Robust estimation for vector autoregressive models

Abstract

A new class of robust estimators for VAR models is introduced. These estimators are an extension to the multivariate case of the MM-estimators based on a bounded innovation propagation AR model. They have a filtering mechanism that avoids the propagation of the effect of one outlier to the residuals of the subsequent periods. Besides, they are consistent and have the same asymptotic normal distribution as regular MM-estimators for VAR models. A Monte Carlo study shows that these estimators compare favorable with respect to other robust ones.