LeSage (1990) runs large-scale forecasting experiments across 50 Ohio industries to compare the error correction model (ECM), unrestricted vector autoregression (VAR), two Bayesian VAR (BVAR) variants (Minnesota and block-recursive), and a Bayesian ECM hybrid (BECM). For the seven industries with confirmed cointegration, the ECM dominates at horizons 3–12 months with ~20% lower mean absolute percentage error (MAPE) at month 12. The BECM, despite carrying a misspecified random-walk prior, performs nearly as well as the ECM — contradicting Engle-Yoo's theoretical prediction. Most surprisingly, both ECM and BECM produce the best long-horizon forecasts even for industries without cointegration, due to the error correction (EC) variable entering with a diffuse prior.
"The improvement in forecasting performance provided by the ECM model is a monotonically increasing function of the forecasting horizon."
"Forecasting experiments may provide an alternative approach to resolving conflicting cointegration test results."
"There may be a role for the ECM or BECM model in forecasting even when the cointegration tests suggest the lack of a cointegrating vector."
The empirical scope — 50 industries, 25-month rolling evaluation, 5 competing models — is unusual for 1990 and lends credibility to the conclusions. The BECM paradox is the most interesting result: it works partly because the diffuse prior on the EC term makes Bayesian shrinkage increase rather than decrease the EC variable's effective weight. The interpretation of forecasting performance as a cointegration diagnostic is pragmatically useful but theoretically loose — forecast improvement from the EC variable could reflect any persistent long-run relationship, not strict cointegration. The absence of Johansen tests (only Engle-Granger ADF) is a limitation by modern standards.