When not to use Prophet for in-app usage

Prophet is a capable seasonal tool. It is also the first library many teams reach for when a PM says “just forecast DAU.” That reflex has wasted more Bangkok evenings than we can politely count.

We teach it in Retention Horizon only after a paper seasonal baseline exists. If the baseline already explains weekday and holiday structure, Prophet has to earn the extra narrative. Often it cannot.

Short, bruised series

Six to ten weeks of daily actives, with two campaigns and a store feature in the middle, is not a gentle yearly cycle. Prophet will still produce a confident ribbon. The ribbon is not a crime; presenting it as a 90-day staffing plan is. For in-app usage we rarely trust a horizon longer than four weeks unless the series is long, stable, and free of unnamed events.

When the calendar is local

Default holiday lists miss Songkran, regional election days, and the specific payday your payroll actually uses. If you do not inject those, the library will treat a festival trough as a trend break. You can fix that inside Prophet. You can also fix it with a simpler weekday-plus-holiday model that a designer can interrogate without reading a parameter dump.

A practical rule

Use Prophet when you have a long daily series, a holiday calendar you actually maintain, and a need to separate trend from seasonality for a memo that will be re-run monthly. Skip it when you are still arguing about event names, when half the days are campaign days, or when the only audience is a weekly product review that needs a 14-day number.

The grocery Atelier we describe on the reviews page declined Prophet on purpose. Their evening prior was wrong; the library would have polished the wrong prior. That is the failure mode we care about.