Net Rating, in full

Our first version carried a per-game context multiplier of our own construction: credit nudged up when a player’s opponents were more heavily used than his teammates, both measured by share of ice time, and trimmed when the opposite was true. It was our reading of the idea that some minutes are harder than others. It was not Dom’s own difficulty measure, which is built differently and rates players differently. We shipped it, flagged the coefficient as hand-set and provisional, and said we’d calibrate it properly. We did — and it didn’t survive the test.

Three separate tests, all pointing the same way. Dropping the multiplier equalled or beat keeping it on every individual predictive target we checked. It improved next-season team goal-differential prediction, .612 to .617, with a bootstrap confidence interval that excludes zero. And it moved the league mean of team rating sums from −5.5 toward the theoretical zero, landing at −2.1. Doubling the multiplier’s strength made every one of those numbers worse — which is the signature of an adjustment that adds noise rather than signal, not one that’s merely under-tuned.

It also failed on inspection. Teammate quality varies over a much wider range than competition faced, and a star’s linemates are fixed for stretches of a season while his opponents average out over it. Our multiplier scored difficulty as opponents’ ice-time share minus teammates’. Nathan MacKinnon’s opponents were heavily used (the 76th to 97th percentile of forwards from 2021-22 to 2025-26), but his linemates were used even more (98th to 100th), so the multiplier ranked his job among the easiest: the bottom 1% of forwards in every season from 2021-22 to 2024-25, and the 5th percentile in 2025-26. It was charging him for having good linemates, a thing his rating already counts as their production, not a debt against his own. That is not a small edge case; it is what the multiplier does to exactly the players it is supposed to be fairest to. Today’s rating uses no such measure; the ice-time pace adjustment below replaced it.

That test was of our multiplier, not of the idea. So we tested the idea again, in September 2026, with a design written down before any result existed. Dom Luszczyszyn shared with us an updated weight set and his own context adjustment. The adjustment splits quality into offense and defense rather than using share of ice time, since a defensive-role player and an offensive-role player can log similar minutes against very different matchups. Teammate quality is measured against the teammate quality expected from how good the player and his team are, and only the difference is used, which strips out what the rating already carries from good linemates. The effects of competition and teammates are fitted game by game (a faint signal, over many games, that weaker competition raises a player’s Game Score) and applied as a per-game adjustment. For past seasons the ratings are run once unadjusted and once adjusted; during a season the adjustment is iterated. We rebuilt that approach as closely as our data allowed. Our reconstruction does not yet reproduce the worked examples we were given, so what follows is a finding about our version of the idea, not about anyone else’s method.

The updated weights lost: they read forwards slightly better and defensemen worse, and the defense cost was the larger. The context question was whether a player coming off a good situation, with strong linemates and soft opposition, is overrated the next season. For forwards he is; for defensemen we found no such effect. Our version of the adjustment rates each teammate and opponent from games without the player in question, and is iterated until the ratings settle. On our own weights it removed about 34% of the forwards’ carried-over credit (90% interval 19% to 97%). But it predicted next season slightly worse, for players and for teams, and it agreed slightly less with a competition-matched measure of how players did against opponents of similar quality. In the first round, on the updated weight set, a single-pass version built the other way, with context rated from other nights rather than other players, removed only about 16%, so the size of the effect depends on how context is measured. So the rating keeps the ice-time pace adjustment below and carries no quality-of-competition adjustment.

A player’s expected on-ice chances against now scale with his own 5v5 ice time. That is distinct from the per-team pace blend above: that one anchors a player’s baseline to his team’s scoring rate; this one anchors his defensive baseline to his own deployment. Bigger minutes are faster minutes: top-line players face other top lines, fourth-liners face other fourth-liners, and the chances-against expectation should reflect that before we count what a player did about it.

The whole argument is that the effect being adjusted for is the environment, not the player. Using player fixed effects across those 9,423 player-seasons, when the same player’s ice time changes, his on-ice xGA(sh)/60 relative to team moves by 88% (forwards) to 100% (defensemen) of the gap you’d see comparing different players at different ice times. An independent event study on the ~870 players whose ice time moved 1.5+ minutes within the same team gives 0.75–1.05 — the same range, from a different design. The constant is measured from within-player variation, not inferred from comparing different players who might simply differ in ways the model doesn’t see.

The chances-for side is deliberately left untouched. Only about two-thirds of the ice-time-vs-pace relationship on offense is environment; the rest is that better offensive players earn the bigger minutes in the first place. Adjusting for it would take away credit those players earned. Offense stays exactly as it was, bit for bit.

Caveats, stated plainly:

• Defensemen come out slightly over-corrected. Their within-player slope is a shade steeper than the across-player one used to set the constant, and it shows: Def now correlates +0.13 with ice time for defensemen, versus 0.00 for forwards. That’s within noise, and we’re watching it rather than re-fitting on a single observation.
• The constant is refit annually on the pooled panel, not per season — a per-season refit adds more noise than it removes. Leaving any single season out of the panel moves the constant by about 10%, which is the honest size of its uncertainty.
• Tied to the current xG model. Both percentages above were measured against the shooter-adjusted xG model as it exists today. An xG model change requires remeasuring them.