Synthetic fixtures · Non-empirical · Versioned

An open conformance benchmark for football prediction scores

Direct answer: this dataset tests whether an implementation reproduces 54 declared 1X2 scoring cases. It is not real match data, a model leaderboard or evidence that any football prediction system is accurate.

Published by Football Proof AI · Published · football-1x2-scoring-benchmark/1.0.0 · Editorial technical note; not externally peer reviewed

Canonical publication record

Abstract

An open, deterministic conformance dataset of 54 synthetic football 1X2 cases with expected top-pick, summed and class-averaged Brier, exact and explicitly clipped log-loss, and normalized H-D-A Ranked Probability Score outputs.

Author and publisher
Football Proof AI
Technical report
football-1x2-scoring-benchmark/1.0.0
Published
Last modified
Release status
Current release
Review status
Editorial technical note; not externally peer reviewed
Version history
  1. football-1x2-scoring-benchmark/1.0.0 : Initial public release.
Immutable artifacts
  1. football-1x2-scoring-benchmark-v1.csv sha256:a2d7dc9aec8faae9b62562b5f50de68f3e6bb787c2619539c057e6beeb4d7128
  2. 1.0.0.json sha256:2481c8c2d42e88333fb2b92494a023692830e8e560dad146f132d45d5dc3fc74
References
  1. https://doi.org/10.1175/1520-0493(1950)078%3C0001:VOFEIT%3E2.0.CO;2
  2. https://doi.org/10.1198/016214506000001437
  3. https://doi.org/10.1175/1520-0450(1969)008%3C0985:ASSFPF%3E2.0.CO;2
  4. https://doi.org/10.1515/1559-0410.1418
  5. https://doi.org/10.1007/s10994-018-5726-0
  6. https://doi.org/10.1007/s10994-024-06625-9

54 fixed cases · Two immutable formats

Download the same benchmark bytes every time

Eighteen probability vectors are crossed with home, draw and away outcomes. Both files declare the formulas, tie rule, numerical boundary and release version; response headers expose their SHA-256 byte identities.

License: CC BY 4.0. Cite the exact version and preserve the artifact SHA-256 when results must be reproducible.

Truth boundary

Conformance is narrower than model validation

Matching the expected cells supports only the claim that code reproduced these cases under these conventions. It cannot prove historical publication, archive completeness, calibration, predictive skill, bookmaker advantage or profit.

  1. 01

    Synthetic

    No row names a real fixture, team, league, provider, price or trained model.

  2. 02

    Non-empirical

    The cases reveal mathematical behaviour; they estimate no population performance.

  3. 03

    No metric winner

    A paired example can show disagreement without proving one scoring rule is statistically superior.

  4. 04

    No external review

    The version and hash identify this publication; they do not create independent replication or peer review.

Representative rows

Where the scoring rules agree—and where they answer different questions

The table shows eight of 54 cases. Full-precision expected outputs and all declared pairs are in both downloads.

Selected synthetic v1 cases; lower is better for every probability score shown
CaseH / D / AResultTop hitBrier avgLog lossRPS
uniform-h0.333333 / 0.333333 / 0.333333Htie—excluded0.2222221.0986120.277778
uniform-d0.333333 / 0.333333 / 0.333333Dtie—excluded0.2222221.0986120.111111
strong-home-h0.8 / 0.15 / 0.05Hyes0.0216670.2231440.021250
ordinal-home-near-h0.6 / 0.35 / 0.05Hyes0.0950000.5108260.081250
ordinal-home-far-h0.6 / 0.05 / 0.35Hyes0.0950000.5108260.141250
certain-home-d1 / 0 / 0Dno0.666667+∞0.500000
certain-home-a1 / 0 / 0Ano0.666667+∞1.000000
tie-home-draw-d0.45 / 0.45 / 0.1Dtie—excluded0.1716670.7985080.106250

Locked conventions

The benchmark never silently changes scale or handles zero for you

Brier sum

Σ(pᵢ − yᵢ)²

The class-averaged companion is exactly the summed value divided by three.

Exact log loss

−ln(p observed)

Zero probability is positive infinity, represented by a blank numeric cell plus an explicit status.

Clipped log loss

ε = 1e-15

The finite numerical variant is a separate column, never a replacement for the exact score.

Normalized RPS

H → D → A · divide by 2

The declared order is an analytical choice, not a claim that RPS is football's uniquely correct metric.

For definitions and sample-level interpretation, read the football prediction accuracy metrics guide. Recalculate individual cases in the Brier scale checker, log-loss calculator and RPS calculator.

Expected invariants

A useful implementation should pass relationships, not only isolated cells

  1. 01

    Complete matrix

    The 18 declared probability vectors crossed with H, D and A create exactly 54 rows.

  2. 02

    Invariant 2

    Home and away mirror cases retain equal Brier, exact log loss and normalized RPS after reversing both labels and outcomes.

  3. 03

    Invariant 3

    The ordinal near and far pairs can hold Brier and exact log loss equal while normalized H-D-A RPS changes.

  4. 04

    Invariant 4

    A one-hot forecast that misses has summed Brier 2 and class-averaged Brier 2/3; normalized RPS is 0.5 for an adjacent miss and 1 for an opposite miss.

  5. 05

    Invariant 5

    The uniform forecast has outcome-invariant Brier and exact log loss; normalized RPS is 5/18 for H or A and 1/9 for D under the declared order.

  6. 06

    Invariant 6

    A tied maximum leaves unique top pick and hit blank while every proper score remains defined.

Machine-readable contract

Column dictionary

benchmark_version
Semantic release of this immutable benchmark.
case_id
Stable test-case identifier within the release.
family
Synthetic scenario family; never a match or league label.
pair_id
Optional identifier for a declared comparison or symmetry pair.
p_home
Input home-win probability on the 0–1 scale.
p_draw
Input draw probability on the 0–1 scale.
p_away
Input away-win probability on the 0–1 scale.
outcome
Synthetic observed result: H, D or A.
unique_top_pick
Unique maximum-probability label, blank for a tie.
top_pick_hit
Boolean hit for a unique top pick, blank for a tie.
brier_sum
Sum of the three squared probability errors.
brier_class_average
Summed multiclass Brier divided by three.
rps_normalized_hda
Two-boundary H-D-A RPS divided by two.
log_loss_nats_exact
Negative natural log of the observed-outcome probability; blank at zero.
log_loss_exact_status
Finite or positive_infinity for the exact log score.
log_loss_nats_clipped_1e15
Separate log loss after clipping to [1e-15, 1−1e-15].

Primary sources and football context

The formulas are established; this fixture set is our editorial release

The cited papers define or analyse the scoring concepts and football forecasting context. They did not review, endorse or independently validate this benchmark dataset.

  1. Brier (1950), probability forecast verification.Original paper
  2. Gneiting & Raftery (2007), proper scoring rules.JASA paper
  3. Epstein (1969), Ranked Probability Score.Original paper
  4. Constantinou & Fenton (2012), scoring probabilistic football forecasts.Journal paper
  5. Dubitzky et al. (2018), open soccer database and prediction challenge.Machine Learning paper
  6. Berrar et al. (2024), 2023 Soccer Prediction Challenge framework.Machine Learning paper