This repository contains a comprehensive Jupyter notebook (CreditRiskEDA_LoanApproval.ipynb) and accompanying visualizations (in /images) that demonstrate end-to-end credit risk modeling for loan approval.
- Records: 51,336 loan applications
- Features: 87 attributes including demographics, financial metrics, credit history, and trade line details
- Target:
Approved_Flagwith four risk categories (P1–P4) - Class Imbalance: P2 (62.7%), P3 (14.5%), P4 (11.5%), P1 (11.3%)
- Visualized class distribution and feature histograms
- Identified demographic and credit-behavior patterns
- Summary: Credit score and age clearly stratify risk; product enquiry behavior differs by category
- Applied ANOVA F-tests to rank 81 numeric features
- Selected top 20 features (e.g.,
Credit_Score,enq_L3m,Age_Oldest_TL) - Encoded 19 categorical features via one-hot encoding
- Summary: Reduced feature set by 55% while preserving predictive power; 42 common features ensure deployment compatibility
- Trained and compared 8 algorithms (Logistic Regression, Random Forest, XGBoost, CatBoost, etc.)
- Addressed class imbalance with sample weights and SMOTE-Tomek
- Implemented ensemble methods: soft voting, weighted voting, stacking
- Generated confusion matrices for all models
- Evaluated class-wise performance for each risk category
- Summary: Stacking ensemble shows the best balance across P1–P4 classes
- XGBoost (Class Weights): F1 (Macro)=70.59%, Balanced Acc=76.21%
- Logistic Regression (SMOTE-Tomek): F1 (Macro)=67.22%, Balanced Acc=72.81%
- Stacking Ensemble: F1 (Macro)=70.69%, Balanced Acc=69.23%
- Hybrid Ensemble (Stacking + Weighted Voting): F1 (Macro)=70.06%, Balanced Acc=68.54%
Conclusion: The stacking classifier remains the optimal model, offering the best macro F1 and balanced accuracy for credit risk prediction.