FPAIFootball Proof AIProbabilities with a public recordEvidence library

Free CSV audit · Runs locally

Audit football prediction confidence against real outcomes

Rank the same settled 1X2 history by maximum probability, top-two margin or entropy. See how much coverage disappears—and whether the remaining calls actually make fewer errors.

Canonical publication record

Abstract

A deterministic browser audit that ranks a supplied settled history of complete 1X2 rows by maximum probability, top-two margin and normalized-entropy certainty, then exposes tie-preserving risk-coverage curves without inventing a safe-pick threshold.

Author and publisher
Football Proof AI
Technical report
football-confidence-coverage/1.0.0
Published
Last modified
Release status
Current release
Review status
Editorial technical note; not externally peer reviewed
Version history
  1. football-confidence-coverage/1.0.0 : Initial public release.
Immutable artifacts
  1. football-prediction-confidence-example-v1.csv sha256:60dd1d60d256ea537e397707335b1a47b770a7d5c04b5d23f6e66c1119c738e7
Machine-readable manifestCSL-JSONBibTeX

Interactive · football-confidence-coverage/1.0.0

Retain less. Show the full trade-off.

Rank one settled 1X2 history three ways, then inspect how retained coverage, errors and probability scores move together. Derived signals are rounded to 12 decimal places before grouping, and equal values stay together.

Browser-only analysis · no prediction rows are uploaded

Required: match_id, published_at, kickoff_at, p_home, p_draw, p_away and outcome. Use one consistent 0–1 or 0–100 scale. Limit: 2 MB and 20,000 data rows. The lab makes no network request with your CSV. The built-in sample is synthetic and is not model performance.

Confidence and coverage report
Ranking accepted rows…

Results appear only after the timing, probability and outcome integrity checks have exposed every excluded row.

Complete-record gate

Coverage starts after malformed, duplicate and late rows are exposed

The audit reuses the same seven-field contract as the complete-record Accuracy Audit. Only valid, genuinely pre-kick-off rows with one unique maximum probability enter the selective-risk curve. Every exclusion remains visible and downloadable.

match_id
Unique stable identity; duplicates are excluded
published_at
Zoned ISO 8601 publication time
kickoff_at
Must be strictly later than publication
p_home / p_draw / p_away
One complete 0–1 or 0–100 probability scale
outcome
H, D, A, home, draw or away
unique top pick
Tied maximum probabilities are excluded from hit-rate risk, not silently broken by outcome order
SHA-256
Identifies the exact local bytes; it does not prove publication time

Download the synthetic confidence example CSV. It demonstrates the contract and curve, not model performance.

Three ranking signals

A large probability is not the only definition of confidence

The tool ranks the same accepted rows three ways and compares them at matched target coverage. Higher values always mean the row is retained earlier. Derived signals are rounded to 12 decimal places before grouping, and tied values are never split to make a curve look smoother.

  1. 01

    Maximum probability

    The largest of home, draw and away probability: max(pH, pD, pA).

  2. 02

    Top-two margin

    The gap between the largest and second-largest outcome probabilities.

  3. 03

    Entropy certainty

    One minus Shannon entropy normalized by ln(3); concentrated distributions rank higher.

  4. 04

    Same target coverage

    Ten fixed coverage targets reveal whether one ranking keeps errors out more consistently.

Risk–coverage contract

Fewer mistakes at lower coverage is a trade-off, not free accuracy

Coverage is selected rows divided by every accepted row. Risk is the top-pick error rate inside that selected set. Silent-failure rate is selected errors divided by the full accepted history, so withholding most matches cannot disappear from the evidence.

Selective risk and coveragerisk = selected errors ÷ selected rows · coverage = selected rows ÷ all accepted rows

Step AURC integrates the displayed, tie-preserving risk over increasing coverage. It is a descriptive sample summary—not a universal pass threshold or statistical guarantee.

Validation boundary

A threshold selected after seeing these outcomes is not out-of-sample evidence

Exploring the curve can generate a future policy. It cannot validate that policy on the same outcomes. Freeze the signal and threshold, then test them on a later untouched period with the walk-forward protocol.

  • No automatic threshold.The lab does not choose the lowest observed error after seeing the results.
  • No calibration shortcut.A ranking can separate easy from hard matches while its probabilities remain miscalibrated.
  • No independence guarantee.Teams, leagues and adjacent fixtures can make Wilson intervals too optimistic.
  • No betting conclusion.Selective accuracy does not establish bookmaker value, profit or future performance.

Primary literature

Confidence needs a visible reject trade-off and a held-out test

These sources ground the ranking and uncertainty definitions. They do not certify this implementation or any uploaded history.

  1. El-Yaniv & Wiener (2010), risk–coverage foundations.JMLR paper
  2. Geifman & El-Yaniv (2017), selective classification.NeurIPS paper
  3. Shannon (1948), information entropy.Journal article DOI
  4. Wilson (1927), binomial interval.Journal article DOI
  5. Gneiting & Raftery (2007), proper scoring rules.Journal article DOI

Cite this tool as: Football Proof AI. “Football Prediction Confidence & Coverage Audit.” Version football-confidence-coverage/1.0.0, 13 July 2026, https://footballproofai.com/tools/football-prediction-confidence-calculator. Editorial technical note; not externally peer reviewed.