METHODOLOGY

Player value, penalty killing and the historical PP team estimate are added in goals. Each piece is shrunk once, where its model estimates the repeatable signal.

PP Team measures scoring above a player-and-opponent baseline in separate games. Games used to evaluate the estimate are excluded from the player and team fits. Three reconstructions are averaged, using the same method from 2021–22 through 2025–26.

Less information means more shrinkage toward no team advantage. The amount is selected using the other seasons. The historical goals already include this shrinkage; the display applies no additional PP multiplier. Player values and penalty-kill shrinkage retain their own methods.

The team contribution is uncertain. Its improvement on held-out games is small, and the measured range includes no improvement. Values depend on the player baseline and the game partition. This is an estimate of historical contribution, with no separate causal claim about coaching.

Win probabilities come from a separate game model, never from adding up GAR. For every game it builds each club's expected goals at five-on-five and on the power play from its expected lineup's player value, its penalty-kill rating and how much power-play time it draws, then turns the two totals into a regulation win, an overtime, or a regulation loss.

• Expected lineup: the skaters a club has dressed in its last ten games, each weighted by how often he played and his usual share of the ice time. Before a club's first game, its roster with last season's usage.
• Player value moves with the season: last season's projection blended with this season's play, weighted toward this season within a few games.
• Overtime and shootouts split close to even (49% to the home side, measured).
• Standings: every unplayed game simulated 20,000 times; each simulation also draws a season-long strength error per club, so the ranges are not falsely narrow. Ties break on points, regulation wins, regulation plus overtime wins, then wins.
• Tested on 2024-25 and 2025-26 using only what was known on each date: final points off by 9.57 on average (root-mean-square) against 10.08 for a regressed points pace; game log loss 0.6748 against 0.6832.

The numbers the display relies on, restated from the frozen model spec with provenance.

• Qualification: 20 GP and 200 even-strength minutes (631 of 684 skater-seasons qualify (2 GP, 22 even-strength minutes — floors scale with the season played))
• Special-teams display floor: 50 PP minutes over a full season, scaled in-season
• On-ice divisors: 5 skaters (PP), 4 (PK)
• PP replacement offset: -0.023 on the projection scale; PK: 0
• G-model: ridge on shift-level ratings; fitted on 2021-22 through 2025-26
• Penalties: one minor = 0.186 net goals; replacement per 60 is the season's league rate for the position plus a fixed offset — drawn F −0.031 / D −0.035, taken F +0.008 / D +0.021; per 60 of all regulation minutes
• Expected goals: MoneyPuck.com's shot model. MoneyPuck rebuilt it for 2026-27 (gradient boosting on 2023-24 through 2025-26 shots) and did not re-score earlier seasons, so 2026-27 chance ratings use the new model while 2021-22 through 2025-26 use the old one, and the five-on-five weights were fitted on the old one. The new model spreads chance value more evenly — fewer very high-danger shots, more value on ordinary ones — so this season's chance ratings can lean slightly toward shot volume. Five-on-five values carry part of last season into this one, so the two models also mix early in the season.

A model is also the things it refused to do.

• Individual goalie ratings — save results did not follow goalies across moves; at team level, save percentage is a club effect and is measured as one.
• Individual PK goal-prevention ratings — the estimate did not repeat across seasons; the zero row is the result.
• A club-level penalty term — after the skaters' projected drawing and taking are removed, what is left does not persist season to season.
• Venue, score-state and opponent adjustments to penalty rates — none improved the next-season prediction; the flat rate ships.
• Crediting a PP unit's goals to its "driver" — permanent linemates make the attribution unidentifiable.
• Win probabilities from summed player value — rates track goals; the game model above does that job, from expected goals by game state.
• Radar charts for style — axis order and enclosed area distort comparison; the percentile bars print the raw values instead.