Accurately predicting extreme tail risk is crucial for China’s banking sector. However, the need for long time series in order to capture extreme events introduces nonstationarity to financial data, posing significant challenges to traditional volatility models, which assume stationarity. Further, the limited availability of extreme tail data can result in unstable estimates, complicating the risk assessment process. To address these issues, we propose the component GARCH-MIDAS-EVT-X model: a model to forecast extreme value-at-risk (VaR) and expected shortfall (ES) that combines generalized autoregressive conditional heteroscedasticity (GARCH), mixed data sampling (MIDAS), extreme value theory (EVT) and economic covariates . The component GARCH-MIDAS volatility model surpasses conventional GARCH approaches by effectively capturing dynamic changes in both conditional and unconditional variance, as is essential for accurately modeling the complex and evolving volatility of China’s banking sector. In addition, the integration of EVT addresses the challenge of data scarcity for extreme events. We apply our model to a stock index representing China’s banking sector and employ a comprehensive set of backtesting methods for VaR and ES. The results demonstrate that our model outperforms other competing models, underscoring its effectiveness in forecasting extreme risks. This study demonstrates the effectiveness of combining component GARCH-MIDAS and EVT methodologies, offering a substantial contribution to extreme risk prediction in the context of China’s banking sector.