Churn prediction estimates which customers are likely to stop buying or cancel, so you can act before they go.
Define Churn Carefully
For subscriptions, churn is a cancellation. For non-contractual businesses, you must define it — for example, "no purchase in 90 days". The definition shapes everything that follows.
Set Up the Prediction
Choose a prediction date and a window: using data up to 1 March, will the customer churn in the next 60 days? Build training data the same way at several past dates, using only information available at each date.
Useful Features
- Recency, frequency and value of activity.
- Changes in usage over time (declines are strong signals).
- Support tickets and complaints.
- Contract details, tenure, price changes.
- Engagement with emails or the product.
Modelling
Gradient boosting and logistic regression are common. Evaluate with precision and recall on held-out time periods, and check calibration if you'll use predicted probabilities.
From Prediction to Action
A churn score is only useful if it changes decisions:
- Prioritise outreach by risk and customer value.
- Match interventions to reasons for risk.
- Measure impact with a control group, because some at-risk customers would have stayed anyway.
Beware Unintended Effects
Contacting some customers can remind them to cancel. Test retention actions with experiments before scaling them.