PULL THEHOW THE MODEL WORKS
Every grade in Pull the Goalie comes from one question: from this moment, how often does the trailing team win? Here's how we answer it, and what it says about the oldest argument in hockey: when should the goalie come out?
THE DATA
The model is fit on 3,936 NHL regular-season games from 2023-24 through 2025-26. For every second of the last five minutes of the third period with a team trailing by one or two, it records what happened: goals for, goals against, whether the goalie was in or out, and every faceoff and whistle.
From that it measures how often goals go in, for and against, in each situation. Then it plays out the rest of regulation second by second for any plan you give it, like "pull the goalie at 2:30," and adds up the chance of winning. A tie counts as half a win, because overtime and shootouts are close to coin flips. It's an exact calculation, not a random simulation, so the same game always gives the same numbers.
The details of how individual calls are graded are on What It Tracks.
THE TRADE
Down one with two to five minutes left, here's what the data shows, in goals per 60 minutes of that situation:
| SITUATION | FOR | AGAINST |
|---|---|---|
| Goalie in, five on five | 2.5 | 1.9 |
| Goalie out, six on five | 4.0 | 11.2 |
Pulling the goalie raises your scoring rate a lot, and raises the other team's far more. That looks like a bad deal until you remember the scoreboard: if you're down one and lose by two instead, nothing changes. You only need the tying goal to happen before an empty-netter. Every second of six on five buys more of those chances. The question is how many seconds to buy.
WHEN TO PULL, DOWN ONE
Each point on this chart is a plan: the time left when the goalie comes out, and how often a team that's down one with 7:00 left wins with that plan. The strip underneath shows the grade the game gives for pulling at that time.
SHOW AS A TABLE
| PULL AT | WIN CHANCE | GRADE |
|---|---|---|
| 7:00 | 12.25% | A BIT COSTLY |
| 6:30 | 12.54% | A BIT COSTLY |
| 6:00 | 12.79% | A BIT COSTLY |
| 5:30 | 13.01% | FAIR |
| 5:00 | 13.20% | FAIR |
| 4:30 | 13.36% | FAIR |
| 4:00 | 13.48% | GOOD CALL |
| 3:30 | 13.57% | GOOD CALL |
| 3:00 | 13.63% | GOOD CALL |
| 2:30 | 13.64% | GOOD CALL |
| 2:00 | 13.62% | GOOD CALL |
| 1:30 | 13.08% | FAIR |
| 1:00 | 12.43% | A BIT COSTLY |
| 0:30 | 11.64% | A BIT COSTLY |
| NEVER | 10.69% | COSTLY |
Three things jump out:
- Never pulling is the worst plan on the board. It wins 10.7% of the time. The best plans win 13.6%.
- The top is flat. Pulling anywhere from 2:50 to 2:20 left is about equally good, and anywhere from 4:25 to 1:50 grades as a good call.
- Late costs more than early. Waiting until 1:00 wins 12.4%, giving up about 9% of your chances. Going a minute and a half early, at 4:00, wins 13.5% and gives up about 1%.
That's the case for the modern pull. For decades the habit was to wait until about a minute left. The data says waiting that long is the expensive mistake, and going too early is the cheap one.
WHEN TO PULL, DOWN TWO
SHOW AS A TABLE
| PULL AT | WIN CHANCE | GRADE |
|---|---|---|
| 7:00 | 2.32% | GOOD CALL |
| 6:30 | 2.33% | GOOD CALL |
| 6:00 | 2.33% | GOOD CALL |
| 5:30 | 2.33% | GOOD CALL |
| 5:00 | 2.32% | GOOD CALL |
| 4:30 | 2.30% | GOOD CALL |
| 4:00 | 2.28% | FAIR |
| 3:30 | 2.24% | FAIR |
| 3:00 | 2.20% | FAIR |
| 2:30 | 2.14% | A BIT COSTLY |
| 2:00 | 2.07% | A BIT COSTLY |
| 1:30 | 1.89% | COSTLY |
| 1:00 | 1.69% | COSTLY |
| 0:30 | 1.47% | COSTLY |
| NEVER | 1.23% | COSTLY |
Down two, everything is smaller: never pulling wins 1.2% and the best plans win 2.3%. In relative terms, though, pulling still nearly doubles your chances. The best plans pull with 6:30 to 5:20 left, and a good call starts as early as 7:00.
Part of that range sits in the shaded area. The model is fit on the last five minutes, so anything earlier than 5:00 uses the 2:00-to-5:00 rates carried back in time. Treat the left side of both charts as an educated guess.
WHAT IT CAN'T SEE
- Every team is an average team. A great power-play group or a hot goalie changes the numbers, and the model doesn't know who's on the ice beyond faceoffs.
- Five minutes of data. Before 5:00 left the rates are extrapolated, as shaded above.
- Penalties are left out. Power plays and penalty kills late in the game aren't modeled separately yet.
- Ties are worth half. In real standings a tie earns at least a point, and a coach chasing a playoff spot may value that differently.
- Units aren't graded. The data can't yet separate six forwards from five and a defenseman, so that call shows a gray square.
Every daily asks you when the goalie comes out. Now you know what the chart says. See if you trust it with the game on the line.