How Our Tennis Tournament Simulations Work — 20,000 Brackets at a Time

"Who wins tonight?" and "who wins the tournament?" sound like the same kind of question. They are not. The first is about one match between two known players. The second is about a maze: seven rounds, over a hundred players, and matchups nobody can name yet because they depend on results that haven't happened.

This article explains — in plain language — how our tournament simulations answer the second question, and why that requires a completely different machine than our daily match predictions.

Two Different Questions, Two Different Machines

Our daily tennis predictions are built by neural networks that study each scheduled match in detail: recent form, serving patterns, surface history, and hundreds of other signals about the two specific players on court that day. That depth is exactly what you want for "who wins tonight?" — and exactly what you can't have for a quarter-final that may or may not happen next Friday, between two players to be determined.

A tournament forecast has to price every possible matchup, including hypothetical ones. For that, we use our player strength ratings — the same Elo, Glicko-2 and surface-specific form systems behind our power rankings, rebuilt from tens of thousands of ATP matches. Ratings have a beautiful property: any two players, at any moment, have a defined win probability against each other. That's the key that unlocks the whole bracket.

Then We Play the Tournament. 20,000 Times.

Here's the whole trick, honestly stated:

1. Lock in reality. Matches already played are facts — their winners advance, always.
2. Flip weighted coins. Every match still to be played gets decided by chance, weighted by the players' ratings. If a player has a 70% chance to win, he wins 70% of the time — and loses 30% of it, because upsets are part of tennis, not a bug in it.
3. Crown a champion, write it down, reset, repeat. 20,000 times.

When the dust settles, the statistics are just counting. Won 6,100 of 20,000 simulated tournaments? Your title chance is 30.5%. Reached the semifinal in 13,400 of them? That's a 67% chance of making the last four. The same 20,000 brackets answer every question at once — title odds, the chance of reaching any round, even things like "how likely is this final?"

Why Not Just Pick the Best Player?

Because the draw is half the story. Imagine two stars of identical strength. One landed in a section full of qualifiers and a top seed who's been injured; the other has to beat three in-form seeds just to reach the semifinal. Same ability, very different title chances — and the simulation sees it automatically, because it walks both of their actual paths thousands of times.

This is also why a player's odds can rise without him hitting a ball: when a dangerous rival is upset on the other side of the draw, the simulations immediately reflect the easier road ahead. Forecasting a champion isn't only about how good you are — it's about who's left standing between you and the trophy.

What "30%" Actually Means

A 30% title chance means: run this tournament from today's situation many times, and that player lifts the trophy in about 3 of every 10. It does not mean "he won't win" — and it doesn't mean we were wrong if he doesn't. Shock champions happen, and an honest forecast gives them honest (small) numbers in advance, rather than pretending the favourite is a sure thing.

What makes a probability trustworthy is calibration: when we say 40%, it should happen about 40% of the time. We tested the engine on three full seasons of historical ATP tournaments — re-forecasting each one from its opening day, with the model only knowing what was knowable then — and the round-by-round probabilities matched reality within about one percentage point across the board. You can inspect how we hold ourselves to that standard on the model transparency page.

Our Track Record, In Numbers

The same historical test, in numbers — 30 ATP tournaments (2023–2025), every one re-forecast from its opening day with a model trained only on data through 2022:

For context: in a 128-player draw, naming the champion at random succeeds 0.8% of the time, always picking the #1 seed succeeds 21% of the time, and always picking the highest-rated player 30%. The simulation's edge over simply backing the favourite comes from pricing the draw. The live track-record section always shows the latest audited numbers — they update whenever the engine is retrained and re-tested, never by hand.

Forecasts That Sharpen Round by Round

A tournament forecast is alive. After every real match, the bracket changes: winners advance, losers' chances go to zero, and everyone else's path gets a little clearer. On every prediction run we lock in the new results, refresh the ratings, and re-run all 20,000 simulations — so the numbers you see during the quarter-finals are conditioned on everything that's already happened, not on last Monday's guesses.

The Fine Print (Kept Short)

• Qualifiers and unknowns get sensible neutral ratings until their identity is known — then they become themselves.
• Withdrawals and walkovers enter the forecast as soon as they're reflected in the official draw.
• Randomness is controlled. 20,000 runs is enough that the numbers are stable to a few tenths of a percentage point; day-to-day changes you see are new information, not noise.

That's the system: ratings that always have an answer, a bracket that gets played out 20,000 times, probabilities you can hold us accountable for — and a forecast that learns something from every completed match. The rest is tennis.