GARCH
Makes conditional variance depend on past innovations and past conditional variances.
How it works
GARCH(1,1) updates conditional return variance from a constant, the most recent squared shock and the previous conditional variance. The shock coefficient measures immediate responsiveness and the variance coefficient governs persistence. For nonnegative parameters with a sum below one, the usual stationary long-run variance is the constant divided by one minus the sum.
Worked example
Set the daily variance constant to 0.000002, the shock coefficient to 0.08 and the lag variance coefficient to 0.90. Starting from variance 0.0001, a 3% shock gives a next-day variance of 0.000164, or about 1.28% volatility. The stationary variance for these coefficients is 0.0001.
Validation and limitations
Check that the variance stays nonnegative and inspect standardised residuals for remaining dependence and tail misfit. Iterate forecasts over the requested horizon. A normal innovation can understate tail risk even if the variance recursion is well fitted.