ARCH
Models conditional variance as a function of past squared innovations.
How it works
ARCH explains volatility clustering by letting today’s conditional variance depend on past squared forecast errors. Large surprises of either sign raise the next variance under the basic symmetric specification. With several lags, each lag can have a different influence. A mean model is estimated or specified before interpreting the residuals.
Worked example
Use ARCH(1) with a constant variance of 0.00004 and a coefficient of 0.3. After a 2% return shock, the next-day variance forecast is 0.00004 + 0.3 × 0.0004 = 0.00016, giving volatility of about 1.265%. After a zero residual, the forecast instead returns to the constant.
Validation and limitations
Check squared standardised residuals for remaining autocorrelation and compare against a constant-variance forecast. Many lags can be needed when volatility decays slowly. GARCH adds lagged conditional variance to represent persistence parsimoniously.