Does co-integration help long-term forecasts? In this paper, we use si
mulation, real data sets, and multistep-ahead post-sample forecasts to
study this question. Based on the square root of the trace of forecas
ting error-covariance matrix, we found that for simulated data imposin
g the 'correct' unit-root constraints implied by co-integration does i
mprove the accuracy of forecasts. For real data sets, the answer is mi
xed. Imposing unit-root constraints suggested by co-integration tests
produces better forecasts for some cases, but fares poorly for others.
We give some explanations for the poor performance of co-integration
in longterm forecasting and discuss the practical implications of the
study. Finally, an adaptive forecasting procedure is found to perform
well in one- to ten-step-ahead forecasts.