On asymptotic inference in AR and cointegrated models with unit roots and heavy tailed errors (Q4355989)

Let \(X_t=\beta_1X_{t-1}+\cdots+\beta_qX_{t-q} +\epsilon_t\) be an AR\((q)\) process such that the roots of the characteristic polynomial \(1-\beta_1z-\cdots-\beta_qz^q\) are on, or suitably contiguous to, the unit circle. It is assumed that \(\epsilon_t\) are iid random variables with a symmetric density, regularly varying tail probabilities, and a positive and finite Fisher information. The author shows that there exists a locally asymptotically mixed normal approximation to the likelihood of \(X_1,\dots,X_n\). Since the density of \(\epsilon_t\) is unavailable in general, it is indicated that its replacement by a suitable estimate based on the entire data does not affect the limiting behaviour of the procedures. The results are applied to a cointegrated model such that the observations \((Y_i,X_i)\) satisfy \(Y_i=\gamma' X_i+\eta_i\) \((i=1,\dots,n)\) where \(X_i\) is a \(q\)-dimensional AR process with unit roots, \(\gamma\) is an unknown parameter and \((\eta_i,\epsilon_i)\) are iid random vectors. For the entire collection see [Zbl 0861.00032].