Predicting Loan Defaults with Machine Learning

This repository contains the code, resources, and final report of my undergraduate thesis: "Conception d’un Modèle de Prédiction des Défauts de Paiement sur les Prêts : Approches, Algorithmes et Évaluation comparative".
The project was completed in 2024 as part of my Bachelor's studies in Mali, and explores various machine learning techniques to predict loan defaults using real-world data.

🔍 Project Overview

This work investigates the use of machine learning algorithms to assess the risk of loan default based on data provided by Home Credit Group, through a public Kaggle competition. The goal is to compare traditional and advanced approaches for credit risk assessment and explore how data-driven methods can improve loan allocation decisions.

Key components of the project include:

• Data cleaning and preprocessing
• Automatic feature engineering using Deep Feature Synthesis (DFS)
• Feature selection using statistical methods
• Model training and evaluation on various algorithms

🧠 Models Implemented

The following machine learning models were trained and compared:

• Logistic Regression
• Random Forest
• Gradient Boosting
• Artificial Neural Networks
• Anomaly Detection using Support Vector Machines (SVM)

Each model was trained using a consistent pipeline and evaluated on the same dataset.

📁 Repository Structure

├── images/                # Visualizations and illustrations used in the report
├── literature/            # Some Research papers referenced in the study
├── notebooks/             # Jupyter notebooks for each modeling and preprocessing task
├── texFiles/              # LaTeX source files for the thesis report
├── UndergraduateFinal.pdf # Final thesis document (in French)
└── README.md              # You're here!

📊 Dataset

The dataset was derived from the Home Credit Default Risk Kaggle competition. After preprocessing, the final dataset contained:

• 1,526,659 samples
• 100 selected features
• Binary target: loan repaid (0) or default (1)

⚠️ Important Note on Licensing:
The code in this repository is released under the MIT License. However, the data (available on Kaggle) are derived from a Kaggle competition and may be subject to different licensing terms as specified by the competition organizers. Please refer to the original dataset page for more information.

🛠 Feature Engineering

One highlight of the project is the use of automated feature engineering via Featuretools and parallelization using Dask. The raw relational data was transformed into meaningful predictors through a scalable pipeline that processed millions of rows in under 24 hours on Kaggle VMs (accessed through kaggle notebooks).

📈 Evaluation

Models were evaluated using:

• Accuracy
• Precision
• Recall
• ROC AUC

A comparative study is provided in the final chapter of the thesis PDF.

📚 References

See the literature/ folder for some key references and academic papers consulted during this project.

💡 Potential future Works

Some perspectives explored include:

• Incorporating fairness-aware ML to reduce bias
• Extending the pipeline to other types of credit products

🤝 Acknowledgments

Thanks to:

• Dr. Maïga Oumar, my thesis supervisor
• The Kaggle community and Home Credit Group for the data