The GAR models and such are interesting, but I never really use them unless for a project where I go through a ton of players.
GAR prides itself on being a much better predictor than any one other metric, but the thing is that most of those metrics are not supposed to be looked at alone. And the problem with a "one metric to rule them all"-kind of solution is that when there's a clear issue, you can't use it to figure out why the results seem inexplicably (un-)flattering. All the things you want to look at is buried inside an algorithm, and whatever you look at outside the metric runs the risk of being double used (forgot the term for that.)
Yesterday I had a discussion about Girgensons, who come across very badly in most metrics. I had said that he shouldn't be any higher than a teams fourth line, because a statistical analysis put him as a weak fourth liner. A poster challenged this, and when I dug deeper to find support for my case I found that he was actually right.
Girgensons numbers get tanked because he played on the fourth line with incredibly limited players that he couldn't carry, so he sunk with them. But when he was placed with skilled players in a complementary role, he was a boon for them all.
That poster was absolutely correct, and I was wrong. Being able to take the numbers apart and look for discrepancies led to that information. Not only that, but it highlighted additional information about what type of player Girgensons is and how to best use him, and it helped bridge impressions and analytics. That's why I'd prefer the statistical community to continue working on providing more tools, not seek the glory of the one solution, as it'll always be faulty without remedy.
Some things I'd like to see:
- Better xPDO (taking into account things like increased competition leading to lower expected save percentage, which isn't taken into account now)
- Better tools to assign credit (to figure out how much credit Tkachuk should have for doing great on a line that's always doing great)
- Better QoC metrics (because the ones we have now are flawed)
- xF% metric (because usage for faceoffs could illustrate the difference between a go-to guy and an ordinary center much better)