Automated Machine Learning and Asset Pricing

Healy, Jerome V.; Gregoriou, Andros; Hudson, Robert

Abstract

We evaluate whether machine learning methods can better model excess portfolio returns compared to the standard regression-based strategies generally used in the finance and econometric literature. We examine 17 benchmark factor model specifications based on Expected Utility Theory and theory drawn from behavioural finance. We assess whether machine learning can identify features of the data-generating process undetected by standard methods and rank the best-performing algorithms. Our tests use 95 years of CRSP data, from 1926 to 2021, encompassing the price history of the broad US stock market. Our findings suggest that machine learning methods provide more accurate models of stock returns based on risk factors than standard regression-based methods of estimation. They also indicate that certain risk factors and combinations of risk factors may be more attractive when more appropriate account is taken of the non-linear properties of the underlying assets.

Citation

Healy, J. V., Gregoriou, A., & Hudson, R. (2024). Automated Machine Learning and Asset Pricing. Risks, 12(9), Article 148. https://doi.org/10.3390/risks12090148

• Journal Article Type: Article
• Acceptance Date: Sep 9, 2024
• Online Publication Date: Sep 14, 2024
• Publication Date: Sep 1, 2024
• Deposit Date: Oct 8, 2024
• Publicly Available Date: Oct 8, 2024
• Journal: Risks
• Electronic ISSN: 2227-9091
• Publisher: MDPI
• Peer Reviewed: Peer Reviewed
• Volume: 12
• Issue: 9
• Article Number: 148
• DOI: https://doi.org/10.3390/risks12090148
• Keywords: Machine learning; Asset pricing; Risk factors; Prospect theory; Peak-End rule
• Public URL: https://hull-repository.worktribe.com/output/4861892