How good are these forecasts?
Every forecast scored on seasons the model never saw while its settings were chosen.
Held-back seasons
| Forecast | Forecasts | Log loss | Favourite won |
|---|---|---|---|
| The model | 9271 | 0.628 | 63% |
| Base rate | 9271 | 0.693 | 50% |
| Recent form | 9271 | 0.674 | 58% |
Against the betting markets
Each market's price before the start, scored on the same finished games as the model. A negative figure means the market was more accurate.
| Market | Since | Priced | Model log loss | Market log loss | Model vs market |
|---|---|---|---|---|---|
| Bookmakers (Pinnacle, DraftKings, bet365) | 2 Feb 2026 | 1936 | 0.645 | 0.606 | −6.4% |
| Polymarket | 13 Oct 2025 | 1801 | 0.644 | 0.611 | −5.4% |
| Kalshi | 22 Jun 2025 | 1789 | 0.647 | 0.598 | −8.2% |
Every match is on neutral ground: tennis has no home side in the model, though a home crowd can matter at Davis Cup and some tournaments.
A win moves the ratings more the more games it was won by (the set scores, from the archive and ESPN): on the held-back seasons that beat a rating that only knows who won.
Matches ended by a retirement before a set was finished, and walkovers, are left out.
Best-of-five matches at the majors favour the better player more than best-of-three; the model doesn't separate them yet.
Injuries and withdrawals are not known to the model until they show in results.
Tournament winner odds (the draw simulated) are not on the page yet.
Questions and answers
How accurate are PodiumOracle's ATP predictions?
On 9,271 games from held-back seasons (2024 to 2026), which the model never saw while it was tuned, the side it favoured won 63% of the time. Its log loss was 0.628, against 0.693 for the base rate and 0.674 for recent form. Against the betting markets on the same finished games, its log loss was 0.645 against the bookmakers' 0.606 over 1,936 games, 0.644 against Polymarket's 0.611 over 1,801 games and 0.647 against Kalshi's 0.598 over 1,789 games (lower is better).