BOCPD — Bayesian Online Change-Point Detection

2026-05-19 · status: promoted · 31.1s

Hypothesis: BOCPD on 15m standardised returns produces P(short-run-length) and E[run-length] features that predict elevated forward 4h realized vol. Spearman IC of P(short-run) vs forward 4h vol ≥ 0.05 walk-forward.

Verdict: PROMOTE — p_short IC = +0.1601, positive in 21/21 windows. Bars in fresh-change-point regime show +26.7% higher forward 4h vol than baseline. Add as feature for vol-forecast / Master-LGBM.

Key metrics

Approach

BOCPD with Normal-Inverse-Gamma prior (conjugate; Student-T predictive) on 15m standardised log returns z = r / rolling_std(r, 96 bars = 24h). Constant hazard H = 1/96 (≈ 1 expected change-point per day). Run-length truncated at R_max = 200 (≈ 50h memory).

Features: - p_short = P(run_length < 4) — fresh-changepoint indicator - e_runlen = E[run_length] — system stability measure

Label: realized vol of next 4h. Sampled every 16 bars (4h) for IC independence. Walk-forward: 12mo train / 3mo test, embargo = 1440 min.

Pooled OOS IC

• p_short → fwd_4h_vol: +0.1601 (n=13,953, 95% CI [+0.1434, +0.1752], p=8.41e-81)

• e_runlen → fwd_4h_vol: +0.0448 (n=13,953, 95% CI [+0.0286, +0.0611], p=1.23e-07)

Change-point regime conditional

• Bars with p_short > 0.5: 3,212 (1.44% of valid samples)

• Mean fwd 4h vol when p_short > 0.5: 0.00339

• Baseline mean fwd 4h vol: 0.00267

• Uplift: +26.7%

Per-window IC — p_short (21 windows)

Per-window IC — e_runlen (21 windows)