The Importance of Expected Goals

[text_output]One of the most common complaints lodged by fans against shot-related stats such as Corsi (shot attempts) and Fenwick (unblocked shot attempts) is that “not all shot attempts are equal.” While most hockey statisticians would point out that this sort of pushback is missing the point of why shot-based metrics are important, the point is still valid. Even the most novice hockey fan can tell you that a shot from the doorstep of the crease off of a rebound is clearly a higher quality shot than a weak wrister from the point with no traffic in front; so clearly not all shot attempts are created equal. This is where the stat “expected goals” (xG) comes into play, and why it is crucial for evaluating a player, game or team.

In layman’s terms, expected goals are shot attempts that have been weighted for shot quality. Before we dig into exactly what expected goals is, it should be noted that multiple models for expected goals exist and they have some variations in their calculations. The two most commonly referenced models are Dawson Sprigings (better known as @DTMAboutHeart on Twitter) and Emmanuel (Manny) Perry, who created and runs the fantastic hockey stats site Cosica. Both Dawson and Manny have published the full methodologies of their models, which you can read here: DTMAboutHeat’s model, Manny’s model. For the purposes of this article, we will use Manny’s model, as it is freely available to all on Corsica at the player and team level and it is provided in (near) real-time as games are occurring.

Now, back to how expected goals works, and why it is so important. Expected goals accounts for both shot quantity and quality by accounting for a variety of shot factors, including shot type (wrist, slap, deflection etc.), distance from the net, shot angle, whether a shot was a rebound or generated off the rush, and if it was taken on the power play, even strength or on the penalty kill. In terms of shot quantity, Manny’s model only considers unblocked shot attempts. For additional information on exactly how the model works and the weights, regressions etc. involved, please reference Manny’s methodology, which I’ve linked to above.

Expected goals is far from a perfect stat. In fact, there is no such thing as a perfect stat, so do both yourself and me a favor and stop complaining about stats, or anything in life for that matter, that “aren’t perfect.” Anyway, back to my point; while expected goals is not a perfect stat, when combined with watching the game and more traditional shot metrics like Corsi or Fenwick, it can be a valuable tool for evaluating games, players and teams. In fact, in Rob Vollman’s Hockey Abstract 2017, he breaks down how adding expected goals to prediction models has a measured impact on the model’s predictive ability, and specifically notes that, “even-strength expected goals, by themselves, slightly outperformed shot differential when predicting next season’s results.” DTMAboutHeart also noted that expected goals serves as a better predictor of future success than Corsi, and provides a graph demonstrating the predictive abilities of expected goals against Corsi and goal differential.[/text_output][image type=”none” float=”none” src=”703″ alt=”” href=”” title=”” info_content=”” lightbox_caption=”” id=”” class=”aligncenter” style=””][text_output]Enough talking notionally about why expected goals is important; you get it, I think the stat is important and I believe you should too. But how do we use it for analysis? As I said earlier, it should be used along with the eye test and other data points. Here are some examples that I believe help provide context to the stat’s usage and demonstrate how it can be used.[/text_output][custom_headline type=”left” level=”h5″ looks_like=”h5″ id=”” class=”” style=””]Individual Game Analysis Example: Rangers vs. Canadians[/custom_headline][line id=”” class=”” style=””][text_output]The Sunday, October 8 game against Montreal is a perfect example of the important role expected goals can play when analyzing a single game. I watched the entire game and came away thinking, “well the Rangers won 2-0, but holy hell did it seem like they were on their heels a lot.” I opened Twitter and saw your typical smattering of talking heads throwing out Corsi numbers to show the Rangers got their asses whooped. To be fair, they did get absolutely smoked in Corsi and Fenwick. But I also saw a number of people trying to argue that it felt that the Rangers got just as many high quality chances, if not more, than Montreal did. Enter expected goals, a stat that considers both shot attempt quantity and quality. According to Cosica’s live game tracking feature, in 5v5 play, the Rangers lost the score adjusted Corsi For % (CF%) battle 66.3%-33.7%—a true ass kicking. As a side note, I always advocate using score adjusted Corsi, as it weights shot attempts to account for the fact that when teams are losing, they are likely pressing, and therefor putting up more shot attempts than their opponents.

However, the expected goals column of Corsica’s stat sheet paints a different picture. Montreal did win the 5v5 xGF% battle (Expected Goals For % = percentage of expected goals a team got compared to the opponent). However, their xGF% was far smaller than their CF% advantage, leading the Rangers just 53.9% to 46.1%. Long story short: all of the eye test advocated on Twitter had a point: in terms of expected goals, which accounts for shot quality, the game was in fact much closer than many would have you believe when only throwing around shot attempt metrics.[/text_output][custom_headline type=”left” level=”h5″ looks_like=”h5″ id=”” class=”” style=””]Team Season Analysis: Rangers 2016-2017 Season[/custom_headline][line id=”” class=”” style=””][text_output]

By this point, most fans are probably well aware of the fact that the Rangers have been a poor possession team for a few years now, and particularly last year. The Rangers finished the season last year with a CF% of 47.9%, good for 5^th worst in the NHL, and better than only the Coyotes, Sabres, Islanders and Devils. In fact, the only other playoff team in the bottom 10 in Corsi last year was Ottawa. However, similar to the Montreal game earlier in the year, a lot of fans believed that the shot attempt numbers were a poor reflection of the Rangers quality of play, and that what they lacked in quantity, they made up for in quality.

The Rangers were 8^th worst in the league in xGF%, so while it still isn’t good, it’s an improvement from their Corsi standing. Keep in mind that quantity still plays a major role in expected goals, so there is only so much a team can improve due to quality when their shot quantity is so low. But again, the point is that when you factor in expected goals into the equation, the numbers start to more closely align with what your eyes are telling you.[/text_output][custom_headline type=”left” level=”h5″ looks_like=”h5″ id=”” class=”” style=””]Individual Player Analysis Example: Michael Grabner 2016-2017 Season[/custom_headline][line id=”” class=”” style=””][text_output]Expected goals data can be used in multiple ways when analyzing an individual player’s performance over the course of a season. One usage method of expected goals data that I’m particularly fond of is comparing a player’s individual expected goals for numbers (ixGF) to the number of actual goals they scored. This type of analysis helps illuminate whether the player under or over-performed in terms of his goal production throughout the season, and helps us get an understanding as to whether the level of goal production is sustainable going forward.

The Ranger player that had the biggest discrepancy between his actual 5v5 goal production and individual expected goals probably will come as no surprise to anyone: Michael Grabner. Grabner led the team with 21 5v5 goals, and posted a 1.4 5v5 goals per-60, which led the team by a significant margin. However, Grabner’s ixGF was only 14.41, second on the team to Kreider and over 6 goals less than his actual total, representing a significant drop off. Further, his ixGF per-60 of 0.96 was second on the team to Rick Nash, representing a 31% drop off from his actual 5v5 goals per-60 of 1.4. Long story short, Grabner got significantly lucky with respect to his goal production last year, a fact that many also witnessed and pointed out on social media while watching the games.[/text_output][custom_headline type=”left” level=”h5″ looks_like=”h5″ id=”” class=”” style=””]Conclusion[/custom_headline][line id=”” class=”” style=””][text_output]Expected goals is far from a “perfect stat,” and like any metric, should not ever be used as the end all be all for player, game or team evaluations. However, it can help us engage in more comprehensive analysis of the great sport of hockey. In my personal opinion, expected goals data serves as an excellent complement to more traditional possession and shot-related metrics, and when combined with other data and the eye test, can help us not only better understand the current level of play for both players and teams, but also help us project the sustainability of play going forward.[/text_output]