Field note · April 27, 2026
When a cohort split hides a seasonal pattern
Join-week cohorts are honest only if the calendar around them is named. Holiday closures and school calendars can look like churn if nobody annotates the weeks.
Cohort reporting likes tidy weeks. Real calendars do not. A studio that closes for a school break, a shop that staffs differently in August, or a membership that pauses in the week after a holiday will show a dip that looks like people leaving for good.
The usual mistake is to split joiners by week and then compare those weeks as if each week offered the same chance to return. If week 52 included three closed days, the return count is a count of open days, not a verdict on the people.
A practical fix is an annotation lane: a second strip under the table that names closures, campaigns, and known staffing gaps. Bundlelogicbase treats that lane as part of the deliverable, not as a slide someone might add later.
When two seasons must be compared, line up the same calendar landmarks rather than the same week numbers. Week 14 in one year may be after Easter; in another year it may still be winter. Customer return follows the local calendar more closely than it follows an ISO week label.
If you cannot reconstruct last year’s closures, say so. A seasonal pattern that cannot be checked should not be used to justify cutting a reminder or a class. Silence in the source files is a finding.