Data Science · End-to-End Project

Customer Churn
Analysis &
Prediction

An end-to-end data science project analysing telecom customer churn — from SQL business queries to a deployed XGBoost model served via FastAPI.

View Live Dashboard View on GitHub

26.5%

Churn Rate

$139k

Monthly Revenue at Risk

84.5%

Model ROC-AUC

84.8%

Model Recall

Project Pipeline

8-Week End-to-End Workflow

From raw data to a production prediction system, every layer is documented and reproducible.

WEEK 01 — 02

Data & SQL

Data cleaning pipeline, 10-section EDA, and 10 SQL business queries against SQLite.

Pandas · SQLite · Seaborn

WEEK 03 — 04

Dashboard

Interactive Streamlit dashboard with filters, KPI cards, and insight banners. Deployed on Streamlit Cloud.

Streamlit · Plotly

WEEK 05 — 06

ML Modelling

5-model comparison, hyperparameter tuning via RandomizedSearchCV, SHAP explainability.

XGBoost · SHAP · Sklearn

WEEK 07 — 08

API & Predictor

FastAPI REST API with Swagger docs. Live prediction UI integrated into Streamlit dashboard.

FastAPI · Uvicorn · Pydantic

Business Intelligence

Key Results

Findings from exploratory analysis and SQL queries across 7,043 telecom customers.

Overall Churn Rate

26.5%

1,869 of 7,043 customers churned

Revenue at Risk

$139,131

Monthly revenue from churned customers

Month-to-Month Churn

42.7%

3x higher than annual plan customers

New Customer Churn

47.4%

Customers in first 12 months

Electronic Check Churn

45.3%

vs 15% for auto-pay customers

Avg Charge — Churned

$74.44

$13.17 higher than retained customers

Machine Learning

5-Model Comparison

Models tuned with RandomizedSearchCV and evaluated on ROC-AUC — the primary metric for ranking churn risk.

Model	Accuracy	Precision	Recall	F1 Score	ROC-AUC	Status
XGBoost (Tuned)	75.87%	46.96%	84.76%	60.44%	84.47%	Production
Random Forest (Tuned)	77.00%	55.30%	69.79%	61.70%	84.12%	Runner-up
Logistic Regression	73.81%	50.42%	79.68%	61.76%	83.99%	Baseline
LightGBM (Tuned)	75.94%	53.41%	73.26%	61.78%	83.54%	Compared
SVM	74.24%	50.98%	76.20%	61.09%	81.66%	Compared

Selection Rationale — ROC-AUC was chosen as the primary metric because the retention team contacts a fixed number of customers per month. The model's ability to rank customers by churn risk determines campaign ROI directly. XGBoost (Tuned) achieved the highest ROC-AUC of 84.47% with a recall of 84.76%, meaning it correctly identifies 85 out of every 100 churners before they leave.

REST API

FastAPI Prediction Service

XGBoost model wrapped as a production REST API with Swagger documentation, input validation via Pydantic, and batch prediction support.

Endpoints

GET / Health check

GET /model-info Model metadata

POST /predict Single prediction

POST /predict-batch Up to 100 customers

Sample Response

{
  "churn_prediction": true,
  "churn_probability": 0.8968,
  "churn_probability_pct": "89.7%",
  "risk_level": "High",
  "recommendation": "Immediate
  retention action required.",
  "model_used": "XGBoost (Tuned)"
}

Business Recommendations

High-Risk Segments

Key findings from SQL analysis mapped to actionable retention strategies.

42.7%

Month-to-Month Contracts

Offer first-month discounts or loyalty credits to move customers to annual plans. Churn drops to 11% on one-year contracts.

47.4%

First 12 Months (New Customers)

Implement a structured onboarding programme with proactive check-ins at months 1, 3, and 6 to reduce early churn.

41.0%

Fiber Optic Without Tech Support

Proactively upsell Tech Support to Fiber Optic customers. Churn drops from 41% to 18% with Tech Support — the highest ROI retention action.

45.3%

Electronic Check Payment

Offer a 5% bill discount to switch to automatic payment. Electronic check users churn at 3x the rate of credit card auto-pay customers.

Technology

Tech Stack

Every layer of the project uses industry-standard tools used in production data science environments.

Data & Analysis

Python 3.11 Pandas NumPy SQLite SQLAlchemy

Visualisation

Plotly Seaborn Matplotlib Streamlit

Machine Learning

XGBoost LightGBM Scikit-learn SHAP

API & Deployment

FastAPI Uvicorn Pydantic Streamlit Cloud

Tuning & Validation

RandomizedSearchCV StratifiedKFold Learning Curves

Version Control

Git GitHub GitHub Pages