Beta Forecasting with Realized Beta Estimators and Machine Learning Algorithms

Abstract

This paper applies machine learning algorithms to the modeling of realized betas for the purposes of forecasting stock systematic risk. Forecast horizons range from 1 week up to 1 month. The machine learning algorithms employed are ridge regression, decision tree learning, adaptive boosting, gradient boosting, random forests and neural networks. We also evaluate forecasts from these algorithms against the forecasts generated from simple constant realized beta estimators. We find that the machine learning algorithms can generate lower beta forecast error, relative to current benchmarks in beta forecasting.

Presented at the 40th International Symposium on Forecasting, October 26-28, 2020.

Keywords: CAPM, Systematic Risk

JEL Classification: C52, C58, G17