💰 Loan Default Risk Prediction & Scoring Engine

End-to-end credit risk system that predicts borrower default probability and converts it into a business-friendly 0–100 risk score — trained on 1.3M+ real Lending Club loans.

---

🚀 Live Demo

👉 loan-default-risk-engine-gxibehstp7sugrrart2cfa.streamlit.app

---

📸 App Preview

---

✨ What Makes This Different

Typical credit risk model	This project
Binary output only (default / no default)	Probability score + 0–100 risk score
Single model	3-model comparison (LR, RF, XGBoost)
No explainability	SHAP identifies key default drivers
No leakage prevention	Post-outcome columns explicitly removed
Academic only	Live interactive Streamlit app

---

📌 Overview

Built on the Lending Club Loan Dataset (1.3M+ finalized loans), this system:

• Predicts default probability with 0.7203 ROC-AUC
• Converts probability into an interpretable 0–100 risk score
• Classifies borrowers into Low / Medium / High risk categories
• Uses SHAP to identify top default drivers per prediction
• Deployed as a fully interactive Streamlit dashboard with sidebar inputs, risk gauge, and risk factor analysis

---

📊 Model Performance

Model	ROC-AUC
XGBoost	0.7203
Logistic Regression	0.7090
Random Forest	0.7066

• Class imbalance (80/20) handled using scale_pos_weight
• Time-based stratified train/test split — no data leakage
• Only pre-application features used

---

🔑 Key Default Drivers (SHAP)

Feature	Impact
Interest Rate	Higher rate → higher default risk
Loan Grade	Grade G borrowers default ~5x more than Grade A
Debt-to-Income Ratio	Strong positive correlation with default
Loan-to-Income Ratio	Engineered feature — captures relative loan burden
Revolving Utilization	High utilization signals financial stress
Loan Term	60-month loans default more than 36-month

---

🧠 Feature Engineering

Four financial stress features engineered from raw inputs:

Feature	Formula	Why
loan_to_income	loan_amnt / annual_inc	Relative loan burden
installment_to_income	installment / (annual_inc/12)	Monthly payment vs monthly income
revol_bal_to_income	revol_bal / annual_inc	Revolving debt load
grade_ordinal	A=1 → G=7	Numeric credit grade encoding

---

🏗️ ML Pipeline

Lending Club Dataset (1.3M+ loans)
        ↓
  Filter finalized loans → Create binary target
        ↓
  Remove data leakage columns
        ↓
  Feature selection + missing value imputation
        ↓
  Feature engineering (stress ratios)
        ↓
  One-hot encoding + stratified train/test split
        ↓
  Model comparison: LR / RF / XGBoost
        ↓
  XGBoost selected (best ROC-AUC)
        ↓
  SHAP explainability + Risk scoring (0–100)
        ↓
  Streamlit deployment

---

🗃️ Dataset

Property	Value
Source	Lending Club Loan Dataset (Kaggle)
Total records	2.26M
Finalized loans used	1.3M+
Features	18 pre-application features + 4 engineered
Target	0 = Fully Paid, 1 = Charged Off
Class ratio	~80% paid / ~20% default

---

🛠️ Tech Stack

Layer	Technology
Model	XGBoost
Explainability	SHAP
Imbalance Handling	scale_pos_weight
Frontend	Streamlit
Visualization	Matplotlib
Data Processing	Pandas, NumPy, Scikit-learn

---

💻 Run Locally

# 1. Clone the repo
git clone https://github.com/Azim521/Loan-Default-Risk-Engine.git
cd Loan-Default-Risk-Engine

# 2. Install dependencies
pip install -r requirements.txt

# 3. Run the app
streamlit run app.py

---

📁 Project Structure

Loan-Default-Risk-Engine/
├── app.py                    ← Streamlit dashboard
├── requirements.txt          ← Dependencies
├── screenshot.png            ← App preview
├── loan_default_eda.ipynb    ← Full EDA + model training notebook
└── model/
    ├── xgb_loan_model.pkl    ← Trained XGBoost model
    └── feature_columns.pkl   ← Feature schema for inference

---

📓 EDA Notebook

Full exploratory analysis, feature engineering, model training and evaluation:

👉 View Notebook

Covers: target distribution, missing value analysis, feature distributions, correlation analysis, leakage removal, engineered features, model comparison, SHAP, and risk scoring.

---

🔮 Future Improvements

• Hyperparameter tuning with time-series cross-validation
• Add SHAP waterfall chart per prediction in the app
• Integrate with credit bureau API for real-time scoring
• Extend to multi-label risk tiers

---

📬 Contact

Built by Azim Sadath