What we’ve measured

A gradient-boosted ranker trained on real bidding behaviour puts 21.9% of a firm’s actual next-window bids in its top ten, against 9.4% for deterministic trade-and-region rules alone. Embedding similarity alone reaches 14.2%; a compact linear model 16.5%.

Splits are temporal, never random — a random split lets a model learn from bids placed after the ones it is asked to predict. Every firm-history feature takes an as-of date and excludes anything dated on or after it with a strict inequality.

The cohort result is the interesting one, and it inverts the expectation. Firms with 20–49 observed bids rank best (recall@10 0.250). The heaviest bidders — 50 or more, and 334 of the 400 firms evaluated — rank worst at 0.215, below even the 5–19 cohort at 0.222. More history does not mean more predictable: a firm bidding hundreds of times a year across many categories is genuinely harder to anticipate than a specialist, because its next bid is drawn from a much wider distribution.

Two figures that are easy to mix up: on a later-window split (settled, 2025-10-01) the same model reaches 2.8× the deterministic baseline, but at recall@10 0.217. The 0.219 and the 2.8× come from different splits — quoting them together overstates the result, so we quote the primary split throughout.

Measured quality is understated by an amount we cannot quantify: only observed bids are labelled, so a tender a firm would have wanted but never saw counts against us.

Under 1% of open notices publish a value — 241 of 48,834 — so we learned contract size from what comparable work actually went for.

The obvious method does not work, and that is the finding. A median lookup by trade × buyer type × region — built from 168,541 real contracts — scores 34.9% exact-band accuracy. Always guessing the single most common band scores 34.4%. The lookup beats a constant by half a percentage point, and is actually worse on the within-one-band measure.

Knowing the trade, the buyer type and the region tells you almost nothing about contract size. A municipal watermain job is $80K or $4M depending on how many metres of pipe, and none of those three fields carries that.

What does carry it is the title. Adding a multilingual sentence embedding of the notice title lifts exact-band accuracy to 41.7% — a 7.3-point gain over guessing, entirely from wording. Scope plurals, facility nouns, phase markers and date ranges are what separate a programme from a job.

41.7% exact means the specific band is wrong more often than right, so 92.1% within one band is the number worth trusting. Roughly 44% of open notices get no band at all, because we would rather say “unknown” than force one.

SEAO publishes in French and CanadaBuys in English, and both have to share one vector space or half the corpus is noise. The model does carry meaning across the boundary: French égouts pluviaux scores 0.51 against English watermain replacement and 0.22 against French mobilier de bureau. Meaning beats language.

But the effect is much weaker than same-language similarity, and that gap caused a real defect. Our open pool is roughly 65% French SEAO notices, so trade centroids built from it are French-dominated. Cosine against them partly measures how French a title is: English notices carry a systematic penalty of about 0.2 — enough that an English watermain notice scored below an English janitorial contract for a watermain firm.

Within one language the similarity is meaningful. Across languages it cannot carry a threshold on its own. So eligibility is decided by trade agreement, which is language-independent, and the similarity floor is only a backstop. Per-language centroids would fix the confound properly and are a known follow-up rather than a solved problem.

Québec publishes public procurement as open data under CC-BY 4.0, including results: for a completed procurement, who bid and what they bid. Ontario publishes nothing systematically comparable.

The consequence is not subtle. Of the contracts we hold with a known winning bid amount, 199,644 are Québec and 9 are Ontario. That is not a sampling artifact or a gap we could close by trying harder — it is what each province’s publishing policy produces.

Everything on this site that learns from bidding behaviour learns it from Québec, because that is the only place the behaviour is recorded. Ontario firms get a weaker product for the same effort, and the ranking and size models are trained on one province and applied to another.

We should be straight about our position: we benefit from Québec’s openness and are constrained by Ontario’s closure, so we are not a neutral party. The number is accurate either way, and you should weigh the argument knowing who is making it.