What goes into those projections?
The projections are entirely at the player level and are based on an all-in-one stat I created a little over a year ago called
Game Score. It combines all the basic box score stats into one number to measure a player’s value. As one season isn’t really enough to get a good read on a player, I use the last three seasons instead. The seasons were weighted by recency using a multi-variate regression and the results of that are regressed to the mean based on the repeatability of each component and the size of each player’s sample. There’s also an age adjustment as we expect players to get better as they move towards their prime and worse as they move away from it. Lastly, there’s a small adjustment for
usage.
2018-19 Edit: The final output is then turned into a win value above replacement level, that I’ve dubbed GSVA, or Game Score Value Added.
Why is this player who is obviously better rated worse than this other player who I don’t like as much?
Game Score, like any stat, isn’t perfect. Especially not with current data. There’s a number of reasons a player can be rated higher, but understanding the weights of each input help uncover why. The two player types that are generally underrated are defensive defencemen (Niklas Hjalmarsson, Marc-Edouard Vlasic) and playmakers (Nicklas Backstrom, Ryan Johansen) while the players that are overrated are point-scoring defensemen (Kevin Shattenkirk, Dougie Hamilton) and shooters (Viktor Arvidsson, Max Pacioretty). The win values are also based on efficiency, so players who do well in a short period of time look better (David Pastrnak, Matthew Tkachuk) as do players who play with other good players. Allocating credit properly is an interesting problem in hockey, and while I do think Game Score does a decent job of it, there are some instances where it may look off.
Do you honestly think Player X is better than Player Y?
Just because a model says this doesn’t make it true, even if I built it. It’s not infallible — I already told you I’m not all-knowing. I don’t agree with every single thing it spits out, but there are many times where it makes me reconsider my feelings about specific players. If the model is showing that Player X is better than Player Y, there is probably a reason for it and it’s an opportunity to dig deeper. It may be wrong, but it creates dialogue and discussion worth having about a player.