Ranking the top-40 goaltenders by Goals Allowed Per Expected Goal Against

[TomasHertlsRooster]

Rank	Player	Team	Goals Allowed Per Expected Goal (Average)
1	John Gibson	ANA	0.85
2	Sergei Bobrovsky	CBJ	0.88
3	Philipp Grubauer	COL/WSH	0.89
4	Corey Crawford	CHI	0.93
5	Ryan Miller	ANA/VAN	0.95
6	Ben Bishop	DAL/T.B/L.A	0.95
7	Frederik Andersen	TOR	0.95
8	Matt Murray	PIT	0.95
9	Marc-Andre Fleury	VGK/PIT	0.96
10	Braden Holtby	WSH	0.96
11	Pekka Rinne	NSH	0.97
12	Jaroslav Halak	BOS/NYI	0.98
13	Darcy Kuemper	ARI/L.A/MIN	0.99
14	Mike Smith	CGY/ARI	1.00
15	Thomas Greiss	NYI	1.00
16	Cam Talbot	EDM/PHI	1.00
17	Robin Lehner	NYI/BUF	1.00
18	Jimmy Howard	DET	1.00
19	Carter Hutton	BUF/STL	1.01
20	Roberto Luongo	FLA	1.01
21	Martin Jones	S.J	1.02
22	Henrik Lundqvist	NYR	1.02
23	Jonathan Quick	L.A	1.02
24	Jacob Markstrom	VAN	1.02
25	James Reimer	FLA	1.02
26	Connor Hellebuyck	WPG	1.03
27	Tuukka Rask	BOS	1.03
28	Andrei Vasilevskiy	T.B	1.04
29	Brian Elliott	PHI/CGY	1.05
30	Semyon Varlamov	COL	1.06
31	Jonathan Bernier	DET/COL/ANA	1.06
32	Carey Price	MTL	1.06
33	Kari Lehtonen	DAL	1.06
34	Keith Kinkaid	N.J	1.07
35	Cory Schneider	N.J	1.08
36	Petr Mrazek	CAR/DET/PHI	1.10
37	Craig Anderson	OTT	1.11
38	Devan Dubnyk	MIN	1.11
39	Jake Allen	STL	1.12
40	Cam Ward	CHI/CAR	1.12

[TBODY] [/TBODY]

I initially just calculated this for my own purposes and had no interest in starting a thread about it, but after seeing the results I felt like it was worth posting. The numbers pass the smell test even more so than I thought.

These are all the goaltenders to play at least 5,000 minutes over the past 3 seasons. This was just an arbitrary number that I picked and once I saw that exactly 40 goaltenders had played that number of minutes, I figured that number sounded about right.

The data used to calculate these numbers is from both Natural Stat Trick and Corsica.Hockey. For each model, goals against were divided by expected goals against. In order to combine the two models, goals against and expected goals against for each data set were added together. (Corsica's data set slightly varies from Natural Stat Trick's because Corsica is missing a few games and goals.)

The full data is here:

Spoiler

Player	Team	GA (Corsica)	xGA (Corsica)	Goals Allowed Per Expected Goal (Corsica)	Goals Against (NST)	xG Against (NST)	Goals Allowed Per Expected Goal (NST)	Goals Allowed Per Expected Goal (Combined)
John Gibson	ANA	399	482.03	0.83	401	462.4	0.87	0.85
Sergei Bobrovsky	CBJ	437	515.7	0.85	438	474.62	0.92	0.88
Philipp Grubauer	COL/WSH	203	234.61	0.87	205	224.23	0.91	0.89
Corey Crawford	CHI	305	341.78	0.89	306	317.63	0.96	0.93
Ben Bishop	DAL/T.B/L.A	301	322.23	0.93	301	309.58	0.97	0.95
Ryan Miller	ANA/VAN	254	281.6	0.90	254	252.02	1.01	0.95
Frederik Andersen	TOR	513	557.13	0.92	513	517.9	0.99	0.95
Marc-Andre Fleury	VGK/PIT	359	388.65	0.92	359	360.52	1.00	0.96
Braden Holtby	WSH	439	459.8	0.95	440	453.6	0.97	0.96
Pekka Rinne	NSH	405	434.3	0.93	408	403.98	1.01	0.97
Jaroslav Halak	BOS/NYI	325	352.25	0.92	326	314.77	1.04	0.98
Matt Murray	PIT	730	786.55	0.93	373	364.82	1.02	0.96
Darcy Kuemper	ARI/L.A/MIN	246	257.46	0.96	248	241.79	1.03	0.99
Mike Smith	CGY/ARI	406	417.14	0.97	406	396.96	1.02	1.00
Thomas Greiss	NYI	308	320.59	0.96	308	296.69	1.04	1.00
Robin Lehner	NYI/BUF	388	396.58	0.98	388	377.51	1.03	1.00
Jimmy Howard	DET	363	373.6	0.97	365	352.54	1.04	1.00
Cam Talbot	EDM/PHI	466	485.2	0.96	467	446	1.05	1.00
Carter Hutton	BUF/STL	256	264.53	0.97	256	241.99	1.06	1.01
Roberto Luongo	FLA	307	322.4	0.95	307	283.4	1.08	1.01
Martin Jones	S.J	471	479.79	0.98	473	447.66	1.06	1.02
Henrik Lundqvist	NYR	480	492.78	0.97	480	450.13	1.07	1.02
Jonathan Quick	L.A	330	325.92	1.01	331	321.31	1.03	1.02
Jacob Markstrom	VAN	379	388.21	0.98	382	356.92	1.07	1.02
Connor Hellebuyck	WPG	479	485.98	0.99	481	449.64	1.07	1.03
James Reimer	FLA	311	321.56	0.97	311	285.66	1.09	1.02
Tuukka Rask	BOS	371	372.04	1.00	371	348.29	1.07	1.03
Andrei Vasilevskiy	T.B	418	419.88	1.00	418	387.74	1.08	1.04
Brian Elliott	PHI/CGY	296	290.46	1.02	302	280.12	1.08	1.05
Jonathan Bernier	DET/COL/ANA	276	264.19	1.04	276	258.03	1.07	1.06
Semyon Varlamov	COL	336	330.37	1.02	340	309.79	1.10	1.06
Carey Price	MTL	445	437.64	1.02	447	404.15	1.11	1.06
Kari Lehtonen	DAL	234	224.39	1.04	234	215.4	1.09	1.06
Keith Kinkaid	N.J	296	291.31	1.02	300	263.15	1.14	1.07
Cory Schneider	N.J	347	339.84	1.02	347	305.56	1.14	1.08
Petr Mrazek	CAR/DET/PHI	344	329.62	1.04	346	299.08	1.16	1.10
Craig Anderson	OTT	433	411.38	1.05	435	372.48	1.17	1.11
Devan Dubnyk	MIN	449	410.68	1.09	449	396.35	1.13	1.11
Jake Allen	STL	411	378.82	1.08	411	358.38	1.15	1.12
Cam Ward	CHI/CAR	389	362.69	1.07	389	334.73	1.16	1.12

[TBODY] [/TBODY]

In my opinion, this is by far the best way to statistically evaluate goaltenders since it accounts for shot quality. Other metrics like GAA, SV%, Wins, etc. don't account for shot quality. Obviously these expected goal metrics aren't perfect but they do have a strong correlation between actual goals and the methodology behind them is solid.

In addition, I think that "goals allowed per expected goal" is the best way to describe and express this metric. Metrics like delta save percentage or GSAA are not quite as simple and do not explain themselves so clearly. "Goals allowed per expected goal" is a lot more simple IMO. For every 100 expected goals that your team allows, John Gibson allows only 85 goals, and Cam Ward allows 112. That is very easy to explain.

Obviously, if I were ranking goaltenders, my final list would be a little different from the list we have here. For example, factors like consistent performance under a heavier workload would lead me to rank Frederik Andersen over Philipp Grubauer, factors like playoff performance would lead me to rank Braden Holtby over Frederik Andersen, and factors like recency bias would lead me to rank Andrei Vasilevskiy over Martin Jones.

However, I still think the list is pretty damn good, and does a better job of ranking these goaltenders than most skater ranking metrics I've seen. Thoughts on the list?

EDIT: A post later in the thread which dives into one unpopular assumption (Rask Vs. Halak). Worth reading.

Using data from Hockeyviz, let's look at Halak and Rask over the past 3 seasons:

Look at the defense that Rask played in front of in 2016-2017 and 2017-2018:

Spoiler

Now compare that to the defense that Halak played under:

Spoiler

Is it not pretty clear, judging from these 8 images, that comparing Halak and Rask side-by-side without adjusting for shot quality will be very flattering to Rask? This is like comparing point totals when one guy plays with Connor McDavid and the other guy plays with Chris Tierney. When we compare players in that manner, we are perfectly fine with adjusting for the environment around these players. Why shouldn't we also do this for goaltenders?

Now look at 2018-2019, when they played for the same team, and therefore played under a very similar defense:

Spoiler

2018-2019 is the only season where they have faced shot quality that is remotely similar, and Rask still faced somewhat easier shots. Despite that, Halak posted a .922 SV%, and Rask posted a .912 SV%.

After looking at this information and seeing, with your eyes, the kind of shots that they faced - as opposed to just numbers on a spreadsheet, which say that Halak faced tougher shots. Does it now make some sense to say that, considering that Rask's SV% is only .001 higher than Halak's over the past 3 seasons, that after adjusting for shot quality, Halak has out-performed Rask in the minutes that he has played?

Click to read more...

 

[izzy]

shockingly jake allen is the poopy

 

[DownIsTheNewUp]

I like the idea of the stat and I think it's well thought out. But the more results of advanced stats that I see, the more convinced I am that, like regular stats, they only provide very limited insight into the game.

There's just no way guys like Vasilevskiy and Rask are below average goalies.

 

[Seanaconda]

Miller could probably have had some more decent starting years if he didn't only want to play for the ducks for his wife or whatever

 

[Jugitsu]

Seeing Dubnyk, Allen and Ward at the bottom of the list and Gibson, Bobrovsky and Grubauer at the top (ironically the goalies I get the most glimpses of, except for Gibson, no one likes the Ducks) validates this method for me. Even though Dubnyk is my team's goalie, I think he's one of the most overrated goalies in the recent history. I have never liked him apart from that miracle run after he got traded to Minnesota.

 

[Kairi Zaide]

I feel like it's inherently flawed, as to me at least, this metric tells me which goalie are less likely to repeat (i.e. similarly to GF and xGF ratio - the closer to one, the less likely it is that it was a fluke year). How consistent is this metric from year N to year N+1?

 

[CantLoseWithMatthews]

Vasilevskiy and Price had no business being ranked #1 and 2. that was pretty bad

 

[Stupendous Yappi]

I guess Binnington and several others lacked the numbers to make the ranking. But then I’m not sure the value of a ranking that leaves out some of the current starters in the league.

 

[Martin Skoula]

I like the idea of the stat and I think it's well thought out. But the more results of advanced stats that I see, the more convinced I am that, like regular stats, they only provide very limited insight into the game.

There's just no way guys like Vasilevskiy and Rask are below average goalies.

They can be below average at one measurement without being below average overall.

If your team does a good job of breaking up dangerous opportunities game in and game out, letting in a fluky deflection is harder to balance out since you're not seeing that much action.

Meanwhile Gibson can let in a few fluky goals and nobody notices because he's still .915 on 40 shots against.

This type of measurement is naturally biased towards good goalies on bad defensive teams, but it still provides valid observations.

 

[heysmilinstrange]

I guess Binnington and several others lacked the numbers to make the ranking. But then I’m not sure the value of a ranking that leaves out some of the current starters in the league.

Like the OP said, the list is made up of goaltenders who have played at least 5000 minutes over the past 3 seasons.

 

[TomasHertlsRooster]

I feel like it's inherently flawed, as to me at least, this metric tells me which goalie are less likely to repeat (i.e. similarly to GF and xGF ratio - the closer to one, the less likely it is that it was a fluke year). How consistent is this metric from year N to year N+1?

I think it doesn't necessarily matter whether or not the metric is especially repeatable so long as we know that it's at least somewhat valid, and I think that it is. It might be less repeatable than raw SV%, for example, or even xSV%, but that would be because defensive skater performance is more repeatable than goaltending performance.

Having said that, this article might give you some insight into the repeatability of each metric:

xSV% is a better predictor of goaltending performance than existing models

You should read the full article to get a better idea, but this image should help answer your question:

(In this photo, the "xSV%" variable actually refers to Fenwick (unblocked shot attempt) SV% - expected SV%.) So, it does seem that goaltending performance metrics are more repeatable and predictive when they account for shot quality.

I guess Binnington and several others lacked the numbers to make the ranking. But then I’m not sure the value of a ranking that leaves out some of the current starters in the league.

Binnington has played only 1856:25 over the past three seasons. 32 games. In my view, it's simply not fair to these other goaltenders to compare them to a goaltender who has only played 1,856 minutes with all of them coming in one season. Goaltender performance can be skewed heavily by small sample sizes, and goaltenders generally do better with a lesser workload. Even just this season, Binnington played fewer games than teammate Jake Allen.

Having said that, in case you were curious, his number here would be 0.88. Just ahead of Bobrovsky and behind only John Gibson. But it dropped to 0.99 in the playoffs.

What other starters did you have in mind?

 

[Kairi Zaide]

So I wanted to answer some of my hypothesis. All data at 5v5 from Corsica. Note that I will call your metric "GApxGA". Also note this was done very quickly because it's late. I'm only using data since the last lockout because I wanted to filter by minutes and didn't want to prorate the shortened season, and I didn't want a season missing in my dataset. Maybe someone can perform a more rigorous statistical analysis.

Is there a link between GApxGA and total shots against?

It seems like the answer is "no". GApxGA isn't inflated by either low or high volumes of shots against the goalie. R² is pretty insignificant, and the distribution is rather scarse and seems to follow a normal distribution.

Is there a link between GApxGA and high danger shots against?

Things get a little bit more interesting here. I asked myself this question because I had an inherent feeling that goalies facing low volume of high danger shots would fare better. R² is still pretty weak, but still you can see a little influence by HDSA60, as in higher HDSA seems to indicate lower GApxGA. Again, the correlation is still pretty weak, so it's more of a tendency than a rule.

How does GApxGA on a given season affect GApxGA on the next season?

The reason I'm asking myself this is that I expect a metric that is telling about how good a player is to show consistency from year to year, from team to team, and so on. Therefore, I would expect to see a high degree of correlation between two periods. Here, we have a pretty low R², which still shows a slight tendency to repeatability, but I would expect much higher for it to be significant.

In conclusion, while I do not think this metric is bad, I do not think it has much predictive power. When talking about how good a player is, we have to consider a time frame that involves the past (which is known) and the future (which is unknown and needs to be projected). I think this metric you propose has some value in evaluating past performance, but I wouldn't bank on it being repeatable. I think this fits in line with my initial remark that being too far away from the average of 1 indicates unsustainable performance, in a way. There's also a very slight link between it and high danger shots against rates, but as I was saying, it's not considerable enough in my opinion.

 

[Kairi Zaide]

I think it doesn't necessarily matter whether or not the metric is especially repeatable so long as we know that it's at least somewhat valid, and I think that it is. It might be less repeatable than raw SV%, for example, or even xSV%, but that would be because defensive skater performance is more repeatable than goaltending performance.

Having said that, this article might give you some insight into the repeatability of each metric:

xSV% is a better predictor of goaltending performance than existing models

You should read the full article to get a better idea, but this image should help answer your question:

(In this photo, the "xSV%" variable actually refers to Fenwick (unblocked shot attempt) SV% - expected SV%.) So, it does seem that goaltending performance metrics are more repeatable and predictive when they account for shot quality.

Thanks for the link. I'm not as much into goalie analytics as I am into skater analytics, so I will definitely read that tomorrow.

 

[TomasHertlsRooster]

So I wanted to answer some of my hypothesis. All data at 5v5 from Corsica. Note that I will call your metric "GApxGA". Also note this was done very quickly because it's late. I'm only using data since the last lockout because I wanted to filter by minutes and didn't want to prorate the shortened season, and I didn't want a season missing in my dataset. Maybe someone can perform a more rigorous statistical analysis.

Is there a link between GApxGA and total shots against?

It seems like the answer is "no". GApxGA isn't inflated by either low or high volumes of shots against the goalie. R² is pretty insignificant, and the distribution is rather scarse and seems to follow a normal distribution.

Is there a link between GApxGA and high danger shots against?

Things get a little bit more interesting here. I asked myself this question because I had an inherent feeling that goalies facing low volume of high danger shots would fare better. R² is still pretty weak, but still you can see a little influence by HDSA60, as in higher HDSA seems to indicate lower GApxGA. Again, the correlation is still pretty weak, so it's more of a tendency than a rule.

How does GApxGA on a given season affect GApxGA on the next season?

The reason I'm asking myself this is that I expect a metric that is telling about how good a player is to show consistency from year to year, from team to team, and so on. Therefore, I would expect to see a high degree of correlation between two periods. Here, we have a pretty low R², which still shows a slight tendency to repeatability, but I would expect much higher for it to be significant.

In conclusion, while I do not think this metric is bad, I do not think it has much predictive power. When talking about how good a player is, we have to consider a time frame that involves the past (which is known) and the future (which is unknown and needs to be projected). I think this metric you propose has some value in evaluating past performance, but I wouldn't bank on it being repeatable. I think this fits in line with my initial remark that being too far away from the average of 1 indicates unsustainable performance, in a way. There's also a very slight link between it and high danger shots against rates, but as I was saying, it's not considerable enough in my opinion.

Good stuff, but remember that this only includes the result of one model, and it only includes results from 5V5 situations. Adding data from a 2nd model and data from all situations would more than double your sample size, which could help stabilize the data a bit.

Do you have p-values on all of the regressions? It could be that the correlation between GApxGA Year N and GApxGA Year N-1 is significant, but just not very large.

In addition, I definitely get what you're saying about predicting future performance. Hypothetically, if we were to use just this one stat to predict future performance next year, it would still be smart to place a heavier weight on the 2018-2019 season than the 2016-2017 season. Having said that, I think you may be placing too much emphasis on the prediction of future results. Hypothetically, let's say we had a metric that had perfect ability to assess goaltender performance. It might be possible that even with this metric, the R^2 value for prediction of year to year performance would not be very strong, simply because goaltender performance varies greatly from year to year. A metric like SV% might be more predictive of future SV% than this "perfect goaltending metric" would be of future "perfect goaltending metric", but only because team defense+goaltending is more repeatable than just goaltending.

And specifically regarding the heavy outlier players being unsustainable, I do agree with that as well. But this kind of just falls back into the idea that goaltending performance in general is prone to great fluctuation.

 

[Kairi Zaide]

Good stuff, but remember that this only includes the result of one model, and it only includes results from 5V5 situations. Adding data from a 2nd model and data from all situations would more than double your sample size, which could help stabilize the data a bit.

Do you have p-values on all of the regressions? It could be that the correlation between GApxGA Year N and GApxGA Year N-1 is significant, but just not very large.

In addition, I definitely get what you're saying about predicting future performance. Hypothetically, if we were to use just this one stat to predict future performance next year, it would still be smart to place a heavier weight on the 2018-2019 season than the 2016-2017 season. Having said that, I think you may be placing too much emphasis on the prediction of future results. Hypothetically, let's say we had a metric that had perfect ability to assess goaltender performance. It might be possible that even with this metric, the R^2 value for prediction of year to year performance would not be very strong, simply because goaltender performance varies greatly from year to year. A metric like SV% might be more predictive of future SV% than this "perfect goaltending metric" would be of future "perfect goaltending metric", but only because team defense+goaltending is more repeatable than just goaltending.

And specifically regarding the heavy outlier players being unsustainable, I do agree with that as well. But this kind of just falls back into the idea that goaltending performance in general is prone to great fluctuation.

Some good points I may address tomorrow if I think about it. I did not compute p-values since I just plotted the stuff and I'm more used with Minitab than Excel for that kind of stuff. I also realized some glaring flaws with my analysis (link with HDSA60 is a given considering the nature of xGA, might be better to see how other metrics vary depending on the distance of GApxGA average, etc.). But well, it's late, so my mind isn't all there.

 

[Stupendous Yappi]

I think it doesn't necessarily matter whether or not the metric is especially repeatable so long as we know that it's at least somewhat valid, and I think that it is. It might be less repeatable than raw SV%, for example, or even xSV%, but that would be because defensive skater performance is more repeatable than goaltending performance.

Having said that, this article might give you some insight into the repeatability of each metric:

xSV% is a better predictor of goaltending performance than existing models

You should read the full article to get a better idea, but this image should help answer your question:

(In this photo, the "xSV%" variable actually refers to Fenwick (unblocked shot attempt) SV% - expected SV%.) So, it does seem that goaltending performance metrics are more repeatable and predictive when they account for shot quality.



Binnington has played only 1856:25 over the past three seasons. 32 games. In my view, it's simply not fair to these other goaltenders to compare them to a goaltender who has only played 1,856 minutes with all of them coming in one season. Goaltender performance can be skewed heavily by small sample sizes, and goaltenders generally do better with a lesser workload. Even just this season, Binnington played fewer games than teammate Jake Allen.

Having said that, in case you were curious, his number here would be 0.88. Just ahead of Bobrovsky and behind only John Gibson. But it dropped to 0.99 in the playoffs.

What other starters did you have in mind?

Hart maybe. I just find it hard to digest a ranking of the aTop 40 or whatever, and some guys slated to be starters are ineligible for the ranking.

 

[Machinehead]

Another thing Carey Price sucks at?

 

[Machinehead]

I like the idea of the stat and I think it's well thought out. But the more results of advanced stats that I see, the more convinced I am that, like regular stats, they only provide very limited insight into the game.

There's just no way guys like Vasilevskiy and Rask are below average goalies.

Rask has had a sub-.920 save percentage four years in a row. Three of those years, .915 or worse.

He had an amazing peak but he is absolutely a below average goalie right now.

I feel for goaltenders in particular, it takes forever and a day for opinions to change on them despite it being the most volatile position.

 

[Machinehead]

Vasilevskiy and Price had no business being ranked #1 and 2. that was pretty bad

This isn't a perfect ranking but it's sure as hell closer than what people are coming up with.

 

[TomasHertlsRooster]

Vasilevskiy and Price had no business being ranked #1 and 2. that was pretty bad

This isn't a perfect ranking but it's sure as hell closer than what people are coming up with.

Just based on this list, I think you get more mileage out of ranking goalies like this than you do by ranking skaters by something like WAR, for example, and probably for the same reason that WAR is considered more useful for ranking baseball players than hockey players. A shot still has many more factors that go into it than a pitch in baseball, but it's a far more static, 1V1 event than anything else that happens in hockey.

Applying this from a practical, I would probably take HF's official top-20 center list over the top-20 centers ranked by an average of WAR between Corsica and Evolving Hockey over a 3-year sample. But I would definitely take this list of top-40 goaltenders over HF's without thinking twice.

 

[VoluntaryDom]

Rank	Player	Team	Goals Allowed Per Expected Goal (Average)
1	John Gibson	ANA	0.85
2	Sergei Bobrovsky	CBJ	0.88
3	Philipp Grubauer	COL/WSH	0.89
4	Corey Crawford	CHI	0.93
5	Ryan Miller	ANA/VAN	0.95
6	Ben Bishop	DAL/T.B/L.A	0.95
7	Frederik Andersen	TOR	0.95
8	Matt Murray	PIT	0.95
9	Marc-Andre Fleury	VGK/PIT	0.96
10	Braden Holtby	WSH	0.96
11	Pekka Rinne	NSH	0.97
12	Jaroslav Halak	BOS/NYI	0.98
13	Darcy Kuemper	ARI/L.A/MIN	0.99
14	Mike Smith	CGY/ARI	1.00
15	Thomas Greiss	NYI	1.00
16	Cam Talbot	EDM/PHI	1.00
17	Robin Lehner	NYI/BUF	1.00
18	Jimmy Howard	DET	1.00
19	Carter Hutton	BUF/STL	1.01
20	Roberto Luongo	FLA	1.01
21	Martin Jones	S.J	1.02
22	Henrik Lundqvist	NYR	1.02
23	Jonathan Quick	L.A	1.02
24	Jacob Markstrom	VAN	1.02
25	James Reimer	FLA	1.02
26	Connor Hellebuyck	WPG	1.03
27	Tuukka Rask	BOS	1.03
28	Andrei Vasilevskiy	T.B	1.04
29	Brian Elliott	PHI/CGY	1.05
30	Semyon Varlamov	COL	1.06
31	Jonathan Bernier	DET/COL/ANA	1.06
32	Carey Price	MTL	1.06
33	Kari Lehtonen	DAL	1.06
34	Keith Kinkaid	N.J	1.07
35	Cory Schneider	N.J	1.08
36	Petr Mrazek	CAR/DET/PHI	1.10
37	Craig Anderson	OTT	1.11
38	Devan Dubnyk	MIN	1.11
39	Jake Allen	STL	1.12
40	Cam Ward	CHI/CAR	1.12

[TBODY] [/TBODY]

I initially just calculated this for my own purposes and had no interest in starting a thread about it, but after seeing the results I felt like it was worth posting. The numbers pass the smell test even more so than I thought.

These are all the goaltenders to play at least 5,000 minutes over the past 3 seasons. This was just an arbitrary number that I picked and once I saw that exactly 40 goaltenders had played that number of minutes, I figured that number sounded about right.

The data used to calculate these numbers is from both Natural Stat Trick and Corsica.Hockey. For each model, goals against were divided by expected goals against. In order to combine the two models, goals against and expected goals against for each data set were added together. (Corsica's data set slightly varies from Natural Stat Trick's because Corsica is missing a few games and goals.)

The full data is here:

Spoiler

Player	Team	GA (Corsica)	xGA (Corsica)	Goals Allowed Per Expected Goal (Corsica)	Goals Against (NST)	xG Against (NST)	Goals Allowed Per Expected Goal (NST)	Goals Allowed Per Expected Goal (Combined)
John Gibson	ANA	399	482.03	0.83	401	462.4	0.87	0.85
Sergei Bobrovsky	CBJ	437	515.7	0.85	438	474.62	0.92	0.88
Philipp Grubauer	COL/WSH	203	234.61	0.87	205	224.23	0.91	0.89
Corey Crawford	CHI	305	341.78	0.89	306	317.63	0.96	0.93
Ben Bishop	DAL/T.B/L.A	301	322.23	0.93	301	309.58	0.97	0.95
Ryan Miller	ANA/VAN	254	281.6	0.90	254	252.02	1.01	0.95
Frederik Andersen	TOR	513	557.13	0.92	513	517.9	0.99	0.95
Marc-Andre Fleury	VGK/PIT	359	388.65	0.92	359	360.52	1.00	0.96
Braden Holtby	WSH	439	459.8	0.95	440	453.6	0.97	0.96
Pekka Rinne	NSH	405	434.3	0.93	408	403.98	1.01	0.97
Jaroslav Halak	BOS/NYI	325	352.25	0.92	326	314.77	1.04	0.98
Matt Murray	PIT	730	786.55	0.93	373	364.82	1.02	0.96
Darcy Kuemper	ARI/L.A/MIN	246	257.46	0.96	248	241.79	1.03	0.99
Mike Smith	CGY/ARI	406	417.14	0.97	406	396.96	1.02	1.00
Thomas Greiss	NYI	308	320.59	0.96	308	296.69	1.04	1.00
Robin Lehner	NYI/BUF	388	396.58	0.98	388	377.51	1.03	1.00
Jimmy Howard	DET	363	373.6	0.97	365	352.54	1.04	1.00
Cam Talbot	EDM/PHI	466	485.2	0.96	467	446	1.05	1.00
Carter Hutton	BUF/STL	256	264.53	0.97	256	241.99	1.06	1.01
Roberto Luongo	FLA	307	322.4	0.95	307	283.4	1.08	1.01
Martin Jones	S.J	471	479.79	0.98	473	447.66	1.06	1.02
Henrik Lundqvist	NYR	480	492.78	0.97	480	450.13	1.07	1.02
Jonathan Quick	L.A	330	325.92	1.01	331	321.31	1.03	1.02
Jacob Markstrom	VAN	379	388.21	0.98	382	356.92	1.07	1.02
Connor Hellebuyck	WPG	479	485.98	0.99	481	449.64	1.07	1.03
James Reimer	FLA	311	321.56	0.97	311	285.66	1.09	1.02
Tuukka Rask	BOS	371	372.04	1.00	371	348.29	1.07	1.03
Andrei Vasilevskiy	T.B	418	419.88	1.00	418	387.74	1.08	1.04
Brian Elliott	PHI/CGY	296	290.46	1.02	302	280.12	1.08	1.05
Jonathan Bernier	DET/COL/ANA	276	264.19	1.04	276	258.03	1.07	1.06
Semyon Varlamov	COL	336	330.37	1.02	340	309.79	1.10	1.06
Carey Price	MTL	445	437.64	1.02	447	404.15	1.11	1.06
Kari Lehtonen	DAL	234	224.39	1.04	234	215.4	1.09	1.06
Keith Kinkaid	N.J	296	291.31	1.02	300	263.15	1.14	1.07
Cory Schneider	N.J	347	339.84	1.02	347	305.56	1.14	1.08
Petr Mrazek	CAR/DET/PHI	344	329.62	1.04	346	299.08	1.16	1.10
Craig Anderson	OTT	433	411.38	1.05	435	372.48	1.17	1.11
Devan Dubnyk	MIN	449	410.68	1.09	449	396.35	1.13	1.11
Jake Allen	STL	411	378.82	1.08	411	358.38	1.15	1.12
Cam Ward	CHI/CAR	389	362.69	1.07	389	334.73	1.16	1.12

[TBODY] [/TBODY]

In my opinion, this is by far the best way to statistically evaluate goaltenders since it accounts for shot quality. Other metrics like GAA, SV%, Wins, etc. don't account for shot quality. Obviously these expected goal metrics aren't perfect but they do have a strong correlation between actual goals and the methodology behind them is solid.

In addition, I think that "goals allowed per expected goal" is the best way to describe and express this metric. Metrics like delta save percentage or GSAA are not quite as simple and do not explain themselves so clearly. "Goals allowed per expected goal" is a lot more simple IMO. For every 100 expected goals that your team allows, John Gibson allows only 85 goals, and Cam Ward allows 112. That is very easy to explain.

Obviously, if I were ranking goaltenders, my final list would be a little different from the list we have here. For example, factors like consistent performance under a heavier workload would lead me to rank Frederik Andersen over Philipp Grubauer, factors like playoff performance would lead me to rank Braden Holtby over Frederik Andersen, and factors like recency bias would lead me to rank Andrei Vasilevskiy over Martin Jones.

However, I still think the list is pretty damn good, and does a better job of ranking these goaltenders than most skater ranking metrics I've seen. Thoughts on the list?

i like it but corsica's xg model overrates tampas team defense and underrates our goaltending since tampa jukes shot location super hard allowing a ton of slot shots and very few crease shots. with that in mind i think vasilevskiy is above average but far from elite

offensively tampa also slot jukes a bit so corsica sort of underrates tampas team offense as well, but thats unrelated to vasy.

 

[Stewie Griffin]

I think if Vasilevsky and Rask were higher no one would complain about this. It's a great way to actually rank goaltenders, after factoring in the defense in front of them. Dubnyk, for example, has stats that I don't think would suggest he's one of the worst starters in the league, but plays behind a great defense.

Where as Gibson plays behind poop

 

[tony d]

Thanks for this, it's quite an interesting stat.

 

[Midnight Judges]

Doesn't xGoals depend heavily on shot distance? As in, if you shoot from closer in, your expected goals per shot would increase?

 

[Merrrlin]

rip Tuuka and Vas.

The list has some flaws obviously but otherwise it looks pretty good to me. Crawford has to be the most underrated goalie of the last decade.