Research On Default Prediction Model Of P2P Net Loan Borrowers

With the rapid development of the Internet finance,P2P online lending has attracted increasing attention from users as a part of Internet finance due to its convenience and efficiency.In the rapid expansion of online lending business,it also exposed a series of problems,especially the bad debts caused by the high default rate of borrowers,which will directly cause losses to the platform and lenders.Based on the risk of P2P network lending platform and the default factors of borrowers,this thesis uses the actual borrower transaction data to construct default prediction model and enhances the platform’s ability to identify default users by reducing the default risk of borrowers.In turn,control the high bad debt rate and improve the survivability of the platform.This thesis explores the default characteristics of Lending Club’s latest borrower data for the problem that the borrower’s high default rate causes the P2P platform to fail to operate normally.Firstly,the data set is extracted,cleaned and transformed.The variables are selected by the feature selection method based on random forest algorithm and the feature selection method of random forest algorithm based on recursive feature elimination.Logistic regression and decision tree model are established,and the confusion matrix is used,ROC curve and other evaluation indicators select the best method.Secondly,based on the traditional statistical model,the machine learning algorithm GBDT is introduced into the default prediction model.The GBDT algorithm is used to select features and use the input variables of the logistic regression model to train the classifier and build GBDT+LR model.Finally compare the prediction effect of each model.The results show that:(1)The characteristics of the random forest algorithm based on recursive feature elimination are better than those of the random forest algorithm.The accuracy of the model constructed by the former is better than that of the latter,and the prediction accuracy of the default of the borrower is also improved.(2)The GBDT algorithm can distinguish different features and feature combinations in the process of feature selection.Meanwhile,the feature outputs of GBDT algorithm can be directly input into the logistic regression model,eliminating the step of manual feature selection.(3)The prediction accuracy of the GBDT+LR fusion model is higher than other models,and has significant advantages in effectively identifying "default users".Finally,the borrower default prediction model constructed in this thesis provides theoretical support for the current domestic P2P network lending platform to solve high default rate and high bad debt rate;it has certain reference value for the risk control of P2P network lending industry.