Correlated Beliefs, Correlated Trades: Agentic AI and Endogenous Volatility Amplification

Abstract

Agentic trading systems are usually evaluated by the accuracy of individual forecasts. That evaluation misses a distinct systemic-risk channel: several agents can be individually well calibrated while forming correlated posterior beliefs and submitting synchronized orders. We develop a controlled market experiment that holds marginal predictive accuracy approximately constant and varies information overlap, common posterior calibration error, and the homogeneity of risk controls. Agents report auditable posterior probabilities over calm, turbulent, and crisis states. Their orders clear in a market where one-sided flow consumes depth and affects subsequent prices and observations. The identifying comparison is therefore not AI versus non-AI. It is low versus high cross-agent belief dependence conditional on comparable individual forecast quality. The mechanism predicts a temporal sequence from posterior dependence to order synchronization, depth depletion, price impact, and tail volatility. A randomized independent-execution placebo preserves correlated beliefs while breaking the belief-to-trade timing channel. In the pilot mechanism experiment, accuracy-matched high-dependence agents produce sharply higher order synchronization, realized variance, tail returns, and a spread/depth proxy, whereas the placebo returns close to the independent benchmark. These results are mechanism validation, not evidence about any particular language model. The evidence is deliberately architectural: it establishes the mechanism in a controlled synthetic market and does not estimate the prevalence or aggregate effect of deployed commercial systems.

Author abstract from the research manuscript. Typography and line breaks have been normalized; the linked source may show a different manuscript version.