AI Trading Convergence Risk Emerges as Wall Street Warns of Crowding Effects
N.R. Finch
Multiple studies find that widespread adoption of similar AI models is pushing institutional portfolios toward convergence — the half-life of a profitable trading signal has shrunk from five-to-seven years to roughly 18 months, meaning AI-driven edge is eroding itself.
What happens when everyone picks stocks with AI?
NYU researchers analyzed nearly one million institutional holding records and found that the more heavily a firm uses AI, the more its portfolio resembles its peers'.
This means → similar models reading similar data end up betting on the same stocks — "edge" becomes "consensus."
Their paper calls this "AI-driven alpha decay": a profitable signal's half-life has compressed from five-to-seven years to roughly 18 months.
Why does more AI kill profitable signals faster?
The core logic: every new AI participant accelerates the expiry of every exploitable pattern.
In plain terms = it is like a shortcut only a few drivers know — once every navigation app recommends it, the shortcut becomes a traffic jam.
The paper stresses that collective outcomes differ in kind from the simple sum of individual gains — advantages that work at the individual level create systemic risk when aggregated.
Can AI trading systems be tricked?
A University of Liechtenstein team built 10 LLM-based trading systems, all of which posted positive returns over a 14-month period.
Researchers then made subtle manipulations to financial news headlines — swapping letters for visually similar characters, embedding hidden text — and every model was successfully deceived.
In the most extreme case, manipulating coverage of a single stock on a single day cut the models' overall returns by roughly 18 percentage points.
This reflects a fragility at AI's information-input layer: one bad decision can cascade through multiple subsequent trading days.
How reliable are these studies — and what are the limits?
The researchers themselves acknowledge that most findings rest on simulations, controlled scenarios, or limited datasets — none proves AI necessarily makes markets more fragile.
Crowded tech trades predated AI and have yet to trigger a systemic crisis.
Still, the body of work points to a shared conclusion: the same technology that boosts information-processing can also make crowded trades, misinformation, and overconfidence spread more easily through markets.
What does this mean for investors?
For active managers, today's winning strategy may turn into tomorrow's crowded trade faster than ever before.
This means → "alpha found by AI" has a shelf life — and that shelf life is shrinking. Sustained edge requires not just better models but differentiated data and strategies.
AI does not deliver a one-way "use it and win" advantage; it ushers in a new landscape where an arms race and convergence risk coexist.
Content is for reference only, not financial advice.