Jimmystats: A Crutin Consensus

Submitted by Seth on January 20th, 2016 at 3:21 PM

Back in the day the recruiting roundups that Ace would put together would show the star ratings from each site of the various Michigan targets. The problem was we kept noticing dramatic differences that weren't really dramatic. For example here's a table of guys given 5-stars by these services since the 2010 class:

Recruit Rivals ESPN 247 Scout
Jabrill Peppers
Derrick Green    
Kyle Kalis    
Ty Isaac    
Ondre Pipkins      
Demar Dorsey      
Dymonte Thomas      
Devin Gardner      
Shane Morris      
Patrick Kugler      
Kareem Walker      

Was Scout ludicrously high on M guys, or giving out more 5-stars? Actually they were all ranking not that far from each other, but Michigan just happened to get a lot of the guys in that 4-/5-star margin. It only looks dramatic because there are only five possible rankings.

This was recruiting until 247 introduced their composite rating. That composite is so amazingly useful for most "how good was he as a recruit" questions.

Rivals Scout ESPN 247 Stars
Top25 Top 25 92+ 100+ ★★★★★
6.1 Top 50 91 98-99 4.75
6.0 Top 75 89-90 96-97 4.50
5.9 Top 150 87 95 4.25
Top250 Top 300 86 90-94 ★★★★
5.8 4-star 80-85 88-89 3.75
5.7 3-star* 79 86-87 3.50
5.6 " 76-78 84-85 3.25
5.5 " 72-75 80-83 ★★★
5.4 " 70-71 78-79 2.75
5.3 2-star* 67-69 75-77 2.50
x " 66 72-74 2.25
5.2 " 63-65 70-71 ★★
5.1 " 62- 69- 1.75

Since forever I've also been maintaining this spreadsheet of data on Michigan players that started as a naming sheet for some iteration of the NCAA game, and just kept gaining columns. My old way of tracking the recruiting ratings on that was to take the stars each service gave out, figuring they all roughly had the same definitions, and average them.

But that was throwing away a ton of information provided by the sites, which typically post national rankings for the top ~250-300 recruits, and in three of their cases have their own more precise star rating systems. For example Rivals's 5-star range includes "6.1" and "6.0", while ESPN (50-95) and 247 (69-102) have numeric scales with the decades roughly coinciding with the next star rating.

They also have position ratings, which don't match up since they split positions differently, but if they can all be turned into percentiles.

So far I've done all but the last bit. Matching table's above. What we end up with is not a composite system like 247's so much as a composite Star Rating system that quadruples the star precision level.

I tried to honor stars and what they mean, but I also took national rankings and position rankings into account when one site's rating spanned multiple ratings of its competitors. So a 5.8 on Rivals will be a 4.00 if he makes the Rivals 250, and a 3.75 if he doesn't. And a 3-star WR on Scout who's ranked just behind the 4-star receivers in the WR rankings is like a 3.5-star.

[After the Jump: charts until I literally break Excel]

Let's have another look at that annual bar graph I produce to compare the classes. The one on the left shows 2016 as it stands and the right is the projected class.


The "projected" 2016 class is just me guessing based on the 'crutin updates Brian publishes, figuring guys he's hinted at aren't in the class, and Michigan picks up, you know, THAT GUY plus some other guys we seem to have heavy leads for. Take it FWIW—if you've got a subscription to anything harder than Twitter you can probably make better guess than that.

You can look at the constituent rankings (still using projected for the 2016 class cause I like to count unhatched chickens shut up!)



Some things may shift once I've gotten the pos ranking percentiles to work.

Best Class Evah?

Going by cumulative stars, Michigan's 2016 class would currently be up there among the better ones in recent Michigan history. Best-ever is within reach.

Class STARs sum Players
2016 (proj.) 108 29
2013 103 27
2012 92 25
2010 90 27
2008 86 24
2016 (cur.) 84 23
2005 84 23
2004 82 22
2009 79 22
2002 71 21
2011 67 20
2006 66 19
2007 66 19
2003 63 16
2014 60 16
2015 50 14

That's good news since the last two classes by this measure were the weakest in that time period. That's a measure of being nearly (or exactly) half the size though. You'll also see that the classes didn't necessarily translate into better teams. The 2015 team certainly benefited from Hoke's excellent 2012-'13 classes, and 2004 and '05 played their part in 2006. Crutin matters! Also holding onto those recruits matters.

Attrition by star rankings.

I just broke Excel. :( I guess play with the spreadsheet and see what you can see.




January 20th, 2016 at 3:32 PM ^

A composite Star Rating system that quadruples the star precision level??


(On a serious note, thanks for posting Seth. Interesting stuff.)


January 20th, 2016 at 4:07 PM ^

2016 (proj.) 108 29 3.724138
2013 103 27 3.814815
2012 92 25 3.68
2010 90 27 3.333333
2008 86 24 3.583333
2016 (cur.) 84 23 3.652174
2005 84 23 3.652174
2004 82 22 3.727273
2009 79 22 3.590909
2002 71 21 3.380952
2011 67 20 3.35
2006 66 19 3.473684
2007 66 19 3.473684
2003 63 16 3.9375
2014 60 16 3.75
2015 50 14 3.571429


Toasted Yosties

January 20th, 2016 at 4:22 PM ^

Each class had no worse than a 3.6 stars/player ratio with an average for those three classes combined at 3.75 stars/player. Turned out pretty good for 2006 and seemingly good for 2007.

I'd like to see the stars-to-player average using this method for national champions and conference champions by conference in the recruiting sites era. Would be interesting to see what the basement ratio for success would be and what an average national champion team's stars-to-player would look like.


January 20th, 2016 at 3:44 PM ^

Seth, I really think you would love using something like R for your analyses. You could make it reproducible so it builds and updates everything for you. (and R has the ggplot2 package which makes all plots prettier).


January 27th, 2016 at 2:56 PM ^

Most of what you're working with (I'm guessing) boils down to some comma separated value (CSV) files, or maybe a few of them linked together based on some primary key. 

You're applying transformations to the data (functions in excel), then you're visualizing the results. 

There is an open-source statistical programming language called "R" that is basically designed for this use-case (almost all the code has been written by statisticians, for better or for worse).

It would require a bit of a learning curve, but I'm guessing your overall productivity would increase if you were using R rather than Excel, since there are are steps in your Excel process you could automate. 

In particular, the "dplyr" package is valuable for data transformations, you can write your own functions (that you can name and call as often as you'd like), and you can visualize (I love the "ggplot2" package).


January 20th, 2016 at 4:09 PM ^

from the Hoke experience. The star ratings are substantially compromised if the players are not coached well. Its an interesting exercise with that caveat.


January 20th, 2016 at 4:41 PM ^

The funny thing is I haven't felt like our players are particularly well coached since about 2001... Tressel & Dantonio & Saban & Urbz' players always seem to be quicker to the point of attack and using better leverage than ours. And often with game plans that precisely expose our shortcomings.

I guess that is a LOT of National & B1G Championships in there, but that's where I'm sure we'd all like to be, right?!

This has become a bit more controversial saying in this board lately, but "In Harbaugh We Trust!"


January 20th, 2016 at 5:00 PM ^

Made a quick model of expected stars by the rank of a recruit in Michigan's class and the size of the class and then took the average difference between the actual and the expected to rank M's recruiting classes in Seth's spreadsheet. E.g if M's best recruit in the class is typically a 4.5 star guy, in 1998 we would have expected him to be rated 4.94 stars.

Just wanted to correct for class size, really. But not sure that actually did that, since class size and starz were positively correlated and I would have expected the opposite. 


January 21st, 2016 at 12:38 AM ^

Class size also correlates with things like producing starters and NFL players and such because moar dudes is good and Michigan has only ever used all of its schollies once in this period (2002). USC once had a small class of almost all 4-stars but it's rare that any team will have more blue chips they can get to commit than spots available.

These are individuals and time to recruit them is limited so I give credit for finding 29 of them cumulatively.


January 21st, 2016 at 12:41 AM ^

If anyone cares it turned out to be a problem with an Adobe update that conflicted with Microsoft's static files. I did some registry finagling to simulate a fresh Office install and it worked.