Bond Default Prediction Based On CNN-SVM Model Research

In 2014,"11 Chaori Bond" broke the history of rigid payment of bonds in China by failing to pay the interest on time,which constituted a substantial default.Since then,defaults in China’s bond market have been frequent,and bond defaults have entered a normalized stage.In the era of artificial intelligence,machine learning methods have been successfully applied to quantitative finance and risk management.In the context of the increasing prominence of bond defaults,how to use machine learning models to comprehensively and accurately assess and predict the default risk of bonds,establish a bond default prediction model suitable for China,and protect the legitimate rights and interests of collaborators is more the core aspect of technology-enabled finance and the use of technology to prevent and resolve financial problems.In this paper,we use substantial default data of Chinese bond market,train the samples based on oversampling algorithm,and try to propose a new integrated machine learning method to analyze and predict the default risk of credit bonds,which enriches the application of machine learning in the field of quantitative financial risk management and has important theoretical and practical significance.In order to better utilize the data of bond issuing entities and solve the current problems of bond default risk research,this paper selects all credit bonds that have incurred substantial defaults from 2014 to June 2022,extracts the data of corresponding bond issuing entities in the first two years of bond issuance year to construct training samples,and in terms of indicators,from profitability,solvency,operational capacity,growth capacity,and statement-derived indicators In terms of indicators,18 financial indicators were selected from five dimensions,and three macroeconomic indicators were selected.To address the problem of data imbalance,the default samples are rebalanced by data collation and synthetic minority oversampling techniques(SMOTE)and generative adversarial networks(GANs),and the new data set generated by the fitting makes the trained model more accurate and reasonable.In terms of models,based on the deep learning framework,random forest(RF),convolutional neural network(CNN),and support vector machine(SVM)are applied to build classification models respectively,and a one-step construction of a fusion prediction model of convolutional neural network and support vector machine(CNN-SVM)tries to improve the prediction effect of single classifier.In terms of evaluation metrics,accuracy,detection rate,recall rate,F-value and AUC value are used to evaluate the accuracy and comprehensive performance of predicting default risk.The empirical results show that,first,based on the prediction effectiveness of each machine learning model for China’s bond market,RF,CNN model,SVM and CNN-SVM fusion model all achieve an accuracy rate of over 93%,indicating that the model constructed in this paper is applicable to the Chinese bond market.Second,in terms of the importance of the selected indicators,when analyzing the importance of the factors based on the support vector institution-building forecasting model,the top three characteristic variables were found to be the return on net assets,net profit(year-overyear growth rate),and net sales margin.Third,in terms of data processing,the comprehensive performance of each model was improved by applying the rebalanced data trained with the oversampling algorithm to the CNN,SVM,and CNN-SVM models.Among them,the F1 value of the fusion model improved by 4.17% and the AUC improved by 2.57% when the SMOTE algorithm was applied,and the GANs algorithm showed a better and more stable improvement,with a 6.81% increase in F1 value and a3.12% increase in AUC.Fourth,from the comparison of the prediction effects of different models,the single classifier RF has slightly higher AUC and F1 values than the CNN and SVM models alone in terms of combined performance indicators,and the improved CNN-SVM model has higher prediction effects than either single classifier in terms of accuracy and combined performance indicators.