BVAR models in the context of cointegration: A Monte Carlo experiment

The kind of prior typically employed in Bayesian vector autoregression (BVAR) analysis has aroused widespread suspicion about the ability of these models to capture long-run patterns. This paper specifies a bivariate cointegrated stochastic process and conducts a Monte Carlo experiment to assess the small sample performance of two classical and two Bayesian estimation methods commonly applied to VAR models. In addition, a proposal to introduce a new dimension to the prior information in order to allow for explicit account of long-run restrictions is suggested and evaluated in the light of the experiment. The results of the experiment show that: the Minnesota -type prior with hyperparameter search performs well, suggesting that the prevalent suspicion about the inability of this prior to capture long-run patterns is not well-grounded; the fine-tunning of the prior is crucial; and adding long-run restrictions to the prior does not provide improvements in the case analyzed.