Research On Risk Control Of Online Loan Based On Stacking Fusion Model

In recent years,Internet finance has been booming on the back of technological progress and inclusive finance.The combination of "artificial intelligence big data + finance" has made Internet finance heated up with capital for a while.It has also become an urgent need to use the power of "fintech" to escort Internet finance,and pre-loan credit assessment is the most important grasp for risk control.In the research of credit risk control model,the traditional rule-based model has been gradually replaced by machine learning and deep learning model.However,for different data sets,the most suitable algorithm models are often different,so it is difficult to obtain both efficient and stable models based on data set training.In this paper,for the network small loan risk control needs,based on the historical data of a Shanghai network loan company platform,to establish a pre-loan personal credit evaluation model based on the Stacking fusion model,aimed at solving the model distinction and stability is difficult to balance the problem.The main research contents and contributions of this paper are as follow:1.Design automatic feature derivation method and fine feature screening strategy.Aiming at the desensitization of some fields in the data set,the feature derived features were generated in batch by means of the Feature Engineering Library of Feature Tools.The refined feature screening strategy was designed,and the layer-by-layer screening was performed according to the IV value,feature importance and feature stability PSI,which solved the possible problems left over by manual feature mining.2.Build a credit scoring model based on Stacking fusion.Aiming at the problem that the single model is incapable of taking into account both the discrimination and stability,a credit scoring model based on Stacking fusion was established,and Logistic regression,GBDT,XGBoost and Light GBM models were fused.The experimental results show that,the accuracy and stability of the Stacking fusion model are improved compared with that of the single model.3.Model validation of the open dataset.Using the loan data of American public dataset Lending Club company for model validation,it was found that the fusion model of Stacking algorithm also has the role of "learning from each other",which can be trained to get both efficient and stable risk control model,and transforms the model probability into credit scores,providing evidence for loan audit.