Advanced Hockey Stats Discussion

[Robin Hood]

I like having confirmation for what my eyes tell me when watching. For example last season I felt Bozak wasn't that great and Grabbo was being misused and wasn't playing as bad as his point total suggested. I thought this from watching all the Leafs games, but reading stats and other opinions that come to the same conclusion further supports you aren't crazy and what you saw with your eyes was reflected in the underlying numbers.

I trust what I see too, but having actual evidence strengthens an arguement more than saying "I saw this so I think this" IMO

I think Bozak >>> Grabo and I've read articles that discuss Grabo + Kessel vs Bozak + Kessel vs Both without Kessel.

[Shep]

Great article on the Minnesota Wild:

http://sports.yahoo.com/blogs/nhl-puck- ... 43871.html

Injuries to 3 of their Top 6 forwards as well as their starting goaltending is a pretty big factor.

[Shep]

I like having confirmation for what my eyes tell me when watching. For example last season I felt Bozak wasn't that great and Grabbo was being misused and wasn't playing as bad as his point total suggested. I thought this from watching all the Leafs games, but reading stats and other opinions that come to the same conclusion further supports you aren't crazy and what you saw with your eyes was reflected in the underlying numbers.

I trust what I see too, but having actual evidence strengthens an arguement more than saying "I saw this so I think this" IMO

Why do you need confirmation? You should feel confident in your hockey analysis.

[Robin Hood]

I like having confirmation for what my eyes tell me when watching. For example last season I felt Bozak wasn't that great and Grabbo was being misused and wasn't playing as bad as his point total suggested. I thought this from watching all the Leafs games, but reading stats and other opinions that come to the same conclusion further supports you aren't crazy and what you saw with your eyes was reflected in the underlying numbers.

I trust what I see too, but having actual evidence strengthens an arguement more than saying "I saw this so I think this" IMO

Why do you need confirmation? You should feel confident in your hockey analysis.

That's not the issue at all. Watching a game does not always give you an unbiased view of what is actually happening.

[Bruyns]

I wouldn't argue against you since Bozak seems to be a better fit for the team. I just have an opinion that Grabovski is better at hockey than Bozak and was given a raw deal by Carlyle. I do understand why the decision was made to buy him out and keep Bozak.

Also I'd say that "stats geeks" (stupid term) on average watch WAY more hockey than the vocal crowd who refuses to have an open mind about them.

[Bruyns]

I like having confirmation for what my eyes tell me when watching. For example last season I felt Bozak wasn't that great and Grabbo was being misused and wasn't playing as bad as his point total suggested. I thought this from watching all the Leafs games, but reading stats and other opinions that come to the same conclusion further supports you aren't crazy and what you saw with your eyes was reflected in the underlying numbers.

I trust what I see too, but having actual evidence strengthens an arguement more than saying "I saw this so I think this" IMO

Why do you need confirmation? You should feel confident in your hockey analysis.

Why wouldn't someone like confirmation? Do you prefer thinking you are right or knowing you are right?

I can have 100% belief in something, but I still value other's opinions on the topic and it is always reassuring when others have the same thoughts on a subject to reinforce your own beliefs. I don't see it as lacking confidence in analysis I view it as further researching an opinion and not being stubborn thinking that only my thoughts are correct and everyone else is wrong.

[Robin Hood]

Also I'd say that "stats geeks" (stupid term) on average watch WAY more hockey than the vocal crowd who refuses to have an open mind about them.

I think the biggest point of frustration of stats guys vs non stats guys is that non stats guys think they can arrive at intelligent hockey conclusions by just watching the game WITHOUT stats. Stats guys meanwhile say do both, it will give you a clearer picture. In response, non stats guys call the stats guys elitist, academic, don't know anything about hockey, have you ever played the game etc.

A statistical revolution happened in baseball under the same conditions and it revolutionized the game. It is happening in the NBA. And it will happen in hockey. Anyone who is too closed-minded to catch-on will get left behind.

[Shep]

Why wouldn't someone like confirmation? Do you prefer thinking you are right or knowing you are right?

I always know I'm right.

[kimmer]

Why wouldn't someone like confirmation? Do you prefer thinking you are right or knowing you are right?

I always know I'm right.

me too

otherwise, try to explain how I flip the following: Burns/Brunner/Garbutt, and instantly upon doing so, they get injured for a week+/suspended 5 games

[Bruyns]

The advent of social media and the accessibility of information and stats I believe is accelerating this movement. We are seeing bloggers with more informative articles and better analysis slowly penetrating the old guard of print media where we read the same tired narratives and hockey cliches in fluff pieces with little substance aimed to create controversy and talking points.

I've had some interactions with Mirtle on Twitter and I find him a little arrogant, but appreciate the angle he brings to the main stream media and believe analysis like he and other bloggers are doing is where hockey analysis is headed and like Shiv mentioned the dinosaurs that refuse to adapt like McGran and Shoalts in our local market will be left behind.

[Bruyns]

Why wouldn't someone like confirmation? Do you prefer thinking you are right or knowing you are right?

I always know I'm right.

You must make a fortune on sports betting

[Nick]

My point on shots being a good measure of possession is on a logical basis - not on a statistical. The original purpose of the measure was to keep track of goalies workload (because they have to respond to every attempted shot) and not on who has the puck. Yes on a team level measured over a long period of time it does closely correlate with possession - but that is where on a logical sense the use of the measure ends. I'm talking theory here - not mathematical correlation.

I believe those measures are being referred to as 'possession' because there is a logical link that having control of the puck, is akin to being superior/ having control of the game (although there are flaws there as well, some of which might even aid in the statistical relation between shot attempts and winning).

the idea of 'team' possession while a player is on the ice is very valid, the reliability of measuring attempted shots while a player is on the ice is somewhat reliable (across teams). Calling shot attempts a possession measure does not make the measure valid and reliable. This is where there is nothing advanced about these statistics, they are using easily available data and looking for meaning - data mining - and there are at least as many errors reached through this method of data extraction as there are useful outcomes.


Anyways. what I would like to do if I had time, or we had enough time, was to actually make a theory, toss out the measures we would use and how they are both valid and reliable measures, and see if we can do a time-sequence model or even a regression to test predictability.

some are easy:
Players that hit more, will get more penalties - all sorts of logical arguments as to why - we can then test that players who hit more do not get get penalized more often,, as well as players who don't hit get more penalties.

part b would be an expansion that teams hit more, get penalized more often, now we've crossed levels and the results would be interesting for a wide variety of reasons.


How it would be of interest to the world would than be if we could cross it with something expected (teams that get more penalties than their opponents lose more often), except if those teams also hit more. Before running the numbers would need to really nail down the logic as why we would expect this - but it's do-able.


Logical connection between measure/cause and outcome, show it cross levels, show an interaction with an even more important measure (winning) at that higher level = strong addition to the understanding of the 2013-2014 NHL season.

edit: oh and suggest the direction of future research - so that you can take on an army of grad students that push your own path to tenure

[Robin Hood]

My point on shots being a good measure of possession is on a logical basis - not on a statistical. The original purpose of the measure was to keep track of goalies workload (because they have to respond to every attempted shot) and not on who has the puck. Yes on a team level measured over a long period of time it does closely correlate with possession - but that is where on a logical sense the use of the measure ends. I'm talking theory here - not mathematical correlation.

I believe those measures are being referred to as 'possession' because there is a logical link that having control of the puck, is akin to being superior/ having control of the game (although there are flaws there as well, some of which might even aid in the statistical relation between shot attempts and winning).

the idea of 'team' possession while a player is on the ice is very valid, the reliability of measuring attempted shots while a player is on the ice is somewhat reliable (across teams). Calling shot attempts a possession measure does not make the measure valid and reliable. This is where there is nothing advanced about these statistics, they are using easily available data and looking for meaning - data mining - and there are at least as many errors reached through this method of data extraction as there are useful outcomes.

I think this is a fairly biased argument.

1. No one is saying Corsi/Fenwick is perfect. But it is a model that provides data with a certain level of confidence.
2. Data mining is the only way to get advanced stats here as well. Whether you use TOI measures, SOG or something else, data mining will be used.
3. Your "theory" goes against what mathematics suggests, which implies that your theory has a level of bias.
4. Corsi numbers can be parsed even further to evaluate portions of the game (e.g. 10 seconds after a faceoff to remove the impact of a faceoff on possession)
5. There are numerous academic papers on the topic about SOG being quite possibly the best indicator on this topic due to their volume. the MIT Sloan conference regularly churns out papers and speakers on the subject.

Not sure if Bruyns wants to add anything here but I feel like your opinion is very biased.

[Bruyns]

Not sure what you feel is the inherent bias in Nick's post. Is he biased against trusting corsi/fenwick as a proxy for possession in your opinion?

The one thing that stuck out to me was this "Calling shot attempts a possession measure does not make the measure valid and reliable". No one has claimed it is a perfect model, but a lot of work has been done in this area to show that it lines up with actual time of possession in the atacking zone and based on that I think the measure can be called reliable, not perfect, but calling it unreliable would go against a ton of academic papers on the topic as Shiv has referenced above. I still agree that logically it seems strange that shot attempts can be such an accurate and reliable proxy for possession, but I can't argue with all the data and math behind it that have shown shot attempts to be the best metric to use when examining puck possession.

[Robin Hood]

Not sure what you feel is the inherent bias in Nick's post. Is he biased against trusting corsi/fenwick as a proxy for possession in your opinion?

The one thing that stuck out to me was this "Calling shot attempts a possession measure does not make the measure valid and reliable". No one has claimed it is a perfect model, but a lot of work has been done in this area to show that it lines up with actual time of possession in the atacking zone and based on that I think the measure can be called reliable, not perfect, but calling it unreliable would go against a ton of academic papers on the topic as Shiv has referenced above. I still agree that logically it seems strange that shot attempts can be such an accurate and reliable proxy for possession, but I can't argue with all the data and math behind it that have shown shot attempts to be the best metric to use when examining puck possession.

You hit exactly why I had a problem with Nick's post. Probably phrased it better lol.

[Nick]

mathematical correlation with flawed logic is not worth anything. your theory and measures must be valid before correlation is worth anything. this is the pursuit of understanding and explaining, with large data sets (such as SOG) it's VERY VERY easy to find correlations, and if too many test are ran, even predictive relationships - the truth is without starting out with your own theory and processing it onward, the error far far far exceeds any reasonable confidence guidelines.


I'm not saying corsi/fenwick are useless, i'm saying they are far from perfect when taken too far; logical leaps are happening because they are being accepted as a measure of possession which is an indicator of team being in-control of the game, which is believed to be related to winning the game. it's too many steps removed.

saying the team that has a higher corsi will win more often, I agree is valid, logical, and supported. it has not been shown to cross levels well (to support comparing players). It was likely realized post-hoc, but than repeated over time and found to be consistent.



I agree individual corsi/fenwick numbers become more accurate from 5v5 starts removed - like I said, I read and follow this stuff too- but now coaching use is at least as big a factor.


I personally love quality of competition/line-mates as adding to understanding, however it actually explains very little - because it partially evens out over a season - aside from saying good teams have more good players, are able to play their good players versus other teams good players more often, and still play their (other) good players against other teams less good players, more often. It's support mathematically, and logically - I have no idea which way it was tested first, i accept it as reasonable use of the data, and strong theoretical understanding.

Your "theory" goes against what mathematics suggests, which implies that your theory has a level of bias.

theory is to be tested, saying fenwick is the best measure for possession is a theory which has supposedly since been tested, within the scientific method supported. - the leafs alone showed serious flaws in this theory, the question shouldn't be -> when will the leafs regress back to what our numbers expect, but rather should be -> what is our theory missing.


We should be clear in not only in what you're saying mathematics suggests, but actually in the role of theory. Especially in use for non-hard-science, the front end is extremely important - and you are down playing that here. there are MBA-esq papers and presentation 100000x's a year, which make use of data in an unfounded manner - because their theory is not sound, or they followed 95% of the scientific method, except they data mined without the proper levels of error (and many many more examples). Hell there are so many research journals out there now that there are thousands of poorly written papers that pass peer review at too low a level - but the academic community knows this, if it's not in a B+ level journal in most fields it's ignored and receives no weighting or further expansion. Once again back to the purpose of the research -> to add to understanding.


Apologies from broken reply - obviously at work and writing in-between actual tasks.

[Nick]

I think it would be fun to craft out a BBKL-made theory that we debate, develop, test and discuss. I'd probably be willing to waste some of my own time and do a portion of the stats.

[Bruyns]

This is an area I'd be interested in contributing to.

I would like to examine what other factors come into play when determining how games are won. We are starting to see that puck possession isn't the be all end all stat that wins games. There are still people out their that believe they are and expect the Leafs to falter this season. If the Leafs continue winning while being hemmed in their own zone and regularly out shot and out chanced it will be a big blow to the advanced stats community.

My problem is I am admittedly lazy and it would take a lot of work to go through data and spearhead any sort of initiative and it is difficult to get motivated to dive into things you brought up like more hits = more PIMs etc. I'd definitely help contribute, examine and discuss any theories you want to look into and it would be cool to come up with a BBKL theory like you mentioned.

I'm a numbers guy and this whole "advanced" stats movement is right up my alley and has added to my enjoyment of the game.

[Nick]

If I could comment quickly and incompletely on the use of advanced stats.

We all know (now) that there has been an acceptance in baseball for use of a statistical analysis in the formation of a team (to some extend at least). They have found a theory, that was first backed up with a numerical breakdown to support the theory, than repeated over seasons, than applied to a team (according to the movie, with complete commitment ).

A crucial step here is an available measure which can be used on an individual level, up to the team level, and across to team comparisons. It is not an approximation of indication of anything - it is the measure for on-base percentage (if i recall correctly). They then used theory/logic to propose that if you get on enough basis, you will get enough runs-in, to win enough games. The correlation/prediction point is from on base to runs-in ; on-base percentage is not an indicator or how often a team gets on base, it is the amount a team gets on base.

Hockey is miles off of this accuracy. I believe we all accept that, we're trying to get to a step in the relationship - though sometimes i question from which side do people think they are approaching it.


Logically possession of puck can be taken to indicate control of the game, which should lead to winning the game - but a major confound (one example) is the final 5 minutes in a 1-goal game when teams change their tactics to puck location rather than puck possession (in the actual sense, not measured sog) where they use the time pressure and physical distance from a position to score as a form of defense - and are actually 'in control of the game' but fenwick is indicating the other team is. So it's not just missing the 'control of game' portion, it's measuring the opposite. I know I've read breakdowns of this exact situation, and for some teams controlling for final 5/2 minutes in 1 goal games can make the measure a more accurate explanation of the outcomes - but then our understanding it rather incomplete, because that control is only being applied post-hoc on an as-needed basis.


IMO there are two sound methods to developing any theory - well not my opinion but the common approaches, inductive and deductive. so we could start as Bryuns just suggested, and work back from what wins games (more goals than your opponent, more often) and then we need measures of more goals, or less goals against, this has a lot of risks IMO to reach a non-repeatable conclusion.

I'd rather begin growing a model outwards, what small plays, that we can measure are steps towards understanding, and if we place them all together, how much of the picture can we paint?

are there different types of PIMS that we can find situations to define? (and record)
what real-time stats form an interaction with goals for/against that we can collect and use? Logically and from out collective hockey knowledge what do we 'know' that isn't yet studied?

[MSP4LYFE]

Only thing more dangerous than ignoring all available information, is using some information incorrectly. A lot of context and sound logic is needed before these in-depth measures will tell you more - to really reach anything useable, finding reliable and valid measures is a first priority, and none of the measures meet strict requirements of prediction. They are great for post-hoc analysis of why something happened- and if the same conditions can be found again, we'd have reason to expect the same outcome but to make a predicative stat (like some say PDO is) takes notably more work to actually be worth a proper 'discussion' in a scientific sense.

Personally I still struggle with the translation of shot attempts (w/ or w/o blocks) to be a good measure of 'possession' - at a team level it has it's uses - as more of the confounds will be randomly balanced out, but at a player level it so full of errors and confounds that finding further value it in is a leap of faith rather than a continuation of logical process (for me).

Now - that doesn't mean I ignore them, and I do follow a several bloggers/writers on the topic - but it's also not why I traded for OEL (just to throw out an example).

I was going to make a post on the subject, and then I realized Nick entered into my brain and did it for me. Great post buddy, I literally agree with every single premise, as well as the conclusions you draw from them.