Machine Learning Early Warning of Extreme Financial Volatility under Distribution Shift: Guaranteed Recall, Balanced Evaluation and a Test of Contagion

Abstract

Early-warning models for financial stress are usually judged by accuracy or a threshold tuned on past data, and neither guarantees how many episodes an alarm will catch. We build and test an early-warning system for extreme volatility in 34 markets across equity indices, currencies, commodities and cryptocurrencies (2000–2026), with a confirmation tier of 162 S&P 500 stocks. The target is whether an asset's realized volatility over the next five trading days exceeds its own causal top-5% level. Seventeen models were compared under purged walk-forward validation with an untouched 2025–2026 hold-out and a future-perturbation leakage test. They included HAR-RV and GARCH(1,1) baselines, machine-learning learners, a volatility-anchored focal gradient-boosting model (VA-Boost), a regime-gated mixture of experts with a contagion-aware expert, and ensembles. Probabilities were calibrated with isotonic regression, and alarms were set by adaptive conformal risk control, which targets a fixed recall under distribution shift. Because a rule that never alerts is 93–94% accurate, models were compared by ROC-AUC, balanced accuracy and the Matthews correlation coefficient. Learned models ranked risk well (ROC-AUC 0.85 in cross-validation and 0.83–0.84 on the hold-out, against 0.78–0.81 and 0.77 for HAR-RV and GARCH), with balanced accuracy of 77% and 76% and Brier skill of 20–23%. VA-Boost matched the best boosters without beating them, and the gains of the gated mixture did not replicate on a second machine. Contagion features did not improve warnings on average, under nested selection, in stress regimes or on single stocks. The adaptive conformal alarm held recall at 0.80 in every validation fold (standard deviation 0.001–0.004), on the hold-out and in all seven crisis windows, whereas F1-tuned thresholds caught between 10% and 64% of events depending on the period. The guarantee required alerts on about 40% of days; VA-Boost made it cheapest, with 12–13% precision against a 6–7% base rate. Volatility warnings reduced drawdowns but not Sharpe ratios. A guaranteed-recall alarm, rather than a more complex model, is therefore the practical gain of machine learning for financial early warning.