On the evaluation of VAR forecasts

In this contribution the use of VAR models for the forecasting of the real GDP in the United States is analysed. A basic model that contains the consumer price index and a short-term money market rate alongside real GDP forms the starting point. The basic model is then expanded by the inclusion of additional macroeconomic variables step by step. The forecasts of the individual models, which are generated as out-of-sample forecasts, are evaluated for different forecasting periods by means of different measures for the testing of the adjustment quality. It is shown that the forecasts of the VAR models turn out to be worse, on average, as those of a univariate AR(2) process. The quality of the forecasts varies strongly in the examined forecasting periods, however, so that the VAR forecasts clearly dominate over the AR(2) forecasts in individual periods. For the preparation of forecasts, it is thus advisable to consider several models simultaneously.