I wanted to write this piece 2 months ago before the season ever started. I sat down, fingers clacking away on my laptop, and began writing. The creative process was running on all cylinders until I went to get some data for the Devils 2017-18 roster and realized… 5 key players expected to contribute to the Devils this year had no data on them. Making assumptions on how guys like Nico Hischier, Jesper Bratt, Will Butcher, Mirco Mueller, and Michael McLeod would measure up statistically would be a frivolous act of data journalism on my part. So, here I am 12 games later (15% of the way through the season) to commit a marginally less egregious act: using small sample sizes to make bold claims! Let’s begin:
To begin, this piece was inspired by a few authors I respect greatly and whose work is among the best in the biz. Prishanth Iyer wrote a piece for The Athletic optimizing lines for the Detroit Red Wings during the offseason. More writers for the Athletic, Murat Ates and Dom Luszczyszyn, wrote articles for optimizing the lineups for the Winnepeg Jets and Toronto Maple Leafs. So, to bring similar content to our favorite small market hockey team from the Garden State, I have performed an exercise to optimize the roster for the New Jersey Devils.
Step One: Clustering Players into Style Types
The articles mentioned above use data from Ryan Stimson’s Passing Project that have players grouped into player type “bins”. There were four major forward groupings, in order of xGF%: Playmakers, Shooters, Balanced, and Dependents. For Defensemen, there were four types as well, in order of xGF%: All-Around, Volume Shooters, Puck-Movers, and Defensive-Oriented.
The problem is that Stimson’s data is not yet available for many of New Jersey’s young roster, and understandably so. Luckily, I performed an analysis similar to Stimson’s clustering exercise last year to categorize forwards and defensemen and can therefore reproduce similar results to use for this article. I performed the cluster analysis (k means for you statheads) on the top 390 forwards (30 teams x 13 forwards) and 210 defensemen (30 teams x 7 defensemen) across six statistics (CF%, GF%, xGF%, G60, A60, P60) to categorize each player in the league. Although these almost definitely are not the same exact inputs or results as was in Stimson’s work, I used the same naming convention based on the xGF% and G60/A60.
HUGE disclaimer: the sample size varies by player. I pulled data from the last 3 seasons for those who have that many games played. Others, like the rookies, sophomores, and fringe players have less. At the time of creation, the Devils have played 12 games, so some players may only have a maximum of 12 NHL games’ worth of data available. I can always reperform this analysis towards the end of the year when we have more data available. The results are as follows:
Marcus Johansson, Taylor Hall, Nico Hischier, and Jesper Bratt provide the Devils with 4 Playmakers; Kyle Palmieri, Adam Henrique, and Brian Gibbons round out the Shooters. Miraculously, this Devils roster only has one Dependent player on the roster in Coleman, and the rest are Balanced.
On the blueline, Will Butcher is the only All-Around dman, while Severson gets labeled a puck mover and the rest are Defensive-oriented. Again, take this with a heaping grain of salt.
So now that we’ve taken inventory of the Devils lineup, who should play with whom?
Step Two: Optimizing Lines and Pairings
-----------------------------------------------------------Forwards--------------------------------------------------------------According to Stimson and Iyer, the breakdown of combinations and expected goals is as follows:
Our objective is to maximize the team’s collective xGF%. To do that, we will weigh each line’s ice time and produce the weighted average xGF%. Here’s what works out to be NJ’s best player type roster:
So now the fun part: slotting in some players. You could arrange these in a few different combinations, but below is one:
Now, I’m not necessarily advocating for this lineup for a few reasons, but this has been a fun exercise nonetheless. Travis Zajac, the centerman rumored to return from IR soon, falls into the Balanced category and therefore doesn’t alter the weighted xGF% of the roster. This shows an inherent weakness in this type of analysis. The team will almost objectively be better come his return, but not according to this exercise. Let this be yet another reason why analytics are just another perspective to be added to a diverse pool of information.
Like we did for the forwards, we will take a look at how to optimize the defense pairings. In the chart below we see the xGF% for each player type duo.
Like we knew heading into the season, the Devils defense corps is far from their strength. Of the 8 defensemen on the roster, 6 have the least flattering classification of Defensive Oriented. As a result the most optimal build is to stack the top pair and leave Cory to fend for himself afterward. If you’ve watched any games in the last two seasons, you might think that the team has already taken this article to heart.
This eliminates a few of the choices and produces the following possible lineup:
Note that Dalton Prout and Mirco Mueller do not make the cut. If we were filling the lineup purely in order of career xGF%, Santini and Mueller would sit in favor of Prout and Lovejoy, but I’d like to maintain the modest reader base I’ve established and therefore won’t publish such lunacy.
Step Three: Gap Analysis
So now that we know what the optimal lines would look like, how do they compare to what Coach Hynes has been deploying this season? Let’s take a look at the most frequent forward and defenseman groupings so far:
The chart above shows forward lines in descending order of 5v5 ice time.The most utilized four lines without repeating players are the first, second, fourth, and last groupings in the chart (outlined in red). Their weighted projected xGF% is 53.5%. Compared to our optimal calculation of 55.6%, one could argue that the coaching staff is leaving an extra 2.13 xGF% percentage points on the table when icing a roster. Ultimately, not awful, but this exercise demonstrates that there may be room for improvement.
The chart above shows defensive pairings in descending order by 5v5 ice time. The most utilized three pairings without repeating players are the first, second, and fifth on the list (outlined in red). Their weighted projected xGF% comes out to be 48.5%. Compared to the optimal calculation above of 49.4%, John Hynes and co. are icing a defense corps that is just .9 percentage points off of optimality.
Ultimately the coaching staff isn’t far off of what this exercise claims is the best combinations of the roster players. For this reason and many others I hope to write about in the near future, I am thankful for a coach like Hynes and I felt the same way last season despite the struggles. I believe in the system he has put in place and believe the fanbase will continue to be delighted as more quality inputs are set in place by Ray Shero. That said, here’s to a defenseman or two.
Cheers, thanks for reading. As always, let me know what you think, good or bad.
Gotta give credit where credit is due/not piss people off:
Ates, Murat. “What Is the Winnipeg Jets' Most Optimal Forward Lineup?” The Athletic, The Athletic, 3 Nov. 2017, theathletic.com/145388/2017/11/03/what-is-the-winnipeg-jets-most-optimal-forward-lineup/.
Iyer, Prashanth. “Iyer: Optimizing the Ideal Lineup for the Red Wings This...” The Athletic, The Athletic, 17 July 2017, theathletic.com/76383/2017/07/17/iyer-optimizing-the-ideal-lineup-for-the-red-wings-this-season/.
Luszczyszyn, Dom. “By the Numbers: What Are the Ideal Forward Lines for the...” The Athletic, The Athletic, 4 Aug. 2017, theathletic.com/81552/2017/08/04/by-the-numbers-what-are-the-ideal-forward-lines-for-the-maple-leafs-this-season/.
*All data from corsica.hockey.com