How good are these forecasts?
Every forecast scored on seasons the model never saw while its settings were chosen.
Every car is the same Dallara chassis and engine, so a driver's rating is mostly the driver; the team part is what a good engineering crew adds that season.
Only the feature races train the model: the sprint grid reverses the top of qualifying, so its order says more about the draw than about pace. Sprints still count in the standings and the title simulation, which plays every remaining race with today's strengths and the real points (25-18-15-12-10-8-6-4-2-1 for a feature race, 10-8-6-5-4-3-2-1 for a sprint).
Formula 2 is hard to call: every driver has the same car, most are in the series for only a season or two, and safety cars and tyre strategy reshuffle feature races. On the held-out seasons the model beats picking at random but not a plain average of recent finishes; both are on the model page.
Pole and fastest-lap bonus points count in the standings but aren't simulated for races still to run.
Seasons before this one come from Wikipedia's standings grid, which has every finish but not the points per race; those are recomputed from the finishing position.
Drivers who move up to Formula 1 or arrive from Formula 3 change the field each year; a newcomer starts from the average.
Questions and answers
How accurate are PodiumOracle's Formula 2 predictions?
On 40 races from held-back seasons (2024 to 2026), which the model never saw while it was tuned, its favourite won 8% of the time in fields of about 22, and it gave the eventual winner 7% on average. Its log loss was 2.836, against 3.085 for treating every driver as equal and 2.898 for recent form.