P2P Pre-lending Anti-fraud Risk Control System Construction

In recent years,with the rise of Internet finance,its derivatives P2P has become popular,and the P2P platform has sprung up.No matter the number of platforms or the transaction volume has increased significantly,the P2P transaction volume in 2016 has reached 1,495.51 billion yuan.The development of P2P meets the needs of a large number of individual consumers for microfinance.However,the current P2P market situation is complex,the qualifications of borrowing users are uneven,and there are problems such as adverse selection and information asymmetry between the platform and the borrower,resulting in the P2P platform.The bad debt rate is still high,and the P2P platform's yield is also decreasing year by year.In order to reduce the default rate of users and reduce the operational risk of the platform,it is very important to establish a reliable pre-lending anti-fraud risk control system.This paper first uses literature analysis method to study a batch of literatures and finds that the domestic credit information system for P2P industry is not perfect.Most of the user's screening still relies on subjective judgment.This paper needs to establish an objective and quantifiable risk control system.The model with good classification results continues to consult the literature,and finally determines three models of GBDT,Xgboost and random forest.It is also mentioned in the literature that certain characteristics of users are very helpful for the classification of users,and provide ideas for modeling work.Then,using the big data modeling method,this paper obtained a total of 609 users' complete data.The user's data is first quantified and 33 features are extracted.The strong rule is then extracted from the feature,and the user's application is rejected directly as long as the user's feature triggers a strong rule.In the next step,the ? values of 33 features are calculated separately.If the IV value is too small,it indicates that the feature is not helpful to the screening users.The deletion of the feature does not enter the modeling process,and finally 32 effective features are obtained.Finally,the above three models were trained with data.The ROC curve and the confusion matrix were used to compare the prediction results of the three models.It was found that the three models had their own merits.Therefore,the three models were combined by voting method,and the combined total model accuracy was obtained.Further improvement.The above two parts constitute the risk control engine of this paper.The specific process is as follows:after the user fills in the information,the judgment of the strong rule engine is first passed.If the strong rule is triggered,it is immediately rejected,and then judged by the wind control model.All of them give money to the user.