In this paper, the CAViaR model was used to measure the HS300, HSI and N225 risk levels and result showed that CAViaR model was able to measure the sample risk value at the 5% confidence level. However, at the 1% confidence level, the CAViaR model severely underestimated the sample risk value. In order to solve this question, this paper introduces the POT model and CAViaR model to construct the extreme risk measure model based on POT-CAViaR model and the test shows that the improved model can effectively improve the accuracy of extreme risk prediction and the model validity is strengthened.