Already member? Login first!

Comments / New

Analytics Corner: We’re Going Streaking

During my daily perusal of #PredsTwitter, I recently saw a spate of suggestions that Kyle Turris should move to wing on the second line, and Craig Smith be moved to the third line.  Any of you that follow me by now are used to my rants—Craig Smith has played consistently well the last few years, and he shouldn’t be moved.

The criticism for him—as well as Filip Forsberg—is that both of them are “streaky” (be warned: I’m going to use the terms “streaky” and “streakiness” a bunch, so settle in). In fact, despite Forsberg being one of my favorites, I’ve said this about him too—he scores in bunches, then goes cold for some time. Still, I wanted to post this question to Twitter, and over 500 people responded:

Being the numbers nerd I am, I wanted to find some way to quantify “streakiness”, and I found that a brilliant person had already done so. Namita Nandakumar (@nnstats)—a Flyers fan, and currently doing analytics for the Philadelphia Eagles—presented her work on the subject at the Vancouver Hockey Analytics Conference last March.

The Method

Her statistical method is fantastic, and in the interest of keeping this relatively simple, I’ll attempt to spend no more than one paragraph summarizing her method—feel free to skip down to the results:

In order to quantify “streakiness”, she wanted to measure it over the entire course of the season, put the streaks in context (did the player score a little or a lot?), and work with a binary outcome: in a single game, did the player score or not? To account for all of these things, she calculated normalized entropy of the full sequence (an 82-game season): in short, seeing if 10,000 randomly generated sequences were more or less streaky than what happened in reality. If Viktor Arvidsson scored goals in 27 games, then 10,000 random sequences had goals scored in 27 games, and each was measured to find what percent were classified as “more streaky” than reality.

The Results

Before I show the results, I do want to preface it with this:

  1. Streakiness is a thing, but there could be external factors at play
  2. Streakiness is not repeatable—the correlations for a player scoring both between seasons and within a season are as unrelated as it gets.
  3. Streakiness is basically worthless as a measure of a player’s skill.

If you’d like to see the proof behind these statements, you can see Namita’s presentation here.

I made some modifications to Namita’s code to make this applicable to current and former Nashville players over the last 4 seasons, including only players who have played 41 games and scored at least 8 goals over each year of that span.

The players are ordered by average “streakiness” over the course of the last 4 seasons, and each colored dot represents a different season. At first glance, the Forsberg voters may not have been far off, as he’s second on the list. But how about if we looked at these same players, but only over the last two seasons?

Adjusting the order to represent the average over the last two years changes a few things (note that the years 2016 and 2017 are simply removed, 2018 and 2019 have not moved). So what can we take away from this?

The Takeaways

  • The 2015-2016 and 2016-2017 seasons for Filip Forsberg were both above 90%—he was pretty darn streaky. This probably has stuck in our minds and colored our perception of him as a goal scorer. But when you compare that to the last two years, he averages about 52%—good for a middle-of-the-pack streakiness value for the team and about the same that Jonathan Toews posted in 2018.
  • Craig Smith is pretty reliable, with his off 2017 season being the outlier. In fact, in 2018 he posted the lowest streakiness percentile of any player who played for Nashville for the full season at 15%. He jumped up to 62% this last season, but still, he’s also not the “streakiest” player on the team.
  • That honor goes to Nick Bonino, who has 2 of the last 4 seasons above 95%—the line at which we can say that his streakiness is statistically significant. He was absurdly streaky in those two years. Below is a visualization of the entire season, with games in which Bonino scored being shown:/

You can see above—he went huge stretches of games without scoring: nearly a month and a half to begin the year, and two nearly month-long stretches at the end, scoring only three after February 1. Remember that his low overall total does not matter—the model compared him to random sequences that had the same number of games with goals as he did. Bottom line: he scored in bunches, then disappeared. Which, all things considered, is about what you would expect from a third line center.

  • It makes some sense that defenders would rank high—they take a lot of very low-percentage shots from the blue line, and the team in front of them providing movement/screens can determine whether or not they go in. However, in both examples above, Ryan Ellis has the second and third best averages on the team—he’s pretty consistent.
  • Kyle Turris played through a lot of injuries during the 2016-2017 season, and playing hurt in Ottawa didn’t help him find any consistency. Limiting it to the last 2 years, he finishes as the most consistent goal scorer on the team—something I was not expecting./

So what does this all mean?  In reality, not very much. A player’s “streakiness” has next to no repeatability between seasons, so pointing at these results and expecting the player to perform the same way is flawed—there are so many other sources of variability that must be feeding into this. However, it is a great way to visualize and quantify how streaky these players really ended up being in a given year.

Oh, and it’s great backup to my personal feelings: leave Craig and Fil alone.

Talking Points