Market Data

How Far Ahead Can You Trust a Manufacturing Forecast? We Publish the Error Bars

A forecast without an error band is a guess wearing a suit. Here is why honest forecasts widen as they reach further out, how conformal bands are built, and why we publish our misses.

Every forecast is wrong by some amount; the only honest question is how wrong, and with what confidence. A point forecast, a single line into the future, hides that entirely and invites false precision. The forecasts here ship with conformal prediction bands instead: a range around the projection with a stated coverage level, so a reader sees not just where a series is expected to go but how much uncertainty surrounds that expectation. For a series like the steel PPI, currently 361.44 index (1982=100) (Jun 2026) and climbing, the band is the difference between a usable planning tool and a number that pretends to know the future.

Why the band widens with the horizon

An honest forecast band is narrow next month and wider a year out, because uncertainty compounds. Next month's value is anchored by the current one and by short-term momentum that rarely reverses instantly. A year out, more can intervene: a policy change, a demand shock, a supply disruption, each adding to the range of plausible outcomes. A forecast whose band does not widen with distance is not more confident; it is less honest, because it is refusing to acknowledge the uncertainty that genuinely grows the further out you look. The widening cone is a feature, the visual statement of how much the horizon costs you in certainty.

What conformal bands actually promise

Conformal prediction is a method for building bands with a coverage guarantee that leans on the model's own track record rather than on assuming the errors follow a neat bell curve. In plain terms: a 90% band is calibrated so that, historically, the true value landed inside it about 90% of the time. That makes the band auditable, you can check whether reality fell inside it as often as advertised, and it is only trustworthy because the misses are published alongside the hits. A forecast track record that shows only the good calls is marketing; one that shows the misses is a measurement.

A point forecast says trust me. A forecast with a calibrated band and a public miss rate says check me. Only one of those is a measurement.

How to use a banded forecast

The disciplined use is to plan against the band, not the line. Budget for a range of outcomes rather than a single point; stress-test decisions at the edges of the band, not just at its center; and lean harder on near-term projections where the band is tight than on far-term ones where it is wide. A banded forecast is not a promise about a single future value, it is an honest map of the plausible ones, and that is exactly what makes it usable for a decision that has to survive whichever outcome actually shows up.

Why a six-year record humbles forecasts

The 36-year record shows steel making a full round trip, which is why point-in-time comparisons mislead so badly here. Its high came at the close of 2021 around 449.71, gave way over the following years to 266.51 by the end of 2024, and has climbed since to 361.44. That leaves it 20% below the peak and well off the floor, so whether today looks high or low depends entirely on which year you anchored to.

View the banded forecasts and their published accuracy track record on the pulse forecast page. See the forecasts

Published 2026-08-05.