Analytics
Trust is won by admitting uncertainty, not hiding it
A forecast that reports a single confident number gets ignored by the people who know the business well enough to know it cannot be that certain.
5 min read
Operational teams have a good instinct for when a number is overstated. Hand a planner a point estimate with no interval, and they will compare it against their own experience, find the cases where it is obviously wrong, and quietly go back to the spreadsheet.
This is usually read as a change-management problem. It is more often a modelling-presentation problem. The model may be perfectly well calibrated internally, and simply be presenting itself in a way that invites disbelief.
Publishing calibrated intervals, segmenting accuracy by condition, and stating plainly where the model degrades all feel like admissions of weakness. In practice they are the thing that earns adoption, because they let an experienced operator work out when to trust the output and when to override it.
The measure that matters is not model accuracy. It is whether the decision that was actually taken improved — which requires the person taking it to understand what the model does and does not know.
