Research On The Risk Assessment Of Bank Loans

The credit has always been an important subject for the bank,which determines its survival and development.The risk control ability is increasingly becoming the invisible threshold of Internet finance i ndustry.T herefore,i t i s v ery i mportant to provide significant risk assessment supports in risk control in p ractice.The encountered data often involves the unbalanced classified data and a large number of count data,also the distribution of these two kinds of data are not in equilibrium state,the study of these two kinds of data has become an important issue of statistics.This paper takes the credit card information of bank customers as the case data,and the customer information is analyzed at the first s tep.Then in order to identify the loan applicants who will default or bad loans in the future,an in-depth analysis of the data is investigated from two aspects:(1)The response variables are unbalanced binary variable,and the traditional twocategory machine learning method ignores the correct identification o f t he few classes that people care about.In this paper,the combination method of sampling technology and integration method,SMOTE integration classification and EUS integration classification,are proposed to improve the classification accuracy of a few classes and reduce the risk of errors.SMOTE and EUS respectively in this paper the data sampling technology combined with integrated classification algorithm to establish the model of comparative model to find t he optimal model.The improvements in this paper,in order to make the final classification decision,adopt the method of adaptive threshold selection to find the optimal threshold for achieving AUC.Fitting results show that this method can effectively reduce the imbalance in the fault of the minority class.(2)In order to further provide the classification support,the number of late payments is added as the new dependent variable,and these variable is the nonnegative integer with a large amount of zero,which is zero-inflation count data.For this kind of data,the traditional statistical method of fitting count data will results in biased statistical inference,so the zero-inflation count data models and the Hurdle model are explored in this paper.After selecting the optimal model,the variables were selected to find the important variables affecting the overdue repayment.Based on the analysis results of the above two aspects,risk assessment support for the risk control of credit management can be provided.