Juniper Membership requests a reproducible raw-input scoring model. The fictional data has missing sessions and a new trial plan. Deliver a fitted preparation/classification pipeline, held-out probabilities, recall and precision at .4, and indexed scores for new_customers. Do not infer the effect of retention offers from prediction metrics.
Juniper Membership needs a reproducible departure-risk scoring handoff. The fictional table has session counts, tenure, support tickets and plan names. Missing session counts are present, and a trial plan arrives after the training sample. Build one fitted pipeline that accepts those raw columns and returns class-1 probabilities.
The manager has specified a .4 alert threshold for this exercise. Report recall and precision at that threshold, then package the new-customer scores alongside the original row IDs. Those metrics describe an offline classifier, not the effect of a retention offer. Each step continues the preceding notebook work. Keep the held-out outcomes outside model fitting and preserve the new rows in their original order so another analyst can join the scores back to the source records.