A deterministic rule: one quarter squad strength, one quarter squad form, one quarter team cohesion, one quarter modelled output. No fitting — every constant is fixed in advance, so the same inputs always give the same card.
10 fixtures on 2026-10-02.
3 fixtures · home edge +0.073 index units
7 fixtures · home edge +0.073 index units
| Predictor | Correct | Hit rate | Log-loss |
|---|---|---|---|
| Composite (1/4 blend) | 51 | 52% | 1.0087 |
| Served model (N1) | 59 | 60% | 0.9210 |
98 graded fixtures · log-loss difference +0.0877 · model ahead
Lower log-loss is better. Both predictors are frozen before kickoff on the same fixtures, so this is a fair comparison rather than a retrospective one. A hit-rate lead with a worse log-loss means the blend picks winners but is poorly calibrated about it.
Minutes-weighted, shrunk mean match rating of the nation’s player pool over the last 730 days.
The same pool’s rating over the last six internationals, minus that nation’s own baseline — so it measures a team playing above or below itself, not its level again.
0.6 × starting-XI continuity + 0.4 × normalised caps-together.
The pooled international rating that serves the ordinary fixture probabilities.
Each is turned into a cross-nation z-score so the quarters are comparable before averaging, then the index difference plus a home edge goes through a fixed curve. A nation missing a quarter takes the pool average for it and the fixture is marked, so you can always see how much of the card is real — presently 165 of 217 nations carry squad strength, 98 carry form and 154 carry cohesion.
Stated here rather than buried: two of these four quarters are built from signals that failed their pre-registered tests. Squad form was worth −0.0012, −0.0018, +0.0004 and +0.0004 across four datasets of growing size; cohesion was worth +0.0140, +0.0010 and +0.0008. Neither has shown predictive value and cohesion has tended to hurt.
That does not make this card wrong — those signals were tested as additions to a rating model, not as a deliberately weighted blend, and there are good reasons to want a view the log-loss does not vote for. It does mean the blend should earn its place rather than be assumed into it, which is what the scoreboard above is for. The ordinary international fixture probabilities elsewhere on the site are unaffected: they remain the pure model.
Built 2026-10-02 21:19. Cards are frozen before kickoff and written to a forward-only ledger with the model’s numbers for the same fixture.