Do Bayesian Vector Autoregressive models improve density forecasting accuracy? The case of the Czech Republic and Romania

Abstract

The supremacy of Bayesian VAR models over the classical ones in terms of forecasting accuracy is well documented and generally accepted in the literature on the grounds of overcoming the short sample and overfitting problems. However, the record is rather limited in case of emerging economies, ¨¨¨and more so for density (as opposed to point) forecasting. In this paper we compare the predictive accuracy of Bayesian and classical VAR models in case of density forecasting Czech and Romanian economic variables. The results show predictive densities are generally well calibrated, especially at shorter forecast horizons (less so for Romania). Log predictive density scores confirm the hypothesis of more accurate predictions of Bayesian VAR over classical VAR and naOve univariate models, the dominance of the former being more obvious in case of Romania. As such, the Bayesian approach to VAR yields a better approximation of the uncertainty surrounding the unknown future and minimizes the prediction errors.