How TennisRace.ai works

Where our numbers come from, how the live ranking and points defending are calculated, and how every analysis is written and checked before publication.

Our data

We keep our own database of professional tennis: every ATP and WTA tour-level match with its draw, round and score, Challenger results, the weekly official rankings, the Race to the Finals and a match archive that goes back to 1968. Live results arrive every few minutes during play. Rankings and the Race follow the official weekly publications.

Live ranking and points defending

The live ranking shows each player's points after the results of the current week. Points defending are what a player scored at last year's edition of this week's tournaments: those points come off when the new official ranking is published on Monday. We compute them from last year's full draws with the official ATP points tables, including the rules for byes and qualifiers.

Expected score

Some match pages compare the score with what the points were worth. For each set we take the share of points each player won on serve and on return, work out how often players at those rates would win a set, and add those chances up over the match. That sum is the expected score, in sets. It uses no odds and no model trained on results.

The line appears only when the expected score is at least 0.75 of a set away from the real one, and not for a winner who was a full set ahead on expected sets anyway. It does not appear when a player retired, when the points of a tie-break are missing, or when the points do not add up to the score. It is a reading of one match, not a verdict on a player: in past matches, a player the score flattered did slightly worse in the next set than the score and the pre-match odds suggested.

How each analysis is written and checked

Every preview and review starts from a fact sheet extracted from our database: the seeds and their rankings, the points on offer, last year's results, the Race standings, head-to-head records and, for match previews, the betting market's view expressed as implied chances. A first model writes the article from those facts only, and every paragraph records which facts it uses.

Automatic checks then compare every number and every player name in the text with the facts the paragraph cites, and block superlatives that no fact supports. A second, independent model reads the article against the fact sheet and rejects anything the facts do not support. Articles that pass both are published straight away, and any that should not stay can be taken down. We never name bookmakers and never encourage betting. The full rules are in our editorial policy.

Match forecasts

We publish our own chance for each player before main-draw matches. It blends three things: Elo ratings by surface, serve and return skill estimated from match statistics with the strength of each opponent built in, and age. Serve and return skill is turned into a chance of winning the match by a model of points, games and sets, and the blend is fitted on our own recent results.

In a backtest on past matches, from before the forecasts went live, log loss was 0.5999 against 0.6073 for Elo alone, on 2,783 main-tour matches from July to November 2026, with weights fitted only on the months before. A backtest is not a live record: the forecast record starts empty and fills forward, one forecast at a time. We do not claim these forecasts are better than anyone else's.