How tennis match prediction works

A prediction is a probability, not a promise

When a model says a player has a 70% chance to win, it is not saying they will win. It is saying that in a large number of similar matchups, a player in this position wins roughly 7 times out of 10 — and loses the other 3. A good prediction is a well-calibrated probability, not a prophecy.

That distinction matters because the other outcome is not a bug. Upsets are priced into the number. If favorites at 70% never lost, the honest probability would be 100%, not 70%.

The factors that actually move a match

Baseline reads each match through concrete, measured factors rather than reputation. The main ones:

• Ranking — the starting point: the higher-ranked player wins most matchups, but it moves slowly and rarely tells the whole story.
• Recent form — current shape, a bit more informative than ranking. Arriving red-hot pushes the probability up; arriving shaky, down.
• Surface — one of the most predictive factors in tennis. The same player can be very different on grass, clay or hard courts.
• Rest and match sharpness — too much rest can cost rhythm; too many matches in a row means tired legs. Both effects are weighed for each player.
• Head-to-head — the stylistic nuance of how two games match up. With small samples it counts for little, but it adds at the margin when the record is clear.
• Context — schedule congestion, stakes, deep-run fatigue and similar situational signals that sit on top of the core matchup.

From factors to a single number

The model does not add these up by eye. Each factor carries a weight learned from history — roughly 186,000 past matches — and they resolve into one calibrated probability for the match.

That calibration is checked the honest way: walk-forward, out-of-sample (the model is only ever tested on matches it never saw in training). On that basis Baseline's read lands around 65% accuracy on ATP and 64% on WTA, with a Brier score of 0.216. Each tour quotes its own audited number rather than one flattering headline figure.

Why a good model looks a lot like the market

Betting markets are efficient: the odds already encode ranking, form, surface and most of what a model measures. So an honest model usually agrees with the market, and when it disagrees sharply, that gap is as likely to be the model's own noise as a real inefficiency.

This is why Baseline never promises a betting edge. The value it offers is a clear, calibrated read of each match and full transparency on how it has actually done — not a claim to beat the bookmakers. You can see the running accuracy on the Transparency page.

How to use a probability

Favorite does not mean winner. Read the percentage as a frequency, remember the losing side is a real fraction of the time, and compare the probability against the price before drawing any conclusion — which is what the odds guide and expected value both cover.

You can see the model's read on today's matches on the daily pick page, or run any matchup yourself in Analyze.