When Judgement Helps
Human adjustment improves a forecast under specific conditions. What they are, how to tell whether yours are met, and how to make the good adjustments more likely.
Six adjustments made in one year, and what each one rested on
Last three releases averaged 11%.
Volume profile in the contract.
Modelled from two past displacements.
Project's own estimate; no comparable.
Confirmed with the send calendar.
No fact cited.
Three of the six carried a measured basis, one was modelled, one was the sponsor's figure and one had no reason at all. Five of them were specific and dated; the sixth was applied to everything. One employer's own record, not a benchmark.
Human adjustment improves a forecast when the person has information the model does not, when that information is specific and dated, and when the adjustment is large enough to be more than noise. Where all three hold, judgement adds a great deal. Where any fails, it subtracts.
The adjustment process in “When Judgement Helps” should preserve the unaltered model result, the human change and the reason for it. A team considering consult this workforce software page for 7 minute rule payroll can add operational time and project context, while the forecast owner remains responsible for documenting the override and reviewing its effect.
The conditions are checkable, which means this does not have to be a matter of opinion.
For an independent reference relevant to “When Judgement Helps”, consult Project Management Institute learning resources. Compare the general material with the organisation’s actual horizon, data definitions and decision process.
Condition one: information the model lacks
A release date, a contract signed last week, an outage at a supplier, a competitor's announcement. Facts that exist and are not in the series.
"There is a release on the fourteenth" passes the test. "It feels low" does not. The difference is whether the person can name the fact, not whether they turn out to be right.
The test is whether the person can name the fact. "There is a release on the fourteenth" passes. "It feels low" does not.
Condition two: specific and dated
An adjustment attached to particular periods, for a particular reason, with an end. "Plus 12% for the week of the fourteenth, because of the release" is specific.
A permanent uplift applied to everything is not an adjustment; it is an uncorrected bias.
Condition three: large enough to matter
An adjustment smaller than the forecast's own error band is almost certainly noise. Moving 126 to 128 inside a band of ±23 changes nothing and adds variance.
A threshold — no adjustment below some fraction of the band without a stated fact — removes most of the harmful ones.
What the evidence generally shows
Large, reasoned, information-based adjustments improve accuracy substantially. Small, frequent, intuition-based ones make it worse. The net effect in a given organisation depends entirely on which kind predominates.
Which is why the record matters more than the policy.
Adjustments in the direction of the last error
A specific hazard. After an under-forecast, the next forecast gets adjusted up; after an over, down. This is chasing noise and it reliably increases error.
It is also easy to detect in the record: adjustments correlated with the previous period's error.
The planner who knows the operation
The strongest case for judgement. Somebody who walks the floor hears about the new process, the broken system, the customer who is about to migrate, weeks before any of it appears in data.
Protecting that channel is worth more than most modelling, and it is usually the first thing lost when planning is centralised.
Making good adjustments more likely
Require a named fact and a date. Require the adjusted periods to be specified rather than the whole horizon. Show the system number alongside. And review the record quarterly in public.
Four small disciplines, and they shift the mix decisively toward the useful kind.
The adjustment that should become a driver
Where the same adjustment is made every month for the same reason, it is not judgement; it is a missing term in the model.
Moving it into the event calendar or the driver register frees the judgement for things that are genuinely new, which is what it is for.
Measuring whether it helped
Error of the published forecast against error of the system forecast, over the same periods. Two columns, one subtraction.
A year of that comparison answers the question definitively for your own operation, which no general finding can.
When it did not help
The honest response is to tighten the conditions rather than to abolish override. Usually the problem is the small frequent adjustments, and a threshold fixes it without losing the valuable ones.
Abolishing judgement entirely loses the release dates and the contract news, which is a worse trade.
Who should be allowed to adjust
Whoever has the information, which is usually the planner and the operations manager jointly. Adjustments from further away are more often about the number being unwelcome than about new facts.
A named adjuster per forecast, in the header, is the simple control.
Judgement about the band
Underused. A planner who knows something unusual is coming but cannot size it has a better option than adjusting the point: widen the band and say why.
That communicates exactly what is known — more uncertainty, no clear direction — and it is the honest response to most vague information.
Judgement at the far horizon
Where it is worth most, because the data supports least. Knowledge of contracts, strategy and planned change is nearly all of what a year-ahead forecast can usefully contain.
At that distance the model is a baseline and the judgement is the forecast, which is the reverse of the near horizon and worth saying out loud.
The summary rule
Judgement is for facts, not for feelings, and for specific periods, not for the whole horizon. Where it is used that way it is one of the most valuable inputs a forecast has, and the record will show it.