How it works

Success is calibration, not profit. When we say a result is 55% likely, it should happen about 55% of the time across many such predictions. We don't publish a betting return, because beating bookmaker prices is notwhat we're claiming to do.

We measure ourselves with the Ranked Probability Score (RPS) on the match result — the standard football metric, lower is better — plus Brier score and log-loss on each market, and reliability diagrams. All of it is public on the Results and Calibration pages, and it updates as fixtures resolve.

One statistical model per competition — a time-decayed Dixon–Colesmodel. It learns each team's attacking and defending strength from recent results (weighting recent matches more heavily) and a home-advantage term, then produces a full grid of exact-scoreline probabilities for the fixture.

Every market we show is just a sum of cells in that grid: the 1X2 result, over/under 2.5 goals, both-teams-to-score, and the most likely scorelines all come from the same place. We never model a market separately, and we never pool different leagues into one model.

In the major leagues where the data exists, we sharpen those strength estimates with expected goals (xG) — a measure of chance quality, not just the goals that happened to go in — blending xG with actual goals only where a pre-registered test showed it genuinely improves accuracy. Everywhere else the model runs on goals alone.

Tournaments like the World Cup and the Eurosare the model's hardest case. A team plays only a couple of group games — far too few to judge its strength from results alone — and matches are at neutral venues, so there is no home advantage to apply.

For these we do two things: we drop the home-advantage term entirely, and we anchor each national side to an external strength rating (World Football Elo) so an elite squad isn't mistaken for a minnow on the basis of two matches. Even so, these are our least-certain numbers, and we mark them low confidence rather than dress them up.

The short explanation on each match is written by a language model — but it only describes numbers the statistical model already computed. The language model never invents a probability. If it quotes a figure, that figure came from the Dixon–Coles grid.

• We don't claim to beat bookmaker odds, and we publish no ROI or tips.
• We don't predict live, in-play, or settled matches — these are pre-match only.
• We don't hide our misses. The track record counts every published prediction, including the ones that were wrong.

The track record is built without cheating on time: a prediction is always scored against the result using only what was known before kickoff, and the backtest fits on the past and predicts the next round, rolling forward and never peeking at the future.

We expect to clearly beat naive baselines (long-run base rates, “always home win”) and to land nearthe market — not above it. That's the honest ceiling for a public model with free data, and we'd rather say so than oversell.

Everything here is public, so the numbers are too. A read-only JSON API serves the same calibrated probabilities you see on the site — no key, no sign-up:

• /api/v1/fixtures — upcoming fixtures with 1X2, Over/Under 2.5, BTTS, and expected goals.
• /api/v1 — the index, including the single-match endpoint.

Use it for your own dashboards or embeds. Same honesty contract applies: these are calibrated probabilities, not betting advice.

Daily prediction digests land on our Telegram channel: t.me/pitchOddsPredictions. Same honest numbers as the site — calibrated probabilities, never tips.