A New Way To Measure Deployment – Expected Faceoff Goal Differential

Zone starts are not that great of a metric. Although certain players do tend to be put out almost exclusively for offensive or defensive purposes, the reality is that for most players’ zone starts have a relatively small effect on a player’s performance. And yet, many hockey writers still frequently qualify a player’s performance based on observations like “they played sheltered minutes” or “they take the tough draws in the defensive zone”. Part of the problem is that we’ve never really developed a good way of quantifying a player’s deployment. With many current metrics, such as both traditional and true zone starts, it’s difficult to express their effect except in a relative sense (i.e. by comparing zone starts between players). So when a pundit says that a player had 48% of his on-ice faceoffs in the offensive zone, it’s difficult to communicate to most people what that really means.

Going beyond that, even if we know that 48% would make a player one of the most sheltered skaters in the league, the question that we should ask is so what? Simply knowing that a player played tough minutes doesn’t give us any information that’s useful to adjust a player’s observed results, which is really the reason that we care about zone starts. We know that if you start your shifts predominantly in the defensive zone, you’ll likely see worse results, but zone start percentages don’t tell us how much worse they should be. Traditional deployment metrics are too blunt of a tool – they provide a measurement, but not one that gives any context to the performance numbers that we really care about.

To create a better zone start metric, one that we can use to adjust our observed results, we have to go back to basics. While most analyses to date have looked at the effect of zone starts on a player’s Corsi Percentages, these studies implicitly assume that, on average, shooting and save percentages are constant over time. This assumption is problematic because in reality, we’d expect the game to look different in the first few seconds following a faceoff – because the winning team generally needs to take time to get organized (or take the puck to the other end of the ice) we shouldn’t expect the offensive opportunities in the first few seconds following a faceoff to be as plentiful, or as dangerous. If teams aren’t shooting in the seconds immediately following a faceoff, or they’re taking rushed, low-percentage shots instead of waiting for better chances, using Corsi as a measuring stick doesn’t make sense.

We can simplify this picture by focusing solely on plain old goal differential. By looking at goals we’re able to capture both the change in shooting and save percentages, as well as in shot generation rates. Because we’re going to be looking at multiple years’ (from 2009-10 to 2014-15) worth of data across all the teams in the league, we don’t need to worry about the sample size issues that usually push us towards using Corsi.

We’ll start by looking at the expected goal differential, by second, for a faceoff in a given zone. We’ll also break it down by which team won the faceoff, as we’d obviously expect the goal differential to be much different following a defensive zone loss than a defensive zone win. The results for the 3 winning faceoff situations are given in the graph below (the results for the losing team are simply the winning team’s results multiplied by -1).


The first observation worth noting is how small the effect is – even after winning an offensive zone draw and continuing play for 45 seconds, the net effect on goal differential is still less than 0.02 goals. This is in line with what others have previously found, although it’s slightly lower because we don’t assume a constant shooting percentage across the whole interval. It also emphasizes how many additional faceoffs a player would need to win to have a signficant effect on a team’s results – as it turns out, faceoffs aren’t as big a deal as we might have guessed.

The other key thing to note here is that the amount of time before we see the full effect of a zone start (i.e. where the graph begins to level off) varies by where the faceoff was won. A player who is on the ice for an OZ faceoff win won’t see the full effect of that win for nearly 60 seconds, while for NZ and DZ wins the effect stabalizes in much less time (approximately 14 and 8 seconds, respectively). This is kind of a big deal, because it means that a player who leaves the ice 15 seconds after an OZ win is really only experiencing about 75% of the “benefit” that a player who played a full 45 second shift follwing an OZ win would experience.

To illustrate this more clearly, look at the difference between the following 2 situations:

  • A player loses a defensive zone faceoff and plays continuously for 30 seconds before leaving the ice on an on-the-fly change.
  • A player loses a defensive zone faceoff, the other team controls the puck for 3 seconds, takes a shot on goal and it’s frozen. The player stays out for the faceoff and loses it, and then plays for 27 seconds before leaving the ice.

It should be obvious from the graph above that these 2 cases don’t represent the same level of danger – in the 2nd situation, where the play lasted only 3 seconds before another faceoff occurred, the risk the player faced was less, since the shot was rushed and unlikely to be the result of his opponents setting up a high quality chance. The faceoff provides a reset, and the seconds immediately following the faceoff are less dangerous than those that occur after a team has had sufficient time to formulate an attack.

All of which is to say that knowing both the result of the faceoff and the amount of time a player played following a faceoff is really important if we want to accurately quantify a player’s usage. A 60% center who leaves the ice immediately following each defensive zone win should have a much smaller zone start adjustment than a 40% center who plays a full 40 second shift, regardless of the faceoff result.

So if we know that we need to change the way we’re measuring and adjusting for zone starts, what’s the best way to do it, considering all that we’ve learned so far? The simplest method is to use a player’s on-ice faceoff results, his shift times, and the graph above to find the aggregate effect on goal differential we’d expect to see for an average player who played in the exact same situation to see.

To create this individual level adjustment, we’ll perform the following steps:

  • Find each 5v5 faceoff for which a player was on the ice
  • Classify that faceoff by zone (Off. Zone, Neu. Zone or Def. Zone) and result (win or loss)
  • Calculate the number of seconds until either:
    1. Another faceoff occurred with the player on the ice; or
    2. The player left the ice.
  • Using 1), 2) and 3), lookup the expected goal differential in the graph above. For example, if a player was on the ice for a neutral zone win and left the ice 12 seconds following it, we’d credit him with +0.0016 goals, while if he was on for a defensive zone loss and a 20 second shift afterwards, that shift would be worth -0.0153 goals to him.
  • Sum the expected goal differentials for each faceoff to come up with an Expected Faceoff Goal Differential, or xFOGD

Expected Faceoff Goal Differential (xFOGD)

xFOGD represents our best estimate, using six years of historical data, of how much a player’s goal differential was impacted by his zone starts and on-ice faceoff results. The table below shows the top and bottom 10 players in total expected faceoff goal differential (xFOGD) from 2009-2010 to 2014-2015.

Season Player xFOGD Season Player xFOGD
20112012 MALHOTRA, MANNY -3.08 20102011 EHRHOFF, CHRISTIAN 2.18
20132014 GORDON, BOYD -2.98 20132014 SHARP, PATRICK 2.20
20142015 GAUSTAD, PAUL -2.84 20132014 TOEWS, JONATHAN 2.28
20132014 KRUGER, MARCUS -2.35 20102011 BURROWS, ALEX 2.45
20142015 GORDON, BOYD -2.26 20142015 TAVARES, JOHN 2.49
20132014 MCCLEMENT, JAY -2.22 20112012 BURROWS, ALEX 2.88
20132014 BOLLIG, BRANDON -2.19 20112012 SEDIN, DANIEL 2.94
20142015 RISTOLAINEN, RASMUS -2.14 20102011 SEDIN, HENRIK 2.99
20142015 KRUGER, MARCUS -2.07 20102011 SEDIN, DANIEL 3.17
20112012 BREWER, ERIC -2.06 20112012 SEDIN, HENRIK 3.37

Here we can see that for certain players, the effects can be rather large, with the Sedins seeing benefits worth about 3 additional goals in both 2010-11 and 2011-12, while their teammate Manny Malholtra’s stats decreased by roughly the same amount in the latter season. While it may seem odd to express our usage stat in goals, the easiest way to think of it is that xFOGD represents the amount we’d need to adjust a player’s plus-minus by, if plus-minus were a perfect measure of player performance.

While the raw stats above are useful in adjusting aggregate performance, it’s also helpful to look at rate statistics (i.e. xFOGD/60) to get a sense as to which players faced the toughest usage on a per minute basis. This metric should align more closely with our traditional zone start numbers, but note that it will still be heavily influenced by a player’s on-ice faceoff results.

Season Player xFOGD/60 Season Player xFOGD/60
20112012 MALHOTRA, MANNY -0.25 20132014 TOEWS, JONATHAN 0.12
20142015 GAUSTAD, PAUL -0.24 20142015 KANE, PATRICK 0.13
20132014 GORDON, BOYD -0.22 20142015 RICHARDS, BRAD 0.13
20142015 GORDON, BOYD -0.20 20112012 TOEWS, JONATHAN 0.13
20142015 NYSTROM, ERIC -0.18 20102011 SEDIN, HENRIK 0.15
20122013 MCCLEMENT, JAY -0.18 20102011 BURROWS, ALEX 0.16
20132014 BOLLIG, BRANDON -0.16 20112012 BURROWS, ALEX 0.16
20132014 KRUGER, MARCUS -0.16 20102011 SEDIN, DANIEL 0.17
20142015 HENDRICKS, MATT -0.16 20112012 SEDIN, DANIEL 0.17
20132014 MCCLEMENT, JAY -0.15 20112012 SEDIN, HENRIK 0.18

Both lists are pretty consistent, which shouldn’t be surprising given that we’re looking at the extreme ends of the spectrum, but it also shows that players do shift around, and that those with highly specialized roles are still identified by our new metric. Full results, including both seasonal xFOGD and xFOGD/60 numbers can be found here.

The most interesting thing to look at is which players’ “usage” change the most when we compare metrics –  that is, players whose usage differs significantly when measured by xFOGD/60 versus a more traditional metric such as OZFO%. Since the numbers won’t be on the same scale, we’ll normalize each stat by calculating each player’s Z-Score for both xFOGD/60 and OZFO%, and then find the absolute difference in Z-Scores between the two metrics. The players with the 10 highest absolute deltas are presented in the table below.

Season Player z-xFOGD/60 z-OZFO% Z Score Delta
20092010 STECKEL, DAVID 0.75 -1.23 1.99
20122013 MARCHAND, BRAD 2.22 0.25 1.97
20102011 HALL, ADAM -0.11 -2.07 1.97
20102011 TYRELL, DANA 0.29 -1.66 1.95
20142015 GRANLUND, MARKUS -1.35 0.59 1.94
20092010 NICHOL, SCOTT 0.69 -1.24 1.92
20122013 SAAD, BRANDON 2.00 0.08 1.91
20102011 TORRES, RAFFI -1.28 -3.12 1.84
20122013 TOEWS, JONATHAN 1.87 0.06 1.81
20122013 VANEK, THOMAS -1.62 0.17 1.80

Here we can see that there are some pretty significant differences in how the two metrics evaluate a given player’s usage. 5 of the 10 players see the sign of their usage change (i.e. they went from having favourable to unfavourable usage, or vice versa), and most of the changes have players going from relatively average usage numbers to much more extreme numbers. The list is mostly populated with players who posted exceptionally good faceoff numbers, which is what we’d expect – their own abilities on the draw tend to remove a lot of the risk they’d seen from their raw deployment numbers.

While some might argue against making such large adjustments to a center’s usage based on their own faceoff results, it’s actually one of the main benefits of using an approach like this. By seperating out the faceoff result from the zone in which it takes place, we can isolate the value of the win from the value (or cost) of the deployment. And because 80% of players on the ice aren’t the ones taking the draw, it’s also a better measure of how they’re specifically deployed – after all, if you’re a defenceman put out for 10 faceoffs a night with Patrice Bergeron you’ll likely see significantly better results than if your most common centerman is Kevin Hayes. Classical zone start measures don’t capture that, but xFOGD does.

Furthermore, the approach presented here is intuitively more accurate – we’re using a more granular view of what actually happened to each player when they were on the ice, so our adjustments should be more reflective of the actual risk or opportunity each player saw from their deployment. Although for most players the overall effect is going to be small, the examples above clearly show that for some players it’s quite significant. It’s those players that justify the additional effort requried to calculate xFOGD, as our assessment of a player’s deployment can go from “extremely tough” to “easier than average” simply by digging a bit deeper into their shifts.


5 thoughts on “A New Way To Measure Deployment – Expected Faceoff Goal Differential

  1. This is helpful, thanks for researching. This is more helpful than OZS/DZS%, although it seems like measuring deployment as a function of time can be problematic. Players can help/hurt themselves by maintaining offensive puck control in the OZ/failing regain puck control in the DZ, correct? ‘Zone-effects’ seems more accurate than ‘Deployment.’

    Maybe you brought this up and I glossed over it, but I’ve watched we too many marathon DZ Brandon Bollig shifts not to mention it.

    • It’s meant more as a baseline upon which to compare a players results to/adjust their observed outputs by. If a player gets stuck in their DZ after a faceoff for an unreasonable amount of time, their results will reflect that. We’ll give them partial credit back because of the fact that they started in the D-Zone, but it won’t be enough to make up for their own poor play.

  2. I think there is an inherent problem in using this methodology in that the talent level and the assigned roles of the players starting in the defensive zone is different than those starting in the offensive zone. Players starting in the defensive zone are typically lesser skilled players whose role is biased towards not allowing the opposing team score a goal over scoring a goal themselves. Players starting in the offensive zone are highly skilled players capable of generating a lot of offense and goals.

    In your running goal differential chart there is a reason why the goal differential after an OZ Win doesn’t stabilize for 45 seconds. It’s because those players are inherently better of players. It is the Henrik Sedin and Patrick Kane types vs the Paul Gaustad and Boyd Gordon types. I’d expect Sedin/Kane to out score Gaustad/Gordon. The initial spike after an OZ Win ends at about 8 seconds which is where DZ Win stabilizes. Everything after that is probably due to talent differential, not the zone start. Even some of the stuff in the first 8 seconds is due to better player talent and not strictly the face off win. Similarly, the goal differential after a DZ win is likely suppressed because of the different talent levels.

    The end result is I think you end up over estimating your adjustments a little because you end up including a talent differential component in your adjustments, which ideally we wouldn’t.

    In 2011-12 Henrik Sedin was on the ice for 62 5v5 goals for and 37 5v5 goals against for a 5v5 goal differential of +25. Is your opinion/evaluation of Henrik Sedin changed any if his goal differential was +22? Incidently, when I remove the first 10 seconds after a zone face off H. Sedin would have a +23 goal differential (57GF vs 34GA).

    The most extreme players might see a goal differential adjustment of 2 or maybe 3. For me that is hardly enough to be concerned about – there are far greater uncertainties in hockey analytics that I’d be far more concerned about.

    • I don’t disagree that there is a potential that we’re overestimating the effects based on role players, but I think the key thing to remember is that the players with extreme zone starts are relatively few and far between. Most faceoffs are going to be contested by two sets of players without very defined roles (i.e. players with relatively neutral zone starts), so the overall effect of those “pure offense” or “pure defense” players is going to be lessened. You can claim that this method is overestimating the effect of zone starts or you can claim that zone starts have little effect, but I don’t think that you can claim both are true because to me the ideas that underlie each argument are at odds with each other.

      The other point is that 2 or 3 goals is a relatively large adjustment – if you credit Sedin with the full +25 (i.e. if you assume +/- is a perfect measure of contribution) then that’s roughly half a win difference in our evaluation. It’s going to be a lot smaller for most players, but it is important to note for the players where it is a big difference.

      • I’ll grant you that 2 or 3 goals is not completely insignificant but it is not any more significant than the uncertainty due to variability in shooting or save percentages which affects every player. The impact of QoT would be far more significant and team playing style is probably a fair bit more important too. For 95% of players it isn’t even worth discussing any ZS adjustment and only a handful of players per year might see an adjustment of 2+ goals. If you are worried about a couple of goals for a handful of players, fine, but for me, zone start adjustments are a long way down on the list of important factors to consider in hockey analytics.

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