Loan Default Prediction with Machine Learning

A comprehensive machine learning project for predicting loan defaults using LendingTree data. This project demonstrates advanced data preprocessing, feature engineering, and model optimization techniques on a large-scale dataset of 2.2 million records.

Overview

This project involved working with a large dataset of 2.2 million records and 150 columns. Through preprocessing, the feature set was reduced by approximately 65%, significantly optimizing the data without sacrificing model performance.

🛠️ Technologies Used

Core Technologies

• Python 3.12+ - Primary programming language
• Jupyter Notebooks - Interactive development and analysis
• Git - Version control and project management

Data Science & Machine Learning

• pandas - Data manipulation and analysis
• NumPy - Numerical computing and array operations
• scikit-learn - Machine learning algorithms and preprocessing
• XGBoost - Gradient boosting framework for classification
• scipy - Statistical analysis and scientific computing

Data Visualization & Analysis

• matplotlib - Basic plotting and visualization
• seaborn - Statistical data visualization
• Point-biserial correlation - Feature-target relationship analysis

Testing & Quality Assurance

• pytest - Testing framework and test automation
• pytest-cov - Coverage reporting
• Comprehensive test suite - 44 tests covering model performance, data validation, and edge cases

📊 Machine Learning Techniques

Data Preprocessing Pipeline

• Feature Reduction - Reduced 150 features by ~65% through correlation analysis and feature selection
• Missing Value Imputation - Strategic handling of missing data with domain-specific approaches
• Outlier Treatment - IQR-based outlier capping to improve model robustness
• One-Hot Encoding - Categorical variable transformation for model compatibility
• Standard Scaling - Feature normalization for improved model performance

Feature Engineering

• Derived Variables - Created debt-to-income ratios and payment-to-income ratios
• Interest Rate Binning - Risk-based grouping using quantile discretization
• Missing Value Indicators - Binary flags for systematic missing patterns
• Correlation Analysis - Point-biserial correlation for feature-target relationships

Class Imbalance Handling

• Scale Position Weight - XGBoost native class imbalance adjustment
• Stratified Sampling - Maintained class distribution in train/test splits
• Performance Metrics - ROC AUC prioritized over accuracy for imbalanced data

Model Selection & Performance

• LazyPredict - Initially attempted but failed due to memory overload with large dataset and unlimited tree depth
• Random Forest - Tested but yielded negligible results for this classification task
• XGBoost (Extreme Gradient Boosting) - Selected as optimal model achieving 72.78% ROC AUC
  • Native class imbalance handling eliminates need for SMOTE
  • Efficient memory usage for large datasets
  • Superior performance on tabular data with mixed feature types

🔍 Model Performance

Final Results

• ROC AUC Score: 72.78%
• Accuracy: 64.15%
• Dataset Size: 2.2M records processed
• Feature Reduction: 65% reduction from 150 to ~53 features
• Class Distribution: 87% good loans, 13% defaults (handled via scale_pos_weight)

Key Findings

The results aligned well with real-world predictors of default risk:

🎯 Strong Predictive Features

• Payment Method - Direct pay demonstrates lower default risk
• Verification Status - Income verification strongly correlates with repayment
• Account Activity - Number of accounts opened in last 24 months indicates credit behavior

📊 Moderate Predictive Features

• Interest Rate - Reflects borrower risk but limited direct predictive power
• Debt-to-Income Ratio - Ranked ~50th percentile in feature importance, less impactful than anticipated
• Credit History Length - Moderate influence on default probability

🚀 Future Enhancements

Model Improvements

• SMOTE Implementation - Apply synthetic oversampling techniques for better class balance handling
• Hyperparameter Tuning - Grid search optimization for XGBoost parameters
• Ensemble Methods - Explore stacking and voting classifiers for performance gains
• Feature Engineering - Create interaction terms and polynomial features

Technical Enhancements

• Model Deployment - API development for real-time predictions
• Feature Store - Centralized feature management and versioning
• MLOps Pipeline - Automated model training, validation, and deployment
• Explainability - SHAP values and LIME for model interpretability

🧪 Testing

This project includes a comprehensive test suite to validate model performance, data quality, and edge cases.

Setup Environment

Create and activate a virtual environment:

# Create virtual environment
python3 -m venv venv

# Activate environment (Linux/Mac)
source venv/bin/activate

# Or use the helper script
source activate_env.sh

# Install dependencies
pip install -r requirements-test.txt

Running Tests

# Run all tests
python -m pytest tests/ -v

# Run specific test categories
python -m pytest tests/test_model_performance.py -v    # Model performance tests
python -m pytest tests/test_data_validation.py -v     # Data validation tests  
python -m pytest tests/test_edge_cases.py -v          # Edge case tests

# Run with coverage report
python -m pytest tests/ --cov=tests --cov-report=html

# Run tests and generate detailed output
python -m pytest tests/ -v --tb=long

Test Categories

Model Performance Tests (12 tests)

• Accuracy, precision, recall, F1-score validation
• ROC AUC score validation (≥0.5 threshold)
• Prediction probability ranges and consistency
• Feature importance validation
• Class imbalance handling

Data Validation Tests (18 tests)

• Schema validation (required columns, data types)
• Value range validation (positive amounts, valid rates)
• Missing value thresholds (<50%)
• Duplicate detection (<5%)
• Outlier detection and correlation analysis
• Class distribution validation

Edge Case Tests (16 tests)

• Empty data handling
• Extreme value inputs (zeros, ones, very large/small numbers)
• Single row predictions and boundary conditions
• High-dimensional data and class imbalance scenarios
• Mixed data types and numerical precision edge cases

Test Results

All 44 tests pass, providing confidence in:

• Model reliability and performance thresholds
• Data quality and preprocessing pipeline
• Robustness against edge cases and unexpected inputs

Deactivate Environment

deactivate

---

This project provided valuable insights into handling large datasets and reinforced the importance of method selection and preprocessing in machine learning pipelines.