Bond Loss Rate Prediction Model Based On Multi-stage Method

The bond loss rate forecast is to predict the future loss rate of the bond in advance by establishing a functional relationship between the indicator vector and the bond loss rate,based on the existing historical data,which has important reference significance for bond investment decision-making.This paper is organized by introduction,the principle of the multi-stage prediction model of loss rate,the construction of the model,the empirical study of Chinese corporate bonds,and the conclusion.The main points of this study are as follows.First,the classification algorithm is preferred,the classification algorithm is different,and the discrimination results of the types of bonds are different,which ultimately leads to different prediction accuracy for the loss rate.Second,the selection of default risk assessment criteria,different standards,and different perspectives on risk.The third is predictive ability.The predictive ability of the model is different,and the use of investors is different.This paper has four academic contributions.Firstly,the paper selects the best prediction method based on the minimum error of the predicted value of the corporate bond loss rate(LR)and the actual value.Secondly,by directly predicting the LR of the bond,it is possible to directly determine the default loss in the case of known bond principal,avoiding double prediction error of the expected loss rate by simultaneously predicting PD and LGD.Thirdly,we use the index data of t-k(k=1,2,...,5)period and the LR of t period to establish models,and predict the LR of t+k(k=1,2,...,5)period by using the indicator data of t period.We have established a default risk prediction model with a forecast period of up to five periods.Fourth,the research shows that the prediction accuracy of the “DT+OLS” multi-stage model is higher than six typical single-stage models,including AdaBoost Regression,Bayesian Ridge Regression,Decision Tree Regression,Kernel Ridge Regression,Lasso Linear Regression,and Support Vector Machine Regression.The Mann-Whitney U test of the LR shows that bond issuance companies in China's financial markets have the best bond credits for “educational industry”,middle for “wholesale and retail industry”,the worst for "manufacturing industry ".