Customer Churn Prediction

A telecom company loses revenue every time a customer cancels their subscription. This project builds a machine learning pipeline to identify customers at high risk of churning before they leave — giving the retention team time to intervene.

---

Results

Model	Test AUC	CV AUC (5-fold)
Logistic Regression	0.764	0.771 ± 0.014
Random Forest ✅	0.768	0.762 ± 0.021
Gradient Boosting	0.767	0.758 ± 0.015

Selected model: Random Forest — best test AUC, robust to outliers, no feature scaling required, and fast to retrain on new data.

---

Key Findings

• Contract type is the strongest predictor — month-to-month customers churn at 44% vs 10–15% for annual/bi-annual contracts
• New customers are highest risk — churn rate drops significantly after the first 12 months
• Fiber optic customers churn more — likely driven by higher monthly charges and price sensitivity
• Recommended decision threshold: 0.35 (vs default 0.5) — catches ~65% of churners at an acceptable false-positive rate for a low-cost retention campaign (e.g. a discount email)

---

Project Structure

├── Customer_Churn_Prediction.ipynb   # Full notebook — EDA, modelling, evaluation
├── customer_churn_realistic.csv      # Dataset (5,880 customers, 20 features)
└── README.md

---

Notebook Walkthrough

Section	What's covered
1. Data Inspection	Shape, dtypes, missing values, class distribution
2. EDA	6 charts testing hypotheses about churn drivers
3. Preprocessing	Encoding strategy, feature engineering, train/test split
4. Modelling	Logistic Regression, Random Forest, Gradient Boosting
5. Evaluation	ROC-AUC, confusion matrix, cross-validation
6. Feature Importance	Top predictors from both tree-based models
7. Business Recommendations	Threshold tuning, retention strategy, monitoring plan

---

Feature Engineering

Two features were engineered beyond the raw columns:

• tenure_group — bucketed tenure into 4 bands (0–12, 12–24, 24–48, 48–72 months) to capture non-linear early churn risk
• charge_ratio — monthly charges divided by tenure, capturing average spend rate and the tenure-charge interaction more cleanly than either feature alone

---

Tech Stack

• Python 3 — pandas, numpy, matplotlib, seaborn
• scikit-learn — preprocessing, modelling, cross-validation, evaluation

---

How to Run

git clone https://github.com/Tinyiko_Mathebula/customer-churn-prediction
cd customer-churn-prediction
pip install pandas numpy matplotlib seaborn scikit-learn
jupyter notebook Customer_Churn_Prediction.ipynb

---

Business Context

At a threshold of 0.35, the model flags approximately 1 in 3 customers as churn risk. For a retention campaign costing $5 per outreach and saving $200 in average customer lifetime value, the model generates positive ROI if it identifies churners with precision above ~2.5% — well within its demonstrated performance.

---

Author

Tinyiko Patience Mathebula