📡 Telco Customer Churn Prediction

Predictive Modelling for Customer Retention

Author: Marisa Oliveira
Project Type: Group Data Analytics / Machine Learning Project
Year: 2026

---

🚀 Interactive App

👉 Launch the Streamlit App

---

📌 Project Overview

A telecom provider was experiencing a churn rate of 26–27%, with the highest attrition among high-value fibre-optic customers.

This project builds an end-to-end churn prediction system to identify at-risk customers early and enable proactive retention strategies.

---

🎯 Key Results

• 📉 Identified ~69% of churn-risk customers (Recall ≈ 0.69)
• 🎯 Improved precision from 0.49 → 0.61 (~24% increase)
• 🔻 Reduced false alerts from 315 → 168 (−47%)
• 📊 Achieved PR-AUC ≈ 0.652 and MCC ≈ 0.51
• 🌳 Tree-based models slightly outperformed on ranking metrics but lacked interpretability

---

💼 Business Impact

The model enables a shift from reactive churn handling → proactive retention strategy.

Operational benefits:

• 🎯 Focus retention efforts on genuinely at-risk customers
• 💰 Reduce wasted outreach and improve campaign ROI
• 🔍 Provide interpretable insights for decision-makers

---

🧠 Technical Approach

📊 Data Preparation

• Cleaned missing values in TotalCharges using tenure consistency
• Handled zero-tenure customers appropriately
• Removed identifiers to prevent leakage
• Applied one-hot encoding to categorical variables
• Standardised numerical features

⚖️ Handling Imbalance

• Dataset: ~73% non-churn vs ~27% churn
• Applied SMOTE on training data only
• Maintained realistic distribution in test set

---

📈 Modelling Strategy

Metrics Used

• PR-AUC → ranking churners correctly
• MCC → balanced performance metric
• Precision / Recall / F1 → operational trade-offs
• ROC-AUC → overall ranking ability

---

🤖 Model Comparison

Model	Precision	Recall	F1	MCC	ROC-AUC	PR-AUC
Logistic Regression	0.49	0.82	0.61	~0.40	~0.79	~0.57
Logistic + SMOTE	0.61	0.69	0.65	0.51	0.80	0.652
Random Forest	0.68	0.51	0.58	0.47	0.845	0.660
XGBoost	0.51	0.79	0.62	0.46	0.846	0.665

---

🏆 Final Model Selection

Logistic Regression with SMOTE

Chosen because it offers:

• ⚖️ Balanced performance
• 🔍 Interpretability for business stakeholders
• 🧩 Simplicity and ease of deployment

Tree-based models remain strong candidates for future iterations.

---

📊 Key Insights

📉 Contract Risk

Month-to-month customers show significantly higher churn risk.

Recommendation:
Encourage migration to longer contracts via incentives.

---

💸 Price Sensitivity

Churners tend to pay ~£15 more monthly.

Recommendation:
Optimise pricing strategy for high-cost segments.

---

⏳ Critical Risk Window

Highest churn risk occurs within the first 12 months.

Recommendation:
Strengthen onboarding and early engagement.

---

⚙️ Operational Impact

Metric	Baseline	Final Model
False Positives	315	168
Precision	0.49	0.61
Recall	0.82	0.69
MCC	~0.40	0.51

✔ Reduced 147 false alerts per cycle
✔ Maintained strong churn detection

---

📂 Project Structure

• 📓 Exploratory Data Analysis Notebook
• 🤖 Modelling & Evaluation Notebook
• 📊 Dataset (Telco Churn)
• 📄 Documentation (README)
• 🌐 Streamlit App

---

🛠️ Tools & Technologies

• Python
• Pandas, NumPy
• Scikit-learn
• imbalanced-learn (SMOTE)
• XGBoost
• Matplotlib, Seaborn
• Streamlit
• Jupyter Notebook