Research On The Method Of Credit Risk Control For Financial Institutions

With the increasing loss of credit risk,it is very important to establish a set of scientific and effective credit risk control scheme for major financial institutions in China.In order to improve work efficiency,many organizations entrust some tedious business processes to machine learning model,but every incorrect prediction may bring serious consequences.At present,the credit risk control scheme based on machine learning has many problems,such as poor explanation and imperfect discrimination index.To solve these problems,this thesis mainly uses multi-source data information to complete credit risk control model and scorecard construction.The main work of this thesis is as follows:In order to meet the requirements of the credit business regulatory department on the interpretability of the final online model,this thesis uses the mainstream logistic regression model to establish the credit risk control model.However,as a generalized linear model,it has the problems of limited expression ability and poor prediction effect.Therefore,this thesis proposes an under sampling algorithm FenbuEasyEnsemble(FEE)based on the improved EasyEnsemble method,which balances the data set by removing noise samples,subset division and set filling.Then,a multi-stage hybrid model XGBOOST_FenbuEasyEnsemble_Logistic Regression(XGB_FEE_LR)based on balanced data set and combined features is constructed.Give full play to the great advantage of good interpretation of logistic regression model,at the same time,make up for the disadvantage of poor prediction effect from two aspects of data preprocessing and feature extraction.In this thesis,three different datasets are used to measure the performance of different models from AUC and G_means.The first mock exam shows that XGB_FEE_LR model performs better than other single models and mixed models in terms of classification effect,and has certain innovative and practical significance.Secondly,the traditional credit risk scorecard only considers the user's own information such as identity,assets,past performance,etc.,and ignores the connections between users,and the score is not complete enough.Therefore,this thesis first dynamically establishes the financial relationship graph of the characters based on the existing multi-source data information.Then on this basis,an influence evaluation algorithm Financial Relations Graph_Anti_Direct_Rank(FRG_ADRank)based on the bidirectional propagation of people's financial relationship graph is proposed to measure the impact of default users on normal users.Finally,this thesis integrates the algorithm into the classic scoring card construction method to build a more comprehensive and effective credit risk scoring card.Compared with the traditional influence evaluation algorithm,the FRG_ADRank algorithm can better reflect the impact of real financial activities.And in this way,financial institutions can find suspicious users who have not shown malicious behavior early,and take measures as soon as possible to make up for the economic losses caused by the lag prevention in the current credit risk control methods.