12 May 2026 · Sample size
Sample size planning when your weekly actives wobble
A calculator that assumes a flat 40,000 weekly actives is a polite fiction for most consumer apps in Thailand. Songkran empties some funnels and floods others. Payday weeks in Bangkok look nothing like the week after a long holiday. If you lock an MDE in January and run the same window in April, you are not being rigorous — you are being convenient.
In Experiment Design for Mobile Releases we ask students to plot thirteen weeks of the actual primary metric’s denominator, not downloads, not registrations. Then we pick the test window from the ugly parts of that plot, not the pretty mean.
Use a corridor, not a point
Instead of a single required sample, write a corridor: the smallest N you will accept if traffic is soft, and the N at which you will stop even if marketing wants “one more week.” The corridor belongs in the brief. If you only write the optimistic number, someone will extend the test until a holiday inflates variant B.
Do not average away festival weeks
Excluding Songkran entirely can be honest. Averaging Songkran into a “typical week” is not. The same goes for 11.11 campaigns and back-to-school. Name the exclusion in the memo so a future reader does not think you cherry-picked.
Thin traffic is a product decision
If your honest MDE on a paywall is 18% relative, you do not have an analytics problem. You have a surface that cannot support the ambition. Variant Atelier critiques often end with “test a cheaper surface” — onboarding copy, not annual pricing. That sentence saves more money than a fancier sequential method.