The Czech Ministry of Finance applies a complex DSGE-VAR model for forecasting purposes. This study concentrates on a presentation of the VAR component which is estimated separately for every macroeconomic variable subject to prediction.
DSGE models are well-known for their compelling long-run predictions while VAR models are prevalent tools for economists focusing on short-run and medium-run forecasts. Final prediction incorporates different versions of VAR models along with an outcome from DSGE-BAYES forecasts with respect to their forecasting accuracy. The final model combines advantages of Bayesian estimation with benefits of regression analysis.
Forecasting performance for VAR models is evaluated based on an iterated rolling-window forecast always estimated one period ahead for forty periods in quarterly frequency from the first quarter of 2012 to the second quarter of 2019. Forecast for the DSGE part of the model is provided by Dynare, the Bayesian method of estimation is designed to secure perfect fit of observable variables on data, this translates into alignment of the modelled series and the historical series.
Final out-of-sample forecast is an outcome of the DSGE-VAR estimation. Selection of suitable VAR models for prediction depends on economic theory, statistical significance of individual variables in the model as well as on a size of root mean squared errors (RMSE). While increasing number of variables is common way to boost in-sample forecast fit, it negatively affects the out-of-sample forecast.