Expected Goals - what a bunch of garbage

[Guffman]

It's honestly embarrassing how these so-called "advanced stats" don't mirror reality.

I was just looking at an article that showed average goal differential to expected goal differential. For 10 teams, the difference was over 0.5 goals/game. What a joke.

Who the heck came up with this garbage? If you're going to push this type of garbage onto us, why not actually factor in the quality of the players (the shooter and the goalie) to give a more accurate tabulation on what the expected goal should be on that particular shot as opposed to being lazy and just applying some non-applicable league-wide average to it?

If you don't have sufficient data on the players involved, how is it more appropriate to assume that the players are league average players?

Click to read more...

 

[Bear of Bad News]

I was just looking at an article

Seems like if you linked to the article that you're complaining about, there might be enough to actually have a conversation.

 

[bossram]

"I have no idea what I am looking at and don't care to know so it's automatically garbage!"

 

[Spirit of 67]

Think of all the characters you could've saved by typing: I hate the xG stat.

That I would've read.

 

[koyvoo]

"I have no idea what I am looking at and don't care to know so it's automatically garbage!"

While I don’t necessarily agree with the OP, there is an assumption from many analytics diehards that the only reason some people don’t buy in is simply for not understanding the formulas. That is just as shortsighted and closed minded as the OP may be. I can assure, there are a good number of people who fully understand what they’re looking at and are skeptical about how much stock to put into numbers and graphs.

 

[bossram]

While I don’t necessarily agree with the OP, there is an assumption from many analytics diehards that the only reason some people don’t buy in is simply for not understanding the formulas. That is just as shortsighted and closed minded as the OP may be. I can assure, there are a good number of people who fully understand what they’re looking at and are skeptical about how much stock to put into numbers and graphs.

I agree there are things to be skeptical about, but when you start a thread based on the premise of "this garbage!!!!" well, you don't build yourself much credibility.

Like, if OP had gone into it any further, he'd realize that pretty much all public analytics figures recognize the impact of shooting and goaltending talent, and that there are separate (or different) models to integrate those things. The issue is shooting and save percentages can be fleeting or prone to hot/cold streaks. Save for the rare elite players, those percentages tend to regress to average whereas xG is more repeatable.

Like sure, team level xG only captures a blend of shot quantity and location, and pretty much nothing else. But it's shown itself to be repeatable and predictive, which is valuable. Creating a thread to complain that a stat that is intended to only measure shot quantity/location and not measure other stuff is just strange IMO. It's like complaining about why a speedometer doesn't also measure air pressure. Some private models are trying to incorporate passing data into xG, but that's not available to us. So this is what we have.

 

[Doctor No]

Without question, the biggest limitation of hockey analytics lies in the communication of the numbers, the correct takeaways, and the caveats surrounding the limitations.

(Of course, that's the hardest part about analytics in general - in my day-to-day job as an actuary, the math I do is certainly difficult, but the business, communication, and strategic aspects are what differentiates.)

This thread, however, isn't that (especially since there are many formulations of "expected goals" and it'd be nice to be able to explore what the original poster is upset about specifically.)

 

[morehockeystats]

There's an interesting question though:
For a given shot situation on ice, how confounded are the shooter and the goalie to it.
Example: A powerplay, the slap shot from the left circle, Alex Ovechkin is a decently likely shooter, but the goalie can be any goalie.

That means that Alex Ovechkin by his own self affects the probability of a PP slap shot from the left circle going in, so the xG value MAY ALREADY HAVE Alex Ovechkin factored in.

We also do not consider the regular NHL goaltenders to have any specific weaknesses, do we? I remember the commentator saying Bobrovsky was expectedly struggling with the blocker shots during his time as a Philly backup (in the 3-10 playoff game vs. Pittsburgh a few years ago)

 

[Fourier]

Without question, the biggest limitation of hockey analytics lies in the communication of the numbers, the correct takeaways, and the caveats surrounding the limitations.

(Of course, that's the hardest part about analytics in general - in my day-to-day job as an actuary, the math I do is certainly difficult, but the business, communication, and strategic aspects are what differentiates.)

This thread, however, isn't that (especially since there are many formulations of "expected goals" and it'd be nice to be able to explore what the original poster is upset about specifically.)

This has always been my position. Most of the "advanced stats" we see have information built into them, but too often people misinterpret what they can and cannot tell you.

 

[Fourier]

While I don’t necessarily agree with the OP, there is an assumption from many analytics diehards that the only reason some people don’t buy in is simply for not understanding the formulas. That is just as shortsighted and closed minded as the OP may be. I can assure, there are a good number of people who fully understand what they’re looking at and are skeptical about how much stock to put into numbers and graphs.

I have been told on many occasions that I just don't understand the math. Usually by someone who obviously does not understand the math. (I use "math" here pretty generously since mostly we are talking basic arithmetic). My experience with this board is that you won't hear that from peole who tend to post here.

 

[Jared Dunn]

While I don’t necessarily agree with the OP, there is an assumption from many analytics diehards that the only reason some people don’t buy in is simply for not understanding the formulas. That is just as shortsighted and closed minded as the OP may be. I can assure, there are a good number of people who fully understand what they’re looking at and are skeptical about how much stock to put into numbers and graphs.

While this is true, does OP strike you as one of these people? If we're being honest the majority who are adamantly against them just don't understand them

 

[Jared Dunn]

I have been told on many occasions that I just don't understand the math. Usually by someone who obviously does not understand the math. (I use "math" here pretty generously since mostly we are talking basic arithmetic). My experience with this board is that you won't hear that from peole who tend to post here.

I feel like there's about 100 people who truly understand the math and the rest just regurgitate it if it fits their agenda

 

[supsens]

I agree there are things to be skeptical about, but when you start a thread based on the premise of "this garbage!!!!" well, you don't build yourself much credibility.

Like, if OP had gone into it any further, he'd realize that pretty much all public analytics figures recognize the impact of shooting and goaltending talent, and that there are separate (or different) models to integrate those things. The issue is shooting and save percentages can be fleeting or prone to hot/cold streaks. Save for the rare elite players, those percentages tend to regress to average whereas xG is more repeatable.

Like sure, team level xG only captures a blend of shot quantity and location, and pretty much nothing else. But it's shown itself to be repeatable and predictive, which is valuable. Creating a thread to complain that a stat that is intended to only measure shot quantity/location and not measure other stuff is just strange IMO. It's like complaining about why a speedometer doesn't also measure air pressure. Some private models are trying to incorporate passing data into xG, but that's not available to us. So this is what we have.

Its not repeatable or predictable. Your dealing with the average of 30 teams over 1000 games. Any given team on any given night ,any given shot from any different shooter and thats not even counting the fact that the actual number of chances is always unknown until after the game is over and that number is not repeatable or predictable.
Its an average that can't predicted much of anything and thats not even thinking about what goalie saves what

 

[Filthy Dangles]

By far the biggest thing unaccounted for are seam passes, east west plays and making the goalie move.

Another problem I've found with xG are they look at Corsi shooting % and not Shooting %. So if a player has a clear shot toward the net from a certain location, shots into traffic that hit shin pads and shots deflected by sticks are going to be counted in that probability of scoring, which doesn't represent the quality of that specific chance.

Third, the actual information these stats come from, have flaws. The distance and angles are often off in the play by play, sometimes by a good amount.

 

[bossram]

Its not repeatable or predictable. Your dealing with the average of 30 teams over 1000 games. Any given team on any given night ,any given shot from any different shooter and thats not even counting the fact that the actual number of chances is always unknown until after the game is over and that number is not repeatable or predictable.
Its an average that can't predicted much of anything and thats not even thinking about what goalie saves what

I don't think you have any idea what you're talking about. The early shift to xG from Corsi was because team level xG was found to be more repeatable and more predictable of future goal difference, after ~25 games something like an r-squared of 0.33.

Yes, everyone knows on any given night anyone can win. That's how luck works. But if you want to predict future team-level performance, the past 25 games of xG actually gives you a pretty good estimate. It's not a coincidence that most of the playoff-projected teams are also in the top-half of the league in xG%.

By far the biggest thing unaccounted for are seam passes, east west plays and making the goalie move.

Another problem I've found with xG are they look at Corsi shooting % and not Shooting %. So if a player has a clear shot toward the net from a certain location, shots into traffic that hit shin pads and shots deflected by sticks are going to be counted in that probability of scoring, which doesn't represent the quality of that specific chance.

Third, the actual information these stats come from, have flaws. The distance and angles are often off in the play by play, sometimes by a good amount.

I'm pretty sure the most popular xG models use Fenwick (AKA unblocked shot attempts). Yes, NHL PBP data is often flawed, but as long as the error is generally randomly distributed, it all just evens out. In particular venues where shot location data is known to be skewed in one direction, the xG can be adjusted (as EvolvingHockey does).

 

[supsens]

I don't think you have any idea what you're talking about. The early shift to xG from Corsi was because team level Corsi was found to be more repeatable and more predictable of future goal difference, after ~25 games something like an r-squared of 0.33.

Yes, everyone knows on any given night anyone can win. That's how luck works. But if you want to predict future team-level performance, the past 25 games of xG actually gives you a pretty good estimate. It's not a coincidence that most of the playoff-projected teams are also in the top-half of the league in xG%.



I'm pretty sure the most popular xG models use Fenwick (AKA unblocked shot attempts). Yes, NHL PBP data is often flawed, but as long as the error is generally randomly distributed, it all just evens out. In particular venues where shot location data is known to be skewed in one direction, the xG can be adjusted (as EvolvingHockey does).

Its not 'luck' its called unpredictable.

 

[bossram]

Its not 'luck' its called unpredictable.

I've literally told you the conditions and general level of predictability for xG. If you want to be willfully obtuse and don't actually want to learn anything about the stats before you criticize them, then go for it. But don't come into a thread and try to make an argument if you don't actually care about the arguments.

I'm a +EV hockey bettor for the last season and half (when I started getting into sports betting), so you can't tell me everything is completely unpredictable lol.

 

[supsens]

I've literally told you the conditions and general level of predictability for xG. If you want to be willfully obtuse and don't actually want to learn anything about the stats before you criticize them, then go for it. But don't come into a thread and try to make an argument if you don't actually care about the arguments.

I'm a +EV hockey bettor for the last season and half (when I started getting into sports betting), so you can't tell me everything is completely unpredictable lol.

I can check the standing pick the top 8 teams and 'predict' they will all play decent for the rest of the season that has about the same prodiction value does it not?

 

[Bear of Bad News]

I can check the standing pick the top 8 teams and 'predict' they will all play decent for the rest of the season that has about the same prodiction value does it not?

We're kind of back to the problem with the initial post in that you haven't quantified the two things (rigorously) that you want us to compare.

What (exactly) are the two things that you're trying to have us compare the predictive value of?

 

[bossram]

I can check the standing pick the top 8 teams and 'predict' they will all play decent for the rest of the season that has about the same prodiction value does it not?

We're kind of back to the problem with the initial post in that you haven't quantified the two things (rigorously) that you want us to compare.

What (exactly) are the two things that you're trying to have us compare the predictive value of?

Exactly. I laid out the conditions (~25 games of xG can explain future goal differential at an r^2 of about 0.33), but supsens doesn't really understand statistics.

Supsens, you'd have to show that something like current standings points can explain future goal differential at a better rate than xG.

 

[supsens]

Exactly. I laid out the conditions (~25 games of xG can explain future goal differential at an r^2 of about 0.33), but supsens doesn't really understand statistics.

Supsens, you'd have to show that something like current standings points can explain future goal differential at a better rate than xG.

So in plain english what does your r^2 of .33 actually represent?

 

[bossram]

So in plain english what does your r^2 of .33 actually represent?

33% of the variation in actual goal differential can be explained by previous xG. 66% of the variation in goal differential is explained by other factors (including luck/variance).

Given hockey is a very luck-driven sport, being able to predict at an r^2 of above 0.3 is quite good IMO.

 

[supsens]

33% of the variation in actual goal differential can be explained by previous xG. 66% of the variation in goal differential is explained by other factors (including luck/variance).

Given hockey is a very luck-driven sport, being able to predict at an r^2 of above 0.3 is quite good IMO.

Thats not much of an answer I guess google didnt help a lot. R2 does not mean a lot and 33% in the real world is not predictive.

 

[bossram]

Thats not much of an answer I guess google didnt help a lot. R2 does not mean a lot and 33% in the real world is not predictive.

Look I told you what it is. If you want a dictionary definition, it's not going to be any more illuminating. 33% of future goal differential can be explained by past xG differential. It's more predictive than other metrics that people were using before, like standings points or goal differential, which is why it's useful.

There's no such thing as "good" or "bad r^2. The interpretation is what you make of it. Real-world data is messy. Getting 33% of the way there, in a game where ~40% is pure luck, is honestly surprisingly effective.

If you want to disagree with how useful it is, go ahead. But hockey has moved past dinosaur thinking and most teams and private analytics companies are conducting analyses in reference to some kind of xG model.

 

[abo9]

I have been told on many occasions that I just don't understand the math. Usually by someone who obviously does not understand the math. (I use "math" here pretty generously since mostly we are talking basic arithmetic). My experience with this board is that you won't hear that from peole who tend to post here.

so your username is Fourier and people just assume you don't know the math?

Idk if your username is related or not, and it should obviously not be a basis for authority, but I find it quite ironic.