No look-ahead
Ratings and ensemble weights are frozen at their pre-match values before the result updates them.
One platform. Every side of the game.
Sand Rating evidence lab
Predict first. Learn second. Prove every gain.
Duna replays beach volleyball history in order. Each model must publish its probability before it sees the result, then earn its place on accuracy, calibration, Brier score, and log loss.
Latest walk-forward run2,336pre-match predictions663 players · Jun 22, 2017–Aug 2, 2026
Real challengers
Simple baselines, Elo variants, Duna ablations, and an online adaptive ensemble compete on the same matches.
Governed promotion
A challenger does not reach production because it sounds advanced. It must improve walk-forward error and remain explainable.
Cumulative learning curves
Run 1090dc4cPrediction error over time
Model comparison
walk-forward-1.0Same history. Same scoring rules.
ModelAccuracyBrier ↓Log loss ↓Calibration ↓AUC ↑
50% even priorbaseline50.0%0.25000.69310.10190.500
Elo team averageelo60.8%0.22570.64160.06030.651
Elo weak-linkelo60.8%0.22470.63900.05540.651
Duna win-only ablationduna60.7%0.22590.64230.0607
Calibration check
When we say 70%, it should happen about 70% of the time.
The diagonal is ideal calibration. Each marker shows predicted probability against the observed win rate; its number is the match count in that bucket.
113217465677837721890
Historical audit trail
Latest 12 rated resultsWhat the model knew before the match.
Megan Kraft / Kelly Cheng3.53 · 3.54 pre-match82%Clémence Vieira / Aline Chamereau3.05 · 3.03 pre-match
Rémi Bassereau / Calvin Aye3.18 · 3.16 pre-match32%Rotar / Gauthier-Rat3.39 · 3.38 pre-match
Rotar / Gauthier-Rat3.38 · 3.36 pre-match67%Timo Hammarberg / Tim Berger3.15 · 3.17 pre-match
Carolina Solberg Salgado / Rebecca Cavalcante Barbosa Silva3.49 · 3.51 pre-match43%Megan Kraft / Kelly Cheng3.56 · 3.56 pre-match
Katja Stam / Raïsa Schoon3.46 · 3.40 pre-match76%
How Sand Rating moves
Doubles is not two singles ratings.
Duna models team strength with extra weight on the weaker partner, reflecting targeting and side-out pressure. Updates consider score margin, uncertainty, evidence quality, and repeat-opponent decay. Weekly display gains are capped, while losses remain uncapped.
A player’s public 1.00–8.00 number is a readable projection of an internal strength and uncertainty state. World ranking points remain a separate official signal; they are never relabeled as Sand Rating.
What “machine learning” means here
Transparent online learning, not marketing fog.
The adaptive ensemble is an online machine-learning model: after each result, it shifts weight toward component models with lower prior log loss. The current lab does not use reinforcement learning, and Duna will not claim that it does unless an actual RL policy is trained, evaluated, and documented.
AI agents may propose new features or challenger models. They cannot silently change the live rating. Every proposal must run through the same chronological evaluation, data-integrity review, versioned configuration, and human promotion gate.
How to read the scorecard
- Accuracy
- Share of winners picked above 50%. Exact 50/50 forecasts count as half a correct pick instead of favoring either team label.
- Brier score
- Mean squared probability error. Lower is better; 0.25 is the 50/50 baseline.
- Log loss
- Penalizes confident wrong predictions more severely. Lower is better.
- Calibration error
- Gap between forecast probability and observed frequency. Lower is better.
- AUC
- How often the model ranks a winner above a loser across probability pairs. Higher is better.