Across 64,669 ATP and WTA event transitions from 2018 to 2026, a player who just won a title underperforms our pooled ELO by 3.5 percentage points in their next first match. The letdown story is still wrong: runners-up dip just as much, and the deficit vanishes on a pure surface rating book. The mechanism is rating overshoot.
Everyone in tennis knows the pattern. A player wins a title on Sunday, flies to the next city, and loses in the first round. The trophy came with a bill, and the bill arrives the following week.
We believed it too. Our betting framework carried a letdown haircut, applied by hand, on every match where one side had just come off a deep run, and it had never been measured on our own data. It was folklore that we had written down and then started charging ourselves for. So we measured it, on 64,669 event transitions. There is a real deficit after a deep run, and it has nothing to do with the trophy, the emotion, or the legs. It is our own rating system overshooting a hot week.
Streak folklore in sport has a long record of surviving right up until somebody counts. The canonical case is basketball's hot hand, where Gilovich, Vallone and Tversky found the belief was near universal among players, coaches and fans, and that the shot sequences did not support it [1]. The lesson that transferred was not "streaks never exist". It was that a vivid pattern needs a baseline before it becomes a finding. Tennis has its own version of that test: Klaassen and Magnus modelled whether points are independent and identically distributed and found real deviations, including a small winning-mood effect, but small enough that iid remains a working approximation for match models [2]. Measured tennis psychology tends to come back smaller than told tennis psychology.
On the letdown itself, the public record is mostly raw win rates, and raw win rates cannot settle the question. Champions in our own sample won 66.6% of their next-event openers, which sounds like the opposite of a letdown until you notice their ratings said they should have won 69.2%. A player who just went deep is a strong player, normally favoured in an early round, so "they usually win" and "they underperform" are both true at once. What has been missing is an expectation to grade the results against, and that is the gap this study fills. We use ELO as that expectation, which is defensible on two grounds: ELO methods rank among the most accurate published tennis forecasters in the one head-to-head comparison of the field [3], and the blended rating book we grade against is ordinary practice rather than a quirk of ours, since the best known public tennis ELO also mixed overall and surface ratings rather than choosing between them [4].
We took 64,669 event transitions from 2018 to 2026, ATP and WTA, main tour and Challenger. A transition is one player finishing one event and starting the next one within 120 days. For every transition we score the first match of the next event as observed minus expected, where expected is the standard ELO formula on pre-match ratings.
That subtraction is the whole point. A title run has already pushed the player's rating up, so "did they underperform their own rating" is exactly the claim a letdown story is making, and it is exactly what this measures. A residual of −3.5pp means the player won 3.5 percentage points less often than their own inflated rating said they would.
Two details keep the number honest. Ratings are strictly pre-match, reconstructed by subtracting each match's own rating change from the stored post-match value, so nothing dated after the match leaks in. And every cell is calibration-matched: big favourites and coin flips are not equally well predicted by ELO, and a champion's next match sits at a much higher expectation than an early loser's, so we net out the rating's own known bias at the same expectation before reporting, because otherwise you mistake the calibration slope for a letdown.
| Prior event | n | Residual vs ELO | z |
|---|---|---|---|
| Won the title | 1,922 | −3.5pp | −3.44 |
| Lost the final | 1,924 | −2.9pp | −2.73 |
| Lost the semi-final | 3,804 | −1.5pp | −2.00 |
| Lost the quarter-final | 7,412 | +0.1pp | +0.11 |
| Lost their opening match | 31,223 | +0.3pp | +0.77 |
A clean gradient, significant at the top, null from the quarter-finals down. If we had stopped here we would have published "the letdown is real, it is worth about three and a half points, and it starts at the semi-final". It would have been wrong.
The folklore is specifically about winning. Emotional release, celebration, the target on your back. So compare the two players who walked off the same court on the same Sunday.
| Same final, same week | n | Residual | z |
|---|---|---|---|
| Played five matches, won the title | 1,595 | −3.7pp | −3.33 |
| Played five matches, lost the final | 1,614 | −3.4pp | −2.95 |
Both groups played a 32-draw, both played five matches that week, both reached the same final. One lifted a trophy and one did not. The difference between them is 0.3pp at z=0.41, which is nothing. Whatever this effect is, it does not know who won.
We keep two rating books. The blended book, 0.7 overall plus 0.3 surface, is the one our model actually consumes, the measured optimum from our surface-ELO study. We also keep pure surface ratings. Same players, same matches, same results, two ways of pricing them. On the surface book, the post-title deficit is −0.8pp at z=−0.82 and the post-final-loss deficit is −1.1pp at z=−1.03, both measured nulls.
This is the finding that ends the psychological reading. A player's emotional state cannot depend on which of our two rating books we chose to grade them against. If the deficit lives in one book and not the other, it is a property of the book. The direction of the split says the same thing: champions whose next event was on the same surface came in at −2.6pp, while champions who switched surface came in at −7.1pp (n=362, z=−3.00). A hangover should not care what surface the next tournament is played on. A pooled rating should, because a clay title's rating gain flows straight into the number that prices a hard-court match the following week, where the player never earned it.
If the mechanism is a rating overshoot, then the size of the rating move should predict the deficit, and players whose rating moved the other way should overperform. Both are true. By how much overall ELO the prior event added:
| ELO the prior event added | n | Residual | z |
|---|---|---|---|
| Lost 20+ | 8,399 | +1.3pp | +2.33 |
| Flat (−20 to +10) | 34,118 | −0.2pp | −0.43 |
| Gained 10 to 40 | 14,354 | +0.2pp | +0.34 |
| Gained 40 to 70 | 5,004 | −0.7pp | −1.04 |
| Gained 70+ | 2,794 | −2.3pp | −2.58 |
That is a monotone line through zero. Nobody tells folk stories about the emotional lift of a bad week, but the data has one, and it is the mirror image of the letdown. This is regression to the mean in a rating, not a mood.
Inside the champions the same gradient holds. Titles that added 10 to 40 ELO gave back 1.1pp, 40 to 70 gave back 2.7pp, and 70 or more gave back 4.8pp (n=796, z=−3.01). The trophy is constant across those three rows. Only the rating move changes, and the deficit tracks the rating move. The cross-section by standing closes it: champions inside the top five by ELO are immune, at −0.7pp and z=−0.29, while champions ranked 21 to 50 are the worst standing band in the study at −8.2pp. That is our own K-factor schedule, which pays roughly double per win to a player with a light rated sample. An elite player's rating barely moves when they win a title, because the model already knew. A rank-40 player's rating jumps, and the jump is too big.
The fallback explanation is physical. Five matches in seven days, tired legs, quick turnaround. Then the worst cell should be the immediate next week. It is the opposite.
| Gap after winning a title | n | Residual | z |
|---|---|---|---|
| 0 to 7 days | 1,087 | −2.7pp | −1.99 |
| 8 to 14 days | 355 | +0.1pp | +0.03 |
| 15 to 21 days | 153 | −4.7pp | −1.37 |
| 22+ days | 327 | −8.8pp | −3.68 |
More than three weeks of rest is the worst spot on the board, at over three times the immediate turnaround. No fatigue story survives that shape, and our fatigue study already showed that schedule load alone barely moves a match. A stale rating does survive it, because a rating carries no time decay: sit out a month and the inflated number is still sitting there, waiting.
Persistence points the same way. After a title the deficit is −2.8pp at the next event and −1.7pp at the event after that, on the same 1,862 players. A hangover is gone by the second tournament. An overshot rating is only half corrected, because it takes results to pull it back down. The one effect that does look psychological runs in the opposite direction and is small: players who lost their opening match and played again within seven days beat their rating by 1.9pp on 13,685 rows at z=+4.16, and that bounce is gone by the second week (−0.5pp at 8 to 14 days).
The deficit is real in the blended book. If you price matches off a pooled ELO, as most public tennis models do [4], your number is too high on anyone who just went deep, by roughly three points at a final and more when the run was large or the player sits outside the top twenty. That is worth correcting.
The cause is not the player. It is the rating. Every version of the story we have all told, the emotional letdown, the champagne hangover, the tired legs, is wrong about the mechanism even where the number is right. If you shade a price because a player looks flat after a big week, you are getting the right answer for a reason that will fail you the moment you apply it somewhere your rating is not inflated.
And the honest limit: we have not shown the market fails to price this. Only our 2026 rows carry odds, which leaves 174 title transitions with a usable de-vigged price, and against the market those rows come in at +1.0pp, which at that sample size means nothing at all. This is a model calibration finding. It is not a demonstrated edge and we are not going to call it one.
We deleted the letdown haircut from the betting framework, because it was a psychological adjustment and the psychology is not there. It is replaced by a symmetric deep-run rating correction, stated as what it is: our pooled ELO overshoots a deep run. Roughly 3pp off anyone who reached the prior event's final, winner and runner-up alike. About 1.5pp for a semi-final loss. Nothing at the quarter-finals and below. Scaled by the ELO the run added rather than by the trophy, up to 4.8pp for a run worth 70 points or more, nothing for a top-five player, largest for ranks 21 to 50. A 1.9pp credit after an opening-match loss, only inside seven days. And a hard rule that it may shade a price and may never be the reason a bet exists, because of the market limitation above.
An independent check landed the same week and agreed. A separate ablation of our machine-learning feature set, with and without the hand-coded human-read features, found the feature group that contains letdown_flag and title_recency worth nothing on a temporal holdout. Two methods, one answer. We publish every pick before the match and grade wins and losses on one public board, and deleting our own rule in public, on our own data, is that same policy applied to ourselves.
Method: 64,669 event transitions, 2018 through 2026, ATP and WTA, main tour and Challenger level; a transition is one player finishing one event and starting the next within 120 days. Each transition scores the first match of the next event as observed minus ELO-expected, pre-match ratings only, reconstructed by subtracting each match's own rating change from the stored post-match value; every cell is calibration-matched (the rating's own bias at the same expectation level is netted out) with Poisson-binomial standard errors. Events are reconstructed by clustering matches into editions on 21+ day gaps and walking the main draw backwards from the unique final, which qualifying rounds cannot survive; editions without a unique final are dropped, draw size comes from the backward walk, and standing is a trailing-365-day overall-ELO rank among same-tour players. Walkovers are excluded as targets, retirements kept. The persistence pair is observed minus expected on the same player panel without the calibration net. The surface-book check reruns the identical study with expectations priced from pure surface ratings. The market panel uses the 174 title transitions from 2026 with de-vigged closing odds.
References:
[1] T. Gilovich, R. Vallone, and A. Tversky, "The hot hand in basketball: On the misperception of random sequences," Cognitive Psychology, vol. 17, no. 3, pp. 295–314, 1985.
[2] F. J. G. M. Klaassen and J. R. Magnus, "Are points in tennis independent and identically distributed? Evidence from a dynamic binary panel data model," Journal of the American Statistical Association, vol. 96, no. 454, pp. 500–509, 2001.
[3] 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.
[4] FiveThirtyEight, "How we're forecasting the 2016 US Open," fivethirtyeight.com, 2016, archived at the Internet Archive (accessed Aug. 12, 2026).
[5] M. E. Glickman, "Parameter estimation in large dynamic paired comparison experiments," Journal of the Royal Statistical Society Series C (Applied Statistics), vol. 48, no. 3, pp. 377–394, 1999.
There is a real deficit: players who just won a title win their next match 3.5 percentage points less often than their own rating predicts, on 1,922 title transitions. But it is not psychological. Runners-up from the same finals dip just as much (−3.4pp vs −3.7pp for champions), and on a pure surface rating book the deficit disappears entirely. The rating overshot the hot week; the player is fine.
No. If it were fatigue, the worst case would be the quickest turnaround. It is the opposite: playing again within 7 days of a title costs −2.7pp, while returning after 22 or more days costs −8.8pp, the worst timing cell in the study. Tired legs recover with rest; an inflated rating just sits there until results correct it.
We have not shown that, and we are not claiming it. On the 174 title transitions from 2026 with de-vigged closing odds, players came in at +1.0pp against the market, which at that sample size means nothing. This is a model calibration finding: correct your own rating if it pools surfaces, but do not treat the letdown as a market edge. Our framework may use the correction to shade a price, never as the reason a bet exists.
See today's picks — published before the match, graded in public →