The predictability of stock returns has always been one of the core research issues in finance. XIN Real Estate Fintech Research Center (THUIFR), attempts to introduce machine learning method to explore the answer to the problem of predictable returns in China. This article reports on research to evaluates the effectiveness of traditional econometric and machine learning models in predicting 108 anomalous stock return characteristics in the A-share market between January 1997 to December 2019. The techniques evaluated include: traditional econometric models and least squares regression, principal component regression, regression, random forests, and gradient boosting. The performance of the six mainstream machine learning algorithms and neural network models on the predictability of performance of A-share stocks were included in the analysis. The main findings of the study were as follows :
(1) The historical transaction data information still had a predictive effect on the return rate of individual stocks in the following month, and the out-of-sample prediction effect of machine learning algorithm is better than that of traditional econometric model.
(2) In China's A-share market, the predictive ability of liquidity characteristic variables is strong, while momentum characteristic variables tend to be weak.
(3) The machine learning algorithms combined with asset pricing research has remarkable economic significance. The performance of two layer neural network weights (market-weighted value), such as long-short portfolio performance in all models, averaged during the out-of-sample test period average could obtain; 3.03% (2.94%) of monthly returns, monthly volatility of 4.65% (6.88%); an annualized Sharpe ratio of 2.26 (1.48); and after adjusting for the FF5 factor, could still get a significant monthly Alpha value of 3.03 (2.95).