If not, do we have sufficient data to make it happen? Is scoring chance information available anywhere?
If not, do we have sufficient data to make it happen? Is scoring chance information available anywhere?
http://www.sloansportsconference.com/?p=648
This is a presentation done by Michael Shuckers on his Defense independent rating of NHL goaltenders. Very interesting and worth watching the presentation and/or reading the paper.
Interested in this too. Also interested in a stat like SV% Close (like when a game is within 2 goals, 1 goal, or tied). Me thinks Pekka Rinne would have a way better SV% close than his overall SV% because when he loses, boy does he lose bad.
Taco, when you get time, I wouldn't mind hearing your take on this if it's of interest to you...
Sorry for the delay in response; I was in Ireland for eleven days.
The notion of evaluating goaltenders on a benchmark shot distribution is an interesting one, and something that's been tried out to a lesser degree in other articles - in particular, I've seen many analyses that rely upon weighting all goaltenders' ESSV, PKSV and PPSV by the same proportion of situations.
I've done something similar (as have others) where one "risk adjusts" each shot faced by a goaltender. In this fashion, instead of molding a goaltender's shot distribution to a league norm, you evaluate a goaltender based upon league-normed expectations. So, for instance, if a goaltender faces a shot that (historically) has a 20% chance of going into the net, he gets credit for 0.2 goals prevented (if he stops it) and gets credit for -0.8 goals prevented (if he doesn't).
Ultimately, both of these types of models are only going to be as good as the RTSS data that feeds into it. Not all wrist shots from location X are the same (and even this assumes that each scorer accurately categorizes shot type and location), and once you start distinguishing between "Brett Hull" and "Chris Dingman", you run into sample size considerations.
Accepting this, the math behind what Michael did appears sound, and it's an interesting way of viewing the problem.
I am most interested in Schucker's way because he seems to have used perhaps the most advanced (mathematically) method I've seen in evaluating goalies. Now advanced doesn't necessarily mean better, but from what I have read of Schucker's work I have come to trust his analysis. Obviously I trust the math (being that he his a professor of Statistics at St. Lawrence), but he seems to do a good job in his past papers.
You'd be surprised at some of the backgrounds of people that we have contributing here.
I was not aware, my apologizes. I do not mean to insult anyone here either. I'm happy if we have some people with a lot of data analysis experience or backgrounds, so that I can learn from those people (I'm only 17 but I plan on majoring in some form of applied and computational mathematics and statistics in college).
Sorry for the delay in response; I was in Ireland for eleven days.
The notion of evaluating goaltenders on a benchmark shot distribution is an interesting one, and something that's been tried out to a lesser degree in other articles - in particular, I've seen many analyses that rely upon weighting all goaltenders' ESSV, PKSV and PPSV by the same proportion of situations.
I've done something similar (as have others) where one "risk adjusts" each shot faced by a goaltender. In this fashion, instead of molding a goaltender's shot distribution to a league norm, you evaluate a goaltender based upon league-normed expectations. So, for instance, if a goaltender faces a shot that (historically) has a 20% chance of going into the net, he gets credit for 0.2 goals prevented (if he stops it) and gets credit for -0.8 goals prevented (if he doesn't).
Ultimately, both of these types of models are only going to be as good as the RTSS data that feeds into it. Not all wrist shots from location X are the same (and even this assumes that each scorer accurately categorizes shot type and location), and once you start distinguishing between "Brett Hull" and "Chris Dingman", you run into sample size considerations.
Accepting this, the math behind what Michael did appears sound, and it's an interesting way of viewing the problem.
The issue is in the accuracy of the RTSS data for shot location. Its brutally inaccurate to the point of uselessness. Which seems to be why these exercises like DIGR don't end up have all that much predictive or explanatory value because of GIGO.
Agreed - I should have said this more clearly.
Its a bummer that the NHL is so bad at that stuff, especially seeing what they are doing in baseball and basketball these days.
I'm friends with a guy that's building his own system of shot value based on video and incorperating both shot location and movement rather than the RTSS stuff that he's supposedly getting closed to finished on. Hopefully he'd be able to crack the problem.