Does Joint Modelling of the World Economy Pay Off? Evaluating Multivariate Forecasts from a Bayesian GVAR

To assess the performance of multivariate density forecasts for the world economy based on a Bayesian global vector autoregressive (GVAR) model, we decompose the predictive joint density into its marginals and a copula term that captures the dependence structure among variables and countries. Moreover, we use the stochastic search variable selection prior (SSVS) on the coefficients in its conjugate form to account for model uncertainty at the national level and augment the GVAR framework to allow for stochastic volatility. Our results are as follows: First, the GVAR systematically outperforms forecasts based on country-specific models in terms of predictive joint density. Second, the good GVAR performance is driven by superior predictions for the dependence structure across variables, whereas the GVAR model does not yield better predictive marginal densities. Third, the relative performance gains of the GVAR model are particularly pronounced during the Great Recession. Finally, our results imply that for some countries a more parsimonious GVAR model can further improve the forecast quality.