Posted by kmasonbx1 on 3/2/2014 7:37:00 PM (view original):

Posted by stinenavy on 2/28/2014 4:03:00 PM (view original):
You don't need the drafted page. If your prestige rises as a result of player getting drafted, the history page prestige doesn't actually change. 

But you still have to factor the impact because let's say a school goes from a B+ to A- as a result of draft picks (which won't be reflected on the history page) and the next season has a season that if it were not for already being an A- would have kept the school at a B+, so the draft impacted the prestige bump that will show up in the history page. 

I feel like that example came off much more confusing then it was in my head. Lol.

I tried to make sense of it 

I'll just respond to what I think your statement was about:

I believe that when prestige gets recalculated after the end of a season, it does not take into account the drafted players any more. So a drafted player is only good for prestige calculation for that one recruiting session.

thats just worlds 8,9,10

the first seven are   1945 to 8832

the ealier worlds have messy numbering patterns such that not all the D3s and D2s and D1s are sorted  together...but not sure that matters to you.
the worlds themselves are sorted together.

i have a table with team name and location in top row and then corresponding team IDs listed for the 10 worlds in the 10 rows beneath.

im just recalling how i got past the " no scrapable team name" problem.... i dont remember it exactly, but somehow the macro found thier opponent in game one,  then looked up thier opponent's schedule and found thier opponent in game one and that returned the name of the school in question.

as i have said... i love digging through the data and figuring things out within HD... but im not clear what the point is here?  

what would we learn if we could figure out the exact equation used for prestige calculation?  doesnt seem like a big deal to me.

i ve got a whole list of things that i think would be more interesting and/or  instructive to dig into.
 but maybe i am missing something? 

I thought about doing exactly what you're talking about, going through the schedule and using the first opponent's schedule to get the team name. Certainly doable but I was just hoping someone might already have a list. As far as getting the range of team IDs, I could actually just iterate starting at 1 and just ignore ones that don't return an actual team, but that would be incredibly inefficient.

I would be curious to discuss some of your other ideas in regards to data mining. I've got a few things I've done in the past but I'm always interested in other new ideas.

Posted by oldave on 3/2/2014 9:55:00 PM (view original):
as i have said... i love digging through the data and figuring things out within HD... but im not clear what the point is here?  

what would we learn if we could figure out the exact equation used for prestige calculation?  doesnt seem like a big deal to me.

i ve got a whole list of things that i think would be more interesting and/or  instructive to dig into.
 but maybe i am missing something? 

I'll admit that half of my motivation behind the project is its completability - once you pull the data, all the variables are there to figure it out. Even the coach change variable can be interpolated from the history datasets.

The first application I see is for attributing more specific prestige grades to each school (i.e. "high A" vs "low A"), which can help make more sense of recruiting effort comparisons. If prestige were known down to the number, instead of just to the grade, that's a huge tool in analyzing recruiting battles and such.

Ok jet... that makes sense... expecially the idea of differentiating highs and lows of same grade..... especially at A+.


Sidebar.... I can't recall if it was always this way but it seems to me there are waaaaaay too many A+ I wish there were only 4 to 6 ish per division. Seems like 20+ these days. Probably would be impossible to put that genie back in the jar at this point... but it would be nice if they created an A++ or since the extra character might screw things up A!

Maybe just call it Z

I have the prestiges, wins, losses, conference champ status, ct champ status, pi round, and nt round of every team in HD in a database, and can export as a spreadsheet and post a link if anyone is interested.  I did this with D3 only because I know that baseline prestige should be irrelevant.  I use the finishing prestige of the last 8 seasons of each program as the thing to predict (normalized to a numerical grade, so e.g. a C is 2.0), and the above variables as the predictors.  This is a simple linear regression and I know I can do better, but this still accounts for 94.3% of the variance (adjusted R-squared, which accounts for over-fitting).  

RPI is actually RPI rank divided by 300, and the coefficient is negative because a smaller ranking number is better.  The CT champ coefficient is negative because the model is probably not very good, and possibly because winning the CT means absolutely nothing, above and beyond the other achievements of a team.  I made separate indicator variables for each of the possible values of NT wins and PI wins, because I didn't want to make a guess about the functional form that the real model might take, e.g. NT wins squared or something like that.  Lastly, you can see the coefficients decrease with the number of seasons back, which makes sense.  

I also used the Lasso method for penalizing over-fitting, and finally rounded coefficients to the nearest tenths place.  Here is how you interpret the following.  A team which wins 2 NT games in the most recent season will, on average, have a prestige that is 2*0.3 + 2^2*(-0.04) = 0.44 better (about half a letter grade) than a team that makes the NT but wins 0 NT games, but is otherwise the same in every way.  I have bolded the two lines I am using to make that calculation.  This will be obscure to many of you, and even those of you that do statistical modeling will probably have some criticism, but that's fine I'm happy to revise the model based on your ideas.  

Lastly, sometimes you get a negative number where you probably shouldn't, e.g. the very first one (CT Champ -0.08).  Taken at face value, this would imply that, all things being equal, it is worse to win the CT than to not.  I'm sure this is not true, so you should just interpret this as noise and assume it probably has no effect.  

1 seasons ago: CT Champ -0.08

1 seasons ago: Conf Champ 0.1

1 seasons ago: Wins 0.44

1 seasons ago: NT 0.63

1 seasons ago: NT wins 0.3

1 seasons ago: NT wins^2 -0.04

1 seasons ago: PI 0.28

1 seasons ago: PI wins 0.03

1 seasons ago: PI wins^2 0.0

2 seasons ago: RPI -0.01

2 seasons ago: CT Champ -0.05

2 seasons ago: Conf Champ 0.03

2 seasons ago: Wins 0.15

2 seasons ago: NT 0.3

2 seasons ago: NT wins 0.18

2 seasons ago: NT wins^2 -0.02

2 seasons ago: PI 0.09

2 seasons ago: PI wins 0.03

2 seasons ago: PI wins^2 0.0

3 seasons ago: RPI -0.01

3 seasons ago: CT Champ -0.0

3 seasons ago: Conf Champ 0.01

3 seasons ago: Wins 0.07

3 seasons ago: NT 0.12

3 seasons ago: NT wins 0.09

3 seasons ago: NT wins^2 -0.01

3 seasons ago: PI 0.03

3 seasons ago: PI wins 0.0

3 seasons ago: PI wins^2 0.0

4 seasons ago: RPI 0.0

4 seasons ago: CT Champ 0.01

4 seasons ago: Conf Champ 0.0

4 seasons ago: Wins 0.02

4 seasons ago: NT 0.06

4 seasons ago: NT wins 0.06

4 seasons ago: NT wins^2 -0.01

4 seasons ago: PI 0.02

4 seasons ago: PI wins -0.0

4 seasons ago: PI wins^2 0.0

5 seasons ago: RPI 0.01

5 seasons ago: CT Champ -0.0

5 seasons ago: Conf Champ -0.0

5 seasons ago: Wins 0.0

5 seasons ago: NT 0.06

5 seasons ago: NT wins 0.05

5 seasons ago: NT wins^2 -0.01

5 seasons ago: PI 0.03

5 seasons ago: PI wins 0.0

5 seasons ago: PI wins^2 0.0

6 seasons ago: RPI 0.0

6 seasons ago: CT Champ 0.0

6 seasons ago: Conf Champ -0.0

6 seasons ago: Wins 0.0

6 seasons ago: NT 0.02

6 seasons ago: NT wins 0.03

6 seasons ago: NT wins^2 -0.01

6 seasons ago: PI 0.01

6 seasons ago: PI wins 0.0

6 seasons ago: PI wins^2 -0.0

7 seasons ago: RPI 0.0

7 seasons ago: CT Champ 0.0

7 seasons ago: Conf Champ -0.0

7 seasons ago: Wins 0.01

7 seasons ago: NT 0.01

7 seasons ago: NT wins 0.02

7 seasons ago: NT wins^2 -0.0

7 seasons ago: PI -0.0

7 seasons ago: PI wins 0.01

7 seasons ago: PI wins^2 -0.0

8 seasons ago: RPI 0.0

8 seasons ago: CT Champ 0.01

8 seasons ago: Conf Champ 0.0

8 seasons ago: Wins 0.0

8 seasons ago: NT 0.0

8 seasons ago: NT wins 0.02

8 seasons ago: NT wins^2 -0.0

8 seasons ago: PI 0.0

8 seasons ago: PI wins 0.01

8 seasons ago: PI wins^2 -0.0

9 seasons ago: RPI 0.01

9 seasons ago: CT Champ 0.01

9 seasons ago: Conf Champ 0.0

9 seasons ago: Wins 0.0

9 seasons ago: NT 0.01

9 seasons ago: NT wins 0.01

9 seasons ago: NT wins^2 -0.0

9 seasons ago: PI 0.0

9 seasons ago: PI wins 0.0

9 seasons ago: PI wins^2 0.0

10 seasons ago: RPI 0.0

10 seasons ago: CT Champ 0.0

10 seasons ago: Conf Champ 0.0

10 seasons ago: Wins -0.0

10 seasons ago: NT 0.01

10 seasons ago: NT wins 0.01

10 seasons ago: NT wins^2 -0.0

10 seasons ago: PI 0.0

10 seasons ago: PI wins 0.0

10 seasons ago: PI wins^2 -0.0

i thought it was only based on the last 4 seasons?
but i could be wrong.


very cool stuff, rg.  

nice, 94.3 is pretty good...i'm not sure if the lower levels work like D1 in that that not all prestige ranges have equal magnitude. if it were true, that could help explain some of your undue variation.