WHAT IT TRACKS

HOW GRADING WORKS

Every call has three options. Each option carries a win probability: the chance your team comes back and wins if you make that call at that moment, with a tie counted as half a win (overtime and shootouts are close to coin flips). Your grade asks one question: how much of your team's chance did your call give up compared with the best option on the board?

GRADE	SHARE OF YOUR CHANCES GIVEN UP
✓ GOOD CALL	2% or less (or it is the best option).
≈ FAIR	2% to 6%.
! A BIT COSTLY	6% to 15%.
✗ COSTLY	More than 15%.

We grade on the share, not raw points, because the stakes change with the score. Down one with two minutes left, a team wins about 7% of the time; down two, about 2%. A call that throws away a sixth of your chances is costly either way. The cut-offs were set from the spread of real options across this season's games.

Grades judge the decision, not the outcome. A good call can still lose, and a bad call can still work out. What actually happened after your call is shown, but it doesn't change your grade.

Some calls show a gray square: no grade. Right now that's the last-unit call (six forwards, or five and a defenseman). The data we use can't tell those apart yet, so we show you the options and the real coach's choice without pretending to know the answer.

WHERE THE WIN PROBABILITIES COME FROM

A model fit on every regular-season game from the last three NHL seasons (2023-24 through 2025-26, 3,936 games). From the trailing team's side, in the last five minutes of the third period, it measures:

• How often goals go in for and against, five on five and with the goalie pulled, down one and down two, in the last two minutes and before that.
• What a faceoff is worth: the chance of a goal either way in the 30 seconds after a draw, by zone and by who won it.
• How long the next whistle takes once the puck drops (about 40 seconds, typically), which prices "pull him at the next whistle."

Each option is a plan: when the goalie leaves, and the draw it starts with. The model plays out the rest of regulation second by second for that plan and adds up the chance of tying. It's an exact calculation, not a random simulation, so the same game always gives the same numbers.

For faceoffs, each taker's chance comes from his season faceoff record plus tonight's, pulled toward 50% because faceoff numbers are noisy in small samples, then matched against the actual center across the dot.

DECISION QUALITY AND THE COMPARISONS

Your game also gets a decision quality score from 0 to 100. Each call scores 100 at the best option and drops to 0 once it gives up 30% of your chances, so it lines up with the grades: good 93 or better, fair 80 to 93, a bit costly 50 to 80, costly under 50. The game score averages the calls, weighted by what was at stake on each one (the gap between the best and worst options), so big calls count for more than small ones.

• The real coach is scored the same way. When his move matches one of the options, he gets that option's value. We judge his decision, not his luck with how fast a whistle came.
• Today's players: how many played, the average and best scores, and the share who picked each option. Counted plays only.
• Simulated callers and real coaches: we score a coin flipper, an always-cautious caller, an always-aggressive caller and the real NHL coaches across every qualifying game this season, and tell you where your score would rank.

REAL VS STAGED

REAL, FROM NHL DATA

• The game, the teams, the score and the venue.
• Rosters and jersey numbers.
• Each player's stats as of the moment you take over: goals, assists, faceoffs and time on ice so far.
• Every play after each call: the time, the players involved and where on the ice it happened.
• Who was on the ice for each play, from the league's shift data.
• When the goalie was pulled and when he went back in, and which timeouts had been used.
• The real coach's call at every decision, and the final result.

MODELED

• Win probabilities and therefore grades: measured from three past seasons, as described above. A model is an estimate, not a certainty.

STAGED

• Where skaters stand between plays. The NHL records where each event happens, not where every player is at every moment. The rink places players sensibly around the real event locations.
• Legs. The meter is a simple formula on time on ice so far, not a measure of real fatigue.
• The option list. We write the three options for each call. The real coach's choice is always one of the real things that happened, but the alternatives are ours.

DATA SOURCES

All game data comes from publicly available NHL sources:

• NHL play-by-play feed: every event, with times, players and locations.
• NHL box scores: player stats and goalie numbers.
• NHL shift charts: when each player was on the ice.
• NHL game reports (the official HTML play-by-play and time-on-ice reports): used to cross-check who was on the ice, and as a fallback when shift data is missing.

When two sources disagree about who was on the ice, we check them against each other before a game is published. Pull the Goalie isn't affiliated with the NHL, and the league doesn't review or endorse how we use this data.

HOW THE DAILY IS PICKED

Each morning a script looks at the previous night's finished games and looks for a team that was trailing late. It favors games with more real decision points, one-goal deficits, chances at the end and a short wait before the action. Games where the trailing team never pulled its goalie, or that don't have at least two real decisions, are skipped. If nothing qualifies, no new daily is posted and the current one stays up.

KNOWN LIMITATIONS

• Your calls don't change what happens. The play you watch is the real game. We grade your choice against the options; we don't simulate an alternate game.
• Your six doesn't play. The players you pick shape your options, like who can take a draw, but the real game uses the real coach's players.
• The model is league-average. It knows the score, the clock, the strength and the faceoff, but not how good these two teams or goalies are. A great power-play team pressing six on five is treated like any other.
• Some calls aren't graded yet. Unit makeup (forwards versus a defenseman on the point) needs on-ice data across seasons that we haven't built yet.
• Only trailing teams for now. Every daily puts you on the team chasing the game. Protecting a lead, line matching and defensive calls are planned.
• No penalties, challenges or timeouts as calls yet. They appear in the play-by-play when they happened, but you don't get to make those decisions yet.
• Faceoff samples. Tonight's draws can be just a handful, so taker odds lean on the season record and are pulled toward 50%.
• Data quirks. League feeds occasionally have timing errors in shift data. We cross-check, but a player might rarely show up on the ice a few seconds early or late.

Spot something wrong? Email hello@pullthegoalie.app with the daily number and what you saw.