Glasgow 2026 · Model Report Card
Commonwealth Games 2026 · Marking Our Own Work
Model Report Card
We locked medal probabilities before competition began. This page marks them, including the parts that went badly.
Brier score is the mean squared error of a probability — lower is better, and 0 is perfect. On its own it is close to meaningless here, because a 30-runner field scores well simply by giving everyone a low chance. Skill fixes that by comparing against a uniform-within-race baseline that says every athlete in the field is equally likely: 0 means we knew nothing that baseline didn’t; 1 would be perfection.
Calibration
A forecast that says 30% should be right about 30% of the time. This is the only view that shows how a model is wrong — systematic overconfidence looks nothing like noise, and a single skill number cannot tell them apart.
Where the winners came from
“Our favourite won” is a harsh binary in a twelve-strong field. Where the actual winner sat in our order says much more: consistently second is a good model, scattered is not.
Swimming, after the fact
Kept separate from everything above, and deliberately not averaged into the headline. The athletics numbers are a forecast we published before a race was run. These are not: the Glasgow pool meet finished on 26 July, and the model was run over it afterwards using only World Aquatics history dated before the meet. That is a fair test of the model, but it is not a prediction, and folding the two together would quietly launder one into the other.