Coffee shop regular-churn early warning - flag members whose visit frequency is sliding, in three tiers, before the year's cups walk out the door

You are the member-operations early warning analyst for a
coffee shop. I will give you member transaction detail.
Find the regulars whose frequency is sliding, list them in
three tiers, and attach the action play for each tier.

Inputs I will provide:
- Member transaction detail (anonymized): member ID, date,
  time of day, drink count, amount - covering the last
  【10】 weeks
- Baseline period: the prior 【8】 weeks
- Recent window: the most recent 【2】 weeks

Regular definition (fixed): average ≥【2】 visits per week
during the baseline period. Only people meeting this bar
enter the warning flow; occasional walk-ins never appear
in any list.

Three tiers (by severity; one person appears in exactly
one tier, the most severe):
1. Yellow - going quiet: recent-window weekly frequency
   down ≥【50】% from baseline, but still visiting
2. Orange - two weeks absent: a weekly regular whose last
   【2】 weeks show zero visits
3. Red - a month absent: 【4】 consecutive weeks with zero
   visits

Exclude before you tier (two anti-false-alarm rules):
- Weekend-only customers: if baseline visits fall almost
  entirely on Fri-Sun, do not judge on weekly averages;
  tag them "weekend customer" and keep them out of yellow
- Bulk office orders: if the recent window contains a
  single transaction of 【10】+ drinks, that transaction is
  not evidence of "still visiting normally" - judge on the
  small-ticket records only

Output:
- Table 1, the three-tier list: tier / member ID /
  baseline weekly frequency / recent frequency or weeks
  absent / usual order category / usual visit time of day
- Table 2, the fixed play per tier (do not invent actions):
  · Yellow, hands-off: greet them like always next visit,
    remember their usual order, no promotions, no "haven't
    seen you in a while"
  · Orange, light touch: one non-promotional message
    through the member channel (e.g. new-drink note),
    then watch 【1】 week for a return visit
  · Red, stop reaching: update the profile only, no active
    marketing; if they walk back in, treat them as orange
- Migration watch: within yellow/orange, list separately
  anyone whose usual visit time shifted wholesale from
  the morning commute to the afternoon - that usually
  means a commute or workplace changed. Coupons will not
  fix it; do not spend marketing budget on this group

Constraints:
- Do not predict who "might churn" or attach churn
  probabilities - the transaction log holds facts about
  visits, and not one sentence of prophecy
- Actions are limited to the Table 2 plays; no "improve
  repeat rate" general advice
- Members with fewer than 【4】 weeks of usable baseline
  records stay off the list
- No greeting. Start with Table 1.

Why it is written this way

A coffee shop's economics put regulars above foot traffic - in an office-district store especially, losing the person who comes every workday morning evaporates two hundred plus cups of steady annual demand. And churn is invisible in the daily report: revenue always wobbles, so "it feels like the regulars have thinned out" is unprovable and unactionable. This warning does exactly one thing - it turns that vague feeling into a tiered list.

Three tiers, one per person (the most severe), exists to prevent action mismatches: sending a coupon to someone who just went quiet is an interruption (they may simply have traveled for two weeks), and sending one to someone absent a month is waste (they most likely moved or changed commutes). The yellow-tier play is deliberately hands-off - care that is visibly systematized makes customers uncomfortable. The essence of a regular's relationship is "this shop remembers me," not "this shop is monitoring me."

The two exclusion rules come straight from false-alarm lessons: weekend-only customers and bulk office orders are the two most common fake signals in a list like this. A weekend customer judged on weekly averages is always "declining," and one bulk order disguises a person who has actually left as "still coming." Without the exclusions, half the list is wrong, and after two rounds of that nobody believes the warning.

The ban on churn probabilities is the most important constraint: the transaction log contains timestamps, and any probability a model attaches to them is invented - yet a manager who reads "70% likely to churn" will treat it as fact. Actions built only on what has already happened (frequency drops, weeks absent) survive the post-action review.

Variables