Roster Turnover

[NoSix]

You're probably right, your suggestion is a better measure of turnover than mine (why didn't you show up earlier :)).

Sorry about that - better late than never?

Anyway, I don't think the problem with 2003-2002 having a different number than 2002-2003 is really such a problem; if you're actually looking for turnover, going backwards in time from 2003 to 2002 doesn't make much sense.

That thought occurred to me, too, after I posted - I agree with you.

[ChrisE]

Sorry about that - better late than never?

Yeah, but I hope you'll understand I'm not going to be urgently trying to try out this new measure of turnover (the process is a little time consuming). As always, if anybody doesn't want to wait till the arbitrary date when I feel like doing it, I'd be happy to share the file.

[ChrisE]

So, I'm reading through my daily bigsoccer threads, and to my surprise, I find myself cited (!) in one. You can read through my brush with fame here (http://www.bigsoccer.com/forum/showthread.php?t=105552). Anyway, this brought about such a strange and satisfying feeling that I felt compelled to do some more work on the topic at hand.

So, anyway, the question is, how is roster turnover related to win percentage change? In particular, do good teams who don't make enough changes on average lose more ground than those who do?

I'm still using the numbers that I drew up, that may not best be described as turnover, but they're something similar. And I doubt using NoSix's idea of turnover would make much of a difference anyway. One day.

Leaguewide, we see an average win % change of -.001 (presumably due to rounding error), and an average turnover (which is actually 1-the stats listed above) of .471. There was a relatively strong correlation between win % change and turnover, about .24.

But I grouped all the league's teams into 3 categories, bottom quartile, top quartile, and middle half, and took their average win % and turnover. Bottom was between .234 and .422, mid was between .429 and .578, max was between .594 and .750. They broke down like:


Turnover Win %
Bot. 1/4 0.593 0.086
Mid 1/2 0.466 -0.005
Top 1/4 0.344 -0.091



It's pretty clear that the deal here isn't that changing your team up is a good thing, but that bad teams tend to improve more than good teams, and that, for obvious reasons, bad teams tend to change their personnel more than good teams.

So, if you do the in-group correlations, you get:


Bot. 1/4 0.01
Mid 1/2 -0.15
Top 1/4 0.00


Which, to my surprise indicates that average teams do slightly better if they stick with what they have; however, it's clear, and this was pretty surprising, that winning % change for the top and bottom teams has very little to do with turnover.

[numerista]

Which, to my surprise indicates that average teams do slightly better if they stick with what they have; however, it's clear, and this was pretty surprising, that winning % change for the top and bottom teams has very little to do with turnover.

Great stuff, Chris. I'm a bit surprised by your last result, though, and I'm wondering if I'm mis-reading it, or if it might possibly be incorrect. Given your sample size, it'd be surprising to see correlations that low, even if the true correlation were zero.

[ChrisE]

Great stuff, Chris. I'm a bit surprised by your last result, though, and I'm wondering if I'm mis-reading it, or if it might possibly be incorrect. Given your sample size, it'd be surprising to see correlations that low, even if the true correlation were zero.

Admittedly, it did happen to look a good deal cleaner than it should have, and maybe I should have admitted that (but the .01 and 0.00 were just so pretty...).


Regardless of where exactly you put the endpoint, the correlation for the lower quartile remains around or below .1. For the top quartile it stays between -.05 and .05. So, perhaps I shouldn't have drawn any conclusions about the middle 1/2, when the top and bottom aren't really too far off (although, interestingly, the correlation gets stronger when you divide it into mid-bottom and mid-top quaritiles). Nevertheless (although it's always possible I've screwed up), it looks to me like turnover isn't a significant factor in improvement/decrement for teams.

[numerista]

Admittedly, it did happen to look a good deal cleaner than it should have, and maybe I should have admitted that (but the .01 and 0.00 were just so pretty...).

Just to add ... I've looked over this data, and ChrisE's work is correct. I would paraphrase his conclusion as follows:

Good teams tend to drop off a little bit, regardless of how much they turn over their rosters. Likewise, bad teams tend to get better, regardless. This is (in part) what statisticians call "regression to the mean."

There is a very slight indication that mid-tier teams do better if they stick together. This wouldn't be too surprising either ... a mid-tier team is more likely to stick together if it knows it has a good chance of improving.

[numerista]

After further analysis, the "very slight indication" I mentioned is actually a very strong indication. A regression of

Change in Win % on Old Win % and Turnover

comes out highly significant for both predictors. Once you've adjusted for your past season's win percentage, the less turnover you have, the better.