Rolling, Not Annual
Why a forecast that always reaches the same distance ahead beats one that reaches the end of the financial year, and what it takes to run one.
A rolling forecast always reaches the same distance ahead. An annual forecast reaches the end of the financial year, which means that in April it looks twelve months out and in February it looks six weeks out, while the decisions it feeds have not changed their lead times at all.
The method in “Rolling, Not Annual” is only as useful as the definitions and source records behind it. A team exploring a practical approach to tips to increase productivity for tips to increase productivity can add timely work and project context, with transparent settings, limited access and a documented correction route around every consequential use.
Almost every organisation runs the second and most would be better served by the first.
As a separate benchmark for “Rolling, Not Annual”, see scikit-learn documentation. Keep source data, model, override and resulting decision visible as distinct steps.
What goes wrong in the fourth quarter
The horizon collapses below the lead time of the longest decision, and recruitment that should have been decided in January is decided against a forecast that ends in March.
| Month | Annual horizon | Rolling horizon |
|---|---|---|
| April | 12 months | 16 weeks |
| September | 7 months | 16 weeks |
| February | 6 weeks | 16 weeks |
The symptom is a predictable annual scramble that everybody attributes to the business rather than to the shape of the forecast.
What goes wrong in the first quarter
The opposite: a twelve-month view for decisions with eight-week lead times, produced at a precision the horizon cannot support, and then reported against for a year.
Both problems are the same problem, and both disappear when the horizon stops being tied to a date.
What rolling means in practice
Each cycle, drop the period that has passed and add one at the far end. The horizon length never changes. In a sixteen-week rolling forecast, every refresh adds a week.
Mechanically this is trivial. What is not trivial is that the numbers change every cycle, and some organisations find that uncomfortable.
The discomfort is the point
A forecast that does not change is not responding to information. The objection "it keeps moving" is usually an objection to seeing that the world moved.
What helps is showing the change rather than hiding it: this period's figure, last period's, and one line on what moved it.
Rolling and the budget
They coexist. The budget is an annual commitment with an owner; the rolling forecast is the current expectation. They answer different questions and the gap between them is the thing worth reporting.
Where the two are merged, the budget wins and the forecast stops being informative, which is the subject of its own note elsewhere here.
The re-forecast cadence
Should match the decision cadence, not the reporting cadence. Weekly for the middle horizon where overtime and agency are decided weekly. Monthly for the long one.
A quarterly re-forecast feeding weekly decisions is three-quarters stale most of the time.
What it costs to run
Less than an annual process, usually, because the work is distributed rather than concentrated. The annual cycle costs three weeks once; the rolling one costs a morning a month.
The saving is real but it is not the argument. The argument is that the decisions get a number with the right distance on it.
The version that fails
A rolling forecast where only the near periods are genuinely re-examined and the far end is copied forward unchanged. It looks rolling and behaves annually.
The giveaway is a far-end figure that has not moved in six cycles while the near end has moved every cycle.
Keeping the record readable
Each cycle's forecast kept as issued, so the error record can be built by horizon: what the twelve-week-ahead forecast said about this week, what the four-week-ahead said, and what happened.
That record is the only way to learn whether the distance is where the error is coming from, and it is impossible to reconstruct later.
The vintage problem
Comparing this cycle's forecast to last cycle's requires knowing which vintage is which. Without a date stamped into the file, two versions of the same quarter become indistinguishable within months.
A date in the filename and a date in the header. It is a small discipline and its absence ruins error records.
Starting one
Pick the decision with the longest lead time, set the horizon, and run it alongside the annual process for two cycles. Do not announce a replacement.
After two cycles it is usually obvious which number people are actually using, and the conversation becomes much easier.
The first cycle is the hard one
Because the far periods have to be built from nothing, and because everybody compares the new output to the annual one and finds it unfamiliar.
By the third cycle the far end is a carry-forward with adjustments and the work drops sharply. Warning people about that shape in advance prevents the process being judged on its worst month.
What rolling changes
The fourth-quarter scramble stops. That is the visible effect and it is the one that gets noticed.
The less visible effect is that the error record becomes readable by horizon, which is the precondition for ever improving the method rather than simply changing it.