How often we are right, and what we are not right about.

01 — Does the range hold the price

Two ranges are published and each is calibrated separately against its own target. Neither is derived from the other.

What the calibration bought

The model's raw quantiles under-cover badly. A single shift applied everywhere corrects the average but widens the range by a third. Splitting by band corrects the average and makes the range narrower than it started — that is the one that is published.

The tight range follows the same path: 41.9% raw against 49.4% published, for a target of 50%.

02 — Band by band

A correct average coverage can hide two halves wrong in opposite directions. So the calibration is done per predicted-price band, and each band is checked separately.

The target we hold ourselves to is coverage between 75% and 85% in every band. One band misses it on this measurement, and it is marked below. We publish it rather than smoothing it away: an accuracy page that shows only its successes measures nothing.

Measured on the tenders whose range rests on the model alone. Where the buyer publishes an estimate, it is built differently and measured separately — that is the next section.

03 — Why the bands are predicted-price bands

Because it is the only one you know at the moment of deciding — and because banding on the real price mechanically produces bad numbers, whatever the model.

Group the tenders by the amount finally awarded and coverage collapses at both ends. That is not a defect of the model: conditioning on the outcome selects, in the low band, precisely the contracts that went for less than expected. Any predictor, even a perfect one, would read the same. The table is here so that nobody has to discover it elsewhere.

04 — On which tenders

The page average mixes two very different situations. Which one applies to you depends on what the buyer published with the notice — and the report tells you which is yours.

The “width” column is the quantiles before calibration: the band-calibrated width is not measured separately on each situation, and giving one for the other would be a correct number under a wrong label. Across the whole set, calibration tightens the median width from 4.3× to 4.2×.

A notice with no published estimate is therefore measured at around 54.3% median gap against 20.2% when there is one. The report shows which of those two situations is its own — the reliability indicator, beside the range, and not in a footnote.

05 — And if the range is too wide

A range that is correct but very wide is no use to you. It is the most well-founded objection to this method, so here is the full distribution rather than its median.

The report places your range inside this distribution rather than leaving you to guess, and it refuses to draw one when the notice does not carry enough information to do better than an order of magnitude. A wide range is an admission of uncertainty; a very wide one is no longer a measurement at all, and it is better to say so than to draw it.

What the report does with that width

That last row is the notices with no buyer estimate and no usable specifications. For those we have no measured accuracy regime at all, and it is that fact — rather than the width of the interval — that decides. Those notices do not necessarily have the widest range; some are narrower than the quarter flagged above. An interval that looks confident on a contract we do not know how to value is the dangerous case, not the one that looks wide.

06 — What it compares against

Two simple methods act as a permanent guard rail. The first is the one anybody can apply; the second beats us on its own ground, and that is why we use it.

Our range is 7.2× narrower than the obvious method, at comparable coverage, and is wrong half as often.

The third row beats us, and we publish it for that reason. On the 48% of notices where the buyer published an estimate, a simple correction applied to that figure does better than our 285 variables. We do not ignore it: the range published on those notices is a blend of the two, weighted on held-out data. A model that refused to use better information because it was not its own would be a bad product.

07 — The report's other forecasts

Price is not the only thing the report forecasts. The others are measured the same way, on the same split.

08 — Where these numbers come from

What these measurements do not say: what the contract costs you, what your order book is worth, or what gets decided outside the public process. The limits are written out in full in the terms of use, and the method in detail is in the report's description.

Get started

The report carries these same measurements beside every figure, on your tender rather than on the corpus.