What a Usable Forecast Looks Like
The assembled answer: what a forecast that can be used and judged actually contains, and what none of it requires.
What one published forecast carried, line by line, at each horizon
2 weeks
118
±4
Daily, with the roster over it; no band published.
6 weeks
121
±11
Weekly; overtime and agency priced beside it.
12 weeks
126
±23
Weekly; recruitment threshold at 131.
12 months
134
±48
Monthly; three cases, not one number.
Each horizon carries its own granularity, its own band and its own decision, and the far one is not a single figure at all. Nothing here required a better method than the one the operation already had. One employer's own published set, not a benchmark.
A usable forecast carries five things: a named decision, a horizon that matches its lead time, a point with a band, the record of how wrong it has been before, and a note of what was adjusted and why. Everything in this collection assembles into those five.
The method in “What a Usable Forecast Looks Like” is only as useful as the definitions and source records behind it. A team exploring explore the platform information here for how to handle multiple clients can add timely work and project context, with transparent settings, limited access and a documented correction route around every consequential use.
None of them is a method. A forecast with all five and a crude method is more useful than a sophisticated one with none, and that is the single most useful thing to know about this discipline.
As a separate benchmark for “What a Usable Forecast Looks Like”, see Board enterprise-planning resources. Keep source data, model, override and resulting decision visible as distinct steps.
The five
The decision, with an owner and a date. The horizon, derived rather than inherited. The point and the band, at a granularity the band supports. The error record, by horizon, against a baseline. And the adjustment record, with reasons.
Five lines of header and two small tables.
What it is not
A prediction. A forecast does not claim to know what will happen; it states what is expected, how much that expectation has been wrong before, and what it is for.
Promising accuracy sets a test that cannot be passed. Promising traceability sets one that can be passed every time.
Being wrong well
A usable forecast that turns out wrong produces an attributable correction: a driver whose size has moved, a calendar entry that was missing, a break that was not spotted.
An unusable one produces a discussion about whether anybody could have known, which improves nothing.
What a reader is entitled to
The basis of any figure, on request, without preparation. Where the answer is "that is the sponsor's estimate, carried as asserted", the forecast is working. Where it is "I would have to look", it is not finished.
A forecast that needs its author present to be understood is a presentation.
The maintenance
A monthly review of half an hour. A quarterly re-estimation of the weekly profile. An annual pass over the drivers and the seasonal shape. A line in the record each period.
A few hours a quarter in total, and it is the whole upkeep.
What usually improves first
Not the method. The calendar, which is a fortnight of conversations, and the horizon, which is an afternoon of arithmetic. Between them they typically move twelve-week accuracy by several points.
Method changes come after those and deliver less.
Where to start with nothing
One decision, one horizon, one simple method, one error record started this week. A page.
That page, honest about what it cannot see, is more usable than a model nobody can explain, and it takes an afternoon rather than a project.
The two numbers to keep apart
What is expected and what is undertaken. Merged, the forecast stops being informative within two cycles and the organisation loses its ability to say what it actually expects.
Everything else here survives most forms of neglect. This one does not.
The habit that holds it together
Writing down what is known, what is estimated and what is assumed, in the same place as the number, each period, as a matter of course.
That is a filing habit rather than an analytical skill, and it is what separates a forecast that improves from one that merely repeats.
What the horizon taught
That distance is the variable almost nobody tracks. A number is not good or bad in itself; it is good or bad at a distance, for a decision, against a baseline.
Reading forecasts that way makes most of the arguments about them resolvable, which is the practical point of organising a whole collection around it.
What to do first on Monday
Save this cycle's forecast as a dated file and do not overwrite it. That single act starts the error record, costs nothing, and cannot be done retrospectively.
Everything else here can be added later. That one cannot.
The forecast nobody questions
Worth noticing, because it usually means nobody is relying on it. A number that is never challenged is a number that is never used.
The first quarter of hard questions is the sign that the document has started to do something, and it is a better indicator than any accuracy figure.
The last thing
The figure at the bottom of a forecast is not the output. The output is the labelled set of statements above it: this is what we expect, this far out, for this decision, with this much error behind us, having adjusted it here for this reason.
A forecast read for its number will be argued with. One read for its labels gets corrected, which is the only way it ever becomes worth having.