How Player Intelligence works

Skaters

A skater's rating is a percentile against his own position group, forwards or defensemen, this season only. Three components:

• Production (60%). Points per game, shrunk toward the position average for players with few games. A 1.2 points-per-game pace over 8 games is treated more skeptically than the same pace over 60.
• Shot share (25%), bonus-only. The share of shot attempts his team generates while he is on the ice. Tilting the ice can raise a rating, it never lowers one. Why bonus-only: applied symmetrically, on-ice shot share punished elite finishers for their linemates and dragged 60-point defensemen below depth forwards. We tested both. Symmetric lost.
• Deployment (15%). Power-play and penalty-kill minutes per game. This measures what coaches who watch every shift actually trust the player to do.

For the precise: skater shot share counts all attempts (Corsi). The goalie model below uses unblocked attempts (Fenwick), because a blocked shot never reaches the goalie. Different questions, different denominators, on purpose.

The blended score is re-percentiled within position and scaled so the ceiling is 99. Nobody is 100.

Goalies

Goalies are rated on goals saved above expected (GSAx). Every unblocked shot attempt a goalie faced this season runs through the same expected-goals model that powers PuckCast game predictions, built from shot location, type, and the play that preceded it. GSAx is the gap between the goals an average goalie would have allowed on that exact workload and the goals he actually allowed. A goalie behind a leaky defense can rate well. A goalie coasting behind a fortress cannot.

What we rejected, and why

• On-ice goals-for percentage. Correlates 0.70 with PDO, which is mostly shooting and save luck. It punished skaters for their goalie's bad month.
• Individual expected goals. Correlates 0.79 with points production. Adding it doubled-counted the same signal.
• Three-year weighted averages. The standard approach elsewhere, and the most common complaint about public player cards: they lag reality. A player who broke out in October still looks mediocre in March. Our ratings are this season, full stop. The cost is volatility early in the year. We take that trade and say so.
• Plus-minus. No.

Validation

Before shipping, we tested the formula on every season pair from 2010-11 forward: build the rating from season N, check how it predicts season N+1.

• Predicting next-season points per game: the blend beat a points-only rating in 13 of 15 season pairs. This is the hard test. Points-only was scored on its home turf, and the blend still won.
• Predicting next-season on-ice shot share: 15 of 15. Partly expected, since the blend includes shot share. The point is that the signal generalizes across seasons instead of fitting noise.

When a test went against us, we kept the result. An earlier candidate formula with symmetric shot share rated Cale Makar the 13th best defenseman. It did not ship.

What we will not do

• Rate players we have not seen. Under 20 games (8 for goalies), a player is unrated, not guessed.
• Show decimals. A rating of 87.3 implies precision the data does not support. You get 87.
• Blend in reputation. If a former Vezina winner is having an average season, he gets an average rating.