How accurate are F1 predictions?
Short answer: it depends entirely on what the number includes. Here is the honest way to read accuracy claims, ours first.
RaceModel: 3.275 positions average error per driver, full field, DNFs included, across 100 races (2022-2026), leave-one-race-out; 68% of podium calls correct (203/300). Every race is graded individually on the accuracy ledger.
The four things that change any accuracy number
- DNFs in or out. Retirements are the biggest single error source: a predicted P3 that classifies P18 is a 15-position miss. Excluding DNFs flatters an error metric by roughly a full position.
- Full field or front-runners. Predicting the top 10 on a processional weekend is much easier than all 20 cars including the volatile midfield.
- Sample size. Five races is a hot streak; a hundred races is a track record.
- Timing and freezing. A prediction made after qualifying, and editable afterwards, is a different claim from one frozen before lights out. RaceModel predictions are frozen pre-race and never edited.
Published accuracy across F1 prediction sites
Taken from each site's own published pages (checked 2026-07-12); we state their numbers as they report them.
| Site | Published accuracy | Basis, as published |
|---|---|---|
| RaceModel | 3.275 MAE, 68% podium (203/300) | 100 races 2022-2026, full field, DNFs included, frozen pre-race, leave-one-race-out; every race graded |
| f1predictor.me | 2.24 mean absolute deviation, 60% podium (1.8/3) | 5 races / 90 predictions (2024 season); site states it cannot predict DNFs and does not model weather |
| f1-predictor.com | none current | ceased regular predictions in February 2024 after seven years |
| podiumprophets.com | no model track record published | community prediction game: users predict, the site scores users, not a model |
| Prediction markets (Polymarket, Kalshi) | self-reported high calibration near resolution | price event probabilities, sharpest close to the race; not a finishing-order forecast |
Editorial "supercomputer predicts" articles (sports media) publish picks but no scored history, so they cannot be placed in this table.
How to read that table
f1predictor.me's 2.24 deviation and RaceModel's 3.275 MAE are not the same measurement: remove DNFs from an error metric and shrink the sample to five races, and the number falls by design, not by skill. On a like-for-like task (full field, retirements counted, 100 races) 3.275 with 68% podiums is, as far as we can find, the only fully published, race-by-race, frozen-before-the-race track record in F1 prediction. If another site publishes one on the same basis, we will happily add it here.
Verify it yourself
The accuracy ledger grades all 100 races individually. Live predictions are frozen before each race on the dashboard and scored on the model changelog. The about page covers the method.