From the “almost too good to be true” department:
Hat tip to whoever the guy was that used that search to find my blog yesterday. See for yourself here.
Note the irony that I’m actually saying the opposite in the quoted snippet.
UPDATE: As of right now, Skeptical Sports Analysis is the #1 result for these searches as well (no quotation marks, and all have actually been used to find the site):
- why was dennis rodman so good at rebounding – Leads to Rodman series, Part 1/4(b). Makes sense, although strictly speaking I’ve never addressed this question.
- skynet does not know the rules of football – It’s a direct quote, so obv it leads to Skynet.
- nfl predictions using regression analysis – Oddly enough, this also leads to Skynet. I guess technically my neural network is a regression.
- how often are sports analysts actually – Leads to main page. Actually what? I don’t know.
Very thorough work makes this an interesting site.
What are your thought on Wages of Wins and wins shares per 48?
I find WoW intriguing, though my own research has led me to some different conclusions. I haven’t tested his results with win differentials yet, but Berri is normally on much better empirical footing than most other basketball analysts.
Also, much respect for Dean Oliver, whose book and website are brilliant (and he was sharp as a tack the one time I’ve talked with him at any length), but WS and WS/48 both perform rather dismally. WS does a little worse than PER, and significantly worse than PER*Minutes, and WS/48 does much worse than WS (as should be expected).
I actually think the data points being gathered in the box score (points, rebounds, assists, etc), aren’t necessarily the observations we want to record. There seem to be “sub-units” that would give you better quantitative analysis.
It seems there is already somewhat of a movement toward this, but what do you think of the thought that “we need better fundamental observations” to build a basketball model? Do you think what we have is good enough to crunch analytically and come up with something that’s not a hot mess? You can probably guess my views on this, since I’m questioning even the base statistics used.
I think there are two separate issues. One is: how you would theoretically build a perfect model? For that, you might be right that more and/or different data is needed. But the other is: how much information can we extract from the data we already have? From my basic research so far, I am led to believe there is a lot more that can be done on that front.
Very interesting stuff. I’ve been reading up on some of these issues, and my personal opinion is that Wins Produced (Berri’s stat) is terrible at doing what it purports to do. I think it takes some findings that make sense (other stats undervalue great rebounding, overvalue raw usage) and goes way off the deep end with them. In fact, so far this year i’ve found predictions of team wins that used the WoW framework to be just about the worst out there. By comparison, Hollinger is about the only one out there whose predictions right now are ahead of Vegas – granted i don’t know how much of a role PER plays in his predictions. Anyway i love the way you present the analysis, everything makes total sense to me, and i’d love to see you put WoW under the microscope.
Does anyone have a link to WoW calculations for players going back more than a few years? I would like to test it using win differentials, but I really don’t want to go through that long-ass process myself.
here are the WoW calcs i believe
http://nerdnumbers.com/automated-wins-produced
Thanks! I’ll check it out.