Forecasting German key macroeconomic variables using large dataset methods

We study the forecasting performance of three alternative large scale approaches using a dataset for Germany that consists of 123 variables in quarterly frequency. These three approaches handle the dimensionality problem evoked by such a large dataset by aggregating information, yet on different levels. We consider different factor models, a large Bayesian vector autoregression and model averaging techniques, where aggregation takes place before, during and after the estimation of the different models, respectively. We find that overall the large Bayesian VAR and the Bayesian factor augmented VAR provide the most precise forecasts for a set of eleven core macroeconomic variables, including GDP growth and CPI inflation, and that the performance of these two models is relatively robust to model misspecification. However, our results also indicate that in many cases the gains in forecasting accuracy relative to a simple univariate autoregression are only moderate and none of the models would have been able to predict the Great Recession.