Quote:
Originally Posted by ASURay
It's not in a very legible format at the moment. I kind of put it together haphazardly. My current formula for predicted_FIP is:
FIP_p = 9.468 - .022*STUFF - .033*MOV - .027*CON
The intercept is biased since I used data mostly from Bronze/Silver/Gold leagues so the resulting FIP estimates are likely too low. Right now I just use the model for comparing SP to one another rather than for predicting their actual performance. The estimates might be pretty good if you're in a lower league, though. I'll update things after this week maybe. Note that this a very simple regression model. Nothing fancy. It assumes the effect of the ratings is linear (e.g., the difference between a 50 and 60 rating for MOV is the same as the difference between a 90 and 100 rating). The R^2 for the model is ~.62, so good but not great.
|
I just went though and fit the model using data from every SP in my current Perfect league that has >=30 career GS at the Perfect level. Unfortunately, this only gave me a sample size of N=89. The new formula is:
FIP_p = 10.18 - .016*STUFF - .043*MOV - .024*CON
Interesting that CON became more effective while Stuff lost effectiveness as compared to fitting the model using data from lower leagues. The estimates almost seem too high now. THe 91 OVR Vida Blue card, for example has a predicted FIP of 4.80. Anyways, the Top 10 using this new model is:
1) 100 Pedro
2) 93 Maddux
3) 96 Brown
4) 97 Hubbell
5) 96 Greinke
6) 94 Maddux
7) 96 Young
8) 97 Pedro
9) 100 Clemens
10) 97 Newhouser