The sensitivity of VAR forecasts to alternative lag structures

Abstract

Although numerous articles examine the effect on VAR forecasts from changes in the variables included in a model, little has been done to examine the effect that changing the lag structure has on forecasting accuracy. The latter analysis is made difficult by the fact that no overriding rule exists for ex ante selection of lag length in such models. This paper examines the sensitivity of forecasts from a VAR model using different lag structures. Holding constant the variables included and the time period studied, different lag structures are used. For example, we use simple ad hoc rules as well as statistical criteria, such as mean square error and Bayesian rules. Our results indicate that the accuracy of VAR forecasts varies dramatically across alternative lag structures. Moreover, our results show relatively short-lagged models to be more accurate, on average, than longer-lagged specifications.