VAR Model Averaging for Multi-Step Forecasting

Given the relatively low computational effort involved, vector autoregressive (VAR)models are frequently used for macroeconomic forecasting purposes. However, the usuallylimited number of observations obliges the researcher to focus on a relatively smallset of key variables, possibly discarding valuable information. This paper proposes aneasy way out of this dilemma: Do not make a choice. A wide range of theoretical andempirical literature has already demonstrated the superiority of combined to single-modelbased forecasts. Thus, the estimation and combination of parsimonious VARs, employingevery reasonably estimable combination of the relevant variables, pose a viable path ofdealing with the degrees of freedom restriction. The results of a broad empirical analysisbased on pseudo out-of-sample forecasts indicate that attributing equal weights systematicallyout-performs single models as well as most more refined weighting schemes interms of forecast accuracy and especially in terms of forecast stability.