The idea
Tennis scoring is hierarchical: points → games → sets → match. Estimate each player's probability
of winning a point on their own serve against this specific opponent, and the standard
Barnett–Clarke Markov math propagates that up to the probability of winning any game, tiebreak, set
and match — and, crucially, every derivative market (set betting, total games, handicap).
Why that's an edge
Sharp books price the match winner tightly. But the derivative markets are looser and are often
internally inconsistent with the moneyline. Since everything here comes from the same two
numbers, when a book's match line and games line imply different serve dynamics, at least one is
mispriced. This tool surfaces that gap.
Where the numbers come from
Ratings are built from real match results (Jeff Sackmann's ATP/WTA datasets): surface-specific Elo
plus rolling serve/return percentages. When a player has enough matches on a surface, the model uses
their actual serve/return stats; otherwise it falls back to an Elo-implied estimate.
Honest limits: the Markov math is exact, but its output is only as good as the
serve/return inputs. Validation should lead with closing-line value, not P&L.
And the discipline of which matches to skip — low tiers carry match-fixing and
adverse-selection risk; retirement settlement rules are a hard gate — matters more than the model.
This is a research tool, not betting advice.