Drivers, and the Ones Nobody Tested
How to tell a driver that moves the number from one everybody believes in, the test that separates them, and what to do with the beliefs that fail it.
Five claimed drivers of one operation's volume, against what each rested on
Four years of daily counts; dates do not move.
Send calendar joined to daily volume, 31 sends.
Modelled from channel mix, not observed.
Sponsor's figure; no comparable launch exists.
Never tested against anything.
Two of the five were measured, one was modelled and two were somebody's belief. All five carried the same weight in the forecast and in the conversation about it. One employer's own register, not a benchmark.
A driver is something that moves the number and can be shown to have moved it. Most driver lists contain two or three of those and several things everybody believes, and the two kinds are treated identically until somebody asks for the evidence.
The method in “Drivers, and the Ones Nobody Tested” is only as useful as the definitions and source records behind it. A team exploring using remote employee monitoring software in workforce forecasts for remote employee monitoring software can add timely work and project context, with transparent settings, limited access and a documented correction route around every consequential use.
The asking is uncomfortable and it is the single most productive hour available in demand planning.
Additional context for “Drivers, and the Ones Nobody Tested” is available from Deloitte human-capital resources. External material is most useful when it prompts a specific validation check rather than replacing local judgement.
The register
One row per claimed driver: what it is, how it is supposed to move the number, what that claim rests on, and how many times it has been observed.
Four columns. Filling them in is the test.
The four states of a claim
Measured: the effect has been estimated from data, with a size. Inferred: it has been modelled from something related. Asserted: somebody with authority says so. Unknown: it is in the list because it has always been in the list.
There is no fifth state called certain, which is the point of having the first four.
Testing a driver
Compare periods with and without it, controlling for the obvious. For a campaign, the days after a send against comparable days without one. For a billing date, the days around it against the rest of the month.
This is arithmetic, not modelling, and it takes an afternoon per driver.
How many occurrences are needed
Enough that the effect is larger than the variation between comparable periods. A driver claimed to lift volume by 20% in a series with 10% daily noise needs four or five occurrences. One claimed to lift by 5% needs many more.
Where the data cannot support the test, the honest entry is inferred or unknown, not measured.
The driver that fails the test
Usually weather, usually something else that feels obvious. The result is not that it has no effect but that no effect can be shown at the size claimed.
What to do with it: leave it in the register as unknown, stop carrying it in the forecast, and stop letting it be used to explain misses after the fact.
The after-the-fact explanation
Which is what untested drivers are mostly used for. Volume was high, the weather was bad, therefore the weather. The same weather in a month with normal volume is never mentioned.
A register with states on it makes this harder, which is most of its value.
The driver that is real but unobtainable
Common with competitor behaviour and with price. The effect is plausible and large, and there is no data to estimate it.
The honest treatment is a named scenario rather than a coefficient: not a number in the model but a case with a consequence.
Drivers that are already in the series
A subtle trap. Where a driver has operated throughout the history at a stable rate, it is already in the baseline and adding it again double counts.
Drivers are only worth modelling separately where they vary, or where they are about to change.
The driver that changed size
Campaign lift that was 20% three years ago and is 8% now, because the audience is saturated. The register entry is still 20% because nobody re-estimated it.
Re-estimating the measured drivers annually is a morning's work and it catches this.
Owning the register
The planner owns the file; each driver has a named source — the person who supplies the calendar or the estimate. Where a driver has no named source, that is itself a finding.
Most registers, on first assembly, have two or three rows with nobody behind them.
What the register changes in a review
The conversation moves from "why was the forecast wrong" to "which driver moved, and was it one we had measured". That is a question with an answer and it produces a correction.
Without the register, every miss is explained by whichever driver is mentioned first.
The driver that is somebody's job
Worth noticing when a driver is owned by the person whose work it describes. Campaign lift estimated by the campaign team, automation savings estimated by the automation team.
Not dishonest, and consistently optimistic. The register entry should name the source so that the reader can apply their own discount rather than the planner applying one silently.
Drivers that work in only one direction
Several do. A campaign raises volume and its cancellation does not lower volume below baseline; a system outage displaces demand forward and nothing displaces it back.
Modelling these symmetrically produces errors on one side only, which reads as bias and is really a specification problem.
What a good register looks like
Five to ten rows. Two or three measured with sizes and dates. A couple inferred, labelled. One or two asserted, carried with a band. And the unknowns listed but not used.
That shape is achievable in a fortnight and it is better than almost any model built without it.