Combining Forecasts: Can Machines Beat the Average?

Abstract

This paper documents benefits of combining forecasts using weights that depend non-linearly of past forecast errors. We propose combining out of sample forecasts from simple models using weights, computed using machine learning algorithms trained on the models' past forecast errors. These nonlinear weights produce more accurate forecasts than conventional approaches based on equal-weighed forecasts, breaking the so-called "forecast combination puzzle''.

Keywords: Forecast Combination; Machine Learning; Non-Linear Models

JEL Classification: C45; C52; C53; C55