COINTEGRATION AND LONG-HORIZON FORECASTING

We consider the forecasting of cointegrated variables, and we show tha t at long horizons nothing is lost by ignoring cointegration when fore casts are evaluated using standard multivariate forecast accuracy meas ures. In fact, simple univariate Box-Jenkins forecasts are just as acc urate. Our results highlight a potentially important deficiency of sta ndard forecast accuracy measures-they fail to value the maintenance of cointegrating relationships among variables-and we suggest alternativ es that explicitly do so.