Application Research Of Internet Finance Credit Risk Rating Model Based On LightGBM Algorithm

With the coming of Big Data Era,the financial industry relying on the continuous development and increasingly perfect Internet platform,has been rapid development and the country attaches great importance to.Among them,risk management control(hereinafter referred to as "risk control"),as the core competitiveness of the financial industry,its key lies in the pricing ability of credit risk,which is mainly reflected in the selection of customers,that is,the evaluation of customer credit risk.No matter how good the loan customers are,there is no risk-free investment,especially for the Internet Finance with high default risk.In case of borrower's default,in addition to initiating necessary safeguard measures and recovering the debt through more effective collection methods and legal means,the platform also needs to analyze the reasons for overdue loan,verify customer's credit qualification from more dimensions,comprehensively evaluate customer's real repayment ability and willingness,and improve the accuracy of personal credit rating,so as to reduce the probability of risk occurrence from the root.Therefore,to do a good job in credit risk rating is the key to ensure the healthy development of the Internet financial industry.At present,in the Internet financial risk control industry,the most widely used credit rating is based on the "Score-Card" model.It is a modern mathematical statistical model with Logistic Regression(LR)as meta algorithm.Its principle is mainly to discover the credit performance hidden in the complex data that can reflect the risk of the applicant by mining,analyzing and summarizing the application information,credit history,business behavior and other data submitted by the applicant.Finally,it is presented in a numerical way as the decision basis of risk control management.Similar statistical algorithms include Decision Tree(DT),K-Nearest Neighbor(KNN),K-means,etc.Other algorithms include Support Vector Machine(SVM),Random Forest(RF),Back Propagation Neural Network(BPNN)and other non statistical machine learning algorithms,which are also widely used in different business of credit rating.Because it can quantify the risk level,has the characteristics of decision-making consistency,fast training and development,high stability,strong business interpretation(ONLY for the statistical method model),easy deployment and so on,it is very suitable for the traditional retail credit business products.However,with the continuous development of big data technology and the overall change of the industry and market,the traditional quantitative model represented by "Score-Card" has been unable to meet the accuracy and efficiency requirements of such massive business data and explosive growth data dimensions.Its shortcomings of high data requirements and low prediction accuracy gradually appear.Therefore,this paper will introduce a new algorithm-LightGBM(Light Gradient Boosting Machine).An algorithm model of gradient boosting framework structure with decision tree as the meta model was first proposed by Microsoft team in 2017.Because of its faster training effect,lower memory utilization rate,higher accuracy,and can support parallel learning mode,it has the ability to process larger data and other characteristics,and has made good response and good performance in the industry.At the same time,these characteristics are also very consistent with the characteristics of data in the current Internet financial credit business.Based on the research on the theory and method of personal credit evaluation at home and abroad,combined with the personal credit evaluation indexes of banks and private financial institutions in China,this paper constructs a personal credit rating model,and studies the key issues involved in the model.Including the exploratory derivation of credit business characteristics and the optimization of model parameters.Using supervised learning algorithm(LightGBM,BPNN,RF,LR)to build the benchmark model.Based on the empirical analysis of the financial credit business of the company and the real business data set of the two platforms of the Kaggle website,to verify:(1)the LightGBM algorithm can improve the prediction accuracy of the credit rating model,and the improvement of the accuracy mainly comes from the improvement of the prediction accuracy of default customers.(2)Based on the horizontal comparison of risk rating models constructed by different algorithms,there are statistical differences between the prediction results,among which LightGBM credit risk rating model has the best prediction effect,and the construction and introduction of this model has practical significance in the field of Internet financial credit risk rating.