Last week’s average is about last week’s kitchen
A seven-day figure spanning a change you made on Wednesday is a blend of two different operations. Our speed board learned to show the current one separately, and it changed what we worked on.
For a while our speed board showed one number per page: the average over the last seven days. It was accurate and it was nearly useless, for a reason that applies to almost every rolling figure in hospitality.
We ship changes most days. A seven-day average therefore blends the version running now with four or five versions that no longer exist. If a screen was slow on Monday and fixed on Tuesday, the blended average stays bad for most of a week — and the fix is invisible in the only number anyone was reading.
What the board does now
Beside the seven-day figure sits a second one: the same average counting only the loads measured on the version that is deployed right now. Above the table is a line that reads like a verdict rather than a table — how many pages are over budget on this release, out of how many have been measured at all, and how many have not been opened since it went out.
The blended column stays, because when nobody has opened a screen recently it is the only thing there is. But it is a number about code that has been replaced, and no job gets taken from it.
The venue version of the same mistake
You changed the menu on Wednesday, or moved a vendor, or put a second till on the door. Now look at the fortnight's average spend, or the average wait, and ask what it is describing. It is describing a blend of the old layout and the new one, weighted by however many days each got. If the change worked, that number will understate it for another week; if it failed, the same inertia will hide it.
The fix is not clever statistics. It is a line on the report saying since the change, and the discipline of writing down what changed and when so the line can be drawn at all. Most venues have the sales data and not the change log, which is why so many arguments about whether something worked cannot be settled.
One caveat worth keeping
A since-the-change figure starts out tiny, and a small sample says very little — so the two rules go together. Split the number at the change, and refuse to read it until enough has landed on the new side to mean something.