Forecasting using random subspace methods

Abstract

Random subspace methods are a new approach to obtain accurate forecasts in high-dimensional regression settings. Forecasts are constructed by averaging over forecasts from many submodels generated by random selection or random Gaussian weighting of predictors. This paper derives upper bounds on the asymptotic mean squared forecast error of these strategies, which show that the methods are particularly suitable for macroeconomic forecasting. An empirical application to the FRED-MD data confirms the theoretical findings, and shows random subspace methods to outperform competing methods on key macroeconomic indicators.