Machine Learning-Based Short-Selling Indicators and the Cross-Section of Stock Returns

This study examines the impact of short-selling volume on future stock returns in the Korean stock market and proposes new short-selling indicators using machine learning techniques. Stocks with higher relative short-selling volume tend to exhibit significantly lower future returns, and this negative predictive power is confirmed through portfolio analysis and Fama-MacBeth regressions. Noting the limitation that existing indicators rely solely on past short-selling volume and thus have a lagging nature, this study develops leading indicators by predicting next month’s short-selling volume using stock loan balance data. In this process, various machine learning models are compared, and the ridge regression-based linear model demonstrates the best predictive performance. By presenting new predictive indicators derived from short-selling data, this study complements the existing literature and offers practical implications for investment strategies.