GARCH Volatility Forecasting

Volatility is not constant and it is not random — turbulent days cluster together. GARCH is the standard way of turning that regularity into a forecast.

The observation GARCH is built on

Volatility clusters. Large moves are followed by large moves, quiet periods by quiet periods. This is one of the most robust empirical regularities in financial data, visible across essentially every asset class and era.

A simple realized volatility calculation ignores this entirely — it treats every day in its window as equally informative about tomorrow. GARCH takes the clustering seriously and builds it into the forecast.

What the model does

GARCH — Generalised Autoregressive Conditional Heteroskedasticity — forecasts tomorrow's variance as a weighted combination of three things: a long-run average variance, the most recent squared return, and yesterday's variance forecast.

In the standard GARCH(1,1) form, those weights are three parameters. The coefficient on the recent squared return governs how sharply the model reacts to a shock; the coefficient on the previous forecast governs how slowly it decays back toward normal. Their sum, typically close to but below one, determines volatility persistence.

What it gives you over simple realized volatility

Faster reaction. A rolling 20-day standard deviation gives a large shock the same weight as a move 19 days ago, and drops it abruptly when it exits the window. GARCH weights recent information more heavily and decays it smoothly.

Mean reversion. The model pulls forecasts back toward a long-run average, matching the empirical fact that volatility spikes subside.

A term structure. Because it is recursive, GARCH produces forecasts for multiple horizons ahead, not just a single number.

Where it falls short

It is backward-looking about catalysts. GARCH knows nothing about an earnings date or a policy meeting on the calendar. It only learns after the event moves prices.

Standard GARCH is symmetric. It treats a large gain and a large loss as equally informative about future volatility, when losses empirically raise volatility more. Asymmetric variants exist specifically to fix this.

Estimation risk. Parameters must be fitted, and the fit is sensitive to the sample. A model calibrated on a calm period can behave badly through a regime change.

Complexity is not free. Research comparing volatility forecasts has repeatedly found that well-specified simple models are competitive with elaborate ones, particularly over longer horizons.

The practical takeaway

GARCH is most valuable at short horizons, where volatility clustering carries the most information. Over longer horizons, forecasts converge toward the long-run average anyway, and simpler blends of realized volatility estimates capture most of the benefit at a fraction of the complexity and with far fewer ways to go quietly wrong.