Does a 70% tennis favourite win 70% of the time?
· Pronocast
When a prediction says “70%” for a player, what does it mean? That out of ten matches like that one, they win seven. That is the promise of a well-calibrated probability. But does it hold? We took the Elo formula as it is, with no correction, and checked it against the real results of 38,071 ATP and WTA singles matches played between January 2022 and September 2026.
The answer is clear: the raw formula is too sure of itself. The favourite it puts at 70% or more actually wins 76.9% of the time, against 80.0% announced.
What Elo announces, what happens
We grouped the matches by the probability Elo gave the favourite, in five-point bands, then took each band’s actual win rate. If the probability were perfectly calibrated, the two columns would match.
| Announced band | Matches | Announced | Actual | Gap |
|---|---|---|---|---|
| 50 to 55% | 6,399 | 52.5% | 52.6% | +0.1 pt |
| 55 to 60% | 6,262 | 57.5% | 54.9% | −2.6 pts |
| 60 to 65% | 5,777 | 62.4% | 60.8% | −1.6 pt |
| 65 to 70% | 5,285 | 67.4% | 65.0% | −2.5 pts |
| 70 to 75% | 4,421 | 72.4% | 68.7% | −3.7 pts |
| 75 to 80% | 3,576 | 77.4% | 74.0% | −3.3 pts |
| 80 to 85% | 2,797 | 82.4% | 79.6% | −2.8 pts |
| 85 to 90% | 1,948 | 87.4% | 84.0% | −3.4 pts |
| 90% and above | 1,606 | 93.5% | 92.3% | −1.2 pt |
Across all matches, the Elo favourite wins 65.1% where the formula announced 67.2%. The gap is small on balanced matches, then it settles in: 1 to 4 points in every band above 55%. Each actual rate is known to within ±1.2 to ±1.6 points (95% interval), so the “70 to 75%” band that wins 68.7% is well below its announced 72.4%, beyond chance.
Taken together, the 14,348 matches where the favourite was announced at 70% or more produced 76.9% wins (±0.7 point) against 80.0% announced. Those favourites lost 3,319 times: one defeat every 4.3 matches, where the formula implied one in five.
Why raw Elo is too optimistic
The Elo formula turns a points gap into a probability, but it does not know that extreme ratings are partly carried by luck: a player with a very high rating has often benefited from good draws and tight results that went their way. Their real strength is a little lower, and the next match reminds them of it. Statisticians call this regression to the mean. We did not isolate that explanation in the data: it is the usual hypothesis, not a result measured here.
What we do measure is the outcome: a raw probability cannot be taken at face value. That is the point of the calibration step in a serious model, which corrects these gaps before displaying a percentage (see how a tennis prediction algorithm works).
The same gap on the ATP and WTA tours
We split by tour after seeing the total: these are checks, not discoveries.
| Announced band | Tour | Matches | Announced | Actual |
|---|---|---|---|---|
| 65 to 75% | ATP | 4,996 | 69.7% | 66.8% |
| 65 to 75% | WTA | 4,710 | 69.7% | 66.6% |
| 75 to 85% | ATP | 3,136 | 79.5% | 75.5% |
| 75 to 85% | WTA | 3,237 | 79.7% | 77.5% |
| 85% and above | ATP | 1,696 | 90.2% | 86.7% |
| 85% and above | WTA | 1,858 | 90.1% | 88.6% |
The gap exists on both tours, a little wider on the ATP in the upper bands. By year it narrows: the Elo favourite wins 63.4% of matches in 2022 against 67.4% announced (−4.0 points), then 66.6% in 2026 against 67.6% announced (−1.0 point, up to 22 September). Our data cannot say whether this improvement will last or is sampling noise.
A 70% favourite still loses three matches in ten
Even with a perfectly calibrated probability, a 70% favourite loses three matches in ten on average. Over a run it shows quickly: across five such matches, the chance that all are won is 16.8%. Across ten matches, 2.8%. Across twenty, 0.1%. In other words, seeing a 70% favourite lose is nothing unusual: the opposite would be strange.
How to read a prediction probability
- A probability is a frequency, not a certainty. “70%” means three defeats in ten, not none.
- Be wary of a raw percentage. The Elo formula alone overstates its favourites by 1 to 4 points depending on the band. Check that a site calibrates its probabilities, and on which matches it publishes its results.
- Judge over hundreds of matches. Over a weekend, a few favourite defeats prove nothing: the verified results page shows accuracy over time.
- Compare with the market. An odds price also reads as a probability, with the bookmaker’s margin on top.
- These figures are about raw Elo, not the Pronocast model, which calibrates its probabilities (see our methodology).
Today’s predictions show each player’s probability. No percentage guarantees a result: if you bet, read our betting guide first.
Method
ATP and WTA singles matches, from 1 January 2022 to 22 September 2026, where both players had already played at least twenty matches: 38,071 matches, each seen from the favourite’s side, meaning the player the Elo formula gives at least 50%. The announced probability is the raw Elo formula’s, computed from each player’s rating before the match, with no calibration or correction: it is the one we check against the result. A band’s actual rate is the share of matches the favourite won; its 95% interval is 1.96 times the standard error of a proportion. The splits by tour and year were made after seeing the total. The probabilities of runs of matches won are exact binomial calculations for a probability of 0.70. The full script is kept with our data.
Frequently asked questions
- Does a 70% favourite really win 70% of their matches?
- With the raw Elo formula, no: a little less. Across 38,071 ATP and WTA matches, the 14,348 favourites announced at 70% or more won 76.9% of their matches against 80.0% announced, with a margin of ±0.7 point. In the 70 to 75% band, the favourite wins 68.7% against 72.4% announced.
- Why is raw Elo too optimistic?
- Extreme ratings are partly carried by luck, and the next match pulls the player back toward their real strength: regression to the mean. We did not isolate that explanation in the data; it is the usual hypothesis. The result itself is measured: a 1 to 4 point gap in every band above 55%.
- Is the gap the same on the ATP and WTA?
- Roughly, yes. In the 65 to 75% band, the favourite wins 66.8% on the ATP and 66.6% on the WTA against 69.7% announced. In the 85%-and-above band, 86.7% on the ATP and 88.6% on the WTA against 90% announced. The gap is a little wider on the ATP in the upper bands.
- How many matches does a 70% favourite lose?
- Three in ten on average. Across five matches, the chance that all are won is 16.8%; across ten matches, 2.8%; across twenty, 0.1%. Seeing a 70% favourite lose is the rule, not the exception.
- Do these figures apply to the Pronocast model?
- No, they are about the raw Elo formula. The Pronocast model calibrates its probabilities before displaying them: its method is detailed on the methodology page.
Other articles
- Tennis prediction algorithm: how an AI model actually worksWhat a tennis prediction model really receives, how it learns without cheating with time, and why its accuracy plateaus around the market’s.
- Elo in tennis: understanding the rating that predicts matchesWhy a rating designed for chess predicts a tennis match better than the ATP ranking, and which adjustments it needs to really work.
- Tennis odds and implied probability: what they really sayHow to read bookmaker odds, strip out the margin, and why the market stays the benchmark most models can’t durably beat.
- How to bet on tennis: bankroll, value betting and riskHow much to stake, what value really means, and why our own model lost 9% betting on its disagreements with the market.
- Clay, hard or grass: what the surface really changes in tennisAcross 36,210 matches, a player’s overall level predicts better than their record on the day’s surface, and grass is where favourites fall most often.
- Head-to-head in tennis: does the H2H really predict the winner?Across 38,123 matches, the head-to-head leader wins less often than the Elo favourite. But a record against the favourite really does count.
- Tennis fatigue and rest: do tired players really lose more?Stringing matches together is barely measurable. Returning after a month without playing is: nearly ten points of win probability less than Elo predicts.
- Indoor tennis: are favourites more reliable under a roof?Indoors, the favourite wins 64.9% of matches, against 65.0% on outdoor hard courts. A good indoor record mostly reflects a good player.
- Best of five in tennis: do favourites really win more?At the Slams, the ATP favourite wins 70.8% of matches, the WTA favourite 70.5% over three sets. What five sets really change, and what they don’t.
- Tennis Challenger vs ATP: are favourites less reliable?The favourite wins 64.0% of men’s Challenger matches and 64.5% on the ATP Tour. At equal probability we measure no gap: what Challenger changes, and what it doesn’t.
- Tennis winning streaks: do players on a run win more?A player on five straight wins takes 56.9% of their matches where Elo expected 60.3%. What a winning streak tells you, and what it doesn’t.
- Tennis surface change: are players really penalised?With global Elo, a player switching surface wins 3.6 points less than expected. With surface Elo, the gap disappears: what that changes for a prediction.
- WTA vs ATP: are there more upsets in women’s tennis?The favourite wins 65.7% on the WTA against 64.5% on the ATP, and at equal probability there is no difference. What the numbers say about upsets in women’s tennis.