Research On CPLE_LightGBM Personal Credit Scoring Model Of Online P2P Lending Based On Reject Inference

Online P2P lending industry is an important part of financial lending.The growth rate was as high as 300% in 2014 and 2015,and returned to reasonable growth rate in 2016 and 2017.But the growth momentum is still strong.In 2017,the total turnover of the online P2P lending industry reached 2.8 trillion.In the process of lending,the number of investors and borrowers reached 17 million and 22 million respectively,which means the online lending mode is popular with a great number of people.The capital cost of P2P lending market determines that it belongs to the secondary market,where the default risk is high.And the loan relationship in P2P online lending can be attributed to personal credit.Therefore,it is urgent and important to carry out personal credit scoring effectively to protect the interests of investors.At present,the construction of credit scoring model generally relies on accepted samples only.In fact,the number of rejected samples in online P2P lending often accounts for 90% of total loan application records each year.When data missed not at random,using accepted samples but abandoning rejected samples builds credit scoring model,which will lead to sample selection bias and reduce the forecast accuracy of credit scoring model.Therefore,it is necessary to use the method of reject inference to integrate the rejected samples into the dataset that will be used for constructing the personal credit scoring model.With different statistical inference algorithms(like extrapolation and augmentation)and semi-supervised learning algorithms(like CPLE_LightGBM,S3 VM,CPLE_RF,CPLE_LR),predicting potential repayment results of rejected samples,to obtain the whole sample dataset.Contrast models were built with supervised learning algorithms(like LightGBM,SVM,RF,LR).Empirical analysis was conducted based on datasets got from renrendai and lendingclub,both of which are P2P lending platforms.The empirical results show that:Firstly,the prediction accuracy of the personal credit scoring model can be improved with the reject inference,and the improvement derived from the better prediction accuracy about bad customers.secondly,in a statistical sense,the personal credit scoring models based on different algorithms do have different prediction performances.Among them,the prediction performance of that personal credit reject inference model of online P2P lending based on CPLE_LightGBM is optimal,so the construction and introduction of this model do make practical significance.Thirdly,in the case of reject inference,the sample structure made an influence on the prediction performance of personal credit scoring model.