Vector autoregressions (VAR's) are an important tool in time series analysis. However, relatively little is known about the finite-sample behaviour of parameter estimators. We address this issue here, by investigating maximum likelihood estimators (MLE's) in the context of a purely nonstationary...