How the GPHG Index Score works

It all started with a simple premise. The official GPHG website listed all watches individually, by category and by year, but there was no way to search the archive or compare brands across categories or years. The GPHG Index reads every watch entered since 2001 — brand, model, complications, price and description. This historical data has been used to train an AI model to rank entrants using a GPHG Index Score in an attempt to quantify the subjective.

What moves the needle?

One signal dwarfs the rest: a brand's own history at the competition. Pedigree, quantified. Price, complications and description themes add a little; case size and water resistance almost nothing. A fine watch from an unknown maker scores low here, statistically a long shot.

A score, not a prophecy

The GPHG Index Score is a calibrated probability put on a 0–100 scale, so a score of 40 really means about a 40% chance. It is a way of asking whether the recorded facts about a watch carry any signal about its chances. It is not a crystal ball. Treat it as an informed guess.

How it is scored, and what the numbers mean

Every accuracy figure on this site is out of sample. For each edition the winner-from-shortlist model is retrained with that year removed, then asked to rank it — so it is never graded on watches it has already seen. In-sample numbers are far higher and are published nowhere on this site, because they measure memory rather than skill.

Sample: 84 shortlists across 6 editions, 2020–2025. Each shortlist counts once; no category is weighted above another.
Favourite: The single highest-scoring watch in a category. There are no probability ties at the top of any shortlist in the record, so no tie-break rule is applied.
A hit: That favourite won the prize for its own category. This is the strict reading, and the one the headline figure uses.
Result: 37% of favourites were hits (31 of 84).
Blind-guess baseline: 17%. Measured, not assumed: for each shortlist, the number of its watches that won that category divided by the number of finalists, averaged across all 84. It lands near one-in-six because exactly one watch wins each category — but it is computed rather than taken on trust.
Edge over chance: About 2.2× better than picking a name out of the six at random. Ranking quality across all shortlisted watches, not just the favourite, is 0.62 ROC-AUC.
Excluded: Shortlists whose winner we could not match back to a finalist — the palmarès has one, our model-name match failed to find its row. Charging those to the model would score our data pipeline, not its ranking. They are removed from the baseline on identical terms.

The looser measure, and why both are shown

Ask instead whether the favourite won any GPHG prize that year, and the model scores 41% against a 21% baseline (35 of 86). The gap is not slack in the model — it is the GPHG's structure. Prizes are also awarded outside the ordinary categories: the Aiguille d'Or above all, plus Horological Revelation, Audacity, Chronometry and the rest, and those go to watches that are also finalists in a normal category. So a shortlist of six routinely contains two or more eventual prize-winners; one 2024 shortlist contains four. That inflates the model's hit rate and the blind-guess baseline together, which is exactly why both halves have to be counted the same way.

An earlier version of this site advertised "34% vs 17%", pairing the loose numerator with the strict denominator — and the 17% was in fact the model's Brier score, read out of the wrong column. Both figures here are recomputed from the prediction records on every request.

How the two vocabularies are matched

Answering the strict question at all requires knowing which prize belongs to which shortlist, and the two lists do not use the same words. The competition's entry pages say TIME ONLY, CALENDAR AND ASTRONOMY, “PETITE AIGUILLE”; the official palmarès says Time Only Watch, Calendar & Astronomy Watch, Petite Aiguille Watch. The scrape's own punctuation drifts between years, too — 2023 wrote MEN where every other year writes MEN'S. A single normalisation reduces both sides to the same key, and it is checked on every data rebuild: every shortlist category from 2020 to 2025 maps onto exactly one prize, and every unmapped prize is a known cross-category award. A new prize the crosswalk has not seen is reported as an error rather than silently ignored.

Category by category the record is uneven, and the samples behind each are small — a handful of editions apiece. The technical categories are read far better than the decorative ones. Treat any single-category figure as indicative, not settled.

Two models

A winner-from-shortlist model (gradient-boosted trees) ranks each year's official shortlist and produces the GPHG Index Score for those editions. For 2026 — before the shortlist is announced — the score instead comes from a separate, archival preselection model (a plain random forest), which asks a different question: which entrants are likely to make the shortlist, not which shortlisted watch wins. It's the weaker of the two models and has no per-watch explanation the way the winner model does, which is why 2026's picks don't come with a "why the model ranked it here" list. The GPHG's own two-round Academy-then-Jury process hasn't changed — the 2020 boundary reflects a gap in what we could record, not a change in how the competition runs. Win counts come from the official GPHG archive and are exact back to 2001; entrant records — and therefore win rates — are most complete from 2005 on.

What the brand numbers mean

Entries: Watches a brand has entered, from our own record of the competition. Complete from 2005; sparse before that.
Shortlisted: Entries that reached the final six in their category.
Prizes: Prizes won, from the official palmarès. Exact back to 2001 (2001–2025).
Prizes per entry: Prizes divided by recorded entries. A ratio, not a probability: because one watch can take both its category prize and a cross-category award, and because prize counts reach back further than entry records, the figure can in principle exceed 100%. It was previously labelled "win rate", which invited exactly the wrong reading.

The two sources have different reach, which is the single most important caveat on any brand page: prize counts run 2001–2025 and are authoritative, while entry and shortlist counts are ours and are only complete from 2005. Any ratio that divides one by the other inherits that mismatch.

Not affiliated

This is an independent data project and neither affiliated, nor endorsed by, the Fondation du Grand Prix d'Horlogerie de Genève. Prize and entrant data come from GPHG public records.

Model details

Model B — winner-from-shortlist

Powers the GPHG Index Score for every edition with a published shortlist. Given a shortlisted watch, ranks it against the others in its category by likelihood of winning.

Algorithm: LightGBM gradient-boosted trees (400 estimators), wrapped in a sigmoid-calibrated classifier so the output reads as a genuine probability.
Training data: Only shortlisted watches — 1,386 rows, 280 winners — spanning 2005–2018 and 2020–2025 (2019 excluded; no usable labels that year).
Inputs: Brand win-rate prior (train-only, leak-free); category; complication flags parsed from the text (tourbillon, perpetual calendar, repeater, chronograph, moonphase, split-seconds, skeleton, GMT, self-winding); complication and function counts; power reserve; size, thickness and price; the watch's price percentile within its own year and category; description length; a 12-dimension TF-IDF → SVD embedding of the marketing description.
Output: A calibrated win probability per shortlisted watch, ranked within its (year, category) — this is the GPHG Index Score. SHAP contributions from this same model drive the "why the model ranked it here" list.

Model A — preselection (archival)

Powers the GPHG Index Score for an edition before its shortlist is published — 2026, currently. Given the full entrant pool, estimates how likely each watch is to make the shortlist in the first place.

Algorithm: A plain RandomForest (300 trees) — a modernised rerun of an earlier research notebook.
Training data: The full entrant pool (not just the shortlisted watches) across the same years as Model B — 2005–2018 and 2020–2025.
Inputs: Brand, Model, Case material, Movement, Functions and Collection (ordinal-encoded); water resistance, size, thickness and price; description length.
Output: A preselection probability per entrant, ranked within category. No per-instance explanation is available the way Model B's SHAP reasons are, so 2026's picks don't come with a "why the model ranked it here" list. It's also the weaker of the two models (around 0.67 ROC-AUC).