Research On Early Warning Of Extreme Risks In Financial Markets Based On Multifractality And Unbalanced Samples

The risk warning of the stock market has always been a key research issue in the field of financial risk management.With the development of information technology,deep learning has gradually become a popular method for time series analysis of stock market.However,due to the imbalance of extreme risk sample data,the model is prone to misclassification of the majority class.This paper is to use multifractality theory,unbalanced data processing method,Res Net neural network and Transformer time series model to study the problem of extreme risk early warning in the stock market.For the identification of extreme risk states,this paper combines the complex characteristics of financial time series data,and introduces a discriminant method based on multifractality for the first time on the basis of EVT and financial crisis period methods,making the risk state samples more representative.In addition,considering the risk of the stock market itself and the external contagion of risk,the methods of factor analysis,KS statistical test and Copula are used to screen and extract 17 representative risk characteristic indicators from both domestic and foreign markets to form an extreme risk sample data set.Finally,in order to solve the problem of unbalanced data,this paper adopts seven unbalanced algorithms of different sampling types such as SMOTE,Adasyn,and Prototype generation(PG)to adjust the categories of financial risk data sets,and introduces deep learning models TS Res Net and Transformer From the model to the study of financial extreme risks,it is compared with the SVM model commonly used in the industry to jointly carry out early warning analysis of extreme risks.The experimental results show that using the unbalanced data algorithm for preprocessing can effectively improve the classification accuracy of the model,among which the PG algorithm in the downsampling algorithm has the best effect,and realizes the relief of data imbalance;combined with the improved SVM of the unbalanced algorithm,The three models of TS Res Net and Transformer can effectively identify risks for the time series characteristics of stock market data and have strong stability.Among them,compared with TS Res Net and Transformer models,the prediction advantage of SVM is not obvious,and the Transformer model has the best performance,followed by TS Res Net.The research results provide a reference for the introduction of deep learning models in financial markets for risk identification.In addition,the research methods and conclusions of this paper also have certain practical significance in financial supervision,Shanghai and Shenzhen stock market trend control,and investor risk decision-making.