Unit-Root Tests Are Useful for Selecting Forecasting Models

We study the usefulness of unit-root tests as diagnostic tools for selecting forecasting models. Difference-stationary and trend-stationary models of economic and financial time series often imply very different predictions, so deciding which model to use is tremendously important for applied forecasters. We consider three strategies: always difference the data, never difference, or use a unit-root pretest. We characterize the predictive loss of these strategies for the canonical AR(1) process with trend, focusing on the effects of sample size, forecast horizon, and degree of persistence. We show that pretesting routinely improves forecast accuracy relative to forecasts from models in differences, and we give conditions under which pretesting is likely to improve forecast accuracy relative to forecasts from models in levels.