Forecasting ultimate, tournament after tournament

How it works, in two steps

The model knows only one thing: the scores of games already played. No player names, no club reputation, no history.

1. 1

   Measure each team's strength

   A strong team wins by a lot. A team's strength therefore shows in its average score margin, computed game after game.

   But not every 15-13 is worth the same. So the computation adds an adjustment that depends on the opponents' strength.

   Reading a strength: "+3" means three points better than the division average. A +3 in Open cannot be compared with a +3 in Women's: divisions never play each other.

2. 2

   Turn a strength gap into a percentage

   The favourite is always the team with the greater strength. If both teams have the same strength, each has a 50% chance of winning. The wider the strength gap, the closer the favourite gets to 100%, without ever reaching it.

   Strength gap between the two teams: 3.0 points

   Favourite's chance of winning: 82%

   Curve set on the tournament shown below.

No cheating: every forecast is dated

Every tournament is replayed round by round. Before each round, the model is recomputed with completed games only, never with the games it has to predict. The accuracy shown is what someone following it live would have seen. Each tournament starts from scratch, with no memory of previous ones.

As long as one of the two teams has not played yet, the model gives no forecast. That covers the whole first round of a tournament.

Is it reliable? The record over 36 tournaments

Two questions. Does the model pick the right winner more often than simple methods? And are its percentages right: when it says 75%, does the favourite really win three times out of four?

Ideas tested, and dropped

The model only looks at scores. Other information looked promising. Each idea was tested like the model itself, using only games already played, and had to hold on independent data: the first two on the close games of two World Championships, 2026 then 2022. None improved the forecasts reliably.

• How scoring is spread across the roster

  Dropped

  Does a team whose offence relies on two or three players lose close games more often?

  Eight ways of measuring concentration were tried. The effect points one way in 2026 and the other way in 2022: that is noise, not a signal.

• Point duration

  Dropped

  Is a team that scores quickly on offence and makes points drag on in defence stronger than its score says?

  The most serious lead: the effect shows in 2026 but fades in 2022, on twice as many games. A real effect would get clearer with more data; this one vanishes. It was a false positive.

• Memory of previous tournaments

  Dropped

  If a team keeps most of its players from one year to the next, do last year's games help forecast it?

  A slight improvement when at least 70% of the roster is the same, but too small to be sure. The only clear benefit: forecasting the first games of a tournament. Within one European season, however, keeping a club's strength from one stage to the next clearly improves forecasts: that is what the seasons do.

What the model does not know

• It only sees scores: no fatigue, no injuries, no wind, no teams easing off once qualified.
• Early in a tournament each pool is a world of its own: comparing two pools assumes they have the same level, which is never guaranteed.
• An 80% forecast is wrong one time in five, by definition. Upsets are part of the game.