Kaizen · Sales forecasting

One looks at 1 thing. Ours looks at 72.

Both try to guess what your restaurant sells tomorrow. We scored them against each other on 18,309 real days — same restaurants, same tills, no do‑overs.

Their forecast $981 off
Our forecast $848 off

How far each forecast landed from the real number, averaged per restaurant per day. A typical day here is about $5,700.

56 / 100 Days our number landed closer than theirs 18,309 days scored
44 / 100 Days we were within 10% of the real number theirs: 39 / 100
0.8% How far off‑center our forecast sits theirs: 4.9% high

That last one matters most. Their forecast sits 4.9% above the truth on an average day; ours sits within 1% of it. Over‑forecasting is the expensive direction — it staffs the floor for guests who never arrive.

Why

Theirs is one recipe. It never changes.

We worked out what the old forecast actually does by testing every plausible recipe against 20,319 of its own published numbers. It averages the last five same weekdays. That's it. That's the whole thing.

Their forecast 1 signal
Ours 72 signals — tap a group to see them

Six kinds of signal

Pick a group above and this will tell you, in plain English, what the forecast is actually reading.

The same Tuesday

What each one knows before it guesses.

Their forecast sees

The last five Tuesdays.

  • Tuesday, four weeks ago
  • Tuesday, three weeks ago
  • Tuesday, two weeks ago
  • Last Tuesday
  • The Tuesday before all of those
Ours sees all of that, plus

Everything else about that Tuesday.

  • It's Lunar New Year
  • Rain in tomorrow's forecast, and that's unusual for here in February
  • A 17,000‑seat arena eight blocks away has a game on
  • This restaurant is running 6% ahead of last year
  • The same Tuesday last year, and the two either side of it
  • How wrong we've been at this restaurant over the past two weeks

A five‑weekday average gives a snowy Tuesday, a Lunar New Year Tuesday and a sold‑out‑arena Tuesday the exact same number. It isn't a bad method. It's a method with a ceiling, and it's sitting on it.

Two details we're proud of

It isn't allowed to cheat.

Yesterday's weather report, not today's truth

It would be easy — and dishonest — to train a forecast on the weather that actually happened. Then it looks brilliant in testing and falls over in real life, because tomorrow you only have a forecast. Ours only ever sees the weather report you'd have had the day before.

“Hot” means hot for your city

75°F is a warm day in Seattle and a mild one in Katy, Texas. So instead of one fixed threshold, every restaurant is compared against its own normal for that date. Unusual‑for‑here is what moves sales; a number on a thermometer isn't.

What happens next

A recipe can't improve. This does.

It rebuilds itself every night

On every day of history it holds, for every restaurant at once. It gets better as your data piles up, without anyone touching it.

It marks its own work

Each night it checks whether it's been running high or low at each restaurant, then corrects — only partway, on purpose. We tested correcting all the way; it measurably got worse.

It doesn't overrule your GM

When a manager sets their own number, we go half and half rather than picking a winner.

model 13.7% · manager 13.9% · both 12.7%

Why you can believe any of this

Most vendors mark their own homework.

We're not asking you to take a number on trust. When these restaurants moved across, their old system's saved forecasts came with them. Ours sit next to them. The real sales are the same real sales for both — so which one was closer is arithmetic, not opinion.

The days
Only days where both forecasts exist, so neither side is judged on days the other sat out. Only days the restaurant actually traded. A holiday closure is not a forecast miss.
The measure
How far the guess landed from the real till, in dollars and as a percentage, then averaged. The plain, unflattering way.
No hindsight
Every prediction is stamped with the date it was trained up to, and every one of those dates falls before the day it predicted.
The window
18,309 restaurant‑days, August 2022 to August 2026, across a two‑brand group. Open restaurants only.
What we don't claim
This is about forecasting sales and nothing else. We haven't published a comparison on labor standards or earned hours, because we don't yet have evidence to the same standard — and we'd rather say so.

Ask us to run this on your restaurants.

The scorecard above isn't a brochure. It's a report the product generates — and we can point it at your restaurants before you commit to anything.

  1. Your sales historyAn export from your POS. The further back it goes, the better every number gets.
  2. Your current forecastsWhatever your system has been predicting. This is the half nobody else can grade — and it's what makes the comparison real.
  3. Nothing elseNo integration, no migration, nothing switched off. It's a measurement, not a trial.