The Limits of AI Trading

Joint with Itay Goldstein, Zigang Li, Liyan Yang

August 2026

We develop a theoretical framework of trading competition between AI-powered investors and rational investors with heterogeneous sophistication. Rational investors have superior private information but differ in their ability to infer information from prices and anticipate others’ trading behavior, modeled through level-(k) reasoning. AI investors do not observe private signals or reason through belief hierarchies; they learn from realized trading profits through reinforcement learning. Despite heterogeneous algorithms and independent exploration, AI investors endogenously converge to common trading rules, generating algorithmic herding. In the limit, their learned demand coincides with the rational-expectations demand of uninformed investors who correctly extract fundamental information from prices. This convergence does not imply algorithmic dominance. The most sophisticated rational investors can outperform AI investors because AI learns from market data generated by average, not frontier, investor sophistication. AI profitability is limited by rational investors’ private-information advantage, the price-stabilizing trades of the most sophisticated rational investors, and the rising price impact of AI trading as algorithmic herding increases AI market share. These forces identify the limits of algorithmic superiority in financial markets.