AI Volatility Forecasts Under Stress: Performance, Tail Risk, and Ranking Reversals

Abstract

Machine-learning volatility forecasts are usually judged by average out-of-sample accuracy, but their practical value depends on whether relative performance survives stress. This paper studies state-dependent forecast ranking rather than unconditional loss alone. Using the full 2,615-observation archived S&P 500 realized-volatility file preserved from the Oxford-Man Realized Library, I compare EWMA, HAR-RV, and XGBoost in a walk-forward design whose state thresholds use only information available through t−1. XGBoost outperforms HAR-RV in normal observations, but the ranking reverses under high volatility. The paired XGBoost-minus-HAR QLIKE differential is −0.0097 in normal states, +0.679 in Stress90, and +1.167 in Stress95; the Stress90-minus-Normal deterioration is +0.689. The pairwise reversal survives log-MSE and log-MAE, Holm adjustment, chronological subsamples, and a pre-specified top-5% one-day realizedvolatility-ratio state with 148 events. Moving-block bootstrap and leave-one-event-out diagnostics indicate that the jump-state result is not driven by any single observation, although the five largest events account for about 42% of the cumulative differential. By contrast, the primary Q50-to-Q90 transition definition identifies only eight events and remains descriptive under the pre-specified power gate. The evidence reveals a forecasting-risk paradox: a model can look competitive under unconditional average-error criteria while becoming materially worse when volatility is high or rising abruptly. The results motivate evaluating AI volatility forecasts by state-conditional ranking stability, not average accuracy alone.

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