Score every customer for churn risk, fitted to your own book.

Build me a customer health / churn-risk model from my Chargebee data, fitted to
my own churn patterns — not a generic template. Work through the phases in
order, show your reasoning, and never present an assumed number as a measured one.

PHASE 1 — Establish the real population.
· Call list_subscriptions with status = active. Paginate until next_offset is
  null. Treat a customer as PAYING only if subscription mrr > 0 — exclude
  $0 / trial / free subscriptions.
· Call list_subscriptions with status in (cancelled, non_renewing) to get the
  previously-paying customers who have since churned.
· Report both counts. If churned paying customers < ~50, say plainly the
  sample is thin and downstream results are directional only.
· Compute annual contract value per customer from plan_unit_price × billing
  frequency (+ addons). Segment: Startup <$15K, Scaleup $15K–$70K,
  Enterprise >$70K.

PHASE 2 — Build candidate signals (compute for active AND churned).
· Contract commitment: from subscription billing_period / billing_period_unit
  (annual vs quarterly vs monthly).
· Payment distress: export_invoices for the trailing 12 months; read
  dunning_status, and pull transactions / list_transactions for failed payments.
· Invoice trend: from export_invoices, compute invoice value and count per
  trailing quarter — flag whether each is rising, flat, or falling.
· Tenure: from subscription created_at / activated_at to today.
· Ramps: list_subscription_ramps with subscription_id[in] batched across the
  segment to flag negotiated escalation deals.
· For each signal, state how many customers you could actually compute it for
  vs. had to treat as neutral.

PHASE 3 — The learning engine. Fit, do not assume.
· For each candidate signal, measure how strongly it separates churned from
  active customers IN MY DATA (report correlation / churn-rate difference, the
  sample size n, and a significance test).
· Drop signals that do not discriminate. Exclude any CIRCULAR signal (one that
  is really the churn event itself, e.g. an action-at-term-end = cancel flag).
· Derive weights from the data via a logistic fit — not intuition. Show each
  signal's weight and the resulting model correlation.
· Check for sampling bias: confirm you are on the FULL paying book, not a
  plan-tier-filtered slice that skews toward large / tenured accounts, which
  inflates every correlation. State which population each statistic comes from.

PHASE 4 — Report and build.
· Give the final weighted parameter list (6–8 strongest signals), the bands
  (Healthy / Watch / At Risk), and the precision/recall tradeoff at a few score
  thresholds — with the threshold you would use for CS triage.
· If billing signals alone are limited, say which usage/consumption signals
  would lift the model, and why.
· Build a single-file interactive HTML dashboard: score distribution and health
  breakdowns by segment, plan, geography, and industry, scored with the fitted
  model. Hover on a band shows the customer count.

If you ever find yourself reframing my request to make a number look better, stop
and tell me instead.