Churn Prediction
Advanced PronunciationEN/tʃɜːrn prɪˈdɪkʃən/
Churn prediction is estimating the probability that a user will stop using a product in the near future, based on patterns in their past behaviour.
Typical inputs are recency of the last activity, usage frequency, the breadth of features touched, and friction signals such as repeated errors; the output is a probability score rather than a verdict. The main pitfall is the base rate: if only three percent of users leave each month, a model that predicts nobody leaves is 97% correct and useless, so precision and recall must be measured on the churning minority. The second pitfall is that acting on a prediction changes the outcome, after which model accuracy can no longer be measured naively.
FAQ
There is no fixed figure, and numbers quoted from other products do not transfer, because the base churn rate, the prediction horizon and the churn definition differ everywhere. The right test is comparison against a naive baseline such as anyone inactive for thirty days, followed by checking whether the retained value of the intervention exceeds its cost.