On 62,744 ATP and WTA matches, the overall (all-surface) ELO rating predicts better than the surface-specific rating on both tours, and a weight sweep finds the optimal mix at 0.7 overall plus 0.3 surface, almost exactly the split FiveThirtyEight published years ago. Our bot now uses that measured blend.
Every serious tennis model keeps separate ELO ratings per surface. Clay-courters and grass-courters are different species, so a rating that mixes surfaces must be worse. That is the intuition, and we believed it too: our bot's ELO probability used the surface-specific rating from day one. Then a run of losing bets made us check the intuition instead of trusting it.
ELO-style ratings are not a fringe choice: in the most cited head-to-head comparison of tennis forecasting approaches, Kovalchik tested 11 published models on a full ATP season and found ELO-based methods among the most accurate available [1]. Later work has pushed the rating idea further; Angelini, Candila and De Angelis extended ELO with margin-of-victory weighting and report Brier scores around 0.212 on tour-level tennis [2], a useful external yardstick for the numbers below.
The surface question has a quieter history. The best-known practitioner system, FiveThirtyEight's tennis ELO, never actually chose between surface and overall ratings: it blended them, and the published blend put roughly 0.71 weight on the OVERALL rating and only 0.29 on the surface rating [3]. The people who ran this at scale had already concluded the pooled rating carries most of the information. We wanted to measure it on our own data, across both tours and down through Challenger and ITF level, where samples are thinner and the noise argument should bite hardest.
We took every match since January 2025 where both players had at least 10 rated matches on the match surface, so both a surface-specific and an overall rating existed for both players. 62,744 matches qualified. For each we computed two win probabilities from the standard ELO formula, pre-match ratings only, and scored both with the Brier score (lower is better; 0.25 is a coin flip).
| Tour | Matches | Surface ELO | Overall ELO | Better |
|---|---|---|---|---|
| ATP | 36,958 | 0.2129 | 0.2095 | overall |
| WTA | 25,786 | 0.2091 | 0.2055 | overall |
The pooled rating wins on both tours, consistently across every cut we looked at. Both numbers sit in the same range as published tour-level ELO benchmarks [2], a sanity check on the rating system rather than a like-for-like comparison, since samples and eras differ.
Pure-vs-pure is the wrong final question, because the practitioner answer was always a mix. So we swept the blend weight: rate each player as w × overall + (1−w) × surface, for w from 0 to 1, and score every step on the same matches (the qualifying pool grows slightly here, 72,112 matches, because the sweep run included late-window rows).
| Tour | w=0.0 (surface) | w=0.3 | w=0.5 | w=0.7 | w=1.0 (overall) |
|---|---|---|---|---|---|
| ATP | 0.2114 | 0.2087 | 0.2076 | 0.2072 | 0.2079 |
| WTA | 0.2080 | 0.2049 | 0.2037 | 0.2034 | 0.2043 |
The optimum is w = 0.7 on both tours. The blend beats pure overall too, and the curve is the same shape for men and women. FiveThirtyEight's 0.71/0.29 folklore [3] replicates almost exactly on an independent dataset that leans much harder on Challenger and ITF tennis than theirs did. Surface information is worth about 30% of the rating, no more.
Surface splits fight over the same finite match history. A player with 300 career matches might have 40 on clay; the clay rating built from those 40 is a noisier estimate of current strength than the rating built from all 300. Surface identity is real, but it is a modest, slowly-changing trait, while small-sample noise is large and immediate. The blend works because it takes the stable signal from everything the player has done and only a measured dose of the surface-specific part.
We also tested a second folk theory: that ratings earned at ITF or Challenger level are inflated, so players stepping up should underperform their ELO expectation.
| Tour | Bucket | n | Expected | Actual | Gap |
|---|---|---|---|---|---|
| ATP | stepping up | 7,936 | 42.1% | 42.1% | +0.0pp |
| ATP | stepping down | 4,066 | 64.2% | 64.3% | +0.2pp |
| WTA | stepping up | 4,733 | 42.9% | 42.1% | −0.9pp |
| WTA | stepping down | 3,264 | 60.7% | 62.1% | +1.4pp |
ATP: no effect at all. WTA: a small real tilt, about one percentage point against players stepping up. Worth knowing; far too small to justify separate rating pools per level.
This is the part most prediction sites would skip. On the morning of this study we switched our bot's ELO probability to the pure overall rating, because round 1 said it was better. By the evening the sweep had a sharper answer, so we moved again: the bot's ELO input is now the measured 0.7 overall + 0.3 surface blend, with a fallback to the pooled rating alone when a player has fewer than 10 matches on the surface. Two model changes in one day, both in public, both because the data said so.
The per-surface ratings still exist and our analyst still sees them, because a 150-point gap between a player's clay and hard ratings is real information about how to read a matchup. They just no longer decide the headline probability on their own, because on that job they measurably lose. The surface itself still enters where it belongs, in the point-level serve and return simulation, which rests on the classic result that tennis points are close enough to independent and identically distributed for such models to work well [4].
We publish every pick before the match and grade wins and losses on one public board. Changing our own model input in public when the data disagrees with our intuition is the same policy applied to ourselves.
One rating system: the K-factors, starting ratings and update rules are ours, and a differently tuned surface ELO (for example with margin weighting [2]) could close some of the gap. One period: matches from January 2025 onward. The 10-match surface threshold biases the comparison sample TOWARD matches where the surface rating should be at its best, and it still loses. The sweep resolution is 0.1, so the true optimum is somewhere near 0.7, not exactly it. And our data leans harder on Challenger and ITF tennis than most published studies, which is our niche but limits comparability with tour-only benchmarks [1], [2].
Method: our own ELO system (K-factor by match count, level-aware starting ratings, updated per match, main tour through Challenger and ITF). Pre-match ratings only, no leakage from the match being predicted. Surface comparison requires 10+ rated matches on the surface for both players. Blend applied at the rating level before the standard expectation formula. Transition buckets use the average level of a player's last 10 matches vs the current match level, minimum 5 recent matches. Brier scores are means over qualifying matches; database snapshot through July 2026.
References:
[1] 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.
[2] G. Angelini, V. Candila, and L. De Angelis, "Weighted Elo rating for tennis match predictions," European Journal of Operational Research, vol. 297, no. 1, pp. 120–132, 2022.
[3] FiveThirtyEight, "How we're forecasting the 2016 US Open," fivethirtyeight.com, 2016.
[4] 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.
No. On 62,744 ATP and WTA matches since 2025, the overall (all-surface) ELO rating predicted match winners better than the surface-specific rating on both tours (Brier 0.2095 vs 0.2129 on ATP, 0.2055 vs 0.2091 on WTA). Surface splits divide a finite match history into smaller, noisier samples, and below a few hundred matches per surface the added noise outweighs the surface signal.
About 0.7 overall + 0.3 surface. Sweeping the blend weight from 0 to 1 on 72,112 matches, w=0.7 minimized the Brier score on both the ATP and WTA sides, beating pure overall and pure surface ratings. This independently replicates the 0.71/0.29 split FiveThirtyEight published for its tennis forecasts.
Barely. On the ATP side there is no effect at all (players stepping up win exactly as often as their rating predicts). On the WTA side there is a small real tilt of about one percentage point below expectation for step-up players and about 1.4 points above for players stepping down. That is worth a note in analysis, not separate rating systems per level.
See today's picks — published before the match, graded in public →