Wednesday 16 September 2026 · times UK
These are calibrated — a 65% call lands about 65% of the time. But the model separates high- from low-scoring matches far less sharply than it separates winners, so there is less information behind a goals number than a result one.
| Home | Draw | Away | |
|---|---|---|---|
| Model | 62 | 23 | 15 |
| Market | 69 | 19 | 12 |
| Over 2.5 | Under 2.5 | ||
|---|---|---|---|
| Model | 53 | 47 | |
| Market | 55 | 45 |
| Home | Draw | Away | |
|---|---|---|---|
| Model | 37 | 31 | 32 |
| Market | 37 | 32 | 31 |
| Over 2.5 | Under 2.5 | ||
|---|---|---|---|
| Model | 42 | 58 | |
| Market | 42 | 58 |
| Home | Draw | Away | |
|---|---|---|---|
| Model | 86 | 10 | 4 |
| Market | 91 | 5 | 4 |
| Over 2.5 | Under 2.5 | ||
|---|---|---|---|
| Model | 78 | 22 | |
| Market | 84 | 16 |
| Home | Draw | Away | |
|---|---|---|---|
| Model | 36 | 26 | 37 |
| Market | 27 | 27 | 46 |
| Over 2.5 | Under 2.5 | ||
|---|---|---|---|
| Model | 57 | 43 | |
| Market | 53 | 47 |
5 of 6 settled since yesterday · 199 of 297 since the model went live.
The model rates the whole card, not just the calls, and learns from all of it. Judged as probabilities rather than picks — a 38% favourite losing is the model working, not failing.
A well-calibrated model's average claim matches how often it happens. That is the number Rule 9 gates on, and it is tracked for results and goals alike whether or not a fixture ever produced a call.
What these numbers are. The model's own estimate of how likely each result is, from form, Elo, rest, shots and xG as they stood at 06:00 UK on match day. One bivariate Poisson produces all of them, so the result and the goals figures share one underlying view of the match. Each published figure is then calibrated against what actually happened in past matches, one market at a time — so they agree closely but are not forced to agree exactly. Market figures are the bookmaker's prices with their margin removed, so the two are comparable.
What they are not. Across 57,633 out-of-sample fixtures the market separated matches better than the model in every league, every season, every outcome and every confidence band — 24 slices, no exceptions, re-tested in August 2026 on the current model. This is a second opinion, not an edge. Where it earns its keep is honesty about its own confidence: when it says 74%, that happens about 74% of the time.
Why a call appears. Thresholds — home 70% · away 60% · over 60% · under 55% — sit where the model has historically been as good as its claim. Raising them would surface fewer calls and be less accurate about them, because its confidence above those levels is inflated everywhere except home wins. Draws are never flagged: the highest one ever rated was 62%.
4 rated · 0 not rated · model goals_v7_cal · built 2026-09-16 07:05 UK