Models
One statistical model sets every published number. Two AI analysts, Claude and Codex, forecast matches in selected leagues in shadow: their views are locked and graded the same way, and they get a weight only if they prove better on enough matches. No betting-market data is used.
0 of 4 conditions met. All four must hold before the AI layer gets a weight (at most 50%), with a new model version and a changelog entry.
Full comparison in the track record| Version | In use | Published model | AI weight | Notes |
|---|---|---|---|---|
v1.3-2026-11 | 2 Nov 2026 – now | Statistical model: Elo + Dixon-Coles, log-pooled and calibrated per league; European club competitions rated across leagues | 0% | Adds the Champions League, Europa League and Conference League. A club's European rating starts from its domestic rating and adds a league-strength offset and a club offset, both learned from European results since 2016/17, qualifying rounds included. Clubs whose domestic league is not in our data are rated from European matches only. Goals use the club's domestic attack and defence plus league and club offsets fitted on European matches. Settings were chosen on 2019/20 and 2021/22 (matches played without fans left out) and tested on 2022/23 to 2025/26: Champions League RPS 0.191 over 628 matches, Europa League 0.198 over 660, Conference League 0.208 over 588. In the Champions League test the model was too cautious about favourites. Domestic league forecasts are unchanged from v1.2-2026-11. |
v1.2-2026-11 | 2 Nov 2026 – 2 Nov 2026 | Statistical model: Elo + Dixon-Coles, log-pooled and calibrated per league | 0% | Calibration per league: each league gets its own temperature, chosen on its validation seasons (2020/21–2021/22; Scotland 2020/21; Saudi Pro League 2023/24) and pulled towards the shared value when the league has little history. Serie A keeps the shared value because its own failed our calibration check. Clubs moving between the Premier League and the Championship now keep their rating across the two leagues (Elo gap 400 points, Dixon-Coles gap 0.1) instead of starting as newly promoted sides. One erroneous result removed from the Saudi history. Other settings unchanged from v1.1-2026-10. |
v1.1-2026-10 | 31 Oct 2026 – 2 Nov 2026 | Statistical model: Elo + Dixon-Coles, log-pooled and calibrated | 0% | First version fitted on real results: open historical data (CC0) from 2016/17 plus live-season results after the open data's last update. Settings chosen on the 2020/21–2021/22 validation seasons, reported on 2022/23–2025/26. Claude and Codex not run yet (no AI layer). |
v1.0-2026-10 | Not published | Statistical model: Elo + Dixon-Coles, log-pooled and calibrated | 0% | Development prototype, run only on sample data while the site was being built. It was never published, and none of its forecasts are in the ledger or the track record. |
Every locked forecast records its model version in the ledger.
Claude is an AI model made by Anthropic and Codex by OpenAI. We use them as tools to read match dossiers; neither company reviews, sponsors or endorses these forecasts.
For information and entertainment only. Not betting advice. 18+. Predictions are probabilities, not certainties. Responsible gambling