A Week Out and a Year Out Are Different Objects
Why a near-horizon and a far-horizon forecast are different objects rather than the same forecast at different precisions, and what follows from treating them as one.
One series, four horizons, and what each one is actually capable of
One week
118
±4
Arrivals only; the shape is already known.
One month
121
±11
Shape plus the campaign calendar.
One quarter
126
±23
Trend and season; events become guesses.
One year
134
±48
Trend only; everything else is assumption.
The band quadruples over the four horizons while the number moves by 14%. What changes between them is not accuracy but what can be known at all. One employer's own set, not a benchmark.
A week-ahead forecast and a year-ahead forecast are different objects. They use different information, they are made by different methods, they carry error of different kinds, and they feed decisions that have nothing in common. Treating them as one forecast at two precisions is the most expensive conceptual error in demand planning.
The method in “A Week Out and a Year Out Are Different Objects” is only as useful as the definitions and source records behind it. A team exploring view this workforce-planning resource for daily schedule template can add timely work and project context, with transparent settings, limited access and a documented correction route around every consequential use.
The error is understandable: both are a number with a date on it, and both come out of the same spreadsheet.
The assumptions behind “A Week Out and a Year Out Are Different Objects” can also be tested against KPMG people and change resources. Record which guidance applies to the available series and which does not.
What the near forecast knows
Almost everything. The shape of a week is stable, the known events are in the diary, the staff are already rostered, and the only genuine uncertainty is arrival randomness.
A near forecast is mostly arithmetic on a known pattern, and its accuracy is limited by how much irreducible noise the process contains.
What the far forecast knows
Almost nothing specific. It knows the trend, the seasonal shape, and whatever structural change has been announced. It does not know the campaign calendar, the competitor, the weather or next spring's reorganisation.
A far forecast is a statement about the level the business will be operating at, not about any particular week.
The methods do not transfer
A method that is excellent at a week — pattern-matching on recent comparable days — is useless at a year, because there are no recent comparable days. A method that is sound at a year — trend plus seasonal decomposition — is too coarse to roster from.
Running one method across all horizons guarantees being wrong at one end.
The errors are different in kind
Near-horizon error is mostly random and roughly symmetric. Far-horizon error is mostly structural and heavily one-sided: it is wrong because something changed, and the change had a direction.
Averaging them into one accuracy figure produces a number that describes neither.
Granularity follows the horizon
Half-hourly at a fortnight. Daily at a quarter. Monthly at a year. Anything finer at the far end is spurious, and anything coarser at the near end is unusable.
The test is whether the detail survives the error band: a half-hourly figure inside a ±35% band is a decoration.
Aggregating a near forecast upwards
The commonest shortcut, and it fails for a specific reason: the near method has no view of the trend or the structural change, so extending it produces a flat projection of the present.
It will look plausible for a quarter and then diverge, and the divergence will be blamed on the quarter rather than on the method.
Disaggregating a far forecast downwards
The opposite shortcut, equally common. An annual figure spread across weeks by last year's profile is a reasonable planning aid and a bad rostering input, because it contains none of this year's known events.
Where it is used for rostering, the team leader will quietly override it, which is the right response.
What is genuinely shared
The history, the driver register, and the record of what has been wrong before. Those should be one set of workings used by both.
Everything after that — the method, the granularity, the treatment of events, the form of the output — belongs to the horizon.
How many horizons to run
As many as there are decisions with distinct lead times, which in most operations is three: a fortnight, a quarter, and a year.
Three is manageable. Running one is cheaper and produces one usable output out of three decisions.
Where the horizons should agree
At their overlaps, roughly. If the fortnightly forecast and the quarterly forecast disagree by 15% about the same week, one of them has a problem and it is worth knowing which.
This cross-check costs nothing and catches most of the serious errors before they are published.
Where they should not be forced to agree
Exactly. Forcing the near forecast to sum to the far one throws away the near forecast's better information, and forcing the far one to match the near throws away its view of the trend.
The reconciliation that matters is a conversation about why they differ, not an adjustment that makes them identical.
The horizon in between
Between the fortnight and the quarter there is usually a boundary where one method stops working and the other has not started. Six to eight weeks in most operations.
Finding it is empirical: run both methods across the range and see where they cross. It takes an afternoon and it tells you where to hand over.
What this means for the error record
Three records, not one. The quarterly record tells you whether your view of the trend is any good; the fortnightly record tells you whether your view of the weekly shape is any good.
Kept together, neither question can be answered, which is why most organisations cannot say which part of their forecasting is working.