How it works
Every knockout tie is rated by eight statistical and machine-learning models, plus a calibrated blend — reading only the football the teams played in the group stage. No betting lines, no market prices. The published number is the chance of each 90-minute result: a win, a draw, or a loss.
Here is the whole chain, from a raw FIFA match report to the published call. Each stage feeds the next; nothing along the way reads a price.
After every match, FIFA publishes a Post-Match Summary Report — around fifty pages of detailed analysis. We read the text on two of those pages: the Key Statistics page and the Phases of Play page. That is the entire source.
We read only on-pitch football: goals, shots, possession, pressing, running distance and the like. We never read a price, an odds line or any betting data — the prediction is generated blind to the market, by design.
Nine numbers per team come off those pages. Here is each one in plain language, with an example value.
How many goals a team is expected to create in a match, based on the quality of the chances it made. Higher is stronger.
How many goals a team is expected to concede, based on the chances it allowed. Lower is stronger.
The share of the match a team had the ball.
The share of play a team spent pressing high up the pitch, hunting the ball near the opponent's goal.
The share of play a team spent in fast counter-attacking phases.
The share of play a team spent sitting deep to defend its own goal.
Distance covered at high speed (sprinting and fast running). A read on how hard a team physically worked.
A pre-tournament estimate of overall team strength, set before a ball was kicked. Higher is stronger.
How many group games we have data for — three, the full group stage.
“Expected goals” (xG) is a shot-quality estimate: it scores how many goals the chances a team created — or allowed — were really worth, rather than just counting the goals that happened to go in.
A team plays only three group games, so any single number is noisy — one freak result can flatter or bury a side. To settle it, we average the three games and then gently pull each number toward a sensible prior. Statisticians call this Bayesian shrinkage: it trusts the data, but not blindly, so one wild game can't dominate.
The result is the team profile — the steady, de-noised picture of each side that every model reads from. From here on, the eight models all work off the same two profiles.
Each model looks at the two profiles through a different lens and turns them into a win/draw/loss. None of them is the final word; their disagreement is the point. Here is what each one leans on and how it decides.
A classic football scoreline model. It takes each team's expected goals for and against — shrunk toward its strength rating — and builds a grid of every plausible scoreline with its probability, including a special correction for low-scoring draws like 0-0 and 1-1, which raw models tend to underrate. Adding up the grid gives the win/draw/loss. It is strong on the realistic shape of football scores and on getting draws right.
A single strength rating that rises and falls across the three group games based on the quality of chances a team created — its expected goals — not just the final score. A team that out-created its opponents but lost still climbs. The gap between the two ratings, plus a term for the draw, becomes the win/draw/loss. It rewards teams whose results flattered or cheated them.
Instead of one rating per team, it keeps a separate attack rating and defence rating, both updated from group-stage expected goals. To forecast a tie it pits one side's attack against the other's defence, and vice versa, to get each team's goal expectation. It shines when a team is lopsided — a great attack with a leaky defence, say — which a single rating would blur.
A strength rating that also weighs physical load. Teams that ran the most high-intensity distance in the group stage get docked a little, on the logic that they are more likely to be tired deep in a tournament. It is built to catch fatigue effects that pure strength ratings miss.
Beyond raw strength, it compares how the two teams actually play — possession, pressing, counter-attacking, sitting deep. A stylistic mismatch can favour the lower-rated side, like rock-paper-scissors layered on top of strength: a great counter-attacking team can trouble a possession side that out-rates it.
A gradient-boosted decision-tree model — a machine-learning method that builds many small decision rules and stacks them. Trained on the group matches and fed chance-creation and defence-penetration stats (shots, line-breaks, ball progressions), it learns patterns that map the gap between the two teams' numbers onto a win/draw/loss.
A random forest — another machine-learning method, an average of many independent decision trees — trained on passing-structure and style features like possession and build-up shape. It is a different statistical lens on the same question, and where it disagrees with the gradient-boosted model is itself informative.
A meta-model that never looks at the pitch at all. Its only job is to learn, from how the seven other models performed across the group stage, how much to trust each one — then weight and combine their views into a single read. It is a model of the models.
The eight views are first pooled into one consensus. Then a small bias correction, learned on the group stage, nudges it where the models were collectively off. Finally a knockout draw calibration applies: knockout matches end level after 90 minutes far more often than group games — about 31% of the time at Qatar 2022 — so on genuinely even ties we lift the draw toward that historical rate. Lopsided ties barely move, and the favourite is always preserved.
The result is the published “Multi-calibrated 8-in-1” win/draw/loss. Every match shows all nine views — the eight models and the blend — side by side on its own page, and we score them in the open as matches finish, so anyone can check the record.
Predict Raven