Desk note · 2 Feb 2026
Sampling without losing signal
Casual puzzle studios in Thailand know the pattern: paid UA around Songkran, a beautiful install-week LTV, then a quiet June. Sampling that week as “too noisy” and dropping it deletes the only signal you had about gift IAP. The model was not “wrong.” It started the clock at install during a week when installs were cheap, curious, and mostly non-paying. Paying users arrived later, or not at all, and the curve still carried the holiday population in the denominator.
First paid receipt as a clock
In the Cohort LTV Workshop we rebuild the same export using first paid receipt as day zero. That excludes the holiday tourists who never opened a wallet. It also makes early LTV look smaller, which is usually closer to cash. Product teams dislike the smaller number. Finance recognises it.
You still need an install-based view for UA efficiency. That view should be labelled as cost recovery, not as customer LTV. Mixing the two on one slide is how burst calendars become “the new baseline.”
Seasonal spikes are not outliers to delete
Some analysts drop Songkran weeks as outliers. That hides a real behaviour: gift IAP, family sharing, and short sessions on shared phones. We keep the weeks and facet them. The question is not “what would LTV be without holidays?” It is “what LTV do holiday starters actually produce by day 30 and day 90?”
If holiday starters pay once and vanish, your live-ops calendar is buying noise. If they convert to a second receipt at a stable rate, the burst may be doing real work — just not the work install-week LTV claimed.
Organic versus paid inside the same holiday
Organic installs during a burst are not a control group. Brand search and word of mouth ride the same creative. Split them anyway, using the campaign parameter you already have. The split is imperfect. It is still better than a single blended curve that UA then uses to request more budget.
We do not teach MMP configuration in this essay. If your parameter is missing, fix that before you model. The workshop assumes the parameter exists.
Related reading: latency budgets for mobile SDKs. Related course: Cohort LTV Workshop.