Credit Risk Inference System

A production-ready machine learning system for credit risk assessment with real-time inference capabilities. Built with Python, XGBoost, Flask, and scikit-learn.

Overview

This system provides automated credit risk scoring for loan applications with the following features:

• ML Pipeline: Scikit-learn Pipeline with StandardScaler, categorical encoding, and XGBoost classifier
• Performance: AUC-ROC > 0.91, Accuracy > 81%
• Low Latency: API responses < 200ms (median)
• Scalability: Handles 1,000+ daily requests with concurrent processing
• REST API: Flask-based API with /predict endpoint

Tech Stack

• Core ML: Python 3.11, XGBoost 2.0.0, scikit-learn 1.3.0
• API Framework: Flask 2.3.3 with CORS support
• Database: PostgreSQL (psycopg2-binary)
• Data Processing: pandas, numpy
• Testing: pytest, requests

Project Structure

CRA/
├── data/
│   ├── generate_data.py       # Synthetic dataset generator
│   └── credit_data.csv         # Generated dataset (1,000 rows)
├── models/
│   ├── credit_risk_model.pkl   # Trained model pipeline
│   └── metrics.json            # Performance metrics
├── app/
│   ├── __init__.py
│   └── api.py                  # Flask REST API
├── tests/
│   ├── test_api.py             # Unit tests
│   └── stress_test.py          # Performance testing (1,000 requests)
├── train_model.py              # Model training script
├── requirements.txt            # Dependencies
└── README.md

Quick Start

1. Setup Environment

# Clone repository (if applicable)
cd CRA

# Create virtual environment
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

2. Generate Data & Train Model

# Generate synthetic credit dataset
python data/generate_data.py

# Train XGBoost model
python train_model.py

Expected Output:

• Dataset: 1,000 samples, 30% default rate
• Test Set AUC-ROC: ~0.91
• Test Set Accuracy: ~0.81

3. Run API Server

# Start Flask API
python app/api.py

The API will be available at http://localhost:5000

4. Test the API

Health Check

curl http://localhost:5000/health

Single Prediction

curl -X POST http://localhost:5000/predict \
  -H "Content-Type: application/json" \
  -d '{
    "age": 45,
    "income": 75000,
    "employment_length": 10,
    "credit_score": 720,
    "num_credit_lines": 5,
    "credit_utilization": 0.3,
    "debt_to_income": 0.35,
    "total_debt": 25000,
    "savings": 15000,
    "loan_amount": 20000,
    "loan_term": 36,
    "num_late_payments": 0,
    "num_bankruptcies": 0,
    "inquiries_last_6m": 1,
    "employment_status": "employed",
    "home_ownership": "mortgage",
    "loan_purpose": "debt_consolidation"
  }'

Response:

{
  "status": "success",
  "prediction": 0,
  "risk_score": 24.56,
  "risk_level": "low",
  "probability": {
    "low_risk": 0.7544,
    "high_risk": 0.2456
  },
  "processing_time_ms": 12.45,
  "timestamp": 1704999999.123
}

5. Run Tests

Unit Tests

pytest tests/test_api.py -v

Stress Test (1,000 Requests)

# Ensure API is running first
python tests/stress_test.py

Expected Performance:

• Median latency: < 200ms
• Throughput: > 100 requests/second
• Success rate: > 99%

Simple API Test

python tests/stress_test.py simple

API Endpoints

GET /health

Health check endpoint

Response:

{
  "status": "healthy",
  "model_loaded": true,
  "timestamp": 1704999999.123
}

POST /predict

Single prediction endpoint

Request Body:

{
  "age": 45,
  "income": 75000,
  "employment_length": 10,
  "credit_score": 720,
  "num_credit_lines": 5,
  "credit_utilization": 0.3,
  "debt_to_income": 0.35,
  "total_debt": 25000,
  "savings": 15000,
  "loan_amount": 20000,
  "loan_term": 36,
  "num_late_payments": 0,
  "num_bankruptcies": 0,
  "inquiries_last_6m": 1,
  "employment_status": "employed",
  "home_ownership": "mortgage",
  "loan_purpose": "debt_consolidation"
}

Response:

{
  "status": "success",
  "prediction": 0,
  "risk_score": 24.56,
  "risk_level": "low",
  "probability": {
    "low_risk": 0.7544,
    "high_risk": 0.2456
  },
  "processing_time_ms": 12.45,
  "timestamp": 1704999999.123
}

POST /predict/batch

Batch prediction endpoint for multiple applications

Request Body:

{
  "applications": [
    { /* application 1 */ },
    { /* application 2 */ }
  ]
}

Features

Dataset Generation

• 1,000 synthetic credit applications
• 17 input features + 2 derived features
• Realistic distributions based on financial data
• 30% default rate with risk-based logic

Feature Engineering

• Numerical Features: age, income, credit_score, debt_to_income, etc.
• Categorical Features: employment_status, home_ownership, loan_purpose
• Derived Features: monthly_payment, payment_to_income ratio

Model Pipeline

1. Preprocessing:

   • StandardScaler for numerical features
   • LabelEncoder for categorical features
   • ColumnTransformer for combined preprocessing

2. Classifier:

   • XGBoost with optimized hyperparameters
   • 100 estimators, max_depth=6, learning_rate=0.1
   • L1/L2 regularization

3. Evaluation Metrics:

   • AUC-ROC: 0.9122
   • Accuracy: 0.8133
   • Precision: 0.7073
   • Recall: 0.6444
   • F1 Score: 0.6744

API Optimization

• Model loaded once at startup (not per request)
• Efficient pandas DataFrame operations
• Response time tracking
• Error handling and validation

Performance Metrics

Model Performance (Test Set)

Metric	Value
Accuracy	81.33%
AUC-ROC	91.22%
Precision	70.73%
Recall	64.44%
F1 Score	67.44%

API Performance (1,000 Requests)

Metric	Target	Actual
Median Latency	< 200ms	~15-50ms*
P95 Latency	< 500ms	~100-200ms*
Throughput	> 50 req/s	> 100 req/s*
Success Rate	> 95%	> 99%

*Actual values depend on hardware

Development

Adding New Features

1. Update data/generate_data.py to include new features
2. Modify train_model.py to handle new features in pipeline
3. Update API input validation in app/api.py
4. Retrain model: python train_model.py

Retraining Model

# Generate new dataset
python data/generate_data.py

# Train model
python train_model.py

# Restart API server
python app/api.py

Running in Production

Using Gunicorn (Recommended)

pip install gunicorn

gunicorn -w 4 -b 0.0.0.0:5000 app.api:app

Using Docker

FROM python:3.11-slim

WORKDIR /app
COPY requirements.txt .
RUN pip install -r requirements.txt

COPY . .

CMD ["gunicorn", "-w", "4", "-b", "0.0.0.0:5000", "app.api:app"]

Future Enhancements

• PostgreSQL integration for data storage
• Model versioning and A/B testing
• Real-time monitoring and alerting
• Feature importance visualization
• Batch processing for large datasets
• SHAP values for model explainability
• API authentication and rate limiting
• Docker containerization
• CI/CD pipeline