Customer Churn Prediction Pipeline

Tools: Python (pandas, scikit-learn, matplotlib, seaborn) · SQL (sqlite3) · Plotly.js · GitHub Pages
Dataset: Telco Customer Churn — Kaggle (7,043 records)
Model accuracy: 82% | AUC-ROC: 0.86
Live Dashboard: 🔗 View Live Interactive Dashboard

---

Problem Statement

A telecom company loses ~27% of its customers annually. Can we predict which customers are about to churn — and translate that prediction into a concrete retention strategy?

---

Approach

Step	Description
1. SQL Exploration	Segmentation queries via sqlite3 — tenure, contract type, monthly charges
2. Data Cleaning	Handle missing values, encode categoricals, scale features
3. Feature Engineering	Create tenure bands, charge-per-service ratio, engagement score
4. Modelling	Logistic Regression + Random Forest — comparison with cross-validation
5. Evaluation	Confusion matrix, ROC curve, precision/recall, feature importance
6. Business Output	Translate model to actionable retention recommendation

---

Key Results

Model	Accuracy	AUC-ROC	Precision (Churn)	Recall (Churn)
Logistic Regression	80%	0.84	0.67	0.58
Random Forest	82%	0.86	0.71	0.63

---

Business Recommendation

Targeting the 3 highest-risk segments (Month-to-month contracts + Fibre Optic + High monthly charges) with a proactive retention offer — a 12-month contract discount — is projected to reduce overall churn by 12–15%, saving approximately £340,000 in annual revenue per 10,000 customers.

Top 3 churn drivers identified:

1. Contract type — Month-to-month customers churn at 3.4× the rate of annual contract holders
2. Tenure — Customers in their first 6 months are at highest risk (42% churn rate)
3. Internet service — Fibre Optic subscribers churn at 2.1× the rate of DSL users

---

Repository Structure

customer-churn-prediction/
├── index.html               # Interactive Plotly.js dashboard (deployed via GitHub Pages)
├── churn_model.py           # Modular ML pipeline — load, clean, engineer, train, evaluate
├── churn_pipeline.ipynb     # Exploratory notebook with full analysis walkthrough
├── WA_Fn-UseC_-Telco-Customer-Churn.csv  # Source dataset
├── requirements.txt         # Python dependencies
├── LICENSE                  # MIT License
└── README.md                # This file

---

How to Run Locally

1. Clone the repository:

   git clone https://github.com/mayankjoshiii/customer-churn-prediction.git

2. Install dependencies:

   pip install -r requirements.txt

3. Run the pipeline:

   python churn_model.py

4. Open index.html in any modern browser to view the dashboard — no server required.

---

Live Dashboard

👉 View the Interactive Dashboard

---

Author

Mayank Joshi — Business Analyst & Data Analyst
MSc Business Analytics (Distinction) · Swansea University
LinkedIn · GitHub