Exploring the Factor ZOO with a Machine-Learning Portfolio
Log in to collectAcademic paper
Exploring the Factor Zoo With A Machine-Learning Portfolio
Halis Sak; Michael T. Chng; Tao Huang
- Shenzhen University
- ?Independent - affiliation not provided to SSRN
- Beijing Normal-Hong Kong Baptist University
- HKHong Kong Baptist University
- Beijing Normal University
- ?Beijing Normal University-Hong Kong Baptist University United International College
Strategy in a nutshell
The study constructs a machine-learning (ML) stock portfolio using 36 years of data, dividing it into a train (1980–1998) and test (1998–2016) sample. ML models (ET, GBDT, and linear regressions) are trained on 106 firm characteristics, and the best-performing models are ensemble-weighted to predict next-month returns. Firms are decile-sorted to form long-short portfolios, with top-decile predicted winners bought and top-decile predicted losers sold. The ML portfolio consistently generates significant alpha (17–29% annualized) across entrenched factor models, driven by time-varying dominant characteristics linked to market, value, profitability, and momentum factors.
Economic rationale
The paper addresses the dynamic nature of stock characteristics within the factor zoo, emphasizing how evolving firm- and investor-level factors affect returns. By applying ML methods to a large database of firm characteristics, the study captures non-linear relationships and generates out-of-sample forecasts that outperform traditional models. Only a few subsets of characteristics, such as investor arbitrage constraints and firm financial constraints, dominate at different times, explaining the ML portfolio’s significant alpha and highlighting the importance of time-varying factor structures in asset pricing.