Identification for univariate time series models with heteroskedastic errors

The SCAN (smallest canonical autocorrelation) method of Tsay and Tiao (17) is modified to become robust to hetero-skedastic data with improved Bartlett's formula for the variance estimation of autocorrelations provided by Berlinet and Francq (3). Data with ARCH(1) variances are typically emphasized in our simulation studies. Our simulation results show that the original SCAN method suffers from severe size distortion for heteroskedastic data, and the distortion is getting larger as sample size is. On the contrary, the modified SCAN method not only has appropriate empirical size but also empirical power for both homoskedastic and heteroskedastic data. Therefore, our modified SCAN method is suggested for identification of univariate time series models with heteroskedastic errors.