The Problem
Customer churn costs subscription businesses 20-30% annually. Acquiring new customers is 5-25x more expensive than retaining existing ones. Without prediction, retention efforts are reactive and unfocused.
Why ML
Rule-based approaches miss complex patterns. Machine learning can identify which combination of factors predict churn, enabling targeted interventions before customers leave.
Key Findings
1
Short Tenure = Highest Risk
New customers (less than 6 months) are most likely to churn. Each additional month significantly decreases churn probability.
2
Month-to-Month Contracts
Customers without annual commitments show dramatically higher churn. Two-year contracts reduce churn risk by 42%.
3
Service Bundling Increases Stickiness
Customers with fewer services more likely to churn. Streaming services and multiple lines increase retention.
4
Premium Service Paradox
Fiber optic customers churn more despite higher prices. Premium customers need differentiated retention strategy.
Feature Importance
Confusion Matrix
Business Impact
| Metric | Value | Implication |
|---|---|---|
| At-Risk Customers Identified | 212 | Proactive retention possible before cancellation |
| Annual Revenue at Risk | $127,200 | Based on $50/month average customer value |
| Retention vs Acquisition Cost | 10-25x | $10-20 retention campaign vs $200-500 acquisition |
Technical Approach
| Component | Technology | Rationale |
|---|---|---|
| Model | Logistic Regression | Chose interpretability over marginal accuracy gains. Stakeholders need to understand which factors drive churn to take action. |
| Data Processing | pandas, StandardScaler | One-hot encoding, feature scaling for normalized inputs |
| Evaluation | scikit-learn | Precision, recall, F1-score, confusion matrix |
| Environment | Google Colab | Jupyter notebook for interactive development |
| Development | Claude AI | Technical collaboration for rapid implementation |
From Insight to Action: Businesses can address potential churn proactively based on machine learning model's identification of which combination of factors predict churn, enabling targeted interventions before customers leave.
- Implement 90-day onboarding for new customers (save ~$200K annually)
- Offer incentives to convert month-to-month to annual contracts (reduce churn from 27% to 22%)
- Cross-sell streaming services to increase customer lifetime value by 20%
What I Learned
- 💡 Interpretability over complexity. CS teams need the "why" behind churn risk, not just a prediction. A model that explains beats one that doesn't.
- 💡 The model's value isn't the 81% accuracy. It's identifying the factors that drive churn, enabling targeted retention programs.
- 💡 Model outputs need translation to drive action. "Contract coefficient: -0.55" becomes "two-year contracts reduce churn by 42%."