Machine Learning-Based Research on Anomalies in China's Stock Market

Abstract

This paper systematically uses machine learning algorithms to assess the predictability of Chinese stock returns while analyzing a comprehensive set of anomalies. Based on the anomaly literature, 108 anomalies are constructed for the A-share market from 2000 to 2021, and eight machine learning algorithms, including ordinary least squares, lasso regression, ridge regression, elastic network regression, principal component regression, partial least squares regression, extreme gradient boosting tree, and feedforward neural network, are used to construct stock return prediction models and portfolios. Compared with previous studies on the US market, the Chinese stock market has higher predictability, with an out-of-sample R2 of 1.48% and a maximum monthly return of 2.79% for long-short portfolios. However, the machine learning models are found to be sensitive to the quality of the data and must be adequately trained and tuned to overcome the high dimensionality and avoid overfitting. This paper further examines the importance of factors in the prediction model and finds that the liquidity factor and risk factor have strong predictive power in the A-share market.