Cointegration and Long-Horizon Forecasting (1997)
Summary
The working paper Cointegration and Long-Horizon Forecasting (1997) argues that, for long-horizon forecasts, imposing cointegration may not improve accuracy when evaluated with standard multivariate forecast metrics; simple univariate Box-Jenkins forecasts can be just as accurate. The authors also highlight a potential flaw in conventional forecast-accuracy measures, which tend to undervalue the value of maintaining cointegrating relationships and propose alternatives that account for them.