Who wins this match, and why?
Pick any two players or pairs from the BWF World Tour. The audited FYP champion model gives the win probability, breaks it into the factors behind it, and updates it once game 1 is in.
Fixture
Why the model leans this way
Each bar is a factor's exact share of the prediction, in log-odds. Bars to the left favour side A, bars to the right favour side B.
Live: after game 1
Enter the game 1 score. A small model fitted on 2025 and tested on all 3,822 matches of 2026 updates the forecast.
Match context
Two effects the model cannot see, measured on 10,852 out-of-sample matches. Tick one to apply its adjustment.
How much does the draw decide?
A 16-entry knockout, computed exactly from every pairwise win probability rather than by random sampling. Start from a seeded draw of the current top 16, then shuffle it to see how much one draw moves each player's title odds.
Title odds
Draw luck
Spread of each top-8 entrant's title odds across 300 random draws. A wide bar means the draw matters more than form.
Bracket
The most likely occupant of each slot, and their chance of getting there.
Rating, form and who beats the model
Every player's Glicko-2 rating history from 2021, current form, and 2026 wins above expectation: actual wins minus the wins the model expected before each match.
Player
Power rankings
Active players (15+ matches, played since May 2026) ranked by Glicko-2 rating. The ± column is the rating deviation: smaller means the model is surer.
Six findings a coach can use
Every number below comes from out-of-sample predictions: the 2025 and 2026 seasons, which the model never trained on. Intervals are 95% bootstrap intervals that resample whole tournaments.
Biggest upsets of 2026
Winners the model gave the lowest chance before the match.
Beating the model in 2026
Wins above expectation, minimum 12 matches. Sustained over-performance is a sign the rating has not caught up yet: a player on the rise.
Can you trust these numbers?
The model was frozen before it saw 2025 or 2026. This page shows how it performed on both seasons, where it is weaker, and a live check that this browser reproduces the served model exactly.
Audit on unseen seasons
ROC-AUC: the chance the model ranks a random winner above a random loser. 0.5 is a coin flip, 1.0 is perfect. Interval: tournament-cluster bootstrap from the FYP audit.
Calibration
When the model says 70%, does that side win about 70% of the time? Points on the diagonal mean yes.
Where it is strong and weak
ROC-AUC by discipline, both audit seasons pooled.
Live reproduction check
How it works, and its limits
Model. Logistic regression on 20 pre-match features: Elo and Glicko-2 rating gaps, recent form over 5 and 10 matches, experience, career record, rest days and head-to-head. Trained on 17,193 matches from 2022–2024.
Order symmetry. The model scores each fixture both ways round and averages in log-odds, so P(A beats B) = 1 − P(B beats A) exactly. Context that is the same for both sides (tournament tier, round) cancels out, which is why the model is a short weighted sum that this page can explain factor by factor.
No leakage. Every match on a given date is rated only from matches before that date.
- Ratings are frozen at 9 August 2026, the end of the scraped data. Results after that date are not reflected.
- Men's singles is the hardest discipline to call (ROC-AUC about 0.73). Women's doubles is the easiest (about 0.86).
- The 0.80 ROC-AUC target was met in 2025 (0.808) but not in 2026 (0.798). Both intervals contain 0.80.
- Probabilities are uncalibrated, and 2026 calibration is weaker than 2025 (ECE 0.038 against 0.014).
- The home and fatigue adjustments are measured effects. The combined adjustment did not clearly beat a simple recalibration on 2026, so treat them as context.
- A doubles pair the data has never seen together starts from a default partnership rating.