Credit Risk Modeling

Overview

This project builds a credit default prediction system that estimates the probability of default for customers using historical repayment and financial behavior.

The focus is not just classification, but:

• Ranking users by risk
• Maximizing recall of defaulters
• Producing calibrated probability estimates for decision-making

---

Problem Statement

Given customer credit history, predict the likelihood of default in the next month.

This is a risk-sensitive problem where:

• Missing a defaulter → financial loss
• False positives → operational cost

Therefore, the system prioritizes:

• High recall for defaulters
• Efficient identification of high-risk users

---

Dataset

The dataset contains:

• PAY_X: Repayment status (delay behavior)
• BILL_AMT_X: Monthly bill amounts (debt)
• PAY_AMT_X: Payments made

Target:

• default (1 = default, 0 = non-default)

The dataset is imbalanced, making accuracy an unreliable metric.

---

Approach

1. Feature Engineering

Constructed behavior-driven features:

• recent_delay → recent repayment severity
• delay_trend → worsening payment behavior
• utilization → credit usage ratio
• pay_ratio → payment discipline

---

2. Models Used

• Logistic Regression (baseline, interpretable)
• XGBoost (final model, non-linear patterns)

3. Threshold Optimization

Instead of default 0.5 threshold:

• Lower threshold → higher recall
• Selected threshold based on business trade-off

---

4. Model Evaluation

Focused on:

• Recall (defaulters)
• Precision
• Confusion matrix

Accuracy was not used as a primary metric.

---

Key Results

• Top 10% users capture ~68% of total defaulters
• Recall improved to ~80%
• XGBoost achieved higher precision at similar recall compared to logistic regression
• Model effectively ranks users by risk

---

Calibration

Initial model showed overconfidence in probability estimates.

Applied isotonic calibration:

• Improved probability reliability
• Did not affect ranking performance

---

Key Insights

• Repayment delay is the strongest predictor of default
• Behavioral signals outperform static financial metrics
• Model is fragile without key delay features (validated via ablation)
• Ranking users is more valuable than binary classification

---

Business Impact

The model enables:

• Prioritization of high-risk customers
• Efficient allocation of risk management resources
• Estimation of expected default rates

Example:

• Top 10% high-risk users → 68% of defaulters
  → Significant improvement over random targeting

---

Tech Stack

• Python
• Pandas, NumPy
• Scikit-learn
• XGBoost
• Matplotlib / Seaborn

---

Project Structure

├── notebook.ipynb       # Full analysis and modeling
├── credit_model.pkl     # Saved calibrated model
├── README.md
├── requirements.txt

---

How to Run

git clone <repo_url>
cd <repo>
pip install -r requirements.txt
jupyter notebook

Future Improvements

• Add more temporal features
• Improve robustness beyond single dominant feature
• Deploy as real-time scoring API
• Integrate with live financial data

Conclusion

This project demonstrates a real-world credit risk modeling pipeline, including:

• Feature engineering from behavioral data
• Model comparison and threshold tuning
• Risk-based ranking
• Probability calibration

The system functions as an early warning mechanism for identifying high-risk customers.