Open tennis data: comeback, hold and tiebreak rates
We counted 116,382 completed singles matches and 2,239,440 service games, and published what came out. It is free to use, cite, chart or argue with under CC BY 4.0 — no account, no API key, no request to us.
Download studies.json · served with an open CORS header, so a browser can fetch it directly.
What is in it
Aggregate findings, not per-match rows. Four studies, each carrying the corpus it was counted from and the date it was counted:
| Study | What it measures | Corpus |
|---|---|---|
| First-set comeback rate | How often the player who lost the first set wins, by first-set score, surface and tour | 116,382 matches |
| Second-set win probability | What the second-set scoreline does to the first-set loser's chances | 116,382 matches |
| Service hold rate | Hold percentage by surface and by tour, tiebreak games excluded | 2,239,440 service games |
| Tiebreak frequency | Share of sets and of matches reaching a tiebreak, by surface | 116,382 matches |
The finding most people do not expect
The margin of the first set predicts a comeback about three times more strongly than the surface does. A player who loses a first-set tiebreak wins the match 21.92% of the time; a player who loses the first set 6–0 wins 7.18%. That is a 14.74-point spread. Across surfaces the spread is 0.73 points (grass 17.20%, clay 16.79%, hard 16.47%), and across tours about three (ATP 18.21%, WTA 17.04%, ITF men 15.18%).
So “lost the first set” is a weak feature and “lost the first set 7–6” is a strong one — which is not how the match is usually discussed.
How it was counted
Completed best-of-three singles matches on the ATP, WTA, Challenger and ITF tours, 2023-01-01 to 2026-09-14, read from the point-by-point record rather than inferred from final scores. Retirements and walkovers are excluded. The figures are counted, not sampled and not modelled.
One check worth knowing about: the same corpus reproduces the widely published ATP (~79%) and WTA (~64%) service hold rates — 78.62% and 64.36% — on a method that was not tuned to match them.
What is wrong with it
Stated here rather than left for you to find:
- Two breakdowns do not sum to the corpus. The by-tour split falls 11 matches short and the by-first-set-score split 508. A re-measurement found no missing seventh tour and no missing eighth scoreline, so this is counting drift in the published rows, not an exclusion.
- The surface split covers 113,288 matches, not 116,382, because roughly 3,100 matches never had a surface stated by the feed. That one is a genuine exclusion.
- The rates are sound. Every tour matches exactly on re-measurement and no first-set rate moves by more than 0.02 percentage points. The conclusions rest on the rates, not on the row counts.
Using it
Take it. Chart it, check it, disagree with it in public. If you re-run any of it and get something different, we would like to know.
curl -s https://blog.livetennisapi.com/studies.json | jq '.studies[0].by_first_set_score'Citation: Live Tennis API (2026). Tennis Match Outcome Studies. CC BY 4.0. https://blog.livetennisapi.com/studies.json
These are aggregates. If you need the per-match records behind them, they are in the API's history endpoints — a free key covers live and upcoming matches, and history is on the paid tiers. The dataset above does not require either.
FAQ
Is the dataset really free to use?
Yes. It is published under CC BY 4.0, which permits any use including commercial use, provided you credit Live Tennis API. No account, no API key and no request to us is needed — the file is served with an open CORS header so a browser can read it directly.
What exactly is in the file?
Aggregate findings, not per-match rows: comeback rates after losing the first set broken down by first-set score, surface and tour; win probability by second-set score; service hold rate by surface and tour; and tiebreak frequency by surface. Every figure carries the corpus it was counted from and an as-of date.
How were the numbers produced?
By counting completed best-of-three singles matches on ATP, WTA, Challenger and ITF from the point-by-point record between 2023-01-01 and 2026-09-14 — 116,382 matches and 2,239,440 service games. Retirements and walkovers are excluded. The figures are counted rather than estimated or sampled.
Are there known problems with it?
Two. The by-tour and by-first-set-score breakdowns fall 11 and 508 matches short of the 116,382 corpus respectively; a re-measurement found no missing tour and no missing scoreline, so that is counting drift in the published rows rather than an exclusion. The surface breakdown covers 113,288 matches because roughly 3,100 matches never had a surface stated by the feed — that one is a genuine exclusion. Every rate reproduces on re-measurement to within 0.02 percentage points.
Can I use it in a paper, an article or a tutorial?
Yes, and we would like to know about it. Cite it as: Live Tennis API (2026), Tennis Match Outcome Studies, CC BY 4.0, https://blog.livetennisapi.com/studies.json. If you want the underlying per-match data rather than these aggregates, the API's history endpoints carry it.