This repository contains the full project for the Home Credit Default Risk case study, completed for the MMF2030 Machine Learning for Finance course at the University of Toronto. The goal of this project is to predict the likelihood of a client defaulting on a loan using advanced machine learning techniques.
The project explores various stages of the data science pipeline—from data preprocessing and feature engineering to model building and business analysis—with a strong focus on practical applications in credit risk management.
The objective is to develop a predictive model that classifies clients into default or non-default categories using historical and transactional data. This classification improves loan approval decisions and helps mitigate credit risk for financial institutions.
Challenges Addressed:
- Reducing false positives and false negatives in credit decisions.
- Enhancing financial inclusion for low-risk applicants.
- Balancing risk management with profitability.
File: MMF2030_Project_Group_Component.pdf
Authors: Shuqi (Serena) Chang, Ruoxu (Ryan) Jiang, Xinyi (Cynthia) Shen, Surui (Alex) Zhang
Highlights:
- Data Cleaning & Preprocessing: Handled missing values, scaling, and encoding.
- Feature Engineering: Created new variables from transactional and credit data.
- Modeling: Implemented XGBoost and Random Forest with hyperparameter tuning.
- Evaluation: Used ROC-AUC, F1-Score, and Profit Curve Analysis for performance metrics.
File: MMF2030_Project_Individual_Component_Xinyi_Cynthia_Shen.pdf
Author: Xinyi (Cynthia) Shen
Note: This report was originally based on Alex's code, but Cynthia's version (MMF2030_Project_Code_Cynthia.ipynb) is cleaner, more efficient, and eliminates excessive warnings. While some numerical outputs differ slightly, Cynthia's code produces more stable results.
Highlights:
- Business Insights: Profit optimization through threshold adjustment.
- Profit Curve Analysis: Maximum profit of $35,513 at a threshold of 0.16.
- Model Evaluation:
- ROC AUC: 0.75
- Accuracy: 86%
- Precision (Non-default): 94%
- Recall (Default): 39%
Recommended: MMF2030_Project_Code_Cynthia.ipynb
- Clean Code: Organized and well-documented Python code with minimal warnings.
- Optimized Workflow: Streamlined data preprocessing, feature engineering, and model evaluation.
- Reproducibility: Clear structure for replicating results and running model evaluations.
- Output: Since we wrote the report based on the other version of code run by Alex, the outputnumbers are slightly different from the individual-written report.
Reference: MMF2030_Project_Code_Alex.ipynb
- Used as the foundation for the individual report but contains more verbose outputs and warnings.
Key Techniques:
- XGBoost and Random Forest classifiers.
- Hyperparameter Tuning with GridSearchCV.
- Profit Curve Analysis for business-aligned decision thresholds.
- Best Model: XGBoost (Optimized for ROC AUC)
- ROC AUC: 0.75
- Accuracy: 86%
- Maximum Profit: $35,513
- Optimal Threshold: 0.16
Business Impact:
- Improved credit risk management and proactive risk mitigation.
- Enhanced financial inclusion for low-risk borrowers.
- Increased operational efficiency via automated decision-making.
- Python
- Jupyter Notebook
- XGBoost, Random Forest
- Pandas, NumPy, Matplotlib, Seaborn
- GridSearchCV for hyperparameter tuning
- ROC-AUC, F1-Score, Profit Curve Analysis
- Credit Risk Assessment: Real-time credit scoring for loan approvals.
- Risk-Based Pricing: Adjust loan interest rates based on risk classifications.
- Fraud Detection: Identify suspicious applications through anomaly detection.
- Portfolio Monitoring: Dynamic adjustment of lending strategies in response to economic changes.
- University of Toronto – Master of Mathematical Finance Program
- MMF2030 Machine Learning for Finance Course
Let me know if you'd like further adjustments or improvements!