The Band, and Publishing It
A point estimate published alone acquires a precision it never had. Where the band comes from, how wide it should be, and how to publish it so it survives.
A point estimate published without a band acquires a precision it never had, and it keeps that precision as it travels. By the third presentation it is a fact, and the decision taken on it is sized as though the number were known.
The error measure in “The Band, and Publishing It” needs both the original forecast and a trustworthy record of what actually happened. Teams researching employee time clock records can use open the relevant platform overview to strengthen the actual-work side of that comparison without treating activity data as proof of demand, performance or intent.
The band is not a hedge. It is the part of the forecast that tells the decision how much weight it can carry.
For an independent reference relevant to “The Band, and Publishing It”, consult UK labour-market statistics. Compare the general material with the organisation’s actual horizon, data definitions and decision process.
Where the band comes from
From the error record, not from theory. The spread of past errors at the same horizon is the best available estimate of the spread of future ones.
- From the error record, not from the model's own interval.
- At the same horizon, because the band at two weeks is not the band at twelve.
- Recalculated each cycle from the rolling record, which costs nothing.
This is another reason the record is kept by horizon: the band at two weeks and the band at twelve are different numbers and come from different rows.
Empirical rather than modelled
Model-derived intervals assume the model is right, which is exactly what is in question. They are almost always too narrow, sometimes by a factor of two.
An empirical band from the last twenty errors at that horizon is simpler, more honest and easier to explain.
What width
Usually the 80th percentile rather than the 95th: wide enough to be honest, narrow enough to be usable. A 95% band in a volatile series is so wide that readers discount it.
Whatever is chosen is stated: "80% band" rather than "±23".
Asymmetric bands
Often correct. Where demand can spike upward much more than it can fall — most service operations — the band should be wider above than below.
A symmetric band in an asymmetric process understates the risk on the side that costs money.
Publishing it so it survives
Inside the number's own line, not in a footnote. "126 ±23 (80%), as at 14 March" is a string that survives being copied into somebody else's slide. "126" with a note at the bottom of the page does not.
This is a presentational detail that determines whether the band exists in practice.
The request to remove it
Will come, usually phrased as a concern that ranges confuse people. The underlying request is for the uncertainty to go away, and removing the band does not do that — it relocates it into the decision, unlabelled.
The workable response is to offer thresholds instead: a point with the level at which the answer would change.
Thresholds as an alternative form
"126; below 118 no action, above 134 recruit." Same information, in a form that reads as a decision rather than as a hedge.
Decision makers accept this where they resist a range, and it is more useful because it makes them state their own tolerance.
Checking the band is honest
Count how often the outturn fell inside it. An 80% band should contain the outturn about eight times in ten. Containing it every time means it is too wide; half the time means too narrow.
This check takes a minute and almost nobody runs it.
The band that is too wide to be useful
Legitimate at the far horizon, and the correct response is not to narrow it. It is to say that the forecast cannot support the decision at that distance, and to offer scenarios or a shorter lead time instead.
A band that is honestly too wide has told you something important.
Bands and aggregation
Narrow faster than people expect when periods are summed, because errors partly cancel. A quarterly total has a proportionally narrower band than any of its weeks.
Which means a band cannot be scaled by multiplying, and doing so overstates the uncertainty of the total considerably.
The band that nobody updates
A real failure: a band calculated once at implementation and carried for years while the series became more volatile.
It should be recalculated from the rolling error record each cycle, which costs nothing because the record is already there.
What to do with the band operationally
More than look at it. The upper end is the staffing contingency question, the lower end is the redeployment question, and both can be planned for in advance rather than reacted to.
A band published without that conversation is information nobody has converted into a preparation.
What the band changes in practice
Decisions get sized to the uncertainty rather than to the point. An operation that knows its twelve-week band is ±18% commits differently from one that believes the number.
That difference is the main thing publishing a band is for, and it shows up as fewer reversals rather than as better numbers.