Subscription Churn Forecast

A software company starts a month with 1,000 paying subscribers, and if it expects 50 of those customers to cancel, its forecast customer churn for that starting group is 5%. It should show whether the projection counts only completed cancellations or also notices that will take effect later.

Define the question first: a customer-count forecast informs support staffing and acquisition targets, while a recurring-revenue forecast informs cash and growth plans, and losing one large account may matter more to revenue than losing many small accounts. Use a consistent cohort and period, because monthly subscribers and annual subscribers face renewal decisions on different schedules, and blending them without adjustment can make a month with many annual renewals look unusually risky.

Stripe's subscriber-cohort guidance groups customers by starting period and tracks their retention over time, which shows whether recently acquired customers behave differently from customers who have stayed for years. Gather several periods of clean history, recording each subscriber's start date, plan, price, renewal date, cancellation date and status, and separate voluntary cancellation from failed-payment churn when possible, because the remedies differ.

The simplest model applies a historical cancellation rate to customers who are exposed to renewal: for 200 accounts renewing next month and an expected 8% cancellation rate, the estimate is 16 lost accounts, which is not a forecast of the entire base unless every customer could leave during that window. A cohort model can use different survival rates for new and mature customers, since first-month users may cancel more often than users in their second year, so forecast each group and add the results rather than imposing one average on all.

An illustrative formula for customer churn is expected cancellations from the opening active cohort divided by opening active customers, so if 30 of 600 opening customers leave, forecast churn is 5%, and newly acquired customers should not be added to the denominator midway through the period. Revenue forecasts need more than counts: multiply expected lost accounts by their relevant recurring contract value, and account separately for downgrades, upgrades, usage, new sales and price changes, since Stripe identifies churn, expansion and contraction among the drivers of SaaS revenue forecasts.

Distinguish gross and net revenue effects, because a company can lose subscriptions and still grow its recurring revenue if remaining customers expand enough, which does not erase the cancelled customers or make churn zero. Choose the measurement date carefully, since a cancellation notice may arrive today while paid access continues until the end of a term, and define whether the model predicts notices, service exits, or recognised recurring value at the renewal date.

Check seasonality and one-off events, because holiday contracts, school-year renewals or a known product shutdown can shift churn, and a historical average is less useful when the customer mix or service terms have changed. Build scenarios rather than a false-precision single number, with base, lower-loss and higher-loss assumptions showing the range of cash and staffing needs, and state which assumptions a manager can influence and which are outside the business's control.

Test forecasts against actual outcomes by comparing predicted and observed churn by cohort, product and billing term, because a forecast that repeatedly understates losses needs recalibration, not just a more optimistic explanation. Watch for misleading data, since a migration between plan IDs may appear as one cancellation and one new sale even when the customer never left, and trial endings and paused accounts should follow explicit inclusion rules.

The forecast should lead to decisions such as earlier customer-success outreach, a cash buffer or a smaller hiring plan, and interventions should be evaluated with actual retention results, not assumed effective because the model improved. Avoid treating a probability as a verdict on a person, using higher-risk segments for helpful service improvements rather than intrusive pressure; for an owner, the useful question is not only how many will leave but what range of loss to plan for, why, and what can be learned when actual renewals arrive, so keep the model updated as those facts change.