Autoencoder-Driven Latent Factor Models for A-Share Market

Thesis research applying autoencoder-based latent factor models to the Chinese A-share market.

Problem: Traditional linear factor models may not fully capture the non-linear risk-return structure in the A-share market.

Method: Constructed PCA, Instrumented PCA (IPCA), and conditional autoencoder latent factor models using 1996–2023 A-share data with 91 firm characteristics. Evaluated portfolio return explanatory power across models.

Result: The conditional autoencoder model significantly outperformed traditional linear models in out-of-sample tests, with superior Sharpe ratio and prediction accuracy, and lower pricing errors. Profitability and market size were identified as core risk factors in the A-share market, validating the effectiveness of latent factor models in China.