Forecast Selection in Unstable Environments

This paper leverages the time-series properties of forecast loss differences for out-of-sample forecast selection. Our framework predicts the conditional distribution of future loss differences while accommodating for time-contingent unstable forecasting environments. We establish distributional theory to quantify the sampling uncertainty of our predictions, enabling the development of advanced selection rules. Through simulations and an empirical application to inflation forecasting, we demonstrate the efficacy of our selection methodology and the potential for our advanced selection rules to achieve second-order forecasting objectives.