How it works
Every match gets a probability before kick-off, from a statistical model trained on 778,454 finished league matches. Every prediction is graded in public afterwards — right or wrong, and kept for good.
The forecasting model
Two statistical models rate every team’s attack and defence, and we publish an even blend of the two.
The first learns from each league’s results, with recent matches counting for more. Where a match’s statistics are recorded, it also learns from expected goals — the quality of the chances a team created, not only the goals it scored. It corrects for low scores such as 0–0 and 1–1 happening more often than chance suggests.
The second treats every rating as a belief that updates after each match, and grows less certain the longer a team goes without playing — so a side back from the summer break is judged more cautiously than one that played on Saturday. Newly promoted teams start below the league average.
One calculation behind every number
Together the models produce one thing: the probability of every possible scoreline. Home, draw and away, total goals, both teams scoring — every figure on the site is read from that single table, so no two of our numbers can ever disagree.
What it learns from
778,454 finished league matches across 786 leagues, for most of them the last 6 seasons: the results, and the match statistics where they exist. No tips, no opinions and no outside predictions — only what happened on the pitch.
Tested before launch
We replayed the model through every league it covers — 198,229 matches, each predicted using only what was known the day before. Its probabilities scored 9.9% better on the ranked probability score (how close each home, draw and away probability came to what happened) than simply quoting each league’s usual home, draw and away rates.
Cup ties between clubs of different leagues — the Champions League, the Europa and Conference Leagues, the Libertadores, a Premier League club away to a League One side — use a second model that compares the leagues through the matches between them. Replayed over 13,126 ties the same way, it picked the right result 57.5% of the time (58.0% over 602 Champions League matches), against 49.9% for a plain rating of every club.
Graded in public
When a match ends, its prediction is graded and stays on the record — never edited, never removed. We publish two measures: calibration (when we say 60%, does it happen 60% of the time?) and the Brier score, which punishes a confident wrong call hardest.
Where we do not predict
643 leagues have the 300 matches of history the model needs, and 554 competitions carry our predictions now — every one is on the leagues page. Elsewhere a match is shown without a prediction, and says so: a guessed number would look exactly like a real one.
AI at Pitchside Lab
Our match previews and post-match reviews, included with a pass, are written with OpenAI’s models. They explain the model’s numbers in plain words — they never change them, and they are given no news, rumours or web searches to repeat.
The forecasting engine and this site are engineered with Anthropic’s Claude. The predictions themselves come from the statistical model above, and from nothing else.
Questions people ask
Is Pitchside Lab free?
Yes. Every prediction, the record of how they did and its download, and 3 questions a day to our AI desk are free. A pass (7 days, £5.99 plus VAT, 70 questions; or 30 days, £19.90 plus VAT, 300 questions; paid once, never renews) adds the questions and the AI match analyses you ask for, each paid from the pass's allowance.
Is there any advertising?
No. There are no adverts, no pop-up offers and no advertising trackers. We are paid only by readers who buy a pass, so the site works for you and not for an advertiser.
Do you take bets or work with bookmakers?
No. We take no bets and have no bookmaker or affiliate links. We publish probabilities, not tips, and our AI never gives betting advice. If gambling is a problem for you, free help is at BeGambleAware.org.
How accurate are the predictions?
Every prediction is graded after full time on the record page, right or wrong, with its calibration, its Brier score and a comparison with the bookmakers' prices on the same matches. We publish that comparison whether we are ahead or behind.
When is a prediction published, and can it change?
Before kick-off. The number can update until the match starts; the one showing at kick-off is the one graded, and it is never edited afterwards.
Which leagues do you cover?
Every league with enough match history for the model to price honestly. Where there is not enough, a match is shown without a prediction and says so. The full list is on the Leagues page.
What does the AI do?
It answers questions about matches from our own data, and writes a preview before kick-off and a review after full time for pass holders. It explains the model's numbers in plain words and never changes them.
Version history
- v1.2 — 17 October 2026. Two corrections to how the model fits its numbers: the low-score adjustment is fitted on exactly the formula that is published, and a fit that would give an impossible probability is withheld. Published probabilities moved by at most 0.05 percentage points; replayed on 3,007 past matches neither version had seen, the two scored the same.
- v1.1 — 22 September 2026. Team ratings also learn from expected goals wherever match statistics are recorded.
- v1.0 — September 2026. The first published model: two statistical models, evenly blended.