Data-driven Premier League predictions, with the reasons. New every Thursday.
A percentage is only worth something if it comes true about as often as it says. So we check every one we've given: the home win, the draw and the away win in every game. If our 60% calls come off about 6 times in 10, you can take our numbers at face value.
Based on 440 games from 2025/26 and 2026/27, replayed week by week using only results known before each game. The model's settings were tuned on 2025/26, so this season's games are the fairer test. Live calls join these as the season goes on.
| We said | Times | Average | Happened |
|---|---|---|---|
| 0% to 10% | 14 | 8% | 0% |
| 10% to 20% | 117 | 16% | 18% |
| 20% to 30% | 556 | 26% | 26% |
| 30% to 40% | 289 | 34% | 37% |
| 40% to 50% | 173 | 45% | 40% |
| 50% to 60% | 110 | 54% | 56% |
| 60% to 70% | 45 | 64% | 53% |
| 70% or more | 16 | 75% | 88% |
43%
Our pick won 79 of 183 games. We'd given it 40% on average.
47%
Our pick won 69 of 147 games. We'd given it 49% on average.
64%
Our pick won 70 of 110 games. We'd given it 63% on average.
Over the same 440 games our pick was right 50% of the time, and the bookmakers' favourite 49%. The score beside this grades the percentages themselves, not just the picks: it is the ranked probability score, where lower is better and the gaps that matter are in the third decimal place.
The bookmakers' numbers are their average odds as collected by football-data.co.uk a day or two before each round, with their margin removed by the power method. We publish on Thursday, so they often have more team news than we do. Odds are a benchmark, not betting advice.
Our ratings lean on the quality of chances rather than the goals that went in. Looking back over past seasons, that seemed to help most when one side's results had run well ahead of, or behind, its chances. It wasn't enough to prove anything, so we've written the test down in advance and will score it on our live calls only.
The rule: take each side's goals minus expected goals, scored and conceded, per game over its last 20 league games. If the two sides differ by 0.45 goals a game or more, the match is in the test. Every published call records this number before kick-off. At the end of the season we compare our percentages in those games with the bookmakers'. The full rule is on GitHub, dated by its commit.
No live games have been scored yet. The first ones count after the next matchweek is played.
Anyone can claim they saw a result coming. Each week's calls are saved as a file, and every file has a fingerprint: a code worked out from its exact contents, so changing a single character gives a completely different one. We post the fingerprint in our Thursday thread on X, and the Internet Archive saves a copy of the file and of this site, both dated by them, not us.
To check a week: download its file from GitHub, run sha256sum on it (or any SHA-256 tool) and compare the answer with the fingerprint below and in our post.