Published forecasts · inspectable evidence

What the forecast record shows.

Live results, retrospective reconstructions and records with unknown provenance are kept separate. Each stage and model version has its own sample.

Live prospective scorecard

Stored probability and interval outputs

Podium reliability counts

These fixed bins are a descriptive summary. Sparse bins and lower probability scores do not establish calibration.

Classification revisions

A corrected classification adds a new evaluation linked to its predecessor. The original forecast and earlier score remain available. Only the current final or corrected revision enters the scorecard.

Metric and publication conventions

Expected-position MAE averages the absolute difference between each stored mean finish and official position, first within a race and then equally across races. Projected-rank error scores the ordering separately. Ties in expected finish use stable driver IDs.

Winner Brier is the sum of (p − y)² over the issued race outcome universe, averaged across races. Winner log loss uses the issued probability of the actual winner. The evaluation-only clipping floor is 0.000001; it never changes a published probability. Podium Brier and log loss are binary scores, averaged across drivers within each race and then across races.

Position intervals include both endpoints. Coverage is the fraction containing the official position; width is upper minus lower position. Discrete 10–90% quantiles need not have exactly 80% coverage. Retirements and disqualifications retain the provider’s official classification order for position error. Disqualified drivers cannot count as winners or podium finishers. Nonstarters, withdrawals and DNQs are excluded from the starter target. If the resulting field differs from the issued universe, the race is unavailable for aggregate evaluation; probabilities are never silently renormalized.

The frozen baseline is recent form before the weekend and starting grid at grid-confirmed. It uses the same issue time, input cutoff and entrant universe as the model snapshot. Missing baseline evidence is unavailable, never a zero error. One latest eligible forecast per event, stage and model version enters each scorecard; provisional classifications do not count as completed races.

Historical comparisons use chronological training and validation. The previously reported 18-race block has already been inspected and is not a fresh holdout. New candidates remain separate from the live record; no retuning or sharpening is implied by this page.

Model details · Report a factual or calculation error