Density Forecasting using Bayesian Global Vector Autoregressions with Stochastic Volatility

This paper develops a Bayesian global vector autoregressive model with stochastic volatility. Three variants of stochastic volatility are implemented to improve the existing homoscedastic framework. In our baseline model, we assume that the variance-covariance matrix is driven by a set of idiosyncratic, country-specific and regional factors. By contrast, the second specification adopted implies that the error variance of each equation is determined by an independent stochastic process. The final specification assumes that the country-specific volatility follows a single factor, which leads to significant computational gains. Considering a range of competing models, we forecast a large panel of macroeconomic variables and find that stochastic volatility influences predictive accuracy along three dimensions. First, it helps to improve the overall predictive fit of our model. Second, it helps to make the model more resilient with respect to outliers and economic crises. Finally, taking a regional stance reveals that forecasts in developing economies tend to profit more from stochastic volatility.