Volatility Regime Detection with ML

One model for all regimes overfits the dominant one. Better: cluster days into calm/trending/stressed via vol features, then train per-regime classifiers. Feature set feeds a regime classifier (vix_z, iv_rv_spread, atm_iv change); once regime is known, the matching sub-model scores direction. Out-of-sample this beats a single global model because it respects that calm and stressed markets behave differently. Regime detection is the unsung hero of robust retail ML -- most '90% accurate' models silently only work in one regime.