Measuring the accuracy of a tennis prediction: what a hit rate does not tell you

A hit rate proves nothing on its own. To judge a tipster or a model, three things are needed: compare it with the market favourite on the same matches, measure it on matches it has not seen, and look at the result in money at flat stakes. Guide published on 05/09/2026.

1. The raw rate measures the market, not the tipster

In tennis, the market favourite wins 68.8% of matches: measured on 220,859 settled singles matches since 2021. A tipster who merely copies the odds therefore already shows "two right calls out of three". A 70ate announced without this comparison says nothing, except that the chosen matches were lopsided.

That is why we publish both figures side by side. Version 1.1.0: 67.39% right winners on 59,960 ATP matches, against 67.07% for the bookmaker's favourite on the same matches. The gap is small, and it is honest: nobody beats the market as a whole, value is decided on the few matches where the model and the market disagree (96.46% agreement, so 3.54% disagreement).

2. Measure on matches never seen

A model that learned on 2022 matches and is measured on 2022 matches is reciting. The rule is walk-forward validation: each season is predicted by a model trained only on previous seasons. Training seasons are excluded from the measurement, or shown separately and labelled as such. Here, 2021 and 2022 were used for training; on the seasons alone unknown to the model (2023 to 2026), the series gives a ROI of 5.27% on 564 bets.

3. The result in money, at flat stakes

The hit rate ignores the odds; ROI includes them. A tipster at 40% on odds of 3.00 makes money (0.40 × 3 − 1 = +20%); another at 70% on odds of 1.30 loses it (0.70 × 1.30 − 1 = −9%). The only measure that matters is the realised expectation: profit divided by the number of bets, at constant stakes. A variable stake (martingale, "staking plan") makes the figure unverifiable, because it can hide losing streaks behind a single large winning stake.

4. The traps that inflate a record

Selection after the fact: showing only winning bets, or deleting bad months. Demand the full, dated history.
Unavailable odds: quoting 2.10 when the market closed at 1.85. Demand verifiable odds, ideally Pinnacle's closing odds.
Small sample: 50 bets cannot separate luck from skill. Under 500 bets, a +10% ROI is consistent with a worthless model.
Shifting scope: measuring on Grand Slams one year, on Challengers the next. The scope must be fixed in advance.
No point of comparison: a rate without the market's rate on the same matches.

The five-point grid

At least 500 bets, on a scope announced in advance.
Verifiable odds, with the closing odds alongside.
A flat stake, and the ROI computed on it.
The full history, losses included, dated.
The market favourite's rate on the same matches, next to the announced rate.

Without any one of these five points the figure cannot be verified, and an unverifiable figure is worth nothing. Our measurement, version by version, with its limits.

Frequently asked questions

Is a 70% hit rate good?

It means nothing on its own. In tennis, the market favourite has won 68.8% of matches since 2021: a tipster who copies the odds already shows that rate. You have to compare on the same matches.

How accurate is the Valuebetennis model?

Version 1.1.0 picks the right winner in 67.39% of cases on 59,960 ATP matches since 2021, against 67.07% for the bookmaker's favourite on the same matches. Measured on 31/07/2026.

What is walk-forward validation?

Each season is predicted by a model trained only on previous seasons. Seasons used for training are excluded from the measurement, otherwise the result is flattering by construction.

How do you evaluate a tipster?

At least 500 bets, odds verifiable at the close, flat stakes, the full history including losses, and dates. Without any one of these five points, the figure cannot be verified.