A Boosted Decision Tree Model for Predicting Loan Default in P2P Lending Communities

Abstract

Loan Default Prediction For Social Lending Is An Emerging Area Of Research In Predictive Analytics. The Need For Large Amount Of Data And Few Available Studies In The Current Loan Default Prediction Models For Social Lending Suggest That Other Viable And Easily Implementable Models Should Be Investigated And Developed. In View Of This, This Study Developed A Data Mining Model For Predicting Loan Default Among Social Lending Patrons, Specifically The Small Business Owners, Using Boosted Decision Tree Model. The United States Small Business Administration (Usba) Publicly-Available Loan Administration Dataset Of 27 Features And 899164 Data Instances Was Used In 80:20 Ratios For The Training And Testing Of The Model. 16 Data Features Were Finally Used As Predictors After Data Cleaning And Feature Engineering. The Gradient Boosting Decision Tree Classifier Recorded 99% Accuracy Compared To The Basic Decision Tree Classifier Of 98%. The Model Is Further Evaluated With (A) Receiver Operating Characteristics (Roc) And Area Under Curve (Auc), (B) Cumulative Accuracy Profile (Cap), And (C) Cumulative Accuracy Profile (Cap) Under Auc. Each Of These Model Performance Evaluation Metrics, Especially Roc-Auc, Showed The Relationship Between The True Positives And False Positives That Implies The Model Is A Good Fit.