Tennis findings, re-tested on 2023-2026 point data

Published 2026-10-28 · data from from 1 January 2023 to 20 October 2026

Four results from tennis research, re-run on 155,144 professional matches from 2023 to 2026, with the code and the numbers free to reuse. The original studies drew mostly on Wimbledon. This record covers the ATP, WTA, Challenger and ITF tours, men and women, on every surface.

Download retest.json (CC BY 4.0) · the analysis code (Python, standard library only) · rerun it on the public sample

The four claims and what we found

The record: 155,144 completed best-of-three singles matches on the ATP, WTA, Challenger and ITF tours, played from 1 January 2023 to 20 October 2026. In 152,748 of them the games that show a server agree on who served first, and the serve-based tests use those. The point tests read 17,177,038 service points: those before deuce, in games recorded with no visible gap. How we counted.

Are points independent and identically distributed?

The claim. Points are neither independent nor identically distributed. Winning the previous point has a positive effect on winning the current point. At important points it is more difficult for the server to win the point. The deviations are small, so the iid model is still a good approximation in many cases.

Franc J. G. M. Klaassen and Jan R. Magnus (2001). Are Points in Tennis Independent and Identically Distributed? Evidence From a Dynamic Binary Panel Data Model. Journal of the American Statistical Association 96(454), 500-509. doi:10.1198/016214501753168217 Data: Wimbledon 1992-1995, almost 90,000 points.

How we tested it. Inside each match we take the server's chance on a point after winning the previous point of the game, and after losing it. We also compare break points (0-40, 15-40, 30-40) with other points. Both are set against the same games with their points shuffled.

What we found. After a won point, the server's chance on the next one is +0.11 points against chance. That is over 13,984,680 points, with an interval of +0.07 to +0.15. On break points it is +0.06 points, an interval that includes zero.

Where we differ, and how far to trust it. The previous-point effect is there across all tours, in the direction Klaassen and Magnus found, but it is about a tenth of a point. It comes from the ITF tours (+0.29 men, +0.25 women), and on grass, the surface of their study, it reverses (-0.26). Break points show no clear effect across all tours, and single tours go both ways: -0.33 on the WTA, +0.17 for ITF men. Our test compares points only inside the same game and stops at deuce, so their deuce and advantage points are not covered. A previous-point effect alone also moves the break-point figure a little.

Is there an advantage in serving first in a set?

The claim. In the men's singles, the first server won 54.7% of first sets. The authors named an advantage in serving first in the first set as the one result that held up. The women's data showed no clear edge. In later sets the first server won less than half the time. The authors put that down to who serves first. The set winner usually closes the set on their own serve, so the weaker player tends to serve first in the next set.

Jan R. Magnus and Franc J. G. M. Klaassen (1999). On the Advantage of Serving First in a Tennis Set: Four Years at Wimbledon. Journal of the Royal Statistical Society: Series D (The Statistician) 48(2), 247-256. doi:10.1111/1467-9884.00186 Data: Wimbledon 1992-1995, almost 90,000 points. Jan R. Magnus and Franc J. G. M. Klaassen (1996). Testing some common tennis hypotheses: Four years at Wimbledon. CentER Discussion Paper 1996-73, Tilburg University (hypothesis 2, the figures quoted here). full text (PDF, University of Amsterdam) Data: 481 Wimbledon singles matches, 1992-1995.

How we tested it. For every set we know who served first and who won it. We count how often the first server wins each set. For sets 2 and 3 we also follow the previous set's winner, as the paper did. Do they win more often when serving first or when receiving first?

What we found. The first server won 50.0% of first sets, so there is no first-set edge here. In set 2 the first server won 47.0%, below half, as the paper found. Its explanation holds too: the previous set's winner served first in only 39.4% of later sets.

Where we differ, and how far to trust it. The 54.7% first-set figure for men does not appear on these tours. On the paper's own comparison, the set 1 winner wins set 2 1.8 points more often when serving first. That is 69.2% against 67.4%, and in set 3 the gap is +0.7, which could be zero. That comparison does not fully take strength out. Who serves first in set 2 depends on how set 1 ended, and that is tied to how strong each player is. So the gap does not show that serving first helps.

After breaking serve, are you more likely to be broken back?

The claim. Hypothesis 12 of seventeen commonly heard statements: after breaking your opponent's service there is an increased chance that you will lose your own service. The authors found the opposite. After a break the hold rate rose from 81.8% to 83.3% in the men's singles. In the women's singles it rose from 64.3% to 68.3%.

Jan R. Magnus and Franc J. G. M. Klaassen (1996). Testing some common tennis hypotheses: Four years at Wimbledon. CentER Discussion Paper 1996-73, Tilburg University. full text (PDF, University of Amsterdam) Data: 481 Wimbledon singles matches, 1992-1995.

How we tested it. We take every service game played straight after the server broke serve in the same set. First the paper's way: its hold rate against all service games. Then inside each match: hold after breaking against hold after the opponent held, set against shuffled matches.

What we found. After breaking, players held 69.1% of the time, against 68.2% for all their service games (902,958 games after a break). Inside each match, against shuffled matches, holding after a break differs by -0.1 points, from -0.2 to 0.0.

Where we differ, and how far to trust it. On the paper's own comparison there is no extra break-back: the hold rate after a break is 0.9 points higher, not lower. On grass it is -0.6 points. Compare each player with themselves inside a match and the rise is gone. So we cannot confirm the paper's reading that breaking serve builds confidence. The rise reflects who breaks serve more than what breaking does.

Do streaks predict the next point or game?

The claim. Momentum is real: recent success raises the chance of winning the next point. The two 2024-2025 papers model the points of the 2023 Wimbledon men's singles and report statistical evidence for momentum. Jackson and Mosurski asked the older form of the question about whole sets: are heavy defeats a sign of psychological momentum or a random effect?

• Jingya Wang, Sihang Guo and Yuanyun Zhou (2024). A multidimensional momentum chain model for tennis matches based on difference equations. PLOS ONE 19(12), e0316542. doi:10.1371/journal.pone.0316542 Data: Wimbledon 2023 men's singles, 7,284 points.
• Chang Du, Caiya Zhang and Likai Zhou (2025). A novel methodological framework for analyzing the momentum effect in tennis singles. arXiv preprint 2509.01243. arXiv:2509.01243 Data: Wimbledon 2023 men's singles.
• David Jackson and Krzysztof Mosurski (1997). Heavy Defeats in Tennis: Psychological Momentum or Random Effect?. CHANCE 10(2), 27-34. doi:10.1080/09332480.1997.10542019 Data: match results by set (see the paper).

How we tested it. Two kinds of streak. Points: did the server win the last two or three points of the game? Games: did a player win the last two or three games of the set, and do they then serve or receive? Each streak is compared with no streak inside the same match and set against independent replays or shuffles.

What we found. Inside a game, after two won points the server's chance on the next is +0.29 points against chance, and after three +0.65. After three won games in a set, a player wins the next 0.74 points more often when serving and 0.25 more when receiving. After two won games the serving figure is +0.08, which could be zero.

Where we differ, and how far to trust it. Both papers fitted their momentum models to points from the 2023 Wimbledon men's singles, with shot-level detail this record does not have. Wang, Guo and Zhou also checked their model on other Grand Slam matches. We test only the core claim, that recent success raises the chance of winning what comes next. It does, by under a point. Jackson and Mosurski's question about whole sets is not tested here.

How we counted

Every match is a completed best-of-three singles match on the ATP, WTA, Challenger or ITF tours. Retirements and walkovers are left out, as on the open tennis data page.

We read each match's point-by-point record forward and keep only what it proves. A row that is not a legal score, contradicts a finished set, or is walked back by the next rows is skipped. Nothing is repaired or filled in.

Service alternates every game, a tiebreak counting as one. So the first server fixes them all. A match is used for serve-based tests only when at least three games show a server and 90% of those agree on who served first.

The point tests use games whose recorded points run from 0-0 to the end with no visible gap. Within them they count only the points before the score first reached deuce. A missing point just before the end of a game cannot always be seen. Tiebreak points are out, because the server changes inside a tiebreak.

A raw comparison misleads in two directions. Pool every server together and strong servers make "won the last point" look predictive when it only reflects strength. Compare inside one match and a run of wins uses up the wins that match had, which pulls the other way.

So every effect on this page is the comparison inside each match, minus the same comparison where nothing depends on what came before. For points, we shuffle the points inside each game, which keeps every game's length and score. For games, we shuffle each player's holds and breaks within each set, keeping only orders under which the set still ends where it did. On simulated matches with no momentum built in, both read as zero within their intervals.

The point shuffle matters for one reason. A game only counts when every one of its points is in the record, and long games miss a row more often. That selection alone makes independent points look streaky, and shuffling inside the kept games cancels it.

Intervals are 95% intervals that count each match as one unit. A row marked thin has an interval wider than 2 points either side, or fewer than 500 cases. Read a thin row as a direction at most.

Rerun it yourself

The code runs on Python 3.10 or newer with nothing to install. Point it at version 1.0.2 of the public sample of our point-by-point dataset on Zenodo. It prints the left column of the table below, to the decimal.

The sample is 4,556 matches from June 2026. In 2,238 of them the games that show a server agree on who served first, so the serve-based tests rest on those. It is small, so most of its tour and surface rows are thin. The full record is free for non-commercial research on application.

The two do not agree on the game tests. On the June sample, holding serve after a break is -2.5 points against chance, and after two won games -1.7. On the full record both are close to zero. We have not found why, so the verdicts above rest on the full record only.

The sample is on Zenodo: doi:10.5281/zenodo.22048730. The commands below fetch version 1.0.2, the one these figures come from.

curl -sO https://blog.livetennisapi.com/tennis-findings-retested/retest_findings.py
curl -sL -o points.csv.gz https://zenodo.org/api/records/23018680/files/points_sample_2026-06.csv.gz/content
curl -sL -o matches.csv.gz https://zenodo.org/api/records/23018680/files/matches.csv.gz/content
python3 retest_findings.py run --points points.csv.gz --matches matches.csv.gz \
    --point-reps 100 --game-reps 200 > sample.json

Data, licence and citation

retest.json holds every figure on this page and more: both runs, every tour and surface, the null values and the counts behind them. It is CC BY 4.0, so you can use it for anything, commercial work included, if you credit Live Tennis API.

The code is MIT licensed. The match-level record behind the full run is not public. Researchers can apply for it free on the academic dataset page.

curl -s https://blog.livetennisapi.com/tennis-findings-retested/retest.json | jq '.studies[] | {id, verdict}'

Citation: Live Tennis API (2026). Tennis findings, re-tested. CC BY 4.0. https://blog.livetennisapi.com/tennis-findings-retested

Related: the open tennis data page (comeback, hold and tiebreak rates, CC BY 4.0) and the point-by-point dataset, free for non-commercial academic research.

FAQ

Is momentum real in tennis?

Yes, but it is small. Inside a game, a server who has won the last three points wins the next 0.65 points more often than chance. After three won games in a set, a player holds serve 0.74 points more often. Counted on 155,144 ATP, WTA, Challenger and ITF matches.

Are points in tennis independent and identically distributed?

Very nearly. After a won point the server's chance on the next one rises by 0.11 points, a real but tiny effect. On grass it reverses. On break points there is no clear change across all tours (+0.06). Klaassen and Magnus (2001) found small effects in both, using Wimbledon 1992-1995.

Does serving first in a set help?

Not in the first set: the first server won 50.0% of 152,575 first sets. In set 2 the first server wins less often, because the player who lost set 1 usually serves first. We found no clean evidence that serving first helps in any set.

How often does a player break back straight after being broken?

A player who has just broken serve holds the next game 69.1% of the time, against 68.2% for all service games. So the break-back rate is 69.1% subtracted from 100, slightly below the usual break rate. Compared inside each match, the difference is about zero.

Can I check these numbers myself?

Yes. Download retest_findings.py and the public sample from Zenodo (DOI 10.5281/zenodo.22048730) and run one command. The sample gives its own figures, which the page lists next to the full record's.