Proof
Model Health
One number judges the model: the Brier score — the average squared gap between what was announced and what happened. 0 would be perfect, 0.25 is a coin flip. It is published here competition by competition, including where it is bad.
- 300
- judged analyses
- 1
- Champion model
- 2015-01-01/2020-12-31
- Trained on
- 0.234
- Brier Score
Where the model is weakest
These competitions have held below what they announced. They are flagged on every fixture, their strongest badge is withheld, and they are excluded from the daily proposals — automatically, without anyone deciding it case by case.
| Competition | Brier | Judged | Gap to announced |
|---|---|---|---|
| Primeira Liga | 0.233 | 22 | -12.0 pp |
| Serie B | 0.234 | 26 | -10.0 pp |
| Premier League | 0.246 | 77 | -10.0 pp |
| K League 1 | 0.277 | 22 | -17.0 pp |
| Primera A | 0.287 | 26 | -22.0 pp |
| 2. Bundesliga | 0.302 | 22 | -11.0 pp |
Where it holds
| Competition | Brier | Judged | Gap to announced |
|---|---|---|---|
| Czech Liga | 0.142 | 24 | +9.0 pp |
| Liga Profesional Argentina | 0.166 | 35 | +5.0 pp |
| UEFA Champions League | 0.175 | 16 | +20.0 pp |
| Superliga | 0.176 | 14 | +15.0 pp |
| Eredivisie | 0.184 | 25 | +11.0 pp |
| La Liga | 0.186 | 24 | +5.0 pp |
| Süper Lig | 0.189 | 22 | +13.0 pp |
| Super Liga | 0.191 | 18 | -5.0 pp |
What this measurement does not say
- Measured on 300 resolved {n|prediction|predictions}. A sample this small moves by several points on a single result — read every figure below with that in mind.
- Only the headline bet of each analysis is scored, against the same verdict as the track record. Markets the model does not price are absent from this page, not hidden.
- The headline bet is by construction the model's best read, so it is never published below 50%: the lower probability bands carry no observation and are left out of the calibration curve rather than drawn through empty space.
- Performance by horizon and drift monitoring need prediction snapshots taken over time. Nothing records them yet, so both are empty rather than estimated.
- Per-competition figures include groups of one or two matches. Read the sample before the Brier score.