Stacking Regression for Time-Series, with an Application to Forecasting Quarterly US GDP Growth
Published in Optimal Transport Statistics for Economics and Related Topics, 2023
Abstract: Machine learning methods are being increasingly adopted in economic forecasting. Many learners are available, and a practical issue now presents itself: which one(s) to use? The answer we suggest is ‘stacking regression’ (Wolpert, 1992), an ensemble method for combining predictions of different learners. We show how to use stacking regression in the time series setting. Macroeconomic and financial time series data present their own challenges to forecasting (extreme values, regime changes, etc.), and this presents challenges to stacking as well. Our findings suggest that using absolute deviations for scoring the base learners performs well compared to stacking on mean squared error. We illustrate this with a Monte Carlo exercise and an empirical application: forecasting US GDP growth around the Covid-19 pandemic.