The tour · stop 1 of 7 · forecasting
Sales are what happened. Demand is what wanted to happen.
The morning buns sold out at 9 a.m. A spreadsheet writes down “26 sold” and quietly learns to order fewer. Butter knows you 86'd them — straight from your Toast stock feed — and works out what the day actually wanted. That one difference compounds every single week.
Fourteen days out,
per item, per day.
Every item you sell gets its own forecast for every one of the next fourteen days — grouped the way your own menu is grouped, with the sky and the rain odds printed on every column. Hover any number and Butter shows its work.
Your weekday rhythm, freshly weighted
Recent Tuesdays count roughly twice as much as Tuesdays a month ago, so a growing café's forecast grows with it — and a softening item gets caught early instead of six weeks later.
Sell-outs, un-censored
A confirmed 86 (from live Toast stock, never guessed from sale timing) gets its demand recovered: sold out with 40% of the morning left, and Butter adds back what that 40% usually sells.
Events that carry receipts
Farmers market Sundays, the street fair, National Coffee Day — each event's uplift is learned from your own history. “Seen 4× · still learning” becomes “confident,” and a venue 3.2 miles away gets honestly ignored.
School's out — Butter noticed
It finds your local district's calendar on its own and folds in the days that move cocoa and croissants: breaks, early release, the first day back.
Closed days can't lie
Holidays, POS outages, and sync gaps are excluded from the math entirely, so a day you were shut never drags a weekday average down.
Statistics, quietly
Trend-damping, shrinkage toward the weekday norm, intermittent-demand handling for the once-a-week movers — the textbook stuff runs under the hood; you just see a number you can trust.
Sales history · the forecast's report card
A forecast you can't grade is just a horoscope.
Every past day is scored against what Butter predicted — a calendar where each date is graded green, amber, or rose, an accuracy line that shows whether the model is getting smarter, and a bias readout that says things like “real demand tends to beat the forecast — we can raise pars.” Click any day for the deep-dive: forecast vs. actual, revenue, and the biggest misses by item.
accuracy per daybias, in plain englishbiggest misses per itembacktested history
Honesty footnote: the backtest re-runs the model as of each past day using only what it knew then — no peeking. It's also how we caught our own mistake: we retired the old weather multiplier on a test that read 0.000 — then found that test's rain history had been silently zeroed by a sync bug, and that it only ever scored same-day predictions. The rebuilt rig reproduces live accuracy to within a point (83.8% offline vs ~84% live, measured August 2026), keeps a weather-off control forever, and weather now earns its way in per category, per cafe, from observed history — engaging on evidence, never by assumption.
Invite-only beta · $99/mo per café when we open
Stop averaging last month. Start forecasting next week.
Get on the list —
We write to every café on the list when it's their turn · no card, no contract