A small Minneapolis empanada story — and the data lessons of running a 200-order-a-week kitchen when every decision is a forecast.
What we learned trying to predict weekly empanada demand. Why ARIMA and Prophet underperform your gut below a certain volume, and what actually worked: a quantile-regression baseline plus a weather-and-weekend correction.
Restaurants treat inventory as last week’s usage times a cushion. That pattern costs real money. Treating every par-level as a probabilistic forecast with an explicit service-level target cut our chicken-filling waste by ~18%.
"We’ll do $40k next month" is a wish, not a plan. Why small businesses should report calibrated ranges instead of point estimates, and a three-step method any owner can run in a spreadsheet.