Market-structure
A trading signal that used to decay over five to seven years now loses half its edge in eighteen months, and AI crowding is the mechanism
An NYU study of nearly one million institutional fund holdings finds portfolios converging as AI adoption spreads, while separate tests show AI trading models beating human accuracy but running 20 to 40 percent daily volatility against a 7 to 15 percent target, evidence that AI is compressing alpha decay faster than it is improving risk discipline.
The clearest empirical case yet that widespread AI adoption is restructuring how markets behave, not just how individual traders perform, comes from an NYU study titled AI-Driven Alpha Decay: Algorithmic Homogenisation, Reflexive Signal Erosion, and the Paradox of Intelligent Markets, by researchers Meng and Chen. Analysing nearly one million institutional fund holdings, they find portfolios becoming measurably more similar to each other as AI adoption spreads through the industry, and their headline estimate is stark: a profitable trading signal that used to take five to seven years to lose half its excess return now loses that much in roughly eighteen months. Their own framing captures the mechanism precisely, that each marginal AI entrant shortens the lifespan of every exploitable pattern at an increasing rate, which describes a market where the tool that helps you find an edge is the same tool eroding everyone else's ability to keep one.
Two smaller studies referenced alongside the NYU work fill in why the homogenisation happens in the first place rather than just how fast. Researchers at the University of Liechtenstein tested ten LLM-based trading models using sentiment analysis over fourteen months through April 2025 and found all of them generated positive returns initially, right up until subtle manipulations of financial news headlines, letter swaps and hidden text invisible to a human reader, fooled every model tested. In the worst case, a single model's return fell by roughly 18 percentage points after a manipulation campaign targeting one stock. Separately, Elm Partners Management tested four AI models including Claude and ChatGPT on simulated S&P 500 and Treasury bond trading and found directional accuracy exceeding 50 percent, matching elite macro traders on that measure alone, but every model ran daily volatility of 20 to 40 percent against a 7 to 15 percent target appropriate to the investor profile being simulated.
That combination, correct direction more often than chance alongside wildly oversized risk-taking, is the more interesting finding than the headline alpha-decay number. Researcher Haghani's own framing is blunt: we train the AIs to be like people, and then we're finding that the AIs are like people, being overconfident and taking too big positions. An industry survey from the Alternative Investment Management Association backs up how fast this is scaling regardless of the risk profile underneath it, with 58 percent of fund managers now expecting increased AI use in their investment processes, up from just 20 percent two years earlier. Put the three findings together and the picture is an industry converging on similar tools, similar training data and similar blind spots simultaneously, faster than it is building the risk controls to manage what happens when a large share of the market is exposed to the same failure mode at once.
The honest caveat, which the underlying researchers themselves flag, is that all three studies rely on simulations, controlled experiments or bounded datasets rather than live market stress tests, and none of them proves AI adoption will destabilise markets outright. Crowded trades in popular names predate AI by decades without producing a structural crisis on their own, and one of the researchers explicitly warns against blindly trusting large language models to make sound decisions rather than treating this as inevitable doom. What the evidence does support, cautiously, is a shift in the right question to be asking. The debate is no longer just whether AI helps an individual investor beat the market, it is what happens to market function itself once a large and growing share of participants are running variations of the same models trained on the same data, converging on the same trades at the same moments, which is a market-structure question rather than a stock-picking one.
Read the original: Bloomberg - A trading signal that used to decay over five to seven years now loses half its edge in eighteen months, and AI crowding is the mechanism. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.