The Ledger

The game model is a Maher (1982) attack–defence Poisson (Maher, Appl. Stat. 31:179–184): each team's goals for/against per game versus league average, multiplied per matchup, home ice +5%. A cross-season carryover prior seeds every team's rates with last season's profile as 18 pseudo-games — the practice FiveThirtyEight used for its club ELOs — to tame early-season noise. Attack rates are blended halfway with shot-volume process and regressed toward career shooting-percentage baselines, the regression logic behind Muffin Regression, rebuilt on official NHL shot counts because expected-goals feeds aren't free to commercialize. Road back-to-backs cost 0.5 expected goals. Most-likely scores are winner-conditioned: tied scorelines are excluded from the pick, because NHL games never end tied.

What we tested and rejected: the Dixon & Coles (1997) low-score correlation and recency-decay weighting both failed our untouched holdout windows and are not in the model. No third-party model or data vendor runs anything here — every number is computed from the NHL's official public feeds by this codebase.

Grading discipline: every parameter is fitted on one date window and only ships if it improves a later, untouched window (currently 108 games). Headline holdout numbers: 59–61% regulation-winner direction; 6.5% exact most-likely score (the world-class ceiling for exact NHL scores is roughly 7–9% — a single goal is irreducible variance). The table above is the live record, not the backtest.