How xG Stat called it
Every fixture gets a verdict before it kicks off. That verdict is frozen at kick-off and graded against the result afterwards, and the archived copy is never edited. This page is what that leaves behind: how often the model's pick was right, and how well its stated confidence matched what happened.
Model accuracy
6% better than guessing the usual home, draw and away split
0.211 against 0.226 for the guess. Lower is better.
Bar is what happened, tick is what the model said. Samples are small this early.
The record starts once 20 matches have been settled. Below that the headline moves by several points a match, which is a coin flip with a decimal point rather than a track record, so nothing is shown until there is something honest to show.
The confidence rows are the part worth reading twice. When the model says 60%, picks in that band should land about 6 times in 10. The other 4 times they do not, and that is the model working, not failing. A model that was right every time it said 60% would be a badly calibrated one.
The error score checks how good the probabilities were, not just whether the pick was right. It rewards the model for being confident and correct, and punishes it for being confident and wrong. A draw counts as a nearer miss than the other side winning, since it sits between the two outcomes. Lower is better, and the score is measured against a guess that always predicts the competition's usual home, draw and away split.
Short runs swing. Judge this over a season, not a weekend.