LAUSR

This paper takes 11-year 5-minute high-frequency trading data of the Shanghai and Shenzhen 300 Index (CSI300) as a research sample. First, it proposes a method to define the normal state and the state of attention of the financial market based on multi-fractal characteristics, and randomly owes it Sampling (RU), synthetic minority oversampling (SMOTE) and traditional support vector machine (SVM) are combined to propose an improved SVM model—RU-SMOTE-SVM model to predict extreme risks in China’s financial market, and compare Traditional SVM, SMOTE-SVM, RU-SMOTE-NN and RU-SMOTE-DT are compared. The empirical results show that the price fluctuations of China’s emerging financial markets have significant multi-fractal characteristics; the normal and concerned states defined based on the multi-fractal feature parameters are not only accurate, but also have obvious statistical test significance and clear practical significance; and traditional SVM and Compared with BP neural network (NN), RU-SMOTE-SVM is not only significantly higher in prediction accuracy, but also in terms of prediction stability. That is, RU-SMOTE-SVM can effectively solve the problems of other early warning models to solve the symmetrical sample problem.