Extreme observations and non-normality in ARCH and GARCH

Most studies employing ARCH and GARCH models document the existence of severe excess kurtosis in the estimated residuals. This non-normality may be due to model misspecifications, structural changes, or outliers. We conduct simulation experiments to examine the impact of extreme observations on the estimated parameters and residuals in the ARCH models. Then, we propose an iterative algorithm to detect and correct for the non-normality generated by extreme observations and additive outliers. Results for the simulated data, US equity returns and $/£ exchange rates are presented. Correcting outliers dramatically reduces the non-normality and bias in the estimated coefficients for small samples.