How the model works

The football model

Each match starts with eight signals: recent competitive form with friendlies discounted, the FIFA/Coca-Cola world ranking, squad market values from Transfermarkt, player-level ratings weighted by league and club, injuries and availability, venue and travel impact, head-to-head history, stakes and motivation.

The signals are folded into a strength gap between the two sides, which becomes the expected goals for each team, the Lambda values you see on every forecast page. Those Lambdas are fed through a Poisson scoreline model with the Dixon-Coles correction for low-scoring games. The output is the probability of every possible result, summed up into the three numbers in the bar: home win, draw, away win. The single most likely cell of the grid becomes the published pick.

Data quality

Each prediction carries a data-quality score between 0 and 1. Where data is rich, the model trusts what it found. Where data is thin, it pulls its estimate back toward the ranking and widens the band. A debutant with no record carries more uncertainty than a regular contender, by design.

Hit and miss

A Hit, marked with a star, is the exact predicted scoreline matching the final result. 2-0 predicted, 2-0 played.

A call, shown in green, is the looser metric: the home win, draw, or away win bucket was right even when the exact scoreline was off. The topbar shows both, the count of correct calls and, with a star, the count of exact scorelines. Every exact hit is also a correct call, so calls are always at least as many as stars.

Penalty shootouts do not move the metric. The score after 120 minutes, regulation plus extra time, is what counts.

What the model never touches

The referee never moves the winner. Card profiles, penalty rates and variance, yes. Who lifts the trophy, no. The referee dossier on each match page is researched in full, but it stays out of the strength gap.

We never invent data. A field we cannot verify is left empty or labelled as unknown. Sources are named on every forecast and stay attached to the page.

The tennis model

Tennis runs on a separate logic. Before the draw is out, the model produces editorial title estimates for the ATP and WTA top ten. Not a simulated bracket. The estimate weighs ranking, current form on grass, recent injury history, and the player record at Wimbledon specifically. Each player page makes the working visible.

Tennis is tracked separately from the football record so a Sinner miss does not pollute the WC 2026 numbers and vice versa.

What can be wrong

The model can be wrong. The sources can be wrong. The result can surprise everyone. Probabilities are not certainties. A 76 percent favourite still loses one in four times. We publish the band, you read it.