View Single Post
Old
08-07-2013, 09:39 AM
  #101
Master_Of_Districts
Registered User
 
Master_Of_Districts's Avatar
 
Join Date: Apr 2007
Location: Black Ruthenia
Country: Belarus
Posts: 1,744
vCash: 500
Quote:
Originally Posted by Trainspotter View Post
So, basically, despite copious hand waving your answer is "no". You can't really argue:

A) that an r^2 value of .3 shows a strong relationship between two variables.
B) Fenwick shows a strong relationship to regulation points in a single season.
C) Fenwick is a good way to predict future regulation points either within a season or season-to-season.

0.5 is not a strong relationship. You may find that a stat allowing one to predict less than half of future results is "relatively impressive", but I'm not so easily sold.
It is relatively impressive - when appropriately contextualized. 47% of the variance in team results over a half season sample is explained by randomness. So there's a fairly sharp theoretical limit on how well any particular metric can predict results from one half of the season to the other.

It's not exceptionally impressive. More nuanced models - like the one MacDonald created - do a better job. Which is what we would expect. If I were predicting future results, I would use a model like MacDonald's. Or a model based on Bayesian inference.

But it is relatively impressive. Most people regard a team's record as a reasonably good proxy for its underlying ability. So for a metric to predict future results better than actual record - especially a very simple one like Fenwick - is relatively impressive.


Last edited by Bear of Bad News: 08-07-2013 at 10:33 AM. Reason: Flaming
Master_Of_Districts is offline   Reply With Quote