How good are these forecasts?
Every forecast scored on seasons the model never saw while its settings were chosen.
Against the betting markets
Each market's price before the start, scored on the same finished games as the model. A negative figure means the market was more accurate.
| Market | Since | Priced | Model log loss | Market log loss | Model vs market |
|---|---|---|---|---|---|
| Kalshi | 6 Jul 2025 | 44 | 3.012 | 2.522 | −19.4% |
A driver's rating is the driver plus the team's cars that season. NASCAR fields are big and close, so even the favourite rarely has more than a one-in-six chance.
The championship is simulated in the 2026 Chase format: the top sixteen on points after 26 races reset to a seeded ladder (2,100 for the regular-season leader, then 2,075, 2,065, 2,060 and down by five to 2,000) and race the last ten; the most points after Homestead wins. Every remaining race is played with today's ratings, paying 55 for a win, 35 down to 1 for the rest, and 10 to 1 for the top ten at the end of the first two stages.
Track type (superspeedway, short track, road course) matters a great deal and is not in the model yet; a superspeedway win is close to a lottery.
Retirements count as finishing behind every car still running, whatever the cause.
The champion record replays each season since 2019 in its own format (the elimination playoffs until 2025) from the real results so far. It doesn't know about points penalties, playoff waivers or a win stripped of its playoff place (Austin Dillon's in 2024), which only matter where it has to work out who makes a later round.
Questions and answers
How accurate are PodiumOracle's NASCAR Cup Series predictions?
On 107 races from held-back seasons (2024 to 2026), which the model never saw while it was tuned, its favourite won 10% of the time in fields of about 38, and it gave the eventual winner 6% on average. Its log loss was 3.089, against 3.631 for treating every driver as equal and 3.287 for recent form. Against the betting markets on the same finished events, its log loss was 3.012 against Kalshi's 2.522 over 44 races (lower is better).