Granularity Is Part of the Horizon
How fine a forecast should be at each distance, the test that decides it, and what spurious precision costs when it reaches a decision maker.
Granularity belongs to the horizon, not to the system. Half-hourly at a fortnight, daily at a quarter, monthly at a year — and the reason is not convention but a test anybody can apply: detail that is smaller than the error band is decoration.
The horizon in “Granularity Is Part of the Horizon” should determine which current records are relevant and how much detail the forecast can honestly support. A planning team evaluating read the detailed setup information for hourly timesheet template can connect recent time and project context with short-term capacity questions, while keeping longer-range assumptions in a separately governed model.
Most forecasting tools will produce any granularity at any horizon, which is why the question has to be answered deliberately.
Teams documenting “Granularity Is Part of the Horizon” may use Anaplan planning resources as an outside reference. Test the method against the real forecast horizon, known events and operational constraints.
The test
Compare the size of the detail to the size of the band. A half-hourly figure of 14 inside a band of ±5 is carrying information. The same figure inside a band of ±9 is not.
Detail smaller than the error band is decoration. A half-hourly figure of 14 inside a band of ±5 carries information; the same figure inside a band of ±9 does not.
Where the detail is smaller than the band, it should be aggregated until it is not.
What spurious precision costs
It is not harmless. A monthly figure broken into weeks at a year's distance will be used to decide which weeks to recruit for, and the weekly pattern in it is an artefact of last year's calendar.
The decision taken on it is worse than one taken on the monthly figure, because the monthly figure would have prompted a question.
Granularity at the near horizon
Half-hourly or quarter-hourly where arrival distribution drives staffing, daily where it does not. The distinction is whether people can be moved within a day.
In operations where they cannot — a fixed shift with no intraday flexibility — half-hourly forecasting is work with no decision attached.
Granularity at the middle horizon
Weekly, almost always. The decisions here are about which weeks to cover, and a monthly total hides the peak that the decision is about.
Daily at twelve weeks is usually spurious; the day-of-week pattern is stable enough to apply later if needed.
Granularity at the far horizon
Monthly, or quarterly where the series is short. The seasonal shape is the only within-year information that survives, and it is a monthly shape.
Anything finer invites the reader to believe something the forecast does not know.
Two kinds of granularity
Time is one. The other is organisational: by team, by site, by skill, by queue. The same test applies and it bites harder, because splitting a forecast four ways roughly doubles the relative error of each part.
A site-level forecast with a ±12% band becomes four team-level forecasts with ±24% bands, and the sum is no better than it was.
Forecast high, split later
The general rule: forecast at the level where the series is most stable, then split using a ratio from recent history. The split introduces its own error, but a much smaller one than forecasting four sparse series separately.
The exception is where the parts behave genuinely differently, in which case they are separate series rather than splits.
When a part really is different
A new site, a queue with its own driver, a team taking a different product. These need their own forecast, and the giveaway is that the split ratio keeps moving.
A ratio that has drifted steadily for a year is telling you the parts are not one thing.
The granularity the requester asks for
Almost always finer than the horizon supports, because finer feels more useful. The answer is not refusal but the test, shown: "at this distance the band is ±23, so a weekly figure of 31 is inside the noise; here it is monthly, and here is what the weekly split would be worth."
Shown once, the argument does not recur.
Changing granularity mid-series
Breaks comparability, which is why it is worth getting right at the start. Where it has to change, both versions are kept for a period so the error record spans the change.
An error record that silently changes granularity is unreadable afterwards, and nobody notices for a year.
What to publish
The granularity the horizon supports, with one line saying what it was and why. Where a finer view exists for internal working, it stays internal.
A working file that escapes becomes the published forecast within a month.
Granularity and the error record
Changing granularity changes every accuracy figure, usually flatteringly. A move from daily to weekly reporting will improve the apparent error by several points with no change to the forecast.
Which is why granularity belongs in the same line as the error, and why a record that has silently changed it is unreadable.
The detail kept for a different reason
Sometimes fine detail is kept not to forecast but to diagnose: the half-hourly actuals are what show that a weekly miss was a single morning.
That is a legitimate use and it does not make the half-hourly forecast worth publishing.
The honest aggregate
Where a decision genuinely needs finer detail than the horizon supports, the right answer is not to supply it anyway. It is to say that the decision needs a shorter lead time, or a different kind of analysis, or a trigger rather than a forecast.
That answer is less satisfying and it is the one that holds up.