It is widely believed that imposing cointegration on a forecasting system, if cointegration is in fact present, will improve long-horizon forecasts. Contrary to this belief, we show that at long horizons nothing is lost by ignoring cointegration when the forecasts are evaluated using standard multivariate forecast accuracy measures. In fact, simple univariate Box-Jenkins forecasts are just as accurate. Our results highlight a potentially important deficiency of standard forecast accuracy measures -- they fail to value the maintenance of cointegrating relationships among variables -- and we suggest alternatives that explicitly do so.