Return predictability via an long short-term memory-based cross-section factor model: Evidence from Chinese stock market

Abstract

This paper proposes a cross-section long short-term memory (CS-LSTM) factor model to explore the possibility of estimating expected returns in the Chinese stock market. In contrast to previous machine-learning-based asset pricing models that make predictions directly on equity returns, CS-LSTM estimates are based on predictions of slope terms from Fama–MacBeth cross-section regressions using 16 stock characteristics as factor loadings. In line with previous studies in the context of the Chinese market, we find illiquidity and short-term momentum to be the most important factors in describing asset returns. By using 274 value-weighted portfolios as test assets, we systematically compare the performances of CS-LSTM and three other candidate models. Our CS-LSTM model consistently delivers better performance than the candidate models and beats the market at all different levels of transaction costs. In addition, we observe that assets with smaller cap are favored by the model. By repeating the empirical analysis based on the top 70% of stocks, our CS-LSTM model remains robust and consistently provides significant market-beating performance. Our findings from the CS-LSTM model have practical implications for the future development of the Chinese stock market and other emerging markets.