How wrong is the model? A leave-one-out test

28 WorldTour riders have a base salary published somewhere in the press. The model uses those as fixed anchors, so on the site they are never "estimated". Here we remove each one in turn, re-run the model (v0.7 (15 Sep 2026)) as if that salary were unknown, and compare the guess with the reported figure. This is the only honest test we can run with public data, and it is published so nobody has to take an error bar on trust.

Rider by rider

Reported figures: Gazzetta dello Sport Jan 2026 / Cyclingnews rich list Apr 2025 / Marca Oct 2023 / Cronica Vasca Mar 2024 / Eurosport-RMC Apr 2024 (Martinez) / NOS Dec 2025 (Kooij) / L'Equipe Oct 2025 (Vauquelin) / bici.PRO Sep 2025 (Ayuso). "Model, salary hidden" is the estimate produced when that rider's reported salary is removed from the anchors and the full model is re-run; the other 27 anchors stay in place. Script and CSV in the data download (validate_v07.py, validation_loo.csv).

What it says

• The typical miss is 24.9%. 14 of 28 riders land within ±25% and 19 within ±40%. The range shown on a rider page is his actual pay band for the job, which is wider than this for a reason: these 28 are the best-documented riders in the sport.
• The bias is gone. The median error is 2.3% and the 28 together are paid €80.6M against €80.4M modelled. Earlier versions, which split a pool by points, ran about 19% low on stars; pricing by job and by the year of signing fixed that.
• The big misses are teams that don't pay the market rate. VAUQUELIN Kévin is the clearest: L'Équipe puts him on €700–900k at Ineos, while the band for a second-tier leader at a €52M team says about €2.9M. That is not noise — it is Ineos buying a rider before his price caught up. Landa, Geoghegan Hart and Carapaz go the other way: paid as leaders on deals signed when they were, at teams whose wage bill can no longer carry it.
• Van der Poel is a structural miss. The model sees Alpecin's top scorer as Philipsen and puts Van der Poel a step below him, so it prices him at half his reported €4M. A rider who is the reason the team exists doesn't fit a pyramid built from results.
• Where results and job line up, it works. Vingegaard, Philipsen, Kooij, Hindley and Evenepoel all come out within 13%.

Why this test is easier than reality

• When one anchor is removed, the other 27 stay. On UAE or Red Bull the remaining reported salaries still pin down most of the team's wage pool, so the model is guessing one number inside a well-constrained box. For a mid-table rider on a team with no reported salaries at all, the box is much bigger.
• All 28 testable riders earn €2M or more. Not one reported salary exists for a €150k domestique, a €300k road captain or a €600k second leader — the zone where most of the peloton lives and where we expect the model to be weakest. This test says nothing about that zone. If you know a number there, the correction form is the most valuable thing you can do for this site.
• The pay bands themselves come from published market reporting, and the tier rules were written while these 28 salaries were known. No parameter was fitted to them numerically, but the design is not blind to them either.

What changes next

Three things would move this the most, in order. More reported figures below €2M — every list that exists stops at the top twenty, which is why the middle of the peloton is untested. Announcement dates for the riders we still infer, since pricing a contract to the wrong year is the single biggest source of error. A rule for riders bought below their market, which is currently just an error we report rather than something the model understands. If you know a figure, the correction form is the most valuable thing you can do for this site.