Bangkok product teams and the late-night usage spike

Imported dashboards often arrive with a polite evening peak around 19:00. That shape is true for some markets. It is a poor default for Bangkok. Commutes run late. Dinner runs later. Grocery, video, and wallet opens keep climbing after many “global” priors have already gone to sleep.

We see the miss most clearly in staffing and in push timing. A 7-day volume forecast that underweights 21:30–23:30 will look accurate in the daily total and still leave night pickers short. A campaign that fires at 18:30 will congratulate itself on open rate while missing the hour people actually have a free hand on the BTS.

Do not average the city into a timezone

Thailand is one timezone. Behaviour is not one timezone. Shift workers, students after evening class, and office staff who reach home after 22:00 do not share a single “night.” If your event stream can hold an hour stamp, plot hours before you plot days. The Thai Market intensive in the curriculum exists mostly for this hour-level reading.

What to put in the prior

At minimum, keep separate factors for weekday evenings and Sunday evenings. Songkran and major retail festivals need their own note; they scramble both hours and days. If you only have daily totals, say so in the memo and refuse to give hour-level advice. Pretending daily data contains an hourly story is how push calendars get mystical.

A small field habit

Once a quarter, sit in the app at 22:40 on a weekday and perform the core task you forecast. Notice which screens feel slow, which payments fail, which copy still assumes daylight. Then look at the hour plot. If the spike is real and the task is painful, the forecast is not your first problem.

More on the sequence we teach: usage forecasting from app data.