AI could cut trading signal lifespan to 18 months

New York University researchers Shuchen Meng and Xupeng Chen analyzed nearly one million institutional fund holdings and found portfolios growing more similar, especially at firms that use AI more intensively. Using a market model, they estimated that a trading signal that once produced excess returns for five to seven years could lose half its excess return in about 18 months as similar algorithms spread across investors. “Each marginal AI entrant shortens the lifespan of every exploitable pattern at an increasing rate,” the authors wrote in their paper “AI-Driven Alpha Decay: Algorithmic Homogenization, Reflexive Signal Erosion, and the Paradox of Intelligent Markets.”

An industry survey showed 58% of fund managers expected to increase AI use, up from 20% two years earlier.

At the University of Liechtenstein, researchers built 10 large-language-model-based trading strategies that used sentiment analysis to forecast stock prices. All models produced positive returns during a 14-month test period ending in April 2025. The team then altered financial news headlines with subtle techniques such as swapping letters for lookalike characters or embedding hidden text. Those changes degraded model performance; in the largest case a model’s overall return fell by about 18 percentage points after manipulation focused on a single stock on a single day. Advije Rizvani cautioned that a single wrong input “could propagate to other days and another decision that the systems are making.”

Elm Partners Management ran a simulated trading challenge in which four popular AI models read advance copies of major newspaper front pages and placed bets on the S&P 500 and Treasury bonds. The models matched the directional accuracy of top macro traders, correctly forecasting market direction more than half the time, but they took on higher than intended risk. Daily return volatility ranged from about 20% to 40%, compared with a recommended range near 7% to 15% for the investor profile used in the test. Victor Haghani observed the models showed tendencies toward overconfidence and oversized positions.

The researchers identified three recurring outcomes across the studies: faster erosion of trading edges as investors converge on the same signals; greater vulnerability to manipulated or noisy information; and systematic risk-taking that can exceed intended limits.

Authors and research teams noted limits to their findings. Much of the evidence comes from simulations, controlled experiments or limited datasets, and crowded trades predate widespread AI use. Giovanni Apruzzese warned that “blindly trusting large language models to make sound decisions would be unwise.”