Carrying regime uncertainty forward in cryptocurrency tail-risk forecasting

Cryptocurrency tail-risk forecasting faces a structural timing problem: contemporaneous cross-asset correlation defines contagion but is unavailable until each hour’s close, forcing pre-close regime classifiers to rely on delayed signals. Lagged volatility proxies therefore breach their thresholds only after correlation shocks peak, so MS-HAR recalls only 14.4% of contagion hours and worsens aggregate QLIKE relative to HAR-RV. We propose RAFT (Regime-Adaptive Forecasting with Temporal filtering), which carries HMM posterior uncertainty forward from realized variance without requiring contemporaneous correlation. Using minute-level data for five USDT perpetual futures on Binance (2020–2025), RAFT matches HAR-RV in aggregate QLIKE while improving contagion-hour QLIKE by 19.2% relative to MS-HAR and reducing contagion-hour 99%-VaR violations to 16.1%, compared with 28.3% for HAR-RV and 38.9% for MS-HAR. The results suggest that carrying regime uncertainty forward can improve real-time tail-risk forecasts when regime signals arrive with delay.