Skip to content
TheChatArchiveThe Chat Archive
GPT 5.1 Codex Max· openaiHugging Face Frontier Reasoning Traces

Contrast GARCH with EWMA RiskMetrics. EWMA is a special GARCH case. Why is GARCH's mean…

  1. User

    Contrast GARCH with EWMA RiskMetrics. EWMA is a special GARCH case. Why is GARCH's mean reversion to long-run variance a key advantage?

  2. Assistant

    Explaining GARCH mean reversion advantage

    GARCH generalizes EWMA RiskMetrics by including a constant term that ensures mean reversion to a finite long-run variance, yielding more stable and realistic volatility forecasts. This mean reversion prevents unconditional variance from being infinite or undefined, allowing risk estimates to revert to a typical level after shocks, unlike EWMA which has infinite memory and no mean reversion. What RiskMetrics calls EWMA is nothing more than a very tightly parameterised (1,1) GARCH. In RiskMetrics you update the one–step variance as [ \sigma^2_t=\lambda,\sigma^2_{t-1}+(1-\lambda)r_{t-1}^2, ] with (\lambda) fixed at 0.94 or 0.97. That is a GARCH(1,1) with (\omega=0), (\alpha=1-\lambda) and (\beta=\lambda); by construction (\alpha+\beta=1), so you have an integrated GARCH and no unconditional variance.

    In a standard GARCH(1,1) [ \sigma^2_t=\omega+\alpha,r_{t-1}^2+\beta,\sigma^2_{t-1} ] the intercept (\omega) and the restriction (\alpha+\beta<1) give you an unconditional variance (\omega/(1-\alpha-\beta)) and an exponential decay of shocks back to that long‑run level. Volatility is allowed to jump when returns are large, but the effect of any single shock dies out and forecasts at longer horizons converge to the average level.

    That mean‑reversion is the key advantage. It reflects the empirical tendency of market volatilities to revert to “normal” levels, prevents the conditional variance from drifting down to zero in quiet periods or staying elevated indefinitely after a spike, and gives you a well‑defined long‑run variance for multi‑period risk forecasts.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.