An Empirical Study of Machine Learning for Filtering Momentum Signals: Daily Frequency Strategy Analysis Based on the A-Share Market

Abstract

Momentum strategies are commonly used in quantitative investing, but they are susceptible to noise signals, false breakouts, and market reversals in the A-share market, leading to significant performance volatility. This paper selects 80 A-shares from January 2020 to May 2025 as samples to investigate whether machine learning filtering can improve the effectiveness of momentum signals. The study generates buy signals based on 20-day cumulative returns and rolling quantile thresholds, and constructs features such as momentum, volatility, trading volume, price position, and trend. A gradient boosting classifier is used to predict the return direction the day after the signal is issued, and the effectiveness of the strategy before and after filtering is compared through single-asset backtesting considering stop-loss and transaction costs. The results show that after machine learning filtering, the Sharpe ratio and win rate of most stocks are improved, and the improvement in Sharpe ratio is statistically significant. However, there is no stable positive correlation between AUC and strategy improvement, indicating that machine learning is more suitable as a signal evaluation tool.