Penalty kill

Step 01 · One shot

Pricing a single shot

Start with a single shot: a missed shot by Nino Niederreiter at P2 9:13 of game 2025020017, DAL at WPG. Net goals never asks whether a shot went in. It asks two questions: how likely was this shot to score, and how much would a goal have mattered at that moment?

How likely was it to score?

The expected-goals model is the same one the whole site uses; the full method is on Net Goals 101. Recomputed here from the stored tables: 0.10307, against 0.10307 in the model’s shot ledger.

How much would a goal have mattered?

At P2 9:13 Lindell’s team was up 1, short-handed. The win-probability table says a goal right then would have cut his team’s chance of winning by 21.05 percentage points. The shot is worth its chance of scoring times that swing:

One hundred WP points is one whole win. The same shot in a 5-1 game would have been worth almost nothing, which is why padding numbers in garbage time does not work.

Step 02 · One shift

Actual minus expected, second by second

Now widen to the whole shift the shot came from: P2 8:23 to P2 10:20, 117 seconds on the ice, 117 of them on the penalty kill. Every penalty-kill shot taken while he was out there is priced the same way:

That is the actual side. The expected side is what an average lineup would have produced in the same seconds. Every second is filed into a context cell by score, clock, skaters on each side and whether a faceoff just happened, and the league’s expected-goal rate in that cell (all sixteen seasons, both teams) is priced with the same win-probability swing as a shot:

Lindell took none of these shots himself. The on-ice components do not care who shot: everyone on the ice shares the result.

The penalty kill adds one correction. The context cell describes the team; on the kill it hides who drew the hard seconds. A killer who jumps on during a clear faces far less than one who starts on a defensive-zone draw or is stuck out for a minute, so each PK skater-second’s expectation is scaled by how his stint began and how long he had been out (shots-against rate on the fly in the first ten seconds: 0.28× the team rate; starting at a faceoff: 1.42×; past a minute: about 1.77×). Here that was worth +2.459.

All five skaters on the ice receive the same credit for the shift. Which of them made the play is handled later, by the Playmaking and Suppression components.

Step 03 · One game

Every shift of Lindell’s best game

Add up every shift. This was his biggest penalty-kill game of the season (DAL at WPG, final DAL 5, WPG 4); the highlighted row is the shift above. Shifts with no penalty-kill time are left out.

Corr. = the re-splits applied to shots against: a shot within ten seconds of a defending-team giveaway is charged 80% to the player who gave the puck away; one within ten seconds of a bad line change is shared with the skaters who just left; on the kill, the shift-phase correction.

Step 04 · The season

82 games added up

Do that for all 82 of his games. Each bar below is one game, best to worst: 54 positive, 28 negative. Summed, they are +214.3 WP points over 22,615 penalty-kill seconds (377 minutes).

Step 05 · Re-centring

Removing the league’s tilt

The context table is pooled over sixteen seasons, so in any one season the league as a whole does not net to exactly zero in each column. In 2025-26 every skater’s penalty-kill value, added up, came to −2662.8 WP points over 3,833,551 skater-seconds: a tilt of −0.0006946 points per second. Each player gives back his share of it, in proportion to his own exposure:

After this the league-wide penalty-kill value is exactly zero in 2025-26.

Step 06 · Into goals

From win probability to goals

Win-probability points are converted to goals by asking the standings what they are worth. A regression of actual team points on the summed inputs of each team’s roster gives 34.95 standings points per win of on-ice priced flow per game. One goal per game over an 82-game season is 82 goals, about 27.33 standings points at six goals per win, so:

That is far below the face value of a win (about six goals) on purpose: every on-ice second is credited to five skaters at once, so a roster’s summed on-ice value counts each event about five times over. The regression learns the marginal value of one player’s share.

Fitted on completed team-seasons each time the chain runs: points = 91.62 + 34.95·on-ice wins + 25.33·finishing + 61.16·penalties + 26.72·GSAx (per game).

Step 07 · Reallocation

Paying back the suppression share

The on-ice number above treats all five skaters alike. Suppression then moves the part of the shot quality he and his partners allowed that can be traced to particular players to those players. That credit is moved, not created: every skater on the ice for those attempts pays his share back out of this column. For Esa Lindell the season’s debit was +0.1664 goals; it too is re-centred on the league (+0.000002 per game × 82 games):

Step 08 · Position

Measured against the average defenceman

After every other step, each component is re-centred once more, separately for forwards and for defencemen, so the average defenceman in 2025-26 is exactly zero in every column. The games-weighted mean penalty-kill value of all defencemen that season was +0.00145 net goals per game, so 82 games of it come off:

Step 09 · Result

Per 84 games

Ratings are quoted per 84 games, one full schedule, so a part season and a full one sit on the same scale.

✓ Recomputed +3.06; his 2025-26 row on the player card and the Players table shows +3.1. They match.

Step 10 · Context

How unusual is this?

Skaters with 40 or more games in 2025-26 (612 players). Esa Lindell ranks 1 of 612.

How much of a season like this is skill? The model estimates, from how much players’ own rates bounce between seasons versus how much players differ, that a career rate of penalty-kill value is half signal and half noise after 384 games. Season rows like this one are shown exactly as they happened; careers and the card’s Regressed view are pulled toward average by that amount.