Credit Risk Prediction — End-to-End ML Benchmark

Real-world machine learning project focused on loan default prediction, model benchmarking, explainability, and cost-sensitive decision making using large-scale lending data.

Built as a research-style pipeline aligned with government R&D and industry data science workflows.

---

Run the Notebook (Colab)

Open and run instantly in browser: https://colab.research.google.com/drive/1mMedWu2dLTOzfZq0OKNsfTpcge6vBF5W?usp=sharing

---

Project Structure

credit-risk-ml-benchmark/
│
│  Credit Risk Prediction Report.pdf
│  credit risk prediction.ipynb
│  requirements.txt
│
└── images/
    corr.png
    Credit History vs Default.png
    DTI Vs Default.png
    Interest_Rate_vs_Default.png
    Loan Amount vs Default.png
    ROC Curve Comparision.png
    ROC LOG.png
    SHAP.png
    Target Default Value Counts.png

---

Executive Summary

Goal: Predict probability of loan default using real lending data and evaluate models from both statistical and business perspectives.

Dataset:

• 270k+ loan records
• 73 engineered features
• Binary target: Default vs Fully Paid

Models compared:

• Logistic Regression
• Decision Tree
• Random Forest
• XGBoost

Best ROC-AUC: 0.75 (XGBoost) Best Business Model (Cost-Sensitive): Logistic Regression

Key takeaway:

The statistically best model is not always the best business model.

---

Problem Statement

Financial institutions must estimate the probability that a borrower will default.

Cost asymmetry:

• Missing a defaulter → very expensive
• Rejecting a safe borrower → less expensive

This project answers:

• Which model predicts default best?
• Which model minimizes financial risk?
• Which features drive predictions?
• How do we move from ML metrics → real decisions?

---

Dataset Overview

After cleaning and feature engineering:

• Rows: 270,385

• Features: 73

• Class distribution:

  • Fully Paid → 79.8%
  • Default → 20.2%

Realistic class imbalance → real industry scenario.

---

Exploratory Data Analysis

Target Distribution

Interest Rate vs Default

Loan Amount vs Default

Debt-to-Income vs Default

Credit History vs Default

Correlation Overview

Key insights:

• Higher interest → higher risk
• Short credit history → higher default probability
• Higher DTI → increased risk
• Longer loan term → increased risk

---

Model Benchmarking

Model	ROC-AUC
XGBoost	0.750
Logistic Regression	0.743
Random Forest	0.734
Decision Tree	0.729

ROC Curve Comparison

---

Threshold Tuning (Real-World Evaluation)

Models evaluated using a risk-aware threshold instead of default 0.5.

Model	Accuracy	Recall(Default)	Precision(Default)
Logistic Regression	0.68	0.68	0.35
Decision Tree	0.76	0.45	0.41
Random Forest	0.77	0.44	0.43
XGBoost	0.77	0.46	0.44

---

Cost-Sensitive Analysis

Assumption: Missing a defaulter costs 5× more than rejecting a safe borrower.

Result: Logistic Regression minimized total financial loss.

Industry lesson: Accuracy ≠ Best deployment model.

---

Explainable AI

Feature Importance (SHAP)

Top risk drivers:

• Loan grade
• Interest rate
• Loan term
• FICO score
• Credit history length
• Mortgage accounts

Interpretability is essential for:

• Financial regulation
• Risk audits
• Responsible AI

---

Tech Stack

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

---

Run Locally

Clone repo and install dependencies:

pip install -r requirements.txt
jupyter notebook

Open:

credit risk prediction.ipynb

Run all cells to reproduce results.

---

Why This Project Matters

Most ML projects stop at accuracy.

This project goes further:

• Realistic dataset scale
• End-to-end ML pipeline
• Model benchmarking
• Threshold tuning
• Cost-sensitive evaluation
• Explainable AI
• Business-driven model selection

Designed to reflect real industry and government research workflows.