Evaluating Forecast Performance with State Dependence

Abstract

In forecasting models, usually no single model emerges as the best overall, as forecasting performance is prone to instabilities because the economic mechanisms providing the data work better on one model during some periods than others (state-dependent). Here, the authors propose a new forecast evaluation methodology to assess models’ absolute and relative forecasting performance when the model is a state-dependent function of economic variables. Results show that these tests uncover “pockets of predictability� in U.S. equity premia and forecasts significantly better than the benchmark forecast when real GDP growth is low. These tests can be applied to relative forecast comparisons, forecast encompassing, efficiency, and, more generally, moment-based tests of forecast evaluation.