Customer_churn_prediction_engine

End-to-end churn prediction engine | XGBoost · K-Means · Kaplan-Meier · Power BI

Customer Churn Prediction & Revenue Recovery Engine

End-to-end ML pipeline for telecom customer attrition analysis

Ashik Mu Asharaf · Data Science Professional ·MSc AI & Data Science (Distinction)
LinkedIn · ashikmasharaf@outlook.com

---

Business Problem

A telecom company is losing 400–500 customers every month.
Each customer is worth ₹850/month on average.
That is ₹4.25 lakh walking out the door — every single month.

This project builds a system that identifies which customers are about to leave, quantifies the revenue at risk, and models the ROI of retaining them before they go.

---

Project Architecture

customer-churn-prediction-engine/
├── notebooks/
│   ├── 01_EDA_Business_Framing.ipynb         # Revenue at risk · churn drivers · loyalty curve
│   ├── 02_Customer_Segmentation.ipynb        # K-Means · 4 customer personas · priority matrix
│   ├── 03_Churn_Prediction_Models.ipynb      # Logistic Regression · Random Forest · Gradient Boosting
│   ├── 04_Revenue_Impact_Quantification.ipynb # ₹ ROI · 3 retention scenarios · CFO summary
│   ├── 05_Survival_Analysis.ipynb            # Kaplan-Meier · intervention calendar · log-rank test
│   └── 06_Executive_Summary.ipynb            # Full narrative · all outputs in one place
├── models/
│   ├── logistic_regression.pkl
│   ├── random_forest.pkl
│   ├── gradient_boosting.pkl
│   └── model_config.json
├── outputs/
│   ├── retention_scenarios.csv
│   ├── segment_revenue_impact.csv
│   ├── intervention_calendar.csv
│   └── km_*.csv                              # Kaplan-Meier curves by segment
├── data/
│   └── telco_churn.csv                       # IBM Telco dataset (place here before running)
├── requirements.txt
└── README.md

---

Methodology

Phase 01 — Exploratory Data Analysis & Business Framing

Baseline churn rate established at 26.5%. Revenue at risk quantified at ₹4.25 lakh/month (₹51 lakh/year). Identified the critical dropout window between Month 12 and Month 18 of customer tenure.

Phase 02 — Customer Segmentation (K-Means)

K-Means clustering (K=4, validated by Elbow + Silhouette) identified four business personas:

Code	Persona	Churn Rate	Priority
ST_M2M	High-Risk New Customers	~43%	CRITICAL
DIG_STRM	Price-Sensitive Churners	~28%	HIGH
DISC_SAV	Engaged Mid-Tier	~8%	MEDIUM
LT_BUND	Loyal High-Value	<5%	PROTECT

Phase 03 — Churn Prediction Models

Three classification models trained and evaluated with recall on the churned class as the primary metric — not accuracy.

Model	Recall (Churn)	AUC-ROC
Logistic Regression	78.4%	0.832
Random Forest	81.1%	0.854
Gradient Boosting ✓	85.2%	0.881

Decision threshold optimised from 0.50 → 0.38 to maximise recall while maintaining precision.

Phase 04 — Revenue Impact Quantification

Model predictions converted to rupees across three retention scenarios:

Scenario	Retention Rate	Monthly Saving	Annual ROI
Conservative (outreach)	20%	₹36,200	~340%
Moderate (₹150 discount) ✓	40%	₹72,400	~520%
Aggressive (upgrade offer)	60%	₹1.27 lakh	~680%

Phase 05 — Survival Analysis (Kaplan-Meier)

Implemented from first principles (equivalent to lifelines.KaplanMeierFitter).
Log-rank test confirms survival curves differ significantly by contract type (p < 0.001).

Key finding: Month-to-month contracts exhibit their steepest survival decline between Month 12 and Month 18.
Action: Deploy retention campaigns at Month 10 — before the decision to leave is made.

---

Key Results

• ₹4.25 lakh monthly revenue at risk identified
• 85.2% recall on churned class (Gradient Boosting, threshold=0.38)
• 4 customer personas with distinct churn trajectories and intervention strategies
• Month 10 established as the universal pre-emptive intervention window
• ₹72,400/month recoverable under moderate retention scenario (₹8.69 lakh/year)
• Top 3 churn drivers: tenure, contract type, internet service tier

---

Technical Stack

Tool	Purpose
Python · Pandas · NumPy	Data pipeline
Scikit-Learn	Logistic Regression, Random Forest, Gradient Boosting, K-Means
SciPy	Kaplan-Meier survival estimator (manual implementation)
Matplotlib · Seaborn	19 production-grade visualisations
Power BI Desktop	4-view executive dashboard (.pbix)
Jupyter Notebooks	6 structured analysis notebooks
GitHub	Version control · public portfolio

---

Dataset

IBM Telco Customer Churn — Kaggle
7,043 customer records · 21 features · No missing values after preprocessing

Place the downloaded CSV at data/telco_churn.csv before running notebooks.

---

Running the Project

git clone https://github.com/Ashik-AI-DataScientist/customer-churn-prediction-engine.git
cd customer-churn-prediction-engine
pip install -r requirements.txt
# Place telco_churn.csv in data/
jupyter notebook notebooks/01_EDA_Business_Framing.ipynb

Run notebooks in order (01 → 05). Each saves outputs used by the next.

---

---

*Report: Available as pdf in report Customer_Churn_Project_IBM/Project Report/Telecom_Churn_Upgraded_Enterprise_Report.pdf https://github.com/Ashik-AI-DataScientist/Customer_churn_prediction_engine/blob/main/Customer_Churn_Project_IBM/Project%20Report/Telecom_Churn_Upgraded_Enterprise_Report.pdf