Forecasting Volatility in the Chinese Stock Market Using Deep Learning‐Based Hybrid Factor Models

Abstract

This paper develops deep learning‐based hybrid factor models for forecasting firm‐specific volatility in the Chinese stock market. The proposed models integrate deep learning‐based instrumented principal component analysis (IPCA) with an autoregressive module, allowing a rich set of observable features to inform the extraction of forecasting‐relevant common factors and factor loadings while accounting for the persistent dynamics of volatility. The empirical results show that the hybrid models can effectively forecast firm‐specific volatility over the full out‐of‐sample period and during extreme market episodes. Furthermore, we conduct a detailed investigation into the marginal contributions of different feature groups over time. The results indicate that factors such as price trends, sentiment measures, money and credit indicators, stock market variables, and cumulative news sentiment significantly improve the accuracy of volatility forecasts.