Goalies are the position this projection is worst at, and it is better to say so up front than to bury it. In last season's backtest the model's typical miss on wins was 7.0 against 8.0 for just writing down last year's total, and on starts it ranked goalies slightly worse than last year's total did (the full backtest). Goalie volume is decided by coaches in October, not by three seasons of history in September.
So this post is three things, in decreasing order of how much the model has to do with them: the projected top 15 from the 2026-09-08 run (93 goalies, 84-game totals); how every team actually split its starts in 2025-26, which is the number a tandem question comes down to; and goals saved above expected last season, split by shot danger, which measures the goalie rather than the team in front of him.
The projected top 15
Read the wins column as two columns: starts and team. Projected starts are a share of the team's 84 games, from three seasons of usage, so a 70% starter who stayed put projects near 59. Wins are the team's expected win rate applied to those starts. That is why Andrei Vasilevskiy projects for 32 wins on 49 starts (TBL at .663) while Dustin Wolf projects for 23 on 51 (CGY at .450). The goalie barely enters into it.
Save percentage is regressed toward the league, hard. The whole board runs from .878 to .914, which is the model admitting it cannot tell goalies apart on save percentage, and the backtest agrees: its projected order of save percentages had almost no relationship to the actual order last season.
How every team split its starts in 2025-26
Regular season, from the game logs. A goalie traded midseason appears under each team; "+1" means a third goalie started games.
Six teams ran a starter at 70% or more. Eleven were below 55%, which is a true tandem, and a tandem goalie's fantasy value is capped no matter how well he plays. The projection's starts column is built from three seasons of these shares, so the places it is most likely to be wrong are the teams that changed goalies this summer and the teams where an injury distorted last season's split. Check the table against camp reports before trusting a starts number.
Goals saved above expected, 2025-26
GSAx is goals saved above expected: the expected-goal value of every shot the goalie faced, added up, minus the goals he actually allowed. Regular season, 20 or more games. Split by shot danger — high is a shot with an expected-goal value of .10 or more, medium .04 to .10, low under .04. Positive is good.
The ten best:
The ten worst:
The danger split is the useful part. Andrei Vasilevskiy was +23.2 on high-danger shots, the best number of any goalie with 20 games. Six of the top ten were positive in all three tiers: Logan Thompson, Scott Wedgewood, Igor Shesterkin, Devin Cooley, Joel Hofer and Jesper Wallstedt. At the other end, Jordan Binnington's -29.1 was spread across all three tiers: -12.3 on high-danger shots, -8.9 on medium and -8.0 on low.
None of this feeds the projection. The model regresses save percentage from box-score results, not from expected goals, so a goalie with a strong GSAx and a middling projected save percentage is a case where two numbers disagree. You get both and decide.
What this does not know
Who wins a camp battle. The starts column gets rebuilt when lines are synced, but goalie usage is a coach's call and stays uncertain past opening night. GSAx is last season's number and says nothing about a goalie who changed teams or defensive structure. Back-to-back counts are the schedule post's job. And 46 goalies on rosters have no NHL history and no projection.
The goalies page has GSAx for every goalie and season; each projected goalie's page shows his starts share and the team rate behind the wins number.