It basically factors in things like:
- Penalties taken
- Penalties drawn
- Shot locations + Passing data for expected goals for and against
- individuals' goal scoring history
etc for each individual player, with a "replacement" level player.
Here were Calgary's forwards' GAR last season, for instance:
A few areas where it might be flawed include:
- players grow (Tkachuk, Bennett could reduce their penalties, Monahan could and should improve his defensive play)
- historical sample sizes may be lacking for some of these younger players
- prospects join teams (Jankowski, Foo, etc)
- it weighs quality of teammate based on past season data, but sometimes teammates improve too
However, like I said earlier, it does have some demonstrable predictive value. We're not talking about rico stat here.