Monday, November 24th, 2014

On Corsi and Faceoffs

27

I’m fairly skeptical of much of the statistical analysis being kicked around these days. I accept that hockey is a clash of probabilities and that events come more or less randomly. That makes statistical analysis possible, but I think all of the efforts to apply them to improve our assessment of individuals fail, partly because the sample size is always small and more importantly because individual numbers are always hopelessly polluted by the team context.

I don’t really object if someone tries to squeeze something interesting from the numbers, but I don’t take them very seriously either. Greg Ballentine likes Corsi rankings enough that he devotes three recent posts to them. In one, he notes that Ryan Kesler improved from 35th to 1st in the league standings, largely because his role changed when Manny Malhotra was signed. In another, he explains why Ryan Clowe finished second to Kesler last season. And finally, Greg explains how the statistic clearly shows that Saku Koivu was really bad last year.

Greg’s examples, it seems to me, make a good case against the Corsi statistic. First, both the Kesler and Clowe stories tell us how much influence context has – neither Kesler nor Clowe could have done it playing on a different team or even playing in a different role on the same team. In other words, this is not really an individual statistic. Second, while Ryan Clowe is a very good player, any statistic that ranks him as the second best player has some credibility problems. Greg explains it, but explanations can’t compensate for a number that ranks Ryan Clowe as one of the best players in the league. Third, does the number fill a real need? I don’t think we need a Corsi number to know that Koivu’s game has fallen off a cliff.

On the other hand, I really like the kind of statistical analysis Tyler Dellow does in this piece. He tests a common claim – that faceoffs are important on special teams – by using relatively reliable team level data. He presents the evidence well and I think the conclusion – that faceoffs aren’t that important – is compelling.

Finally, Tyler’s study has a practical application applies the conclusion to a choice the Oilers have on the power play next season. He thinks it would be fine to go without a natural centre on the first power play unit. (I agree. I don’t think the Canucks were hurt because the second unit power play didn’t have a centre for much of the season.)

That the conclusion jibes with a position I’ve held for a long time does not surprise me but that’s not really why I liked the post. If the data had shown that the faceoff was important, I would have had to reconsider my position.

Tyler and I do disagree on one point. When explaining why winning the first faceoff is insignificant, Tyler writes:

I’ve put some thought into why the difference might be so small and the answer that I’ve come up with is that, when you’re facing an elite PP team, whether you win the draw or lose the draw, you’re going to have to taste the poison. If you lose the draw, they get to set up and throw their best at you for a minute and a little more, subject to any clears. If you win the draw and ice the puck, what happens?

The reason the difference is small is because it is a game of puck position, not puck possession. Whether you win or lose the draw, the puck is deep in their end and you have an extra man. We mash together the words “win the draw and ice the puck”. Those are two separate things and the easier part by far is winning the draw.

Generally, a defenceman hurries back while the forwards take a leisurely skate back to the red line or far blue line. They aren’t hard seconds to play. It would be interesting to take a look at the average shift length on the PP for stars in shifts where they won the initial faceoff versus those where they lost. I strongly suspect we’d find that the shifts tend to be longer when they’ve lost the initial faceoff.

If the penalty kill wins the draw and clears the puck, it is a very good thing. A good PK will harass the puck the entire 200 feet back. If the power play can’t get set up, it won’t score often.

Be Sociable, Share!

Comments

27 Responses to “On Corsi and Faceoffs”
  1. Hawerchuk says:

    “any statistic that ranks him as the second best player has some credibility problems”

    Tom, I’m normally a fan of your work, but this is so lazy it sounds like Don Cherry is writing your blog.

    You know, David Backes was #2 in the league in +/- last season. Does that mean +/- has credibility problems? Jeff Schultz was #1 in 2009-10. Blake Wheeler was #2 in 2008-09. We could go on.

    Corsi, like any other statistic, needs to be understood in the context of other factors. Sneering at it because, like any other simple statistic, it doesn’t provide a unified measure of a player’s complete value doesn’t contribute anything to the larger discussion of hockey analysis.

  2. I tend to agree with Gabe on Corsi. The fact that Corsi demands more context to understand what’s going on in the evaluation of individuals is, to me, a feature rather than a bug. I think people are looking for “the magic bullet” statistic far too often. I mean, every statistic is, to some extent, a team statistic since we’re talking about a team game. It’s quite true that Ryan Kesler wouldn’t have had the best Corsi number in the league if he were playing a different role, but he probably wouldn’t have had over 40 goals either. That doesn’t make either of those statistics useless, it just means that both of them demand the use of other information for a better interpretation. If you want Corsi to rank players from best to worst, it’s going to be terribly disappointing (just like most statistics). If you want it to help you understand what’s going on in the game, it can be very helpful, especially when used in concert with other statistics and visual observations.

    As to the faceoff question, I tend to agree with you. It would be really interesting to see which teams are more efficient at clearing the puck within the first ten seconds of a PK (either by winning the draw and icing the puck, or winning another battle), and how much impact that has on whether or not the penalty is killed. I’d bet that it would be much greater than what Tyler found from looking at faceoff wins and losses.

  3. Tyler says:

    Whether you win or lose the draw, the puck is deep in their end and you have an extra man. We mash together the words “win the draw and ice the puck”. Those are two separate things and the easier part by far is winning the draw.

    See the thing is, there’s data to support the idea that goalscoring falls a lot when you lose a faceoff, although this was at evens, not on the PP. See this post.

    You get the extra man on the PP but the other team can’t hammer it down the ice consequence free.

    I think I can test this though and see if the data shows that shots and goals are way down in the first X seconds after you lose a PP faceoff. I’m sure that they will be.

    • IamJoe says:

      Found it!!!

      This is the two pieces I always think of whenever a discussion of faceoffs come up:
      http://www.puckprospectus.com/article.php?articleid=239
      http://www.arcticicehockey.com/2010/8/20/1633298/offensive-zone-faceoffs-and

      The first is basic 5v5 play, the second incorporates data for 5v4. It’s basic shots for data, no goal scoring or SH% info. That was part of the issue I had with your faceoff post the other day – This shows that each faceoff is important for about 5-10 seconds, and (at least the way I understood it) you were only looking at the first faceoff of each PP, so going into it, if we accept Gabe’s data here as true, you already know that the first faceoff will only effect ~10s of the PP (less if it generates an immediate shot), so you’re only looking at 1/12th of the total possible time on the PP, and making conclusions from that. The way I’m seeing it is that you have to include all faceoffs on the PP, and that could amplify any effects to become significant, for example if there are 3 faceoffs on a given PP, you’re now talking about 1/4 of the PP time being impacted by faceoff results, instead of 1/12th.

      I don’t know that a league wide average would necessarily work here (i.e. “there are 2.81 faceoffs per PP in the NHL”) because I would think some teams freeze the puck more than others. Maybe good PK’s have fewer faceoff opportunities because the PP spends more time passing to get a better shot, rather than just jacking shots at the net? I don’t know, seems like an interesting idea.

      (btw Gabe, you should add that Faceoff piece to your FAQ so its easier for me to find and show to people)

    • Tom says:

      See the thing is, there’s data to support the idea that goalscoring falls a lot when you lose a faceoff, although this was at evens, not on the PP.

      Am I missing something here? I think this is compelling evidence that there is a huge difference as to where the faceoff is taken, but it doesn’t say much – anything? – about who won the faceoff. Whether I win or lose a draw in the offensive end, I think I am much more likely to get the next shot, the next chance, the next goal.

      I expect winning the draw will will improve my percentages somewhat but “who wins” will be much less important than “where is it?” and when you factor in the fact that winning or losing the draw comes down to close to 50-50 anyway, the overall impact will be small.

      I think this is largely why we see “momentum”. When a team is playing well, its like all four lines are winning their shifts. We’ll see one team dominate for a stretch through several line changes. Then there is a shift and the other team carries the play for a while. When a team really sucks, almost everybody sucks. When a team is going great, everybody is going great. I think this happens because what happens on this shift depends to some degree on what happened on the previous shift.

  4. Tom says:

    You know, David Backes was #2 in the league in +/- last season. Does that mean +/- has credibility problems? Jeff Schultz was #1 in 2009-10. Blake Wheeler was #2 in 2008-09. We could go on.

    Of course the plus minus has credibility problems. These three examples scream “problem” with the statistic. What does this number purport to tell us? It seems to me that both the plus minus and Corsi plus minus purport to do exactly what you say – provide a unified measure of an individual player’s value. They don’t do it very well because it is impossible to separate the individual skater from the team. I realize that it is possible to place either number into a context, but in the end that just leaves you with a number that may or may not say anything. The context is that this number is polluted by the team and the players role within the team.

    My contribution to the discussion of hockey analysis is buried in this post. Obviously, Greg’s time is Greg’s time and far be it from me to tell him how to spend it. However, I wish a lot less time was spent trying to do the impossible – measure individual accomplishment – and a lot more time was spent doing studies like Tyler did here.

    • I think the mistake is in thinking that either Corsi tries to give a unified measure of an individual player’s value. It doesn’t. At the individual level, Corsi tells us where the puck is spending most of its time when a given player is on the ice, which I think is very useful information but its’ obviously not any kind of comprehensive evaluation of a player’s ability. We seem to agree on that last bit, but disagree on whether knowing where the puck spends most of its time when a give player is on the ice is a useful piece in player evaluation. If that’s the case (and please correct me if I’m wrong), why don’t you find that information useful?

      • Tom says:

        At the individual level, Corsi tells us where the puck is spending most of its time when a given player is on the ice, which I think is very useful information but its’ obviously not any kind of comprehensive evaluation of a player’s ability.

        But it is used to evaluate a player’s ability. That’s why it is broken down by player and why Greg listed the league’s best and the league’s worst. If it is not supposed to reflect the player quality, what does it reflect?

        It would be very useful information if we could tell how much of the result is attributable to the given player and how much of it is attributable to the other players on the ice and how much of it is attributable to the position of the puck when the given player begins his shift.

        • Corsi is used to evaluate a player’s ability most effectively in concert with a number of other statistical measures and visual observations. On its own, Corsi makes a point about where the puck is when a player is on the ice. That alone tells you something about a player, but it’s a very incomplete picture, much like using goals alone would offer a very incomplete picture. Michael Grabner, for instance, had a very good year, but I don’t think a reasonable argument can be made for him as one of the twenty best players in the league last season. Nor do I think it’s reasonable to think that the twenty best players in the league were all forwards. The use of goals is flawed too unless we add some context.

          If we use Corsi (or goals or whatever) in combination with other information like the role of a given player (who he plays with, who he plays against, where he starts his shift), then we get a fuller picture; using “With/Without” analysis can help to single out players who tend to drag linemates down or lift them up; and the “how much” questions are being worked through with a few general rules of thumb finding some currency already. This article, for instance, tries to correct for where players begin their shifts:

          http://objectivenhl.blogspot.com/2009/08/corsi-corrected-for-starting-shift.html

          Even with that adjustment, the picture obviously isn’t perfect, but I don’t think we’ll arrive at one magic number for player evaluation, so I don’t use perfection as the standard for useful. Numerical analysis and observational analysis both have the challenge of disentangling individual and team performance. It’s not easy to do with either method (or even with both working together), but I don’t think that makes the efforts in either arena futile.

  5. Hawerchuk says:

    “It seems to me that both the plus minus and Corsi plus minus purport to do exactly what you say – provide a unified measure of an individual player’s value.”

    Where, exactly, do they do this? I’ve written numerous times that Corsi predicts about 40% of a *team’s* winning percentage. The other 60% is due to many other issues. Similarly, I’ve written about the difficulty of using Corsi to evaluate individual players even after you’ve adjusted for faceoff zones and quality of competition. If that’s “purporting to provide a unified measure of an individual player’s value,” then we own different English dictionaries.

    The great irony of your claim is that aside from Tom Awad, statistical hockey analysts have made no claims to even having a unified player rating – and Tom readily admits the flaws in his work. People come to the table expressing the exact opposite of the confidence that you claim they have.

    • Tom says:

      Similarly, I’ve written about the difficulty of using Corsi to evaluate individual players even after you’ve adjusted for faceoff zones and quality of competition. If that’s “purporting to provide a unified measure of an individual player’s value,” then we own different English dictionaries.

      Fair enough. If Corsi does not evaluate the quality of an individual player, what do we make of Kesler’s Corsi? What makes it a useful statistic?

      • Ralph says:

        His QoC and zone start tell us he wasn’t used in as defensive a role as before. His Corsi tells us he (or, at least, the Canucks while Kesler was on the ice) hammered his opposition.

      • Hawerchuk says:

        I think you conclude first that Kesler/Raymond/Samuelsson/Hamhuis/Bieksa is a hell of a unit – the forwards, in particular, had very positive puck possession despite roughly 50% offensive zone starts as a group and relatively tough competition. When you dig deeper into that crew, I think you find that Kesler and the D are the big drivers of the results. We can put some dollar figures on Kesler if you want, but they’ve got huge error bars on them.

  6. Name says:

    I agree with Tom about corsi. Faceoffs not mattering for PP effectiveness, or in general, I disagree with fully. I believe teams like the Sharks and Canucks who dominate teams in the faceoff circle are afforded huge advantages.

    But as for Corsi, it is way overused, and way overvalued. I went to fear the fin the other day, the Sharks SB Nation blog, to read analysis on the Havlat vs Heatley trade, and it was all corsi. That’s how they came to their conclusion about the trade, that Havlat is better 5 on 5, and a superior two way player (they even started referring to him that way, “the Sharks traded Sniper Dany Heately for two-way forward Martin Havlat….are you freaking kidding me?? Havlat a two way player?”) But that was the whole analysis, everything was based around these inherently flawed corsi statistics. Even the basis of corsi, that shot quality always evens out, so we just have to measure shot quantity, is inherently flawed. The strategies and styles some teams play lead to giving up a greater number of shots, but reducing quality ones, whereas some teams strive to block or prevent every single shot, no matter where it comes from. Therefore some players, just by playing on a certain time, will inherently be on the ice for more shots against. It doesn’t mean they’re giving up more quality scoring chances, or making lots of defensive mistakes, or failing to control play offensively.

    That’s why corsi is flawed. In addition to it being way overused right now (and Hawerchuk, come on, you know as well as I do that it IS being used to draw conclusions, not just yield supplemental information. Just go look at every single Fear the Fin trade and/or player analysis), and way overvalued, it is also inherently flawed at its core. That means that even if you only used it for supplemental information, you would run into anomalies and bad information. Compounding this problem even further is that you would have no idea when the corsi number you’re looking at is a fair representation of the player, and when it’s an anomaly. There are however many players playing in the NHL, hundreds of them. If the base underlying logic behind corsi is that shot quality evens out, even if that were true 99% of the time, for 99% of the players–and it’s nowhere near that high, but let’s just say it was—that would mean one out of every 100 players would still have very misleading corsi numbers. If 600 or so skaters play in the NHL in the every year, that’s 6 players. Odds are, at least one of them is in the top 100 NHL players, maybe even 2. That’s a very significant player with a very misleading corsi. That, in and of itself, is a problem, because every time you look at a player’s corsi, there is no way to know if you’re looking at the misleading one. And that’s just with 1 in 100 players misrepresented. In reality it’s much, much higher than that.

    That’s the problem with Corsi. You’re better off looking at scoring chances created vs scoring chances against where the player was at fault, although that also has the inherent flaw of what’s a scoring chance and what isn’t, and you still run into the qualcomp and teammate quality problem. If you don’t want to take someone else’s word for what’s a scoring chance and what isn’t, you could try some mix of regular +/- combined with a more reliable quality of competition and quality of teammates. But you can’t determine quality of competition and quality of teammates by corsi. So how do you determine it?

    Goals scored while on the ice and goals scored against while on the ice are both objective statistics, so if you could combine +/- with taking into account teammate quality and opposition quality, you might have something there, although shooting percentages of teammates and goaltender quality against and all that would have to be accounted for as well. Regardless, corsi is too flawed to be trusted for most things.

    • Buttercrunch7 says:

      C’mon Fear the Fin rarely uses CORSI. Let’s not make stuff up, okay?

    • Hawerchuk says:

      This entire comment was pulled from your posterior.

      1. Fear the Fin is not a stats blog. Never has been, never will be. They can talk about whatever they want without it reflecting one iota on statistical analysis.

      2. “Even the basis of corsi, that shot quality always evens out, so we just have to measure shot quantity, is inherently flawed. The strategies and styles some teams play lead to giving up a greater number of shots, but reducing quality ones, whereas some teams strive to block or prevent every single shot, no matter where it comes from. Therefore some players, just by playing on a certain time, will inherently be on the ice for more shots against. It doesn’t mean they’re giving up more quality scoring chances, or making lots of defensive mistakes, or failing to control play offensively.”

      Name these teams and players. Thanks.

      3. “Hawerchuk, come on, you know as well as I do that it IS being used to draw conclusions, not just yield supplemental information.”

      Since when does it matter that someone else misuses a statistic? Holy crap, Corey Perry led the league in goals, OMG, MVP!!1! Wow, goals must be an inherently-flawed statistic.

      4. “You’re better off looking at scoring chances created vs scoring chances against where the player was at fault”

      Where’s your evidence for this statement?

      5. “Goals scored while on the ice and goals scored against while on the ice are both objective statistics, so if you could combine +/- with taking into account teammate quality and opposition quality, you might have something there”

      I already did that. Turns out Corsi is better.

      You should spend a few minutes reading about the metrics you’re bashing and then take another stab at this.

      • “Name these teams and players. Thanks.”

        Ok. Here are a few names for you to consider.

        Some guys who can drive shooting percentage: Sidney Crosby, Marian Gaborik, Nathan Horton, Bobby Ryan, Martin St. Louis. Henrik and Daniel Sedin. Alex Tanguay. Jason Spezza.

        Some guys who can suppress shooting percentage: Marco Sturm, Travis Moen, Tyler Kennedy, Taylor Pyatt, Shawn Thornton, Chris Drury, Jeff Carter. Torrey Mitchell. Kamil Kreps.

        More details on my website. http://hockeyanalysis.com/2011/07/19/some-names-for-gabe/

        • Hawerchuk says:

          Oh, David Johnson…If there was a “statistical” “analyst” that Tom Benjamin should be mad at for filling the internet with sophistry, it’s you…

          I told you I wouldn’t engage with you unless you took a tack other than “write the opposite of what Gabe writes.” So we’re done…

        • Magdelina says:

          I feel staiesfid after reading that one.

  7. Name says:

    team, not time.

    and goaltender quality, period, not just against. I realize it reads wrong.

  8. As the poster who wrote the first Corsi posts that Tom is commenting upon, I thought I should comment – though I am a little late to the party.

    I believe that Corsi based analysis is one of the strongest tools we have to understand hockey. I believe that I am showing it on my blog. Tom jumped off on some of the more early posts in a summer of sabermetrics posts. I intend to make contextual adjustments to the players to come up with a better set of rankings. I am well aware that they will not be perfect even after i have corrected for the effects that I have some idea how to correct. Today, I looked at team Corsi ratings and will begin the adjustments by trying to remove team effects in the future. This drops Ryane Clowe and so do other contextual adjustments.

    The main reason for posting the raw unadjusted Corsi ratings are because they are interesting. they are something that most people havent seen (Tom – You are a big Canuck fan – did you have any idea Ryan Kesler led the league in Corsi last year? I doubt it). It also allows me to show the effects of various adjustments and how they change rankings – hopefully making things better fit the “eyeball test”.

    Asking the raw rankings to fit with our impressions of how good players are and throwing them out when they don’t makes as much sense as throwing out goals as a meaningful stat because last year Blake Comeau outscored Henrik Sedin and we should all know Henrik is a better player.

  9. Roberto says:

    It’s been a while since I read Moneyball, but as I recall, there wasn’t simply one stat that dictated that one baseball player was better than another. Defensive and offensive stats were weighed together, and collectively one could estimate to a fairly high degree of accuracy how many runs a team would score and give up over a season, and therefore how many wins one could expect for the team. The statistics turned out to be so much more relevant in predicting a player’s quality than how good a player looked while doing it, that Billy Beane took it to the extreme of never watching the players play so as not to be unduly influenced by what his eyes told him.

    Since Corsi numbers are so highly dependent on a) the players the player plays with the most, and b) the role the player plays on a team, it would seem to me to be a useful stat for determining how a player might fit in with a team in the role required of him. This is assuming, of course, that the stats are available for that player within that role.

    This being the case, I don’t know how any GM interested in signing or trading for a player could ignore this data.

Trackbacks

Check out what others are saying about this post...
  1. [...] interesting statistical debate sprung up today started by Tom Benjamin who wrote about his skepticism of the Corsi statistic.  In it Tom comments on the fact that Ryan Kesler and Ryan Clowe ranked so highly in corsi in [...]



Speak Your Mind

Tell us what you're thinking...
and oh, if you want a pic to show with your comment, go get a gravatar!