Why We Publish Predictions
Every day we publish a probability for every match on our board. Every day after, we score what we published against what actually happened, and the result goes on a public page whether it flatters us or not.
That loop is the whole project. Everything else here is scaffolding around it.
Nobody asked us to build this. We built it because we wanted somewhere a football match could be looked at the way it deserves to be looked at: carefully, with the numbers laid out in the open, and without anyone trying to sell you something on the way out.
What we publish, and when
Before kickoff, every match on our board gets a full probability split — home win, draw, away win — along with the most likely scoreline and the fair odds those probabilities imply. Fair odds are simply the probability expressed differently: 100 ÷ the probability. They are not a bookmaker's price, and we do not compare them to one.
The next day we go back and check. Not selectively, not only when it went well. Every prediction is written to a log before the ball is kicked and settled against the real result afterwards. It goes into the record whether it looked clever or foolish, and the log is never edited.
Every match has its own page
We did not want a site that gives you a percentage and nothing else. Each fixture has a dedicated page: the probability split, the expected goals behind it, both sides' form going into the match, head-to-head history, and every scoreline the model considered, with the likely ones picked out.
Some of the week's bigger fixtures also carry a written breakdown; the rest carry the model's one-line call. We want to be clear about what those are. They are not punditry, and not a journalist's opinion dressed up in statistics. They are the model explaining, in plain language, the numbers it arrived at. If it says two teams cannot be separated, that is because the probabilities genuinely sit on top of each other.
Live scores, and live probabilities
Football matches are not static, and neither are our estimates. We publish live results, and alongside them the probabilities as they move through the match. A red card, an early goal, a game that opens up in the last twenty minutes: the numbers respond as it happens, rather than leaving you with the version we published hours before kickoff.
It is one of the more instructive things on the site. Watching a number swing and swing back tells you a great deal about how much of football is genuinely uncertain — which is, in the end, the thing we are trying to measure.
We publish where we are wrong
This is the part we care most about, and the part most prediction sites leave out. So rather than tell you the page exists, here is what it currently says.
Across 1,661 settled predictions we have called the right result 47% of the time and the exact scoreline 9% of the time. Our Brier score is 0.640, against 0.667 for a model that learned nothing and gave every match an even split. We are ahead of knowing nothing. That is a lower bar than it sounds, and we would rather say so than let the number do work it has not earned.
| Probability | |
|---|---|
| We said | 74.6% |
| It happened | 60.8% |
When we put a team at 65% or better, it wins about 61% of the time. We are too sure of ourselves at the top end, and the same error runs the other way at the bottom: outcomes we price at 10–20% come in closer to 23%. We have known about this, tried a fix, measured it again, and found the fix barely moved it. That is where it stands.
By competition it is uneven, sometimes badly. Our worst is the French second division, where we have called 25% of 36 matches correctly — worse than picking at random from three outcomes. We would rather print that than average it away.
One thing has gone the other way. We wrote on our methodology page that the model backed away favourites too hard, and that we would not claim it was fixed until we had re-measured it on the current parameters. We have now re-measured it, on 342 predictions: the gap has closed from about fourteen percentage points to under three. It is fixed. We are recording that in the same place we recorded the fault, because a page that quietly deletes its weaknesses as they go away is a page you cannot trust about the ones that remain.
We also keep league-by-league accuracy standings, so you can see where the model is on firm ground and where it is guessing more than we would like to admit. Some competitions have deep, reliable data behind them. Others do not, and the results reflect that.
The methodology page sets out the whole approach, including its known limitations — the ones we have closed and the ones still open, such as the goal totals our model currently runs a little high. We treat those limitations as part of the documentation, not as something to bury.
The rest of the site
- Fixtures & Predictions — the full board, several days ahead and behind, across every competition we cover.
- Results — final scores next to the call we published before kickoff, so the comparison is always one click away.
- Tables — league standings, kept current.
- Title races and relegation odds — where every league we project is heading, and what moved since the last round.
- Best XI — the standout performers, selected on the numbers.
- Analysis — the model's read on the biggest matches of the coming week.
- Track record Accuracy by club Accuracy by competition — how we have actually done, updated automatically after every match.
We currently cover 50 competitions and over 800 clubs, and we add to that when the underlying data is good enough to justify it. We would rather cover fewer leagues well than many badly — and the accuracy standings above are there so you can judge whether we are living up to that.
No bookmakers, no affiliates
Everything here is published for informational and entertainment purposes. Our estimates are a statistical model's output, not betting advice, and they guarantee no outcome. We carry no gambling-affiliate advertising and no operator sponsorship, and we do not intend to. It is the only way we know to keep the numbers pointing at what is true rather than at a conversion rate.
Tell us what to build next
The model is not finished, and it never will be. We are constantly refining it, testing new inputs, and adding things we think will genuinely be useful rather than merely impressive.
That is where you come in. If something is unclear, if a number looks wrong, if there is a competition you want covered or a statistic you keep wishing were there, write to us at [email protected]. We read everything, and a fair amount of what is on this site started as somebody's suggestion.
Thanks for reading. Enjoy the football.
Figures on this page are recomputed from our own prediction log — 1,661 settled pre-match predictions between 19 July and 2 September 2026, across 44 competitions in 36 countries. They move as the sample grows; the track record page always carries the current ones.