Application of Unbalanced Data Classification Algorithm in Quantitative Financial Risk Management. Issue 4 (October 2020)

• Record Type: Journal Article
• Authors: Zhu, JinPeng; Wang, HanChen
• Is Part Of: Journal of physics. Volume 1648:Issue 4(2020)
• Publication Date: 2020-10
• Subjects: Financial Risk -- Intelligent Early Warning -- Unbalanced Data Classification Algorithm -- SVM Model; Physics -- Congresses; 530.5
• DOI: 10.1088/1742-6596/1648/4/042093
• Languages: English
• ISSNs: 1742-6588

Abstract

Abstract: The global financial crisis that has erupted many times in recent years has not only harmed the financial systems of various countries, but even severely hit the economies of various countries. Therefore, the importance of financial risk early warning has become more prominent and the challenges facing it have become more severe. In view of this, it is very important to explore an extreme risk early warning model that suits the reality of the Chinese financial market, and to accurately predict and prevent extreme financial risks. This article first constructs the model's early-warning indicator system, and determines the state indicators by synthesizing the two state indicator definition methods based on the crisis period and EVT, in order to determine whether extreme financial risks occur in the financial market at a certain time. The prediction performance of the improved SVM under different unbalanced sample data sets is compared. It is highly feasible to use it for early warning of extreme risks in China's financial market. This paper presents a BADASYN algorithm based on boundary sample adaptive synthesis at the data level. The algorithm first finds a small number of samples in the class boundary region, then adaptively synthesizes some samples according to their distribution, and adds the newly synthesized samples to the training set. In the data set sampled by BADASYN, the support vector of the trained SVM model is mainly composed of newly synthesized samples, and finally the separation hyperplane is close to multiple types of samples. Experimental research shows that after testing SVMs constructed with four kernel functions, the prediction accuracy of each model is very high, reaching more than 93%, reflecting the stability of SVM prediction performance.