How the numbers are made

The work behind the site's core numbers, in short: how each is built, what was left out and why, and how it was tested

Expected goals (xG)

How likely a shot was to go in, judged from shots like it.

  • Histogram gradient boosting trained on 1,125,610 unblocked shots, every season from 2016-17 to 2025-26
  • 32 inputs per shot: where it was taken, its angle and type, the play before it and how soon, the score, strength, period and clock, the shooter's position and hand
  • Penalty shots and shots at an empty net valued by models of their own
  • Blocked shots carry no value: the league records where a shot was blocked, not where it was taken
  • Who shot and who was in goal left out, so finishing and goaltending can be measured against it
  • Tested on seasons it had never seen: each season scored by a model fitted only on the seasons before it
  • Older timing withheld: until 2023-24 the league's clock gave the result away, so for those seasons the model never sees it

Competition

How strong the opponents and team-mates a player shared the ice with were, and whether he faced the other team's best.

  • Built from about 480,000 five-on-five stints in a recent full season, each a stretch of play in which nobody comes on or goes off
  • Each opponent and team-mate counted for the time shared with him, goaltenders left out
  • Rated on the share of ice time their coach gave them and their share of expected goals, unadjusted, so the answer is not fed into the weights
  • Before each game every team's skaters are ranked into line- and pair-sized tiers by how their coach had used them, so a top line is that team's own
  • Why tiers: within one team the league-wide average barely moves from player to player, while the share of ice against the opposing top line varies widely
  • Home and road games added together, though the home coach has the last change and so picks the matchups

Adjusted rating

How much a skater adds to his team's five-on-five chances, for and against, once his team-mates, his opponents and his deployment are allowed for.

  • A ridge regression over every five-on-five stint on the site, about 480,000 in a recent full season, each crediting the five attackers and the five defenders
  • Home ice, the zone of the opening faceoff and the score allowed for, so a checking line is not charged for where it started
  • Offence and defence rated apart, each pulled towards last season's rating, how hard chosen by testing on games it had not seen
  • Judged on players who changed teams: their rating holds up across the move better than their on-ice share of expected goals does
  • Recomputed as of every date a game was played, from earlier games only, so no game is rated with its own future
  • Five-on-five only, by measurement: a power-play unit is nearly always the same five players, so a regression cannot tell them apart

Goaltending (GSAx)

How many goals a goaltender kept out beyond what the shots he faced were worth, with a season's luck allowed for.

  • Each shot's expected goals set against what went in, missed shots counted, shots at an empty net and the shootout left out
  • Each shot judged against shots of its own season, strength and rebound type, so a season in which more shots went in lifts or sinks nobody
  • Pulled towards last season by the shots faced: a starter faces about 2,200 unblocked shots a season, and even that many leaves a season's figure carrying luck
  • Regular season and playoffs pooled, because a playoff run alone is too few shots to read
  • The pulled reading predicts a goaltender's next season better than the raw one, though only about as well as calling every goaltender average
  • The raw difference, which other sites call GSAx, is shown beside it as GSAE

Faceoff skill

How much more often a taker wins his draws than an average taker would against the same opponents.

  • Every draw modelled as one taker against another, about 79,000 in a recent full season, so beating a strong taker counts for more than beating a weak one
  • Home ice, the zone and the strength on the ice allowed for before any skill is credited
  • Every zone counts, the neutral zone included: a taker who wins at centre ice tends to win in the end zones too
  • Pulled lightly towards last season's reading, since a regular's own draws decide most of it; how hard chosen by testing on games it had not seen
  • Stated as how often he wins, not how many draws he wins, so taking more draws ranks nobody higher
  • Tested on the next season: the pulled reading predicts a taker's next-season draws better than his plain faceoff percentage does
  • Only who won is recorded, not where the puck went, so a clean win and a scramble count alike, each worth a sliver of an expected goal
How the numbers are made | Primitive Stats