A short thought on my usage of models

Yesterday @Woodguy55 asked me a question in the comment section of Allan Mitchell's blog Lowetide.ca:

"How do you feel about Corsica giving Larsson and Benning an almost identical Player Rating?"

I added a bunch of additional information(to the preamble directly before the first infograph) about how and why I use models, that should be posted here for reference.

Benning is a guy who we know has better macro shot share impacts, so this is a guy we know who is getting a boost.

Manny’s model we know rewards guys who offensively outperform their xGF, but doesn’t (or sometimes inversely) reward guys who defensively outperform xGA.

If a model makes you double-take, just use general rules about the strengths and weaknesses of it, and see if there’s consistency in players it underrates and overrates. I know Larsson and Benning are gonna have opposite-ended directional pulls just from what the model values before I even peruse their numbers, I’m not surprised by it.

I look at the methodology guys are using /then/ grab their data, every time.
Some of the language gets lost in translation, I do what I can.

Another example, Dom Luscycyzn’s game score loves point production. His model doesn’t like Larsson either. Guys who do more than produce points as play drivers get underrated, Nino-Hall at forward, Tanev-Hjalmarsson at D, to steal Ian Tulloch’s distillation.

You can see the weighting in action here

Player Game Score = (0.75 * G) + (0.7 * A1) + (0.55 * A2) + (0.075 * SOG) + (0.05 * BLK) + (0.15 * PD) – (0.15 * PT) + (0.01 * FOW) – (0.01 * FOL) + (0.05 * CF) – (0.05 * CA) + (0.15 * GF) – (0.15* GA)

Heavy on points.

Larsson’s two seasons in 15-16 and 16-17 tell me models like those two aren’t going to like him because he’s a usage freak who can not get murdered in roles 95% of Dmen would get murdered in.

The purpose of the models in the exercise isn’t really to dig deep on Edmonton’s D corps but to get a cursory glance at the other Pacific teams without having to do the level of research I feel was necessary to get the full scope of things.

I am reasonably comfortable in thinking I have a more detailed understanding of the individual Dmen on Edmonton than what the model can tell me, not entirely so with the other 40+ guys in the division.

The reason I’m not gonna give Larsson a bump in the numbers in the excercise or say that he should get one is because then you have to give that to Tanev and every other guy in that player type, and then you also have to ding Burns et ceter et cetera.

The other part is, especially with the GAR I was curious to see what the projected net of pairings were.

Like I said earlier in the thread, I’m not super confident in my ability to project a /pairing/ because of issues weighting stuff, especially in pairings where there’s one guy who’s good and one guy who’s bad. The GAR gives you a little picture like hey, the guy who’s good adds eight and the guy who’s bad only takes away two, so they might be alright.

That’s an interesting thing to note and come back on, to see if the model was right, wrong, and how often either. Or in what ways they were right and wrong. Then I take that and maybe I’m closer to projecting pairings than I am now.