Retention budgets go to the customers who matter most, and "matters most" is usually read off this month's bill. That ranks a new high-paying line that will leave at its first contract end above a loyal mid-paying household. Customer lifetime value combines two forecasts — how long a customer stays and how much they pay while they do — and the ranking changes.
Churn is modelled as a monthly hazard that depends on tenure, plan, bundle, support contacts and — most sharply — whether a contract ends that month. The product of monthly survival probabilities gives each customer's own retention curve, not a single churn score. In production this was Multi-Task Logistic Regression and a Cox model with elastic net.
Monthly revenue drifts: upgrades, downgrades, add-ons. A model of the trajectory from the customer's history and attributes replaces the flat line — a Temporal Fusion Transformer over static, past and known-future inputs in production, a regression on the same ideas here.
Expected revenue in each future month times the probability of still being there, summed over the horizon. Ranking by this number instead of by ARPU moves retention effort from lines that are about to leave anyway toward customers whose long-term value is at risk — and the holdout shows how much value that recovers.