Every team has a number it quotes without checking. The conversion rate. The churn figure. The one slide that gets reused for a year. It feels settled because everyone says it, and that is exactly what makes it dangerous.
Here is the problem in a setting nobody argues about, because nothing is at stake in it. Professional tennis.
Across 116,382 completed professional singles matches, a player who loses the opening set goes on to win 16.52% of the time. One in six. The figure is correct. It is also close to useless, and the reason it is useless is the same reason your conversion rate might be.
The number is an average of things that are not alike
Split those same matches by how the first set was lost and the single figure comes apart:
– lost the first set 7-6: wins 21.92% of the time, from 14,265 matches
– lost it 6-4: 19.41%, from 25,546
– lost it 7-5: 18.89%, from 10,252
– lost it 6-3: 17.39%, from 25,970
– lost it 6-2: 13.30%, from 19,739
– lost it 6-1: 11.06%, from 14,269
– lost it 6-0: 7.18%, from 5,833
Top to bottom, that is a threefold spread. And look at the two groups of almost identical size, 14,265 and 14,269. One recovers twice as often as the other. Same sport, same measurement, same period. The only difference is a detail the average threw away.
The ordering never breaks once across all seven scorelines. That is what a real pattern looks like, and it was invisible until the number was split.
Four words recover it
The useful version of “one in six” is not a better average. It is one question: by how much. The margin of the first set is the thing that carried almost all the signal, and it was sitting in the data the whole time.
The business equivalents are not hard to find. A 3% conversion rate across all traffic hides the channel that converts at 9% and the one that converts at 0.4%. An average handling time hides the ticket type that takes eleven minutes. In every case the fix is the same: find the dimension you averaged over, and put it back.
Two tests for whether your number is hiding something
**Does the split hold its order?** Take the metric apart by one sensible dimension and read the groups in order. If they line up and keep lining up as the dimension increases, the pattern is real. If they jump around, you are probably reading noise. In the tennis data the rate falls every single time the margin widens, across seven steps, with no exception.
**Are the group sizes similar?** Two groups of 14,265 and 14,269 matches give you a clean comparison. Size cannot be the explanation for the gap between them. When your segments are 400 and 40,000, the small one will swing for reasons that have nothing to do with what you are measuring.
Context changes the baseline, too
The same data has a second lesson, which is that the baseline itself moves with conditions. Service hold rates across 2,239,440 service games run 74.43% on grass, 69.95% on hard courts and 65.98% on clay. A hold rate that is unremarkable on one surface is poor on another.
By tour the spread is wider still. Men’s professional matches hold 78.62% across 230,715 games, women’s 64.36% across 266,503. Both reproduce the hold rates published elsewhere, on a method tuned to neither. That is how you know the measurement is sound.
None of that is reachable from a single pooled number. It needs the raw events, kept at the level they happened.
Which is the actual lesson
These splits were available for one reason. The underlying data was stored point by point, with every event kept, not as a match summary. Nobody decided in advance that first-set margin would matter. It mattered, and the detail was still there to check.
That is the practical takeaway for anyone building reporting. Aggregate late. Keep the dimension you think you will never need. The question that makes your number useful is usually one you have not thought of yet, and pre-aggregated data answers only what someone already anticipated.
For this analysis the match-level events came from the Live Tennis API. It serves live scores, fixtures, players and match events as JSON, with completed results and point-by-point history. The figures above are published openly at blog.livetennisapi.com/studies.json under CC BY 4.0, free to reuse with attribution. Anyone can check them.
Method, briefly: completed best-of-three professional singles matches from 1 January 2023 to 14 September 2026. Retirements and walkovers excluded. Counted from the point-by-point record, not estimated.
And that is the shape of the whole thing. The number was never wrong. It was just answering a question nobody had asked.
