Two kinds of DNA
| Layer | Built from | Examples |
|---|---|---|
| Match DNA | Results only: sets, games, tiebreaks, opponent ranking at match time | Match, set and game win rates; straight-sets and deciding-set win rates; tiebreak win rate; average game margin; win rates against top-10 / top-50 / top-100 opponents |
| Technical DNA | Match statistics only — never derived from results | Ace rate, double-fault rate, first serve in, first- and second-serve points won, service points won, hold rate, break points saved, return points won, return games won, break-point conversion |
Both layers are descriptive: they summarise what has happened. Neither is a forecast. The product's statuses say so directly — "Descriptive only" — and a rating movement is described as a fact about results, not a prediction.
What a percentile means
A 90th-percentile hold rate means the player's hold rate is higher than about 90% of qualified players on the same tour. Three qualifiers matter:
- Same tour only. ATP players are compared with ATP players and WTA players with WTA players. Tours are never pooled.
- Qualified only. A player needs a minimum sample on that metric — each metric has its own minimum denominator (a game-level rate needs far more games than a match-level rate needs matches) — and the comparison needs enough peers.
- Surface is explicit. Surface splits use the surface recorded by the source, never an inferred one, and need a minimum number of matches on that surface.
- Illustrative: a player shows hold rate at the 90th percentile, with a High confidence tier, on hard courts.
- Reading: across qualified players on the same tour, this hold rate is near the top — on hard courts, over the sample the tier describes.
- Not the reading: "this player wins 90% of service games" or "this player is 90% to win the match". A percentile is a rank among peers, not a rate and not a probability.
Confidence and missing data
Every value carries a confidence tier based on how many matches and points it rests on: insufficient below the metric's minimum, then low, medium and high as the sample grows. A percentile is shown only for medium or high confidence values.
Missing is never zero. When the source has no statistics for a match, or a player is below the minimum, the value is blank and marked insufficient — "we do not know", not "never happens". Technical statistics are not available for every tour and event, so a player can have full Match DNA and partial technical DNA. Where a tour comparison has too few qualified players, the product holds that comparison and says so, while Match DNA is still shown.
From description to prediction
A matchup probability is a different object. Matchups shows a win probability from the PBE Rating — an Elo-style rating built from results — only where that tour's rating has beaten a ranking-based baseline in backtests, and only when both players have enough rated matches. Head-to-head records are shown as context, not used as a model feature. See Calibration for how to judge such probabilities.
DNA describes past opponents and conditions. A strong return percentile built mostly on clay may not carry to grass; a short sample after a long absence can mislead; and descriptive ranks do not account for who a player will face next. Tennis DNA and some matchup surfaces are All Access features; the product's methodology page is the source of truth for current gates.