Definition
Subscriber churn prediction is the application of predictive analytics to identify subscribers who are at high risk of stopping engagement — whether through unsubscribing, marking email as spam, or simply ceasing to open and click. By predicting churn before it happens, marketers can take proactive retention actions.
Churn prediction typically uses machine learning models trained on historical behaviour patterns to identify early warning signals that precede disengagement.
Churn Indicators
| Indicator | Warning Signal | Predictive Value |
|---|---|---|
| Declining open rate | Gradual decrease over 4-8 weeks | High |
| Cessation of clicks | Previously clicked, now doesn't | Very high |
| Increased unread time | Taking longer to open emails | Medium |
| Segment changes | No longer engaging with previously preferred content | Medium |
| Reduced scroll depth | Opening but not scrolling | Low-medium |
| Lower click density | Fewer clicks per open | Medium |
Building a Churn Prediction Model
- Define churn: What counts as churned for your programme (e.g., no open in 60 days)
- Gather training data: Historical behaviour of subscribers who have and have not churned
- Select features: Engagement metrics, recency, frequency, content preferences
- Choose a model: Logistic regression, random forest, gradient boosting
- Train and validate: Split data into training and test sets
- Set a threshold: Define the probability score that triggers intervention
- Deploy and monitor: Score subscribers regularly and track model performance
Proactive Retention Actions
| Churn Risk Level | Action | Timing |
|---|---|---|
| Low risk (10-30%) | Maintain current strategy, monitor | Ongoing |
| Medium risk (30-60%) | Change frequency, offer preference update | Within 2 weeks |
| High risk (60-80%) | Send re-engagement campaign | Within 1 week |
| Very high risk (80%+) | High-value win-back offer | Immediate |
| Critical (95%+) | Transition to reactivation programme | Immediate |
Measuring Prediction Effectiveness
- Precision: % of predicted churners who actually churn
- Recall: % of actual churners correctly predicted
- F1 score: Harmonic mean of precision and recall
- Lift: Improvement over random selection
- Lead time: How far in advance churn is predicted
Was this useful?
Related Glossary Terms
AI Inbox Summary
An AI inbox summary is an AI-generated digest that condenses unread email — often highlighting news, actions and senders — changing how clearly your marketing email reaches and engages subscribers.
Attention Rate
Attention rate is the percentage of email opens that last longer than 5 seconds, distinguishing genuine reads from passive opens, preview-pane views, or Apple MPP auto-loads.
Click-to-Engage Rate
Click-to-engage rate measures clicks per minute of active session, providing a deeper measure of content engagement than standard click-through rate.
Email Attention Score
An email attention score is an analytics measure estimating how much attention a subscriber gives your emails, combining factors such as recency, frequency and engagement. Learn how attention scores prioritise your list.
Email Community Building
Email community building is the practice of using email to foster connection, discussion, and belonging among subscribers, transforming a one-way broadcast into a two-way conversation.
Email Content Freshness
Email content freshness measures how recently content was created or updated and its impact on engagement, providing strategies for content refresh cadences, evergreen libraries, and performance decay analysis.
Frequently Asked Questions
For reliable predictions, you need at least 6 months of engagement data on at least 10,000 subscribers. A larger dataset with clear churn patterns produces more accurate models. Smaller lists can use rule-based churn prediction (engagement thresholds) instead of machine learning.
Yes, but with simpler approaches. Small lists (under 10,000) can use rule-based churn indicators — if a subscriber hasn't opened in 30 days and previously opened weekly, flag them for intervention. Machine learning requires larger datasets to produce meaningful predictions.
A well-built churn prediction model can achieve 70-85% accuracy in identifying subscribers who will churn in the next 30-60 days. Accuracy depends on data quality, feature selection, and the distinctiveness of your churn patterns. Models should be retrained every 3-6 months as subscriber behaviour evolves.