Untitled Document
Brown Columbia Cornell Dartmouth Harvard Penn Princeton Yale



 Page 22 of 22 « First<19202122
Username Post: This proves that Covid can be managed at Universities        (Topic#24887)
palestra38 
Professor
Posts: 32682

Reg: 11-21-04
Re: This proves that Covid can be managed at Universities
03-24-21 02:44 PM - Post#322503    
    In response to mrjames

And it's your view that this is any more likely to result in statistical accuracy (or on-court success) than is the NFL or NBA draft, based on far more developed metrics than high school basketball? Because both NFL and NBA success is notably hit or miss when it comes to draft position.

 
mrjames 
Professor
Posts: 6062

Loc: Montclair, NJ
Reg: 11-21-04
Re: This proves that Covid can be managed at Universities
03-24-21 03:01 PM - Post#322504    
    In response to palestra38

Generally recruiting rankings have predictive power at the top and bottom end. That is, the highest end of Ivy recruits in terms of rankings have very high hit rates and the bottom (no rankings at all) have had very low hit rates, while everything in between tended to be meaningless.

IIRC the last model I ran basically had win share credit for the top recruits (say 2* average and up) and then win share credit for the *number* of recruits.

So, I don't think the model is telling us a ton we don't already know: Really highly rated recruits are generally good, and anything beneath that is a relative crapshoot.

Anecdotally, I think the model could be greatly improved with *real* offers, but those are hard to come by in a bulk fashion. My own guesses at contributions aren't really based on this model but rather offers and what the programs are seeing from their players.

 
OldBig5 
Masters Student
Posts: 639

Age: 66
Reg: 02-18-18
Re: This proves that Covid can be managed at Universities
03-24-21 04:07 PM - Post#322505    
    In response to mrjames

  • mrjames Said:
Generally recruiting rankings have predictive power at the top and bottom end. That is, the highest end of Ivy recruits in terms of rankings have very high hit rates and the bottom (no rankings at all) have had very low hit rates, while everything in between tended to be meaningless.

IIRC the last model I ran basically had win share credit for the top recruits (say 2* average and up) and then win share credit for the *number* of recruits.

So, I don't think the model is telling us a ton we don't already know: Really highly rated recruits are generally good, and anything beneath that is a relative crapshoot.

Anecdotally, I think the model could be greatly improved with *real* offers, but those are hard to come by in a bulk fashion. My own guesses at contributions aren't really based on this model but rather offers and what the programs are seeing from their players.



That makes lots of sense based on what I have seen over the years. Thanks for your analysis and rankings also.

 
mrjames 
Professor
Posts: 6062

Loc: Montclair, NJ
Reg: 11-21-04
Re: This proves that Covid can be managed at Universities
03-30-21 04:34 PM - Post#322645    
    In response to OldBig5

Fired up the model again. Here's the class-by-class Win Share projections at present (win shares projected, rank among all classes since 2002, out of 160):

1. Harvard 2019 (28.3, #6)
2. Harvard 2018 (24.3, #10)
3. Harvard 2020 (22.8, #14)
4. Princeton 2018 (21.3, #21)
5. Penn 2019 (20.9, #25
6. Columbia 2021 (17.9, #43)
7. Yale 2020 (16.9, #49)
8. Princeton 2019 (16.3, #51)
9. Dartmouth 2019 (15.4, #62)
10. Cornell 2020 (14.7, #69)
11. Princeton 2021 (14.5, #74)
12. Columbia 2019 (13.4, #84)
13. Penn 2021 (13.0, #89)
14. Brown 2021 (12.8, #92)
15. Brown 2020 (11.6, #109)
16. Columbia 2018 (10.7, #115)
17. Brown 2018 (10.7, #116)
18. Columbia 2020 (10.5, #119)
19. Penn 2020 (10.2, #120)
20. Princeton 2020 (10.2, #121)
21. Cornell 2019 (10.0, #123)
22. Dartmouth 2020 (10.0, #124)
23. Yale 2021 (9.3, #131)
24. Penn 2018 (9.0, #133)
25. Cornell 2021 (8.8, #137)
26. Dartmouth 2021 (8.8, #138)
27. Yale 2018 (8.3, #140)
28. Harvard 2021 (7.1, #143)
29. Brown 2019 (7.0, #145)
30. Dartmouth 2018 (6.6, #146)
31. Cornell 2018 (6.0, #149)
32. Yale 2019 (5.9, #150)

And at a total team level, here's where the predicted win shares stand for the 2018-2021 classes:

Harvard 83
Princeton 62
Penn 53
Columbia 52
Brown 42
Yale 41
Dartmouth 41
Cornell 40

The top and bottom two are pretty unsurprising, but the middle is incredibly interesting. Brown and Yale dramatically underachieve for two different reasons. The model hates small classes, so Yale gets dinged in 2019 and 2021 (2 and 3, so far, respectively). And despite Brown's classes having some heft, there is a tremendous mismatch for how those classes are viewed in New England versus by the national recruiting services.

Meanwhile, Princeton and Penn recruited four more players in the past three classes than Yale and Columbia recruited six more, which is much of the reason why two of Yale's last three classes are among the six worst over these four years while only Penn 2018 is in the Bottom 10 from those three schools from 2019-2021.

Then, if you look at just the two classes we haven't seen (2020 and 2021), here are the expected Win Shares:

Harvard 31
Columbia 28
Yale 26
Princeton 25
Brown 24
Cornell 24
Penn 23
Dartmouth 19

Clearly, if you take out Penn's substantial 2019 class, it sinks quickly. Yale's two-person 2019 class gets removed, so it subsequently rises.

The watchout, though, is that the average win share year across the sample is 13.7, so only two teams are seen has having recruited "above average" for this cycle. I don't think that matches reality, and I do think that there are two reasons for it:

1) Many ratings systems just aren't going as deep as they used to, leading to a lot more "0s" that would have been 1 or 2 stars when, say, Future150 and ESPN were more active. For instance, from 2010 to 2017, ESPN recognized 240 out of 298 Ivy recruits (many as NR, but at least recognized). But it pivoted its strategy to only focus on the top 200 or so in 2018, and since has only recognized 48 over 139. This has a slight impact on the model, as the model rewards team for the sum of average recruit scores and the difference between a half-star and zero-star rating over a handful of recruits can meaningfully change expected WS.

2) Ivies aren't loading classes as much as they used to be. The Ivies recruited 39 kids per class over the first six years of the 2010s, but only 35 per class during the 2018-2021 cycle. Since the model sees value in having more D1 credible recruits to choose from, lowering the number of recruits lowers the average number of win shares.

For these reasons, I think the model can still be a useful relative assessment tool, but it struggles at comparing across cohorts given the changes in the ratings systems strategies. All previous caveats apply, including, most importantly, the notes that regional recruiting rankings and true offers would lend a LOT more predictive ability to this model if they could be reasonably attained.

Edited by mrjames on 04-01-21 11:52 AM. Reason for edit: No reason given.

 
Penndemonium 
PhD Student
Posts: 1877

Reg: 11-29-04
04-01-21 01:00 AM - Post#322711    
    In response to mrjames

Thanks, mrjames. You may have explained this in the past, but is your model based on regressions from recruiting data and player contributions, or based on an intuition of most likely factors and their relative weight?

 
mrjames 
Professor
Posts: 6062

Loc: Montclair, NJ
Reg: 11-21-04
04-01-21 11:37 AM - Post#322766    
    In response to Penndemonium

The model is based on team-level win shares by class compared against the aggregated ratings for that class. It is indeed a multiple regression analysis with the features being number of recruits, sum of recruiting scores and recruits between 1.5 and 2.0 and 2.5 and 3.0 stars (which the model sees as generally over-rated when controlling for their contribution to the sum of scores).

A zero-star recruit is worth 2.8 average win shares, while a one-star recruit is worth 4.4 average win shares. But then, while the stars keep adding to the Sum of Scores, for 1.5 to 3 there is a penalty that takes away some of that benefit, as traditionally, recruits rated in that range don't perform increasingly well. That stops at 3-star and above where the performance is very strong.

I made a couple of tweaks to the model and will adjust the numbers above, but that's basically what's happening. Still... I think that having regional recruiting rankings and offers would MASSIVELY improve model performance (especially at the individual level, where this model isn't very strong at all).

 
 Page 22 of 22 « First<19202122
Icon Legend Permissions Topic Options
Report Post

Quote Post

Quick Reply

Print Topic

Email Topic

24968 Views




Copyright © 2004-2012 Basketball U. Terms of Use for our Site and Privacy Policy are applicable to you. All rights reserved.
Basketball U. and its subsidiaries are not affiliated in any way with any NCAA athletic conference or member institution.
FusionBB™ Version 2.1 | ©2003-2007 InteractivePHP, Inc.
Execution time: 0.224 seconds.   Total Queries: 17   Zlib Compression is on.
All times are (GMT -0500) Eastern. Current time is 10:30 AM
Top