Final Project for Applying Machine Learning – Spring 2025
Columbia University, New York City, USA

Credit ratings influence billions in capital flows — but do raters behave differently in times of crisis?
This project investigates whether human credit rating agencies add value or diverge from machine learning-based predictions during periods of financial instability. Using XGBoost and real-world corporate ratings data, we explore:

• 📉 Prediction accuracy during crisis vs. normal periods
• 🤖 Model-human disagreements and their underlying drivers
• 📊 Feature analysis of override patterns
• 🔍 Behavioral insights into rating decisions across time

• ml_final_project.ipynb – the full analysis notebook
• corporate_ratings.csv – cleaned dataset used in modeling

• XGBoost classifier with hyperparameter tuning
• Cross-validation with custom metrics
• Confusion matrix + risk-adjusted scoring
• Precision-Recall and ROC analysis
• Time series resampling and crisis segmentation
• Visualization with matplotlib & seaborn

• Michelle Ren
• Louis Sellier

• Columbia, Professor Björkegren, and TAs for guidance
• Original dataset and prior academic literature on credit risk modeling
• XGBoost and scikit-learn libraries