🏦 Banking Loan Default Prediction

An end-to-end machine learning system to predict whether a loan applicant will default, built with XGBoost, SHAP explainability, customer segmentation, and an interactive Streamlit dashboard.

---

📁 Project Structure

banking-loan-default-prediction/
├── data/
│   ├── raw/           loan_data.csv (5,000 synthetic records)
│   └── processed/     train.csv, test.csv, final_dataset.csv
├── notebooks/
│   ├── 01_EDA.ipynb
│   ├── 02_Feature_Engineering.ipynb
│   └── 03_Model_Training.ipynb
├── sql/
│   ├── schema.sql          Database schema
│   └── business_queries.sql  Analytical SQL queries
├── src/
│   ├── data_loader.py
│   ├── preprocessing.py
│   ├── feature_engineering.py
│   ├── eda.py
│   ├── train_model.py
│   ├── evaluate_model.py
│   ├── prediction.py
│   ├── customer_segmentation.py
│   ├── explainability.py
│   └── utils.py
├── models/            Trained artifacts (.pkl)
├── app/
│   └── streamlit_app.py
├── reports/charts/    All EDA + evaluation charts
├── requirements.txt
└── main.py            Pipeline orchestrator

---

🚀 Quick Start

1. Install dependencies

pip install -r requirements.txt

2. Generate raw data (already included)

cd data/raw && python generate_data.py

3. Run the full pipeline

python main.py

Flags:

python main.py --skip-eda        # skip chart generation
python main.py --skip-shap       # skip SHAP (faster)
python main.py --threshold 0.4   # custom decision threshold
python main.py --predict-only    # score test set with saved model

4. Launch the Streamlit App

streamlit run app/streamlit_app.py

---

📊 Model Performance

Metric	Value
ROC-AUC	~0.71
F1 Score	~0.59
Accuracy	~0.65
Algorithm	XGBoost
Imbalance	SMOTE
Validation	5-Fold Stratified CV

---

🔧 Key Features

• 34 engineered features including loan-to-income ratio, payment-to-income, risk score composite, credit utilisation proxy, and more
• SMOTE oversampling to handle class imbalance (~44% default rate)
• SHAP global (beeswarm + bar) and local (waterfall) explanations
• K-Means customer segmentation (4 clusters) with PCA visualisation
• Streamlit dashboard with: portfolio overview, single applicant assessment, batch CSV scoring, and model performance charts
• SQL schema + 10 business queries ready for PostgreSQL

---

🧩 Pipeline Modules

Module	Description
data_loader.py	Load & validate raw CSV
preprocessing.py	Clean, encode, split
feature_engineering.py	14+ derived features, scaling
eda.py	7 chart types saved to reports/
train_model.py	XGBoost + SMOTE + 5-fold CV
evaluate_model.py	Metrics, ROC, PR, confusion matrix
prediction.py	Single & batch inference API
customer_segmentation.py	K-Means clustering
explainability.py	SHAP global & local explanations

---

📋 Input Features

Feature	Type	Description
age	int	Applicant age
income	float	Annual income
loan_amount	float	Requested loan
loan_term	int	Months
interest_rate	float	Annual %
credit_score	int	300–850
employment_type	cat	Salaried/Self-Employed/Business/Unemployed
loan_purpose	cat	Home/Auto/Education/Personal/Business/Medical
loan_grade	cat	A–F
debt_to_income	float	DTI ratio
num_late_payments	int	Historical late payments
previous_defaults	int	Prior defaults
…	…	+ 8 more raw + 14 engineered

---

📜 License

MIT