Ensemble Machine Learning and Stock Return Predictability
AsianFA 2019, AMES 2019, FMND 2019
50 Pages Posted: 8 Jan 2019 Last revised: 17 Sep 2019
Date Written: March 1, 2019
Many, even sophisticated, models cannot beat a simple mean combination of univariate stock market return forecasts. We introduce an ensemble machine learning method, which averages forecasts from sophisticated models (like BMA, WALS and LASSO) based on random subsamples and which learns from its mistakes by adaptively changing sampling distributions. Empirically, our novel method improves the simple mean forecast with statistically significant monthly out-of-sample R2 of around 2-3% and annual utility gains around 3%. Our approach benefits from predicting well in volatile periods and especially from extreme market drops. The forecasting gains of our new method stem from improved diversity among individual forecasts. We obtain similar gains in forecasting accuracy when we use our method to predict factor portfolios and other macro economic variables.
Keywords: Equity Premium Prediction, Machine Learning, Forecast Combination, Parameter Uncertainty, Diversification
JEL Classification: G17, G12, G02, C58
Suggested Citation: Suggested Citation