Churn prediction
Churn prediction uses customer data to estimate who will stop buying, cancel a subscription or leave a service within a defined period. Businesses use these estimates to focus retention efforts before customers leave.
How it works
First, define churn and the prediction window: for example, a paying customer canceling within the next month. Then use earlier data, such as declining usage, failed payments or support history, to estimate each customer's risk.
Only use information available when the prediction would be made. A cancellation request received afterward would make a model look better in testing without helping you act earlier.
For a subscription startup, the practical question is whether these scores help save customers at a cost that makes sense.
Flagging 100 customers before 50 cancel
A hypothetical business starts the month with 1,000 paying customers. Its model flags 100 as high risk. By month-end:
| Model's assessment | Customers | Customers who canceled |
|---|---|---|
| High risk | 100 | 30 |
| Lower risk | 900 | 20 |
| Total | 1,000 | 50 |
The flagged group contains 30 ÷ 50 = 60% of all cancellations. That is the model's recall.
But only 30 ÷ 100 = 30% of flagged customers canceled. That is its precision. The team would contact 70 customers who would not otherwise have canceled.
Overall churn was 50 ÷ 1,000 = 5%. Predicting that nobody would leave would be 95% accurate, yet miss every cancellation. Accuracy alone is a poor way to judge this model.
Questions people ask
Can AI predict churn?
Yes, machine-learning models can estimate churn risk, but they cannot identify future cancellations with certainty. Compare them with simple rules and test on later customer data. Separately, test retention outreach against a control group: predicting risk does not prove an intervention prevents churn.
What does 10% churn mean?
For customer churn, it means 10% of the customers present at the start of a stated period left during that period. If 20 of 200 starting customers cancel in a month, monthly customer churn is 10%. Revenue churn measures lost revenue instead, so specify which metric you mean.
