Tennis Advanced

Held per tour and per surface: clay says little about grass.

• EloElo ratingMoves with the quality of who you beat.
• SPWServe points won ratingSmoothed, pulled to tour average until earned.
• RPWReturn points won ratingWhere the largest style differences live.
• Matches ratedMatches behind the ratingSo a thin sample is visible.
• Serve ptsServe point sampleThe gate the point level model depends on.

Two engines: one from the ratings, one from serve holds. Completed matches only.

• P(win) EloWin probability from ratingsFrom the surface rating gap alone.
• P(win) MarkovWin probability from serve holdsEmpty under two thousand serve points a side.
• Straight setsStraight sets probabilityA different question from who wins.
• Bo3 scoreBest of three set scoresTwo nil, two one, and the reverses.
• Bo5 scoreBest of five set scoresThe same over the men's major format.
• E[games]Expected total gamesThe total most often traded.
• SD gamesTotal games spreadHow safe that total actually is.

How a serve has performed over a recent window.

• Ace%Ace rateServe points ended by an ace.
• DF%Double fault rateThe cheapest way there is to lose a set.
• 1st%First serves inHow often the second has to be used.
• 1SW%First serve points wonWhat holds a service game together.
• 2SW%Second serve points wonThe best single predictor of a close match.
• Hold%Service games heldWhat everything above rolls up into.

The handful of points that decide sets.

• BP save%Break points savedA good server against a clutch one.
• BP conv%Break points convertedThe gap here is worth whole sets.
• Break%Return games wonThe return twin of hold percentage.
• BP facedBreak points faced per matchThe context a save rate needs.

Splits against the opponents that trouble a given player.

• vs LHVersus left handersA persistent problem for some players.
• vs Top 10Versus top tenThe only sample that matters late in a draw.
• vs Top 50Versus top fiftySolid tour level, with a larger sample.
• vs big serversVersus big serversTop quarter of ace rate. Most upsets live here.
• Ace quartileServe power bandWhich quarter his own ace rate falls in.

Ratings rebuilt as they stood, which is what a backtest needs.

• As of dateRating as of a dateUsing only what was known before that day.
• Leak checkLeakage guardEnforced in the pipeline, not trusted.
• SnapshotStored rating snapshotSo a historical claim reproduces exactly.
• Replay windowBacktest windowWhat a reported hit rate actually measured.