Customer Churn Prediction

End-to-end ML system predicting telecom customer churn
CV AUC-ROC: 0.845 | XGBoost + SHAP + FastAPI + Streamlit

---

Business problem

~27% of telecom customers churn annually, costing the industry billions in lost revenue. This project builds a production-ready ML system that identifies high-risk customers so retention teams can act proactively — reducing churn by targeting the right customers with the right offer at the right time.

---

Live demo

Resource	Link
Streamlit app	customer-churn-prediction-dazrvuydozwk3mm4m8ov9k.streamlit.app
GitHub repo	github.com/Efrrowini/customer-churn-prediction

---

Results

Metric	Score
CV AUC-ROC (5-fold)	0.845
Test AUC-ROC	0.835
Precision	0.642
Recall	0.537
F1-score	0.585

---

Key findings from EDA

• Month-to-month customers churn at 42% vs 11% for 2-year contracts
• Fiber optic users are 2x more likely to churn despite higher bills
• Customers without tech support churn at 41% vs 15% with support
• New customers (tenure under 3 months) are the highest risk segment

---

SHAP explainability

Global feature importance — top churn drivers:

Per-customer explanation — why one high-risk customer is flagged:

---

Architecture

Raw CSV
  -> EDA (notebooks/01_eda.ipynb)
  -> Feature engineering (src/preprocess.py)
  -> sklearn Pipeline + XGBoost (src/train.py)
  -> MLflow experiment tracking
  -> SHAP explainability (notebooks/03_shap_analysis.ipynb)
  -> FastAPI endpoint (api/main.py)
  -> Streamlit demo (app/streamlit_app.py)

---

Tech stack

Layer	Tools
Data	pandas, numpy
Modelling	scikit-learn, XGBoost
Explainability	SHAP
Experiment tracking	MLflow
API	FastAPI, Pydantic, uvicorn
Frontend	Streamlit
Deployment	Streamlit Cloud
Version control	Git, GitHub

---

Project structure

customer-churn-prediction/
├── data/raw/                  # original CSV (gitignored)
├── data/processed/            # train/test splits (gitignored)
├── notebooks/
│   ├── 01_eda.ipynb           # exploratory analysis
│   ├── 02_modeling.ipynb      # pipeline verification
│   └── 03_shap_analysis.ipynb # explainability
├── src/
│   ├── preprocess.py          # feature engineering
│   └── train.py               # pipeline + MLflow
├── api/
│   ├── main.py                # FastAPI app
│   └── schemas.py             # Pydantic models
├── app/
│   └── streamlit_app.py       # demo UI
├── models/                    # saved pipeline artifact
├── reports/                   # plots and screenshots
├── requirements.txt
└── README.md

---

Run locally

git clone https://github.com/Efrrowini/customer-churn-prediction
cd customer-churn-prediction
python -m venv venv
venv\Scripts\activate        # Windows
pip install -r requirements.txt

# Train model
python -m src.train

# Run Streamlit app
streamlit run app/streamlit_app.py

# Run API
uvicorn api.main:app --reload
# then open http://localhost:8000/docs

---

Dataset

IBM Telco Customer Churn dataset — 7,032 customers, 21 features, binary target.
Available on Kaggle.

---

Built by Efrrowini | Presidency University Bangalore | Data Science Portfolio