Forecasting in factor augmented regressions under structural change

Factor augmented regressions are widely used to produce out-of-sample forecasts of macroeconomic and financial time series. However, these series are subject to occasional breaks. We study the effect of neglected structural instability on the forecasts produced by factor augmented regressions when the latent factors are estimated by cross-sectional averages from a large panel of variables. Our results show that neglecting structural instability can be very costly in terms of forecasting performance. We derive analytical results to show that instability in the factor model and in the forecasting equation impacts the produced forecasts. We further provide numerical results showing that conditioning upon the most recent break tends to produce more accurate forecasts than unconditional estimation methods based on expanding or rolling windows. However, the actual gain depends on the location and the magnitude of the breaks. Finally, an application to out-of-sample stock return forecasting using liquidity proxies illustrates the empirical relevance of our results.