This paper predicts stock returns in the Chinese market by using an improved autoencoder machine learning approach and financial big data encompassing approximately one hundred firm characteristics. The findings demonstrate that the autoencoder factor can extract predictors from a large amount of information containing company characteristics to forecast returns and can achieve significant excess returns in the cross sections. Additionally, the analysis on the significance of factors reveals that the anomalies are timevarying in the Chinese stock market. Additionally, the predictive efficiency of the autoencoder method correlates with macroeconomic conditions and economic policies. The autoencoderbased longshort portfolios can effectively mitigate market risk especially during substantial market bubbles and heightened speculative periods, demonstrating well resilience to the shifts of fiscal and monetary policyinduced economic conditions.