Find leading indicators of long-term churn

WHAT THIS TESTS The interviewer wants to see if you can convert a slow lagging indicator into early leading signals that the team can act on while it still matters. It also probes your discipline around cohorts, leakage, and causality.

A GOOD ANSWER COVERS Start by framing the target: label users who have reached the 12-month mark as renewed or churned, ensuring everyone in the analysis had a fair chance to renew. Build a feature set strictly from the first 30 days so signals are usable early. Good features include login frequency, breadth of distinct features touched, time-to-first-value, completion of an onboarding milestone, and counts of a core habit-forming action. Compare distributions between renewers and churners using effect sizes, not just averages, and segment by plan and acquisition source to avoid Simpson-style reversals. Then validate that promising signals hold on a fresh cohort before trusting them.

COMMON WRONG ANSWERS Immediately fitting a black-box model without exploring the data, leaking features that occur after day 30, treating correlation as proof of cause, or ignoring survivorship and tenure so churned and active users are compared unfairly.

LIKELY FOLLOW-UPS How would you confirm a signal is causal rather than correlated, perhaps via an onboarding experiment? How would you operationalize the indicator into a dashboard or alert? How do you handle class imbalance and seasonality across cohorts?

ONE CONCRETE EXAMPLE Suppose churners average two distinct features used in week one while renewers average five. You hypothesize that early feature breadth drives stickiness. You confirm the gap on a new cohort, then run an onboarding experiment nudging new users toward a second and third feature. If 12-month renewal improves for the treated group, you have a validated leading indicator the team can monitor weekly instead of waiting a year.