NoFuckery AI logo

NF-023 / Evidence brief / version 1.0 / 2026-08-06

NoFuckery AI

Verdict

Four supply-chain planning vendors publish forecast-accuracy figures. None names the formula that produced the number — and by one of those vendors’ own guide, the formula alone can treble the error reading on a single forecast.

Accountable editor: Chris McNosky · No commercial relationship with any vendor named · Five vendor pages, one each, read 2026-08-05 and 2026-08-06

The record

StatusClaimReceipt
direct observationRELEX states that formula choice alone moves the reading: A forecast might show 95% accuracy by one measure while simultaneously revealing a 15% error by another. That is 5% error against 15% — threefold, on one forecast.Measuring forecast accuracy: The complete guide
direct observationThe same guide says two valid readings answer different questions and should never be compared.Same guide
direct observationThe guide also publishes typical accuracy ranges: high-volume stable products 75–85%, slow-movers 50–70%, fresh weather-sensitive 70–80%.Same guide
direct observationRELEX publishes: Biyoute has achieved weekly forecast accuracy of 89% at the regional level, with grocery forecasts exceeding 92%. Both figures sit above the ceiling RELEX itself publishes for the most forecastable class of product. Level and period stated; formula not stated; no baseline.Biyoute announcement
source reportedA UNFI speaker, quoted by RELEX: a twenty five percent improvement in our forecast error rate — error rate, not accuracy, and a customer statement rather than a vendor measurement. Same speaker, same passage: It’s still early, though.UNFI case study
direct observationBlue Yonder, under Key Benefits — This is what you could achieve: businesses can achieve a 12% improvement in forecast accuracy. Formula not stated; relative or absolute not stated.Blue Yonder
direct observationo9: Forecast accuracy improved by more than 11 percentage points to 87%. ToolsGroup: Lift forecast accuracy 5-15 points. Both use points rather than percent, which is the more precise form. Neither names the formula.o9 · ToolsGroup
direct observationKinaxis publishes no quantified accuracy claim on its demand planning page.Kinaxis
inferenceOf the four vendors publishing a quantified accuracy claim, none names the formula. No published figure can be placed on RELEX’s own stated spread, or compared with another vendor’s.Follows from the rows above.

The gap

RELEX’s methodology guide prescribes matching aggregation to the planning decision. Biyoute’s figure names its level and period. o9 and ToolsGroup express gains in points rather than percent, which removes the relative-or-absolute ambiguity Blue Yonder leaves open.

None of the four quantified claims names the formula.

These figures are published to be compared. RELEX’s guide states that two valid readings of one forecast should never be compared.

Naming the formula makes the numbers comparable. Not naming it makes them incomparable. A category whose published figures cannot be checked costs buyers and protects sellers.

Between vendors the effect inverts. A vendor whose figure holds up under a strict formula would gain from naming it; non-disclosure removes that advantage. The fix is one clause in the same sentence as the number.

Who carries it

The vendor ran the calculation and knows which formula produced its number. The buyer cannot determine it from the published material, and makes the selection anyway.

A percentage without its formula cannot be interpreted or compared to another. Four of the five vendors reviewed publish them that way.

Four questions before signing

  1. Which formula produced this number — MAPE, WMAPE, bias, something else — in writing?
  2. At what aggregation level and horizon, and does that level match the decision this system will automate?
  3. What is the baseline: the prior system, or a naive forecast?
  4. Over what period, and does it include peak?

The first is the one that settles it, and the only one none of the material reviewed here answers.

Operator decision

Do not enter a forecast-accuracy percentage into a vendor comparison until the formula behind it is stated in writing. Where two figures came from different formulas, treat them as different quantities and refuse the comparison — on the vendors’ own published reasoning.

Tell me where this is wrong

This brief reads one public page per vendor. It does not know what is in a datasheet, an RFP response, or a room. If you have sold these systems, bought them, or measured forecast accuracy for a living and something here is wrong, say so in public and I will record the correction with your name on it.

The two weakest claims here: the Biyoute figure is set against a range whose own formula is unstated, so the comparison may not be one; and absence of the strings MAPE and WMAPE from a marketing page is thin evidence about what a vendor measures.

Provenance

Every quotation was taken from the raw HTML of the cited page, retrieved and searched on 5 and 6 August 2026, not from search summaries. Vendor pages change without notice; those are read dates. Corrections are recorded at Corrections.