Presenting Uncertainty Honestly
P14.communication.03 · Audience: guest, it-ml, language-pro · Prerequisites: The One-Slide Answer
The forecast said 40,000 units. The team planned for 40,000 units, ordered parts for 40,000 units, and was blindsided when the real number came in at 31,000 — inside the range the analyst had computed but never shown. The single number was not wrong; it was the honest midpoint of a wide interval. What was wrong was presenting it naked, as if the analysis knew the future to the unit. Stripping the uncertainty off a result does not make the result more certain — it just moves the surprise from the slide to the warehouse. This module is about putting the uncertainty back on, honestly, and treating "we don't know yet" as a finding rather than a failure to conceal.
Report "40,000" and you have made a claim: that the truth is close enough to 40,000 that the difference does not matter for the decision. Sometimes that is fair. Often it is not — the estimate carries a range so wide that 40,000 and 31,000 are equally consistent with what you actually know, and reporting only the midpoint hides that fact from the person planning around it. The point estimate is not a lie, but presented alone it lets the audience believe in a precision the analysis never had.
ⓘ Concept: The number without its range overstates what you know
Why it matters — Every estimate from finite, noisy data has a spread, and that spread is part of the finding — arguably the more decision-relevant part, because it says how much to trust the headline. A point estimate presented alone quietly asserts that the spread is negligible, and the audience, having no other information, plans as if it were. When the outcome then lands somewhere else in the range the analyst always knew about, the failure looks like a bad forecast but is really a communication failure: the uncertainty existed, it was simply not shown. Reporting the range is not hedging or a lack of confidence; it is reporting the actual state of your knowledge instead of a flattering compression of it.
🗣️ From a linguist's perspective: Confidence you can't honestly certify
Presenting uncertainty honestly is the one-slide answer applied to what you do not know: the range leads alongside the number, the display is calibrated to the decision, and "we don't know yet" is allowed to be the finding. The final module turns to the words around the numbers — the prose, and the discipline of never letting it claim more than the data supports. Step back any time to the P14 pillar overview to see the chart-craft and communication tracks together.
Going deeper (technical) — prediction interval vs confidence interval
Interview-grade notes for the technical (it-ml) track. The framing above is audience-broad; the it-ml distinction below is one that presenters routinely get wrong, and mislabelling the band is a quiet way to overclaim.
A confidence interval and a prediction interval answer different questions, and the prediction interval is almost always the one your audience cares about. A confidence interval brackets a parameter — where the true mean or true slope lies — and it narrows toward zero as the sample grows, because with enough data you pin the average precisely. A prediction interval brackets the next observation — where a single future value will fall — and it does not shrink to zero however much data you have, because it must include the irreducible spread of individual outcomes around the mean. Confusing them is how a forecast comes to be shown with an implausibly tight band: someone plotted the confidence interval of the mean when the planner needed the prediction interval for the actual quantity they will provision against.
A fan chart's widening comes from two stacked sources. Part is estimation uncertainty (you do not know the model's parameters exactly) and part is process uncertainty (even a perfectly known model has random shocks that accumulate over the horizon). For most forecasts the second dominates far out, which is why the band keeps widening even when the model is well estimated — and why a fan chart that stops widening is usually one that has quietly dropped a source of uncertainty it should have kept.
State the coverage, not just the band. "80% interval" and "95% interval" are different pictures of the same forecast, and a band shown without its coverage level is uninterpretable — the reader cannot tell whether one point outside it is expected (a 95% band should be breached about one time in twenty) or alarming. The honest caption names the coverage; the forecasting pillar (P7) covers how to check that a claimed 80% band actually contains 80% of outcomes, which is calibration, and an uncalibrated band is worse than none because it is confidently wrong about how confident to be.
Ask the mentor about this module
Ask a question about this content. The mentor explains and grounds its answer in what you are studying; asking is recorded as a learning signal, not a grade.
Try it yourself
A scratch console for this page's ideas — ungraded, nothing you run here is recorded.
Scratch console
A scratch console with the scientific stack (pandas, numpy, scikit-learn). Runs on the server — no network, resource-limited and measured.
Output appears here.
Where next?
Later in Communication