Model performance
Chronological walk-forward evaluation. For each held-out race the model predicts from ratings available before that race, then updates only after scoring.
Headline backtest (out-of-sample)
Task type: field-event (winner among many). Win Brier is not a clean head-to-head number for a field event, so the headline head-to-head metric is pairwise finish-order accuracy.
Pairwise accuracy
60.7%
coin flip 50.0%
Win Brier
0.97
Naive baseline Brier
0.9736
uniform 1/N
Races scored
72
Primary 路 win log loss
3.577
Secondary 路 Brier
0.9704
Pairwise accuracy context
60.7%
Backtest details
- Races evaluated
- 72
- Pairwise samples
- 50,569
- Winner in model top 5
- 23.6%
- Winner in model top 10
- 48.6%
- Field sigma
- 200
- Pairwise K scale
- 0.06
- Samples are chronological. Binary pairwise samples are oriented by stable driver id, not by winner/finisher.
- Probabilities use only completed NASCAR Cup live-feed fields available from public cf.nascar.com JSON feeds.
- Current forecast reuses the latest known Cup field until a future entry-list adapter is added.