How our tennis prediction model works
Every probability on TennisTips comes from two sources: our own player ratings and the prices offered by bookmakers. This page explains both, shows how the model performed in a full-season test and lists what it cannot see.
1. Surface-blended Elo ratings
Each player has an Elo rating that goes up after a win and down after a defeat. The size of the change depends on how surprising the result was: beating a much stronger opponent moves the rating more than beating a weaker one.
- Overall rating built from every singles match a player has played.
- Surface ratings kept separately for hard, clay and grass courts.
- Blend: for a given match we combine the overall rating and the rating for that surface 50/50.
- Shrinking K-factor: a newcomer's rating moves quickly, and the adjustment gets smaller as a player builds up more matches, so established ratings are more stable.
Ratings are built from ATP, WTA and Challenger singles results (men's and women's Challenger level).
2. Market odds
We collect prices for the same match from several bookmakers. Each price includes the bookmaker's margin, so we convert the odds to implied probabilities and remove the margin so the two players add up to 100%.
3. Combining the two
The market-based probability is the base, and the Elo model is combined with it. The result is the probability you see on each match page, together with the fair odds that probability implies.
Backtest: one full season
Period: 1 November 2025 to 19 November 2026. We included 17,945 matches in which both players already had at least 10 earlier matches in our data. Every prediction was made using only information available before the match.
The model's favourite won 63.1% of these matches.
| Tour | Favourite won |
|---|---|
| ATP | 64.3% |
| WTA | 64.6% |
| Challenger (men) | 62.5% |
| Challenger (women) | 60.6% |
| All matches | 63.1% |
Calibration
A good model is not only right often; its percentages should mean what they say. If it gives a player 70%, players in that range should win about 70% of the time. The table groups matches by the probability the model gave its favourite.
| Favourite probability | Matches | Predicted | Actual |
|---|---|---|---|
| 50-60% | 7,430 | 54.9% | 54.8% |
| 60-70% | 5,779 | 64.6% | 63.3% |
| 70-80% | 3,298 | 74.3% | 73.0% |
| 80%+ | 1,438 | 85.1% | 82.3% |
In the higher ranges the model was slightly overconfident: its favourites won a little less often than predicted.
What the test told us honestly
In the same test, the bookmakers' closing prices were more accurate than the Elo model on its own. That is why we use market prices as the base and add the model on top, rather than the other way round. It is also why we do not promise profit.
Our public pick record starts from the day the site launched. No pick is ever deleted, whether it won or lost.
What the model is not
- It does not see injuries or withdrawals announced shortly before a match.
- It does not account for weather, conditions on the day or a player's motivation.
- Its probabilities are estimates, not guarantees. A 75% favourite is expected to lose roughly one match in four.
Betting is for adults aged 18 and over. Read our responsible gambling page and our FAQ for more.