Forecasting Czech GDP using Bayesian dynamic model averaging

Forecasting future path of macroeconomic aggregates has become crucial for monetary and fiscal policymakers. Using Czech data, the aim of this paper is to demonstrate the benefits of the Bayesian dynamic averaging and Bayesian Vector Autoregressive Models (BVAR) in forecasting real GDP growth. Estimation of richly parameterized VARs often leads to unstable estimates and inaccurate forecasts in models with many variables. Bayesian inference and proper choice of informative priors offers an effective solution to this problem by shrinking the variance of model parameters. Bayesian dynamic model averaging (DMA) then makes it possible to account for model uncertainty by combining predictive abilities of many competing VAR models considered by a researcher. Since forecasting performance of individual models may vary over time, the DMA can adapt their weights in dynamic and optimal way. It is shown that the application of DMA leads to substantial forecasting gains in forecasting Czech real GDP.