Implications of Cointegration for Forecasting: A Review and an Empirical Analysis

Cointegration has different theoretical implications for forecasting. Several empirical studies have compared the out of sample forecasting performance of cointegrted VECMs against unrestricted VARs in levels and in differences. The results of these studies have been generally mixed and inconclusive. This paper provides a comprehensive review over the subject, and also examines the effects of cointegration rank restrictions on forecasting performance of VAR models through conducting an empirical exercise in the framework of a new twocountry (Canada-US) model. The results show that a VAR/DVAR model forecasts as well as the best cointegrated VAR model (and even better). Therefore, it seems that using cointegration techniques does not pay a dividend.