Accurate but Correlated: Common AI Forecast Errors and Endogenous Market Volatility

Abstract

Artificial intelligence can improve each institution's risk forecast while reducing the financial system's ability to diversify forecast errors. This paper separates these two objects. Individual accuracy is a marginal property of an error distribution; systemic model risk is a covariance property. With N institutions, equicorrelated log-risk forecast errors leave an aggregate variance floor of ρσ² that does not vanish as N grows. We derive the exact threshold above which an AI technology with lower individual mean-squared error nevertheless creates greater aggregate forecast risk than noisier independent legacy models. Risk-sensitive position rules then map common errors into synchronized net demand. With linear price impact, forecast-error covariance enters return variance; with GARCH updating, the initial variance increment is propagated by the multiplier (1-β)/(1-α-β). Prospectively separated simulations validate the signs under smooth volatility targeting and a discrete risk limit. Holding marginal error variance fixed, raising cross-institution error correlation from zero to 0.8 increases return variance by 3.15% under the smooth rule and 6.26% under the discrete risk limit. The stronger precommitted claim that both variance and one-percent expected shortfall rise by at least 5% is rejected. Common errors robustly amplify two-sided volatility under the model's conditions, but symmetric errors do not necessarily create comparably large downside-tail losses. Model diversity should therefore be evaluated through error covariance rather than model counts, vendor counts, or average predictive accuracy.

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