Selecting an innovation distribution for Garch models to improve efficiency of risk and volatility estimation

J. H.Venter and P. J. de Jongh

Abstract

ABSTRACT

It has become common practice to fit Garch models to financial time series by means of pseudo-maximum likelihood. In this study we investigate the behavior of several maximum likelihood-based methods for estimating the Garch model parameters and for estimating volatility and risk measures (VAR and expected shortfall). We consider normal inverse Gaussian (NIG), skewed-T, T and non-parametric kernel densities for this purpose and compare the efficiencies of the resulting estimates with those based on the normal distribution. The NIG-based approach is found to be competitive with the other methods in most of the cases considered.