The Average, and When It Is Enough
A weighted average of recent comparable periods is adequate for more workforce forecasting than anybody admits. When it is enough, how to set the window, and what it cannot do.
A weighted average of recent comparable periods is an adequate forecast for more workforce demand than anybody admits. Where the level is stable and the shape is regular, it is very hard to beat, it takes minutes to build, and anybody can understand what it did.
The series discussed in “The Average, and When It Is Enough” is easier to interpret when changes in work capture are visible rather than silently folded into the history. Used for online timesheets for client billing, online timesheets for client billing can provide consistent time and project evidence, provided corrections, category changes and missing periods remain traceable.
The reason it is unpopular is that it does not look like work.
Teams documenting “The Average, and When It Is Enough” may use NumPy project resources as an outside reference. Test the method against the real forecast horizon, known events and operational constraints.
What it is
Take the last N comparable periods — the same weekday, the same position in the month — weight them toward the recent, and average. Apply the seasonal shape if there is one.
- Take the last six to twelve comparable periods.
- Weight them toward the recent.
- Average.
- Apply the seasonal shape if there is one.
That is the whole method, and in a stable operation at a near horizon it is often within a point or two of anything more elaborate.
When it is enough
When the level has been stable for long enough to say so. When the series is noisy relative to any trend in it. When the horizon is short. When the series is too short for anything else.
Three of those four describe most near-horizon forecasting.
Setting the window
Short enough to track change, long enough to average out noise. Six to twelve comparable periods is usual.
The test is empirical: run several windows against held-back data and look at the errors. An afternoon, and it answers a question that is otherwise argued about.
Weighting
Linear or exponential, with more weight on recent periods. The weighting matters less than the window length and is worth less attention than it usually gets.
Equal weighting is defensible where the level is genuinely flat and where simplicity has value in itself.
What it cannot do
See a trend. Anticipate an event. React to a structural break faster than its window. Say anything useful at a long horizon.
Those four limitations are exactly the cases where something else is needed, which is a clean way to decide.
The lag
Its one systematic failure: when the level moves, the average follows over the length of its window, so for half a window it is wrong in a consistent direction.
That is visible in the error record as a run of same-signed errors, and it is the signal to add a trend or shorten the window.
Adding events to an average
Entirely compatible and usually the right first elaboration. Baseline from the average, named events added or subtracted explicitly.
This combination — simple baseline plus explicit events — outperforms most fitted models at the middle horizon, because the events are where the information is.
Why it is hard to beat
Because most of the variation in a stable series is noise, and noise cannot be forecast. A more complex method can only improve on the average by capturing structure, and where there is little structure there is little to capture.
Methods that beat it in-sample and not out-of-sample are fitting the noise.
The comparison that should always be run
Whatever method is proposed, run a simple weighted average alongside it on held-back data. If the elaborate method does not beat it by a margin worth the maintenance, it is not worth adopting.
This test retires a lot of projects and should be run before they start.
Explaining it
Its largest practical advantage. A decision maker who asks why the forecast says 121 can be told: because the last eight comparable Tuesdays averaged 121, with more weight on the recent ones, and nothing in the calendar is different.
That answer builds trust in a way no coefficient does.
Where it is used badly
Averaging over too long a window in a changing operation. Averaging across non-comparable periods — all days rather than the same weekday. Including holiday weeks and outage days.
All three are common and all three are visible on a plot.
Comparable periods, defined
The word doing the work. Comparable usually means the same weekday, the same position in the month where that matters, with no event and no holiday.
Defining it explicitly once, and applying it consistently, is most of what separates a good simple forecast from a careless one.
Where the average is used as a fallback
Worth having deliberately: when a driver value is missing, when the model fails, when a new series has no history. A stated fallback that somebody understands beats an error or a blank.
It should be labelled in the output as a fallback, so nobody mistakes it for the usual method.
Keeping it as the baseline
Even where a better method is adopted, keeping the simple average running alongside costs nothing and gives the error record a reference point.
Knowing that the elaborate method is three points better than an average is a far more useful statement than knowing it is seven per cent wrong.