Abstract: In order to capture time-varying features of volatility of asset price, multi-resolution analysis (MRA) was used to decompose financial returns into orthogonal components in different time domains. For each component, a certain ARMA-GARCH model was built. Extreme value theory (EVT) was then introduced so as to model the fat-tail of financial returns, and an MRA-EVT model was constructed. Finally, the proposed model was applied to predict VaR of CSI 300 index, and compared with traditional models, such as ARMA-GARCH model, unconditional EVT model and MRA model. Empirical results show that the MRA-EVT model significantly improves the accuracy of VaR estimation.