Credit Card Default Prediction Model

A machine learning project to predict credit card defaulters using various models and evaluation strategies. Focuses on domain-specific metrics like Recall (Class 1) and Precision (Class 0), beyond just accuracy.

Project Structure

• eda/: Exploratory data analysis notebook
• data/: Data loading utilities
• preprocessing/: Feature encoding, SMOTE, and other transformations
• modeling/: Model training and tuning scripts
• evaluation/: Evaluation and metric plotting

Models Used

• Logistic Regression
• Random Forest
• XGBoost
• Hyperparameter tuning using:
  • Accuracy (default)
  • Domain-specific custom score

Evaluation Focus

• ROC AUC Score (model's ability to distinguish between the 2 classes: default/non-default)
• Recall of Class 1 (catching defaulters)
• Precision of Class 0 (preserving creditworthy leads)

Key Learnings

• Accuracy can be misleading in imbalanced datasets.
• Business context determines the right evaluation metric.
• Tuning of parameters has to be aligned with the right evaluation metric. In this case, custom metric has been used during tuning, with focus on AUC, Recall (1), and Precision (0).

How to Run

# Create environment (optional)
pip install -r requirements.txt

# Run full pipeline
python main.py