This project aims to predict whether a borrower will default on a loan based on features such as credit grade, employment length, loan purpose, and more. We used supervised machine learning techniques on real-world data from Lending Club to build and evaluate models that can help lenders reduce risk.
To identify potential loan defaulters before issuing loans, enabling better credit risk assessment and minimizing financial losses for lending institutions.
- Source: Lending Club (public datasets)
- Used:
accepted_loans.csv - Target Variable:
loan_status(0 = Fully Paid, 1 = Charged Off) - Default Rate: ~13%
- Data Cleaning: Removed data leakage features and irrelevant columns
- Exploratory Data Analysis (EDA): Analyzed trends behind defaults
- Feature Encoding: Converted categorical variables using One-Hot and Label Encoding
- Class Imbalance Handling: Applied SMOTE to balance the target variable
- Model Training: Trained Logistic Regression, Decision Tree, Random Forest, and XGBoost
- Evaluation: Compared models based on Recall, Precision, F1-score, and ROC-AUC
| Model | Recall (Default) | Precision | ROC-AUC |
|---|---|---|---|
| Logistic Regression (SMOTE) | 0.03 | 0.40 | 0.66 |
| Decision Tree (SMOTE) | 0.26 | 0.21 | 0.65 |
| Random Forest (SMOTE) | 0.20 | 0.27 | 0.68 |
| XGBoost (SMOTE) | 0.01 | 0.43 | 0.69 |
Random Forest after SMOTE was selected as the best model due to its balance of recall and ROC-AUC — critical in catching actual defaulters.
Plots include:
- Confusion Matrix
- ROC Curve
- Feature Importances
You can find them in the plots/ folder.
Install packages using:
pip install -r requirements.txt