On 22,048 matches from the 2026 season, the betting market's favorite won 71.5% of the time, and its probabilities are honest at the low end while slightly underrating heavy favorites: a 90-95% favorite actually wins 96.4%. Our own model picks winners less accurately than the closing market, which is normal.
"How accurate should a tennis prediction be?" has no single answer, because accuracy depends mostly on which matches are picked.
The same 2026 season and the same method (pick the favorite), sliced by price:
| Only bet when the favorite is... | n | "Accuracy" |
|---|---|---|
| Priced under 1.30 | 8,549 | 85.3% |
| Priced 1.30 to 1.60 | 8,822 | 66.7% |
| Priced 1.60 to 2.00 | 4,677 | 55.2% |
Picking only favorites under 1.30 gives 85% accuracy and still loses money, because a 1.15 price already implies 87% (our favorite-longshot study shows those bets return minus 3% and worse). An 80%+ accuracy claim without the odds behind it cannot be told apart from match selection. Independent tipster reviewers likewise warn bettors to ignore anyone promising 80-90% win rates.
We stripped the bookmaker margin from every closing price (no-vig, both sides scaled to 100%) and checked calibration by band: when the market says 70%, does the favorite win 70% of the time?
| Market says | Matches | Actually won |
|---|---|---|
| 50-55% | 2,335 | 53.2% |
| 55-60% | 3,498 | 58.7% |
| 60-65% | 3,435 | 63.3% |
| 65-70% | 2,601 | 69.1% |
| 70-75% | 2,557 | 73.2% |
| 75-80% | 2,918 | 80.8% |
| 80-85% | 1,598 | 84.9% |
| 85-90% | 2,056 | 91.9% |
| 90-95% | 950 | 96.4% |
Where matches are uncertain (55-65%), results land within a point or two of the market's number. An early version of our script counted exact even-odds matches as favorite wins and inflated the coin-flip band by about 11 points; those pairs are now excluded, since a 50/50 has no favorite.
At the top, heavy favorites win 4-5 points more than implied. This is the favorite-longshot bias seen from the favorite's side: margin is loaded onto underdogs, so favorite prices sit slightly long. The vig absorbs that gap at prices you can actually take (heavy favorites returned minus 3% flat-staked).
Overall the market's favorite wins 71.5% of matches, with a Brier score of 0.187 (lower is better; a coin-flipper scores 0.25). The widest published comparison of tennis forecasting models, eleven approaches on a full ATP season, found the best of them, plain ranking regressions and Elo, in the same range [1]. Roughly a quarter to a third of tennis matches are won by the player the predictor did not favour.
On the 4,477 matches our forward pipeline predicted and settled this season:
| Predictor | Winner-pick accuracy | Brier score |
|---|---|---|
| Devigged closing market | 67.6% | 0.204 |
| Our model blend | 62.8% | 0.225 |
The close aggregates everyone's money and information minutes before play; our model prices hours earlier from historical data. Market accuracy is lower on this subset than the 71.5% above because our slate skews toward closer matches. Some of our model inputs are partly anchored to market prices, so this is not a fully independent comparison. Our closing-line value against the fair close is covered in the devig-CLV study.
Method: 22,838 finished 2026 matches with closing odds; 22,048 after excluding exact even-odds pairs (no favorite exists) and a handful of unpriceable quotes. Implied probabilities devigged by proportional normalisation. Calibration is the devigged favorite probability, bucketed, vs actual favorite win rate. Model comparison on the 4,477 settled forward predictions with stored model probabilities, market Brier computed on the identical matches. Calibration table re-verified 2026-08-07 on the grown sample (22,634 matches): favorite wins 71.4%, Brier 0.188, every bucket within 0.3 points. Related: favorite-longshot, devig-CLV, head-to-head and fatigue studies.
References:
[1] S. Kovalchik, "Searching for the GOAT of tennis win prediction," Journal of Quantitative Analysis in Sports, vol. 12, no. 3, pp. 127–138, 2016.
Depends entirely on the matches. Picking heavy favorites gives 85%+ with zero skill and negative profit. On genuinely uncertain matches (both players priced 40-60%), the market itself manages about 53-63%, so anything sustainably above that range on those matches would be remarkable. Judge predictions against the odds, never against a bare percentage.
Not at the closing line, on the evidence so far. Our own model picks winners at 62.8% where the devigged close manages 67.6% on the same matches. The realistic goal for a model is beating a subset of prices early, before the market corrects, which is measured by closing-line value, not accuracy.
Easily, and that is the problem. Our data shows picking short favorites yields 85%+ accuracy with zero skill and negative profit, because a 1.15 price already implies 87% wins. High accuracy is cheap to produce honestly, so the number proves nothing by itself. Independent tipster reviewers warn bettors to ignore anyone promising 80-90% win rates and put realistic tennis strike rates at 40-60%, which matches what we measure for the market itself on competitive matches. If a claim comes without the odds and a full graded history attached, you cannot tell skill from a menu choice.
See today's picks: published before the match, graded in public →