Tennis
Do American men actually perform better at the US Open?
No, the evidence does not support a general US Open advantage.After accounting for player ability, surface, and draw, the main model finds no statistically significant US Open advantage over any of the other three majors in either recent window.
American men made four US Open semifinal appearances in2021–2025. This analysis compares those results with the number of American entries, pre-tournament ratings, and each tournament’s actual draw.
Entries and late-round appearances
Quarterfinal appearances
Appearances per 100 entries
Values used in this figure
Actual and expected bracket advances
An Elo rating estimates playing strength from earlier results and the strength of the opponents. Here, a blend of overall and surface-specific Elo ratings estimates each possible matchup before a tournament begins. The actual draw then determines which opponents a player could meet.
The model adds up each player’s chances of progressing through the draw to get an expected number of bracket advances. One advance means moving past one round. A walkover counts as an advance; a retirement counts as a played match.
Expectations use the recorded entrants and draw, with ratings held fixed throughout each tournament. These are model averages, not predictions of an exact total.
Values used in this figure
US Open performance relative to other Slams
Divide actual minus expected advances by the number of entries at each Slam. Then subtract the other Slam’s result from the US Open’s.A positive difference favors the US Open. The Australian Open is the main comparison because both tournaments use hard courts.
Statistically significant means the 95% interval excludes zero (a nominal 5% threshold). These exploratory comparisons are not corrected for testing multiple Slams, windows, or models. A non-significant result does not prove the true difference is zero.
Estimates, intervals, and significance
How to read significance and uncertainty
An interval entirely above zero indicates higher adjusted US Open performance; one entirely below zero indicates lower performance. If it includes zero, the difference is not statistically significant by this rule. Significance does not measure the size or practical importance of a difference. TheNIST handbook explains the connection between confidence intervals and hypothesis tests. Here the intervals come from a bootstrap, so the 5% threshold is approximate and depends on the sampling and model assumptions.
The bootstrap resamples whole players or whole years 5,000 times, retaining the outcomes grouped within each unit. The fitted model stays fixed. These intervals describe sampling variation in this comparison, conditional on that model; they do not include every source of rating error or establish a causal home effect. Year intervals have only five or ten clusters.
Results for individual players
Values used in this figure
Compare every American with a US Open entry
Ordered by total actual minus expected US Open advances. Different players have different numbers of entries; this is not a rating of player ability.
Results by year and sample exclusions
Annual US Open results vary around the model’s expectation. In 2024, Americans advanced 22 times against 14.5 expected. In 2025, they advanced 17 times against 23.2 expected.
Positive values mean more bracket advances than expected. These are annual totals, before the sample exclusions below.
Values used in this figure
The selected player or year is removed from both tournaments. Ratings and draw expectations remain fixed. These are point estimates.
Methods and remaining checks
This discovery run does not support a broad recent US Open advantage over the Australian Open. The narrower question is whether the recent impression reflects extra entries and an exceptional Tiafoe record. That question still needs confirmation.
The comparison does not isolate crowd support, travel, scheduling, injuries, or a causal home advantage. A negative estimate against Australia is also not proof of a US Open disadvantage.
Population, coverage, and outcome definitions
One entry is one player in one men’s singles Grand Slam main draw. “American” means the source’s recorded International Olympic Committee country code is USA; it does not describe birthplace, training location, or crowd support.
The five-year sample has 306 American entries across 20 tournaments. The decade has 560 across 39; Wimbledon was canceled in 2020. The wider audit verified 143 complete draws from 1990–2025, of which 103 fall in the 2000–2025 evaluation period. No 2026 results are included.
A played win excludes walkovers and includes retirements. A bracket advance includes walkovers. These totals coincide for recent Americans at the US Open and Australian Open. An all-American match contributes one win and two player match appearances to the aggregate win percentage.
Model, actual draw, and predictive checks
The surface-aware model blends overall and surface Elo equally. Ratings use earlier tour-level matches, with tuning in 1990–1994 and calibration in 1995–1999. Evaluation begins in 2000. A conservative event-completion buffer of 21 days for Slams and 14 days for other events protects against using unfinished events; ratings stay fixed throughout each target tournament.
The recorded main draw includes replacement entrants. Exact bracket calculations allow every opponent who could progress through that draw. Actual later-round winners do not determine simulated winners. Expected totals are exact under the model; 20,000 simulated draws per tournament supply conditional prediction intervals and null checks.
Across 4,932 played Slam matches in 2016–2025, the surface-aware model has a Brier score of 0.1825, versus 0.1902 for the ranking model. Brier score is the average squared error of predicted win probabilities; lower is better, and predicting 50% for every match gives 0.25. Better average prediction does not guarantee unbiased estimates for every subgroup.
Limits that still matter
- Limited prior match history. In the decade, 105 American entries had fewer than 20 earlier tour-level matches. Lower-level results are missing from the rating histories. Requiring 20 prior matches reduces the US–Australia estimate to about −0.063 advances per entry, with an interval spanning zero.
- Incomplete wildcard labels. An official 2022 draw audit corrected entry labels. US Open labels from 2016–2020 remain incomplete. An unlabeled entry cannot be assumed to have qualified directly, so the known-wildcard exclusion is only a partial check.
- Exploration is not confirmation. The 2000–2025 outcomes have already been inspected. Earlier years cannot now be presented as untouched holdout data. The Tiafoe pattern needs a specified replication check that accounts for having selected him after seeing the results.
Sources, attribution, and reproducibility
Match and ranking data: Jeff Sackmann / Tennis Abstract,ATP Tennis Rankings, Results, and Stats, retrieved frompreserved revision 712be0c, licensed underCC BY-NC-SA 4.0. The original repository was unavailable when the research was retrieved. The analysis normalizes names, repairs audited entries, removes two exact duplicate outcomes, derives ratings, and aggregates or simulates draws. It is not endorsed by the source author.
Source hashes, the measurement contract, official draw checks, model specifications, and the full discovery report are available in the research code and documentation. The numerical run passed 16 tests and reproduced 26 checked outputs byte for byte. This page contains derived summaries only. The derived data and charts are shared underCC BY-NC-SA 4.0.
Reproducible evidence SHA-256