Football Proof AIProbabilities with a public recordEvidence library

Free CSV audit · Runs on your device

Football prediction accuracy calculator

Paste or upload a complete 1X2 prediction history. The lab checks every row before calculating hit rate, uncertainty, Brier score, log loss and calibration—without uploading your data.

Interactive · football-accuracy-audit/1.0.0

Audit the rows before trusting the rate.

Use the synthetic example, paste a CSV or choose a file. Every calculation runs locally with a fixed, inspectable contract. The tool makes no network request with your prediction history.

Browser-only processing · no upload

Required header: match_id, published_at, kickoff_at, p_home, p_draw, p_away, outcome. Limit: 2 MB and 20,000 data rows. The built-in rows are synthetic and must never be cited as model performance.

Deterministic audit report
Ready for a complete record

Run the audit to validate the rows and reveal the metric report.

Input contract

Seven fields make each forecast auditable

The calculator refuses to turn malformed or post-kick-off rows into flattering metrics. Extra columns are preserved in the input fingerprint but never change the fixed formula.

match_id
A stable, unique identifier; duplicates are excluded
published_at
ISO 8601 time when the probability became public
kickoff_at
ISO 8601 kick-off known for the audited forecast
p_home
Home-win probability in 0–1 or 0–100 format
p_draw
Draw probability using the same scale as every row
p_away
Away-win probability; all three must total 1 or 100
outcome
Settled result: home, draw or away
Optional slices
league and model_version for context outside the core score

Download the synthetic example CSV, then replace its rows with a complete public record. The example is a format demonstration, not model performance.

Metric convention

One Brier result, two commonly reported scales

Σ3
Sum the three squared class errors. A uniform 1X2 forecast scores 0.6667.
÷3
Divide the same sum by three. The identical uniform forecast scores 0.2222.
×3
Convert class-averaged Brier back to the summed convention before comparing reports.

The lab publishes both values together so two mathematically equivalent results cannot be marketed as if the smaller-looking convention were a better model.

Interpretation order

Read exclusions before reading accuracy

A high percentage is meaningless when difficult matches, duplicates or late publications silently disappear. This report keeps data quality, sample uncertainty and probability quality visible beside the headline.

  1. 01

    Coverage

    Compare total, accepted and excluded rows; inspect every exclusion reason.

  2. 02

    Timing

    Require publication before kick-off. A result known at publication is not a forecast.

  3. 03

    Hit rate

    Count the largest-probability outcome, then keep its Wilson interval beside the point estimate.

  4. 04

    Probability error

    Use Brier and log loss to score the full home-draw-away distribution, not only the top pick.

  5. 05

    Calibration

    Read confidence against observed frequency with the sample size shown for every bucket.

  6. 06

    Fingerprint

    Store the input SHA-256 and formula version so a later report can be tied to the same bytes.

Primary references

Forecast scoring and statistical uncertainty

These sources define the metrics. They do not certify a dataset, prove that rows were historically public or imply future profit.

  1. Brier (1950), probability forecast verification.Original paper and DOI
  2. Gneiting & Raftery (2007), proper scoring rules.Journal paper and DOI
  3. Wilson (1927), score interval for a proportion.Original paper and DOI
  4. scikit-learn, official log-loss definition.Technical documentation