THE BOOK--Playing The Percentages In Baseball

Great work.

In all fairness to BP, Colin was talking about batters and not pitchers when he said that they used almost equal weights for prior years. He said nothing about pitchers, I don’t think.

Surely they don’t think that pitchers should have their prior years get equal weight.

Hopefully, Rob will do the same thing with batters. I’m not sure why he chose pitchers when BP only discussed batters, but it is great work nonetheless.

Also, I’m not sure, but I think that if pitchers in the “high” group had higher FIP’s in year x-4, they would also tend to have higher FIP’s in year x-3 (and x-2, etc.). I think.

So, even though both groups had a similar 3 year FIP, their 3-year Marcel or other projection would have a smaller gap between the groups.

IOW, Rob got this:

Group FIP IP K/9 BB/9 HR/9
Low 3.46 162 9.0 3.4 0.8
High 3.52 160 8.3 3.5 0.8

However, if the high group had higher FIP’s in the earlier years, a forecast based on only 3 prior years might be something like this:

Low 3.5
High 3.48

Since he found that the higher group was .19 more in FIP in year X, it would be actually .21 more than expected rather than .13, giving even more weight to year x-4.

Again, I am not sure if a higher FIP in year x-4 means a higher FIP in year x-3 (and x-2 and x-1), but if Rob is reading this, perhaps he can tell us the average FIP in each year for each group…

I’d bet that Colin has something similar for pitchers.  His top ranked starter was Tim Lincecum, and Kershaw was much lower than Marcel had him.  That’s what happens when you count Lincecum’s two Cys so high, and you include the very young Kershaw’s high BB rates.

Nice work.  I think the component breakdown is interesting.  It seems that x-4 HR and BB rates have a small amount of predictive power for year X, in the expected direction.  However, x-4 K rate adds nothing (as we would have guessed), and may even be a ‘reverse predictor’—those with a low x-4 K rate actually improve in year X.  Because the pitcher has improved his K rate in recent years, his expected K rate in year X may actually be HIGHER than that of the pitcher who was worse in x-4.  Has anyone (Brian?) studied whether the trajectory of K rate adds any predictive power?  (Alternatively, this may just be a function of age.)

MGL/1: I think the Kershaw example got stuck in my head, so I was thinking only about pitchers. (I also didn’t read Colin’s article fully enough, clearly.)

As for the FIP breakdowns by year:

Group   t-4    t-3    t-2    t-1      t
Low    2.75   3.40   3.48   3.54   3.47
High   4.55   3.38   3.36   3.67   3.66

I guess I was wrong about the fact that we might see a higher t-3 FIP for the high group. 

BTW, you said that the average FIP (for 3 years) for the low group was 3.45 and the high group was 3.51.  I get 3.47 for both groups just taking the simple average of all 3 years. Can different number of IP in each year make that much difference?

Also, you base line (or expected FIP) needs to be a weighted average of the 3 years, not an un-weighted average (assuming that a weighted average is correct for pitchers, which I doubt that Colin would argue with).  The reason is this:

What happens if we happen to get these numbers (just by chance) for the two groups:

Group t-4 t-3 t-2 t-1 t
Low 2.75 3.43 3.51 3.47 3.47
High 4.55 3.18 3.56 3.67 3.53

Even though both groups have the same simple average FIP in years t-1 to t-3, the higher FIP in year t in the “high” group would be explainable by the higher FIP in year t-1 and t-2, even without including year t-4…

Given the number of pitchers he got, I think it was a foregone conclusion that he was not going to get that kind of result.

And he didn’t get that kind of result.

So, I’m not sure why we need to incorporate more complexity into something that gives us such starkingly clear results.

I brought that up because I was concerned that the high group was 13 point higher in year t-1 and thus their expected FIP in year t without year t-4 might be higher than a simple average of the 3 years would suggest (and the simple average numbers do not seem to jive with the individual year numbers as well). Although it is the opposite in year t-2 which somewhat mitigates the weighting versus simple average problem…