Goals saved above expected inherits the shot model
Hockey analysts spent years explaining why save percentage is a poor way to judge goaltenders. Not all shots are equal, they pointed out, and a goalie who faces a steady stream of point shots through traffic is doing a different job than one who faces breakaways and backdoor tap-ins. The answer that emerged was goals saved above expected, usually shortened to GSAx. It compares the goals a goalie actually allowed to the number an average goalie would have allowed on the same shots. GSAx is a real improvement. But it is only as good as the expected goals model underneath it, and that model has blind spots that pass straight through to the goaltender’s number.
How the number is built
The calculation itself is simple. An expected goals model assigns every unblocked shot a probability of becoming a goal, based on features such as distance from the net, shot angle, shot type, whether the shot came off a rebound or a rush, and the game situation. Add up those probabilities for all the shots a goalie faced and you get his expected goals against. Subtract the goals he actually allowed and you get GSAx. A positive number means he stopped more than an average goalie would have; a negative number means he stopped fewer.
Everything interesting happens in the first step. The goalie’s rating is the gap between reality and the model’s prediction, so any error in the model’s prediction becomes part of the goalie’s rating. If the model underrates how dangerous a certain kind of shot is, goalies who face a lot of those shots will look worse than they are. If it overrates a certain kind of shot, goalies who face those will look better.
What the public models can see
Most public expected goals models, including the ones behind widely used analytics sites, are built on the NHL’s play-by-play data. That data records where each shot was taken from, what type of shot it was, and the sequence of events around it. From the sequence, models can infer whether a shot was a rebound or came shortly after a turnover or zone entry.
What the play-by-play does not record well is what happened immediately before the shot. A one-timer off a cross-ice pass is far more dangerous than a shot from the same spot after the shooter carried the puck in alone, because the goalie has to move laterally and reset. A shot through a screen, with the goalie unable to see the release, is more dangerous than an unobstructed one. A deflection off a stick in front changes direction late. These factors matter enormously to the save, and the basic location data captures them only partially.
Tracking data from the league’s puck-and-player tracking system can fill some of those gaps, and models that use pre-shot movement generally do better. But the most commonly cited public GSAx figures still lean heavily on location and event data. That means they treat a royal-road one-timer and a harmless wrist shot from the same spot as roughly the same chance.
Some teams give up the dangerous version of every shot
The missing information is not random. Defensive systems differ in how well they prevent cross-ice passes, clear the front of the net, and keep shooters from walking into open ice. A team that defends poorly in those respects will give up shots that look ordinary in the location data but are more dangerous in reality. Its goalie will face more goals than the model expects, and his GSAx will suffer for problems that belong to the skaters in front of him.
The reverse is true too. A team that is excellent at forcing shots from the outside and clearing rebounds may give up shots that the model rates as more dangerous than they really are. Its goalie can post an impressive GSAx partly because the model is overestimating the difficulty of his workload. GSAx was designed to separate the goalie from the defense, and it does that much better than save percentage, but it cannot fully separate them on the things the model does not see.
Where the shot was recorded is not always where it happened
There is a second source of error that analysts have documented for years: rink bias. Shot locations in the play-by-play are entered by scorers in each arena, and different scorers have historically recorded shots systematically closer to or farther from the net. If a building’s scorers tend to log shots as closer than they really were, the model will think the home goalie faced more dangerous shots, and his GSAx will be inflated. If they log shots as farther away, it will be deflated.
Model builders try to correct for this, and the league’s tracking data has reduced the problem. But a goaltender plays roughly half his games in one building, so any remaining bias in that building’s scoring can follow him through a whole season.
Small samples, big swings
Even with a perfect model, goaltending performance is noisy. A starter faces somewhere around fifteen hundred to two thousand shots in a full season, and the difference between an elite season and an average one is a relatively small number of goals. A few bad bounces, deflections, or screened shots can shift a goalie’s GSAx noticeably. That is why GSAx often looks very different from one season to the next for the same goaltender, and why multi-season samples are much more reliable than a single year, let alone a single month.
How to use GSAx well
None of this means abandoning the stat. GSAx is a better starting point than save percentage or goals against average, both of which ignore shot quality entirely. It just needs to be read with its limits in mind:
- Check which model: different sites use different expected goals models, and they can disagree about the same goalie by several goals over a season.
- Prefer larger samples: multiple seasons tell you far more than one, and a month tells you almost nothing.
- Consider the team: a defense that allows a lot of cross-ice passes or screens can drag down its goalie’s number.
- Look at splits: how a goalie performs on high-danger chances versus low-danger shots can reveal whether the model is the problem.
- Watch the games: rebound control, positioning, and puck tracking are hard to capture in any public data set.
GSAx answers the question it was designed to answer much better than the old stats did. But its answer is always relative to a model, and the model is making its own guesses about every shot. When the model is wrong in a consistent way, the goalie’s number is wrong in the same way.
GSAx does not measure a goalie against the shots he faced. It measures him against a model’s description of those shots.