How accurate is the Formuradar F1 forecast?
A public test, race by race and without seeing the future. The figures are recalculated automatically; probabilities are estimates, not certainties.
In short
- With the grid already known, the model gave the real winner 42.2% probability on average, over 149 Grands Prix since 2020.
- “Whoever starts first wins” gave 42.9%; an equal split, 4.9%.
- Winner error (log-loss, lower is better): 1.272 for the model vs 1.518 for “whoever starts first wins”. The average probability is similar; the difference is in how the rest of the probability is spread.
- The model clearly improves on the simplest baseline.
- Test last updated: .
How well does it call the race winner?
After qualifying, with the grid already known. “Average probability on the real winner” is better the higher it is; the error (log-loss) and the top-10 error (Brier) are better the lower they are.
| Method | GPs | Avg. probability on real winner | Winner error | Top-10 error |
|---|---|---|---|---|
| Formuradar model (grid already known) | 149 | 42.2% | 1.272 | 0.1505 |
| “Whoever starts first wins” | 149 | 42.9% | 1.518 | 0.1648 |
| Everyone equally likely | 149 | 4.9% | 3.007 | 0.2497 |
And season by season?
An improvement visible in a single year would be suspicious (the 2026 regulations changed the cars a lot). That is why it is broken down by season.
| Season | GPs | Formuradar model (grid already known) | “Whoever starts first wins” |
|---|---|---|---|
| 2020 | 17 | 47.0% | 44.2% |
| 2021 | 22 | 40.8% | 38.1% |
| 2022 | 22 | 28.6% | 35.2% |
| 2023 | 22 | 57.2% | 46.6% |
| 2024 | 24 | 35.7% | 37.2% |
| 2025 | 24 | 43.2% | 49.9% |
| 2026 | 18 | 44.9% | 50.7% |
And pole position?
Before qualifying, the model gave the eventual pole-sitter an average probability of 27.7%, over 151 Grands Prix; an equal split would give 5.0%.
The real test: forecasts frozen before each session
There is no hindsight here: each forecast is saved before the session starts and scored afterwards with the real result. It is the one test that cannot be dressed up. The sample is still small (few Grands Prix), so these are hints, not a verdict. Brier error: lower is better.
| Market | GPs | Model error | Baseline error | Better than baseline? |
|---|---|---|---|---|
| Teammate head-to-head | 5 | 0.2027 | 0.2500 | Yes |
| Pole position | 5 | 0.0348 | 0.0340 | No |
| Podium | 5 | 0.0657 | 0.1038 | Yes |
| Race winner | 5 | 0.0259 | 0.0331 | Yes |
Teammate head-to-head: for each pair of drivers at the same team, the model says who will finish ahead. The baseline is a coin flip (50%). Bookmakers rarely offer this market, so here we can only measure against chance.
How is the test done?
For each Grand Prix since 2020, the model is fitted only on earlier weekends and forecasts that weekend; its forecast is then compared with the real result. The baselines are deliberately simple: if the model does not beat them, it adds nothing. The data comes from public sources.
Unofficial fan project. Probabilities are estimates: they are not financial advice or an invitation to bet.