Vol forecasting: persistence vs HAR-RV vs GBM

2026-05-19 · status: promoted · 20.3s

Hypothesis: A richer model (HAR-RV or gradient boosting) improves on naive persistence (forecast = trailing rv) as a forward-vol forecaster, walk-forward at 1h / 4h / 1d horizons.

Verdict: SHIP — at the 4h horizon, GBM beats persistence: IC +0.836 vs +0.745 (R² +0.700 vs +0.546). Wire ml.forecast.predict_vol_4h into strategies/ for position sizing and entry filtering.

Key metrics

Approach

We forecast the log of forward realised volatility at three horizons (1h, 4h, 1d). Three models compete on the same hourly panel:

1. Persistence — forecast = trailing log(rv_h). The Exp 1 baseline.
2. HAR-RV (Corsi 2009) — linear regression of log(forward rv) on log (trailing rv_1h, rv_4h, rv_1d). The classic econometric vol forecaster.
3. GBM — gradient-boosted regression on all features (rv at 4 windows, ret_24h, ret_4h, range_4h, range_1d, log_vol_z_1d, hour, dow).

Walk-forward: 12 months train, 3 months test, 1-day embargo. Out-of-sample predictions are concatenated across all windows then scored against actual forward vol.

OOS comparison (concatenated across all walk-forward windows)

Horizon 1h

Horizon 4h

Horizon 1d