Our methodology

The model

Court Shark's win probabilities come from a match model trained on completed ATP and WTA matches. It is built from each player's ratings, form on the surface, serve and return statistics, head-to-head record and recent schedule. The output is a probability for each player, and the two always add up to 100%.

A probability describes how often a player in this position wins. A player given 65% is expected to lose about one match in three like it, so a single result never proves the number right or wrong.

Published before play, never changed

The model's number for a match is recorded before the match starts and is never changed afterwards. That is what makes it possible to grade the model honestly, and it is why an article written before a match and one written after it quote the same figure.

How the model is graded

Every week the Tennis Desk publishes model notes: how often the side the model favored went on to win, and how well its probabilities matched what happened. A well-calibrated model's 70% picks win about 70% of the time. The notes show each confidence band with its sample size, because a band with few matches can swing on chance alone. The full record of graded predictions is on the performance page.

The market consensus

Where an article mentions the market consensus, it means the pre-match price for each player converted to a probability. We show it next to the model as context.

Charts and tables

Charts in Tennis Desk articles are drawn from the same data the article cites, and every chart has a data table behind it. They are a snapshot taken when the article was written and do not update afterwards. For live numbers, open the player and head-to-head pages linked from each article.