The Term Structure of Machine Learning Alpha
Log in to collectAcademic paper
The Term Structure of Machine Learning Alpha
David Blitz; Matthias X. Hanauer; Tobias Hoogteijling; Clint Howard
- ?Robeco Institutional Asset Management
- ?Robeco Quantitative Investments
- DETechnical University of Munich
- ?Technische Universität München (TUM)
- University of Technology Sydney
- ?Abu Dhabi Investment Authority
Strategy in a nutshell
The strategy uses machine learning models—OLS, Elastic Net, Gradient Boosted Regression Trees, and a three-layer deep neural network—to predict U.S. stock returns across multiple horizons (1-, 3-, 6-, and 12-months). Portfolios are constructed using a 10/50 buy-hold rule: long positions in top-ranked stocks and short positions in bottom-ranked stocks based on predicted returns and prior rankings. An ensemble model aggregates predictions for improved accuracy. Trading costs are accounted for, and the optimal prediction horizon is selected based on net returns. Continuous monitoring of performance metrics ensures the strategy adapts to changing market conditions while maintaining cost efficiency.
Economic rationale
The strategy’s effectiveness arises from machine learning models’ ability to capture complex, nonlinear relationships among stock characteristics and future returns. By analyzing multiple prediction horizons, it adapts to both short-term and long-term market dynamics, extracting predictive signals from diverse features. Efficient portfolio construction rules mitigate trading costs, enabling the strategy to convert predictive accuracy into real-world profitability. Its flexibility across time scales and ability to exploit both high-turnover short-term signals and more stable Quality/Value signals underlie its consistent performance and competitive advantage over traditional linear methods.