Credit Risk Model Stability Project

This project develops a predictive model for credit risk assessment, specifically designed to evaluate loan applicants with limited credit history. The model balances predictive performance with temporal stability, making it suitable for real-world financial applications.

Overview

The project addresses the challenge of financial inclusion by:

• Developing robust credit risk assessment models for individuals with limited credit history
• Maintaining model stability over time to reduce retraining costs
• Implementing advanced feature engineering and machine learning techniques
• Providing an interactive interface for real-time predictions

Features

• Advanced data preprocessing using Polars for efficient data handling
• Comprehensive feature engineering pipeline including temporal and statistical aggregation
• Implementation of state-of-the-art machine learning models (LightGBM and CatBoost)
• Interactive visualization dashboard built with Streamlit
• Model stability monitoring using Population Stability Index (PSI)

Technical Architecture

• Data Preprocessing: Polars for high-performance data manipulation
• Model Training: LightGBM/CatBoost for gradient boosting
• Interface: Streamlit for interactive user interface
• Visualization: Plotly for dynamic data visualization
• Model Serialization: Joblib for efficient model storage

Performance

• Achieved AUC score of 0.86 in validation
• Maintained model stability with only 0.9% AUC decay over time
• Superior performance compared to traditional logistic regression methods
• Effective handling of imbalanced datasets

Usage

Data Preprocessing and Model Training

python wrangling_all_1202.py

Running the Demo Interface

python streamlit_final.py

Results

• Successfully handles high-dimensional, temporal, and hierarchical data
• Implements robust feature engineering and selection
• Provides interpretable predictions through SHAP values
• Maintains model stability for long-term deployment

Future Improvements

• Implementation of dynamic updates via online learning
• Enhanced explainability through SHAP-based visualizations
• Expansion to new regions and demographics
• Integration of real-time data streams

Contributors

• Qianran Wu - Electrical Engineering and Computer Sciences, Columbia University
• Pengyu Tao - Statistics, Columbia University

License

This project is under no license.

Acknowledgments

Based on the Kaggle competition "Home Credit - Credit Risk Model Stability"