📡 Customer Churn Prediction — Telecom

A machine learning project that predicts which telecom customers are likely to cancel their subscription, enabling the business to take proactive retention action before it's too late.

---

🧩 Problem Statement

Customer churn is one of the most expensive problems in the telecom industry. Acquiring a new customer costs 5–7x more than retaining an existing one. This project builds a classification model to identify at-risk customers before they leave, so the business can intervene with targeted offers or support.

---

📊 Dataset

• Source: IBM Telco Customer Churn Dataset
• Size: 7,043 customers × 21 features
• Target: Churn — Yes (churned) / No (stayed)
• Class Distribution: ~73.5% No Churn / ~26.5% Churn (imbalanced)

Key Features Include:

• Demographics — gender, senior citizen, partner, dependents
• Services — phone, internet, streaming, security, tech support
• Account — contract type, payment method, paperless billing
• Charges — monthly charges, total charges, tenure

---

🔧 Tech Stack

Area	Tools
Language	Python 3
Data Processing	Pandas, NumPy
Visualization	Matplotlib, Seaborn
Machine Learning	Scikit-learn
Model Saving	Joblib
Web App	Streamlit

---

🚀 Project Workflow

1. Data Cleaning

• Converted TotalCharges from string to numeric (contained hidden spaces)
• Filled 11 missing values in TotalCharges with median
• Dropped customerID (irrelevant to prediction)

2. Exploratory Data Analysis (EDA)

• Visualized churn distribution and confirmed class imbalance (26.5% churn)
• Found month-to-month contract customers churn ~3x more than two-year customers
• Plotted distributions of Tenure, MonthlyCharges, TotalCharges
• Correlation heatmap showed TotalCharges highly correlated with Tenure

3. Preprocessing

• Applied pd.get_dummies() for One-Hot Encoding of all categorical features
• Used StandardScaler for numerical feature normalization
• Addressed class imbalance using class_weight='balanced' — preventing the model from simply predicting "No Churn" every time

4. Model Building & Comparison

Model	Accuracy	Churn Recall
Logistic Regression	74.7%	82%
Random Forest (baseline)	~79%	~76%
Tuned Random Forest	~80%	85%

5. Hyperparameter Tuning

Used GridSearchCV with 5-fold cross-validation, optimizing for Recall (not accuracy) because missing a churner costs more than a false alarm.

param_grid = {
    'n_estimators': [100, 200, 300],
    'max_depth': [5, 10, 15],
    'min_samples_split': [2, 5, 10]
}

6. Evaluation

• ROC-AUC Score: 0.86 — excellent class separation
• Plotted ROC Curve and Confusion Matrix
• 5-Fold Cross-Validation for reliable performance estimate

---

📈 Key Results

• ✅ Final model: Tuned Random Forest
• ✅ ROC-AUC: 0.86
• ✅ Churn Recall: 85% — model correctly identifies 85% of actual churners

Top 5 Features Driving Churn:

1. Tenure (longer = less likely to churn)
2. Total Charges
3. Monthly Charges
4. Contract Type — Month-to-Month customers churn most
5. Internet Service — Fiber Optic users show higher churn

---

💡 Business Recommendations

1. Incentivize long-term contracts — offer discounts to move month-to-month users to 1 or 2-year plans
2. High-value customer monitoring — proactively reach out to customers with high monthly charges
3. Tenure-based loyalty rewards — implement offers at key tenure milestones where churn risk spikes
4. Fiber Optic service audit — investigate whether high churn in Fiber Optic is a pricing or quality issue
5. Predictive retention team — use model scores to flag high-risk customers for the customer success team

---

🖥️ Streamlit Web App

Built an interactive web app where anyone can enter customer details and get an instant churn prediction with probability score and suggested retention actions.

👉 [Click here to try the app] (https://your-link.streamlit.app)

---

📁 Project Structure

customer-churn-prediction-ML-Python/
│
├── app.py                                      # Streamlit web application
├── Customer_Churn_Prediction_Improved.ipynb    # Full ML notebook
├── customer_churn_model.pkl                    # Saved trained model
├── scaler.pkl                                  # Saved StandardScaler
├── requirements.txt                            # Python dependencies
└── README.md                                   # Project documentation

---

📬 Connect

GitHub: Shahnawaz-analytics

Linkedin: Shahnawaz-khan