DETECTING LEVEL SHIFTS IN THE PRESENCE OF CONDITIONAL HETEROSCEDASTICITY

The objective of this paper is to analyze the finite sample performance of two variants of the likelihood ratio test for detecting a level shift in uncorrelated conditionally heteroscedastic time series. We show that the behavior of the likelihood ratio test is not appropriate in this context whereas if the test statistic is appropriately standardized, it works better. We also compare two alternative procedures for testing for several level shifts. The results are illustrated by analyzing daily returns of exchange rates.