HMM-Regime — Causal 2-State Volatility Regime

2026-05-19 · status: inconclusive · 9.1s

Hypothesis: A 2-state Gaussian HMM on (1h return, 4h trailing RV) provides a causal regime feature (forward filter, not smoothing). P(state=Turbulent) predicts forward 4h RV at IC ≥ 0.05 walk-forward, and state-conditional vol means differ by ≥ 50% (regime separation).

Verdict: PURSUE — Pooled IC +0.3905, separation 1.42×, stability 100%. Includes signal but partially redundant with trailing rv_4h. Worth keeping for Master-LGBM (Exp #22) as a categorical regime feature, not as a standalone replacement for vol.

Key metrics

Approach

1h resampling of 1m close, observation = (log return, log of trailing 4h RV). For each walk-forward split (12mo train / 3mo test, embargo = 1440 min), fit GaussianHMM(n_components=2, covariance_type='full') on the training segment, then run a manual causal forward filter over (train+test) so the state at any test-time t depends only on observations up to t.

State 1 is post-hoc renamed to be the higher-mean-RV one ('Turbulent'). Label is forward 4h realized vol starting at t+1.

Crucial methodological point: hmmlearn's predict_proba is smoothed (uses future observations) — that's the standard source of inflated HMM results in published crypto papers. We use only forward filter (alpha-only).

Pooled OOS IC (P(state=Turbulent) → forward 4h RV)

• Pooled Spearman IC: +0.3905 (n=11,496)

• Bootstrap 95% CI: [+0.3729, +0.4065]

• p-value: 0

Regime separation (state 1 / state 0 mean fwd 4h vol)

• OOS median separation ratio: 1.42×

• Windows with separation ≥ 1.5×: 7/21

Per-window OOS metrics (21 windows)