Statistics Glossary → Conditional variance (volatility) and GARCH orders
Conditional variance (volatility) and GARCH ordersconcept
The way a series' uncertainty for the next period changes over time with past shocks and past volatility; the quantity that GARCH models describe.
In financial returns, calm periods tend to follow calm ones and turbulent periods turbulent ones (volatility clustering). In a GARCH(p, q) model the conditional variance at time t is σ²ₜ = ω + Σ αᵢ ε²ₜ₋ᵢ + Σ βⱼ σ²ₜ₋ⱼ: q lagged squared shocks (the ARCH terms, α) and p lagged conditional variances (the GARCH terms, β). In practice GARCH(1,1) is often enough. The sum α + β shows how persistent volatility is; if it is below 1, the variance returns in the long run to ω / (1 − α − β).
Standard GARCH reacts in the same way to positive and negative shocks. In stocks, bad news tends to raise volatility more (the leverage effect); GJR-GARCH models this asymmetry by adding an extra coefficient for negative shocks, and EGARCH by separating sign and size effects in an equation for the log variance. The unconditional variance can stay constant while the conditional variance changes from period to period; the two should not be confused. The choice of error distribution (normal, t) matters especially for tail risk.
An everyday example
Daily returns of a stock index are examined. In a calm week daily moves stay around 0.5%, while after major crisis news there are swings of 3-4% for days. A GARCH(1,1) model finds α = 0.08 and β = 0.90; since α + β = 0.98, the effect of a shock on volatility fades only slowly, over weeks.
In MerQur
The 'Variance model (GARCH / EGARCH / GJR-GARCH)', 'ARCH order q (α) and GARCH order p (β)', 'Volatility forecast horizon' and 'Variance initialization (Sample variance / Backcast)' options in the GARCH analysis.