Methodology
Rankings here come from a Bayesian model fitted to IFSC lead results, 2021–2026. A lead result says how far a climber got, so the model works move by move.
One move at a time
Each climber has a rating and each route has a difficulty that rises from move to move. The chance of passing a move depends on the gap between the climber and the move. Some moves are much harder than their neighbours: these cruxes are fitted for each move and are shared by everyone who climbs the route, which is why a whole field can fall at the same hold. A climber's height on a route is the first move they fail.
Good days and bad days
Climbers do not perform at exactly their rating. The model has a climber-by-event term and a climber-by-route term, so someone can be sharp all weekend or simply suit one route.
Across seasons and events
Each season starts from a climber's last rating, with room to move. A climber the model has never seen starts from the ratings of earlier debutants, which sit well below the established field but vary widely. Routes at continental events are placed against World Cup routes through the climbers who appear at both. Those offsets are uncertain in any one season.
Elo and thin evidence
Elo is the rating on a chess-style scale. Men and women are fitted separately, so compare within a sex. Each rating has a 95% range. A rating built on few routes can be high but it is not settled, so climbers with fewer than 8 routes in a season are greyed and left unranked.
Forecasts
A forecast simulates the rest of an event 20,000 times under the competition's ranking and cut rules. Each simulation draws climbers' form, then draws routes from those of the same round and level in recent seasons, with fresh cruxes. Probabilities come from how often each outcome occurs. The simulator and the model share one description of a climb, checked against each other.
Forecasts are made before the event, after qualification and after the semi-final. All of them on this site are backtests. They were computed afterwards, with the model refit on only the results available at that stage. None was posted in advance.
How well it works
31 events and 62 competitions from 2024 to 2026, 20,518 predictions. Skill is the improvement in Brier score over a base rate.
| Outcome | Pre-event | After quali | After semis |
|---|---|---|---|
| Reach semi | +53% | ||
| Reach final | +40% | +35% | |
| Podium | +38% | +40% | +35% |
| Winner | +28% | +32% | +36% |
Forecasts are close to observed rates:
| Forecast band | Mean forecast | Observed |
|---|---|---|
| Under 5% | 0.7% | 0.9% |
| 5–15% | 9.2% | 9.5% |
| 15–30% | 21.8% | 22.8% |
| 30–50% | 38.9% | 36.7% |
| 50–70% | 59.7% | 61.3% |
| 70–90% | 80.0% | 81.6% |
| 90%+ | 96.2% | 95.2% |
The model was built on 2024 and 2025, then checked once on 2026, though the skill and calibration tables above pool all three seasons. Ratings beat a simple recent-form measure in all 31 events, calling 1,199 of 1,504 semi places against 1,039. That measure is a modest opponent, and we have no history of the official world ranking to compare with.
What it does not do well
- The simulation adds little beyond the ratings: ranking climbers on rating alone does about as well. The value is in the ratings.
- Long shots reaching finals still came in about twice as often as forecast on 2024 and 2025, though not on 2026.
- Debutants are a little overrated, because recent debut classes have been weaker than earlier ones.
- One competition is a small sample, so a single event's score says little.