Reproduce “Deep Learning for Forecasting Stock Returns in the Cross-Section” on OpenResearch

About this paper

Masaya Abe, Hideki Nakayama

Researchers from Nomura Asset Management and The University of Tokyo developed and evaluated deep neural networks for forecasting one-month-ahead cross-sectional stock returns in the Japanese market. Their deep learning models consistently achieved higher rank correlation and directional accuracy compared to shallow neural networks, and often outperformed or matched the performance of Support Vector Regression and Random Forests, resulting in favorable risk-adjusted returns in simulated long-short portfolios.

Key insights

• Deep neural networks generally exhibit superior performance in cross-sectional stock return prediction compared to shallower neural networks, demonstrating the value of multi-layered architectures for extracting complex features from financial data.
• The best-performing deep learning models demonstrated competitive or improved predictive power when benchmarked against established machine learning methods such as Support Vector Regression and Random Forests.
• Combining diverse machine learning models through an ensemble approach can further enhance prediction accuracy, yielding the highest rank correlation in the study.

Summary by alphaXiv.