Every prediction is written down before kickoff and graded afterwards. The first prediction stands — nothing is revised once a result is known. The figures on this page count only calls committed before kickoff (the full-round record runs from 2026-09-10); the opening rounds, rebuilt after the fact, are reported apart.
Scored properly
A hit rate rewards calling the favourite and says nothing about whether 46% meant 46%. Log loss and Brier score punish a confident miss more than a hesitant one, which is what a published probability has to answer for. Lower is better. Two baselines say what the numbers mean: a third each is knowing nothing; the base rate is knowing only how often home sides win this season (47% home, 22.3% draw, 30.7% away, over these same matches, which flatters the baseline).
| Forecast | Log loss | Brier |
|---|---|---|
| This model | 0.9721 | 0.5759 |
| Base rate | 1.0517 | 0.6349 |
| A third each | 1.0986 | 0.6667 |
On the frozen seasons the same model scores 0.984 log loss over 11,311 fixtures against a base rate of 1.073. How it compares with the bookmakers' closing prices is on Versus the market.
Is it honest?
Accuracy can be gamed by only predicting easy matches. Calibration cannot: to be calibrated the model has to be right about its own uncertainty, so its 35% calls must fail about 65% of the time. Each row groups predictions by how confident the model was, and compares what it claimed against what happened.
- 30-40%116 callssaid35.3%landed31.9%
- 40-50%103 callssaid44.8%landed51.5%
- 50-60%74 callssaid54.4%landed64.9%
- 60-70%48 callssaid64.1%landed70.8%
- 70-80%14 callssaid74%landed100%
Grey is what the model claimed, blue is what happened. Bars that match are the point. Landing above the claim is not a virtue: it means the probability was too low, and every fair price built on it was too high. Buckets with few calls will move a lot; treat small samples as noise.
Where the probability sits
The buckets above group calls, so a bias in one outcome's level can hide inside them. This table takes every live fixture, not only the calls, and compares the mean probability the model gave each outcome with how often that outcome happened. A gap beyond two standard errors is a bias, not noise.
| Outcome | It said | Happened | Gap | ± 1 s.e. |
|---|---|---|---|---|
| home | 41.5% | 47% | -5.6 | 2.6 |
| draw | 27% | 22.3% | +4.7 | 2.2 |
| away | 31.6% | 30.7% | +0.9 | 2.4 |
Where a draw gap comes from: the model has priced these matches at 2.73 goals a game and 2.99 have been scored. A scoreline grid turns every goal it fails to expect straight into draw probability. On the eight frozen seasons the same model says 25.0% draw against 25.4% drawn, and a draw correction fitted season by season comes out between 1.00 and 1.02 — so no correction is applied, and the gap is reported here instead.
The reconstructed rounds
The full-round record started on 2026-09-10. The 146 fixtures played before that (2026-08-15 to 2026-09-07) were rebuilt on that date, walk-forward: each priced only from data available before its kickoff, by the model as it stood then. The method is the same, but the rows were written after the results were known, and a reader cannot verify the difference — so they are kept out of every headline figure and shown here on their own. The weekly picks ledger, a separate record, was committed before kickoff from the first round.
Both sets together: 501 fixtures, 53.3% landed, 0.9776 log loss.
By competition
Open a competition for every prediction it has made, graded match by match.