Nobody publishes the score

Every probability we publish is written into a JSON file before kickoff and hashed. Each file carries the SHA-256 of the file before it, so the whole directory is a chain: alter one past prediction without rebuilding every later hash and verification fails — publicly, on our own ledger page, which runs the check on every build. The chain is pushed to a public git repository on every run, so hiding a change would mean rewriting every later file and force-pushing, in public, where anyone who cloned it earlier can see.

A git history alone does not prove when — a commit date is a setting. So each day's entry is also anchored with OpenTimestamps, which commits its hash into the Bitcoin blockchain: proving a file existed before a block is not a claim about our honesty, it is a property of the chain with the most expensive clock in the world. Anchoring has run since 28 August 2026; the first two days of the ledger predate it and are marked, on the ledger page, as chain-only. We would rather say that ourselves than have somebody else say it for us.

You can recompute all of it in under a minute:

git clone https://github.com/GitSimaao/proofodds && cd proofodds
python -m proofodds.verify

That needs nothing installed. The verifier imports only the standard library, and rewrites the hash rule rather than calling the code that wrote it — our code checking our code could agree with itself while both were wrong.

The part most sites leave out

Our model loses to the closing line. A walk-forward backtest over seven Premier League seasons puts it roughly 0.019 nats per match behind the market-average close, and the live record since 26 August 2026 currently reads:

That number is on the front page and it stays there whichever way it points. We are not selling a model that beats the market. We are publishing a measurement, and the measurement includes us.

Why the closing line. Not our own opinion of a good result, not units won, not return on investment. The market-average closing price is the number left after every informed opinion has moved it, right up to kickoff, and log loss cannot be cherry-picked the way ROI over three hundred bets can. Two reference points frame every figure: predicting 1/3-1/3-1/3 every week scores 1.0986, and the closing line scores about 0.9639. Everything anyone knows about a football result is worth the gap in between. Our job is to say how much of that we capture, and to be exact about it.

What is covered. Eight divisions — Premier League, Championship, La Liga, Serie A, Bundesliga, Ligue 1, Primeira Liga, Eredivisie — with one Dixon-Coles goals model fitted per division on that division's matches only. Two markets per match: the result, and whether it goes over 2.5 goals, both out of the same fitted model and both graded against the market-average close. The model itself, the backtest behind it, and the reason the benchmark is the average rather than a single book are on the method page, measurement included.

What is not covered, and why. No both-teams-to-score, no over 1.5, no most-likely-scoreline as a graded market, no Champions League. Not because they are hard — most fall out of the same scoreline grid for free — but because no closing line is published for any of them. We would be stating numbers we could never be shown to be wrong about, and a page where every figure is scored is worth more than a page with more figures on it.

A warning about comparing the two markets. A 1X2 closing line holds about 0.135 nats of knowledge over guessing. A closing total holds about 0.020 — there is roughly seven times less to know about goals, and a closing total barely beats a coin flip. Any model therefore sits closer to the market on goals almost regardless of how good it is. Reading a smaller gap there as "better at goals" is exactly backwards, so the scorecard reports the share of what was available in each market rather than inviting the comparison.

No affiliate links. No ROI headline. No "since we improved the model" start date. Every graded match since the first publication is in the numbers, and the ledger is there to prove it.

Statistical forecasts, published for information. Not betting advice. 18+.