Corrections to our tennis data
This log lists 17 corrections to values we published, newest first. Three are in our point-by-point sample and 14 are in data our API served. They run from 22 August 2026 to 29 October 2026, and each one links the public record of the change. The same rows are a free CSV under CC BY 4.0 and an Atom feed.
Download the CSVAtom feedLicence: CC BY 4.0
Corrections, newest first
Each row gives the day of the record, the dataset or API data it touches and what changed. Before and After show the values the record states. The last column links the record itself. New corrections go at the top of this table and into the feed.
| Date | Asset | What changed | Before | After | Version |
|---|---|---|---|---|---|
| API / | Fixtures for 29 October showed as live before they started. A digit in the day-ahead schedule was read as a set in play, which marked 298 fixtures live. The stored rows were restated the same day. | live | Not in the record | Docs 1.13.77 | |
| API / | A fixture could disagree with its own match for hours. A fixture bound to a match kept its status and start time until the next 6-hourly refresh. It now takes both from the match when it is read. A fixture marked cancelled or postponed keeps that status. | Not in the record | Not in the record | Docs 1.13.77 | |
| Point-by-point sample | Version 1.0.2 shipped the 1.0.0 README by mistake. That brought back the wrong phrase fixed in 1.0.1, and version 1.0.3 restores the corrected wording. Its players file also drops one ID column. Every other column is unchanged. | recorded live since January 2023 | Live observation began 2025-10-12 | Sample 1.0.3 | |
| API / | Recent matches could show one set fewer than the match itself. Recent scores on a player's stats read the stored games raw, so a match decided in a match tiebreak lost its last set. They now show the games the match shows. | Not in the record | Not in the record | Docs 1.13.67 | |
| API / | The statistics endpoint and the player totals judged some finished matches differently. A deciding match tiebreak stored as points in one place and as one game in another got null or false from the statistics endpoint. Both now read it as one finished set. | Not in the record | Not in the record | Docs 1.13.67 | |
| Point-by-point sample | Timestamps labelled UTC were written in local time. In the points file, timestamp_utc moved back 3 hours on all 951,064 rows. In the match index, scheduled_time_utc moved back 3 hours on 120,758 rows and 2 hours on 52,813. Every other value is unchanged. The example in the record is match 9. | 2026-03-22 19:40:00 | 2026-03-22 17:40:00 | Sample 1.0.2 | |
| API / | The final score of some tiebreak sets was missing from the tape. The row that closes a tiebreak set carried no tiebreak_final when the set counter moved one row late. Of 9 live matches whose tiebreaks stated a final, 2 had it on no row. The games on each row now decide it. | Not in the record | Not in the record | Docs 1.13.48 | |
| API / | A completed match could lose its point tags and clocks. A reconstruction replaced the stored stream even when the stream was complete. The stream is now kept when it is complete, or tagged and legal end to end. Such a match now reads basis live instead of reconstruction. | Not in the record | Not in the record | Docs 1.13.35 | |
| API / | Some points were tagged as a double fault on a first serve. The serve and outcome tags contradict each other, and they were served as stated. Such a point now reads null for serve and outcome, and stored tags with the contradiction were cleared. | double fault on serve 1 | null | Docs 1.13.35 | |
| API / | Tapes could show impossible point sequences as fact. Some recorded games moved backwards, and the tape builder fitted them into a forward chain. A score such as 30-30 could then repeat three times in a row. In August, 547 of 586 ATP and WTA matches had at least one such game. Those games are now withheld whole and counted in games_withheld. | Not in the record | Not in the record | Docs 1.13.28 | |
| API match surface | Team event matches had no surface. All 268 Davis Cup, Billie Jean King Cup and Laver Cup matches in the 60 days to 20 September served null. Each match now takes the surface of its tie when exactly one tie fits. Matches from those 60 days and upcoming ones were backfilled, and ties with no announced venue stay null. Match 191867, for example, now reads hard. | null | hard | Docs 1.13.26 | |
| API match surface | Some ITF matches had no surface while their tournament had one. From 21 June to 4 September 2026, 2,533 matches served null. They now carry their own tournament's surface, and the 90-day share of matches with no surface fell from 19.8% to 11.8%. Match 186637, for example, read null and now reads hard. | null | hard | Docs 1.13.25 | |
| API / | Points in a game we joined late were numbered from 0. A 0 claims the opening state of a game, which was false there. Over three days, 1,341 rows carried 0 at a score other than love. The number now reads null there, and a late row from another game no longer resets the count. | 0 | null | Docs 1.13.18 | |
| API history packages | Some tape packages gave the wrong court surface. A rule based on tournament names had set 19,439 Challenger matches in the packages from February 2023 to December 2024 to grass. The packages were repaired and rebuilt. Each repaired package now lists its fixes in a corrections file. | grass | Not in the record | Docs 1.9.6 | |
| API / | Ranking rows could carry a retired player id. When player ids were merged, the weekly ranking tables kept the old one. A merge now repoints them, so ranking rows and match records carry the same player_id. | Not in the record | Not in the record | Docs 1.9.6 | |
| API prices | On roughly half of all pairings, price side 1 was the other player. Side followed the market's own display order instead of players.p1 and players.p2. Side 1 on those markets meant players.p2 and now means players.p1, and the fix applies to past ticks too. | players.p2 | players.p1 | Docs 1.9.4 | |
| Point-by-point sample | The 1.0.0 description said the data was recorded live since January 2023. The match index starts in 2023, but live point observation began on 12 October 2025. The data files did not change. | recorded live since January 2023 | Live point observation started 2025-10-12 | Sample 1.0.1 |
What counts as a correction?
A row goes in when a data value we published was wrong and we changed it. That can be a value in one of our dataset files or on our data pages, or a value our API served or stored. A null or missing value counts when the record calls it a fix or a repair. So does a wrong statement in a dataset's own description or README, because it ships with the data.
Where the record states no value before or after the change, the cell reads "Not in the record". These changes are not listed here:
- New data, new fields and counts we measure again and date, such as the figures on our about page.
- Files we rebuild from live data, such as the tournament map, which change as the data changes.
- Fixes to how the API reads a request, such as a filter it ignored or a time window it read wrongly. These are in the docs changelog, with fixes to the reference text, schemas and parameters.
- Corrections to a single match result or player profile, which the API flags on the match or the player. The section below explains how to find them.
Which datasets have had no correction?
We also read the history of every other free dataset we publish. None has had a correction so far.
- Line-calling dataset. Version 1.0.1 changed the metadata, the licence text and the README. No row changed between its releases.
- 2015 results sample. One data release, on 18 August 2026. The only later change was to the licence text and the README.
- October 2026 point-by-point sample. One release on Hugging Face, on 13 November 2026.
- Match outcome studies. Released on Hugging Face on 20 September 2026. The only later change updated links in the README.
- The /data pages on this site. We read their change history and found no published figure that we later corrected.
Where are corrections to single matches and players?
They are not listed here, because the API flags each one on the match or player itself. After a match is published as completed, its status, winner or final score can still change. Each such change adds a row with basis restatement to the match's status history. The match also carries result_restated_at and result_version.
To find completed matches whose result was corrected since a given time, call GET /history/matches with restated_since. These flags start on 22 September 2026. Earlier corrections were not flagged. The changelog entry for 1.13.40 describes them in full.
A correction to a published player profile, such as a birth date, sets profile_restated_at and raises profile_version on the player. GET /players with restated_since lists the players corrected since a given time.
A tape package repaired after it was published also carries a corrections file, with one row per changed field. The package manifest from GET /history/packages/{period} lists it with format corrections. The first such repair is in the table above.
How do I report a wrong value?
Write to hello@livetennisapi.com with the dataset or endpoint, the value and why you think it is wrong. If we correct it, we record the correction here, or on the match or player as the section above describes.
Can I reuse the log?
Yes. The CSV is under CC BY 4.0, so you can copy and change it, commercial use included. Credit Live Tennis API. A credit line that works is "Corrections log: Live Tennis API (livetennisapi.com/data/corrections), CC BY 4.0."
The CSV has one row per correction, with the columns id, date, asset, what_changed, before, after, version and version_url. An empty before or after means the record states no value. It is UTF-8 with a header row, newest first.
The Atom feed carries the same rows. Add it to a feed reader to see each new correction as it is published.
The datasets themselves are our point-by-point dataset for researchers and the line-calling dataset. Every change to the API, corrections included, is in the docs changelog.
Every record was read on 19 November 2026.