Predicting Loan Approval with LASSO Logistic Regression

Abstract

We build a predictive workflow for loan approval using a Kaggle-sourced synthetic dataset. After cleaning and exploratory analysis, we fit a LASSO logistic regression with 5-fold cross-validation to select informative features and reduce overfitting risk. The final model achieves test AUC ≈ 0.935, and we select an operating threshold to minimize false positives while keeping TPR ≥ 0.60 (FPR ≈ 0.053), balancing risk control with access to credit.

Methodology

1. EDA & Interactions — age/gender weak; interest rate & previous defaults strong; notable age × defaults、amount × intent.
2. Preprocessing — standardize numerics, one-hot categoricals, remove age outliers (>100).
3. Model — LASSO (liblinear), λ grid with 5-fold CV; metric: ROC-AUC; check multicollinearity.
4. Thresholding — pick threshold with TPR > 0.60 minimizing FPR (~0.053 at ~0.597).

Key Results

• AUC (test): ~0.935; robust generalization.
• Top signals: previous default (−), interest rate (+), loan amount (+); intent varies by category.
• K-Means (k=3): segments align with risk profiles; one group shows higher approvals despite lower income but moderate amounts/high rates.

Discussion & Limitations

Synthetic data limits external validity; future work: add DTI, LTV, utilization, compare forward/backward CV selection, test higher-order interactions.

Downloads

• 📄 Paper: STAT207 Final Project
• 🖼️ Poster: STAT207 Presentation