Methods & limitations · reviewed 18 October 2026
How forecasts and fan picks work.
Results and missing information
Results include their source and update time. If an update cannot be verified, the previous result remains available. Missing values are shown as unavailable.
What the model predicts
Pre-weekend forecasts use recent race form. Grid-confirmed forecasts use starting positions. An experimental model also considers qualifying, team and championship results.
Forecasts estimate the official classification order among starters. Retirements keep their position in the published result; drivers who do not start, withdraw or fail to qualify are excluded.
Historical archive reconstructions were created after the races and may include later corrections to grids, results and entrant lists. They do not show what a forecast published before those races would have predicted.
The evaluation uses 87 race weekends. Historical update times are approximate, and the probabilities have not been shown to be calibrated. Retirement risk and circuit-specific strengths are not modelled.
Technical model details
A regularized Plackett–Luce model fits trailing race, qualifying, team and points-order features. Grid position is added only to the grid-confirmed research stage. Preprocessing and regularization are fitted on earlier chronological blocks. The last block is held out. Seeded full-field permutations produce coherent distributions; every entrant appears once.
Race-block intervals, rank metrics, winner probability scores and calibration summaries are recorded in the machine-readable evaluation.
Measured baseline comparison
| Stage | Learned model error | Published baseline | Baseline error | Test weekends |
|---|---|---|---|---|
| Pre-weekend | 3.974 | Recent form | 3.945 | 18 |
| Grid-confirmed | 3.525 | Starting grid | 3.340 | 18 |
The learned model did not beat the selected baseline. F1Fix publishes the baseline and keeps the learned model experimental.
Voting and the deadline
Pick one driver to win each race ballot. No account is required. One IP address has one saved pick per ballot; choosing another driver replaces it. The results show each driver's share of those picks. The list starts in model order, then follows pick totals; model order breaks ties.
People sharing an IP address share one pick, and changing networks can allow another pick. This is a casual fan poll. Read how voting records are stored in our privacy policy.
Pre-weekend voting closes five minutes before first practice. Grid-confirmed voting closes five minutes before the race. Picks cannot change after the deadline, even if result processing runs later. A changed entrant field or cancelled event voids the affected ballot.
The last accepted pick and its matching forecast are saved at the deadline. Earlier account-based comparison archives use their original methodology.
Circuit maps
The Austin circuit outline comes from OpenStreetMap. Its colours group corners for this guide.
© OpenStreetMap contributors · ODbL 1.0 · Download source data
Circuit map source details
The approximate outline uses all 347 coordinates from 29 ordered ways in OpenStreetMap relation 6537729, version 13. Local metric projection, uniform scaling and SVG Y inversion preserve its shape. The computed length of 5,502.23 m is compared with the official 5,513 m reference as a size check; it is not a certified survey. The coloured groups are not official timing sectors.
Sources, terms and credits · Privacy and participation policy