QuarcHome
Model catalogueBack to the model index

GARCH–EVT tail estimation (McNeil–Frey)

Fits an extreme-value tail to filtered innovations for conditional tail-risk estimation.

How it works

The approach filters returns with a conditional volatility model and fits an extreme-value tail to the standardised residuals beyond a threshold. A generalised Pareto distribution models tail exceedance size. Rescaling the estimated tail by the current volatility produces conditional VaR or ES.

Worked example

Suppose the fitted standardised loss quantile is 3.2 and tomorrow's predicted volatility is 1.25%, with negligible conditional mean. The loss quantile is about 4.0%. If the same standardised tail estimate is used with volatility 2%, the implied loss quantile rises to 6.4%.

Validation and limitations

Choose thresholds using stability diagnostics. Too low a threshold can bias the tail fit and too high a threshold leaves few observations. Check residual dependence and uncertainty in the tail shape. A fitted Pareto tail with shape at least one has infinite mean excess, which challenges a finite ES calculation.

Research sources