GARCH(1,1) says variance(t) = w + a x shock(t-1)^2 + b x variance(t-1). A typical equity fit has a near 0.08 and b near 0.90. Because a + b is close to 1, shocks decay slowly and the model reproduces the observed persistence of volatility.
The long-run variance implied is w/(1 - a - b), and forecasts revert to it geometrically. With a + b = 0.98, a doubling of volatility takes roughly 34 days to decay halfway back, which matches the lived experience of a post-crash market.
Honest limits: GARCH forecasts volatility, not direction, so it improves sizing rather than selection. It is slow to react to genuine regime breaks, it assumes symmetry unless you use a variant like GJR-GARCH that treats down moves as more volatility-inducing, and its parameters are unstable when refit often.
Related: ewma, volatility-clustering, arima, volatility-targeting