The Role of Cointegration in the Forecast Accuracy of Var Models

This paper assesses the forecast performance of a set of VAR models under a growing number of restrictions. With a maximum forecast horizon of 12 years, we show that the farther the horizon is, the more structured and restricted VAR models have to be to produce accurate forecasts. Indeed, unrestricted VAR models, not subjected to integration or cointegration, are poor forecasters for both short and long run horizons. Differenced VAR models, subject to integration, are reliable predictors for one-step horizons but ineffectual for multi-step horizons. Cointegrated VAR models including appropriate structural breaks and exogenous variables, as well as being subjected to over-identifying theory consistent restrictions, are excellent forecasters for both short and long run horizons. Hence, to obtain precise forecasts from VAR models, proper specification and cointegration are crucial for whatever horizons are at stake, while integration is relevant only for short run horizons.