The four models, explained and interactive

A linear-weighted composite of four inputs rated 0–10 — the same shape as a correlation-weighted regression, just with fixed rather than fitted weights. This is the base layer AngelScore builds on, generalized to eleven feature families elsewhere on the site.

Each runner's competitive score becomes a Poisson rate parameter λ — the same family of model used for expected-goals-style sports forecasts. A runner's win share is its rate relative to the field's total rate.

Start 0 (the most recent) always carries full weight. Each start further back is discounted by the decay factor raised to its distance — a lower decay forgets old form faster.

Where models 1–3 are closed-form, Monte Carlo is empirical: it draws randomized outcomes repeatedly and counts what happens. More trials narrow the estimate without changing what it's estimating.