Machine Learning Interpretability Research For Loan Default Prediction

With the rapid development of machine learning,various industries have applied the complex algorithm of machine learning model to judge and make decisions.Although the accuracy of the model is greatly improved,it also brings business application risks.The black-box attribute of the model leads to the lack of clear interpretation of the model results in business applications,and it is difficult for business users to provide criteria for model decision making.In addition,as the financial field has increasingly strict requirements on risk control,regulators have also put forward higher requirements on the stability,security,fairness and interpretability of machine learning models,so we need to be more cautious about the black box model and its potential risks.As an important part of the financial field,how to predict the default probability according to the pre-loan information of users and explain the corresponding results to users at the same time is an urgent problem to be solved,so as to bring trust to users as much as possible on the premise of guaranteeing their own risk control.The data in this paper are from 110,000 real user loan information on Lending Club.XGBoost model is used to predict loan default and compared with the results of various models.The AUC value of XGBoost model after parameter optimization is 0.70 and KS value is 0.32,both of which are better than other models.After that,feature importance,partial dependence diagram,LIME(Local Interpretable Model-Agnostic Explanations)and SHAP(Shapley Additive Explanation)are used explanation A variety of explanation methods are used to explain the default prediction results,and the most influential factors are identified,including the application month of the lender,the time interval between the opening date of the latest installment account and the loan application date,the monthly repayment amount,debt-to-income ratio,the ratio of total annual repayment to annual income,etc.This paper summarizes the characteristics and application scope of different interpretation methods and gives the thinking and steps of choosing interpretation methods under different business requirements.Finally,recommendations are made for financial institutions so that the model can balance accuracy and interpretability.The possible innovation of this paper lies in that most current researches on machine learning focus on model accuracy.In this paper,model interpretation is taken as the research focus on the premise of ensuring accuracy,which is applied to the actual problem of loan default prediction,providing a practical basis for the research on machine learning interpretability.