Accuracy
How well the model's probabilities hold up against real results — measured, never assumed.
We publish real backtest metrics — Brier score, log-loss, accuracy and calibration — as soon as enough Global Cup 2026 matches have been played and scored. No results are fabricated before then: until the tournament produces outcomes to grade against, this page stays empty by design.
What we measure
Brier score
The mean squared error between predicted probabilities and actual outcomes. Lower is better — it rewards being both confident and correct.
Log-loss
Penalises confident wrong calls more harshly than Brier. Lower is better; a model that is sure and wrong is punished hard.
Hit rate
How often the most likely predicted outcome (home, draw or away) matched the real result. Simple, but coarse.
Calibration
Whether outcomes given a 70% chance actually happen about 70% of the time. A reliability curve shows where the model is over- or under-confident.
Underdog detection
How well the model flagged upsets in advance — measured on the matches it rated as high upset risk.
Compared against baselines
Raw metrics mean little in isolation, so every score is reported next to two reference points: the bookmaker consensus (margin-free implied probabilities) and a plain Elo model. Beating a baseline is the bar; matching it is honest; trailing it is reported just the same.
No live tournament results yet — this page will populate automatically after the group stage begins.