Representation Homogeneity and Systemic Instability in AI-Dominated Financial Markets: A Structural Approach

A new structural study by Yimeng Qiu and Qiwei Han, published on arXiv (2604.22818), reveals a critical vulnerability in AI-dominated financial markets: when AI trading agents process market data in similar ways, they can collectively destabilize the system. The researchers built a multi-agent model calibrated with high-frequency microstructural data, where each AI uses a two-layer architecture: a nonlinear 'representation' layer that encodes raw market states into high-dimensional feature vectors, and a linear 'readout' layer that generates return forecasts for trading. Crucially, they distinguish between 'representation homogeneity' (how similarly agents encode information) and 'forecast overlap' (how similar their predictions are), showing these are related but not equivalent.

Through controlled factorial experiments varying representation homogeneity, risk aversion, and learning rates, the paper demonstrates that increased representation similarity amplifies synchronization in beliefs and positions among agents. This leads to volatility clustering, liquidity stress, and elevated tail risk. The mechanism is insidious: during low-volatility periods, hidden leverage accumulates through 'position stickiness,' which then collapses catastrophically when a shock triggers synchronized deleveraging. The findings provide a structural foundation for macroprudential policies that monitor and preserve diversity in how AI systems represent market information, suggesting that regulators may need to enforce algorithmic diversity to prevent AI-driven flash crashes.