Research
A new theoretical model combining performative prediction, algorithmic herding and cognitive dependency argues that widespread AI adoption paradoxically thins market depth by making supposedly independent trading signals converge, withdrawing liquidity exactly when stress makes it most needed
Researchers Shuchen Meng and Xupeng Chen build a unified framework showing that AI systems trained on similar data and comparable methods develop overlapping common-signal interpretations that collapse diverse strategies into herd-like positioning, while AI predictions reshape the very fundamentals they forecast in a self-reinforcing loop that decouples price discovery from underlying value.
The paper's core move is combining three mechanisms that are usually studied separately into one model: performative prediction, where a forecast changes the reality it is trying to predict; algorithmic herding, where independently built systems end up trading the same way; and cognitive dependency, where human traders lean on AI output rather than forming an independent view. Studied in isolation each is a known concern. Combined, the authors argue they compound into something worse than any one alone, because herding removes the diversity that would otherwise limit a bad prediction's damage, while performativity means the bad prediction can validate itself before anyone notices it was wrong.
The mechanism worth sitting with is what the authors call the common signal component. AI systems trained on similar datasets using comparable methodologies do not stay diverse the way human analysts with different backgrounds and incentives tend to. Their predictive signals partially collapse into shared interpretations, which means what looks like many independent trading decisions is closer to one decision executed many times. That is precisely the condition under which liquidity vanishes fastest in a stress event, since the buyers a seller is counting on are running the same model and reaching the same sell decision at the same moment.
The performative half of the model is the part regulators have less vocabulary for. When an AI prediction moves prices, and the moved price then feeds into whatever fundamentals the next model iteration uses as input, the forecast is no longer a passive read of reality, it is participating in writing the reality it will next be tested against. Prediction accuracy becomes secondary to prediction coordination, in the authors' phrasing, which is a genuinely different failure mode from the mispricing problems traditional market microstructure theory was built to describe.
This connects directly to the reproducibility review and the NBER asset-pricing benchmark this site has already covered, both of which flagged that AI trading research often can't distinguish genuine predictive skill from artefacts of shared training data. Meng and Chen's model gives that empirical concern a theoretical mechanism. Watch for whether regulators start citing common-signal convergence specifically, rather than AI adoption in the abstract, as the basis for concentration limits or diversity requirements on trading models, since a model naming the exact mechanism is a much easier thing to write a rule against than a general worry about AI risk.
Read the original: arXiv - A new theoretical model combining performative prediction, algorithmic herding and cognitive dependency argues that widespread AI adoption paradoxically thins market depth by making supposedly independent trading signals converge, withdrawing liquidity exactly when stress makes it most needed. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.