What is churn prediction?
Churn prediction is the use of a predictive model to score each customer on the likelihood of leaving within a defined window, such as the next quarter. It is usually a binary classification model trained on past customers, comparing those who left with those who stayed. Common inputs are falling orders or usage, late payments, support complaints, a recent price increase, and an approaching contract end date.
The first decision is what counts as churn. A subscription business sees a cancellation. Many other businesses never do, because customers drift away, so churn has to be defined as a drop in spend or a gap in orders.
Churn prediction in practice
Contractor accounts at a building-supply distributor rarely close. They move to a competitor one product line at a time. The distributor defines churn as a quarter in which an account spends less than half its trailing average. A model trained on three years of order history gives each sales rep a weekly list of at-risk accounts. Each account shows the features that drove its score, such as fewer lumber orders or a run of late deliveries.
What to watch for
- A risk score identifies who may leave but says nothing about who will respond to an offer, so test retention actions with an A/B test against a control group.
- Exclude anything recorded only after a customer decides to leave, such as a cancellation request, because it causes information leakage.
- Retrain when prices, products, or competitors change, since the reasons customers leave change with them.