Across 218,890 player-matches, next-day fatigue is a smooth gradient in the total games of the previous win, from a +3.4 point bonus after an easy win of 16 games or fewer to a -4.4 point penalty after 31 or more, on one day of rest. Our flat -4.5 point three-set rule was wrong in both directions.
Everyone in tennis knows the marathon story. A player survives an epic, and in the next round they are a shell. It is one of the few fatigue claims that comes with a real number attached, a collapse of roughly 14 percentage points in win probability, and it is repeated in previews and podcasts every Grand Slam.
We repeated a version of it ourselves. Last month we published a study of 209,000 matches which found that a three-set win the day before costs about 4.5 percentage points, and we wrote that figure into our betting framework as a flat haircut on the tired side. We have now re-measured the question on 218,890 player-matches, scoring the outcome against our ELO expectation instead of against a raw win rate. Two things came back. The cliff is not there. And our own flat rule was wrong in both directions.
The 14 point figure has a specific origin. In 2012 Jeff Sackmann counted the 146 Grand Slam matches since 2001 whose deciding fifth set reached 6-6, and found that the winners went 43-103 in the next round, a 29.5% win rate [1]. Against the post's own rating-based forecast of 43.4% for that same group, the shortfall of roughly 14 points is where the number comes from. It is a real, carefully framed result, and it is the source almost every later repetition traces back to.
What the folklore dropped is that the same author revisited it five years later with a continuous measure. Bucketing by time on court in the previous match, the ratio of actual to expected wins declines steadily from 1.07 for matches under 90 minutes to 0.76 for matches of 4:30 or longer, with no threshold anywhere in the sequence [2]. That is a gradient, not a cliff, published by the person who found the cliff. The headline survived in the culture and the correction did not.
The sports science points the same way. Gescheit and colleagues put trained players through four consecutive days of four-hour matches and measured progressive declines in sprint and jump capacity alongside elevated muscle damage markers, with serve velocity preserved and accuracy degraded [3]. Reid and Duffield's review of fatigue in match-play tennis describes pronounced and prolonged physiological and neuromuscular perturbations that build within and between matches and worsen on consecutive days of play, and nowhere does it identify a discontinuity at a particular match length [4]. Nothing in the physiology predicts that a set reaching 6-6 is categorically different from one ending 7-5.
The unit of observation is one player in one match, where that player won their immediately previous match between 1 and 14 days earlier. Beyond 14 days it is a layoff, which is a different study. The outcome is a residual, not a win rate: for every observation we take the pre-match ELO expectation, expected = 1 / (1 + 10^(−(elo_player − elo_opponent) / 400)), and subtract it from the result, so residual = (1 if the player won, else 0) − expected. Reported numbers are the mean residual in percentage points.
This matters because a raw win rate rediscovers the fact that better players win more. The residual asks the only question worth asking, which is whether the load signal moves the outcome beyond what the rating already knew. ELO-based methods sit among the most accurate published tennis forecasts in head-to-head comparisons [5], so the rating's expectation is a fair bar for a signal to have to clear rather than a nuisance term to sweep out.
Ratings are leak-free: our stored ELO row is the post-match rating, so the pre-match value is the stored rating minus the change that match applied. The basis is the all-surface rating, following our August 11 recalibration. One baseline is needed to read the tables: a player who just won carries an ELO bump from that win, so "just won something" is already a mildly negative pool before any fatigue exists. The baseline residual for any previous win is −0.13pp, close enough to zero that every cell below reads directly.
Sample: 218,890 player-matches across 125,432 matches, ATP and WTA, main tour through Challenger and ITF. Retirements are removed structurally by requiring that the previous match's winner actually closed out two sets (best-of-three) or three (best-of-five). Standard errors are clustered by match, because the two observations a single match contributes are exactly anti-correlated.
| Games in the previous win | 1 day rest | 2 days | 3+ days |
|---|---|---|---|
| 16 or fewer | +3.42 (n=31,798) | +2.13 | +2.33 |
| 17-19 | +1.11 | +0.35 | −0.46 |
| 20-22 | −0.48 | +0.24 | −1.38 |
| 23-26 | −2.12 | −0.93 | −2.52 |
| 27-30 | −2.59 | −1.22 | −3.01 |
| 31 or more | −4.36 (n=15,445) | −2.36 | +1.46 |
All figures are percentage points of win probability against ELO expectation. The bottom right cell (+1.46) is not statistically significant and should be read as "no measurable deficit left", not as a rebound. Read down the first column and there is no step anywhere. The effect grows one bucket at a time from a genuine bonus at the top to a genuine penalty at the bottom, exactly the shape the time-on-court study found [2] and exactly the shape the physiology implies [3]. The right way to read this table is as a replication of that 2017 result at scale, on a database roughly twenty times the size and reaching down to Challenger and ITF tennis, plus two things it did not measure: the rest interaction and the easy-win bonus.
Total games also separates better than the binary flag we had been using: the spread from the lightest to the heaviest bucket is 6.7pp, against 3.4pp for the two-set versus three-set split. In a regression on the residual, adding the games gradient and the rest interaction to the flat three-set rule is worth chi-square(5) = 249.7, p < 0.0001.
If marathons were categorically different, the flag for a deciding set that reached 6-6 should carry information that a plain game count does not. It does not.
| Previous win, deciding set | Residual |
|---|---|
| Reached 6-6 (tiebreak or extended set) | −3.60pp |
| Ended 12 games (7-5) | −3.68pp |
Those are the same number. A player who won 7-6 in the third is in identical shape the next day to a player who won 7-5 in the third. Once continuous games are in the model, adding the 6-6 flag on top is worth chi-square(1) = 1.72, p = 0.19, which is nothing.
We want to be fair to the original result rather than score a point off it. Our database is roughly 99% best-of-three; there are 573 ATP five-setters in 318,000 matches, and the best-of-five fifth-set 6-6 cell has n=92, which is far too thin to report. We therefore cannot test the exact population the 2012 study measured [1]. What we can say is precise and limited: in the population we can test, which is essentially all best-of-three professional tennis, the deciding-set 6-6 flag is a slice of a smooth games gradient and adds no independent effect. A 14 point discontinuity does not appear anywhere in 218,890 observations.
The games slope is not constant. It decays as rest accumulates:
| Rest before the next match | Fatigue slope |
|---|---|
| 1 day | −0.411 pp per extra game |
| 2 days | −0.226 pp per extra game |
| 3 or more days | −0.118 pp per extra game (not significant) |
Rest moderating load is chi-square(2) = 30.9, p < 0.0001. This is the single most useful line in the study, because games are not a clean physical measure. A long win also says the player only just got past that opponent, so game count doubles as a form signal. A pure form signal would not care how many days pass before the next match. This one halves after one extra day and disappears after two, which is what a physical recovery channel looks like and not what a quality signal looks like.
The sharpest specification is not the absolute load at all. It is the load differential against the opponent's own previous match: −0.523pp per game of differential, with 2.4 times the R-squared of the absolute version. That makes sense. Tennis is a two-sided contest, and a 30-game win matters less if the other player also played 30. The tours also differ. ATP recovery is complete by three days, where the slope goes flat, while WTA keeps a −0.364 slope even at three or more days of rest. We do not have a mechanism for that and we are not going to invent one, but the split is large enough that we treat the two tours separately.
Our previous published rule was a flat −4.5pp for a three-set win the day before. It failed twice over. First, we over-penalised the ordinary case. The honest three-set versus two-set gap on this sample is −3.36pp, not −4.5pp. The old estimate came from a binary straight-sets versus three-sets split among day-after winners, opponent-adjusted and averaged per surface, on a smaller sample; measured as a continuous ELO residual with errors clustered by match, the same contrast comes out smaller. Every median three-setter we haircut, we haircut about a third too hard.
Second, and worse, we scored the largest single cell in the study as zero. An easy win of 16 games or fewer the day before is +3.42pp, a bonus, on n=31,798 observations. A flat penalty rule has no way to express that. We treated fresh, dominant winners as neutral when the data says they are the strongest cell on the board, and that one cell holds more observations than anything else in the study.
One artifact runs against us here rather than for us. Our rating update already applies a margin-of-victory multiplier of up to 1.5x on a blowout, so the quick-win group enters its next match on a rating that was inflated harder than the grinder's. It beats that rating anyway. The gradient survives an adjustment already working against it.
The flat −4.5pp three-set haircut is deleted from our betting framework. In its place is a continuous games-by-rest term, applied to the load differential against the opponent's previous match wherever both are known, with the WTA slope kept at all rest levels. The tiebreak and marathon special case is deleted outright, because it was measuring a slice of the gradient twice. We are also leaving the correction where people can see it: the earlier article stays up with a pointer to this one. Publishing a number and then quietly replacing it is how a research page turns into marketing. Every pick our model makes is published before the match and graded in public, and correcting our own inputs in public is the same policy applied to ourselves.
Check the gradient against closing prices, which is the only test that separates "our model was wrong" from "the market is wrong". Split the games slope by age, where recovery should differ and where our framework currently applies a qualitative adjustment with no measured number behind it. Investigate the WTA recovery gap directly instead of encoding it as a tour dummy.
Method: 218,890 player-matches across 125,432 ATP and WTA singles matches (main tour, Challenger and ITF), where the player won their previous match 1 to 14 days earlier. Outcome is the residual against the pre-match ELO expectation (stored post-match rating minus that match's change; all-surface basis). Load features parsed from set scores: total games, sets played, deciding set reached, deciding-set games, deciding set reached 6-6. Rest is calendar days between matches, capped at 14. Both players of an index match can qualify; when both sit in the same cell their residuals cancel, which is correct behaviour for equal load but reduces effective power below the raw n. Model comparison is OLS on the residual with CR0 cluster-robust covariance by match. Reproducible via the study command with --tour, --since and --json flags; database snapshot of 318,000 matches, run 2026-08-12.
References:
[1] J. Sackmann, "The hangover effect of a marathon fifth set," Heavy Topspin (Tennis Abstract blog), Aug. 9, 2012. http://www.tennisabstract.com/blog/2012/08/09/the-hangover-effect-of-a-marathon-fifth-set/ (accessed Aug. 12, 2026).
[2] J. Sackmann, "The negative impact of time of court," Heavy Topspin (Tennis Abstract blog), Jun. 2, 2017. https://www.tennisabstract.com/blog/2017/06/02/the-negative-impact-of-time-of-court/ (accessed Aug. 12, 2026).
[3] D. T. Gescheit, S. J. Cormack, M. Reid, and R. Duffield, "Consecutive days of prolonged tennis match play: performance, physical, and perceptual responses in trained players," International Journal of Sports Physiology and Performance, vol. 10, no. 7, pp. 913–920, 2015, doi: 10.1123/ijspp.2014-0329.
[4] M. Reid and R. Duffield, "The development of fatigue during match-play tennis," British Journal of Sports Medicine, vol. 48, suppl. 1, pp. i7–i11, 2014, doi: 10.1136/bjsports-2013-093196.
[5] S. A. Kovalchik, "Searching for the GOAT of tennis win prediction," Journal of Quantitative Analysis in Sports, vol. 12, no. 3, pp. 127–138, 2016, doi: 10.1515/jqas-2015-0059.
There is a hangover, but it is a smooth gradient, not a cliff. The more games a player needed to win their previous match, the worse they perform against rating expectation, from +3.42 percentage points after a win of 16 games or fewer to -4.36 after 31 or more on one day of rest. A deciding set that reached 6-6 (-3.60pp) costs the same as one that ended 7-5 (-3.68pp), so the marathon flag adds nothing once total games are counted.
The honest three-set versus two-set gap on 218,890 player-matches is -3.36 percentage points against ELO expectation, not the flat -4.5 we published earlier. And an easy win is not neutral: winning in 16 games or fewer the day before comes with a +3.4 point bonus, the largest single cell in the study (n=31,798), which a flat penalty rule scores as zero.
The rest decay is the evidence it is physical: the slope is -0.411 percentage points per extra game at one day of rest, -0.226 at two days, and a non-significant -0.118 at three or more. A pure form signal would not care how many days pass before the next match; this one roughly halves per extra rest day, which looks like recovery. It is evidence rather than proof, because our data has no match-duration column and total games also encodes how comfortably the player won.
See today's picks — published before the match, graded in public →