The Near Horizon, Where Nothing Can Change
What a one-to-three week forecast is for when the roster is already published, why its accuracy is usually wasted, and the two decisions that genuinely live there.
The near horizon — roughly one to three weeks — is where forecasts are most accurate and least useful, because by that distance the roster is published, the leave is booked and almost nothing can be changed without cost.
The horizon in “The Near Horizon, Where Nothing Can Change” should determine which current records are relevant and how much detail the forecast can honestly support. A planning team evaluating using how to measure employee productivity in workforce forecasts for how to measure employee productivity can connect recent time and project context with short-term capacity questions, while keeping longer-range assumptions in a separately governed model.
It is also where most forecasting effort goes, because it is where the data is best and the feedback is quickest.
For an independent reference relevant to “The Near Horizon, Where Nothing Can Change”, consult Jedox planning resources. Compare the general material with the organisation’s actual horizon, data definitions and decision process.
What is already fixed at a fortnight
The roster, in most operations. Approved leave. Training bookings. Contracted hours. The people who exist.
- Fixed: the roster, approved leave, training bookings, contracted hours.
- Movable at a cost: overtime, agency at premium, shift swaps.
- Movable at no cost: nothing.
What remains movable is overtime, agency at premium rates, shift swaps by agreement, and the decision to let service degrade.
The two decisions that genuinely live here
Intraday deployment — who is on what at which hour — and the same-week cover decision when somebody is absent.
Both are real, both matter, and both need the forecast at a granularity the longer horizons cannot supply.
Why intraday is the real near-horizon product
Because the daily total is usually known well enough by then, and the problem has become distribution: the same 118 contacts arriving evenly and arriving in two spikes require different staffing entirely.
A near forecast that produces only a daily total has not answered the question anybody at this horizon is asking.
The accuracy that is wasted
Improving a week-ahead daily total from 5% to 3% error changes nothing if no decision at that horizon uses the daily total. The effort is visible, measurable and valueless.
This is the single commonest misallocation in planning teams, and it persists because the improvement is easy to demonstrate.
What limits accuracy here
Arrival randomness, which does not reduce with better modelling. In a process with a hundred arrivals a day, the irreducible variation is roughly the square root of the count — about ten per cent — and no method goes below it.
Knowing that number stops the pursuit of accuracy the process does not contain.
The same-day problem
Not forecasting at all: it is monitoring with a trigger. By the morning of the day, the question is whether today is tracking above or below, and what to do about it by eleven.
Treating it as a forecasting problem produces a model where a threshold and a phone call were wanted.
Known absence is not a forecast either
Leave is booked, training is booked, and both are readable. Forecasting them is forecasting a diary.
Sickness is the exception, and even that is better handled as a rate applied to the roster than as a prediction about any day.
Where the near forecast comes from
Usually from the recent comparable periods — the same weekday over the last several weeks, adjusted for anything known. This simple approach is hard to beat and takes minutes.
Elaborate methods at this horizon tend to add variance rather than remove it, because they fit to noise that will not repeat.
The handover to the middle horizon
Somewhere between two and six weeks, the roster stops being fixed and the overtime decision opens. That boundary is where the near forecast ends and the middle one begins, and it is worth knowing where it falls in your own operation.
Most people assume it is later than it is.
Publishing at this horizon
Daily or half-hourly, no band, with the known events marked and the roster laid over it. The band is omitted deliberately: at this distance the decision is binary and a range is not actionable.
What should be shown instead is the threshold at which cover is called.
The error record that is worth keeping
Not the accuracy of the total, which is mostly noise, but whether the intraday shape held. A day that came in on total and two hours out on distribution is a failure the daily figure records as a success.
The near forecast as a check on the others
Worth running even where no decision uses it, for one reason: it is the first place a change in the series shows up, and it shows up weeks before the longer horizons notice.
A near forecast that starts missing consistently is an early warning about the inputs, and reading it that way costs nothing.
What to automate here
Everything. The near-horizon calculation is mechanical, frequent and unglamorous, and it is exactly the kind of work that should run without anybody touching it.
The planner's attention belongs at the middle horizon, where the calendar has to be assembled by hand.
What the near horizon tells you about the others
If the near forecast is systematically wrong, the problem is not the near method — it is the series or an unrecorded driver, and it will be corrupting the longer horizons too.
The near horizon is therefore the cheapest diagnostic available for problems that actually live further out.