Tennis Tournament Simulations
For every live ATP tournament we simulate the entire remaining bracket 20,000 times on our Elo & Glicko-2 player ratings. Completed matches are locked in, every match still to play is sampled from its win probability — what comes out is each player's chance of reaching every round and lifting the trophy.
's-Hertogenbosch 2026
Eliminated (7)
Stuttgart 2026
Eliminated (12)
Track record — tested against history, no hindsight
Before going live, the engine was made to re-forecast 30 historical ATP tournaments (2023–2025) from each one's opening day — with the model trained only on data through 2022, so it could never peek at the answers. These are the results, straight from that test.
pre-tournament (top pick)
top 5 picks
at Grand Slams
picking the #1 seed
By tournament tier
| Tier | n | Top-1 | Top-5 |
|---|---|---|---|
| Grand Slams | 12 | 42% | 100% |
| Masters 1000 | 10 | 30% | 50% |
| ATP 500 | 5 | 20% | 40% |
| ATP 250 | 3 | 33% | 33% |
Do the probabilities mean what they say?
| When we predicted… | It happened | n |
|---|---|---|
| 0–1% (avg 0.2%) | 0.1% | 6,650 |
| 1–2% (avg 1.5%) | 0.7% | 1,347 |
| 2–5% (avg 3.3%) | 3.0% | 1,985 |
| 5–10% (avg 7.2%) | 6.9% | 1,641 |
| 10–20% (avg 14.4%) | 13.8% | 1,721 |
| 20–30% (avg 24.5%) | 23.9% | 961 |
| 30–50% (avg 40.4%) | 42.0% | 1,597 |
| 50–70% (avg 58.8%) | 59.6% | 1,379 |
| 70–90% (avg 78.5%) | 79.4% | 528 |
| 90–100% (avg 98.7%) | 98.9% | 379 |
Every round-reach probability from the test, grouped into bands and compared with what actually happened — predicted and observed match within about a point in every band. That's what makes a 30% title chance a number you can act on, not a vibe.
Backtest generated Jun 09, 23:41 · 20,000 simulations per tournament · methodology in how the simulations work · more model audits on model transparency.
How the simulations work
Where do these probabilities come from?
Every player carries Elo, Glicko-2 and surface-specific strength ratings rebuilt from 39,000+ ATP matches. A calibrated model turns any rating gap into a match win probability, and we then play out the entire remaining bracket 20,000 times — sampling every not-yet-played match from its probability. A player's title chance is simply the share of those 20,000 tournaments he wins.
Why simulate instead of just picking the favourite?
Because the draw matters. Two equally strong players can have very different title chances if one has byes and qualifiers ahead while the other faces seeds from the second round. Simulation prices the whole path — including the matches your rivals might lose for you — which a single head-to-head number can't capture.
How accurate are these forecasts?
Backtested on 30 historical ATP tournaments (2023–2025): the pre-tournament favourite by our model won 33% of events and the champion came from our top 5 in 67% (100% at Grand Slams). More importantly, the probabilities are calibrated: when we say a player has a 40% chance of reaching a round, it happens about 40% of the time. See model transparency.
When do the forecasts update?
On every daily prediction run while a tournament is live: finished matches are locked in, ratings refresh, and the remaining bracket is re-simulated — so the odds sharpen round by round until the final.
1 · Rate
Every player carries continuously updated Elo, Glicko-2 and surface-specific strength ratings, rebuilt from 39,000+ ATP matches since 2015.
2 · Simulate
The remaining bracket is played out 20,000 times. Each hypothetical match is sampled from the rating model's win probability; rating uncertainty is resampled per simulation so volatile players get honest long shots.
3 · Verify
Backtested on 30 historical tournaments: round-reach probabilities are calibrated within a point of observed frequencies. See model transparency.