ARCH/GARCH Volatility Modelling

Description

Test whether a metric has calm and turbulent spells, and model and forecast its changing variance rather than assuming constant noise.

A series can show no correlation in its values yet strong dependence in its variability: large changes follow large changes and quiet periods follow quiet ones. ARCH models make the current variance depend on recent squared shocks; GARCH adds past variances, so a GARCH(1,1) usually suffices. Variants handle persistent shocks (IGARCH, the basis of exponentially weighted volatility), asymmetry between rises and falls (EGARCH) and explanatory variables. The workflow tests residuals from a simple mean model for ARCH effects, fits a low-order GARCH by conditional maximum likelihood, often with Student-t errors, checks standardised residuals, and forecasts volatility, which reverts towards its long-run level. For UX metrics the value is in describing and anticipating periods of instability, such as latency or error rates whose variability clusters, so alert thresholds and confidence bands widen and narrow appropriately.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Effort
Days

Steps

  1. Fit a simple mean model (constant or low-order autoregression) to the series and keep the residuals.
  2. Test for ARCH effects with Ljung-Box on squared residuals or Engle's Lagrange multiplier test (regress squared residuals on m lags; chi-squared on m degrees of freedom).
  3. If present, fit a low-order GARCH such as (1,1) by conditional maximum likelihood, with normal or Student-t innovations (3 to 6 degrees of freedom are common).
  4. Check standardised residuals and their squares for remaining autocorrelation, and inspect skewness, kurtosis and a QQ plot.
  5. Consider asymmetric or persistent variants if bad shocks raise volatility more than good ones or if shocks barely decay.
  6. Forecast conditional variance recursively and use it to set time-varying control limits or prediction bands.

Sources

  • Analysis of Financial Time Series (Tsay, 2002)

Also known as

  • ARCH/GARCH models
  • GARCH
  • Volatility modelling
  • Conditional variance modelling
  • EGARCH
  • IGARCH
  • Conditional heteroscedasticity (volatility clustering) testing and GARCH modelling

Categories

  • Analytics & Behaviour
  • Statistics & Modelling