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 usually limited number of observations obliges the researcher to focus on a relatively small set of key variables, possibly discarding valuable information. This paper proposes an easy way out of this dilemma: Do not make a choice. A wide range of theoretical and empirical literature has already demonstrated the superiority of combined to single-model based forecasts. Thus, the estimation and combination of parsimonious VARs, employing every reasonably estimable combination of the relevant variables, pose a viable path of dealing with the degrees of freedom restriction. The results of a broad empirical analysis based on pseudo out-of-sample forecasts indicate that attributing equal weights systematically out-performs single models as well as most more refined weighting schemes in terms of forecast accuracy and especially in terms of forecast stability.