Customer Churn Prediction System

This project focuses on building a machine learning model to predict customer churn for a telecom company and developing a data-driven retention strategy. The goal is to identify customers at high risk of churning and implement targeted interventions to reduce churn rate and maximize customer lifetime value.

The project follows a standard data science pipeline:

1. Data Preprocessing & Feature Engineering: Cleaning the data, handling missing values, and creating relevant features.
2. Model Training Setup: Splitting data, scaling features, and handling class imbalance.
3. Regularization Methods: Comparing L1 (Lasso), L2 (Ridge), and Elastic Net regularization for feature selection and model stability.
4. Logistic Regression Modeling: Training a final Logistic Regression model using the best regularization approach, optimized for recall.
5. Churn Scoring Pipeline: Creating a production-ready pipeline to score new customer data.
6. Customer Segmentation: Segmenting customers based on churn risk and behavior.
7. ROI Analysis: Calculating the financial return on investment for different retention campaign scenarios.
8. Retention Playbook: Generating a detailed action plan and timeline for segment-specific retention tactics.
9. Monitoring Dashboard: Setting up a framework for tracking campaign performance and KPIs.

Data

The project uses the IBM Telco Customer Churn dataset, available here. It contains customer demographic, service, and account information, along with churn status.

Methodology

• Exploratory Data Analysis (EDA): Initial data inspection, quality checks, and target variable analysis.
• Feature Engineering: Creating 9 new continuous features from original data (e.g., average_monthly_spend, loyalty_score).
• Data Preprocessing: Handling missing values in TotalCharges, converting data types.
• Categorical Encoding: One-Hot Encoding for categorical features.
• Train/Test Split: Stratified split to maintain churn distribution.
• Feature Scaling: StandardScaler applied to continuous features.
• Class Imbalance: Using class_weight='balanced' in Logistic Regression. (SMOTE was explored but the non-SMOTE model performed best on the test set for the target metric.)
• Regularization: Comparing Lasso, Ridge, and Elastic Net Logistic Regression to prevent overfitting and improve interpretability.
• Model Selection: Choosing the Elastic Net Logistic Regression model with C=0.1 and l1_ratio=0.5 based on superior recall (86.9%) on the test set and interpretability benefits.
• Threshold Optimization: Identifying an optimal probability threshold to maximize recall (achieving ≥85% target).
• Prediction Pipeline: Developing a script to load new data, preprocess, engineer features, scale, and generate churn probabilities and predictions.
• Customer Segmentation: Defining strategic segments based on churn probability and risk level.
• Financial Analysis: Calculating ROI for targeted retention campaigns using defined costs and success rates.
• Retention Strategy: Building a playbook with segment-specific tactics, owners, timelines, and KPIs.
• Monitoring: Creating a simulated dashboard to track campaign performance metrics.

Key Results

• Best Model: Elastic Net Logistic Regression (C=0.1, l1_ratio=0.5)
• Test Set Performance (Optimized for Recall):
  • Recall: 86.9% (Exceeded 85% target)
  • Precision: 47.7%
  • F1-Score: 0.616
  • ROC-AUC: 0.847
• Feature Selection: The Elastic Net model removed 10 out of 39 features, simplifying the model while maintaining performance.
• Feature Importance: Key churn drivers include loyalty_score (protective), InternetService_Fiber optic, revenue_velocity, PaymentMethod_Electronic check, and PaperlessBilling_Yes (risk factors).
• Customer Segmentation: Customers are segmented into actionable groups like 'Champions at Risk', 'Needs Attention', 'Potential Risk', and 'Loyal & Stable'.
• ROI Analysis: Targeted intervention by segment is projected to yield an ROI of 1632%, saving an estimated $1.15M in net benefit compared to doing nothing.
• Retention Playbook: Specific tactics, priorities, budgets, and expected outcomes for each segment.
• Monitoring: A framework for tracking campaign KPIs like contact rate, engagement rate, retention rate, and financial impact.

Elastic Net Logistic Regression is the most suitable model for this dataset, providing a strong balance of predictive performance (meeting the recall target), interpretability, and efficiency. The segmentation, ROI analysis, and playbook provide a complete, actionable solution for proactively managing customer churn.