How the WinningScore Model works

What the number is

Every win-probability on the site is an analytical estimate — home, draw and away add up to 100%. It is not betting: no odds, no bookmaker, no stake, no tip. It is a way to read a match, nothing more (see the disclaimer).

The model

The World Cup runs on a de-vigged statistical consensus built from public data, not our own model: a national-team cup gives each side too few matches to fit a club-style model reliably. Below is how the model works on the league editions.

For league editions the number comes from a Dixon-Coles model — a refined Poisson model of goals. From every finished match it fits each team's attack and defence strength (measured from goals blended with expected goals (xG), so a side that created the better chances is rated on chance quality, not just the final scoreline) plus a home-advantage term, then turns those into the expected goals for each side and the full 1X2 split. Two refinements matter: a low-score correction (independent Poisson under-rates draws and 0-0/1-0 results), and time-decay — last week's form weighs far more than a result six months old.

It is deterministic and reproducible: same data in, same number out, every time — which is exactly what lets us measure it (below). No LLM, no hand-tuning per match.

What it uses: goals scored and conceded, expected goals (xG), recent form, home/away split, time-weighting. What it deliberately ignores: injuries, suspensions, line-ups, motivation, weather, news. Those are real and they matter — so they live in our written match analysis, where a human can weigh them, rather than being baked into the number where they can't be measured.

Our model vs the reference

Behind the scenes we also compute a reference consensus — an aggregate of public data, de-vigged and kept only as a probability (never shown as a price). Where both exist you can switch between them with the source selector and compare. Being honest: that reference is the single most accurate 1X2 estimate that exists — no standalone model reliably beats it. We publish a model-only number by choice (we never touch gambling), not because it is sharper. The goal is a transparent, reproducible read — not a claim that we beat the reference.

How we measure it (RPS)

We score every pre-kickoff analysis against the real result using the Ranked Probability Score (RPS) — the standard metric for ordered 1X2 outcomes (lower is better; it rewards being close, e.g. reading it as a draw when the away side wins). A good score sits around 0.20, the level of a sharp reference line; the realistic 1X2 accuracy ceiling is about 50–53%, not the 70–80% people expect. We freeze each analysis before kickoff so the score is honest — the model never sees the result it is graded on.

For analysis and entertainment only. Not betting advice. See the matches →