"Outlier Detection in GARCH Models"  


Jurgen A. Doornik and Marius Ooms

	

Abstract:

We present a new procedure for detecting multiple additive outliers in 
GARCH(1,1) models at unknown dates. The outlier candidates are the 
observations with the largest standardized residual. First, a 
likelihood-ratio based test determines the presence and timing of an 
outlier. Next, a second test determines the type of additive outlier 
(volatility or level). The tests are shown to be similar with respect to 
the GARCH parameters. Their null distribution can be easily approximated 
from an extreme value distribution, so that computation of p-values does 
not require simulation.

The procedure outperforms alternative methods, especially when it comes 
to determining the date of the outlier. We apply the method to returns 
of the Dow Jones index, using monthly, weekly, and daily data. The 
procedure is extended and applied to GARCH models with Student-t 
distributed errors.