How good are these forecasts?
Every forecast scored on seasons the model never saw while its settings were chosen.
Held-back seasons
| Forecast | Forecasts | Log loss | Favourite won |
|---|---|---|---|
| The model | 2723 | 0.869 | 60% |
| Base rate | 2723 | 1.052 | 47% |
| Recent form | 2723 | 0.981 | 53% |
At the 2026 World Cup (104 matches) the model's log loss was 0.857 against 1.067 for the base rate, and its favourite (win, draw or loss) was right 64% of the time. Every one of those forecasts was made before the match, from the ratings as they stood.
Every national team is rated on every men's international since 1872 — World Cups, continental championships, qualifiers, Nations Leagues and friendlies — with one Elo scale for all of them.
A match moves the ratings by how much it mattered: a World Cup match counts in full, a continental final 0.85, a qualifier or Nations League match 0.65, another tournament 0.5 and a friendly 0.35, the proportions the World Football Elo Ratings use. Home advantage applies only when a side plays in its own country.
A draw is its own outcome. Knockout matches settled on penalties count as draws for the ratings.
The next World Cup is in 2030 (Spain, Portugal and Morocco, with opening matches in South America), with 48 teams. Qualifying hasn't started, so there is no tournament simulation yet; it will appear once the field and groups are known.
Squads change more between international windows than club sides do; the ratings don't know who was picked, and friendlies are often used to try players out.
The champion record replays every World Cup since 1998 from the ratings as they stood, but the model's settings were fit on the internationals of 2016 to 2023, so only the 2026 tournament is a true holdout.
Women's internationals are not included.