Who will win?

In plain language: pick the model's top watch in a category and it wins that category about one time in three — against one time in six for a name drawn out of the six finalists at random. That is roughly 2.1× better than chance, on 29 hits from 84 shortlists. It is an edge, not a crystal ball.

Counted more generously — the favourite winning any GPHG prize that year, including the cross-category awards it might also be up for — the model hits 40% against a 21% baseline. Both measures are shown because they answer different questions; the headline above is the stricter one.

Every figure here is out of sample: for each edition the model was retrained with that year removed, then asked to rank it. The blind-guess baseline is measured from the same shortlists under the same rule, never assumed to be one-in-six. Shortlists whose winner we could not match to a finalist are excluded from both sides. Ranking quality across all shortlisted watches is 0.62 ROC-AUC. Full definitions →