Forecasting vector autoregressions - the influence of cointegration: a Monte Carlo study

This paper investigates the forecasting performance of cointegrated systems by simulation. it builds on the investigation by engle and yoo (1987) and extends their work in two important directions. first, we also consider a conditional maximum likelihood procedure due to johansen (1988) and find improved forecasting performance relative to previous suggestions, but the gain remains small. second, examination of a three-dimensional system shows that, even for long period forecasts, those vars which overestimate the true number of common trends (overdifferenced systems) perform only slightly worse relative to the correct model but considerably better than vars on which too much cointegration is imposed.;