Start Here

What you are looking at

Imagine a restaurant that has to decide each morning how much of each dish to prepare. Prepare too little and it runs out; too much and it throws food away. Several computer models try to guess, day by day, how much of each dish will sell. This site lets them compete on a month of invented sales, and shows which one guessed closest. A café, a petrol station, a travel agency or a hotel faces the same question, and the site runs the same tournament for each.

A fair test

Each model guesses one day at a time, and may look only at the days before it. It never sees the day it is guessing, or any day after. Once the day is over, its guess is set beside what really sold. Every model gets the same days and the same information, so none has an unfair start. Some businesses plan further ahead; there, the guess is made several days before.

How far off

On 7 December 2025, Arancini sold 28. SARIMAX + weather guessed 29.95. It was off by 1.95, which is 7.0 % of what sold: that is its MAPE for that item on that day. A model’s figure for a day is the average over its items. 0 would be a perfect guess, and lower is better. It is a percentage, not a mark out of 100: a guess of three times what sold is off by 200 %.

Reading the table

The table on the home page has one row per day and one column per model. Each square is one model’s MAPE for that day, and its colour only repeats the number. The tomato corner marks the model that came closest that day. Here are three days and three models from the latest run:

A square that reads ND is blank on purpose: there was nothing to compare that day, for example because the business was closed.

Two winners, and blank squares

There are two kinds of winner. The day winner is the model that came closest on one day, and it can change from day to day. The run winner has the lowest average over all the days scored, and needs a score on at least 80 % of them. If two models tie, the one listed first wins. The simplest model, Same weekday, guesses each item as its average on the same weekday over the last four weeks: it is the one every other has to beat.

• ND · nothing to compare: there were no sales to compare with, so there is no answer. It is never a zero, and it is left out of every average.
• fail: the model broke, or gave no answer that day. That is about the model, not the day.
• short history: the model did not yet have enough past days to make a guess.

Your business

This site shows a restaurant, a café, a petrol station, a travel agency and a hotel, but the method belongs to none of them. It fits any business that sells countable things every day: portions, litres, rooms, tickets, appointments. A good forecast means less food thrown away, the right staff on the right day, and stock that does not run out. To see it on your own sales, try it with your own file: it runs in your browser, and the file never leaves it. If you would like a full tournament on your sales, with every model, get in touch on LinkedIn.

• The method, in full
• Where the sales come from
• Your own sales