Churn prediction

A churn model is trained on historical customers, labelled by whether they left in the period after a snapshot date, with features describing their state at that date: usage trends, support contacts, payment issues, tenure, plan, engagement with key features. It outputs a probability per customer, refreshed daily or weekly, and typically an explanation of the top contributing factors so a retention manager knows what to talk about.

Two design choices decide whether it is useful. First, the definition of churn and the window: for a subscription SaaS it might be non-renewal within ninety days; for a bank it might be balance falling below a threshold. Second, the action: a score nobody acts on is a dashboard. The best programmes pair the model with uplift estimation, so effort goes to customers whose churn an intervention can actually prevent, and measure the whole thing by A/B test.

It is not a list of unhappy customers, and it is prone to data leakage: a feature such as "cancellation request submitted" predicts churn perfectly and uselessly. Features must be restricted to what was knowable at the snapshot date, before the outcome was decided.

• Founder / CEO: for subscription businesses, retention is the cheapest growth; a model that finds at-risk accounts early is leverage on the whole revenue base.
• Support manager: the model tells you which accounts to reach proactively, and your interactions are among its most predictive features.
• Data lead: the snapshot-and-label construction is where these projects go wrong; get the time discipline right before tuning any algorithm.
• CFO: measure the programme on retained revenue against a holdout, not on the model's accuracy score.

Churn prediction exists because by the time a customer tells you they are leaving, it is usually too late and always expensive. The signals of disengagement appear weeks earlier in usage and support data, but no team can watch every account by hand. A model turns those signals into a ranked list and an explanation, so limited retention effort goes where it matters. The trade-off is that prediction is not prevention: scores must feed an action, the action must be tested for uplift, and the model must be built carefully to avoid learning from information that only exists after the decision to leave.

• A B2B SaaS company flagging accounts with falling seat activity ahead of renewal for customer-success outreach.
• A telecom or broadband provider identifying subscribers likely to port out and testing targeted retention offers.
• A bank predicting dormancy in savings accounts and prioritising relationship-manager calls.
• An ed-tech subscription predicting learners likely to lapse and triggering mentor contact.
• A D2C subscription box predicting skip-and-cancel behaviour and adjusting the next box or cadence.

Technique / practiceA classical ML technique with a strong operational component. Eazyware builds churn models under AI and ML development, with leakage-safe feature construction, explanations per account and a controlled test of the retention action.

Eazyware service that covers it: AI/ML Development. Starting prices are on the pricing page.