| The strong development of Internet finance has profoundly changed the way people consume.As an innovative Internet financial product,P2 P network lending is widely recognized by the public for its convenient borrowing,small transaction volume,fast transaction,and meeting the diversified needs of different customers.Online lending has thus gained rapid development in China.However,compared with the western developed countries,China’s P2 P network lending is limited by the late sprouting,short development time,imperfect credit system,and unsound relevant laws.The incident of defaulting bad debts is not uncommon,which makes investors face serious financial security risks.Avoiding the risk of borrowing and improving credit security has become a bottleneck in the development of the P2 P network lending industry.Therefore,how to establish a safe and reliable credit risk assessment system and evaluation model for borrowers to accurately identify default users has very important practical significance.This paper uses the historical transaction data of Renrendai to conduct a credit evaluation study of P2 P network borrowing borrowers.The random forest classification model is mainly used,and the logistic regression and K nearest neighbor classification models are established.By comprehensively comparing the evaluation effects of each model,we aim to find a model with higher classification accuracy and excavate important features that affect the default.Trying to provide useful information for the current P2 P network lending market in China to better avoid borrowing risks and realize the healthy development of the online lending market.As an integrated algorithm,random forest has better classification accuracy than traditional classification model,and can effectively avoid over-fitting,tolerate noise,good model stability and high operational efficiency,so it is more suitable for the evaluation of credit risk.In this paper,we first establish a random forest classification model.In the modeling process,through the calculation and sorting of the importance of explanatory variables,we found that the seven variables of annual interest rate,total borrowings,borrowing period,working hours,education,number of applications for borrowing,and company size have a significant impact on predicting whether users have defaulted.Subsequent elimination of the less important variables and re-modeling,the simplification of the model and theimprovement of the model accuracy are realized.By optimizing the two important parameters of the number of variables selected in the decision tree branch node of the random forest and the number of trees in the forest,we obtain a random forest model with high classification accuracy and compare it with the logistic regression and the K nearest neighbors.Based on the classification accuracy,precision,recall rate,model stability and AUC value,the performance of the three classification models in classification prediction was evaluated.We found that random forests performed best in classification accuracy,recall rate and AUC value,and had the highest recognition accuracy for default customers and the best overall performance.At the end of the article,we summarize the conclusions obtained from the empirical evidence,and combine the issues that should be noticed in the actual credit evaluation,and look forward to the improvement of the research. |
