Grabovski Hypothesis – Regression Refined (offense)

28 Aug

1st Regression and Post

I received a good comment yesterday about whether ‘shots for’ being related to Fenwick and Corsi were messing with the regression.  So, I separated all of that out.

Corsi – Fenwick = Blocked shots (BS)

Fenwick – Shots for = Missed shots (MS)

I re-ran the regression with and without the lockout year from the 07-08 season…..

Results

Missed shots and blocked shots have no statistically significant affect on goals for.  That seems pretty obvious, right?  If a shot is blocked or missed the net, then why would it have an affect?  I think the idea with Fenwick and Corsi “tilting the ice”, or having a lot of possession, does not affect offense statistically according to this regression.  On offense, the only things that seem to matter significantly are shots that actually hit the goalie and shooting percentage.

Again, if a player can individually put more shots on goal or be more accurate than a person is replacing, he would benefit the team.  He could also assist his players by getting them the puck in certain positions to be more accurate or to get the puck on net.  However, it appears the possession portion as defined by Corsi and Fenwick are irrelevant on offense.

More on defense coming soon….

Details

With lockout season….

Estimate              Std. Error             t value Pr(>|t|)

(Intercept)          -98.299082          25.991111            -3.782    0.000218 ***

SF                           0.076925              0.002881              26.704  < 2e-16 ***

SA                           0.003641              0.002532              1.438     0.152307

BS                           0.002288              0.004109              0.557     0.578428

MS                         -0.003707             0.005360              -0.692    0.490186

Sh.                          16.144123            0.307568              52.490  < 2e-16 ***

Sv.                          -0.283125             0.290146              -0.976    0.330600

OZFO.                   0.111255              0.262291              0.424     0.672000

DZFO.                    -0.340800             0.209741              -1.625    0.106112

east                       1.269434              1.335174              0.951     0.343124

west                      1.531076              1.402439              1.092     0.276556

yr13                       0.659509              2.735007              0.241     0.809751

yr12                       -0.013865             1.072697              -0.013    0.989703

yr11                       -0.152273             1.074736              -0.142    0.887503

yr10                       0.352642              0.957111              0.368     0.713017

yr9                          -0.045403             0.849630              -0.053    0.957448

yr8                          NA                          NA                          NA          NA

Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Without lockout season….

Estimate              Std. Error             t value                  Pr(>|t|)

(Intercept)          -1.464e+02          1.207e+01           -12.135                 <2e-16 ***

SF                           7.962e-02            1.167e-03            68.241                   <2e-16 ***

SA                          -6.315e-04           1.086e-03            -0.582                    0.562

BS                           2.436e-03            1.658e-03            1.470                     0.144

MS                         -2.519e-03           2.191e-03            -1.150                    0.252

Sh.                          1.827e+01           1.409e-01            129.625                 <2e-16 ***

Sv.                          7.535e-02            1.303e-01            0.578                     0.564

OZFO.                   -9.744e-02           1.195e-01            -0.815                    0.416

DZFO.                    -7.999e-02           9.176e-02            -0.872                    0.385

east                       -1.961e-01           5.760e-01            -0.340                    0.734

west                      -8.845e-02           6.023e-01            -0.147                    0.883

yr13                       NA                          NA                          NA                          NA

yr12                       3.841e-01            4.270e-01            0.900                     0.370

yr11                       3.014e-01            4.280e-01            0.704                     0.482

yr10                       4.671e-01            3.778e-01            1.236                     0.219

yr9                          9.021e-02            3.310e-01            0.273                     0.786

yr8                          NA                          NA                          NA                          NA

Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

About these ads

4 Responses to “Grabovski Hypothesis – Regression Refined (offense)”

  1. Anon August 29, 2013 at 5:08 am #

    Just a note: when people say Grabovksi will improve the Caps’ Corsi and Fenwick, the assumption is that a lot of that improvement will show up in shots on goal. (No one thinks that Grabovski will add a lot of missed shots and blocked shots by the other team without anything ending up on net.)

    Shots, Fenwick, and Corsi are used somewhat interchangeably because they’re all similar. Shots obviously have a more direct connection with goals, which is what we’re ultimately trying to measure here. Corsi and Fenwick expand the sample size, so they’re slightly better mid-season predictors of team success (http://objectivenhl.blogspot.com/2011/02/shots-fenwick-and-corsi.html).

    Sometimes, people prefer Fenwick because a) it maintains the sample size advantage without adding shot-blocking skill into the equation like Corsi does, b) it correlates better with hand-tracked scoring chances, and c) it seems like arena biases (i.e. Florida and Nashville seem to count 5-10% more shots on goal than the average arena) aren’t as big if you combine missed shots. (That last point is easy to think about–a shot going just wide that the goalie might have touched on its way by will count as a shot on net in one place and maybe as a missed shot in another–but is, as far as I know, unverified.)

    I think the people who use Fenwick would move to shots on net if you could show that shooting and missing the net is in large part due to ability, not randomness; one recent look says ability (http://www.broadstreethockey.com/2013/8/19/4631928/why-all-shot-attempts-arent-created-equal), the older one everyone cited before then says no (http://vhockey.blogspot.com/2010/05/blocked-shots-luck-or-skill.html).

    To reconcile that with what you’ve done here: Since you’re looking backward at an entire season, it wouldn’t surprise me if, after 80 games, the sample size advantage of Fenwick and Corsi over shots no longer outweighs the added noise you get from including missing shots and blocked shots (since, obviously, shots, shots + misses, and shots + misses + blocks are not all exactly the same).

    The reason Corsi is so often cited is in large part because that’s what Behind the Net uses, and for a few years BtN was the best resource for these numbers.

    It’s probably worthwhile to examine differences between the three at an individual level–Grabovski may not have the same Corsi impact as he does shots on goal impact. But there should be a big shots on goal impact, too–if you shoot enough, some of those shots will find their way to the net. I don’t remember exactly what the average conversion rate is between the three, but I think it’s something like 1 Corsi ~ 0.75 Fenwick ~ 0.6 shots on goal. So if the second line this season is 15 Corsi / 60 better than the second line last year (Ribeiro had a relative Corsi per 60 of ~ -10, Grabovski +1 but +15, +21, and +14 the three years before that), if shots and Corsi follow the average conversion rate, that should be an extra 9 shots in shot differential per 60 (~4 games).

    • bourciertm August 29, 2013 at 12:21 pm #

      The regression shows that on average there is a goal for every 14 shots on goal (obviously there is defense not accounted for here. So, if you say they get an extra goal from Grabovski every 8 games (this would include him improving other players too) then this would be an extra 10 goals he produces in a season. Not 10 more goals for him, but for the people around him too. It is still a pretty low number. I think it makes sense to go to shots-for as the regression shows this as the best predictor. Corsi and Fenwick are useful, but they only (mostly?) are good for predicting outcomes because they have shots-for in the equation.

      I appreciate your input! More to come as I go forward….thank you.

      • Anon August 30, 2013 at 2:30 am #

        10 goals doesn’t sound like a lot, but if the Caps had that on their team this year (well, six goals, since it was a short season), they would have finished second in the conference and third in the league in goal differential. Nothing to sneeze at.

        I wouldn’t go so far as to call the Caps “Cup Contenders” (like others have) after this signing, but I do think it helps them solidify their spot among dark horses (the tier below Pittsburgh/Boston/Chicago/Los Angeles etc), looking at it in terms of goal differential.

        As it so happens, the Caps finished a respectable 6th in goal differential in 2013. (I think I commented the other day that GD and wins are highly correlated– something like r = 0.9.) I think we can all agree they were probably a little (or a lot) fortunate to have such a nice GD, so we’d expect a lower ranking next year even if Ribeiro had stayed. But if that +10 per full season from Grabovski can cover up for the shooting% drop off (which will probably hit the power play pretty heavily) then instead of the Caps overperforming by ranking 6th, they can rank 6th legitimately, by “true talent.” That would give them a GD around +30-35.

        An average team by definition has a GD of zero; the Stanley Cup Finalists since 07-08 have been +73*, +31, +51, +25*, +62*, +11, +77, +51*, +19, +15*, +38 (rated to 82 GP), and +91* (rated to 82 GP). (* denotes winner)

        +30 is well back of the recent elite teams (Detroit x2, Vancouver, Chicago x 2), but it’s still good enough to be within striking range.

        The Capitals may not be a top-5 team next year, but they have the pieces in place to be knocking on the door of that group. There are other teams in the same situation, just outside the top-5, so the team isn’t quite yet a standout. But at least they seem to have separated themselves from the mediocre class of the league.

        It may still lead to a first-round exit–in fact, that’s probably the most likely scenario, given that 22/30 teams are eliminated by May. But the way I see it, in theory, the top eight teams in the league advance to the second round, and I think the Caps are one of the best eight teams in the league. In theory, the top four teams advance to the conference finals. I don’t think the Caps are there yet, but they can be if they get to face a worse team–which is a possibility now that they’re about as good or better than at least half the teams that will make the playoffs in the first place.

Trackbacks/Pingbacks

  1. Grabovski Hypothesis – Regression Refined (defense) | Sport Exec in Training - August 28, 2013

    […] Regression Refined (Offense) […]

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

Follow

Get every new post delivered to your Inbox.

Join 174 other followers

%d bloggers like this: