Pittsburgh analytics conference recap

Arik Parnass attended the Pittsburgh Analytics Conference. He is a contributor to HP and writes for Eyes on the Prize.

The death of hockey’s public stats movement was greatly exaggerated.

That was the overall takeaway from Saturday’s WAR-On-Ice Hockey Analytics Conference in beautiful — if confusingly laid out — Pittsburgh, Pennsylvania. There had been a worry, following the poaching of those names you’ve now heard so many times — Eric Tulsky, Tyler Dellow, Sunny Mehta, Vic Ferrari — that hockey’s suddenly booming analytics movement would stall, that the brain power now had no public review, and that the public now had none of its leading innovators.

One of the early presentations at the conference was made by James Santelli, a Pittsburgh-based award-winning Sabermetrics writer, on the shift in baseball and the importance of communication when it comes to convincing players and managers of analytic truths. One has to look no further than baseball — Santelli being a prime example — to find that while the initial innovators may move on, others will always be there to replace them. Just as there are always a new crop of promising hockey players when the NHL Draft rolls around in June, there will now always be another crop of analysts. Hopefully, building off of Andrew Thomas and Sam Ventura’s work to put this event together, there will now also always be forums for new ideas to be shared and reviewed.

Here were some of the highlights of the conference, which can be found in its entirety here:


Sam Ventura, one of the co-founders of WAR-On-Ice and a PHD student at Carnegie Mellon, led off the conference with some data from a pre-conference survey of the attendees. He shared some of the best reasons given for attending the conference:

“My fantasy team needs some help.”

Don’t we all. This was the part of the conference in which I really wished I had my laptop. Felt bad for fake Zach Parise forced to chug it out in the starting lineup for “Tyutin in the Staal” even with a concussion.

“I like being right more than I like being wrong.”

Respect for the honesty. Being right is also a hobby of mine.

“I have been manipulating numbers to tell stories stories for years, way before it was cool to do it.”

We know, Sean. We know.

“I take a class taught by the organizers, and I would like an A.”

Nobody likes a suck-up. Especially one who gets to have multiple-regression explained to him through references to FenClose and even-strength shooting percentage.

“Gots to get stat-wise so I can defend myself a little better in HFBoards threads.”

This is an entirely valid reason 

James Santelli on Analytic Implementation and the Shift

If you’ve read much of my stuff, you know I love poaching analytic advances from other sports and applying them to hockey. As Santelli put it, “baseball is easy.” It’s a stop-and-start sequential game with discrete outcomes and a small number of moving parts. Hockey is about as far from that as you can get. That said, there are still a large number of lessons hockey can learn from baseball analytics, and there are ways to adapt the shift to a hockey setting.

The most basic hockey equivalent of the shift builds on the principles of zone entries that have been uncovered over the past few years. We know that on average, carry-ins lead to greater offense than dump-ins. We have learned which players carry the puck in more, what we need now is location data. Once player and puck tracking is implemented for teams, learning which players like to carry the puck in where can lead to defensive shifting to best defend the blue line.

Let’s say, for example, that we know that Phil Kessel is responsible for 50% of his team’s carry-ins when he’s on the ice, and that 80% of those are on the right side. That means that in order to deny him entry into the zone, a team could theoretically stack their left side, and force either a dump in or a tough pass, something out of Kessel’s comfort zone. It could resemble something like this (presuming a team was defending with a left-wing lock, the C could be an L, or the C and RD could switch spots). Kessel is the green “R” with the puck, and the defense shifts to prevent his carry-in.

This is the future of hockey analytics, and just because baseball is such a different sport doesn’t mean it doesn’t have lessons to offer.

arik 1

Matt Cane on Using Shot Location for Additional Information

Cane, of PuckPlusPlus.com, gave a very insightful presentation on further utilizing the NHL’s shot location x,y data to make conclusions about players, handedness, and defensive responsibility. He created a metric called Player Side Bias, which divides a player’s shots from one side of the ice by their total shots (and then subtracts 50%) to determine where they prefer to shoot from.  Using this, one can make assumptions about which players are playing LW and which RW.

He used this data to come to a few different conclusions. First, making the assumption that defenders who generally play more on the left are left-handed, he found that same-handed defense pairings tend to post lower shots for percentages than opposite-handed pairings. My concern with this conclusion would be that coaches try to put lefties with righties when possible. For teams that have an imbalance of handedness, it is most likely that the third pairing features two lefties or two righties, so that the top pairings can be to the coach’s preference.

In St. Louis, for example, defensemen play on their proper side with the exception of the third pairing, where both Ian Cole and Barret Jackman are lefties. The result of this, therefore, is that same-handed pairings would be expected to have lower shot-for percentages simply by virtue of them being inferior pairings. It doesn’t rule out the possibility that opposite-handed pairings are preferable — in fact I’m sure that’s true, at least marginally — but it does shed some doubt on Cain’s first finding.

arik 2

The more significant use of Cain’s side bias was in determining a player’s “% Shots Against from Side” metric, which combines the NHL’s x, y data with side bias to determine which player was responsible for more shots coming from their side while on ice. For Pittsburgh last year, for example, Paul Martin had a team-low 46.8% of opposing shots come from his side, which Brooks Orpik had a team-wrost 56.8%. I would be curious to see these numbers in the 10 seconds following zone entries, to account for the fact that once the opposition sets up, it is the wingers that are responsible for preventing shots from the point. Using only shots from 20 feet or closer might also be interesting, using Greg Sinclair’s Super Shot Search, to include only shots for which the defenders are most responsible.

Cain’s other finding was that players shoot for a higher percentage on their off-sides (ie righties from the left and lefties from the right). This is intuitive, considering that not-only do one-timers go for a higher percentage, but shooting from the inside also opens up a far better angle on a goalie. Shots from closer to the middle are always going to be more valuable, and shooting from one’s off-side inherently means shooting from closer to the middle. The next thing to account for is shot rate (at even strength most importantly) from each side, and obviously defensive impact. It is my opinion that as time goes on, we will see a shift towards players playing more on their off-sides, not just on the power play — where it’s obvious — but at even-strength as well. We’ll see if that comes to fruition.

Sam Ventura on Zone Transition Times

Ventura introduced a pair of new metrics using the league’s RTSS data that he hopes will paint a picture of which teams are fastest and slowest at transitioning from defense to offense and at maintaining offensive zone possession. ZTTOD and ZTTDO show the average time between offensive and defensive events (faceoffs, shot attempts, ends of periods, hits, etc.) as marked by the NHL. Ventura presented only 2014-15 data at the conference, which led to a classic #HockeyTwitter sample size freak-out, but my biggest concern with the metric was whether it would be severely skewed by high attempt teams like Ottawa (for which both metrics would appear on the quick side because of the high volume of events) and low attempt teams like New Jersey (for which both would appear slow). For the 2014-15 data, my skepticism appeared accurate, but the 2013-14 data (later released on Twitter) painted a different picture. It’s still a concern of mine, but the idea is interesting. This could be a metric that gains more accuracy and relevant with the introduction of tracking technology.

Benjamin Zhang on Elo Ratings in Hockey

Zhang, a student of Thomas — and the suck-up from earlier? We’ll never know — gave an impressive presentation on adapting chess’s Elo and Glicko ratings into a model to rate NHL teams based on contextual factors like home ice and fatigue, awarding points for wins weighted based on whether or not a team was favored. An interesting takeaway was that the 2008-09 San Jose Sharks — a team that won the President’s Trophy before losing in the first round of the playoffs to the Anaheim Ducks — tailed off significantly in the final quarter of the season by this metric (which Zhang calls “Expected Shot Probability Rating”).


I would be curious to see whether there was any type of a pattern there that could explain the Sharks’ perpetual playoff losing. Maybe it hasn’t simply been bad luck after all.

Kody Van Rentergen on Strategic Adjustments Through Video Analysis

For a systems fiend like myself, this presentation was very informative. Van Rentergen is a coach for the Robert Morris University hockey team. Following a disappointing defensive season, the RMU coaching staff took a look at every goal scored and allowed, and using fairly basic metrics, came to some conclusions about the team’s system.

Van Rentergen found that most of the team’s goals allowed were the result of players getting beat down low in the team’s heavy double-down system, or swarm. As a result, RMU changed to a more passive layered system similar to the one used by the Boston Bruins at the NHL level, and after some early adjustment struggles, improved rapidly down low. This change also allowed the forwards to pressure the points in the defensive zone with more rigor and to block more shots, which led to a huge improvement in goals allowed on point shots.

On offense, the team decided it needed to play to its strength — speed — more deliberately, and had its forwards retreat farther into the neutral zone to receive passes with pace, as well as having the weak-side defensemen remain available for outlet passes behind the puck-mover so that dump-ins could be made more of a rarity. Through these changes, RMU managed to survive the loss of a talented, professional-caliber senior goalie, and make the NCAA tournament last season.

The Future

The conference was a great success, and the hockey analytics community is in fact alive and well. It was great to finally meet noted Twitter personalities Jen Lute Costella and Stephen Burtch, who both gave insightful presentations on their recent work, as well as the organizers themselves, Thomas and Ventura, who did a great job putting it all together, and of course a number of other intelligent fans, writers, and fantasy hockey GMs.

It was the first of its kind in the United States, but certainly not the last. Hopefully, in 10 years, we’ll look back and laugh at the 120 or so person turnout, the awkward sometimes-malfunctioning microphone and camera, and the excess cheese-stuffed pizza, as we dine on steak and caviar in some fancy convention center or NHL arena, and laugh at the Toronto Maple Leafs’ 57 year cup drought. After all, as Ventura noted early on Saturday, “if you want to get any attention with fancy stats, you should always include Toronto.”

2 thoughts on “Pittsburgh analytics conference recap

  1. Pingback: Shot Location Data and Strategy I: Off-Hand Defencemen | puck++

  2. Pingback: Shot Location Data and Strategy II: Evaluating Individual Defensive Play | puck++

Leave a Reply

Your email address will not be published. Required fields are marked *