Free paired benchmark · Runs locally

Test whether your 1X2 model beats the bookmaker market.

Compare one complete model forecast with the same match's de-vigged market probabilities. Keep proportional and Shin baselines side by side instead of choosing the friendlier result.

Canonical publication record

Abstract

A deterministic browser-local benchmark of complete football 1X2 model probabilities against proportional and Shin de-vigged bookmaker odds on the same settled matches, with separate evidence timestamps, paired proper scores, overround diagnostics and explicit information-set limits.

Author and publisher
Football Proof AI
Technical report
football-market-benchmark/1.0.0
Published
Last modified
Release status
Current release
Review status
Editorial technical note; not externally peer reviewed
Version history
  1. football-market-benchmark/1.0.0 : Initial public release.
Immutable artifacts
  1. football-market-benchmark-example-v1.csv sha256:2f9840ba2ec65487aae74e5f0c0e5b325da05ea847204729c17f19f84d757bef

Interactive · football-market-benchmark/1.0.0

One model. Two market baselines. The same settled matches.

Remove the supplied 1X2 overround proportionally and with Shin, then compare both baselines with the exact same model rows. The lab makes no network request with your probabilities or odds.

Browser-local calculation · forecast quality, not profit

Required: match_id, model_published_at, odds_captured_at, kickoff_at, model_home, model_draw, model_away, odds_home, odds_draw, odds_away, outcome and odds_stage. Odds stage is pre_match_snapshot or closing. Optional: league and bookmaker. Use one model scale throughout. Limit: 2 MB and 20,000 rows. The synthetic example is a format demonstration, not performance. Download the immutable example CSV.

Model versus market report
Removing margin and scoring paired rows…

The comparison appears after probability, odds, timing and duplicate checks expose every excluded row.

Same-match evidence contract

Pair the forecast, available odds and result before scoring

Every accepted row needs one complete model 1X2 distribution, three finite decimal prices greater than 1, one settled outcome and times proving the forecast and price snapshot were available before kick-off. Invalid or late rows stay visible as exclusions.

match identity
Stable unique key so one fixture cannot be counted twice
publication time
Zoned time when the tested model probability split was fixed
odds time
Zoned time for the exact bookmaker snapshot used as the baseline
kick-off time
Strictly later than both pre-match evidence timestamps
model probability
Complete home, draw and away distribution on one consistent scale
decimal odds
Home, draw and away prices greater than 1 from the same real snapshot
odds stage
Self-declared pre-match snapshot or closing label; the browser cannot authenticate it
source label
Optional bookmaker and league fields preserve provenance but do not verify the provider
outcome
One settled result: home, draw or away
SHA-256
Identifies the exact local bytes; it does not independently timestamp them

Two de-vig assumptions

Raw inverse odds are not a probability forecast until the margin is handled

Decimal odds imply raw weights of 1 ÷ odds. Their sum usually exceeds one. Proportional normalization divides every weight by that sum. Shin instead fits z under an insider-trading model and redistributes margin non-proportionally. The fitted parameter is not an observation of insiders. Neither method is ground truth, so the report keeps both.

Proportional baselinepᵢ = (1 ÷ oddsᵢ) ÷ Σ(1 ÷ odds)

Transparent and deterministic; assumes the margin scales every outcome equally.

Shin baselinesolve z, then normalize pᵢ(z)

Allows favourite–longshot asymmetry; report fitted z as a model parameter—not observed insiders—and any solver exclusion.

Forecast-quality claim

Score the full probability split, not just who finished top

The model and both market baselines face the same settled matches. Class-averaged Brier score measures squared error over all three outcomes; log loss strongly penalizes tiny probability on what happened. Lower is better for both. Publish paired differences, coverage and exclusions—not only the winning cell. For finite deterministic exports, the implementation floors an actual-outcome probability of zero at 1e-15 before taking the logarithm; that epsilon is included in every JSON and CSV report. Reproduce that boundary on a single 1X2 forecast in the football log-loss lab.

  1. 01

    Same rows

    No fixture-mix advantage: every score is calculated on the same accepted matches.

  2. 02

    Two baselines

    A superiority claim must survive both proportional and Shin margin removal.

  3. 03

    Paired losses

    Inspect per-match score differences and their uncertainty, not unrelated averages.

  4. 04

    Held-out period

    Freeze the model and protocol before evaluating a later, untouched match window.

Claim boundary

Better probabilities do not automatically create a profitable strategy

This lab evaluates forecasts against a supplied market snapshot. It does not know whether those prices were executable, which limits applied, how the market moved, or what selection and transaction costs a strategy would face.

  • No odds shopping.Combining the best home, draw and away prices from different times or books creates a baseline that never existed.
  • No closing-line shortcut.Odds observed after the model publication can contain information the model never had; timestamps must be reported.
  • No selective coverage.Omitted leagues, matches or failed forecasts remain part of the completeness question.
  • No profit claim.Forecast quality, expected value, executable price and realized return are separate claims.

Primary literature

The benchmark is grounded in scoring rules and declared margin removal

These papers ground the scoring, odds-setting, market-comparison and Shin assumptions. They do not certify this implementation, any uploaded odds, or future model performance.

  1. Brier (1950), verification of probability forecasts.Original paper and DOI
  2. Shin (1992), insider trading and favourite–longshot bias.Economic Journal paper and DOI
  3. Gneiting & Raftery (2007), strictly proper scoring rules.Journal paper and DOI
  4. Franck, Verbeek & Nüesch (2010), bookmaker and betting-exchange prediction accuracy.International Journal of Forecasting paper and DOI
  5. Koning & Zijm (2022), bookmaker odds setting and prediction.Annals of Operations Research paper and DOI

Cite this tool as: Football Proof AI. “Football Prediction Model vs Bookmaker Odds Calculator.” Version football-market-benchmark/1.0.0, 13 July 2026, https://footballproofai.com/tools/football-model-vs-bookmaker-calculator. Editorial technical note; not externally peer reviewed.