Prediction band

A calibrated uncertainty range, not a single guessed number. A prediction (or conformal) band is a range the true value should land inside a stated share of the time — 80% coverage means the real number should fall inside roughly 8 times out of 10.

Why it matters

A single forecast number hides how sure it is; a band shows it. Width is the honesty: a tight band claims a lot, a wide one admits doubt. The Cost Index publishes its derived forecasts as a band targeting 80% coverage (a split/EnbPI conformal method, widened until it holds). Walked forward across 84 ingredients and 5,929 scored steps, the band actually held about 84% of the time — covering a touch more than promised, which is the honest direction to err.

Watch the 90-second explainer

Frequently asked

What is a prediction band?

A prediction band is a range, not a single number. Instead of guessing one figure and pretending it's exact, you publish a low-to-high range and a promise about how often the real value should fall inside it. A band built for 80% coverage is saying: the true number should land inside this range about 8 times out of 10. The honest part is that the range admits what it doesn't know — a tight band claims a lot, a wide band admits more doubt, and either one is more useful than a lone number with no error bars.

What does conformal mean, and why 80% coverage?

Conformal is a way of setting the band's width from the method's own past misses rather than from an assumption about the shape of the data. You score how far off the forecast has been historically, then size the band so it would have covered the target share of those cases. Eighty percent is the coverage target the Cost Index aims for: wide enough to be honest about week-to-week noise, tight enough to still be a useful range an operator can plan against.

How does the Muntin Cost Index use a prediction band?

The Cost Index publishes its derived forecasts as a band targeting 80% coverage, built with a split/EnbPI conformal method that is widened by adaptive conformal inference until it holds. Walked forward across 84 ingredients and 5,929 scored steps, the band actually contained the next print about 84% of the time — slightly more than the 80% target. Covering a touch more often than promised is the honest direction to err: the index would rather be caught being a little too cautious than caught being overconfident.

Prediction band vs confidence interval — what's the difference?

They answer the same question — how much could this be off? — but build the range differently. A classic confidence interval assumes a distribution; a conformal prediction band, the kind the Cost Index uses, sets its width from the method's own past misses instead. That is why it can promise a coverage rate (80%) and then verify it walking forward (about 84%).