Among others, using Monte Carlo experiments, Engle and Yoo (1987), and Lin and Tsay (1996) have argued that imposing cointegration constraint produces superior long-term forecasts. However, Christoffersen and Diebold (1998) have noted that these earlier studies had misinterpreted their Monte Carlo results. Christoffersen and Diebold (1998) analytically showed that the cointegration constraint does not improve long-term forecasts. Instead, they argued that only integration constraint matters in long-term forecasts.
In this paper, we show that when using the MSE criterion, neither cointegration nor integration leads to superior long-term forecasts. Explicitly, we analytically show that as the forecast horizon goes to infinity, the accuracy of the cointegrated system forecast approaches that of forecasts which impose neither integration nor cointegration, essentially following the analytical methodology in Christoffersen and Diebold (1998). Our Monte Carlo experiments yield results which corroborate our analytical results even for the forecasts with estimated parameters in finite samples. The experiments further reveal that the correct specification of the drift term plays a key role in the accuracy of long-term forecasts, particularly when we construct the forecasts based upon estimated parameters.