How the tennis prediction model works

One thing only: the probability that each player wins the match. A likely set score is published too, but it is a secondary output and less reliable than the win probability.

The model covers the ATP, WTA, Challenger men's and Challenger women's tours, in singles and doubles. ITF tournaments are not predicted, but their results feed the strength ratings of the players who compete in them.

The model is a gradient boosting model trained on professional match history since 2020, in four steps:

• an overall Elo and a surface Elo, the latter seeded from the overall rating rather than from a neutral value, without which it converges poorly and ends up less predictive than the overall rating alone;
• recent form and a fatigue measure drawn from matches played in the preceding days;
• the head-to-head record, weighted by recency: a meeting five years ago does not count as much as one last month;
• the market's implied probability where odds are available, with the bookmaker's margin removed.

No information from after the match can enter the calculation. It is the one rule that, when broken, produces a model that looks excellent in testing and is useless in production, without ever raising an error.

In practice, each player's state is read and updated in that order, by a single shared implementation used both for training and for prediction. A match's detailed statistics, which are only known once it has finished, are never used as they are: they enter only as rolling averages computed over the player's earlier matches.

The accuracy shown in the application is computed differently, and it is the figure that really counts: it compares the predictions actually published before the matches with the results that followed. It is not reconstructed after the fact and cannot be filtered retroactively.

Accuracy alone is not enough to judge a model that publishes probabilities: it ignores confidence. Saying 51% or 95% for the same correct call gives the same accuracy, though the two are not worth the same.

Finally, the probabilities produced are capped: no prediction shows absolute certainty. A calibration left unbounded produces values of exactly 0 or 1 in thinly covered regions, and a single one of those proved wrong costs as much as hundreds of moderate errors.

Where the data comes from

Schedules, results, live scores and odds come from the provider api-tennis.com. An interruption or an error on its side can leave some information temporarily inaccurate. See the legal notice and the terms of use.