garch-volatility-toolkit

GARCH Volatility Toolkit

Self-Evolving skill — if a recipe drifts from what actually reproduces (model spec, cost regime, campaign Sharpe numbers, library API), fix this SKILL.md in the same change; see the Post-Execution Reflection at the bottom.

Walk-forward GARCH(1,1) and GJR(1,1) recipes for volatility forecasting and portfolio construction. Tested on 20 seeds, 2025-26 test window, 2bps and 7bps cost regimes.

Campaign Results (honest): A negative-to-marginal result set, not a deployable edge. GJR inverse vol-sizing helps at LOW cost only (ensemble Sharpe +0.49→+1.00 @2bps) but is COST-FRAGILE — its +35% turnover erases the benefit at retail 7bps (Δ+0.02, p=0.54, coin-flip). DCC de-weighting is economically immaterial (+0.05–0.07). GARCH-as-features is a flat null. Per-seed p<0.0001 figures are inflated by applying a deterministic overlay to correlated seeds — use ensemble PSR instead. See CAMPAIGN_VERDICT.md. This skill's value is the METHODOLOGY (fitting recipes + leakage traps), not an alpha claim.

Quick Start

1. Univariate GARCH(1,1) Forecast Volatility

import pandas as pd
import numpy as np

def garch_forecast(returns_series, window=252):
    """
    Walk-forward GARCH(1,1) fitting and forecast.

    h_t = ω + α·r²_{t-1} + β·h_{t-1}

    Parameters:
    - window: Rolling estimation window (default 252 bars ≈ 1 year)
    - Returns: pd.Series of forecast volatilities (sqrt(h_t))
    """
    ret = returns_series.values if isinstance(returns_series, pd.Series) else returns_series
    T = len(ret)
    h = np.full(T, np.var(ret), dtype=np.float64)
    fc = np.full(T, np.nan, dtype=np.float64)

    for t in range(1, T):
        r = ret[max(0, t-window):t]
        if len(r) > 10:
            om = max(np.var(r) * 0.01, 1e-6)
            al = min(0.1, max(0.01, 0.1 * np.mean(r**2) / (np.var(r) + 1e-10)))
            be = 0.8
        else:
            om, al, be = np.var(ret) * 0.01, 0.05, 0.85

        h[t] = max(om + al * ret[t-1]**2 + be * h[t-1], 1e-8)
        fc[t] = np.sqrt(h[t])

    return pd.Series(fc, index=returns_series.index if isinstance(returns_series, pd.Series) else None)