Football is a distribution, not a prediction.

At the centre of Metis is an independently developed generative match model. The current system estimates home and away scoring intensity using regression and chronologically updated measures of team strength and match context. A fixture must be forecast using only information available before that fixture — no hindsight, no retro-fitting.

From the estimated scoring parameters, Metis generates a joint scoreline distribution. This means the probabilities shown across the site are internally connected rather than independently generated headline predictions: the win probability, the over/under split, the both-teams-to-score figure and the clean-sheet chance on a match page all come from the same underlying distribution.

Betting markets are exceptionally strong football forecasters, and Metis says so plainly. Two forecasts appear on Metis, and they are kept deliberately distinct:

• Raw MetisThe independent statistical forecast, produced before any market information is applied. Stored separately, so the model can be evaluated honestly against the market.
• Metis ForecastThe public forecast. Where fresh de-vigged market prices are available, the headline 1X2 probabilities are anchored to them while retaining the scoreline structure required for everything downstream — a market-anchored Metis distribution. Match pages label the anchor status.

The point of the anchoring is coherence: a public forecast that benefits from the market's information, expressed as a full distribution the market does not publish.

Player projections are built on top of the match distributions. For Fantasy Premier League, the engine combines the match-level scoring environment with player-level estimates: start probability, expected minutes, goal and assist involvement, clean-sheet contribution and expected points.

Because the engine works with distributions rather than point estimates, it can also report a floor, median and ceiling for a player's score, and the probability of a blank or a haul. Two players with identical expected points can have very different risk profiles — that difference is often the real decision.

A generative model can be asked conditional questions: how does the forecast move if a particular team scores first, or if a key player does not start? Where the site exposes scenario tools, they are restricted to conditioning the production engine genuinely supports — hypotheticals the model cannot answer correctly are not offered.

Single-match accuracy is a misleading measure of a probabilistic forecaster: a good model is often "wrong" in any one match and right in the long run. Metis is evaluated the way probability forecasts should be:

• Calibration — when the model says 60%, it should happen about 60% of the time.
• Ranked Probability Score — a proper scoring rule over the ordered home/draw/away outcome.
• Market comparison — the raw model scored against de-vigged closing prices on the same matches.
• Walk-forward evaluation — out-of-sample only; a fixture is always forecast with data available before kickoff.

Every published forecast is stored with its timestamp, model version and market-anchor status, and results are settled against it. The long-term aim is visible performance — including the bad periods. A forecast that cannot be checked is not particularly useful.

Matches — match distributions and scenarios. Fantasy — player-level decisions. Markets — future quantitative market research and backtesting.

The World Cup 2026 model that started this project is preserved as a tracked public archive, including its full forecast-vs-reality retrospective.