The autoencoder asset pricing model in the Chinese stock market

We apply the autoencoder model to study asset pricing in the Chinese stock market and compare it with the instrumented principal component analysis (IPCA). The IPCA assumes a linear relationship between factor loadings and stock characteristics, while the autoencoder allows for nonlinear relationships, offering greater flexibility. Using data from nearly all stocks on the Shanghai and Shenzhen exchanges (2006–2023), we find the autoencoder outperforms the IPCA in return prediction, as measured by predictive ${R^2}$R2 and zero-net long-short portfolio performance. However, the IPCA better explains cross-sectional return variation with a higher total ${R^2}$R2, consistent with Gu, Kelly, and Xiu’s (2021) findings in the U.S. market. Given short-selling restrictions in China, we also evaluate long-only portfolios, where IPCA predictions achieve higher Sharpe ratios. While liquidity and trend factors dominate in the U.S. we find liquidity, fundamentals and valuation factors most important in China. Additionally, state-owned firms’ returns are easier to predict but harder to explain in cross-sectional terms than non-state-owned firms. Our study offers new insights into asset pricing in the Chinese stock market.