Skip to content
Presaira

We lock in our predictions before kickoff. You challenge them with yours.

Presaira forecasts football today, with basketball and Formula 1 coming soon. Every call goes on record, and it's free to challenge and play against.

See the World Cup forecastSee every sport

48 nations104 matches

World Cup 2026

The road from kick-off to the Final, complete. Tap any stage to look back at it.

Every match the model called, and what happened.

See the forecast

All 104 matches were forecast before kickoff and scored afterwards. None was omitted.

The 2026 World Cup is the first with 48 teams and 104 matches, the largest in the tournament's history.

Where we forecast

The same model runs every competition below. Each one gets the whole engine and the same published record.

World Cup 2026

Archived
48teams
104matches
12groups
0.528Brier score, across all 104 matches
View forecast

Champions LeagueUCL

Open
36clubs
8matchdays
189matches
8 of 36qualify direct from the league phase
See the model

Top five leaguesTop five leagues

Roadmap
not publishedclubs
not publishedmatches
not publishedmatchdays

No figures published yet.

5 leaguesone club engine, weekly cadence
Email me when this opensNotify me

Formula 1

Roadmap
not publishedrounds
not publisheddrivers
not publishedteams

No figures published yet.

Every roundgrid to podium, uncertainty shown
Email me when this opensNotify me

NBA

Roadmap
not publishedteams
not publishedgames each
not publishedrounds

No figures published yet.

Full seasonregular season games modelled
Email me when this opensNotify me

Another competitionAny competition

Roadmap
any fixture listinput
any formatrules
same enginemodel
Yoursany fixture list can be modelled
Get in touch

The full record

Every prediction Presaira has published, and what happened next.

0.528Brier score, all matchesdistance from what we published to what happened, lower is better
0.898Log loss, all matchespunishes a confident answer that turned out wrong, lower is better
64.4%Most likely outcome came first
104Matches scored, none omitted

These figures cover the 2026 World Cup only. Each match was scored against the last forecast published before its kickoff, and the totals were taken after the tournament.

The curve1/6

FIFA World Cup 2026

Every band of probability the model published during the tournament, against how often the thing actually happened. On the line is a model that was right about how often it would be right.

World Cup 2026, published probability against what happenedEach point is a band of published probabilities. On the line means the model was right about how often it would be right. Below the line means it claimed more than happened, and above it means it claimed less.Perfect calibration0%25%50%75%100%0%25%50%75%100%Probability we publishedHow often it happened128

Each point is a band of published probabilities. On the line means the model was right about how often it would be right. Below the line means it claimed more than happened, and above it means it claimed less.

World Cup 2026, published probability against what happened
Probability we publishedHow often it happenedMatches in this band
10%0%2
17%3%34
26%18%128
34%29%62
44%63%41
55%81%32
64%83%12
71%0%1

FIFA World Cup 2026

One row per band of published probability. The hollow mark is what we published, the solid one is how often it happened, and the bar between them is the distance. Thicker bars carry more matches.

Below 40 percent, it claimed more than happened

27%19%

226 scored observations

Above 40 percent, it claimed less than happened

51%72%

86 scored observations

  • 40%the error changes direction here

what we publishedwhat happenedbar thickness is the number of matches in the band

FIFA World Cup 2026

The tournament was not one thing. Brier score, where lower is better, for the group stage and the knockout separately, and for all 104 matches together.

Axis truncated: the bars run from 0.450 to 0.667 rather than from zero. 0.667 is the Brier score of even odds on all three outcomes, the worst score on the panel, and 0.450 is the nearest twentieth below the best. Lower is better.

FIFA World Cup 2026

Brier score, where lower is better. Every row is scored on the same matches with the same rule, and each scope is scored against a base rate computed for that scope.

All matches104

Group stage72

Knockout stage32

Axis truncated: the bars run from 0.450 to 0.667 rather than from zero. 0.667 is the Brier score of even odds on all three outcomes, the worst score on the panel, and 0.450 is the nearest twentieth below the best. Lower is better.

FIFA World Cup 2026

The same bands, sized by how much of the record sits in each. The weight sits low: 226 of the 312 observations were published below 40 percent, and that is the range where the model claimed more than happened, 27% published against 19% observed. The other 86 sit above 40 percent, where it claimed less.

  • 40%the error changes direction here

Where it was most confident, and what happened

  • 71%0%1 scored observation
  • 64%83%12 scored observations
  • 55%81%32 scored observations

Read the count before the rate. The highest band the model ever published in holds a single match, and that match did not happen, so the most confident call of the tournament was wrong. The band below it holds twelve and came in above what was said. Both are here.

2018 and 2022 backtests, not the 2026 record

The 2018 and 2022 backtests, on the same pooling. They are here to show that the pattern above belongs to 2026 rather than to the model, and every figure on this panel describes a tournament that is not the 2026 World Cup.

2018prior tournament192 match and outcome pairs

This repository holds a reliability curve for this tournament and no headline metrics, so no Brier score is shown. None is borrowed from another year.

2022prior tournament192 match and outcome pairs

Brier 0.621 across 64 matches, against 0.667 for even odds on all three outcomes.

The unit on both rows is the match and outcome pair: one match contributes three observations, one for a home win, one for a draw and one for an away win. Both years read 192 because both tournaments were 64 matches, not because a figure was repeated. The 2022 block states its 64 matches beside it, and 2018 divides the same way.

Both are backtests: the model was run against a tournament that had already finished. The 2026 figures on every other panel are live walk forward, where each match was scored against the last forecast published before its own kickoff. The two are not comparable and are deliberately never placed on one chart.

Every run stores its model version, git commit, data hash and RNG seed.

Read the methodology

How it works

The model publishes first and cannot change its answer. You answer the same question, and both are marked the same way.

Coming soon

  1. 01

    The model goes first

    We publish a probability for every match before kickoff, and it never changes.

  2. 02

    You go second

    You give your own probability for the same match. Free, nothing staked.

  3. 03

    The gap is your score

    Both answers go through the same rule, called Brier, against what happened.

  4. 04

    Confidence costs

    Saying 90 percent and being wrong costs more than saying 60.

  5. 05

    The table

    Everyone is ranked on the same matches, with the model in the table.

  6. 06

    Mini leagues

    Run a private table with friends over the same matches.