Research On The Prediction Of Bond Default Risk Based On Machine Learning

After 40 years of development,the China bond market has become an important financial market.However,since the rigid redemption situation was broken in 2014,bond defaults have also entered a normalization stage,and the default risk of bonds has gradually been exposed to market participants in a more obvious way.In the context of noticeable bond default risks,however,the quality of China's credit ratings is controversial.Therefore,it is of practical significance to establish new and effective methods to reasonably identify and control default risks.With the help of widely used machine learning models,I study the problem of predicting bond defaults and related events,in order to effectively identify and reduce the bond default risks faced by investors,and provide ideas for improving the quality of credit ratings.Therefore,this paper raises three research questions:first,can the machine learning model effectively predict bond defaults and related events? Second,if so,then which type of model has better generalization performance? Third,what kind of features play a more important role in identifying default bonds?I use eight machine learning models,including Logistic Regression,Lasso,Ridge,Elastic Net,Support Vector Machine,Random Forest,XGBoost,and Deep Neural Network,to study bond defaults and related event prediction.This paper first examines the generalization performance of each model to answer question one.I find that all of the machine learning models can effectively predict defaults and related events.The random forest and XGBoost have the best performance,with prediction accuracy of 84.16% and 79.87% respectively,and the ability to identify positive cases(F1)of 0.8118 and 0.7495 respectively.As for the second question,through Friedman test and Nemenyi post-hoc test,this paper finds that random forest and XGBoost have better performance,while Logistic regression,which is used in most literature in other related fields,is not the optimal model for bond default prediction.As for the third question,this paper uses random forest and XGBoost to conduct research,and finds that more attention should be paid to the profitability and characteristics of the company and the characteristics of the bond when assessing the risk of bond default.Finally,this paper changes the ratio of default and healthy sample and the time span of financial data respectively,in order to construct two new data sets for robust test,and finds that machine learning can still effectively predict bond defaults and related events.The conclusion is still stable.Random forest and XGBoost are still the most outstanding models.The contribution of this paper has three main points.First,the paper enriches the literature research on the application of machine learning models for bond default prediction.I use 8 types of machine learning models to study the problem of bond default prediction and find algorithms with outstanding performance,which provides a reference for further development of machine learning applications and theoretical research on bond defaults.Second,the paper enriches the literature research on the factors of China bond defaults.This paper comprehensively studies the contribution of various types of variables to predicting bond defaults,and identifies the most important types of factor,which provides ideas for further research on bond default factors,transmission mechanisms and other related issues.Third,the paper quantitatively reveals the contribution of the non-linear model setting to the bond default prediction.This paper constructs two types of cross-product variables to explicitly characterize the nonlinear characteristics of the model,and quantitatively explores the importance of nonlinearity in default prediction through factor importance analysis.Fourth,from the perspective of practical meanings,this paper achieves good results in the prediction of bond defaults by applying machine learning models,providing ideas for improving the quality of China's credit ratings,and also providing effective tools for bond investors to identify default risks.