Total Hockey Rating (THoR): A comprehensive statistical rating of NHL Players

[schuckers]

Indeed one can think of this as like Corsi in that it puts emphasis on shots and shot differential. We do that because we're tying events to the probability they lead to a goal. But unlike Corsi we give some weight to shots by location (x,y) and type of shot according to their probability of being a goal. (Before I get hammered on RTSS and shots, we adjust shot location by rink and then we break the O zone in to ~50 grids to get our probabilities of a goal.)

As Scott Cullen has noted we don't give credit for scoring a goal and that over long periods that seems to matter. We're thinking about ways to account for this that give credit for scoring but also recognize short term fluctuation in scoring rates.

 

[Megalodan]

good idea poor execution

 

[Bear of Bad News]

good idea poor execution

Please elaborate.

We tend to expect a higher level of discourse in this subforum.

 

[Mathletic]

Indeed one can think of this as like Corsi in that it puts emphasis on shots and shot differential. We do that because we're tying events to the probability they lead to a goal. But unlike Corsi we give some weight to shots by location (x,y) and type of shot according to their probability of being a goal. (Before I get hammered on RTSS and shots, we adjust shot location by rink and then we break the O zone in to ~50 grids to get our probabilities of a goal.)

As Scott Cullen has noted we don't give credit for scoring a goal and that over long periods that seems to matter. We're thinking about ways to account for this that give credit for scoring but also recognize short term fluctuation in scoring rates.

Hello Mr. Schuckers! I wasn't aware you were a member of the board.

Been a fan of yours at the MIT conference the past couple years. Not to mention Milbury and Burke were complete dicks with you last year. I like your idea behind your DIGR stat as well. Much more representative of a goalie efficiency than the various point shares stats.

Keep up the good work, don't forget that "creative people must be stopped" as coined by David Owens.

 

[overpass]

Indeed one can think of this as like Corsi in that it puts emphasis on shots and shot differential. We do that because we're tying events to the probability they lead to a goal. But unlike Corsi we give some weight to shots by location (x,y) and type of shot according to their probability of being a goal. (Before I get hammered on RTSS and shots, we adjust shot location by rink and then we break the O zone in to ~50 grids to get our probabilities of a goal.)

As Scott Cullen has noted we don't give credit for scoring a goal and that over long periods that seems to matter. We're thinking about ways to account for this that give credit for scoring but also recognize short term fluctuation in scoring rates.

Adding data on the puck movement before the goal was scored may help. Getting goalies moving side to side with lateral puck movement is an important part of shot quality that isn't incorporated by shot location data.

 

[schuckers]

Hello Mr. Schuckers! I wasn't aware you were a member of the board.

Been a fan of yours at the MIT conference the past couple years. Not to mention Milbury and Burke were complete dicks with you last year. I like your idea behind your DIGR stat as well. Much more representative of a goalie efficiency than the various point shares stats.

Keep up the good work, don't forget that "creative people must be stopped" as coined by David Owens.

Thanks. Been reading here awhile and decided to post recently that arcticicehockey does much less analytics that aren't Jets related and given that THoR got some attention here. Burke was playing his role and Mibury was quite nice after.

 

[schuckers]

Adding data on the puck movement before the goal was scored may help. Getting goalies moving side to side with lateral puck movement is an important part of shot quality that isn't incorporated by shot location data.

There is plenty of information like this that would be very useful. In addition to your list, the degree to which the goalie is screened and the percentage of passes completed/tipped would be a start.

 

[Cunneen]

There is plenty of information like this that would be very useful. In addition to your list, the degree to which the goalie is screened and the percentage of passes completed/tipped would be a start.

Just wanted to let you know Mr. Schuckers that I'm a huge fan. Your work on Thor and DIGR is awesome. I'm just a junior in highschool but your no doubt one of my role models (I'm planning on majoring in either statistics or Applied/computational Mathematics and Statistics and I want to work heavily in the area of Hockey analytics).

 

[schuckers]

Just wanted to let you know Mr. Schuckers that I'm a huge fan. Your work on Thor and DIGR is awesome. I'm just a junior in highschool but your no doubt one of my role models (I'm planning on majoring in either statistics or Applied/computational Mathematics and Statistics and I want to work heavily in the area of Hockey analytics).

Thanks. I'd recommend taking as much programming and CS as you can and don't neglect the communication side. Good writing and public speaking are important.

 

[Cunneen]

Thanks. I'd recommend taking as much programming and CS as you can and don't neglect the communication side. Good writing and public speaking are important.

thanks, really appreciate the advice.

 

[Cunneen]

Thanks. I'd recommend taking as much programming and CS as you can and don't neglect the communication side. Good writing and public speaking are important.

Would you recommend double majoring in computer science and stat in college, or Major in one and minor in the other.

I have no programming or computer science experience at all (taking a Java programing course next year as a senior). I do have a strong math background and am taking AP Calc BC and AP Stat next year as a senior.

Thanks,
Pierce Cunneen

 

[Aela*]

I really look forward to this being a great thing that works very well and isn't a bit subjective and prone to human error and that all of the statistics are there, including the 'non obvious ones' like say, time on ice spent in front of the goalie. Doesn't seem like it's the 'be all end all' to hockey analytics yet, but I hope it gets there.

 

[schuckers]

More on THoR

Since we introduced the Total Hockey Ratings or THoR back in February there has been plenty of interest and some skepticism. Since it is a fairly in-depth system it has taken some time but we have put together an evaluation of THoR.

Part I on the reliability, i.e. how consistent it is to itself, can be found LINK HERE. The year to year correlation of THoR values for players is about 0.65. For CorsiRel it is about 0.5.

Part II on the validity, how close it is to 'ground truth', of THoR. This is a bit tricky since we don't have 'true' player values. THoR is based upon probabilities of events leading to goals and so it should have some inherent validity. We looked at a proxy for the validity of THoR by looking at the probability that winning a game that ends is regulation is a result of out producing your opponent on the things that THoR counts. By this metric THoR does well but slightly less than Fenwick or Corsi. The link is THIS ONE.

Finally, in Part III we look at some of the players that have been associated with THoR namely Tyler Kennedy and Alexander Steen. We also look at some other recent players and their THoR values. Here is THAT LINK.

Schuckers

 

[hella rights]

I was looking at your NP20 regression model and was wondering, given that on-ice events rarely result in a goal in the subsequent 20s, if you ever considered using a zero-inflated model?

http://en.wikipedia.org/wiki/Zero-inflated_model

 

[wgknestrick]

Are the "units" of THoR (win shares)?

Personally, I still like GVT's units of goals above replacement. It is easier to relate to IMO.

 

[schuckers]

I was looking at your NP20 regression model and was wondering, given that on-ice events rarely result in a goal in the subsequent 20s, if you ever considered using a zero-inflated model?

http://en.wikipedia.org/wiki/Zero-inflated_model

We thought of a bunch of alterations but not ZIP's explicitly.

 

[schuckers]

Are the "units" of THoR (win shares)?

Personally, I still like GVT's units of goals above replacement. It is easier to relate to IMO.

Units for THoR are Wins Above Replacement.

 

[wgknestrick]

Units for THoR are Wins Above Replacement.

Is there any way to convert the values into goals above replacement? I would be more than excited to compare them to GvT and study the differences (especially for defensemen and goaltenders - where I suspect GVT is off).

Do you have any spreadsheets to post?

 

[schuckers]

Is there any way to convert the values into goals above replacement? I would be more than excited to compare them to GvT and study the differences (especially for defensemen and goaltenders - where I suspect GVT is off).

Do you have any spreadsheets to post?

6 Goals above replacement = 1 Win above replacement.

Will post some results shortly now that there is enough data from the current season.

 

[hella rights]

We thought of a bunch of alterations but not ZIP's explicitly.

I'm not sure it would be the most appropriate model since it assumes a separate data generating process for the zeros, but it might be worth checking out.

Also, I was poking through the THoR 2011-2013 spreadsheet and noticed Ovechkin was ranked 335, sandwiched between Adam Pardy (?) and Jared Spurgeon (???), how does that happen?

 

[schuckers]

On Ovechkin, there is no doubt that he has world class talent. Performance was an issue in 2011-12. Remember these articles:

What's Wrong with Alex Ovechkin

A Look at Ovechkin's Decline

Ovechkin's Drop Off

 

[Bear of Bad News]

New (yesterday) article on THoR from a St. Louis fan site:
http://www.stlouisgametime.com/2014/1/26/5347356/universal-measures-of-player-value-thor

 

[schuckers]

Yikes, I knew that was coming but did not expect such a poor job.

 

[schuckers]

Here's my reply in the comments to the post at stlouisgametime

1. We do NOT treat all events equally. By assigning each play a different value, the NP20, for the response of our regression, we weight them differently. Thus, a shot in close is worth much much more to a player's value than a faceoff win at neutral ice since the former is much more likely to lead to a goal.

2. We looked at events and saw not much change in the cumulative probability of a goal after 10 seconds and we doubled the time frame to 20 seconds to be conservative. We're looking at different responses from Dellow so we're likely to have different results.

3. The results you've presented here are not the latest. We've updated the model and put out the results from this newer model several times since the results from the paper (which is what you use here). Tyler Kennedy comes out very well (on a per play basis) since he is 2nd in the league in SOG/60 over the time period for these results.

4. We've put out the results for the full model with events from event strength, powerplay and penalty kill. The link is here:
www.statsportsconsulting.com/thor. At that same link are some other details including the fact that we are adjusting for rink effects, score effects, zone starts, and home ice.

5. Running the regressions is not easy. It takes at least 12 hours on a very fast computer. Our design matrix is often 300000 rows by 1200 columns.

6. What makes THoR different is that we adjust for the other players on the ice when the event is occurred. Thus we are trying to estimate the value of a player while accounting for who they play against and who they play with as well as the other contextual factors mentioned in #4 above. We think we've had some success in this regard since THoR has a very high year to year reliability
THoR All Events.

7. Schuckers is spelled with two c's.

 

[Bear of Bad News]

Neil Greenberg takes on Michael:

http://www.washingtonpost.com/news/...150-players-in-the-nhl-not-according-to-thor/

I see Michael's already responded.