Outliers and conditional autoregressive heteroscedasticity in time series

This paper reviews the literature on GARCH-type models proposed to represent the dynamic evolution of conditional variances. Effects of level outliers on the diagnostic and estimation of GARCH models are also studied. Both outliers and conditional heteroscedasticity can generate time series with excess kurtosis and autocorrelated squared observations. Consequently, both phenomena can be confused. However, since outliers are generated by unexpected events and the conditional variances are predictable, it is important to identify which one is producing the observed features in the data. We compare two alternative procedures for dealing with the simultaneous presence of outliers and conditional heteroscedasticity in time series. The first one is to clean the series of outliers before fitting a GARCH model. The second is to estimate first the GARCH model and then to clean of outliers by using the residuals adjusted by its conditional variance. It is shown that both approaches may result in different estimated conditional variances.