Moneyline record

One logistic regression, fitted on priced fights from before 2021-01-01, predicting whether the first corner wins. Every input is a difference between the two fighters, so the model never sees a fighter on their own — only how they compare.

1. market price — the book's own de-vigged closing price
2. Elo — tuned Elo, with win attribution and the promotion prior
3. Glicko — a second rating that tracks how settled it is
4. rating certainty — Glicko's rating deviation: how sure the rating is
5. KO rating — a rating built only from knockouts
6. submission rating — a rating built only from submissions
7. decision rating — a rating built only from decisions
8. age — which of seven age brackets each fighter is in
9. career length — years since the fighter's first recorded bout

The result is then pulled back towards the market by λ = 0.89, because a blend that strays too far from the closing line is usually wrong rather than early.

The rule. Back a fighter when the model's probability beats BetOnline's own de-vigged price by more than 2%. One flat $100, no staking ladder, no discretion. Nothing else is considered — not camp, not injury, not styles, not who is on a run.

Because it is a sum, every pick can be taken apart: each input pushes the log-odds up or down by coefficient × value, and those pushes add up to the prediction exactly. That is what the Why column shows. The market term is the baseline — what the book already thought — so the reasons the model disagrees are the ones after it.