Who’s on the rise?

1. Compare like with like

Each class and event pool is modeled separately. Every scored class result becomes comparisons against the other classified drivers in that event. A win is 1, a tie is ½, a loss is 0. No-time entries, solo classes and duplicate driver entries within one class are excluded.

2. Estimate opponent-adjusted strength

We fit a Bradley–Terry model: P(A beats B) = 1 / (1 + exp(θB − θA)). Our implementation minimizes weighted logistic loss plus ½Σθ². Each comparison weighs 1/(field size − 1), so a driver contributes one unit per event. Regularization pulls short or perfect records toward the baseline. Display score = 1000 + (400 / ln 10) × θ; a 400-point gap means 10:1 fitted pairwise odds. This is separate from F1 points and the legacy field score.

3. Show uncertainty and test forward

We refit 64 times after resampling whole events with replacement and show the middle 90% of scores. Resampling keeps opponents from one race together. Missing drivers in a resample return to the baseline. These are stability ranges, not Bayesian credible intervals; sparse records can remain poorly identified. We also fit on the earliest 80% of event dates and evaluate later results using weighted Brier loss. Test results cover known, connected drivers only and can vary with this single split. Parameters were fixed before this check.

4. Respect the gaps

Drivers without a chain of shared opponents form separate competition groups; we do not predict between them. The model assumes stable ability across its selected history and cannot observe practice, weather, tires, mechanical failures, or unindexed events. Five starts is a display threshold, not proof of reliability. The ranking is an experimental statistical estimate, not an official SCCA championship or an assessment of a person.

5. Cars and win probabilities

The optional driver–car model adds a strongly regularized car term (penalty 2Σβ²). A car label needs at least eight entries and three drivers before receiving an adjustment; otherwise it adds zero. Labels remove the year and normalize punctuation, but do not resolve every trim or spelling. Shared labels are not guaranteed identical equipment. Driver choice and car preparation are confounded: a positive term does not prove the car caused better results.

Field win chances use the Plackett–Luce choice formula, exp(θdriver + βcar) / Σfield exp(θdriver + βcar). It is a scenario conditional on this field and all entries posting comparable scored results. It excludes breakdown/DNF risk, is not calibrated race-winning odds, and must sum to 100% before rounding. Adding a competitor changes everyone’s chance. The car model’s holdout is reported separately.