VAR, Error Correction and Pretest Forecasts at Long Horizons

This paper focuses on the construction of forecasts over long horizons where a typical, long-horizon forecast might span four years using twenty to forty years' data. It is argued that the presence of persistence in the form of unit or near-unit autoregressive roots poses substantial difficulties for long-horizon interval and point forecasting. These difficulties may not be overcome even by efficient pretesting or model-selection procedures and might, in general, lead to point forecasts with large asymptotic root mean squared errors and undesirably wide prediction intervals. Copyright 1996 by Blackwell Publishing Ltd