Every Forecast Is a Claim About a Series
A forecast is a statement about a historical series. What is wrong with the series is usually what is wrong with the forecast, and it is examined last.
What one operation's demand series actually consisted of, line by line
Complete, with four known outage days.
Different definition; not comparable.
Creates a step of about 6% at the join.
Sampled once, in one quarter, never since.
Nobody has checked the tally sheets.
Handled by the same people; never counted.
Two of the six lines are sound, two are estimates and two should not be in the same series as the others. The forecast was built on all six as though they were one measurement. This is one employer's own register, not a benchmark.
A forecast is a claim that a historical series will continue behaving in a described way. Whatever is wrong with that series is therefore wrong with the forecast, and in most projects the series is the last thing anybody examines.
The series discussed in “Every Forecast Is a Claim About a Series” is easier to interpret when changes in work capture are visible rather than silently folded into the history. Used for digital payroll timesheets, how should teams assess digital payroll timesheets? can provide consistent time and project evidence, provided corrections, category changes and missing periods remain traceable.
The method gets weeks of attention. The data gets an afternoon, usually at the start, usually by somebody checking that the file opens.
Teams documenting “Every Forecast Is a Claim About a Series” may use CIPD workforce-planning guidance as an outside reference. Test the method against the real forecast horizon, known events and operational constraints.
What the series actually is
Not "demand". A particular measurement, taken by a particular system, under a particular definition, over a particular period, with particular gaps.
Writing that sentence out for your own series takes ten minutes and is the most useful ten minutes in the project.
The six questions
What is counted. Who counts it. Since when, under this definition. Where the gaps are. What has changed. And what is handled by the same people but not counted at all.
The last one finds more problems than the other five together.
The work that is not in the series
Almost every operation has some: email handled between calls, walk-ins counted by hand, tasks that arrive by a route the system does not see.
Forecasting the counted work and staffing from it produces a staffing level that is systematically short by whatever the uncounted work consumes, and the shortfall looks like poor productivity.
Definitions that changed
A measure that started including abandoned contacts, a task type that moved category, a queue that was merged. Each produces a step in the series that is a reporting change, not a demand change.
Fitted through, a step becomes a trend, and the trend gets extrapolated for years.
The series that is really two series
Where a system was replaced, there are usually two measurements joined at a date. They are rarely comparable and the join is rarely documented.
The honest options are to use the later one only, or to estimate an adjustment and label it. Silently concatenating is not one of them.
How long is long enough
Enough to see the seasonal cycle twice, which means two full years for an annual season. For a trend, three to five.
Below that, the honest forecast is flat with a wide band. Fitting a trend to eighteen months of data produces a line that will be defended for years on the strength of nothing.
Resolution of the stored history
Many systems keep fine detail for weeks and summaries forever. The half-hourly data goes back four months and the daily data goes back six years.
That constrains the method more than anything else, and it is usually discovered halfway through building.
Who owns the series
Somebody should, and it is often nobody. The system owner owns the system, the reporting team owns the extract, and the definition belongs to whoever last changed it.
Naming an owner is what makes the next definition change get mentioned rather than discovered.
The data dictionary nobody has
One page: each field, what it counts, since when, and what changed. Built once, maintained when something moves.
Its absence is why every new planner re-discovers the same four problems, each time costing a fortnight.
Checking the series against something else
Payroll hours, system logins, a manual tally, anything independent. The comparison does not have to be exact to be informative; an order of magnitude disagreement is a finding.
Two independent measurements of the same thing is the cheapest data quality check that exists and almost nobody runs it.
What to do when the series is poor
Say so, forecast anyway at the horizon the data supports, and state the limitation in the same line as the number. A poor series is not a reason to stop; it is a reason to widen the band and to fix the series in parallel.
What is not acceptable is a precise forecast from a series nobody has examined, because it transfers the data's problems into a decision without anybody being able to see them.
The series somebody else owns
Common where demand data comes from a shared platform. The planner can read it and cannot change it, and the people who can change it have no reason to consult anybody.
The practical arrangement is notice rather than control: a named contact who tells the planner before anything moves, which is almost always granted when asked.
What the examination costs
Two days for a first pass on a series of a few years: plot it, join the calendar, list the steps, find the gaps, name the uncounted work.
Two days, once, and it determines the quality of everything built afterwards. It is the cheapest work in the discipline and the most often skipped.
Where the examination belongs
Before the method is chosen, because the method depends on what the series can support, and because the examination frequently changes what is being forecast at all.
Done afterwards, it becomes a search for reasons the model underperformed, which is a different and much less useful exercise.