NinetyNinety vs Polymarket
Polymarket's football prices are some of the best forecasts around. Ninety shows them on every match, and adds its own: on 619 matches of 2025/26, Ninety's best estimate (90Machine plus the market) beat Polymarket on its own and Pinnacle's closing line.
The numbers
619 matches of 2025/26, five big leagues, graded on the result.
| Forecast | Probability given to the result | Log loss |
|---|---|---|
| Ninety best estimate (90Machine + market) Best | 38.27% | 0.9605 |
| Polymarket | 38.25% | 0.9611 |
| Pinnacle, closing | 38.15% | 0.9636 |
Why combine them
A prediction market and a football model see the game differently. When they disagree, each has information the other is missing. Blending the two beat both, measured 30 minutes before kick-off, the exact moment the app updates.
See it on every match
Each match in Ninety shows the market's win chances, 90Machine's own forecast and where they differ, free, with no account needed.
How it is measured
Each match is graded by how much probability a forecast gave to what actually happened: home win, draw or away win. That is log loss, the standard measure for forecasts, and it punishes confident misses. Brier score, the average squared error, tells the same story. Higher "probability given to the result" is better; lower log loss and Brier are better.
Nothing is graded on matches the model was fitted on. Every forecast here came from a model trained only on earlier seasons, the same way it runs live. Since late October 2026 the 2025/26 season has also been used to decide which versions ship, so it is no longer an untouched test; the six-season figure is the fairer long-run view. The 2025/26 tables were measured with 90Machine 1.7; 1.7.1 changed only its calibration.
90Machine 1.7.1. Updated 4 November 2026. Forecasts are information, not betting advice.
Questions
How does Ninety compare with Polymarket?
Ninety's best estimate (90Machine combined with the market price) beat Polymarket and Pinnacle's closing line on 619 matches in 2025/26, measured exactly as the app shows it, 30 minutes before kick-off.
How is accuracy measured?
With log loss: how much probability a forecast gave to what actually happened, punishing confident misses. It is the standard way to compare forecasts. Brier score (squared error) tells the same story. Each season is forecast by a model trained only on earlier seasons, so nothing is graded on matches it was fitted on.