Walk me through how you would build a churn model for a subscription product from scratch.

Define churn and the prediction horizon first, since the label determines everything else. Build a point in time feature set from behavior before the cutoff, split by time, start with a regularized logistic or gradient boosted baseline, and evaluate with precision at whatever retention capacity exists. Then design the intervention alongside the model, because a score nobody acts on has no value.

I start with the label, because churn is not obvious. For a monthly subscription it might be failed to renew at the next billing date; for a usage product it might be no activity for thirty days. I also fix the horizon: are we predicting churn in the next thirty days or the next quarter? Those are different models with different interventions. Then I build features as of a cutoff date, using only behavior before it, so engagement trend, support tickets, payment failures, seat count changes, tenure. Point in time correctness is the part I am most careful about. I split by time, training on earlier cohorts and validating on later ones. Baseline is regularized logistic regression so I have something interpretable, then gradient boosting to see how much headroom exists. For evaluation I ask the retention team how many customers they can contact a week and report precision in that top slice rather than overall AUC. And I insist on designing the intervention with them, ideally with a holdout, because otherwise in six months nobody can tell whether the model saved a single customer.