Ask for one number and you will get one. The question that matters is what range it was pulled from, and how wide that range is. A tennis dataset makes the gap between the two easy to see.
The corpus holds 116,382 completed best-of-three singles matches from January 2023 to September 2026. In those, the player who lost the opening set went on to win 16.52% of the time. That figure is correct and it describes a match that nobody played.
The Range Behind the Single Number
Sort those matches by how the first set actually ended and the one rate separates into seven.
A first set lost on a tiebreak at 7-6 is followed by a comeback 21.92% of the time, across 14,265 matches. Lost 6-4, it is 19.41% over 25,546 matches. Lost 7-5, 18.89%. Lost 6-3, 17.39% over 25,970.
Then it falls away. Lost 6-2, 13.3%. Lost 6-1, 11.06%. Lost 6-0, 7.18% across 5,833 matches.
So the honest range runs from about 7% to about 22%, and the quoted 16.52% sits in the middle of it.
Why the Middle Is the Weakest Place to Stand
An average is a balance point, not a typical case. A balance point can sit exactly where very few observations are, and here it does.
The mechanism is plain. The margin of the first set measures how far apart the two players were. Six games apart means something different from one game apart, and the comeback rate tracks that reading closely. The seven rates fall in the same order as their margins, with no exception.
That ordering is the part worth trusting. A pattern holding across every category, with a reason you can state in a sentence, is hard to explain away. A difference between two groups is not.
A Second Cut That Behaves Differently
Sort the same matches by competition tier and almost nothing happens. The men’s top tour returns 18.21% across 10,945 matches. The men’s Challenger tier gives 17.32% over 30,406. The women’s Challenger tier 17.25%, the women’s top tour 17.04% over 13,543. The women’s ITF circuit 15.97%, the men’s 15.18%.
Three percentage points across six groups, with confidence intervals between 0.42 and 1.08 points. Some of those gaps are real. None would change a decision.
So one dataset yields a split worth acting on and a split barely worth mentioning. Only looking at both tells you which is which.
The Control That Makes It Believable
Sort by court surface and the rate hardly moves at all. Grass 17.2%, clay 16.79%, hard 16.47%, across 113,288 matches. Less than a percentage point.
That matters because it shows the method is not manufacturing differences wherever it looks. A cut that finds nothing, and reports it, makes the cuts that find something easier to believe.
Three Questions for Any Single Figure
What is it an average of? If the group contains obviously different cases, it is a summary and not a description.
What does the breakdown look like? Whoever computed the headline computed the parts first. Reluctance to show them is itself an answer.
Do the parts order themselves sensibly? Here comeback rate falls as the first-set margin widens, every time. A breakdown that jumps around has a sample problem or a definition problem.
All of the above comes from the Live Tennis API, a JSON service for developers. It covers live scores, fixtures, players and match events, with point-by-point history behind them. The figures are published at blog.livetennisapi.com/studies.json under CC BY 4.0, free to reuse with attribution.
