Predicting Loan Default Among Borrowers Using Bootstrapped Digital Lending Platforms in Metro Manila, Philippines: A Survey-Based Gradient Boosting Approach

Abstract

This study developed a predictive model of loan default among 512 borrowers using bootstrapped digital lending platforms in Metro Manila, Philippines. Data were collected using a validated survey instrument via quota-controlled stratified random sampling. Quotas were allocated based on the bootstrapped digital lending platforms respondents used. CatBoost was used to predict loan repayment outcomes from self-reported survey data. Model evaluation employed a 70/30 training-test split and 5-fold cross-validation on the full dataset. Feature importance analysis and SHapley Additive exPlanations (SHAP) were used to interpret the model's predictions. Loan characteristics accounted for 90.81% of the model's reported feature importance, with loan amount alone accounting for 82.85%. SHAP analysis further indicated that higher loan amounts were generally associated with higher predicted default risk. Behavioral traits and financial capacity indicators accounted for 2.91% and 6.28% of the reported feature importance, respectively. These findings were leveraged in the development of credit risk assessment practices among bootstrapped digital lending platforms.