Case study
F1 Podium Predictions
Before each 2026 race I predict the podium from the qualifying result and publish it. There is one notebook per race, so every prediction can be traced and checked against what happened.
- Podium drivers right per race (best model)
- 2.19 / 3
- Baseline: top 3 on the grid
- 2.07 / 3
- Races in the walk-forward test
- 43
- Models compared
- 7
How it works
- 1
Download
Race results from Jolpica through fastf1; weather and practice times from OpenF1
- 2
Build inputs
Grid position, form over the last 5 races, podium rate, team points, circuit history
- 3
Walk-forward test
7 models, each race predicted using earlier races only
- 4
Select
Keep the model with the most podium drivers right per race
- 5
Predict
Apply it to the new grid, with penalties added by hand
- 6
Publish
Notebook, podium card and a public scorecard
Results
| Race | Predicted podium | Actual podium | Right |
|---|---|---|---|
| Mexico City (1 Nov) | Russell, Verstappen, Hamilton | Verstappen, Hamilton, Antonelli | 2 of 3 |
| Brazil (8 Nov), first version | Russell, Norris, Verstappen | Norris, Verstappen, Hadjar | 2 of 3 |
| Brazil (8 Nov), corrected version | Russell, Norris, Leclerc | Norris, Verstappen, Hadjar | 1 of 3 |
Mexico was re-run after the race using only pre-race data. The first Brazil version ignored grid penalties; the corrected run was committed during the race, so it is not a true pre-race prediction.
What I found
- The edge over the grid rule is small (2.19 against 2.07) and I would not call it proven.
- The models mostly agree with the grid, then move a few drivers up or down based on recent form.
- Practice pace and how wet qualifying was did not help. Recent form without the grid was clearly worse.
Limitations
- Penalties are entered by hand from news and race control messages.
- The race-day weather forecast is not an input.
- With about 43 test races, small differences between models are mostly luck.
- Drivers starting far back with strong recent form are where the models disagree most.