Leading on Brier so far: OddsIntel (own model) at 0.615 across 5 settled.
Same matches, same outcomes, four predictions — scored identically. Re-computed nightly as matches resolve.
Sorted by Brier (lower = better). Re-computed nightly as matches resolve.
| # | Source | N | Brier | Log-loss | Hit-rate | CLV |
|---|---|---|---|---|---|---|
| 1 | OddsIntel (own model) | 5 | 0.615 | 0.999 | 60% | -0.2pp |
| 2 | OddsIntel (blended with market) | 5 | 0.640 | 1.025 | 60% | -0.1pp |
| 3 |
Same fixtures, same outcomes. Each source is scored only on the matches it had a row for, then sorted by mean Brier (lower is better). N tells you the sample size — early in the tournament a strong-looking source on 3 matches isn't conclusive.
Brier = mean squared error of the (home, draw, away) prob vector against the one-hot actual outcome. Log-loss = −log(p) on the eventual outcome, averaged over matches. Same definitions as the Summary page.
CLV= avg of (source's prob on the actual outcome) − (market's prob on the actual outcome). Market is the reference baseline, so it has no CLV against itself.
Opta: we don't scrape Opta's daily-updated articles yet. The row is included with a "Coming soon" note so the table reflects the full intended comparison set.
Re-computed nightly as matches resolve.
| 5 |
| 0.727 |
| 1.145 |
| 60% |
| baseline |
| — | Opta supercomputer Coming soon — scrape not yet wired up. | 0 | — | — | — | — |