---

📞 Telecom Churn Prediction

This repository contains code and models for predicting customer churn in the telecommunications industry. 📊 The project uses machine learning and statistical techniques to identify customers at risk of leaving.

---

🧠 Summary

The eda.ipynb notebook performs an in-depth Exploratory Data Analysis (EDA) on a telecom churn dataset to uncover:

• 📆 Service Agreements: Month-to-month customers churn more than those on annual/bi-annual plans
• ⏳ Tenure: Newer customers show higher churn risk
• 💸 Charges: High monthly charges and sudden increases are churn signals
• 📡 Services: Fewer service add-ons (e.g., no internet/security) = higher churn
• 💳 Payment Methods: Customers using electronic checks churn more frequently

---

🚀 Demo

▶️ View the Streamlit App

⚡ Quick Start:

1. 🔹 Go to the Demo Data tab
2. 🔹 Click Run Prediction
3. 🔹 (Optional) Click Evaluate Predictions

🧍 Manual Prediction:

1. 🟢 Go to the Manual Input tab
2. 🟢 Fill in customer details
3. 🟢 Click Predict Churn

📁 Upload Your Data:

1. 📤 Go to the Upload Data tab
2. 📤 Upload your CSV file
3. 📤 (Optional) Upload true labels for evaluation

---

📈 Model Performance During Training (AUC-ROC)

⚙️ Model	🎯 CV Mean AUC (± std)	🧪 Test AUC
🐱 CatBoost	0.8511 ± 0.0146	0.8472
🐱 CatBoost (tuned)	(best trial) — 0.8385	—
⚔️ XGBoost	0.8503 ± 0.0147	0.8481
⚔️ XGBoost (tuned)	(best trial) — 0.8407	—
💡 LightGBM	0.8482 ± 0.0154	0.8506
💡 LightGBM (tuned)	(best trial) — 0.8383	—
🧠 Stacking Ensemble	—	0.8491

---

📊 Streamlit App Evaluation

🟦 CatBoost

• ✅ Accuracy: 81.23%
• 🔢 Confusion Matrix:

	Predicted No	Predicted Yes
Actual No	4,697	477
Actual Yes	845	1,024

---

🟨 XGBoost

• ✅ Accuracy: 81.58%
• 🔢 Confusion Matrix:

	Predicted No	Predicted Yes
Actual No	4,733	441
Actual Yes	856	1,013

---

🟩 LightGBM

• ✅ Accuracy: 81.56%
• 🔢 Confusion Matrix:

	Predicted No	Predicted Yes
Actual No	4,725	449
Actual Yes	850	1,019

---

🟩 Stacking Ensemble

• ✅ Accuracy: 81.16%
• 🔢 Confusion Matrix:

	Predicted No	Predicted Yes
Actual No	4,776	398
Actual Yes	929	940

---

📂 Dataset Description

This dataset includes 7,043 telecom customers with 21 features: demographics, service usage, account info, and churn status.

🧾 Feature Overview

🏷️ Column	📝 Description
customerID	Unique customer ID
gender	Gender: Male or Female
SeniorCitizen	1 = Senior citizen, 0 = Not
Partner	Whether the customer has a partner
Dependents	Whether the customer has dependents
tenure	Months with the company
PhoneService	Phone service subscription
MultipleLines	Has multiple phone lines
InternetService	DSL, Fiber optic, or None
OnlineSecurity	Has online security add-on
OnlineBackup	Has online backup add-on
DeviceProtection	Has device protection add-on
TechSupport	Has technical support add-on
StreamingTV	Has streaming TV
StreamingMovies	Has streaming movies
Contract	Contract type: Month-to-month, One year, Two year
PaperlessBilling	Uses paperless billing
PaymentMethod	Payment type (e.g., Electronic check)
MonthlyCharges	Monthly bill amount
TotalCharges	Total charged amount (as string; needs conversion)
Churn	Whether customer churned: Yes or No

---

🔍 Exploratory Data Analysis

1️⃣ Descriptive Analysis

• 📦 Categorical feature review
• ❌ Duplicate detection
• 🔍 Unique value profiling

2️⃣ Data Wrangling

• 🧼 Handle missing values
• 🧱 Feature binning

3️⃣ Univariate Analysis

• 🧪 Normality tests

  • D’Agostino-Pearson
  • Anderson-Darling

• 📊 Individual visualizations

4️⃣ Bivariate Analysis

• 🔗 Correlations:

  • 📉 Numerical vs Numerical (Spearman)
  • 📋 Categorical vs Numerical (Kendall’s Tau, Mann-Whitney U)
  • 🔘 Dichotomous (Phi Coefficient)
  • 🔀 Categorical vs Categorical (Chi-Square, Cramér’s V, Uncertainty Coefficient)

• 🔄 Collinearity Checks

• 🖼️ Feature Pair Visualizations

6️⃣ Multivariate Analysis

• 🧠 Multicollinearity
• 📊 Frequency Distribution

7️⃣ Feature Engineering

• 🔥 One-Hot Encoding

8️⃣ Data Preparation

• ✂️ Train-Test Split
• 🧮 Encoding + Scaling

9️⃣ Model Training

• 🐱 CatBoost
• ⚔️ XGBoost
• 💡 LightGBM
• 🧠 Stacking Ensemble

---