On 299,508 matches, 3.59% end in a retirement, climbing from 2.57% at ATP level to 4.09% at men's ITF level, with grass the safest surface. The strongest predictor is a prior retirement, a confound-stratified 2.24x odds ratio that our framework uses only as a brake, never as an edge.
This study started with a bug. Our bet-grading function is_retirement() exists to notice that a match ended early, so an alt-market bet (a handicap or totals position) is voided instead of graded on an unfinished score. Validated against hand-checkable data, it caught 272 of 10,757 real retirements: 2.5% recall. It read a status field populated only since June and looked for the string "RET" in a score field that never contains it. One in 28 matches in our database ends in a retirement; the grader saw fewer than one in a thousand.
Below: the bug, the fix, and a study on the corrected labels covering base rates, who retires, and the one predictor our betting framework now uses.
The first version put the headline rate at 3.74%, one in 27, on 304,179 matches. A re-audit found the replacement detector had a smaller version of the same problem: 618 rows with a "0 - 0" tally and zero recorded games, walkovers and cancellations the feed labels "Finished" (clustered 2021-2023), were counted as retirements, and roughly 4,000 more zero-game rows with a completed-looking tally sat in the denominator as ordinary matches. All zero-play rows are now excluded and every number here comes from the corrected rerun; the detector's validation on the ground-truth era is unchanged (precision 1.000, recall 1.000). The headline moves from one in 27 to one in 28; every gradient and the prior-retirement hazard survive, slightly smaller. We also cross-validated the labels row by row against an independent open dataset (99.93% agreement on 10,691 overlapping matches; see limitations).
Match.status is reliable only from June 2026 onward (before that, everything reads "Finished"), and our score field stores a set tally with no retirement marker. The fix reads the scoreline: a completed best-of-three win requires the winner to hold two legally completed sets, with guards for padding artifacts and super-tiebreak deciders. On the window where the status field can serve as ground truth (July 2026 onward, 3,534 matches), the corrected detector scores precision 1.000, recall 1.000 (147 true positives, 0 false positives, 0 false negatives).
We re-inspected every settled alt-market bet on the public Claude track: all 20 completed normally and none was a retirement. No bet was mis-graded, by luck rather than design.
Palau and colleagues counted 19,314 retirements across 584,806 second- and third-tier men's matches (Challenger and ITF) and 7,306 across 267,380 women's matches: 1.56 per 1000 games for the men, 1.36 for the women, about 3.3% of matches, with hard and clay riskier than grass [1]. Oliver and colleagues covered four decades of ATP and WTA main-tour events and report a rising trend, with surface, tournament type, round and ranking difference mattering and age difference not [2]. Casals and colleagues counted Davis Cup singles from 2000 to 2019: 3.2% of matches, 1.05 per 1000 games, best-of-five format carrying a large risk multiplier (IRR 5.52), and 63% of retirement injuries in the lower limbs [3].
On 299,508 matches with the corrected labels, 10,757 ended in a retirement: 3.59% of matches, 1.70 per 1000 games.
| slice | matches | retirements | per match | per 1000 games |
|---|---|---|---|---|
| ALL | 299,508 | 10,757 | 3.59% | 1.70 |
| ATP main tour | 24,109 | 620 | 2.57% | 1.06 |
| WTA main tour | 23,980 | 673 | 2.81% | 1.33 |
| Challenger (M) | 53,997 | 1,941 | 3.59% | 1.63 |
| ITF (M) | 94,604 | 3,870 | 4.09% | 1.96 |
| ITF (W) | 93,450 | 3,343 | 3.58% | 1.78 |
Our WTA main-tour rate (1.33/1000) is close to the published women's tier-2/3 figure of 1.36 [1], and our Challenger and ITF men's rates (1.63 and 1.96) bracket the published men's 1.56. Retirements rise further from the top tier, with men's ITF highest at 4.09% per match; prize money that requires showing up, best-of-three volume scheduling and thinner medical support all point that way. Our ATP main-tour rate reads lower than published main-tour work partly because our detector misses some best-of-five retirements (see limitations).
Per player-match (baseline 1.80%):
| predictor | cells | rate |
|---|---|---|
| Age | under 21 → 33+ | 1.49% → 2.28%, monotone |
| Surface | grass vs hard vs clay | 1.11% vs 1.86% vs 1.79% |
| Tier | ATP tour vs ITF (M) | 1.29% vs 2.05% |
| Draw stage | finals vs earlier rounds | 1.46% vs 1.83–1.88% |
| Load | 100+ games in prior 7d vs fresh | 2.89% vs 1.60% |
Grass as the safest surface matches the published direction on both tours [1], [3]. Players rarely retire from finals, which reads as selection (a finalist was healthy enough to win all week) plus stakes. The load cell is computable before the match: 100+ games in the previous seven days nearly doubles the base rate.
Davis Cup research found best-of-five format carrying a risk multiplier above five [3]. Our data cannot confirm it, and that is not a refutation. Raw, our best-of-five cell (Slam main draws, 5,178 player-matches) retires at 1.35%, below the best-of-three rest at 1.80%; held against its own tier it is null (ATP tour best-of-three sits at 1.29%). But a best-of-five retirement after the leader has banked two completed sets is invisible to a scoreline detector, so our cell is an undercount of unknown size. The naive pooled comparison, where best-of-five looks safer, is a tier and measurement artifact, not a fact about format.
Retirement history is the strongest signal we found. Raw, players with two or more career retirements retire at 2.70% against 1.27% for players with none, but career length confounds that comparison. Stratified jointly by tier, age band and career matches played, the Mantel-Haenszel pooled odds ratio across 101 strata is 2.24, replicated by tour (2.18 men, 2.33 women). Within every career-length band the lift is 1.8x to 3.9x. Most retirements are lower-limb muscle and tendon problems [3], which recur.
Little. A void returns the stake, so on alt markets even the worst realistic cell (prior retirements plus heavy recent load, roughly 4% void probability against the 2.3% baseline) costs forgone edge on about 1.7% more bets: roughly a 2% relative EV haircut, inside staking noise. Alt-market sizing does not change. On the moneyline a retirement is not voided: the bet stands and the retiring player loses. When both flags fire on a side we back (prior retirement history and 100+ games of recent load), our framework shades that side's win probability down by up to one percentage point. Risk context can shrink or remove an edge, never create one.
Method: 299,508 completed matches (ATP, WTA, Challenger, ITF; walkovers and zero-play rows excluded), 10,757 retirement-flagged by the corrected scoreline detector: winner must hold two legally completed sets, guarded against score-padding artifacts, super-tiebreak deciders and zero-game phantom rows (fixes in commits e0e27c8 and the 2026-08-29 zero-play guard). Detector validated on the 2026-07+ era where Match.status is populated: TP 147, FP 0, FN 0, TN 3,387. Per-player-match predictor rates use 599,016 player-match rows, baseline 1.80%, 95% CIs by Wilson score, p-values two-sided against the baseline. Prior-retirement effect is a Mantel-Haenszel pooled odds ratio across 101 joint tier x age-band x career-matches strata (per-tour reruns over 53 and 48 strata respectively). All cells from the 2026-08-29 corrected rerun of study_retirement.
References:
[1] A. Palau, E. Baiget, J. Cortés, L. Martin, M. Crespo, and M. Casals, "Retirements of professional tennis players in second- and third-tier tournaments on the ATP and WTA tours," PLOS ONE, vol. 19, no. 6, e0304638, 2024, doi:10.1371/journal.pone.0304638.
[2] A. Oliver, E. Baiget, J. Cortés, L. Martin, M. Crespo, and M. Casals, "Retirements of professional tennis players in ATP and WTA tour events," European Journal of Sport Science, vol. 24, 2024, doi:10.1002/ejsc.12177.
[3] M. Casals, J. Cortés, D. Llenderrozos, M. Crespo, T. E. Hewett, L. Martin, and E. Baiget, "Epidemiology and factors influencing Davis Cup retirements over the past twenty years," International Journal of Sports Physical Therapy, vol. 19, no. 9, 2024.
On 299,508 matches across ATP, WTA, Challenger and ITF levels, 3.59% ended in a retirement, about one match in 28, or 1.70 retirements per 1000 games. The rate climbs as you go down the ladder: 2.57% at ATP main-tour level but 4.09% at men's ITF level. Published sports-medicine studies of the middle tiers report 1.56 (men) and 1.36 (women) per 1000 games, right where our equivalent slices land.
The strongest predictor is having done it before: stratified by tier, age and career length, players with prior retirements carry a 2.24x odds ratio, replicated on both tours. Older players retire more (2.28% at 33+, 1.49% under 21), heavy recent load nearly doubles the rate (2.89% after 100+ games in seven days), and grass is the safest surface (1.11% vs 1.86% on hard).
Davis Cup research found best-of-five format carrying more than five times the retirement risk. Our data cannot confirm or deny it: our detector infers retirements from scorelines, and a best-of-five retirement after two completed winner sets is invisible to it, so our Slam cells undercount. What we can say is that the naive pooled comparison, where best-of-five looks safer, is an artifact of tier and measurement, not evidence about the format.
See today's picks: published before the match, graded in public →