What matters
Most matches are best of three sets; men's Grand Slam singles are best of five, which gives the stronger player more time for an edge to show. Each game is either a service game or a return game. Servers win most of their games, so a set usually turns on one or two breaks of serve — or goes to a tiebreak when neither player breaks.
Surface is not interchangeable. Hard, clay and grass change how much the serve is worth and how long rallies last, so a player's clay record says limited things about grass. Match context matters too: retirements and walkovers end matches without a completed result, the draw decides who a player can meet, and opponent quality varies enormously across a season.
Important data
| Input | Why it matters |
|---|---|
| Service and return points | The underlying unit; games and sets are built from them |
| Surface (hard, clay, grass) | Changes the value of serve and return; samples split by surface get small fast |
| Match format | Best-of-3 vs best-of-5, tiebreak rules |
| Opponent quality | Ranking or rating of who a result came against |
| Recent form and sample | How many matches, on which surface, how recently |
| Completion status | Retirements and walkovers are not normal results |
Common metrics
| Metric | What it measures | How to read it |
|---|---|---|
| Ace rate / double-fault rate | Aces and double faults per service point | Serve weapon vs serve risk |
| First serve in | Share of first serves landed | Context for the two serve-points-won numbers |
| First- / second-serve points won | Points won behind each serve | Second-serve points won exposes weak serves |
| Service points won | All points won on serve | Drives hold rate |
| Hold rate | Service games won | Compounds from service points won |
| Break points saved | Break points faced and saved | Small samples; noisy |
| Return points won | Points won on the opponent’s serve (first and second return) | Drives break rate |
| Return games won / break-point conversion | Breaks of serve, and breaks per break-point chance | Where sets are usually decided |
Model concepts
Because points compound into games, a small difference in service points won becomes a large difference in hold rate — see the worked example. Match models therefore lean on serve and return, adjusted for surface and opponent.
Ranking is context, not probability. Rankings reward points accumulated over a rolling year, including tournaments a player chose to enter; they say little about a specific match on a specific surface. A rating built from match results is a better input — PropBetEdge Tennis uses an Elo-style PBE Rating, and publishes its matchup win probability only where a tour's rating has beaten the ranking baseline in backtests.
Tennis DNA describes what a player's record shows: Match DNA from results (set, game and tiebreak win rates, margins, opponent-quality splits) and technical DNA from match statistics (serve and return rates). Percentiles compare players only within the same tour, need a minimum sample, and are withheld rather than shown as zero when data is missing. Where a tour comparison has too few qualified players, the product holds it and says so. A DNA percentile is not a win probability. See Tennis DNA.
The PBE Tennis pick model has not been trained or validated out of sample, so PropBetEdge Tennis publishes no picks and no pick track record. The rule is the same in every sport: no published call without out-of-sample evidence.
Live intelligence
A tennis match is read tournament → match → set → game → point. Tennis PBEcast shows each level the source actually covers: some feeds carry the point score and the server; others carry only set and game scores, and the missing point data is left blank rather than invented. Updates are observations every several seconds, not a point-by-point stream, and replay walks only the stored events. See Tennis PBEcast.
PropBetEdge Tennis tools
Tennis DNA ↗
Match DNA and technical serve/return DNA with same-tour percentiles and sample gates. All Access.
TennisPBEcast ↗
Live analytical court and replay, at the granularity each source actually provides.
TennisMatchups ↗
Matchup DNA and PBE Rating win probabilities where a tour is validated. All Access.
TennisPlayers ↗
Player profiles and results.
TennisRankings ↗
ATP and WTA singles rankings, archived weekly.
TennisTournaments ↗
Events with dates, draws and results.
TennisNews ↗
Stories built from PropBetEdge data, checked against frozen evidence.
TennisMethodology ↗
Tennis DNA definitions, sample sizes, confidence tiers and as-of rules.
Worked example
- Illustrative, not a real player: a server wins 64% of service points. If points are independent, the chance of holding a game is P(hold) = p⁴(1 + 4q + 10q²) + 20p³q³ · p² / (1 − 2pq), with q = 1 − p.
- p = 0.64 → P(hold) = 81.3%. Drop to p = 0.60 → 73.6%.
- A 4-point drop in service points won costs 7.7 points of hold rate — roughly one extra lost service game every 13 service games.
Glossary
Related lessons
Tennis DNA
What does a Tennis DNA percentile actually tell me?
Tennis PBEcast
What is Tennis PBEcast showing me, and how fine-grained is it really?
Sample size
How many results before a record means something?
Calibration
When a model says 70%, does it happen 70% of the time?
Game state
What exactly does a live model know that a pregame model does not?