A customer success team builds a churn prediction model, labeling every customer in the training data as either "churned" or "active," and training a standard binary classifier on that label. This setup contains a specific, easy-to-miss statistical problem: every customer currently labeled "active" is being treated as a confirmed non-churner, when the honest fact is that some meaningful share of them simply haven't churned yet and may well do so later — their true outcome isn't actually known at the time the model is trained.
What right censoring actually means here
Right censoring, a term from survival analysis, describes exactly this situation: an outcome (churn) that hasn't happened yet for some observations by the time the data is analyzed, meaning the true final outcome for those observations is genuinely unknown, not simply unfavorable. A customer who signed up two weeks before the model was trained and hasn't churned yet isn't evidence of a loyal customer — they're evidence of a customer who hasn't had time to churn yet, which is a fundamentally different piece of information than a customer who's been active and stable for two years.
A binary classifier can't represent this distinction, and that's the actual problem
Forcing every active customer into a single "not churned" label collapses two very different situations — a long-tenured, genuinely stable customer, and a brand-new customer who simply hasn't had the opportunity to churn yet — into the same training label, as if they carried equally strong evidence of loyalty. A model trained this way tends to systematically underweight the risk associated with recent signups specifically, because their short, uneventful tenure so far gets treated as equally strong evidence of stability as a long-tenured customer's much longer uneventful tenure, when it clearly isn't.
Why this distorts the model's actual predictions in a specific, consequential direction
A model trained on right-censored data this way will tend to underestimate churn risk generally, because a meaningful share of the customers labeled "not churned" in the training data are functionally mislabeled — they simply hadn't churned yet by the time the snapshot was taken, not because they were genuinely safe. This bias is worse the more the customer base has grown recently, since a larger share of the "active" label in that case is made up of customers who haven't had much time to reveal their true churn behavior at all.
What actually handles this correctly
Survival analysis methods — the Kaplan-Meier estimator, Cox proportional hazards models, and related approaches — are specifically built to incorporate right-censored observations correctly, treating a customer's censored (not-yet-churned) status as partial information about survival time so far, rather than forcing it into a false binary outcome label. These methods produce a churn risk estimate that appropriately accounts for how long a customer has actually been observed, rather than treating a two-week-old account and a two-year-old account as equally strong evidence of "not churning," which is exactly the distinction a naive binary classifier collapses.
What this means for building or evaluating a churn model
- Check whether the model treats every non-churned customer as an equally confirmed non-churner, regardless of tenure — that's the specific sign of the right-censoring problem
- Prefer survival analysis methods over naive binary classification when a meaningful share of the customer base is relatively new or otherwise has limited observed tenure
- Be specifically cautious of a churn model's risk estimates for recently acquired customer cohorts, since that's where the right-censoring bias has the most room to distort predictions
- Re-frame churn as a time-to-event problem (when, if ever, does this customer churn) rather than a fixed-outcome classification problem where possible
None of this means binary churn classifiers are useless — many produce genuinely useful signal despite this bias. It means the bias is specific, predictable, and correctable, and a model that ignores it is quietly understating risk in exactly the customer segment — the newest ones — where getting the estimate right often matters most.