Bimodal Characteristic Returns and Predictability Enhancement via Machine Learning

• 成果类型: Article
• 署名作者: Han, Chulwoo
• 署名单位: Durham University
• 刊物名称: MANAGEMENT SCIENCE
• ISSN/ISSBN: 0025-1909
• DOI: 10.1287/mnsc.2021.4189
• 发表日期: 2022
• 页码: 7701-7741
• 关键词: bimodality; deep momentum; Machine Learning; Deep neural network; reclassification

摘要

This paper documents the bimodality of momentum stocks: both high- and low-momentum stocks have nontrivial probabilities for both high and low returns. The bimodality makes the momentum strategy fundamentally risky and can cause a large loss. To alleviate the bimodality and improve return predictability, this paper develops a novel cross-sectional prediction model via machine learning. By reclassifying stocks based on their predicted financial performance, the model significantly outperforms off-the-shelf machine learning models. Tested on the U.S. market, a value-weighted long-short portfolio earns a monthly alpha of 2.4% (t-statistic = 6.63) when regressed against the Fama-French five factors plus the momentum and short-term reversal factors.