A response to YOhio

[Indy Coug]

I do all my programming in COBOL.

I hope you leveraged that ability to the hilt during the Y2K craze.

[Katy Lied]

download play by play data and parse it into various fields.

I dont see a provision for onsides kick; is it defined as a function of kickoff distance or you just skip that one?

[Indy Coug]

I dont see a provision for onsides kick; is it defined as a function of kickoff distance or you just skip that one?

This isn't a finished product: it's about 90%-95% complete. I keep thinking I've finished it only to find I need new categories/refinements the more I audit and massage the data.

It would help if the NCAA had more uniformity to their play-by-play language. The biggest problem, however, is that some of the info is flat out wrong and/or plays are missing altogether. A noteable example is that the Beck to Watkins OT touchdown pass against TCU in 2005 doesn't exist as far as the play-by-play text is concerned.

An audit shows that overall the yardage totals come within 98% to 99% of actual. The purpose of this exercise is to put together a database that can be used to churn out situational statistics that simply aren't available elsewhere and I believe it's accurate enough on a macro level to do that.

[Tim]

maybe Ruby too

I discovered in college that Ruby was created by an LDS guy here in Japan:

http://en.wikipedia.org/wiki/Yukihiro_Matsumoto

[ByuMba90]

Tim, isn't that japanese guy also the Iron Chef? (nm)

(nm)

[YOhio]

Have you made any changes to your program in the past couple of years? Also, Santos silent in this thread? What's up with that? Not willing to share the secret sauce that makes the Santos Power poll?

[Jeff Lebowski]

Have you made any changes to your program in the past couple of years? Also, Santos silent in this thread? What's up with that? Not willing to share the secret sauce that makes the Santos Power poll?

I miss the Santos Power Rankings.

[pelagius]

Have you made any changes to your program in the past couple of years? Also, Santos silent in this thread? What's up with that? Not willing to share the secret sauce that makes the Santos Power poll?

Reminds me of the golden days back on cougarguard; Remember, my computer model thread:

http://cougarguard.com/forum/showthread.php?t=5335

Given the talk about computer models I wanted to explain a little bit about how most computer models actually work. To do this I am going to write down a pretty simple model that still captures (I think) the essential features of what computer models such as Massey and Sagarin do (Jay and Indy, let me know if I botch anything). In order to understand my model you do need to know what a regression is. If you have ever taken a statistics class that included regression analysis or any kind of econometrics course you should be ok.

Let's suppose you wanted to predict by how much BYU would win by if they play TCU again. A simple way to model would be the following:

MOV = r_{byu} - r_{tcu} + e

1. MOV = margin of victory
2. r_{byu} = BYU's rating (this is a power rating not an ordinal ranking)
3. r_{tcui} = TCU's rating
4. e = luck (bad ball bounce, poor call by the official, team gets sick the night before, etc)

Thus is if your estimates are r_{byu} = 25 and r_{tcu}=13.5, then your model is predicting the BYU would win the rematch by 11.5.

The trick is how do we estimate these ratings. In order to do this you have to notice that you can rewrite the above equation as the following:

MOV = r_{byu}(1) + r_{tcu}(-1} + r_{wyoming}(0) + r_{unlv}(0) + e

This may seem like a trivial and unnecessary rewriting of the equation, but it isn't because it also makes the equation a regression: MOV is the dependent variable and the 1,-1,0,0 are the independent variables. If we rewrite the equation like the following it really looks like a regression equation:

MOV = r_{byu}x1 + r_{tcu}x2 + r_{wyoming}x3 + r_{unlv}x4 + e
y =b1x1 + b2x2 + b3x3 +b4x4 + e

Now we just need enough data and we can estimate the regression and consequently the rating system. Note: I am only going to rank these 4 teams relative to each other but if you want to rank more all you do is add more terms to the regression and add data involving all the teams.

MOV  BYU   TCU   Wyo    UNLV
14     1    -1     0        0          (BYU vs TCU
48     1    0      -1       0          (BYU vs Wyoming)
45     1    0      0        -1         (BYU vs UNLV
23     0    1      -1       0          (TCU vs Wyoming)
15     0    1      0        -1         (TCU vs UNLV)
8      0    0      1       -1          (Wyoming vs UNLV)

The MOV is the dependent variable and the rest of the columns are the independent variables in the regression. You throw all this in a regression and the estimated coefficients are the power rating and we can rank teams this way.

Technical notes: (1) you have to drop one of the columns for the regression to work (linear dependence). So let's drop the UNLV column. This implicitly makes UNLV rating = 0 and all the other coefficients are measured relative to that. (2) Don't include an intercept in the regression.

Okay, let's estimate:

. reg mov byu tcu wyo, noconstant;

      Source |       SS       df       MS              Number of obs =       6
-------------+------------------------------           F(  3,     3) =   27.34
       Model |      5154.5     3  1718.16667           Prob > F      =  0.0111
    Residual |       188.5     3  62.8333333           R-squared     =  0.9647
-------------+------------------------------           Adj R-squared =  0.9294
       Total |        5343     6       890.5           Root MSE      =  7.9267

------------------------------------------------------------------------------
         mov |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         byu |      43.75   5.605057     7.81   0.004     25.91221    61.58779
         tcu |         23   5.605057     4.10   0.026     5.162206    40.83779
         wyo |       1.25   5.605057     0.22   0.838    -16.58779    19.08779
------------------------------------------------------------------------------

The regression spits out the ratings and now we can rank the teams:

1. BYU (rating = 43.75)
2. TCU (rating = 23)
3. Wyoming (rating = 1.25)
4. UNLV (rating = 0)

Note, if BYU and TCU were to play again this model (which doesn't use very much data so don't take it seriously) predicts a MOV of 43.75-25 = 20.75 points.

Note: Strength of schedule doesn't show up as a variable in the model. However, it is important and the model does take it into account. Strenth of schedule is implicit in the regression fitting process. We are estimating BYU's rating controlling for the rating of TCU, UNLV, and Wyoming.

[YOhio]

Let's back up even one more step. I turn the computer on. Type in my password and I'm looking at Windows 7. Tell me, what button do I click to start writing my computer power ranking model program.

[jay santos]

Must not have noticed this thread before. I try to avoid threads with YOHIO in the title.

I've explained my model before. It's not really an original model. I tried doing regression models and never felt satisfied with the results. Geez now I can't hardly remember how it works, it's been a while since I played with it.

I went to the Massey compilation page. They are ranked in order of which model is closest to the entire average. I believe an average of many models is better than any one model.

So I built a model with variables:
SOS, W-L, MOV, dampening effect for large wins, home-away factor

I'm not a true statistician and can't remember exactly how I did it, but I think I optimized the variables by doing a simple least squares (but I think it was probably least absolute values cuz I never liked the logic of converting variance to least squares) on the final season Massey compilation.

I tweaked it from year to year, but not much. And I might have not used the entire Massey compilation but maybe cherry picked the best 20 or so models.

At any rate, mine's not original but because it optimizes on an average of many models and an average of models predicts better than any one model, mine is THE BEST!

[NorthwestUteFan]

I miss the Santos Power Rankings.

I miss Green Acres' avatar. Hawt!

[TripletDaddy]

[YOhio]

Insomnia bothering you still?

[TripletDaddy]

Insomnia bothering you still?

that is sleeping avatar.

[YOhio]

that is sleeping avatar.

Sometimes insomnia makes people very sleepy during the day.