Maximum likelihood estimation for [formula omitted]-stable double autoregressive models

Abstract

The paper investigates the maximum likelihood estimation (MLE) for a first-order double autoregressive model with standardized non-Gaussian symmetric α-stable innovation (sDAR) within a unified framework of stationary and explosive cases. It is shown that the MLE of all parameters, including the stable exponent in the innovation, are strongly consistent and asymptotically normal (with the exception of the intercept for the explosive case). Particularly, the MLE of the parameter in the conditional location is always asymptotically normal, regardless of the stationary or explosive case. This point totally differs from that for linear α-stable AR models in Andrews et al. (2009). Furthermore, it is the first time to provide exact values of the quantities related to the innovation in the asymptotic covariance matrices when the true innovation is the standard Cauchy distribution. Additionally, a modified Kolmogorov-type test statistic is proposed for model diagnostic checking in the stationary case. Monte Carlo simulation studies are conducted to confirm our theoretical findings and assess the finite-sample performance of the MLE and the modified Kolmogorov-type test. An empirical example is analyzed to illustrate the usefulness of sDAR models.