What a $200K job looks like vs a $2M job

Here is a question that should be easy and is not: looking at a tender notice, how big is the job?

For almost every notice, the document does not say. Under 1% publish a value — 241 of the 48,834 notices we hold. Everything else is a title, a buyer, and a closing date. So we tried to learn size from what past contracts actually went for.

The obvious approach barely works

The natural first move is a lookup table: group past contracts by trade, buyer type, and region, take the median of each cell, and quote it. It is transparent, explainable, and easy to defend.

It is also, on a held-out time window, almost worthless.

Read the first two rows again. A lookup table built from 168,541 real contracts beats always saying “$100–500K” by half a percentage point — and it is actually worse on the within-one-band measure.

That is not a tuning failure. It is the finding: 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: the title

The model that beats the constant does so by reading the title. Adding a multilingual sentence embedding of the notice title lifts exact-band accuracy from 34.4% to 41.7% — a 7.3-point gain over guessing, entirely from wording.

What a big job says about itself

• Plurals and scope words. “diverses rues” — various streets — is a programme, not a job.
• Facilities rather than components. “station d’épuration”, “treatment plant”, “centrale”. A plant is an order of magnitude above a pipe run.
• Phases and lots. “Phase 2”, “lot 3” — a job large enough to be cut up.
• Multi-year spans and standing offers. A date range in the title is a programme value, not a project value.

And a small one

• A single named object. One culvert, one street, one structure.
• Study, inspection, expertise. Professional services on a specific question price in the tens of thousands.
• “Travaux mineurs”, “minor works”. The document is telling you directly.

The honest limits

Two further limits, stated because they bound what any of this is worth:

• The corpus is Québec. 199,644 of 199,714 priced awards are Québec; Ontario has nine. An estimate on an Ontario notice is an inference from Québec comparables. See why that is.
• The inflation adjustment uses the wrong sector. We restate amounts in current dollars with StatCan’s building construction price index (table 18-10-0289-01, Québec non-residential, 2023=100) because Statistics Canada publishes no active engineering construction index — 18-10-0022 ended in 2019 and 18-10-0096 in 1993. The adjustment matters a great deal ($1.00 in early 2018 is $1.62 today), and it is a proxy.

What to do with this

The practical takeaway is not our accuracy number. It is that the title is the most informative thing on an unpriced notice, and most people skim it. Before you open a document, read the title for scope words, facility nouns, phase markers, and plurals. You will sort a list of notices by likely size faster than any lookup table can.

You can see the estimator working on today’s open notices — with every band labelled by where it came from — on the ranker.