Santiago Blandón Ferro·Architect, technology for poultry production·More about me
A forecast without a band is not a forecast: it is an opinion with decimals
A bare number invites belief. A number with a band states how much is known and how much is not, and lets you check afterwards whether the band was right as often as it promised.
What is missing from "this flock will weigh 2,640 g on day 40"?#
What is missing is how much of that is knowledge and how much is ignorance.
That sentence is indistinguishable from one saying 2,641. The last two digits are pure decoration: nobody knows a flock's weight to the gram twelve days out, and presenting it that way invites decisions taken at a precision the figure does not have.
The honest version is 2,420 – 2,860 g, centred on 2,640. It takes more room and says considerably more.
Why does the band change the decision and not just the presentation?#
Because which edge matters depends on what the number is for.
| Decision | Edge to read | Why |
|---|---|---|
| Committing kilos to a customer | the lower one | if you miss, you pay |
| Filling a truck or a plant slot | the upper one | if there is spare, people wait |
| Deciding whether a target is reachable | both | if the target falls inside, the honest answer is "unknown" |
That third case is the one most often abused. When the target sits inside the band, any decisive verdict — "yes it will" or "no it won't" — is being invented by the system. The third possible answer is not an excuse: it is the actual state of knowledge, and some decisions change depending on which of the three it is.
What is the band made of?#
Three things that add up, worth separating because they shrink in different ways:
- Measurement error. How many birds are weighed and how. It shrinks with better weighing: larger sample, calibrated scale, same time of day.
- Residual ignorance. What the model does not capture about that particular farm's management. It shrinks with history: more closed cycles from the same complex.
- Distance to horizon. Twelve days out is less knowable than two. It does not shrink at all; it is what makes the band widen as the projection reaches further.
And a fourth that almost nobody models: the quality of the input data. If that week's weighings came in flagged — a weight that falls with no thinning to explain it, a record identical to the previous day's, a capture outside the shed's working hours — the band has to widen. A projection computed on doubtful data with the same tightness as on clean data is lying twice.
How do you know whether a band is any good?#
You check, and the check is deliberately uncomfortable.
An 80% band promises to contain the real value 80% of the time. No more, no less. You have to store every projection with the date it was made and the day it points at, wait for that day, record the actual value and count.
- Coverage well below what was promised: the band is too tight. The system appears to know more than it does, which is the fastest way to lose a customer's trust.
- Coverage well above: the band is so wide it carries no information. "Between 1,800 and 3,400 g" is always right and helps decide nothing.
That report has to be showable even when it looks bad. Ask anyone selling you forecasts for their coverage: if they cannot show it to you, what they are asking for is an act of faith.
And when there is no history yet?#
That is the real test, because it is precisely when the projection is looked at most and known least.
The right approach is to start from the genetic standard adjusted by whatever little has been observed of the flock, and shrink that adjustment back towards the standard the weaker the evidence. With two weighings the system should lean mostly on the reference curve; with twelve, on the flock.
The wrong approach is to train a model on three closed cycles and present it as learning. With three flocks you can fit anything to anything. An honest engine has a gatekeeper: if the trained model does not consistently beat the standard under cross-validation, it is rejected, and the screen says it was rejected. Rejecting more often than adopting is not a failure of the mechanism: it is the mechanism.
The rule, in one line#
If a system hands you a number with no band, the question to ask is not "how wrong are you?". It is "how often have you been right so far, and where can I see it?". If there is no answer, it was not a forecast.
In the demonstration you can see how every projection is recorded against the value that was later measured.