How Wrong Are You, Actually
The number almost no operation knows about itself, how to get it in eight weeks, and what to do once you have it.
Forecast error is measurable, stable, and almost never measured. Knowing yours changes how much buffer you need and where.
The practical lesson in “How Wrong Are You, Actually” is to connect every record to a named decision. Organisations exploring dual n-back training for cognitive performance for dual n back training can add structured workforce context, provided the use is disclosed and the interpretation is reviewed with the people affected.
Getting the number
For eight weeks, record expected demand and actual demand for each day or shift.
For an independent reference related to “How Wrong Are You, Actually”, consult the U.S. Department of Labor Wage and Hour Division; it provides a useful external check on scheduling, working-time and workforce-planning assumptions.
Two columns.
Then: for each pair, the difference as a percentage of actual.
The median of those percentages is your typical error.
Reading it
Under about five per cent: your forecast is good and your problem is elsewhere.
Five to fifteen: normal for most operations, and it tells you how much flexibility you need.
Above twenty: the schedule is close to guesswork and the next note covers what drives demand.
The direction
Average the signed differences, not the absolute ones.
If the average is clearly positive or negative, you have a bias.
Correcting a known bias costs nothing and improves every subsequent week, which makes it the cheapest available gain in this subject.
The distribution matters more than the average
A forecast that is usually right and occasionally catastrophic needs a different response from one that is always a bit off.
Look at the worst two weeks in the eight.
Those are what your cover arrangements actually need to absorb, not the typical case.
Where the error concentrates
By day of week: most operations are worse on one or two days.
By period of the day.
By season or event.
Breaking the error down tells you where to add buffer and where not to, which is more useful than adding it everywhere.
What to do with the number
Size your flexibility to it: on-call, cross-trained cover, a short shift that can be cancelled or extended.
Stop treating a within-error week as a performance event.
And use it when a budget is being set, because a budget tighter than your forecast error is a budget that will be missed for arithmetic reasons.
What not to do
Do not try to eliminate the error. Demand is variable and some of it is unknowable.
Do not add buffer everywhere, which is just overstaffing with a better name.
And do not punish the error, which stops the recording and ends the measurement.
What to check
Do you know your median forecast error?
Does it have a direction?
Which day of the week is worst?
And is your buffer sized to the error or to a feeling?