"I've made a huge mistake" - GOB

In a recent article on the role of luck and skill in shooting percentage, I used Steven Stamkos as an example of a player who shot for a very high percentage last season and who should see his goal tally drop this year as a result. Why I didn't choose Corey Porey, I have no idea. I don't play fantasy hockey, but I looked over a few projections and noticed that they all had Stamkos over 50 goals. Since I suggested he'd have a drop from the 45 he made last year, it seems like a deeper look at is in order.

To forecast Stamkos's goals for this season, I am using a method many others have used - finding comparable players to see how their numbers changed. For comparable players, I took all players that had a 40+ goal season at or below age 20. My source for this list is hockey-reference.com. There are a couple guys that fizzled out and some current players we'll have to wait on, but this is an elite list: Brian Bellows, Rob Brown, Jimmy Carson, Wayne Gretzky, Dale Hawerchuk, Ilya Kovalchuk, Pierre Larouche, Mario Lemieux, Eric Lindros, Rick Martin, Rick Nash, Owen Nolan, Alex Ovechkin, Barry Pederson, Luc Robitaille, Geoff Sanderson, Craig Simpson, Tony Tanti, Pierre Turgeon and Sylvain Turgeon.

Because Stamkos is about to start his age-21 season, I will forecast his goal total by looking at the relationship between goals per game for the above players for the seasons where they were 20 or younger and the season where they were 21*. Fortunately, there is a strong relationship, as you can see in the graph below:

Let's pause for a second to marvel at that 92-goal season for Gretzky. I left it zoomed out so you can see just how amazing that was. The other outlier was Mario Lemieux scoring 54 goals in only 63 games.

The regression spits out this equation:

GPG21 = 0.0362 + 1.0019*GPG1820

where GPG21 is the goals per game in the age-21 season and GPG1820 is the total average goals per game for all seasons where the player was younger. Notice how close the coefficient is to 1. That tells us that typically there is a direct relationship between average goals per game before 21 and at 21. The average difference between two guys is almost exactly the same the season they turn 21 and their careers to that point. Because that coefficient is very close to 1, we can see that the average elite young goalscorer improves by about 0.036 goals per game in his age-21 season. So if a guy only played full seasons, we'd expect him to score about 3 more goals than his career average for all years before age 21.

Thus far Stamkos has scored 119 goals in 243 games, or 0.49 goals per game. Using the above formula, this forecasts him at 43.2 goals for next year, a small drop from last season but up from his career average of just under 40 goals per season.

You might argue that Stamkos has improved a lot since his rookie season - he only scored 25 goals his rookie year and has put up 51 and 45 the last two. To look into that, I used the same methodology but only used the seasons where the player was 19 or 20 to forecast the year 21 season. The downside of this method jumped out - dropping that first season makes the earlier data far less descriptive and predictive of the 21-year-old season. Here's a graph:

While they look similar, you can see that there is less of an up-and-right pattern. The R^2 is 0.125 for this regression and was 0.292 for averaging all three seasons. So instead of explaining 29% of the variation in goal scoring at age 21, the average of the previous two seasons only accounts for 12.5% of it. So for that reason, I don't put as much stock into this regression. This forecast puts Stamkos at 49 goals this season.

Conclusion

Like most forecasts, it's all messy and things change when you tweak the model a bit or change the group of comparables. I didn't include it, but I changed the comparable group to be all players who scored at least 80 goals before age 21, which excludes several of the above names and includes Crosby, Trevor Linden and Steve Yzerman. That model also came out with a prediction of about 43 goals. While I think the first model above is the best, his rookie year Stamkos both took substantially fewer shots - 181 compared to 297 in 2009-2010 and 272 last seasons and his shooting percentage was substantially down - 12.7% compared to 17.2% and 16.5%. There is some reason to think that first year should be given less weight.

I think somewhere in between the two predictions is best and something very close to last year's 45 should be expected. This is an expectation, it's definitely possible that he could go on a tear and get over 50 and about as likely that he scores below 40, especially with injuries possible even for a guy with few injury problems.

* ages as defined by hockey-reference - the age the player is on February 1st of the season. Ilya Kovalchuk would have been 21 the year of the lockout so I used the season after for his count. Rick Nash turned 20 the season of the lockout so I shifted everything back a year for him as well.

let's not crown him just yet. take out gretzky/lemieux 92/54 seasons (also lol 80s/90s hockey), and see what the regression really looks like.

ReplyDeleteThanks for the comment, Sylar. Taking out the two best forwards ever puts him at 41 goals next year if he plays all 82.

ReplyDeleteOn Twitter Greenberg suggested I do more to take era into account. That also lowers his predicted-goal total, I'll post a follow up soon.

80's and 90's had higher scoring games so you simply cannot use those stats in direct comparison with today's game. EX. In 2003-2004 Kovalchuk led the league with 41 goals, but in 1984-1985 there were over 20 players with 40+ goals.

ReplyDeleteHere's a list that has taken league wide goals-per-game into account and put Ovechkin's recent 65 goal season into context:

http://michael-dell.suite101.com/nhl-adjusted-goal-totals-a107169

Yeah, Ovechkin's total is similar to Gretzky's absurd 92.

ReplyDeleteYou've reminded me that I never did that follow up. When I ran the numbers with era adjustments I came up with between 40 and 41 goals. IIRC Greenberg got something similar.