Vvshantal/credit_scoring_ml

Machine Learning-powered loan eligibility platform with Uganda-specific financial data and professional banking interface. Built with FastAPI, React, and Scikit-learn. # Credit Scoring ML Models Machine Learning models for credit risk prediction using behavioral features from mobile money transaction data. ## Overview This project focuses on predicting credit risk using transaction patterns from the PaySim mobile money dataset. It includes comprehensive feature engineering, model training, and evaluation for multiple ML algorithms. ## Machine Learning Models - **Logistic Regression**: Interpretable linear model with coefficient analysis - **Random Forest**: Ensemble method with feature importance ranking - **XGBoost**: Gradient boosting with multiple importance types (gain, weight, cover) - **LightGBM**: Fast and memory-efficient gradient boosting ## Feature Engineering **58+ Behavioral Features** engineered across 7 categories: - **Income Stability** (10 features): income patterns, consistency, and trends - **Expenditure Patterns** (12 features): spending behavior and transaction types - **Balance Maintenance** (10 features): balance management, volatility, and thresholds - **Transaction Diversity** (6 features): recipient diversity and transaction entropy - **Temporal Patterns** (8 features): timing regularity and activity patterns - **Rolling Window Features** (9 features): 24h, 168h, and 336h aggregations - **Risk Indicators** (5 features): overdraft attempts and suspicious patterns ## Project Structure ``` credit_scoring_ml/ ├── data/ │ └── raw/ # PaySim data and sample data ├── notebooks/ # Interactive model analysis │ ├── 00_quick_demo.ipynb # Quick demo with sample data │ ├── 01_logistic_regression_model.ipynb │ ├── 02_random_forest_model.ipynb │ ├── 03_xgboost_model.ipynb │ └── 04_lightgbm_model.ipynb ├── scripts/ │ └── train_paysim_model.py # End-to-end training pipeline ├── src/ │ ├── data/ │ │ └── loader.py # Data loading utilities │ ├── features/ │ │ └── paysim_engineer.py # Feature engineering (58+ features) │ └── models/ │ …