GLOSSARY
Every term on the site, in plain language — the Methodology carries the math.What a skater's number is, how it is priced, and the rows it itemizes into.
A skater's value for one season, in goals. It is built in two steps: the model first projects how many goals per hour his play should be worth next season, above what a freely available replacement would give in the same minutes, then multiplies that rate by the minutes he actually played this season. Read it as this season's ice time priced at next season's expected rate. It does not predict how many minutes he will play next year, so it is a rate forecast priced on real minutes rather than a forecast of his next-season total.
Every rate on this site is an estimate of what a skater will do next season, and it has been scored against the seasons that actually followed (Methodology, chapter A). A season-story number tells you what happened; a projection tells you how much of it to expect again. Projections read smaller than raw season totals for exactly that reason.
Pulling a noisy number toward the average. A hot month of shooting says less about a skater than a full season of shot attempts, so the model keeps the part of each season that tends to repeat and trims the rest — more trimming for thin ice time, less for a full season. It is why projections are smaller than season totals, and why they predict better.
Five skaters against five. A skater's even-strength value splits into an offense row (helping his own line score) and a defense row (holding the other line down). Both are even-strength minutes only.
Scored 5-on-4 only. A skater's PP row is his own projected share of what a league-average power play would do with him on the ice. Whatever a club's power play produces beyond its skaters' shares is a team matter and lives on the team page (see PP residual).
One net row: minor-equivalent penalties drawn, less minor-equivalent penalties taken, each projected from this season's rate and priced at what a league-average power play gains from one minor — about 0.19 goals. Drawing and taking both follow a player when he changes clubs, which is why the row is his. Shown per 60 minutes of all regulation ice time.
The unit penalties are counted in. A two-minute minor is one; a double minor is two; matching minors that cancel each other count for neither side; majors, misconducts and penalty shots count zero; bench and goalie minors belong to the club, never to a skater. Each call is checked against the manpower that followed the whistle.
Every skater's PK row is exactly zero because the model measured individual goal prevention on the kill and found it does not repeat from one season to the next (Methodology, chapter H). The row stays on the page because it is a result. The kill itself is real and appears on the team page as one team term.
The bar a skater is measured above: what a freely available player would give in the same minutes. At even strength it averages two views of the NHL boundary — players who washed out of the league and players in their final NHL season — both priced with the same forecast as everyone else. On the power play nobody outside the league receives those minutes, so the bar is the deployment margin: players moving into or out of power-play time, failed auditions included. It is re-measured on every publish and must sit inside its own band.
Enough ice time for the board: a full season requires 20 games and 200 even-strength minutes, and the floors scale down with the season played so the board works from opening week. Sub-qualified skaters keep their pages; the board, the percentile chips and the style labels require the floor.
The same value as a rate: each row's GAR divided by that row's own ice time, times sixty. Even-strength rows use even-strength minutes, the power-play row uses power-play minutes, and penalties use all regulation minutes — so the rates never add up to one total; the goals figure does that job. Power-play rates need 50 PP minutes over a full season (scaled in-season); below that the cell shows a dash.
The season's GAR spread over games played — the one rate that needs no game state, useful for comparing skaters with very different ice time.
Seasons are written split-year: 2025-26 is the season that opened in the fall of 2025. Every board and page carries the label it was measured on.
What a team page adds up, what it leaves as an empty row, and how a traded skater is split.
A team page's total: every skater's GAR plus the team channels, itemized, summing exactly to the printed figure. What the model cannot measure — goaltending, individual PK goal prevention — stays on the page as a labeled empty row rather than being folded into the total. Teams are ranked on this figure.
A player channel is value attributed to one skater: even-strength offense and defense, the power-play row, penalties. A team channel is value the club produced that no skater's row can claim, and it is measured only after the skaters' projections have been removed — so nothing is counted twice.
The club's power-play value left over after its skaters' projected shares are accounted for — the persistent part of a team's power play that the five men on the ice do not explain. It is shrunk by how reliably it repeats from season to season.
The club's penalty kill as one team term. Because individual PK goal prevention does not repeat, there is nothing to double count: this row carries the whole kill.
Where a club's measured team value stands among the 32. The Teams table opens sorted on the total, and the rank moves only when the total does.
Under a team filter, a traded skater's row becomes the value he produced for that club, attributed game by game from the season's record. The slices add back to his season row exactly; his league rank is kept, and the percentile chips are hidden because they belong to the season total.
The percentile bars, the archetype labels and the usage figures on a player page.
The percentile bars on a player page: how a skater plays at even strength, grouped as creation, finishing, suppression, penalties and usage. Each bar ranks him against qualified skaters at his position, with 99 the best, and the raw figure prints beside it so the bar never has to be trusted alone.
His effect on his own line's rate of shot attempts (shot creation) and of expected goals (chance creation) while he is on the ice, shown as a percentage above or below a league-average line. +8% shots means his line attempted 8% more with him out there.
Primary assists per 60 minutes of even-strength ice time — the last pass before a goal, counted at a rate so a fourth-liner and a first-liner can be compared.
His effect on his line's shooting percentage — how often the line's shots go in with him on the ice. It is a line statistic, distinct from his personal shooting percentage, and it earns a place only for forwards (Methodology, chapter D).
Goals per 60 minutes of even-strength ice time.
His effect on the other line's shot attempts and expected goals while he is on the ice. Fewer is better, so a good defender reads negative: −5.8% shots means opponents generated 5.8% fewer attempts against his line. The bar still fills to the right for good defense, because the ranking runs the way a reader expects.
A shot's chance of becoming a goal given where and how it was taken, summed over shots. It credits creating dangerous chances whether or not the puck went in, which makes it steadier than goals from one season to the next.
Minutes per game in one game state, shown as minutes:seconds. On the board, TOI/gm is all-situations time per game in regulation — overtime is outside the model's scope, so it reads a few seconds to about half a minute under the league's official figure; the style profile shows even-strength, power-play and penalty-kill usage as separate rows.
A power-play or penalty-kill usage row reads 'not deployed' when a skater averaged under half a minute a game in that state. Below that floor a coach has made a yes-or-no decision rather than a quantity, so the raw number prints with no bar and no rank.
The raw rates behind the penalties row, per 60 minutes of all regulation ice time, ranked against qualified position peers. Taken is ranked so that fewer reads better, which keeps every bar in the profile pointing the same way.
The small numbers on the count card: where a skater's value ranks among qualified skaters at his position, 99 being the best. They appear on season totals in both the goals and per-60 views, and they are hidden on a single club's slice of a split season, where the peer group would be the wrong one.
A label for how a skater plays at even strength (e.g. F-Play Driver), produced by a grouping method that places skaters with similar creation, finishing, suppression and usage figures together. It describes a way of playing; his value stays on his player page.
A role label that includes special teams (e.g. F-PP1 Engine, D-Shutdown): how much he plays at even strength, on the power play and on the kill, together with his production and value in each. Built the same way as the style label, it reads mostly as a depth-chart role.
The depth-chart position the model assigned from season-long usage — first line, second pair, first power-play unit — as deployed over the season rather than any given night's lines. A skater with a regular slot is placed; one without is a spare.
The statistical words, each in plain language with the number it produces here.
The stretch of play between substitutions when the same five skaters face the same five. Every attempt, chance and goal is charged to the ten skaters on the ice for it, so the credit and the ice time always come from the same slice of the game.
A skater's running estimate of his effect on his line's scoring (offense) or on the other line's scoring (defense). Before each stint the ten ratings on the ice set an expectation; whatever happens beyond it moves each rating a little. Linemates and opponents sit inside the expectation, so they cannot quietly become his credit. Each season starts fresh from the position average.
Testing a forecast on data it never saw. The model is fit on some skaters' seasons and scored on the others, with all of one skater's seasons kept together; then it is fit on the early seasons and scored on the latest one. A number that only works on the data it was built from is memorizing; a held-out number is predicting.
How much of the season-to-season variation a forecast explains, from 0 (no better than guessing the average) to 1 (perfect). Forwards' even-strength offense predicts next season at about 0.48 held out; defense at about 0.13. Hockey is noisy — one season is a small sample of a skater — so those are strong figures, and the site prints the small ones as honestly as the large.
Whether the forecasts are the right size. Line up what actually happened against what was projected: a slope of 1 means a projected goal was, on average, a realized goal; below 1 the projections were too bold; above 1, too timid. The published model sits at 1.0 within its margin of error for every rate it ships.
How much of a rating survives from one season to the next — the slope of next season's figure on this season's, so 0.6 means 60% of the spread carried forward. It is checked separately for skaters who stayed with their club and skaters who moved: a skill that only carries over on the same roster is a team effect wearing a player's name, and it is removed from his row.
Whether a number agrees with itself inside one season. Split a skater's games into two halves and ask if the halves tell the same story; low reliability means the measurement is mostly noise, and the model shrinks it hard. Individual PK goal prevention scores about 0.14, which is why its row is zero.
The fitting method that turns a skater's ratings into a single forecast. It is an ordinary line of best fit with a penalty that pulls each weight toward zero, so a thin sample cannot produce a wild weight. How hard to pull is chosen by held-out testing with each skater's seasons grouped together.
How an input earns its place: refit the forecast without it and measure what the held-out prediction loses. Inputs are judged on that loss rather than on the size of their weight, because two inputs that move together share a weight between them.
The team test that prices each club's next-season roster using only the projections available a season earlier, so the sum reflects who actually played. Summed skater projections explain 39% of next season's regulation goal differential on that footing (Methodology, chapter I).