On the determination of lag order in vector autoregressions of cointegrated systems (Q1965982)

Dynamic modelling which uses maximum likelihood cointegration methods requires the specification of a lag order of the vector autoregressive model we deal with. In the literature this is carried out by using univariate Box-Pierce statistics. In this paper the multivariate nature of the problem is recognized and an attempt is made to evaluate by means of simulations the functioning of multivariate order determination criteria and portmanteau statistics in a cointegrated setting. More precisely, the behavior of information criteria, AIC, the log criterion, BIC, the iterated log criterion, LIL, and the multivariate portmanteau statistic in a simple bivariate cointegrated system using Monte Carlo simulations is investigated. The outcome of the study indicates that some of these statistics behave favorably and could be useful as an alternative or a complement to univariate testing.