Real model. Real backtest. Activation still fails closed.
LightGBM football prediction model backtest
2,075 evaluations · 49.30% hit · 0.2054 Brier · 3.41% ECE
The numerical gate passed. Candidate activation did not: pointInTimeProvenance is required. Latest-state CSVs have no correction observed-at revision history.
Published by Football Proof AI · Published and updated · lightgbm-football-prediction-model-backtest/1.0.0 · Editorial technical note; not externally peer reviewed
Canonical publication record
Abstract
An aggregate-only, reproducible EPL research backtest of the fixed fourteen-feature LightGBM plus one-vs-rest isotonic candidate across 2,075 chronological evaluations. The numerical gate passed, but production activation failed closed because latest-state source files do not provide correction observed-at revision history.
- Author and publisher
- Football Proof AI
- Publication version
lightgbm-football-prediction-model-backtest/1.0.0- Published
- Last modified
- Release status
- Current release
- Review status
- Editorial technical note; not externally peer reviewed
- Production method
- Editorial dataset maintained from the declared method, linked sources and any listed reproducible artifacts.
- Software and AI assistance
- Software, including generative AI, may assist drafting, transformation or quality checks. It is not treated as a source, author or independent reviewer; factual and quantitative claims must remain bound to cited sources, declared methods or reproducible artifacts.
- Why this exists
- Enable readers to inspect assumptions, reproduce calculations and reject claims that exceed the published evidence.
- Commercial boundary
- No bookmaker, prediction provider or paid placement controls the evidence rules, calculations or conclusions on this page.
- Version history
lightgbm-football-prediction-model-backtest/1.0.0: Initial public release.
- Immutable artifacts
- 1.0.0.json
sha256:1918b88313dd13367e33c961828748d00c549bb676745ca65e34c03263a49e4c - lightgbm-football-prediction-model-backtest-1.0.0.csv
sha256:f133ea4bf213ef9419849891d5eee377e30576c8cc848d47857da792a5d133d9 - 1.0.0-manifest.json
sha256:87e00eb63c6400b5af6ae1871fde27e81d095c22c619228852d849e6fc042ff0 - generate-lightgbm-football-prediction-model-backtest-1.0.0.py
sha256:97452e6951374124c34fd3163d505b5889b69db18c99d7154d81c9e98310d27d
- 1.0.0.json
- Release resources
- References
- https://proceedings.neurips.cc/paper_files/paper/2017/hash/6449f44a102fde848669bdd9eb6b76fa-Abstract.html
- https://icml.cc/Conferences/2005/proceedings/papers/079_GoodProbabilities_NiculescuMizilCaruana.pdf
- https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.TimeSeriesSplit.html
- https://doi.org/10.1175/1520-0493(1950)078%3C0001:VOFEIT%3E2.0.CO;2
The scored result
How accurate was this Premier League AI prediction model?
Direct answer: Across 2,075 Premier League walk-forward evaluations, the historical LightGBM research backtest recorded a 49.30% top-pick hit rate and a 0.2054 class-average Brier score. These are aggregate backtest results, not live prediction accuracy or evidence of betting profit.
The existing 14-feature pipeline trained LightGBM multiclass models, fitted one-vs-rest isotonic calibrators on separate chronological slices and scored three later walk-forward folds. The evaluation runs from 2021-02-02 to 2026-05-24.
2,075 matches
Three time-ordered folds from 3,800 EPL source fixtures.
49.30%
Direction-only accuracy; it does not account for quoted odds or profit.
0.2054
Mean squared probability error across home, draw and away, divided by three.
3.41% ECE · 1.0797 log loss
ECE is top-label, ten-bin expected calibration error; log loss penalizes weak probability assigned to the observed result.
The activation boundary
A metric pass is not permission to publish predictions
Frozen source hashes establish which publisher files produced this aggregate. They do not reconstruct every earlier correction and the time it became observable. Without that history, a historical row cannot carry complete point-in-time provenance.
pointInTimeProvenance is required
The production model gate was not modified, bypassed or activated.
Reproducible protocol
Fourteen locked features, separate calibration, future-only scoring
Form, venue splits, Elo, head-to-head and rest-day features are generated by the same proofxi_ml implementation used by the candidate pipeline. A 24-hour publication lead and 24-hour result-availability lag prevent a result from entering a feature snapshot at kick-off.
- 01
Hash-pinned inputs
Ten EPL seasons, 2016/17–2025/26. Every publisher response must match the SHA-256 inherited from the empirical benchmark manifest.
- 02
Three walk-forward folds
Raw training, isotonic calibration and evaluation slices are chronological and separated by availability-safe purges.
- 03
Aggregate-only release
No raw rows, fixture-level probabilities, fitted LightGBM model or calibrator are redistributed.
Exact feature names and configuration are included in the JSON. For the full boundary design, read the walk-forward validation protocol and leakage auditor.
Machine-readable evidence
Download the aggregate, manifest and generator
Each immutable response exposes a content SHA-256 and supports conditional requests. The manifest binds the generator and all ten source fingerprints.
Aggregate JSON
1918b88313dd13367e33c961828748d00c549bb676745ca65e34c03263a49e4c
Aggregate CSV
f133ea4bf213ef9419849891d5eee377e30576c8cc848d47857da792a5d133d9
Release manifest
10 source hashes and three public artifact hashes.
Generator source
97452e6951374124c34fd3163d505b5889b69db18c99d7154d81c9e98310d27d
Read before reusing
What this release does not claim
- BoundaryThis is an aggregate research backtest for one league and ten seasons, not live forecast accuracy.
- BoundaryLatest-state CSVs lack correction observed-at revision history, so point-in-time provenance is incomplete and activation fails closed.
- BoundaryNo raw source rows, fixture-level probabilities, fitted model or calibrator are redistributed.
- BoundaryHit rate, Brier score, log loss and ECE do not establish betting profit, value or a universal edge.
Source files remain at Football-Data.co.uk. The aggregate release does not assert a licence to redistribute their raw rows.
Questions answered directly
LightGBM football model backtest FAQ
Is this Football Proof AI's live prediction accuracy?
No. It is an aggregate historical research backtest on hash-pinned latest-state EPL CSV files. It is not a live prediction ledger and no production model was activated.
Why did activation fail if the numerical thresholds passed?
The source files do not provide a versioned correction observed-at history. Their frozen bytes prove which latest-state files were used, but not when every historical correction first became knowable. The production candidate gate therefore fails closed on point-in-time provenance.
Does a 49.30% 1X2 hit rate prove betting profit?
No. Hit rate ignores odds and price availability. Brier score, log loss and ECE describe probability performance, but none of these figures proves expected value, obtainable prices or profit.