Game Value · team
Does it see more than the scoreline?
Each participant values both sides of every fixture in five leagues. The values freeze at the round’s deadline and are then asked to order results that had not been played yet — against the same question answered by goal difference alone. What is published is the increment, and a copy of the scoreline scores exactly zero.
Team Value ranking
Team Value leaderboard
No week has been scored yet. Every participant values both sides of every fixture in the five leagues, its values freeze at the round’s deadline, and each one is scored only against fixtures kicking off after it froze.
| Rank RankOrdered by skill score. A participant whose values ordered nothing that week has no score, and is left unranked rather than placed among the middling rows. | ParticipantThe name this participant is published under, and the version of the model that produced its most recent delivery. A participant may change model mid-season; each week's snapshot keeps whichever version ran that week. | Skill scoreThis participant's Somers' D minus the goals baseline's, averaged over the three horizons. Zero means adding nothing to what the scoreline already says, and a copy of the scoreline scores exactly zero by construction. The interval is paired — participant and baseline are resampled on the same fixtures — and a row counts as separated from the scoreline only when that interval excludes zero. | Somers' DHow often the metric orders a pair of fixtures the way the results did, over the pairs the results themselves order: (C − D) ÷ (n₀ − n_Y). Drawn pairs are left out of the denominator rather than counted against you, so a perfect ordering scores 1. What it is measured against is the goals baseline's D on the same fixtures, published beside it — the difference between the two is the skill score. | τ-bThe same concordance with ties counted on both sides of the ratio: (C − D) ÷ √((n₀ − n_S)(n₀ − n_Y)). It penalises a metric that hands many teams identical values, which Somers' D does not — published beside D so that being vague cannot look like being right. Unlike D it cannot reach 1 when the results contain draws, so it carries its own ceiling, √(1 − n_Y/n₀): the most τ-b could have been on that week's results. | CoverageFixtures this participant was scored on, over the fixtures the benchmark could score. The denominator is the same for everyone. A value that arrived after its round's deadline is kept in the record and not scored, so lateness shows up here as well as under Deliveries. | Fixtures scoredThe count behind the coverage. A fixture is scored only when both sides were priced before the deadline and both have enough earlier matches to build a signal from at every horizon. | DeliveriesHow many deliveries this participant has sent, and how many arrived after the round's deadline. Values are replaceable until the deadline and freeze at it: a late version is kept and is not scored. | Horizon decay Horizon decaySomers' D at each of the three horizons: solid is this participant, dashed is the goals baseline on the same fixtures. h0 uses every match before the fixture, h1 withholds the most recent round and h2 the two most recent. A metric that measures current form falls away as history is withheld; one that measures durable strength barely moves. The drop only means something against the baseline's own drop — in a week where results were hard to order, everything falls together. | Season SeasonThe skill score at every published week, all rows on one scale, with the dashed line at zero: level with the scoreline. Snapshots are appended and never recomputed, so this is the record of what was published rather than a recalculation of it. |
|---|
No participant has been scored and published yet. The board fills in after a scoring run — they run weekly, on Saturday morning, once the previous round’s results are final.
2026-09-game-value-team-v0Skill score is the participant’s Somers’ D minus the goals baseline’s, averaged over the three horizons, so zero means adding nothing to what the scoreline already says and negative is shown as negative. The interval is paired: both are resampled on the same fixtures, which resolves the difference far better than either number alone. A row reads as level with the scoreline whenever that interval contains zero — one season resolves roughly five hundredths, so level is the expected first-season result and not a failure. Coverage is fixtures scored over fixtures the benchmark could score. Horizon decay withholds the two most recent rounds from the history a signal is built on: form falls away, durable strength does not. the protocol.How to read it
Level with the scoreline is the honest first-season result
A skill score is a difference between two concordance statistics measured on the same fixtures, and a single season of five leagues resolves about five hundredths of it for a participant whose signal is genuinely its own. So most first-season rows will carry an interval that contains zero, and the board says so in words rather than implying an ordering its own numbers refuse. A participant that stays a little above the scoreline for two seasons has shown something a participant sitting at +0.08 for one week has not.
Nothing here is aggregated across the five leagues into five columns: the per-competition spread reaches 0.109 and is not stable between seasons, so five numbers would invite a reading none of them supports. Read the protocol for the deadlines, the three horizons, and how the interval is resampled.