We’re back to work on the NHL model when we have moments (still hoping to have it done in time for the playoffs), and we made a moment for it today.
One of our curiosities about hockey is how predictive shot volume is of goals scored. We’ve read a little on the topic, but after starting that, we realized it was quicker to do a regression, so we did that instead.
Our regression isn’t perfect. It only counts shots, not blocked shots or missed shots, and it includes playoff games and overtime and treats all of that equally (it does stop the count before shootouts, so we aren’t dealing with shootouts with these). There are arguments to be made for or against our methods on all of these fronts, but the goal here wasn’t to deeply explore the relationship between shots and goals in individual games. The goal was to, in the broadest sense, figure out what general sort of correlation there is between the number of shots a team takes in a given game and the number of goals they score.
The answer?
A pretty weak one.
We put five seasons of data in (with the 2019/20 data a bit odd because of the pandemic break and subsequently unusual playoffs), which means there are 6,563 games in here, with two data points per game (home shots vs. goals, away shots vs. goals). So, over 13,126 pieces of data, we got the chart above, where we can see that there’s a relationship of some sort (we went with the box-and-whisker plot, in which fifty percent of data points are within the solid blue boxes for each item on the x-axis, because a scatter plot was making outliers look more significant than they are), as well as a correlation coefficient of 0.116 and an r-squared value of 0.013, implying some positive correlation, as we’d expect, but very little predictivity. Tell me a team took 30 shots and I can tell you the most likely number of goals they scored is three, but that they very well might have scored two, or four, or one or five or etc. Tell me a team took 35 shots and I can tell you the most likely number of goals they scored is three, but that they very well might have scored two, or four, or one or five or etc. Tell me a team took 40 shots and I can tell you the most likely number of goals they scored is three, but that they very well… Helpful, right?
What’s more, we found that the team that took more shots, over these five seasons and playoffs of play, actually lost more often than they won, posting just a 49.7% win percentage in games that didn’t end in shootouts or with an equal number of shots by each team. Even more helpful, right?
It actually is helpful for the model, though it may sound strange, with the basic reason for that being that this tells us to not rely too heavily on shots, and to specifically not count shots as too strong a proxy for goals in individual games (and to certainly not count shots as a proxy for wins).
It’s possible, when we compare shots vs. goals and shots vs. wins on a team-by-team basis over entire seasons or other stretches of play (yes, we’re sure this has been done somewhere, but again—it’s quick enough that it makes more sense to just do it ourselves), that we’ll find shots are important, or are a good reflector of how good a team is, or should at least be included in whatever eventual formula or formulas we use to reflect the strength of a team. But at a high, broad level, we know this: Shot volume is not going to tell us as much in hockey as I’d guess it does in soccer or basketball.
So, that was today’s moment of NHL understanding. We’ll try to keep them coming. And hopefully, in the next few weeks, have a model to send your way.
Note: If I messed something up terribly here, please correct me. I keep checking and not finding anything, though. It seems so counterintuitive, but it also seems to be true?
I’d be interested to know how this changes by period, and when a team is leading/trailing. It seems similar to the fallacy that “running the ball in the fourth leads to wins,” when the opposite is more true – leading teams are burning clock by running. My hunch would say that a trailing team would be throwing a flurry of pucks at the net, especially late, whereas a team protecting a lead would be trying to play keepaway.
That would make sense! Hopefully we’ll look into it one day (don’t have period-by-period stats at the moment, but would be a good offseason dig).