Are Stock Returns Predictable in China? A Machine Learning Approach
huihang wu,xingkong wei,Xiaoyan Zhang
DOI: https://doi.org/10.2139/ssrn.3971419
2019-01-01
Abstract:The predictability of stock returns has always been one of the core research questions in finance. This paper attempts to introduce machine learning method to answer whether stock returns are predictable in China. With 108 characteristics data in Chinese stock market from January 1997 to December 2019, this paper compares the out of sample predictability of the traditional econometric model with that of 6 major machine learning models, including partial least squares, principal component regression, elastic net regression, random forests, gradient boosted regression trees and neural networks. The main findings of this study are as follows: (1) historical trading data can predict individual stock returns in the next month, and the out-of-sample prediction of machine learning algorithm is better than that of traditional econometrics model; (2) in Chinese stock market, liquidity characteristics have strong predictive power, while momentum characteristics are weak in out of sample prediction; (3) the combination of machine learning algorithm and asset pricing research can generate significant economic value. During the out of sample test period, the performance of two-layer neural network equal-weighted (value-weighted) long-short strategy is the best among all models, with an average monthly return of 3.03% (2.94%), the monthly volatility of 4.65% (6.88%), the annualized Sharpe ratio of 2.26 (1.48), and the significant monthly adjusted Alpha of 3.03 (2.95) in terms of FF5 factor. We present results that demonstrate that machine learning algorithm have clear merit over traditional techniques in China.