isharane/Credit-Risk-Loan-Default-Prediction

Credit Risk Prediction using Ensemble Learning

This project aims to predict the probability of credit default using a combination of Logistic Regression, Random Forest, and XGBoost classifiers combined in a soft voting ensemble. The dataset contains financial, demographic, and credit history attributes of loan applicants. The objective is to identify high-risk applicants and support better lending decisions.

Dataset

Source: Give Me Some Credit.csv Target variable: SeriousDlqin2yrs

• 0 → No serious delinquency in the past 2 years
• 1 → Serious delinquency occurred

Key Features:

• RevolvingUtilizationOfUnsecuredLines – Ratio of credit card balance to credit limit
• NumberOfTime30-59DaysPastDueNotWorse – Count of 30-59 days late payments
• age – Age of the applicant
• NumberOfTimes90DaysLate – Count of 90+ days late payments
• DebtRatio – Monthly debt payments to income ratio
• MonthlyIncome – Monthly income of the applicant
• NumberOfOpenCreditLinesAndLoans – Number of open credit lines and loans
• ...and other relevant credit-related attributes.

Data Preprocessing

• Handling Missing Values: Median imputation for numerical columns.

• Outlier Treatment: Clipping based on IQR method.

• Feature Engineering:

  • DebtToIncomeRatio = RevolvingUtilizationOfUnsecuredLines / MonthlyIncome
  • Age binning into categories (<30, 30-40, 40-50, 50-60, 60+)

• Scaling: StandardScaler applied to numerical features.

• Class Imbalance Handling: Stratified train-test split to maintain target distribution.

Model Training

Three base models were trained and tuned using GridSearchCV with ROC-AUC as the scoring metric:

1. Logistic Regression (with L1 regularization)
2. Random Forest Classifier
3. XGBoost Classifier

These were combined into a VotingClassifier with voting="soft" to leverage predicted probabilities from all models.

Model Performance

Metric	Class 0	Class 1
Precision	0.94	0.60
Recall	0.99	0.18
F1-Score	0.97	0.28
Accuracy	0.94
ROC-AUC Score	0.864

Macro Avg F1: 0.62 Weighted Avg F1: 0.92

Feature Importance (SHAP Analysis)

The top features influencing the model’s predictions are:

1. RevolvingUtilizationOfUnsecuredLines
2. NumberOfTime30-59DaysPastDueNotWorse
3. Age
4. NumberOfTimes90DaysLate
5. NumberOfOpenCreditLinesAndLoans
6. DebtToIncomeRatio

📌 Conclusion

• The ensemble approach achieved a ROC-AUC of ~0.864, showing strong discriminatory power.
• SHAP analysis highlighted that credit utilization, late payment history, and age are key determinants of credit risk.