I really wanted to get this post out last night but life sometimes intervenes. I was all set to start writing and then my wife’s computer went belly up (stupid Vista). I was up till 3 am working on it, using what my wife calls the black and white little nerd screen.
Now, I get to write it during my lunch break. In fact it’s a fun little challenge in trying to write something coherent and good in less than 45 minutes.
The last couple of weeks have been extremely productive (both professionally and hobbywise). I’ve been learning a lot and doing a lot. The first is obviously a result of my recent life changes. The second is the the fault of Andres Alvarez, his mad skills (All Powered by Nerd Numbers). He’s placed game to game Wins Produced splits at my disposal and placing this kind of data at my disposal even when I’m swamped, is like leaving cookies for the cookie monster.
So as a result, I’ve been taking these numbers out for a spin (WARNING: This is a math intensive zone feel free to go read and review the Basics if you need to). The results have been very surprising. Now my fun time right now is somewhat limited (so I’m not going to hit every bullet today and my a.a. milne post stays on the queue) but I really wanted to get this particular fun fact out there for the world.
A common discussion in this and the other WOW blogs is that the WP coefficients are off. I decided to check this by looking at the game to game correlation between between Wins Produced margin (Sum of Home team ADJP48*MP -Sum of Visiting team ADJP48*MP for each game) vs. the actual point margin for each game. I figured that if there was skew it would show up upon detailed inspection (and we could then work on correcting/refining it). I looked at all the splits for the 2009-2010 season . The results are as follows:
How about that, it lines up perfectly. I get the fun equation, Point Margin =0.0377 + 15.5 WP margin (I then round to get a whole number). This has an Rsq of 99.8%. Alternitavely, WP margin= (Point Margin -0.0377)/ 15.5.
The absolute average absolute error is .32 points and a maximum absolute error of 2 points (for every game after rounding).
Out of 1230 games I got:
A 2 point differential 6 times (.48%)
A 1 point 392 times (31.8%)
and spot on 67.6% of the time
We really need to work on improving those coefficients!