Reverse Engineering the Rivals Ranking Equation

#1
kidbourbon

kidbourbon

Disgusting!
Joined
Nov 12, 2005
Messages
5,040
Likes
20
#1
Has anyone ever taken a stab this? I was feeling nerdy today and started messing around on excel trying to model the data and come up with a way of reproducing the rankings. I got decent but imperfect results. Excel is limited to modeling the data in a linear or logarithmic fashion, and the rivals formula itself is obviously a bit more complicated than that.

I think I could get it quite close if I had matlab and a couple hours, but, alas, I do not.

Anyway, to the extent anyone is interested in putting this stuff in a spreadsheet and predicting where UT will end up if they get any number of guys, the last set of coefficients I came up with (which are just about as good as the 10 sets I had already come up with are:
#5 stars * 238.77 == A
#4 stars * 133.85 == B
#3 stars * 29.63 == C
#2 or 1 stars * -2.087 == D
-9 == E
A + B + C + D + E == approximation of rivals ranking.

To get this data, I just did a line of best fit on teams 1-100. The logarithmic model didn't come up quite as well. If anyone else is feeling nerdy and wants to take a stab at it, I'd be interested in seeing alternative approaches.
 
#3
#3
It might get more complex than that, as each guy has another score that designates his star ranking. Also, you can't tell how much better #1 is than #2 and #2 than #3. So your numbers can't really be calculated exactly.
 
#6
#6
kidbourbon said:
Has anyone ever taken a stab this? I was feeling nerdy today and started messing around on excel trying to model the data and come up with a way of reproducing the rankings. I got decent but imperfect results. Excel is limited to modeling the data in a linear or logarithmic fashion, and the rivals formula itself is obviously a bit more complicated than that.

I think I could get it quite close if I had matlab and a couple hours, but, alas, I do not.

Anyway, to the extent anyone is interested in putting this stuff in a spreadsheet and predicting where UT will end up if they get any number of guys, the last set of coefficients I came up with (which are just about as good as the 10 sets I had already come up with are:
#5 stars * 238.77 == A
#4 stars * 133.85 == B
#3 stars * 29.63 == C
#2 or 1 stars * -2.087 == D
-9 == E
A + B + C + D + E == approximation of rivals ranking.

To get this data, I just did a line of best fit on teams 1-100. The logarithmic model didn't come up quite as well. If anyone else is feeling nerdy and wants to take a stab at it, I'd be interested in seeing alternative approaches.
Click to expand...

If they really wanted to do this right they would match a teams needs against their signees, give premiums for lineman over skill positions, etc. Right now the ranking penalize teams with few schollys to give.
 
#8
#8
Lots of stars + 3 years + Kiffin & Co. = SEC Champs.

Can't help with your formula but maybe you could help the guys on the OKU thread understand why the Carrier Dome could use some air conditioning.

Tons of air conditioning = # of people times x # of btuh's, or something to that effect.
 
#9
#9
s1216 said:
If they really wanted to do this right they would match a teams needs against their signees, give premiums for lineman over skill positions, etc. Right now the ranking penalize teams with few schollys to give.
Click to expand...

That would be way to complicated. A formula you suggest would benefit teams who recruited poorly in the past because they would have a lot of need. It would be an endless see-saw. No, the current system is better.
 
#10
#10
VOLS2BSTACKED said:
Lots of stars + 3 years + Kiffin & Co. = SEC Champs.

Can't help with your formula but maybe you could help the guys on the OKU thread understand why the Carrier Dome could use some air conditioning.

Tons of air conditioning = # of people times x # of btuh's, or something to that effect.
Click to expand...

I was wondering if he could do my taxes.
 
#14
#14
kidbourbon said:
Has anyone ever taken a stab this? I was feeling nerdy today and started messing around on excel trying to model the data and come up with a way of reproducing the rankings. I got decent but imperfect results. Excel is limited to modeling the data in a linear or logarithmic fashion, and the rivals formula itself is obviously a bit more complicated than that.

I think I could get it quite close if I had matlab and a couple hours, but, alas, I do not.

Anyway, to the extent anyone is interested in putting this stuff in a spreadsheet and predicting where UT will end up if they get any number of guys, the last set of coefficients I came up with (which are just about as good as the 10 sets I had already come up with are:
#5 stars * 238.77 == A
#4 stars * 133.85 == B
#3 stars * 29.63 == C
#2 or 1 stars * -2.087 == D
-9 == E
A + B + C + D + E == approximation of rivals ranking.

To get this data, I just did a line of best fit on teams 1-100. The logarithmic model didn't come up quite as well. If anyone else is feeling nerdy and wants to take a stab at it, I'd be interested in seeing alternative approaches.
Click to expand...

You should absolutely, no reservations about it be banned for mentioning MATLAB on here. I have daily (not by choice) contact with it and hate it.

I come to this board to escape the world of engineering :)
 
#16
#16
kidbourbon said:
Has anyone ever taken a stab this? I was feeling nerdy today and started messing around on excel trying to model the data and come up with a way of reproducing the rankings. I got decent but imperfect results. Excel is limited to modeling the data in a linear or logarithmic fashion, and the rivals formula itself is obviously a bit more complicated than that.

I think I could get it quite close if I had matlab and a couple hours, but, alas, I do not.

Anyway, to the extent anyone is interested in putting this stuff in a spreadsheet and predicting where UT will end up if they get any number of guys, the last set of coefficients I came up with (which are just about as good as the 10 sets I had already come up with are:
#5 stars * 238.77 == A
#4 stars * 133.85 == B
#3 stars * 29.63 == C
#2 or 1 stars * -2.087 == D
-9 == E
A + B + C + D + E == approximation of rivals ranking.

To get this data, I just did a line of best fit on teams 1-100. The logarithmic model didn't come up quite as well. If anyone else is feeling nerdy and wants to take a stab at it, I'd be interested in seeing alternative approaches.
Click to expand...

As someone already mentioned, it's going to go off of their ranking score instead of just their star rating. It may be some sort of step function that jumps at the different star cutoffs, but I would bet that the effect on ranking can change within a star level.
 
#18
#18
You could also narrow down the accuracy by adding in the last few years' recruiting rankings so the sample size would be greater.

Aren't the number of stars based on a numerical rating? You'd have a numerical value worth a certain number of points like you already have but for each different number rating. If you go by those you'd definitely have more coefficients but your final numbers would be more accurate. This would mean the difference between say having two 5.8 4* and two 6.1 4*
 
Last edited:
#19
#19
Oh and you wouldn't have to use matlab you could use matrices for higher and lower ranked recruiting classes so you get roughly all the coefficients for all the numerical ratings.
 
#20
#20
wwilso29 said:
You could also narrow down the accuracy by adding in the last few years' recruiting rankings so the sample size would be greater.

Aren't the number of stars based on a numerical rating? You'd have a numerical value worth a certain number of points like you already have but for each different number rating. If you go by those you'd definitely have more coefficients but your final numbers would be more accurate. This would mean the difference between say having two 5.8 4* and two 6.1 4*
Click to expand...

1)
Yes, I could. I actually did an analysis on the class of 2008 alone and got similar numbers to what I was getting from the 2009 data. But I didn't go a step further and combine the two. Not a bad idea, and might not take too too long to do.

2)
Yes, the stars are based on a numerical rating and by just using the stars as the variables (rather than the numerical rating), I have a built-in margin of error. If I narrowed the range of categories to make more categories (as you suggested) I would increase the number of coefficients and the prediction would be better. Even then, though, I feel like it wouldn't be dead on. I have a hunch that rivals has some aspect of diminishing returns in their formula. In other words, the difference between a class of 15 3-star guys and a class of 14 3-star guys would be bigger than the difference between a class of 30 3-star guys and 29 3-star guys. Or at least that is probably how I would do it if I were designing the rating.

Maybe I'll save further tweaks for a rainy Saturday or something. It was much easier just to copy over the table from the website and have the 4 variables.
 
#21
#21
CSP said:
You should absolutely, no reservations about it be banned for mentioning MATLAB on here. I have daily (not by choice) contact with it and hate it.

I come to this board to escape the world of engineering :)
Click to expand...

I was once an engineer and am now a lawyer. While I'm fairly certain that I can no longer design a three-stage amplifier or solve a K-map, I do think that I could probably still dish out some matlab. I hated it at first, but it grew on me quite a bit. I think matlab is one of the few engineering-related tasks that I could still do.
 
#23
#23
Say just for 4* you would have this:

(# of 5.8s)*5.8*A + (# of 5.9s)*5.9*B + (# of 6.0s)*6.0*C+(# of 6.1s)6.1*D = a number

The matrix for this range would be like this:

[(# of 5.8s)*5.8 (# of 5.9s)*5.9 (# of 6.0s)*6.0 + (# of 6.1s)*6.1] = [a number]

If you have a TI-89 or similar it solves for each coefficient. By doing something like this for higher and lower ranked classes you get the coefficients for a wider range of the numerical ratings but the matrix would be much much bigger
 
#24
#24
kidbourbon said:
I was once an engineer and am now a lawyer. While I'm fairly certain that I can no longer design a three-stage amplifier or solve a K-map, I do think that I could probably still dish out some matlab. I hated it at first, but it grew on me quite a bit. I think matlab is one of the few engineering-related tasks that I could still do.
Click to expand...

Interesting, I was once pre-law, now I'm an engineer.

I'm assuming you were a gEEk?
 
New posts
You must log in or register to reply here.
Advertisement



Back
Top