Credit Scoring & Expected Loss - Lending Club

This is a comprehensive credit risk pipeline built on Lending Club loan data. Estimates the Expected Loss of a loan portfolio using the standard formula, with LGD (Loss Given Default) as a percentage of exposure.

EL = PD × (LGD / 100) × EAD

---

👉 Live Demo Here

---

Project Overview

The project is structured in 4 phases:

• Data Preparation: The raw Lending Club dataset comes with noise, nulls and post-default variables that would cause data leakage to the target variable. We impute those, delete some irrelevant variables and engineer features based on domain knowledge and statistical insights.
• PD Model: Binary classification to estimate the probability of not paying for each loan.
• LGD Model: Regression trained exclusively on defaulted loans to predict loss magnitude.
• Expected Loss: Loss estimation combining both models into a single dollar figure, applying the formula above.

Estimated Models

PD: Classification (Focus on maximizing ROC-AUC)

Model	ROC-AUC
LightGBM	0.75
Logistic Regression	0.74
Random Forest	0.73
Decision Tree	0.55

LightGBM is the best model here. Appart from having the highest ROC-AUC, it is also the fastest and most complete model.

Business Decision: Threshold Selection

The threshold is set at 0.10 to prioritize risk reduction.

• Recall = 0.95 → 95% of defaulters are identified
• Precision = 0.54 → moderate false positives

In credit risk, missing a defaulter (FN) is far more costly than rejecting a good client (FP). In fact, it is worse to lose $100,000 from one client than $10,000 from 10 clients each This threshold minimizes costly defaults, aligning with a conservative risk strategy.

Precision-Recall Trade-off: lower approvals, higher portfolio quality.

LGD: Regression (Focus on minimizing MAE & RMSE)

Model	MAE	RMSE
Random Forest	0.25	1.38
LightGBM	0.41	1.48
Decision Tree	0.56	2.65
Linear Regression	11.39	20.09

Random Forest is the best regression model here, with a MAE of 0.25%.

Expected Loss Result

Metric	Value
Portfolio (test set)	$174,008,000
Expected Loss	$7,815,756
% of portfolio	4.49%

4.49% expected loss. Acceptable range for consumer credit.

Dataset

The dataset is not included due to size constraints.

Download it from Kaggle and preprocess it using ETL.ipynb.

File Structure

.
├── .github/                     # CI: lint with `ruff` on every push. CD: periodic ping to keep the Streamlit app awake.
├── app/
│   ├── __init__.py
│   ├── main.py                  # model training
│   ├── modeling.py
│   ├── utils.py
│   └── plots.py
├── models/
│   └── metadata.json
├── viz/
├── .gitignore
├── ETL.ipynb                   # Data quality, feature engineering & risk analysis
├── LICENSE
├── README.md
├── requirements.txt
└── streamlit_app.py            # deployed application

How to Run

1. Install dependencies

pip install -r requirements.txt

2. Run ETL (Data Preparation)

Open and execute the Jupyter notebook to generate cleaned data:

jupyter notebook ETL.ipynb

This generates Data/cleaned_data.csv

3. Train Models

python -m app.main

This trains PD (classification) and LGD (regression) models and generates models/metadata.json

4. Launch Dashboard

streamlit run streamlit_app.py

Note: For production, the app is already deployed at credit-risk-pp.streamlit.app

License

MIT License