The New York Islanders’ 2021-2022 season has been a mess. Dud hockey. But let’s not focus on that disappointment today. Go read one of the hundreds of articles, blog posts or angry tweets that have already been written on the subject if you want to relive that discussion. No, today, an analytics nerd – yours truly – is going to discuss Kyle Palmieri and his abnormal 2021-2022 season through the lens of expected goals.

If you aren’t familiar with expected goals, don’t worry, I’m going to be giving an overview of the concept before we apply it to Palmieri. That’s right. We are going to be talking about nerd stuff! Please don’t close this tab! Give me a chance – not every analytics writer is a condescending know-it-all who cherry picks bad takes and dunks on them with quote tweets.

If you’re still here, welcome to Analytics Toolbox. Let’s dive in…

Expected Goals

Expected goals is my favorite, and – I believe – one of the most useful analytics concepts. The concept has gained popularity in hockey circles in recent years. For those who haven’t heard of it or have, but brushed it off because “the dang nerds are ruining hockey,” or just want a refresher on it, you have come to the right place.

(If you are comfortable with the concept of expected goals, feel free to skip this section).

Expected goals (xG) is a model-based metric used to isolate the evaluation of play-driving and chance-creation/suppression ability from things a player cannot control such as bounces, quality of goaltender, etc. How does this work? The models make use of the public data tracked by the NHL. The NHL tracks every unblocked shot attempt and collects over 100 pieces of information per unblocked shot attempt (shooter, location on ice, type of shot, etc.). Analytics nerds leverage this historical data to train the data-driven expected goals models.

As more NHL games are played and new shot attempt data is collected, the expected goals models are used to calculate the “probability of goal” value from that new data. For example, a one-timer from the slot might have an expected goals value of 0.25, because that type of high-danger play results in a goal 25% of the time. Common jargon is, “that scoring chance was worth 0.25 expected goals.”

For the visual learners, here’s a figure! The left image describes how an expected goals model that only considers shot location would work, and the right image shows the model features (e.g. input variables) of Evolving Hockey’s expected goals model – one of the best publicly available models.

The length of the blue lines, or “gain,” represent the magnitude of the impact that each feature has on the model. We can see that “shot distance” (how far away from the net the shooter is), “seconds since last” (how long it has been since the last shot was taken – think rebounds) and “shot angle” (how centralized the shooter is) are the most impactful features. I think we would all agree that those three features being the most impactful makes a ton of sense.

Before we move on to evaluating Palmieri’s 2021-2022 season through the lens of expected goals, I want to address one aspect of the model that people tend to get hung up on. I find this to be especially true when someone is hearing/learning about expected goals for the first time. This aspect is the fractional/decimal nature of expected goals.

Going back to the figure, we calculated “Freddy” had 0.66 expected goals over the course of the game. This sounds a bit silly the first time you hear it. How can a player have 0.66 goals? Are you saying we expect “Freddy” to have scored 0.66 goals? You can’t score 0.66 goals!

Imagine in the first period of “Freddy’s” next game he had a shot attempt from the low slot and from the side wall – so over the course of his last four periods of hockey he has 0.66 + 0.04 + 0.3 = 1 expected goal! We can interpret this as “based on the scoring chances Freddy accumulated over the past four periods he should have scored one goal.” By “should,” I mean that him having scored one goal would be the most likely scenario assuming he has league-average shooting talent and is shooting against league-average goaltenders.

If “Freddy” scored more than one goal in this timeframe, we could infer he is either a talented shooter or just getting lucky (to determine if it is luck or talent-driven you should use a multiple-season sample size). If “Freddy” had less than one goal (so in this example, zero goals) in this timeframe, we could say the opposite; he is untalented and/or unlucky.

Alright, I think we’re all on the same page now. This analytics stuff is cool, right? Even if you’re not convinced yet, stick around a little longer – we are finally going to talk about how this applies to Palmieri and his 2021-2022 season with the Islanders!

Palmieri and the Islanders

Talking about the Islanders is more painful this season than it has been in recent history. Between you and me, I did not enjoy this part either. I ended up giving myself an ultimatum: no food until I finished the article. The good news is, (or bad news, if you haven’t enjoyed this piece) I love pizza more then I hate writing about the Islanders, so here we are.

Islanders fans — even hockey fans in general — thought this season would play out very differently. The Isles have been to back-to-back final fours! A fall this drastic was completely unexpected. Even most analytics nerds – a group that has been low on the Islanders in the past – had them as a top team in the Metropolitan Division in their preseason predictions.

One specific player the fans, media and management expected more from this season is Palmieri. Palmieri came over from the Devils at the last years NHL Trade Deadline and was a useful piece for the Islanders’ playoff run. He picked up seven goals and two assists with a plus-5 rating in 19 games. Lou Lamoriello must have seen something he liked because he inked Palmieri to a four-year, $20 million extension. Wait, Lamoriello committed a ton of cap-space to a player who will be in his mid-30s by the end of the deal? That’s so unlike him. Who could have seen this coming?

This contract looked like an absolute disaster for the first-half of this season. Palmieri’s statistics through the Islanders’ first 41 games? One goal and six assists. Ouch. Granted, he was injured for a portion of these games and only actually appeared in 29 of them. Still, you would expect more from a player with his pedigree.

Throughout this rough start, Palmieri’s expected goals were telling a different story. At the halfway mark of the season, Palmieri held the fourth highest ixG/60 (individual expected goals per 60 minutes of ice time) on the Islanders with 0.95 – a number that put him among the league’s elite goal scorers. His actual goals per 60 minutes was just 0.13 – a number that put him in the same company as stay-at-home defensemen.

Islanders Expected Goals Leaders Through 41 Team Games Played (Courtesy of Natural Stat Trick):

Maybe this wasn’t a case of a bad contract and a washed-up player. Maybe Palmieri was just getting unlucky. When a player underscores their expected goals over a medium-sized sample size (e.g. about 20-30 games) it can usually be chalked up to bad luck. This isn’t true in all cases, of course. In fact, there are well-known players who consistently underscore their expected goals. Brady Tkachuk of the Ottawa Senators and Brendan Gallagher of the Montreal Canadiens are two examples. Both these players tend to jam the puck in tight into the goalies’ pads – a play that expected goals models typically overrate even though it is not a high percentage scoring chance.

Palmieri is not that type of player. He is a sniper, and this is evident when comparing his historical numbers for actual and expected goals. Over the last three seasons, Palmieri has outscored his expected goals by 13% on average. This overscoring of expected goals is not uncommon among the league’s elite shooters. Because certain players consistently over/underscore expected goals contextualizing a player’s actual goals to expected goals ratio using historical data is an essential step when using expected goals models for analysis.

I run the Instagram page called @DataDrivenHockey, and back in early February we made a series of posts where we predicted one player from every NHL team to heat up in the second half of the season, and one to cool down. Palmieri was the clear heat-up choice for the Islanders.

View this post on Instagram

A post shared by datadrivenhockey • NHL hockey data/modeling (@datadrivenhockey)

Since his abysmal first half of the season Palmieri has picked up 10 goals and added six assists in 23 games played. Those aren’t world-beating numbers by any means, but it’s a 36-goal, 57-point pace over a full 82 games. Certainly the numbers of a player worth his $5 million cap hit. Other writers might have gloated about nailing this prediction, but fortunately I am much too humble and good-looking to boast. Look, another picture!

This graph compares Palmieri’s first half points, on-ice goals for, and on-ice goals against to the same statistics in the second half of the season (what do I mean by on-ice? That is a discussion for the next article, so stay tuned for that!) Everything has significantly improved for Palmieri in his second half – goals for and points increased, and goals against decreased.

I have one final picture I want to share with you — the timeline of Palmieri’s goalscoring this season. Wow. Just wow.

Palmieri has been remarkably consistent at generating scoring chances this season. His cumulative expected goals have climbed at a steady rate all year. His goals? Not so much. But regression comes for even the farthest of outliers, and Palmieri’s second half scoring is an excellent example of this. Palmieri’s rebound proves that even at the age of 31 he is still an above-average goal scorer in the NHL, despite the hard luck he faced early this year.

Final Thoughts

We evaluated Palmieri’s season by utilizing a single metric: expected goals. Generally, this is not recommended – formulating an opinion based off a single input is not good practice. Analysis is only worthwhile if you are building up a picture of the player using multiple inputs – multiple statistics-based inputs and watching the games. Notice I said and, not or. We focused on expected goals because this article was intended to introduce that concept but, moving forward, you will need more analytics in your toolbox.

Good news – this is a series! Come back to Analytics Toolbox soon for a discussion on the definition and utility of “0n-ice” metrics.

John Smark