Payday in Thailand is not evenly sprinkled across a week. Tourism weeks thicken some districts and empty others. Songkran moves both labour and leisure. Monsoon weeks delay delivery and therefore delay the “successful order” event many marketplaces use as return. None of that cares that a Silicon Valley template drew a line at 168 hours.
When a Bangkok grocery app celebrates a Day-7 lift after a free-delivery campaign, we ask three calendar questions before we look at the chart. Did the window include a payday Friday? Did it include a public holiday that closed wet markets? Did rider supply actually match the demand the campaign created? If those answers are “we did not check,” the percentage is not a retention finding.
Liquidity is part of the metric
On two-sided surfaces, a buyer cannot return if there is nothing worth returning to. Seller-side quality and rider availability are not “context.” They are part of whether the user had a chance to perform the behaviour you named. A holdout that randomises buyers while ranking still concentrates good listings in treatment is not a buyer experiment. It is a ranking leak.
Marketplace Retention Field Methods, one of our quieter programmes, spends a session redrawing the unit: listing cluster, store, or zone, depending on where leakage actually lives. The flagship course sketches this; the field-methods bench goes further.
A working alternative
We often move primary windows to a full billing cycle or to a four-week block that excludes named holiday weeks. That is slower. It is also harder to fake. Secondary metrics can still be daily for operational monitoring. The decision rule should not be.
If your executive still wants Day 7 on a slide, put it in an appendix labelled “early operational read, not the stop rule.” That sentence has saved more than one alumni team from shipping a novelty spike.
For the holiday calendar we actually teach, see module 04 of Holdout Design or write the studio from contact.