A Churn Risk Score Without a Time Window Is a Label, Not a Prediction
'High churn risk' means nothing until it answers: at risk of not buying within how long, compared with what baseline? Scores that skip the window get applied to everyone, which is the same as applying them to no one.
Ask a team what its high churn risk segment means and you will usually get a description of the score: recency stretched, frequency falling, engagement down. Ask the question the score exists to answer, how many of these customers will fail to buy in the next ninety days who would have bought with an intervention, and the room goes quiet, because the score was never built with a window in it. A churn prediction is a claim about a period. Without the period, the score cannot be tested, since 'eventually churns' is true of every customer who ever lived, and it cannot be acted on, since an intervention timed for a thirty day risk is a different spend from one timed for a year. The window is not a refinement of the score. It is the score. Everything else, the features, the model, the dashboard colour, is machinery in service of a claim that either names its timeframe or does not deserve the decisions being made from it.
Build the window from the buying cycle, not the calendar
The right window comes from the store's own purchase rhythm: roughly one and a half to two times the typical gap between orders for the segment in question. A replenishment business with a sixty day cycle can meaningfully predict ninety day churn, because a customer ninety days silent has missed a cycle and a half and the silence itself is the evidence. A considered purchase business with a nine month cycle needs a window so long that the score's practical use changes: less a save play trigger, more a portfolio view of how much future revenue is cooling. Both are legitimate. What is not legitimate is one window, usually a quarter because reporting likes quarters, pasted across categories whose cycles differ by a factor of five.
Related reading on this site: SegmentSage and A Segment Too Small to Act On Is a Fact, Not a Segment cover the neighbouring problems in more depth.
Backtest in the open or do not ship the score
A windowed score can be checked, which is its virtue. Take last year's cohort at the moment the score would have fired, and count how many actually failed to buy inside the window, against the base rate for comparable customers the score ignored. If the high risk band churned at twice the base rate, the score earns its interventions. If it churned at the base rate, the score is a recency sort wearing a lab coat, and the honest move is to use the recency sort and say so. Publishing the backtest next to the score, in the same dashboard, changes how the whole team talks about it: risk becomes a measured quantity with an error rate, instead of a colour that ends arguments.
Match the intervention to the window or waste both
A thirty day risk window buys urgency: a save offer now, priced against the order that is about to not happen. A twelve month window buys patience: better onboarding of the next purchase, category education, a reason to return that is not a discount fired in panic at a customer who buys twice a year and was never late. Teams that run one save play against every risk band spend the urgent money on the patient customers and send the patient content to the urgent ones, and conclude that churn programs do not work. The program was fine. The clock was wrong.
Frequently asked questions
Why does a churn score need a time window?
Because churn is only a meaningful prediction against a period: the probability a customer fails to buy within the next ninety days can be tested and acted on, while 'at risk' with no timeframe is true of everyone eventually and cannot guide the timing or size of an intervention.
How should a store choose its churn window?
From its own purchase cycle, typically one and a half to two times the normal gap between orders for that segment. A sixty day replenishment cycle supports a roughly ninety day window; long cycle categories need longer windows and different, less urgent interventions.
How do you validate a churn risk model?
Backtest it: score a past cohort as of a fixed date, then compare actual in window churn in the high risk band against the base rate for similar customers the score passed over. A useful score separates clearly from the base rate; if it does not, it is adding cost without information.
Should every high risk customer get a discount offer?
No. The intervention should match the window: short windows justify immediate save offers, long windows call for lower cost engagement that gives the customer a reason to return. Firing urgent discounts at long window customers mostly subsidises people who would have bought anyway.