Measuring Goalie Decisions

A Different Approach to Micro Stats for Goalies

Back in October, I published an article on a skill-based goalie metric. It’s a wishlist and a pipedream rolled into one. I’m proud of that article, and maybe there’ll be the day where the ideas can be transformed into something actionable, but not yet.

In the meantime, I’ve done a lot of thinking recently about micro stats, which typically try to measure everything leading up to shots. They involve a granularity that is something I have found extremely attractive when thinking about goalies this summer. One question I keep returning to is how real was Jake Allen’s 2024-25. 19.07 GSAx according to Evolving-Hockey is frankly incredible value for a guy who only played 34 games. But it’s not in line with his aging curve and his previous results. What changed? Did something change? Is this just an outlier season, or did an older dog learn new tricks?

Goalies are infamous for their season-to-season variation in results. Frankly, there’s no statistical way to project future goalie performance, even using GSAx. GSAx is the best publicly-available model we have, but it is entirely blind to goaltending action as I describe in my skill-based goalie stat article:

Towards a Skill-Based Goalie Statistic

As things stand we can’t think of any results a goalie achieves as wholly reliable by itself because, beyond the eternal problem of fully separating team defense, goaltending is a fundamentally passive and reacting position. Let me clarify what I mean, because there are many passive and reactive aspects of other positions in other sports. When quarterbacks drop back to make a decision on a play, they make decisions that influence directly the flow of offense. Yes, they react to what is happening around them, but they control the ball and have a say in how they avoid sacks and turnovers, if they pass or run, and who they target. When pitchers throw a baseball, they control where it goes and when and how it’s delivered; when they throw strikes, they not only prevent offense, but also (except for 2-strike fouls) actively affect the game state by reducing potential of future offence. Batters likewise react to pitches, but in doing so, they make decisions that directly affect offense in some way.

So what do we do? How to we track what a goalie actually does — their process — if we can’t assign value (results) to it? The simple answer is don’t assign value to it. Just track the process. Baseball does this with swing decisions — but swing decisions correlate to results. Baseball Youtuber Foolish Baseball reminds us that “it is good to swing at strikes and not swing at balls”. For goalies, it’s not clear which actions best accomplish their goal of stopping the puck.

We will assume that a goalie’s decisions describe who they are and how they choose to stop the puck. Since we cannot accurately assign value to actions, we will instead use the actions to understand the decisions being made. The goal is not to ascribe value to these decisions, but rather track them as a way to highlight a change or maintenance of approach.

Basic Metrics

These are metrics that, theoretically, anyone with enough time¹ and access to 1312 NHL games could record so long as there are consistent standards.

• Drops/CA — the idea is simple: I want to know how frequently goalies drop to their knees relative to all shot attempts. A drop would be considered any time a goalie brings at least one knee to the ice after having had neither knee on the ice. For these purposes drops include feet-to-knee slides. Bringing Corsi back to its original purpose of tracking goalie workload, this metric posits that in modern goaltending one of the main decisions a goalie makes is to drop to one or both knees. It’s a fairly major commitment for a goalie to make: being on your feet allows you to move backwards or forwards (via c-cuts) or laterally, so there must be some reason to drop. Typically, that reason is to deal with a shot or the threat thereof. The point of this is metric is simple: different goalies have different levels of comfort for when they want to make that trade-off. This measures the rate a goalie makes that trade-off without judging whether or not it was the correct decision. The tough part for this metric would be if there isn’t a sample large enough for this metric to stabilize — what if game scenarios are too random to get meaningful results from this method. Upsetting, but we’d still be in the same boat we are right now. The other potential issue is the issue of multiple shot attempts occurring when the goalie has already dropped. The solution is to count “unique drops” (ie instances of a knee touching the ice, when neither knee had previously touched the ice) and shot attempts that happen with the goalie already on one or both knees together. The only stipulation being that a “unique drops” ‘ends’ when a shot attempt is recorded or the neither of the goalie’s knees touch the ice. This should help distinguish what are shot or location reactions from rebound and scramble scenarios. Furthermore, atomizing this data by location could be useful. Using the zone map used by NHL EDGE as the demarcation points, we’d expect the highest ratio to occur in the lowermost sections of the zone. Fenwick might also be useful for this exercise, but I would argue that goalies react to blocked shots enough that it can’t really be thrown out here. The alternative is to break down the decisions by shot attempt outcome: blocked, missed net, and shot on goal, but for these purposes, non-attempts or discretionary drops would have to be removed from those samples.
  This clip of Anthony Stolarz in this March 3rd, 2025 game versus the Sharks provides a quick example. I count 2 unique drops and 2 shot attempts, but 1 shot attempt came with Stolarz already having made a unique drop, so we only count the first one.

• Feet/knees lateral movement ratio — It sounds simple, but it is pretty complicated: what percent of goalie lateral movement involves the goalies knees touching the ice. Goalie movement is hard to quantify, which is why I’m not counting depth-taking movement (c-cuts or glides) and just focusing on lateral movements. Even still, feet-to-feet movement can be pretty minute and rapid — shuffles specifically— but we can’t just discount shuffles because shuffles (toes are basically parallel) and t-pushes (toes begin perpendicular, before snapping back to parallel to stop the movement) actually exist on more of a spectrum —especially in mid-range push lengths (think anything from an angle to the middle of the net) — when you reach very high levels of goaltending. Over a large enough sample size, this should reveal who is most comfortable moving laterally on their feet, and who is most comfortable doing so on their knees. Neither provides a “right” answer, but for some goalies, they may benefit from trying to use one over the other in more situations. Furthermore, what you might see is younger goalies having a ratio that might be seemingly non-sensical: a giveaway perhaps that they’ve yet to adjust to the speed of the game and are making decisions that aren’t correct for them, even if they aren’t being burned for it.

• Rush Feet/Knees movement ratio — the same as above, but only for rush chances, which we can define as scenarios that begin with an entrance with possession and fewer defenders ahead of the puck (closer to the net) than attackers between the puck-carrier and the nearest defender that is behind the puck, and ends with either a shot attempt, a turnover, or when the maximal number of skaters on the ice have entreated the zone. Geez! Odd-man rushes are so easy to intuit but hard to explain. If there are more defenders closer to the net than there are attackers who are between the puck and the closest defender that is still behind the puck. We’d expect this metric to have more feet than knees² within goalies, but not necessarily across goalies.

This ends the portion of stats that could just be tracked by theoretically anyone. We, unfortunately, aren’t left with a very full picture of goaltending decisions. So, I’m left returning to the statistical wishing well, all in the name of creating a fuller profile for goalies that can be used to track their tendencies in the net. As we go along, it becomes clear that there are far more tangible benefits to using those than just ‘drop decisions’.

NHL EDGE-adjacent Stats

This is where I mix things up a bit. Some of these are positioning decisions but others are skating-based.

As you may or may not know, I’m not overly impressed with NHL EDGE’s goalie stats. I wrote an instant reaction note in March that spelled out my issues with its more dubious stat-tracking.

I’m still mad about that. But there’s other things I’m not a big fan of. I don’t like its save percentage by location. In my opinion, it unnecessarily atomizes save percentage by something meaningful, like shot location, while being stubbornly agnostic about shot quality — thus, actually making it unmeaningful. As flawed a stat as save percentage is, it’s unapologetically encompassing and evaluative. If I say a goalie has a .900 save percentage, it means that goalie has save 9 out of every 10 shots they faced. It gives a statistic that I can translate into both a rough shots-to-save rate or adjust and compare using GA%-. It suggests an overall quality. Telling me a goalie has a save percentage of .900 from slot shots is great, but that’s not the only place shots come from. SV% by location is in some ways descriptive, but it’s also simultaneously meaningless, because I now need the full picture again. And it actively removes the weight of shot selection. High-, medium-, and low- danger save percentage cause me similar headaches. Please. We have xGoals. They account for all of this. Just use that, NHL EDGE. It’s actually infuriating that an organization with access to each skaters speed can’t make an xGoals model that could potentially improve upon all current models using their skater speed data. But no, I suppose we can’t have nice things.

Another way to look at the zone SV% issue is through the lens of Connor Hellebuyck’s zone-by-zone SV% breakdown for 2024-25:

Hellebuyck’s SV% by zone courtesy of NHL EDGE. Yes, the 0.926 SV% from the left boards is strange, but I posit if that is actually meaningful?

A .926 SV% from the middle-of-nowhere on the blocker side is below average in 2024-25. Ask yourself these two questions: (1) does this factoid add meaningful knowledge to our understanding of Hellebuyck’s performance last season? (2) Should opposing teams target this “relative weakness” when playing Winnipeg or would they be better off attacking the places where Hellebuyck posted SV%’s below .850 because it turns out that that that ice is where most goals will always be scored. In case you’re still wondering, shooting above 15% is better than shooting 7.4%. This is why I don’t believe in describing goalies as having “weaknesses” — because things like “get screens and tips”, “shots just above the pad (especially to the blocker-side”), and “shooting before the feet are set” are good keys to success no matter who is in net. There may be technical deficiencies or what have you, but at the NHL level, goalies with significant shortcomings usually don’t hang around.

Now that I have endeared myself by criticizing the entire NHL EDGE stats coverage of goalies, I’m going to come up with ideas that perhaps only NHL EDGE has the technology to implement. I’m not beating the “ideas guy” allegations implied by my previous article, am I? I’m also going to looking to leverage the zone divisions EDGE already employs. Why? Because where I suggest it, it will actually be contextually important to know where the puck is relative to the goalie, indicated by the phrase “by zone-section”. When I do so, ask yourself, “what benefit would not doing that have?” Typically, the answer will be none, which is what differentiates my usage versus using it for save percentage.

• Average crease depth per o-zone section — I want to know how far a goalie will challenge on average when the puck is in a certain part of the zone. As mentioned above, the different zones are useful because having an overall measure of this would actually not say anything much of substance. This is probably one of the more immediately noticeable aspects of process.
  Mike Smith, for example, famously made an adjustment to stay further back in his crease in essentially all situations upon his arrival in Phoenix in 2011-12. This lead to a monster breakout season at age 29 and while that season was undoubtedly his best, his new process allowed him to still have that same dominant ceiling through the rest of his career (which lasted until his age 41 season. Call him a late bloomer, I guess). This first clip is from game 4 of the Boston-Tampa series in 2011 in which Smith relieved fellow 21st-century Oilers legend Dwayne Roloson after the later allowed 3 unanswered goals in the first period (to be fair to Roloson, while not his best performance, it still reads like a placebo pull). Here’s Mike Smith on a cross-crease opportunity: the shot isn’t the best, but Smith gets good depth on his feet before snapping down. Ideally, he would have stopped himself before making the save, but it doesn’t kill him here, because — again — not the best shot. The closest 2011-12 equivalent I could find was this play from game 1 of the 2012 Conference Finals versus the Kings in which Smith starts nearer to the middle than the first clip, respecting Dustin Brown’s shot, and the eventual pass from Brown is a little slower than the fist clip, but Smith stops about half-way up the crease just for Doughty to fan even worse on the shot. Now, to be clear, if this pass is faster, I’d expect most goalies to have ended up here just to beat the pass and be set for the shot, but since this pass is so slow, and actually is functionally stopped by the time Doughty skates into it, Smith kinda camps in his spot — which is kinda an underlying theme of his 2012 breakout.
  I also want to compare where he is in the net for the faceoff. I typically wouldn’t classify this as a major adjustment, but there is a noticeable difference in Smith’s case.

  

  Mike Smith lines up for a faceoff in Game 4 of the 2011 Conference Finals versus the Bruins.

  I don’t think I can overstate how uncontroversial Smith’s positioning is in the above picture. This is where I used to set up for faceoffs — all 5’7” of me. But it’s not just me: all 6’7” of Mads Søgaard likes this spot as well, as you can see below.

  

  Søgaard (far right) at the top of his crease directly off the draw in the second period of an October 14th meeting between the Sens and Kings.

  You don’t want to challenge too much here — faceoffs are typically not credible shot threats from the opposing centre — but getting your toes to the top of the crease is typically a good idea as it sets you up to make a lateral pushes out if it goes towards the middle off some sort of set play. So when you see this next screenshot, it should be clear that something unorthodox is at foot.

  

  Mike Smith lines up for this faceoff just under 12 months after his previous picture in Game 3 of the Conference Semi-Finals versus the Nashville Predators.

  This is crazy. Mike Smith is on his post — not in the sense that he’s trying to seal it, but he does have his foot there. What’s interesting, though, is that the in-zone faceoff dots are typically a demarcation of what is dangerous ice and less dangerous ice, so, in a sense, this makes sense: the puck position on the dot isn’t really as dangerous as the puck actually moving to the middle. It’s emblematic of Smith’s stylistic adjustment (which we will explore even further in the next section) that is based on only challenging head-on shots rather than a bit of everything as many goalies do.
  At this point you might be saying “I thought this was about goalie microstats, not Mike Smith”. I know. But there’s no better before-and-after demonstration of the value of process-tracking than Mike Smith. Mike Smith’s breakout in 2011-12 is obvious because the change is so drastic — and everyone was talking about his work with Sean Burke being the catalyst, so we were even contemporaneously aware of this. Like, this is knowledge I’ve had for so long, I can’t remember where I picked it up. But Smith’s style in 2012 is an extreme, and I’m not sure if he maintained it for the rest of his career. I might need to do a follow-up film study.

• Average crease depth per o-zone section (rush) — Rush chances are different than those that arise from inside the zone. The downward pressure of attack requires a management of depth and momentum. Also, the puck carrier needs to lead a rush for it not to be offside, thus goalies will typically be more aggressive in their initial depth, and use backwards momentum as they accept the rush. The consensus on this method changed over the time. In the late 00s to early 2010s, there you might find more active c-cutting to generate this momentum whereas nowadays the orthodoxy is to just use 1 cut and glide. Regardless, since depth is typically in conversation with lateral movement having the carrier be the lead, or close to the lead, is an important distinction.
  Mike Smith plays this 3-on-2 with fairly a fairly orthodox level of depth — in fact, what’s pretty outlier-ish is how he maintains his position at the top of his crease, likely due to the actual pace of the rush and position of the puck carrier. His D also do a good job limiting options to the middle, thus the pass back is the only legitimate option for a pass after the puck-carrier is forced wide. Contrast that against this play, not even half a year later, where we see Smith start at the top of his crease when the puck is in the middle, but quickly adjusts his depth back as the puck travels in more-or-less a straight line to the boards. The angles are slightly different, yes, but Smith is much more willing to lose depth in the second clip. For our purposes, we might even consider sub-dividing NHL EDGE’s zones (depicted below) with a more goalie-centric idea thereof.

  

  Jake Allen’s 2024-25 shot location breakdown. The dark grey indicates below 50th-percentile. I don’t know why they’d make such a choice— especially when measuring this by volume. It frankly bonkers to me

  

  Here is my adjustment to the EDGE model. The thinner lines are ones I added. I would consider actually having “Centre zones” — the ones in the direct middle — and “mirror zones” to concatenate congruent zones under the assumption that goalies won’t drastically prefer a level of depth on one side of the ice over another. I’d also ignore anything outside the blueline and below the goal line.

  The idea, however, is obviously not so simple. For as little credit as I’ve given NHL EDGE, no one just draws random lines on a rink and expect to draw meaningful conclusions from it. They need to mean something for their specific purposes. And for NHL EDGE, they do: if we’re thinking about shooters and where in the ice they shoot from most frequently. They explicitly used shot data (specifically location volume and frequency) to demarcate this. And it is useful for those ends. Draisaitl has spots on the ice that he likes more than others, creating incongruencies in his shot distribution. Examining Jake Allen’s shot chart (it’s the one with numbers), we see rough congruencies of all non-middle sections, with the exception of the middle-left (82) and -right (62) sections. Even still, 20 more shots over the course of a 30+ game season isn’t necessarily something to interpret as meaningful. That’s why, for goalies, we need to group these areas together by their “mirror zones”. Mirror zones are the zones that would exist if you chopped the ice in half and held up a mirror to them. So, the left and right points would be grouped together, as would the left and right side-boards.

• Average PK crease depth by zone —Many goalie will alter their style slightly on the PK given that opposition’s capacity for passes and movement require a more respect be given to threats in better ice, even as each individual shooter is likely more effective than average. It’s a tight-rope walk. If a goalie’s PK approach and movement-execution (elaborated upon below) are identical year-over-year, then we’d expect the results to be tied more-or-less to the shot-suppressing abilities of the actual PK, or, for playoff scenarios, shooting talent of specific opposition players. We also might expect less regard for these scenarios among certain goalies. Is this because their approach is invariably equally successful (or not) regardless of strength? Who knows.

• “Stop-to-stop” on-feet movement maximum speed (minimum 1 ft travelled) — basically, this measures how quickly goalies can travel cross-crease on their feet when they really need to. When moving cross-crease, the first thing that must move is the head: it tracks the puck. There’s typically a point, when the puck has moved past square to the shoulder when you should have the trajectory lined-up and ready to execute the movement. For most plays this happens so fast it’s not even conscious, but what we need to realize here is that there’s a delay, a slight lag between tracking and moving. Once opened up, knowing the trajectory, the goal is follow the path of the puck to its destination and be stationary by the time the recipient can do anything with puck. Keep in mind, these are generally longer passes that we’re talking about. The puck can travel faster than the goalie, but, luckily for the goalie, the distance is much shorter. But what we also need to take away from this exercise is not only the process of movement, but the fact that movement of all kind in goaltending is puck-informed. We can’t take the average of these movements because that won’t tell us as much as the upper limit since goalies should never move faster than the puck. You see this even in high-stakes situations — overtimes, last 2-minutes of a game: goalies will still move at a leisurely pace simply because that’s how fast the puck is moving. A nth-percentile speed might also be useful for similar purposes.
  You might be wondering why I called this category “stop-to-stop” and not just plain “foot speed” or “movement speed”. The answer is that I would only care about a certain type of movement — a movement with a definite endpoint: a stop of some kind. Ask yourself, how would you measure the speed of Mike Smith’s head-first dive on this play? Or a slide that turns into the splits? This pedantry is just as much about ease of measurability as it is about the simple fact that these desperation moves are fundamentally about compensating for the whole body being unable to catch up by putting some of the body in front of the puck. In fact, I’d suspect in most cases, it’s marginally slower to move across in the splits, but you can horizontally cover more space quicker.
  What this could tell us who is not only who is moving fast, but who is moving effectively. While it’s fairly intuitive to assume that goalies with better raw speed will be more effective, all things equal — but we all know that’s not how things work at all. Additionally, changes in foot-speed could be a canary in the coal mine for a potential decline or who has potentially had one hip surgery too many.

• Average “stop-to-stop” on-feet max speed — For as much as I touted measuring a top speed over an average speed, an average speed could also be a useful tool if used in conjunction with an upper speed measurement. This is because we can account for movement selection. Goalies faster on their feet are more likely to select foot movement. Sliding is for when you’re behind the play or anticipating a pass to arrive and be released faster than you can set your feet and react. To be clear, arriving on one’s feet is not is merely theoretically optimal. As I’ve said, the essence of a save is an interception of trajectory. So long as that is achieved we have functionally reached optimal goaltending on the micro-level. Arriving on one’s feet simply leaves movement options more readily available, like responding to a return pass.

• Average Head height pre-release (non-rebounds only) per o-zone section — This measures exactly what it says on the tin, but why measure it at all? The head is simply a very easy, consistent, and visible thing to measure, being always on top of the chest. The purpose is to measure stance height using the head as a proxy — likely using percentage of full on-skates height. Goalies will typically vary their stance depending on where the puck is in-zone, but sometimes to different degrees. Using this metric can indicate the height of the stance, but also the butterfly. Looking at pre-release, we can see how goalies do things like fight through screens, cover net on slot shots, and manage their height when the puck is on the perimeter.

• Average Stance width per o-zone section—Like head height, this measures stance shape and how it may or may not vary throughout the zone. The metric is simply the distance between a goalie’s skates while on their feet. Though it is simple, it might not end up being easy to actually measure. Skates are typically blocked from immediate camera view and thus might need to be estimated. Interestingly, the advice I typically hear these days is that stances should optimally have the knees under one’s body to help both movement and reactions. This said, Henrik Lundqvist was famous for both of those things and he typically used a very wide stance. Yet another example of why these stats resist an outcome-oriented approach from the outset.

Meaningfulness and Feasibility

The one major hang-up I have on these is will there be enough difference in behaviours to derive anything meaningful from the data? Will these metrics even be stable enough to get consistent results absence conscious changes in decision-making. In a word, I worry about potentially inherent noise. If these numbers are too noisy, they won’t be able to track anything. The reason Mike Smith managed to hijack so much of the focus of this article is simply because of how radical his changes were from 2010-11 to 2011-12. It’s up there with Jonas Hiller’s hyper-butterfly style in terms of it extremeness. Both are actually rather antiquated and quaint in today’s goaltending landscape, so who can tell if there are consistent and measurable differences between the goaltenders of today. There’s also a danger that this will simply reveal defensive zone patterns rather than actual, useful data. Hopefully by covering both rush and non-rush decision making, we can piece together something consistent and less noisy, but it’s not immediately clear.

Then there’s feasibility. I return to the NHL EDGE conundrum: is EDGE simply unable or unwilling to record meaningful stats for goalies? I find both equally likely, with the secret third option of “both” also looming large. It’s hard for me to say 100% that the lack of substantial data from the NHL is actual indifference. But it’s things like goalie goal differential per 60 and points % as actual categories that suggest this indifference — like, I think things like OT SV% and shootout SV% would actually be things that aren’t really easy things to track, that aren’t tracked elsewhere and would be fairly interesting. I understand that there’s no way to accurately track a puck on the ice, but EDGE has been measuring player speed for a while now. Why can’t we measure goalie speed, even in only specifically delineated movements? Do they have goalies or goalie coaches on staff that are feeding them ideas?

The idea of this article is that there is more to measure for goalies than just results even if we need to conceive of such measures being wholly independent of results. If goalies cannot dictate meaningfully the play around them then we need to measure what they decide to do with that information and how well they can execute their decisions and transform them into results. If we cannot model successful goalies statistically, then we should try to describe individual goalies to see what could make them, specifically, successful.

1. Very basic rough math tells me that this would take over 2000 hours, or ~83 days of solid footage, if we assume 90 minutes of broadcast footage.

2. Ignore the uncomfortable image