ABSTRACT. Average monthly price data from twelve hinterland markets and the Houston port price for wheat are studied in a cointegration framework using the EngleGranger twostep procedure and Johansen's maximum likelihood procedure. Outofsample forecasts from an error correction model are compared to those from a vector autoregression fit to levels and a univariate autoregression fit to first differences. This comparison suggests that modeling these (cointegrated) data as a levels vector autoregression, rather than as an errorcorrection process, results in significantly higher error bias, but lower error variance, at long horizons. Copyright 1993, Wiley Blackwell. All rights reserved
