Research On Early Warning Of Credit Bond Default Risk Based On Logistic-stacking Model

As an important part of China’s financial market,the bond market shoulders the important responsibility in supporting and promoting the real economy.Super November,however,since the 2014 "11 Chaori Bond" default since the incident,the bond market default event frequent,especially since 2018,the market credit contraction and economic structure transformation environment,bond defaults and defaults began to rise sharply,at the same time,are no longer limited to small and medium-sized enterprises can find default subject,The default risk has also begun to spread to listed companies,which should attract people’s attention,especially some listed companies with great social influence.In this context,it is particularly important to establish a bond default warning model to measure its credit risk.Based on Chinese listed companies from 2014 to 2021 of 61 only 537 not only default default bonds and bonds as the research sample,on the basis of the analysis of affecting factors of bond defaults screening credit rating index,from the company’s profitability,capital structure and debt paying ability,operation ability,growth ability and cash flow level six aspects to select indicators,Feature selection and balance with sample data,respectively,on the establishment of the early warning model,in view of the current China’s model of machine learning algorithm using relative single stack and lack of multi-model fusion method research,this paper constructs the logistic regression,support vector machine(SVM),random forests,XGBoost and Light GBM5 groups algorithm model by contrast,Random Forest,XGBoost,Light GBM and logistic regression are selected to establish a two-layer bond default risk early warning model through Stacking fusion algorithm.Finally,the Logistic-Stacking model was compared with logistic regression,support vector machine,random Forest,XGBoost and Light GBM.The results show that:(1)strong cyclical industries such as manufacturing,wholesale and retail trade,and real estate have become the industries with high incidence of bond defaults.In addition,defaults are mostly concentrated in private enterprises,mostly distributed in the eastern region.(2)Bond default risk has a significant chain effect,which is continuously transmitted and aggregated on the whole risk chain,and eventually causes bond default.(3)In view of the non-equilibrium of sample data,the extended data set after the balance treatment of Borderline SMOTE algorithm significantly improves the prediction ability of the model.(4)According to the results of model comparison,the Logistic-Stacking model constructed based on Stacking fusion algorithm is superior to the single Logistic-Stacking model in terms of accuracy,accuracy,F1 value and AUC value.It indicates that the multi-model Stacking fusion strategy based on Stacking fusion algorithm has stronger prediction performance.(5)Total asset turnover ratio,asset-liability ratio,return on equity,EBITDA/total liabilities,and accounts receivable turnover rank high in the comprehensive contribution of the characteristic importance analysis,which can be used as key risk indicators for monitoring in daily life.The innovation of this paper mainly lies in the following:(1)In the previous researches on bond default,the default samples of ST companies are mostly the default samples,which are not rigorous to a certain extent.Therefore,this paper selects substantive default samples in the bond market to make the empirical results more reliable.(2)In order to solve the non-equilibrium problem of sample data,this paper uses Borderline-SMOTE algorithm to balance and optimize the sample data.(3)The Logistic-Stacking bond default warning model was established by Stacking fusion algorithm,and the prediction performance of the logistic-stacking bond default warning model was better than that of the single model.