What this system does, and where it breaks.
A model card is the document a serious team publishes so users can judge whether to trust a system. Ours includes the failures — that is the point of it.
At a glance
- System
- Forecast Foundry
- Task
- 1X2 + goal markets
- Coverage
- 17 leagues
- Forecasts
- 7,913
- Settled
- 7,861
- Horizon
- published pre-lineup
What it predicts
The probability of a home win, a draw and an away win for 17 football leagues, plus derived goal markets — over/under 2.5 goals, both teams to score, and a full scoreline distribution. Outputs are probabilities, never picks.
What it is fed
- Historical results, goals, shots, shots on target and corners (99,567 matches).
- Bookmaker prices, opening and closing, de-vigged with the power method.
- Confirmed lineups where a feed is available, converted into availability adjustments.
- Availability claims extracted from public match reports and press coverage, each with a source URL, a confidence and a method (rules or llm).
It is not fed expected goals, player tracking, travel distance, referee tendency or weather.
How it performs
Walk-forward, out of sample, 7,861 settled forecasts. RPS is the proper scoring rule for three ordered outcomes; lower is better.
| Measure | This model | Reference |
|---|---|---|
| RPS | 0.202 | 0.202 (market) |
| Accuracy | 50.9% | 43.5% (always home) |
| Calibration error | 0.009 | 0.000 (perfect) |
| Mean closing-line value | -0.37% (n=305) | 0.00% (no edge) |
The model is a shade worse than the market on RPS and indistinguishable on closing-line value. That is the honest reading: it produces a well-calibrated distribution, not an edge.
Known failure modes
- It cannot beat the closing line. The benchmark comparison is deliberately unfair to the model and the model loses it.
- 85% of each probability is the market. The blend weight is learned per league and lands at 0.80–1.00. Where the model disagrees with the price, the disagreement is usually wrong.
- Promoted and unseen clubs get league-average priors. A newly promoted side is treated as an average team until it has enough matches.
- Early-season forecasts are the weakest. Fewer than five matches of form is mostly noise, and the model does not widen its uncertainty enough to say so.
- A transfer can still look like an injury. Where the injury feed covers a match, only a listed player counts as absent and unlisted ones are ignored. Where it does not, a missing regular is downgraded to doubt at half weight rather than claimed as injured — contained, not eliminated.
- Extracted claims do not move forecasts yet. Text extraction is only trusted once it has been scored against confirmed lineups: until measured precision clears 85% over at least 30 scored claims, claims are collected and audited but never applied.
- Injury coverage is partial. The feed is queried per day and only covers fixtures we track. Matches outside it fall back to the conservative squad-absence reading.
- Lineups arrive after we publish. The headline forecast is made before confirmed XIs; a late surprise is not priced until a second forecast is written.
- Small leagues are less reliable. Fewer matches means wider parameter uncertainty and a larger gap to the market.
Intended use
Understanding how likely each result is, comparing that with a price, and holding the system to account against a public record. Research, journalism, and personal analysis.
Not intended for
Guaranteed returns, stake sizing advice, or any decision where a wrong probability causes harm. It is not betting advice and it is not a tipster service.
Reproducibility
Every forecast is written to an append-only, hash-chained ledger with its model version, the data cut-off, the parameters hash and the price at publication. The daily Merkle roots are public. 7,913 rows, chain intact.
See the record →