When Judgement Makes It Worse
The five patterns of harmful adjustment, how each shows up in the record, and the small controls that remove them without losing the useful kind.
Harmful adjustment follows five recognisable patterns, and each leaves a distinct trace in a record that keeps the system number alongside the published one. None of them requires anybody to be acting in bad faith; all five are things reasonable people do.
The adjustment process in “When Judgement Makes It Worse” should preserve the unaltered model result, the human change and the reason for it. A team considering review the available controls directly for fte meaning can add operational time and project context, while the forecast owner remains responsible for documenting the override and reviewing its effect.
The controls that remove them are small, and they do not touch the adjustments that help.
For an independent reference relevant to “When Judgement Makes It Worse”, consult Association for Project Management resources. Compare the general material with the organisation’s actual horizon, data definitions and decision process.
Pattern one: chasing the last error
The forecast was under last period, so this period's is nudged up. The previous error was mostly noise, so the nudge is in a random direction relative to what happens next.
- Chasing the last error.
- The small frequent nudge with no stated fact.
- The standing uplift that became habitual.
- Optimism about one's own project.
- Adjusting toward the comfortable number.
In the record: adjustments correlated with the previous period's error. It is the single commonest harmful pattern.
Pattern two: the small frequent nudge
Adjustments of two or three per cent, made most periods, with no stated fact. Individually harmless, collectively they add variance and subtract accuracy.
In the record: many adjustments, mostly small, reasons mostly blank or generic.
Pattern three: the standing uplift
An adjustment applied every period in the same direction, originally for a reason, now habitual. It is a bias with a human face.
In the record: an adjustment column with a non-zero mean and almost no variation.
Pattern four: optimism about one's own project
Adjusting the forecast to reflect what a change is expected to achieve rather than what comparable changes have achieved. The automation will remove 25% of contacts; the record of past automations says 11%.
In the record: adjustments associated with change programmes, usually in the same direction, usually too large.
Pattern five: the comfortable number
Adjusting toward whatever makes the plan work, the budget balance or the conversation easier. This is the one people worry about and it is not the commonest.
In the record: adjustments that move the forecast toward a target or a previous commitment.
Why they are hard to see without the record
Because each individual adjustment has a reason that sounds fine at the time, and because the published number is the only one kept. With one column there is no pattern to see.
With two columns and a reason field, all five patterns are visible within a year.
The controls, in order of value
A threshold below which no adjustment is made without a named fact. A requirement to specify which periods are adjusted. The system column kept. A quarterly review of the two columns together.
Four controls, none onerous, and together they remove most of patterns one, two and three.
Handling pattern four
Not by refusing the sponsor's figure but by carrying two cases and the record of past comparable changes. "The last three automations removed a median of 11%; the project expects 25%; here are both."
That is a conversation rather than a confrontation, and it usually ends with a sensible middle.
Handling pattern five
The hardest, because it is usually not the planner doing it. The defence is structural: separate the forecast from the target, name an owner for each, and publish the gap.
Where the two numbers are separate, the pressure moves to the target row where it belongs.
What not to do
Remove the ability to adjust. The valuable adjustments come from the same channel as the harmful ones, and the valuable ones are worth more.
An organisation that bans override loses the release dates and keeps the model's blindness.
The review
Quarterly, half an hour, two columns and a chart of adjustment against subsequent error. Not an appraisal: a reading of a record.
Done in public, it changes behaviour quickly, mostly by making the small nudges feel unnecessary rather than by forbidding them.
The pattern that looks harmful and is not
Worth guarding against in the review: large adjustments in a period that then turned out badly. The adjustment may have been correct and the period affected by something else entirely.
Judging individual adjustments by their outcome is as unreliable as judging individual forecasts. The record is read in aggregate or not at all.
Adjustment under time pressure
Most harmful adjustments are made in the last hour before publication, without the calendar open and without checking the system figure against recent actuals.
Moving the adjustment step earlier in the cycle, with the register on the screen, removes a surprising share of them.
The finding to expect
In most first reviews: a handful of excellent adjustments carrying real information, a larger number of small ones costing a point or two, and one standing uplift nobody could account for.
Removing the second and third categories, while protecting the first, is usually worth more than a change of method.