The base signal is a player's transfer-market value — the price the football market puts on him. It sounds crude, but a decade of research (and our own testing) keeps landing on the same result: the wisdom-of-the-crowd baked into market prices is very hard to beat for predicting results. Thousands of clubs, scouts and bidders price a player more accurately than most hand-built rating systems can.

Two corrections sit on top. First, age: the market prices resale value, so it over-discounts a 33-year-old who is still excellent now and over-prices a teenager who might be excellent later. A position-specific age curve pulls the number back toward present-day ability. Second, a small recent-form adjustment from club expected-goals and defensive numbers — capped at 15% of the rating, because form is noisy and the market has usually priced it in already.

Everything is then expressed as a z-score within position, so a goalkeeper and a striker sit on the same scale.

Not everyone contributes equally. The model estimates expected minutes for each player — better players are likelier to start and finish — and takes a minutes-weighted roll-up of the squad into separate offensive and defensive ratings. Those are mapped onto the Elo scale, the same points system used to rank national teams, so a squad built from market values becomes directly comparable to a team's results-based rating.

The last move is a 50/50 blend: half the rating comes from the players, half from the team's actual Elo — its track record of results. Why both? Market value is excellent for European sides but undersells some South-American and results-driven teams whose players are cheaper than their results deserve. Blending hedges "who's on paper" against "what they've actually done," and it was the single biggest accuracy gain in testing.

A match is mostly a count of rare events — goals — and the natural tool for counts is the Poisson distribution: give it an expected number of goals (called λ, lambda) and it returns the chance of 0, 1, 2, 3 … goals. Each team's λ comes from a small regression fit on 898 World Cup qualifiers, where expected goals depend on just two things: home advantage, and the Elo gap between the sides. Bigger favourite → higher λ → more goals expected.

Multiply the two teams' goal distributions together and you get a grid of every scoreline and its probability — which collapses into win / draw / loss and an expected score. One real-world wrinkle: independent Poisson slightly under-counts 0-0s and 1-1s, because level teams play it safe. The Dixon–Coles correction nudges those low draws up to match reality. This grid is exactly the score matrix you can expand on the Matches page.

Knowing each match as a probability is not the same as knowing who lifts the trophy — you have to chain 104 matches through a group stage and a knockout bracket, where every result reshuffles who meets whom next.

So the model runs a Monte Carlo simulation: it plays the entire tournament from first whistle to final 10,000 times, each match resolved by a random draw from its score distribution, with extra time and penalties included. Shootouts are treated as a coin flip — the evidence says they essentially are. Counting how often each team reaches each round gives the percentages on the Tournament page.

The regression above is fit on qualifiers — but a World Cup is not a qualifier. A point-in-time backtest of recent major tournaments surfaced two systematic gaps, and each is corrected with a single dial:

Host advantage is smaller than an ordinary home game. A qualifier home side gets a large boost; a World Cup host gets about half of it. The co-hosts USA, Mexico and Canada still get a real edge — roughly +5 percentage points on home soil, not the +11 a naïve home bump would hand them.

World Cups score more. Recent World Cups average about 2.7 goals a game; the qualifier-fit model expected nearer 2.3. A small lift to the goal level closes the gap — and, helpfully, also improved who-wins accuracy, because World Cup games score more and draw less than qualifiers.

Everything above assumes expected lineups. About an hour before kickoff the real elevens are published — and a model built on players can use them directly. An automated agent reads the confirmed lineups from ESPN, re-rates both teams with the actual starters, and republishes the prediction.

The swings are real. In the USA–Paraguay group game the USA rotated into a weaker back line while Paraguay named their strongest attack; the forecast moved roughly eight percentage points toward Paraguay between the provisional and confirmed lineups. A team-level model could not see that coming.