Some Thoughts on Goaltending Statistics

Stephen Burtch
5 min readMar 17, 2021

I have noticed a recent shifting in the conversation away from traditional Goaltending metrics such as save percentage and goals against average, towards more focused dissection of High Danger save percentages.

Much of this is being driven by the Toronto sports media spotlight being turned on Frederik Andersen — who is in the midst of a down season in a contract year — who is again being found wanting in his usual areas of difficulty.

What many fans likely weren’t aware of prior to this season, is that Andersen has to some extent always struggled with Higher Danger shot attempts, and largely built his solid Save Percentage results in prior years on the “skill" of very very rarely allowing Low Danger shots to score.

Those seasons, early in his Toronto tenure, where he was anointed “Steady Freddie" for consistently posting a .918 save percentage were a bit of a mirage statistically.

Toronto for many years, has been a defensively porous outfit that consistently allowed very large volumes of shots on net. NHL clubs will typically surrender a larger number of Low Danger shots from distance than high danger shots from the slot or off the rush.

Andersen’s persistent excellence in stopping lower danger shots means that higher shot totals (with a larger proportion being Low Danger) would lead to his save percentage rising considerably.

Since Sheldon Keefe has taken over behind the Leafs bench, they have sought to stem the bleeding defensively and cut their shots against considerably.

Unfortunately for Andersen, when NHL clubs restrict shot share, this typically means they allow fewer Low Danger chances, and High Danger attempts still happen at a similar pace.

I’m not a big fan of the idea of retaining Andersen’s services because he has consistently struggled in stopping High Danger shots. I’ve made this point repeatedly in the past… but I digress.

All of this is my very roundabout way of highlighting a significant problem with more modern Goaltending metrics.

So before I address how I think we can fix the issue, let’s summarize the key issues as I perceive them.

  1. Save Percentage is bad because not all shots are equally difficult to save.
  2. Binning shot types into High, Medium and Low Danger categories is also problematic statistically because it is arbitrary where you delineate the bins. One website’s high danger chances aren’t equivalent to another’s. You can see where this becomes problematic.
  3. Less dangerous shots are essentially meaningless from a Goaltending analysis perspective. There is no reliable information contained in whether a goalie lets in Low Danger point shots or not. Even if they don’t let them in, long term virtually all NHL goalies stop the vast majority of Low Danger attempts.

If we accept the above 3 points as a starting position you get to some interesting places analytically.

If we know the majority of information is contained in a goalie’s ability to stop the most dangerous shots, why do we let Low Danger shots determine how we rate goalies to such a high degree?

Virtually all current advanced Goaltending metrics function on a system that penalizes goalies for the goals they allow rather than crediting them for the saves they make.

Consider the case of Carey Price this season. If you look up his High Danger Sv% he again ranks near the top of the NHL (as is often the case). But if you look up his GSAx or dFenSv% he looks abysmal. Why? Because he has let in an abnormally high number of Low Danger shots, and the penalty for doing so is large.

GSAx is literally just the difference between expected and actual goals a goalie allows. The difference between what types of shots a goalie lets in affects GSAx drastically. If a goalie allows goals on harder to stop shots and not easier to stop shots they will look better than if the reverse happens.

The disparity between the xG value and the actual goal surrendered is far higher for a Low Danger shot than a High Danger shot.

We penalize goalies far more for what amounts to random noise than we reward them for actual skill.

We also talk about “back breaking" bad goals that goalies allow, while excusing difficult to stop shots going in (“he had no chance on that one").

In reality the opposite should be our approach. Long distance randomness? “Meh that sucks, but flukes happen and he almost never lets those in…”

High Danger slot shot gets buried again? “Why are we paying this guy so much if he can’t stop dangerous shots?”

We remember goalies for random mistakes far more than their great saves. It seems unfortunate that our statistical measures are reinforcing the same tendency.

A solution is possible

We don’t praise defenders crazily for scoring on random long distance point shots. We don’t suddenly decide they are star scorers because their goal was worth so much more than the expected value of the shot.

We know shooting is highly variable and prone to fluctuation. We know the same about goaltending.

Maybe we could account for this by crediting Goalies for the xG of the shots they stop, rather than penalizing them for the goals they allow.

We can fall back on the idea of save percentage, which is simply saved shots divided by shots faced.

I propose the idea of xGSv%.

All we need to do is divide a goaltender’s total xG saved by the total xG of all shots they faced.

Low danger shots won’t inflate the numerator or denominator considerably. Saving easy to stop shots won’t make a goalie look amazing. Allowing the occasional random point shot to go in won’t make them look horrible.

The main skill in Goaltending is saving the hard to stop, high xG shots. The only way to post a good xGSv% would be to stop more of those shots.

The litmus test here is reliability in the metric. Do past numbers predict future numbers?

More to come on this I promise (data, charts, etc.) — just want to get the idea out there for discussion at this point.



Stephen Burtch

Stephen Burtch is a Math and Physics Educator from Toronto who has written about and analyzed hockey for over a decade.