Two-window evidence · Runs locally

Audit football prediction performance change before lifetime averages hide it.

Put an earlier reference window beside a later recent window. Keep performance, prediction shape and sample composition separate so one flattering lifetime average cannot erase a change through time.

Canonical publication record

Abstract

A deterministic browser-local monitor that requires model-version provenance, compares fixed reference and recent windows, and withholds like-for-like attribution when version or sample-composition gates fail.

Author and publisher
Football Proof AI
Technical report
football-model-drift/1.0.0
Published
Last modified
Release status
Current release
Review status
Editorial technical note; not externally peer reviewed
Version history
  1. football-model-drift/1.0.0 : Initial public release.
Immutable artifacts
  1. football-model-drift-example-v1.csv sha256:27fadfb53299cbf1a7482133afc57fab47250a930bd3b78c55b6c598fc36eb9e

Interactive · football-model-drift/1.0.0

Fix the windows before reading the change.

Paste a complete settled 1X2 history or choose a CSV. The audit runs locally and makes no network request with your prediction record.

Browser-only · recent minus reference · no automatic retraining

Required: match_id, published_at, kickoff_at, model_version, p_home, p_draw, p_away and outcome. Optional: league. Accepted rows are canonically sorted by kick-off, publication time, then match ID in code-unit order. The latest rows fill the recent window; the declared gap and immediately preceding reference rows remain explicit. Limit: 2 MB and 20,000 rows. The synthetic example is format-only. Download the immutable example CSV.

Comparing the declared windows.Two-window evidence report
Scoring fixed windows…

The report appears after enough rows pass integrity checks and the declared windows fit.

One archive · Two time windows

Ask whether later settled forecasts behave differently from earlier ones

The engine sorts accepted matches chronologically, assigns the latest rows to the recent window, leaves the declared gap out of both estimates, and uses the immediately preceding rows as the reference. Every reported difference is recent minus reference.

match identity
A stable unique key prevents one fixture appearing twice
publication time
Zoned timestamp strictly before kick-off
kick-off time
Orders the accepted record without inspecting the result
model version
Required provenance; a version change can confound a time comparison
1X2 probabilities
Complete home, draw and away distribution on one file-wide scale
outcome
One settled home, draw or away result
league
Optional composition label; missing labels stay explicit
window settings
Reference, gap and recent row counts are part of the receipt
SHA-256
Binds the report to exact local bytes, not to historical publication

Three signals, three meanings

Do not collapse error, prediction shape and sample mix into one badge

A recent loss increase is outcome-based performance evidence. A changed mean predicted 1X2 mix describes model outputs. A changed league or result mix describes the evaluated sample. They can move together without having the same cause.

01 · Performance

Did forecast error change?

Compare class-averaged Brier score and log loss first. Hit rate and ten-bin ECE add context, not a replacement verdict.

02 · Prediction shape

Did the probabilities move?

Total-variation distance compares the mean predicted home, draw and away mix without calling the movement good or bad.

03 · Composition

Did the tested football change?

Outcome and league distributions can make an aggregate score move even when the underlying forecasting process did not degrade.

Directional comparison

Recent minus reference keeps every difference interpretable

Positive Brier or log-loss differences mean the later window has higher average error. A negative hit-rate difference means fewer uniquely ranked top choices were correct. ECE is the sample-weighted mean absolute accuracy–confidence gap across ten fixed confidence bins: it is bin-sensitive, is not a proper scoring rule and receives no universal pass line.

  1. 01

    Full probabilities

    Tied top labels still receive Brier and log loss; they leave only hit-rate and top-label calibration denominators.

  2. 02

    Visible gap

    Gap rows separate the adjacent scored windows and remain counted instead of disappearing.

  3. 03

    Approximate intervals

    The displayed 95% intervals assume independent rows and can be too narrow under team, league or serial dependence.

  4. 04

    No threshold hunt

    The browser does not search for the split, window size or metric that produces the most dramatic result.

Causal boundary

Performance drift is not the same as covariate or concept drift

Covariate drift concerns the distribution of model inputs. Concept drift concerns a changing relationship between inputs and outcomes. This CSV contains probabilities, results and optional labels—not the complete feature stream—so it can expose a performance change but cannot identify either hidden process.

  • Version changes confound time.If the model version differs within or between windows, the conservative verdict is downgraded instead of crediting age, retraining or football conditions.
  • Composition changes confound averages.A league, outcome or mean prediction-mix shift can change aggregate loss without proving model degradation.
  • Short histories amplify noise.A small recent window can move sharply after a few unusual matches. Version 1 uses 30 settled rows per window as a disclosed editorial floor for a directional label, not a universal statistical threshold.
  • Monitoring is not an action policy.The result does not automatically recommend retraining, changing a threshold, stopping a model or placing a wager.

Primary literature

Proper scores and drift terminology set the claim boundary

These publications ground probability scoring, calibration and drift language. They do not certify this implementation, an uploaded archive or the cause of an observed change.

  1. Brier (1950), verification of probability forecasts.Original paper and DOI
  2. Gneiting & Raftery (2007), strictly proper scoring rules.Journal paper and DOI
  3. Gneiting, Balabdaoui & Raftery (2007), calibration and sharpness.JRSS Series B paper and DOI
  4. Gama et al. (2014), concept-drift adaptation survey.ACM Computing Surveys paper and DOI
  5. Franck et al. (2025), monitoring probability calibration over time.Open preprint and DOI

Cite this tool as: Football Proof AI. “Football Prediction Model Drift Monitor.” Version football-model-drift/1.0.0, 13 July 2026, https://footballproofai.com/tools/football-prediction-model-drift-monitor. Editorial technical note; not externally peer reviewed.