Cell2Cell, a major telecom company in the U.S., was facing a pressing marketing challenge: high customer churn. The company realized that while it was spending heavily on acquiring new users, it was losing a significant number of existing ones each month—cutting into its revenue and long-term customer value.

This project is focused on solving the issue using marketing analytics and predictive modeling. I aimed to:

• Predict which customers were most likely to churn
• Understand why they were likely to churn
• Design targeted, data-driven retention strategies to proactively reduce churn and increase loyalty

This was not just a data science exercise—it was a full-fledged marketing strategy project grounded in customer insights and ROI-focused decision-making.

I used a mix of predictive modeling and segment-based marketing strategy to tackle the churn problem. Here's how I approached it:

1. Data Preparation

I worked with a rich dataset of 71,047 customers, split into:

• A calibration set with ~50% churners to train the model more effectively
• A validation set that mirrored real-life churn (~2%) for unbiased testing

The dataset included 75+ variables like equipment usage, tenure, customer service interactions, and pricing—typical features marketers could act on.

2. Modeling Techniques

I developed and compared two models:

• Logistic Regression: Interpretable and useful for explaining churn drivers
• Decision Tree: More flexible, better at capturing nonlinear relationships

The decision tree model outperformed logistic regression in both recall and lift, making it the better option for identifying churners in a marketing context where missing a potential churner is more costly than targeting a loyal customer.

3. Evaluation Metrics

I used business-relevant metrics:

• Recall (finding churners): 65.6%
• Lift in Top Decile (targeting efficiency): 1.33
• Accuracy: 60.03%

These metrics helped justify model effectiveness to marketing stakeholders.

The analysis revealed powerful behavioral and financial signals that indicated churn risk:

Factors that increased churn risk:

• Retention calls: Customers who had more interactions with support were 76.6% more likely to churn—often a signal of frustration
• Older phones: 36.5% higher churn odds—many users left due to outdated hardware
• Overage charges: 12% higher churn odds—pricing dissatisfaction was a key trigger
• Multiple subscriptions: Customers with multiple plans were more likely to leave, likely due to complexity or high costs

Factors that reduced churn risk:

• Longer tenure: More loyal customers showed lower churn (–19.2%)
• Responses to mail offers: Suggesting they were engaged with the brand (–21%)
• Low credit score: These customers were less likely to switch due to fewer alternatives

These insights helped shape targeted messaging and personalized offers in our retention campaign.

Challenge: Marketing Waste from Targeting the Wrong Customers

A naive strategy would be to offer incentives to all customers predicted to churn—but what if they wouldn't have churned anyway? This leads to wasted marketing dollars.

Solution: I used lift analysis to focus only on the top decile—customers who were 75%+ more likely to churn than average—making our outreach both effective and cost-efficient.

Challenge: Complexity in Communicating Model Outcomes

Predictive models can be technical, but our stakeholders were marketers.

Solution: I chose interpretable models, visualized results through decision trees, and translated insights into simple, actionable personas and offers.

This project delivered measurable business impact:

• Churn dropped from 2% to 1.9%
• Customer lifetime value (LTV) increased by $4.39M
• Return on investment (ROI): 757%

I learned that:

• Model precision is great—but recall is king in retention marketing
• Simplicity in communication drives stakeholder buy-in
• Combining marketing intuition with data science leads to smarter, scalable campaigns

Retention Campaign Examples

High-Risk Customers (Churn Probability > 75%)
Offer: Phone upgrade + personalized outreach
Impact: 30–40% drop in churn risk

Low-Risk Customers (Churn Probability < 20%)
Offer: Simple loyalty perks or upsell opportunities
Impact: Stable churn, improved satisfaction

Tools & Techniques

• Python (pandas, scikit-learn, matplotlib)
• Logistic Regression, Decision Trees
• Confusion Matrix, Precision-Recall, Lift Analysis
• Excel Simulators for ROI modeling
• Segmentation & offer targeting based on churn scores