Forecasting with a Random Walk

The use of different time-series models to generate forecasts is fairly usual in the fields of macroeconomics and financial economics. When the target variable is stationary, the use of processes with unit roots may seem counterintuitive. Nevertheless, in this paper we demonstrate that forecasting a stationary variable with forecasts based on driftless unit-root processes generates bounded mean squared prediction errors at every single horizon. We also show that these forecasts are unbiased. In addition, we show via simulations that persistent stationary processes may be better predicted by driftless unit-root-based forecasts than by forecasts coming from a model that is correctly specified but is subject to a higher degree of parameter uncertainty. Finally, we provide an empirical illustration of our findings in the context of CPI inflation forecasts for a sample of industrialized economies.