Aggregate Density Forecasting from Disaggregate Components Using Large VARs

When it comes to point forecasting there is a considerable amount of literature that deals with ways of using disaggregate information to improve aggregate accuracy. This includes examining whether producing aggregate forecasts as the sum of the component’s forecasts is better than alternative direct methods. On the contrary, the scope for producing density forecasts based on disaggregate components remains relatively unexplored. This research extends the bottom-up approach to density forecasting by using the methodology of large Bayesian VARs to estimate the multivariate process and produce the aggregate forecasts. Different specifications including both fixed and time-varying parameter VARs and allowing for stochastic volatility are considered. The empirical application with GDP and CPI data for Germany, France and UK shows that VARs can produce well calibrated aggregate forecasts that are similar or more accurate than the aggregate univariate benchmarks