bank of England)

The case, then, is prospective rather than retrospective. Finance is adding a powerful synchronisation technology to a market structure whose crowding and fire-sale mechanics are already well documented.

Cheap intelligence changes where alpha lives

Sanford Grossman and Joseph Stiglitz described the underlying economics in 1980. If prices perfectly reflected all available information, nobody would pay to acquire information in the first place. Markets need some reward for becoming informed. (Grossman and Stiglitz, American Economic Review)

AI changes what counts as expensive information.

Reading public filings, screening securities and testing standard hypotheses are becoming cheaper. As those tasks lose scarcity, the return attached to doing them should fall. The remaining edge has to come from information that still costs time, access, trust, expertise or capital to obtain and interpret.

Some of it will be proprietary data; much of it will be less glamorous.

Joshua Coval and Tobias Moskowitz found that active mutual fund managers earned significant abnormal returns in nearby investments, consistent with an informational advantage from geographic proximity. The result is old, but the logic is newly relevant: information retains value when access is uneven. (Coval and Moskowitz, Journal of Political Economy)

A supplier may report a healthy order book while customers stretch payment terms. A founder can present the same spreadsheet as ten competitors, yet an investor who has watched that founder operate through two crises may judge the survival odds differently. These advantages are economic: the information is costly to create, verify or interpret.

Silicon Valley Bank offers a useful warning about the limits of data alone. Rising rates were public. The bank's securities portfolio was disclosed. Its reliance on uninsured deposits was knowable. What proved lethal was the interaction among duration exposure, depositor concentration, liquidity and the speed of the run. More than $40 billion left on March 9, 2023, and management expected another $100 billion in outflows the next day. (Federal Reserve)

SVB's lesson is simpler: possessing the inputs does not guarantee understanding of the mechanism connecting them.

That distinction becomes more valuable as access to the inputs gets cheaper.

The market starts trading the trader

The next source of alpha may sit one level above the asset.

Traditional security analysis asks what an asset is worth and what could change that value. A crowded, highly automated market adds another question: what will the current holders be forced to do before value has time to matter?

There is already a large literature on this problem. Joshua Coval and Erik Stafford found that mutual funds hit by severe investor outflows were forced to sell existing holdings, creating price pressure in stocks commonly owned by distressed funds. Because flows were partly predictable, other investors had an incentive to anticipate those trades. (Coval and Stafford, Journal of Financial Economics)

Markus Brunnermeier and Lasse Pedersen showed how funding constraints can turn that pressure into a spiral. Falling prices weaken funding liquidity. Tighter funding forces sales. Those sales reduce market liquidity and push prices lower, which can force still more deleveraging. (Brunnermeier and Pedersen, NBER)

AI gives investors better tools for mapping those constraints. A fund does not need to know another manager's exact model if it can infer the conditions under which that manager must reduce risk: leverage, volatility targets, margin calls, redemption pressure, regulatory limits or concentration caps.

That shifts the object of prediction from price alone to the behaviour of constrained holders. The asset still matters, but so do the owner's balance sheet and mandate.

The BIS has warned that AI trading systems may learn to exploit vulnerabilities in other firms' algorithms in ways that are rational for each firm but destabilising in aggregate. (BIS) The implication for investors is uncomfortable. As machines improve at analysing companies, more of the durable edge may come from analysing the constraints around the people and machines that own them.

Keynes's beauty contest becomes a little harsher. The question is no longer only what everyone else will prefer. It is who can still hold the position when everyone wants the door.

AI becomes an admission fee

There is a precedent for technologies that are essential to compete yet poor sources of lasting advantage.

Eric Budish, Peter Cramton and John Shim showed how high-frequency traders spent heavily on speed because tiny latency differences created arbitrage opportunities. Each firm had a rational reason to invest. Collectively, the race consumed resources without giving the industry a permanent edge over itself. (Budish, Cramton and Shim, Quarterly Journal of Economics)

AI can follow the same economic path.

Every serious investment firm will need good models, good data, enough compute and people who know how to use them. For a while, the best implementations will produce material advantages. Competition will copy the useful parts, vendors will package them and employees will carry methods from one firm to another. What begins as edge becomes infrastructure.

The technology remains essential, but its economics start to resemble an admission fee.

The more public-information processing is commoditised, the more valuable the remaining differences become: proprietary information, unusual mandates, patient capital, better liquidity, distinctive domain knowledge and the ability to recognise a crowded trade before the crowd needs to leave it.

A healthy market needs investors to discover errors quickly. But speed and similarity are not the same thing as resilience. If many institutions ingest similar information and react to stress through similar controls, the mechanism that corrects prices can also amplify the correction.

AI will probably remove a great deal of mediocre alpha. That is good for markets. It will also make disagreement harder to manufacture from the same public facts, which raises the price of having a different source of information or a different ability to wait.

When intelligence becomes cheap, difference becomes expensive.

The durable edge will belong less to the investor with the smartest standalone model than to the one who can see when a crowded model consensus has become fragile, understand what will force it to break, and still have the liquidity to remain on the other side.