ELOCorsi Oilers

Stat Guy · Sep 6, 2014

Hey all, I recently created a macro discussed in a previous thread about determining a players ELO rating taking faceoffs.

I figured I could easily adapt it to develop a ELO rating using Corsi.

First...
For those who don't know ELO, check out here for an indepth and confusing explaination.
For the borderline imbeciles like myself, in a nut shell this is ELO.
It is a ranking system that takes away points from the loser and gives them to the winner. All players start with a 1200 ELO rating. Higher than 1200 is good....lower is bad. How many points to change hands is determined by their ranking going into the matchup.
If a higher ranked player wins, he gains fewer points from the lower ranked player because he was "suppose" to win. The opposite is true. If the lesser ranked player wins, he will gain greater points.
ELO was created in a 1 vs 1 matchup. Since hockey is a team sport, I use the average ELO of a line, versus the average ELO of an opposing line.
This is an effort to weigh the "Level of competition" versus the "Level of Teammate" in combination with Corsi.
If you don't believe in "Corsi", you might as well not read any further as this stat isn't for you as that is what my ELO rating is based on.
I match a line versus another line in a 5 vs 5 battle.
For every shot, missed shot, blocked shot and goal based off of the NHL realtime stats, I do a ELO calculation.
The line that has the scoring chance gets credited with a win, and steals some of the losers points.
How many points is determined by both lines average ELO rating.
Also...for the ELO experts, I have issued a different "K Factor" for the different events. K Factor is a weighting of importance, and is basically the "maximum ELO Points" available.
Goal= K factor of 5
Shot= K factor of 3
Miss&block= K factor of 2
This is to give greater emphasis to a goal, over a blocked shot.

Ok...an example
Here is a line matchup for a specific event in a game between Tampa and Florida in 2009

Initial ELO

So this event, or play, pits Tampa's line with an average of 1207.75 versus Florida's line rated at 1174.18
On this play Florida's Nathan Horton took a shot on goal, but missed the net. (giving a K factor of 2 for this play)
So despite Horton's poor marksmanship, this does count as a win for Florida in the ELO Corsi department.

So I do a ELO calculation,
Florida will steal 1.10 ELO points from Tampa.
Since Florida was a slight underdog, they will take more points then Tampa would have had they registered the scoring chance.
Had Tampa had a missed shot on net, they would have taken .90 points from Florida.
These lines were pretty evenly matched so the variance was little.
Had Tampa had a line worth 1300 and Florida 1100, and Florida won the event, Florida would have swiped 1.5pts and if Tampa had won with the same ratings, they would only steal .5pts.
That is where the K factor come in. For a goal (with a k factor of 5) the same 1300 vs 1100 match up would see Florida steal 3.8 ELO points while Tampa would steal only 1.2 if there heavily favoured line scored.

So back to our initial example, the NEW ELO Corsi rating for the lines after Horton's missed shot on goal are

Anyways, that is how I work the calculations.
I have so far run every 5 vs 5 event from 2007 until today with the ELO Corsi Rating carried over from one year to the next, and here are the how the Oilers look.
Keep in mind, this is a CAREER stat (or at least from 2007 until today), not a stat for last season. 1200 is the starting point for each player, and is therefor also the league average.

Here are the Top 25 players in the league

#|Pos|Player|Team|ELO Corsi
22 | L | DANIEL SEDIN | VANCOUVER CANUCKS (VAN) | 1528
37 | C | PATRICE BERGERON | BOSTON BRUINS (BOS) | 1476
11 | L | ZACH PARISE | MINNESOTA WILD (MIN) | 1463
13 | C | PAVEL DATSYUK | DETROIT RED WINGS (DET) | 1455
14 | R | JUSTIN WILLIAMS | LOS ANGELES KINGS (L.A) | 1436
20 | C | ALEXANDER STEEN | ST. LOUIS BLUES (STL) | 1433
27 | R | PATRIC HORNQVIST | NASHVILLE PREDATORS (NSH) | 1431
19 | C | JONATHAN TOEWS | CHICAGO BLACKHAWKS (CHI) | 1422
67 | R | MICHAEL FROLIK | WINNIPEG JETS (WPG) | 1421
87 | C | SIDNEY CROSBY | PITTSBURGH PENGUINS (PIT) | 1418
15 | R | BRAD RICHARDSON | VANCOUVER CANUCKS (VAN) | 1415
18 | L | JAMES NEAL | PITTSBURGH PENGUINS (PIT) | 1410
40 | L | HENRIK ZETTERBERG | DETROIT RED WINGS (DET) | 1403
93 | R | JAKUB VORACEK | PHILADELPHIA FLYERS (PHI) | 1399
11 | C | ANZE KOPITAR | LOS ANGELES KINGS (L.A) | 1397
27 | D | ALEX PIETRANGELO | ST. LOUIS BLUES (STL) | 1396
62 | L | CARL HAGELIN | NEW YORK RANGERS (NYR) | 1394
9 | C | MIKKO KOIVU | MINNESOTA WILD (MIN) | 1392
68 | R | JAROMIR JAGR | NEW JERSEY DEVILS (N.J) | 1388
7 | L | DAVID BOOTH | VANCOUVER CANUCKS (VAN) | 1387
12 | C | DEREK ROY | ST. LOUIS BLUES (STL) | 1385
65 | D | ERIK KARLSSON | OTTAWA SENATORS (OTT) | 1385
11 | C | MIKAEL BACKLUND | CALGARY FLAMES (CGY) | 1384
71 | L | NICK FOLIGNO | COLUMBUS BLUE JACKETS (CBJ) | 1383
11|C|JORDAN STAAL|CAROLINA HURRICANES(CAR)|1381

I'm interested in hearing feedback and suggestions for tightening it up.
Every argument against Corsi is valid!
So is the ignoring of Zone Starts which is huge.
But I think this does a decent job of relating Corsi, to the level of competition and level of teammate.

Thanks for reading!

Aceboogie · Sep 7, 2014

This looks pretty cool. I would love to look into this more in the upcoming weeks to test it out and give feedback.

Are you looking to factor in zone starts into this in the future? This is the first time Ive seen this and more initial thought was how would a third linne sutdown line fair against a top line stationed in their own (the 3rd lines) defensive zone? The events (shots/goals) are far more likely to happen in the O zone for the top lines, and the top lines normally get alot more offensive zone starts. You state that a line heavily favoured will "steal" less points from a less favoured line, so this could account for the difference. But a potential issue could still arise because a situation where the face off is in the N-zone will result in the top line stealing as many points as they would steal had they started in the O-zone for a shot/goal. To me it would make sense for a N-zone shift start to count for more points than O-zone start. However this is nitpicking the stat and probaly gets evened out through a larger sample with more events (starting in D zone)

That is my initial thought, but I will research this more when I got time. Thanks for sharing

Stat Guy · Sep 7, 2014

I have thought a lot about zone starts.

One thing suggested to me is to "ignore" any events that happen within the first 10 seconds of a faceoff.

I'm also working on an ELO rating for goalies, and trying to give a heavier weighting to forwards in the offensive zone and dmen in the d-zone when calculating a lines ELO.

Replacement* · Sep 7, 2014

Stat guy.

It isn't just about not believing in Corsi. (although in honest it is that as well) Its that your stat takes data that is already derivative and further derives it. With end results mathematically manipulated to a degree that they become very susceptible to variable noise. Through several calculations of data each time Corsi is used in the equation potentially magnifies a coefficient of how much distortion could be contained in that corsi variable. In a sense multiplying distortion coefficient times distortion coefficient. Not meaning to single out just your stat either, just that its the topic at hand here. I think in general the stat community doesn't reflect on whats being raised.

A further concern for me is you didn't adequately explain why the specific numerical manipulations of the date are done in your elo stat. I kind of missed the logic of what your stat is in response to or how or why its trying to redefine.

ScrillaVilla · Sep 7, 2014

Hey Stat Guy,

Just wondering how the Elo for an individual player is calculated to get the line average. Since its based off 1-1 match ups (like chess) that don't occur in hockey how is it calculated?

oobga · Sep 7, 2014

Hall at the top, Jultz at the bottom? It definitely passes the simple sanity test

Agree that zone start adjustment would help make the stats a bit more fair. Maybe make a "win" that comes within a certain amount of time after an offensive zone start worth less.

doulos · Sep 7, 2014

I find it interesting how many of the names seem to fit well with watching them. There are a few that don't seem quite right, but that could be because my own perception of the players is off as well.

Have to admit that seeing Eberle there at #2 on the list doesn't sit well with me from a 'saw him good' standpoint - not a criticism, but more of a 'Huh' on my part.

I love this stuff, and while the real value of it is still being worked out, keep it up. This work, and the refinement of it, is what will keep pushing this work forward.

nexttothemoon · Sep 8, 2014

Always interesting seeing different takes on various stats and analysis derived from them.

Kudos for doing what had to be a massive amount of work on this.

Personally I'm not a big believer in Corsi/Fenwick/Shots based stats because like Replacement said above... they are somewhat derivative and tangential to what really matters... goals for and against.

Once games are decided by the shot tally then I'll give Corsi more credit.

I do realize over time the teams that outshoot others will likely come out ahead on the scoreboard more often as well... but not always and when dealing with certain players this is even more apparent.

Some players are more accurate or more "discerning"... they simply don't generate a high number of marginal scoring chances... but the ones they do generate are often deadlier and more often lead to goals.

I also agree that zonestarts is a huge factor. A guy like Boyd Gordon is pretty much doomed to be seen as merely mediocre because of his time spent in the defensive zone.

Adjusting for 1st 10 seconds after a faceoff in the d zone is a good start.

http://stats.hockeyanalysis.com does exactly that with their 5on5 ZS adjustment... for all of their goals, shots, fenwick and corsi stats.

I find that simply looking at goals for/against when a player is on/off the ice and then also actually looking at what each player contributes offensively is a decent "quick" test of how each player is doing compared to others... rather than getting too derivative with corsi/shots/fenwick etc.

It's an interesting stat though and like anything likely gives some insight and when taken with what you actually see out there on the ice... gives a fuller picture of the value and contributions of each player.

Stat Guy · Sep 8, 2014

Replacement said:
Stat guy.

It isn't just about not believing in Corsi. (although in honest it is that as well) Its that your stat takes data that is already derivative and further derives it. With end results mathematically manipulated to a degree that they become very susceptible to variable noise. Through several calculations of data each time Corsi is used in the equation potentially magnifies a coefficient of how much distortion could be contained in that corsi variable. In a sense multiplying distortion coefficient times distortion coefficient. Not meaning to single out just your stat either, just that its the topic at hand here. I think in general the stat community doesn't reflect on whats being raised.

A further concern for me is you didn't adequately explain why the specific numerical manipulations of the date are done in your elo stat. I kind of missed the logic of what your stat is in response to or how or why its trying to redefine.

I totally agree with you.
"Let's take a vague snapshot stat like Corsi, and multiply it by an even vaguer stat."
We end up with vague squared.

But I think there is a small bit of truth to the end result .
There just isn't a way for an average Joe to measure the 1000's of little things that happen in a game.
Corsi and Fenwick show us how the ice tilts when a player is on the ice.
But it doesn't show the level of teammate or level of competition. If I were to play RD with Crosby, Datsuk, Ovi and Chara, I'd have decent Corsi numbers too.
That is my motivation for trying to redefine Corsi. Remove the ability the pad stats by tilting the ice, and collecting Corsi points when blessed with skilled line mates or Vise versa

Stat Guy · Sep 8, 2014

ScrillaVilla said:
Hey Stat Guy,

Just wondering how the Elo for an individual player is calculated to get the line average. Since its based off 1-1 match ups (like chess) that don't occur in hockey how is it calculated?

The ELO rating is done line average vs line average.
The rating change for the line average is then assigned to each of the players ratings.

For example
Line A average is 1234
Line B average is 1167

Line B wins the event giving these lines new ratings of
Line A= 1231
Line b= 1170

Line a average increased by 3 pts, so I increase each player on that lines rating by 3 pts.
Line b has all players rating decrease by 3 pts.

I considered a couple ways to do change the individual ratings.
A)I debating dividing the points won/lost among the players based on how many points they brought into the event.
B)I considered dividing all points won/lost evenly among the players on the line.
C)I considered the way I did it.

Option A was my 1st choice, but I couldn't decide how to best make it work.
Should the top rated player get the lions share of the points for an event won, or should the lowest ranked person because he exceeded expectations? I believe the top ranked player, should get the most for the win, as he will also lose the most for a loss. The calculation is what made me choose option C, as it was way simpler for my first run through.
I think I know how I'm going to do it, and I was hoping no one asked this question as I wasn't comfortable with the way I did it.
I was in the 2010 season when I decided, but I wanted to get some other feedback before re-running seasons. (It takes about 12hrs per season for the macro to run on my POS laptop)

nexttothemoon · Sep 9, 2014

Whether we collectively believe in this stuff or not... keep it up and you'll probably land a job with an NHL team soon.

Stat Guy · Sep 9, 2014

nexttothemoon said:
Whether we collectively believe in this stuff or not... keep it up and you'll probably land a job with an NHL team soon.

And if I can't get a with an NHL team, maybe I can get one withe the Flames.

I am Lorde · Sep 10, 2014

Stat Guy said:
I totally agree with you.
"Let's take a vague snapshot stat like Corsi, and multiply it by an even vaguer stat."
We end up with vague squared.

But I think there is a small bit of truth to the end result .
There just isn't a way for an average Joe to measure the 1000's of little things that happen in a game.
Corsi and Fenwick show us how the ice tilts when a player is on the ice.
But it doesn't show the level of teammate or level of competition. If I were to play RD with Crosby, Datsuk, Ovi and Chara, I'd have decent Corsi numbers too.
That is my motivation for trying to redefine Corsi. Remove the ability the pad stats by tilting the ice, and collecting Corsi points when blessed with skilled line mates or Vise versa

I love the idea of using ELO to improve (presumably) the corsi stat, and I disagree with your comment about ending up with vague squared. If anything it's the opposite no?

Based off my tiny knowledge of ELO or any other statistical analysis, it seems to me this is a stat that will "pull" players back towards 1200 the further up or down they go. Meaning that if anything the "vagueness" or error is convergent (reduced)... correct me if I'm wrong?

But then again it's not a linear scale so I see where you and replacement are coming from.

Mc5RingsAndABeer · Sep 11, 2014

Fantastic work! This is just awesome. Based upon the top ranked players, it seems like it also has good predictive value (which is really all that matters)

How does this handle special teams?

Stat Guy · Sep 11, 2014

5RingsAndABeer said:
Fantastic work! This is just awesome. Based upon the top ranked players, it seems like it also has good predictive value (which is really all that matters)

How does this handle special teams?

Right now it is purely 5 on 5 rating.

It will take some creative thinking to create a special team ELO.
Because it is based off of shot attempts, and teams on the PK seldom get any, (and really, shot attempts on the PK or not very important to PKers)
Perhaps I can get count a shot on goal as a win for the PP team, and a "Zone clear" as a win for the PK team, then players can be rated based on the level of competiton vs level of teammate.

It might work. I'll fiddle with it!

Chubaka7 · Sep 11, 2014

Based on a cursory reading of the ELO system it seems to me that one of the major problems is that the values of K given appear to be very arbitrary. I understand that goals are seen as more valuable than shots, misses or blocked shots, but if the system is based on probabilities then shouldn't the relative value of each be based the relative probability of each event occurring within a game or shift?

If I were to base the K value of a goal I could easily imagine it being somewhere around 10 times higher than that of a shot, given that league average shooting percentages are somewhere around 10% or less (i don't know the actual stat). I would imagine that it should be possible to estimate relative probabilities for shooting %, shots, and misses/blocks and test new K values quickly enough and see how that changes the results.

I would also take issue with the idea of assigning a greater change in points to a higher rated player in the case of a "win or loss", as they are not always the lead contributor to the event. I know you didn't go that route and that was probably wise.

Overall, this stuff is always interesting and leads to fun and stimulating discussion.

Mc5RingsAndABeer · Sep 11, 2014

Stat Guy said:
Right now it is purely 5 on 5 rating.

It will take some creative thinking to create a special team ELO.
Because it is based off of shot attempts, and teams on the PK seldom get any, (and really, shot attempts on the PK or not very important to PKers)
Perhaps I can get count a shot on goal as a win for the PP team, and a "Zone clear" as a win for the PK team, then players can be rated based on the level of competiton vs level of teammate.

It might work. I'll fiddle with it!

I'm not sure how often zone clearances happen.

Can you do something like this?

+ to the 5 players on the PP if a goal is scored
- to the 4 players on the PK if a goal is scored

+/- balanced so that the net gain is 0
+/- allocated so that a team performing at league average on the PP and PK for that season end up with 1200 ratings for special teams

Would that work?

Chubaka7 · Sep 12, 2014

5RingsAndABeer said:
I'm not sure how often zone clearances happen.

Can you do something like this?

+ to the 5 players on the PP if a goal is scored
- to the 4 players on the PK if a goal is scored

+/- balanced so that the net gain is 0
+/- allocated so that a team performing at league average on the PP and PK for that season end up with 1200 ratings for special teams

Would that work?

Technically, that should work. This would probably work more like a traditional chess ELO system so that PP is expected to successfully score at a league average rate and the opposing PK team is expected to successfully kill a penalty at a league average rate.

Points would have to be awarded to the PK players should they kill the penalty and then penalize the PP for not scoring. Since PK is generally much more successful than PP, the K values for PP would need to be much higher than PK K values to balance things out.

Stat Guy · Sep 12, 2014

You gave me a good idea for how to do PP/Pk.

The 1st step of ELO is to figure out a players chance of winning given the current ratings of the players competing.
Since we know the success rate of an average power play, we can work in reverse to see what ELO ratings are required to give us that success rate.

So rather then starting with an ELO of 1200, I will start the PKers at approx 1400 and the pp at 1000.
This will project a success rate of about 90% for 1min of pp time for the PKers. (I will fine tune the starting numbers to precisely the right ratio.)
Then by using the shift charts from NHL.com, I can "battle" line versus line. A goal is a win for the PPers, no-goal is a win for the PK. I will then multiply the change in ELO by the length of shift relative to 1 minute.
Eg, a line goes for 1min10sec without scoring, if they had the initial ratings of 1000 (pp) and 1400(Pk)
For a 1minute shift the PKers gain 0.45 pts for minute, so multiply the 0.45 x 1.16 and the actual rating change for a successful 1:10 kill is .522pts.
The next line takes over and gives up a goal 30 seconds later.
That line will get credit for 30 seconds of kill, .45 x .5, but then loses 4.45pts when the goal is scored.
Of course for every point the Pk gains or loses, the PP does the opposite.

Not perfect yet, but I think this a good place to start! We can see how accurate it is by seeing how much the PP ratings drift from the 1400 average and we can see how the pp guys drift from the 1000 average.

Beerfish · Sep 12, 2014

I hear Jeff Lynne has pretty good puck possession stats.

Stat Guy · Sep 12, 2014

Beerfish said:
I hear Jeff Lynne has pretty good puck possession stats.

I had to google it... But it made the lol more genuine!

Stat Guy · Oct 14, 2014

I'm going to update this thread on a game by game basis.
So here is how the ELOCorsi changed for each player in the two games so far this season. This will give us a broad overview of how a player is trending.
Who is improving, who is slumping. (According to possession ELO ratings)

Why are the starting numbers different from the chart posted earlier?
I've incorporated Zone Starts into the ELOCorsi rating.
So to summarize, The ELOCorsi rating:
a) Is based on Corsi...Shots attempted while a player is on the ice for versus shot attempts against. Only in 5vs 5 situations
b) Includes level of teammate
c) Includes level of opposition
d) Includes zone starts
e) 1200 is an "average" player

Coming soon
An ELOPowerplay (90% done) and ELOPenaltyKill (90%done) along with an ELOFaceoff rating (99.9% done).
ELOGoalie is in the brainstorming stage (0.1% done)

Player | Team | Starting ELO | Game1 Elo | Change Gm1 | Game2 ELO | Change Gm2
JEFF PETRY | EDMONTON OILERS (EDM) | 1178.7 | 1196.5 | 17.8 | 1197.1 | 0.6
TAYLOR HALL | EDMONTON OILERS (EDM) | 1280.4 | 1288.2 | 7.8 | 1287.1 | -1.1
MARK FAYNE | EDMONTON OILERS (EDM) | 1288.5 | 1280.6 | -7.9 | 1286.2 | 5.6
JESSE JOENSUU | EDMONTON OILERS (EDM) | 1168.1 | 1182.2 | 14.2 | 1186.9 | 4.6
NAIL YAKUPOV | EDMONTON OILERS (EDM) | 1169.7 | 1169.4 | -0.3 | 1172.5 | 3.1
JORDAN EBERLE | EDMONTON OILERS (EDM) | 1297.1 | 1306.2 | 9.2 | 1302.7 | -3.6
TEDDY PURCELL | EDMONTON OILERS (EDM) | 1317.5 | 1319.4 | 1.9 | 1319.6 | 0.3
JUSTIN SCHULTZ | EDMONTON OILERS (EDM) | 1106.1 | 1111.2 | 5.1 | 1113.5 | 2.3
ANDREW FERENCE | EDMONTON OILERS (EDM) | 1070.9 | 1082.8 | 11.9 | 1076.8 | -6.0
MATT HENDRICKS | EDMONTON OILERS (EDM) | 1162.9 | 1175.2 | 12.3 | 1179.8 | 4.6
MARK ARCOBELLO | EDMONTON OILERS (EDM) | 1197.8 | 1195.6 | -2.2 | 1192.8 | -2.8
BOYD GORDON | EDMONTON OILERS (EDM) | 1232.0 | 1247.4 | 15.3 | 1253.8 | 6.4
LEON DRAISAITL | EDMONTON OILERS (EDM) | 1200.0 | 1195.5 | -4.5 | 1197.9 | 2.4
DAVID PERRON | EDMONTON OILERS (EDM) | 1261.2 | 1259.2 | -2.0 | 1256.0 | -3.2
BRAD HUNT | EDMONTON OILERS (EDM) | 1188.2 | 1202.1 | 13.9 | 1197.6 | -4.6
BENOIT POULIOT | EDMONTON OILERS (EDM) | 1330.2 | 1329.2 | -1.0 | 1330.7 | 1.5
NIKITA NIKITIN | EDMONTON OILERS (EDM) | 1156.9 | 1156.9 | 0.1 | 1163.9 | 7.0
RYAN NUGENT-HOPKINS | EDMONTON OILERS (EDM) | 1208.0 | 1221.3 | 13.4 | 1216.5 | -4.9

Jumptheshark · Oct 14, 2014

I will just add my old rant here of how people are looking too closely at the numbers and not enough on the chemistry and make up of the team and the play on the ice.

Only thing I have liked so far this year, other posters are now looking at the Hall-Nuge-Eberle line and realizing that it is not as good as it should be

ekcut · Oct 14, 2014

jumptheshark said:
I will just add my old rant here of how people are looking too closely at the numbers and not enough on the chemistry and make up of the team and the play on the ice.

Only thing I have liked so far this year, other posters are now looking at the Hall-Nuge-Eberle line and realizing that it is not as good as it should be

You are 100% right that some people do look too closely at numbers.
But that doesn't mean you should completely ignore them.

Advanced stats are like sugar in a coffee.
You have to find just the right amount.
Too much or too little ruins the cup.

Of course there is always "that guy" who has to loudly proclaim that he drinks it black, and anyone who adds sugar isn't a real coffee aficionado.

Stat Guy · Oct 17, 2014

Here is the current top 10

ELOCorsi Oilers

Registered User

Registered User

Registered User

Checked out

Registered User

Tier 2 Fan

Registered User

and again...

Registered User

Registered User

and again...

Registered User

Registered User

5-14-6-1

Registered User

Registered User

5-14-6-1

Registered User

Registered User

Registered User

Registered User

Registered User

Rebooting myself

The Refs shot JFK.

Registered User

Ad

Ad

Ad