Forecasting Volatility in the Chinese Stock Market: A Study Based on GARCH-Type Models and Machine Learning Algorithms

This study focuses on the return series of the Shanghai Composite Index, and ChiNext Index over the past decade, using realized volatility calculated from 5-minute high-frequency data as a proxy for true volatility. To enhance prediction accuracy, the study incorporates training set residuals into machine learning algorithms such as XGBOOST, Random Forest, and GRU. Compared to the forecasting results of the GARCH model, machine learning models significantly capture volatility fluctuations during periods of high market turbulence, demonstrating notable improvements in predictive accuracy. Furthermore, by calculating utility gains, the prediction results assist investors in optimizing their investment strategies based on the mean-variance model. The findings aim to provide investors and regulatory bodies with more forward-looking and adaptive tools for anticipating market dynamics, thereby offering valuable insights for academic research in financial markets.