121 possible scores

Model 01 · World Cup 2026 · finished

The World Cup model.

A Bayesian scoreline model that forecast all 104 matches of the 2026 World Cup before kickoff, and replayed the whole 48-team tournament 20,000 times to price every team's path.

104matches forecast before kickoff
49k+international results it learned from
20,000full tournaments simulated
~40daily refits, with no leaks of future data

Method

How one World Cup forecast was made.

Every day of the tournament the model ran this pipeline from scratch, using only results that were already known.

  1. Point-in-time data

    A result counts only if it was observed before the forecast. Matches on the forecast day itself are excluded.

  2. Elo, rebuilt

    Elo ratings are recomputed from scratch each time: start at 1500, K = 40 scaled by match importance, a 100-point home edge.

  3. Bayesian fit

    Attack and defence for every team, fitted with ADVI over 30,000 iterations, keeping 1,000 plausible versions of every team.

  4. Venue

    Hosts at home get a 1.4× home edge. Everyone else plays on neutral ground, where the home edge is split evenly.

  5. 121 scores, then 1X2

    An 11 × 11 score grid for each of the 1,000 draws, averaged, then summed into home, draw and away.

  6. 20,000 tournaments

    Groups, tiebreakers, the eight best third-placed teams, extra time and penalties, all played out 20,000 times.

Contributing factors

Everything that moves a World Cup forecast.

Follow the final through the machine, from the data it reads to the forecast it publishes. Every number is a real setting from the published model.

All 104 forecasts

The whole tournament in one shape.

Scroll to replay the World Cup from the opening match to the final. Each dot is one real forecast, placed by its three probabilities.

Two roads to the final

Spain and Argentina, match by match.

Before every knockout match the model split its belief three ways: win, draw (penalties) or loss. Watch each team's stream divide at every gate, and follow the channel that really happened.

Team wonDraw, went to penaltiesTeam lostPercentages are the model's, issued before kickoff. In the final: Spain win · draw · Argentina win.

How it did

The record, stage by stage.

Every forecast was scored once the match was over. RPS rewards confident forecasts that turn out right and punishes confident ones that turn out wrong. Lower is better.

Mean RPS by stage

The Round of 32 was its sharpest stage. The single third-place match was its worst.

Said versus happened

Draws and goal margins were almost exactly right. Favourites won more often than it said: it erred on the cautious side.

Called

Spain as champion

Spain and Argentina were its top two title picks from the group stage on. It rated the final itself close to a coin flip.

Called

All four semi-finalists

Each was already in its projected last eight, published before the Round of 32 ended.

Called

The final, at 90 minutes

Its single most likely score for the final was 0–0, at 13.9%. Spain won it in extra time.

Missed

Spain 91.5% v Cape Verde

Its most confident group pick ended 0–0. In the Round of 16, Norway beat a Brazil side it had at 60.4%.

What was tried

Three ideas changed the model.

Each idea was tested on its own, mostly on past matches the model had never seen. Three were adopted; the other four on this map were tested and dropped.

    Also tested and dropped Counting friendlies for less when learning attack and defence, and rest days between games.

    Still open In the most one-sided 10% of past matches, big wins came only 73% to 87% as often as the model expected, and no fix has been adopted.

    From build to archive

    Eight days to build. Thirty-nine days of World Cup.

    From the first commit on 3 June to a public archive on 28 July. The two gates, shown as diamonds, are changes with the owner's sign-off on record.

    1. M1 · 3 Jun

      Build the engine

      A results database, Elo ratings, a goal model and a 20,000-run simulator; first forecast on 7 June.

    2. M2 · 8 Jun

      Test, then adopt

      Ideas were checked one at a time against real past matches; three passed and were kept.

    3. Gate A · 10 Jun

      Owner signs off

      The host boost went live only after the owner approved it at a checkpoint.

    4. M3 · 10 Jun

      Dress rehearsal

      A practice run caught a bug that hid results typed in late at night; then settings were frozen.

    5. M4 · 11 Jun

      A forecast every day

      About 40 daily re-runs; when the results feed ran late, scores were typed in by hand.

    6. Gate B · 1 Jul

      Matching bug fixed

      Swapped home and away teams left Canada at 71% for a round it had reached; the owner approved fixes.

    7. M5 · 19 Jul

      Spain won the final

      All 104 games were then scored; on 28 July, 24 forecast snapshots went into a public archive.

    NextThe Premier League model→