Credit Risk Prediction Using Machine Learning

This repository contains an academic machine learning project focused on credit default prediction using the LendingClub dataset.
The objective is to design, evaluate, and compare several models under realistic banking constraints, including temporal validation, class imbalance, and economic interpretability.

The project combines:

• exploratory data analysis,
• robust preprocessing,
• multiple predictive models,
• out-of-time validation,
• and business-oriented model selection.

Final Version of the Project

Final Notebook

The final and authoritative implementation of the project is:

notebooks/IF4_Derasse_Paviet_Minard_Project_CreditRisk.ipynb

This notebook contains:

• the full data preprocessing pipeline,
• leakage detection and removal,
• temporal (OOTV) validation,
• model training and evaluation (Logistic Regression, Random Forest, XGBoost, TabNet, Decision Tree),
• economic interpretation of results,
• and final model selection.

This notebook should be considered the definitive version for grading and evaluation

Final Report (PDF)

The final written report is available in:

report/IF4_Derasse_Paviet_MinardCredit_Risk_Prediction___Machine_Learning.pdf

The LaTeX source is located in:

• report/main.tex
• report/figures/

The report presents:

• the business context,
• methodological choices,
• results analysis,
• operational recommendations,
• and conclusions aligned with banking practices.

Previous / Intermediate Versions

The repository also contains earlier or intermediate notebooks created during the development process:

• Lab5_Project_CreditRisk.ipynb
• updated version of project.ipynb
• Essai_Tuning.ipynb

These files correspond to:

• early experiments,
• hyperparameter tuning attempts,
• partial implementations.

They are kept for traceability only and should not be used for evaluation.

Repository Structure

Modeling Strategy (High-Level)

• Validation protocol: strict time-based split (≤2016 / 2017 / 2018 OOTV)
• Main challenge: strong class imbalance and temporal drift
• Key metrics: ROC-AUC, Recall (default), Precision (default), F1-score
• Excluded approach: SMOTE (too many false positives for banking context)

Final model recommendations:

• Production credit scoring -> Random Forest
• Early-warning / monitoring-> TabNet
• Governance et explainability reference --> Optimized Decision Tree

Authors

• Sherynne Derasse
• Raphaël Paviet
• Basile Minard

Dataset

The dataset used is the LendingClub Loan Dataset (2007–2018)
Publicly available at: https://www.lendingclub.com/