THE BOOK--Playing The Percentages In Baseball

How does 20%-10%-70% compare to the average new AAA call-up?
That seems like an important point of comparison. I don’t have a sense whether a 70% failure rate is extremely bad or just a little worse than average.

Excellent point!  If they had someone much better in the minors, presumably he would have been called up.

I don’t have an answer for you, but it’s a great question.

Gotta love #9 also.  Dayton will never learn.

Eric writes this:

It is without question that, over a smaller sample, ERA estimators inform on more levels about a pitcher’s effectiveness than ERA itself. But over a longer period of time, and over multiple seasons, ERA tends to emerge as a more predictive source of information. In Davies case, I would argue that his 768 innings is a sufficient sample to determine that his run prevention skills are quite poor, and that his estimators don’t paint a more accurate portrait of his performance.

What is the evidence for this?  Have we studied all pitchers with a gap (in one direction or another) between their FIP and ERA (park and hopefully defense adjusted of course) after 700 IP or so and then looked at the gap subsequently (to see if the gap persists, and to what extent)?

Even that is not a fair question.  I hate when we take continuums and make them out to be either/or, black/white, digital rather than analog, etc.  (For example, “At what point does UZR stabilize?” I hate that question, unless you specifically want to define “stabilize” as exceed the 50% - or some other number - regression point.)

IOW, what sort of regression do we apply to the gap between FIP and ERA?  Is it only 10% after 768 IP?  If it is, then Eric’s statement is valid.  If it is still 50% or more (or even 30%), then I don’t like the statement....

It’s an excellent question, because it allows us to talk about systematic bias.

If you had one million PA, I would take actual (park adjusted) RA9 over (park adjusted) component RA9 (i.e., LWTS or BaseRuns or FIP, etc).  The reason should be clear, but I’ll say it anyway: component RA9 intentionally ignores things that are not completely random.  Pitching with men on base, how you handle BIP, and even baserunning things (SB, CS, PK, BK), etc.

All these things all stabilize (i.e., r=.50) at some point, say no more than 5000 PA.  If you ignore this data, and you have one million PA, then you have a systematic bias.

You don’t mind a systematic bias when you have few PA (say under 1000), because the random variation dwarfs the systematic bias.  That is, better to reduce random variation rather than worry about systematic bias.

This is why something like UZR is really good if you have less than 2 years of data, and why something like WOWY might be preferred if you have more than 10 years of data.  At some point, the systematic bias not handled by UZR (i.e., stringers marking LD outs as FB but LD hits as LD) will dwarf the random variation.

So, going back to MGL’s quote of Eric’s claim: it’s a legitimate question to ask.  And that is, when does knowing his actual runs allowed actually give us more information than his components?

My guess… just pure guess… would be at 5000 PA.

I’m slightly too young to remember Mike Scott’s career, but as a Met fan I know him as MIKE F’N SCOTT.  I had no idea his career shape was so unusual.  4.5 K/9 over 885 IP, and then all of a sudden 10 K/9 in that ridiculous ‘86 season.  I’ve heard the development of the splitter was the tipping point for him (as well as grumblings about scuffing the ball), but sheesh.

...and as a result, I wonder if the other “20%” had a similar breakthrough.

Probably an obvious point, but if you’re talking about, for example, Jim Palmer, the answer to the question “when does knowing his actual runs allowed actually give us more information than his components?” is “never”.  Unless you go back and calculate that Jim’s fielders saved him about .35 r/g or so (which is what you get when you compare Bal’s DER to the league average and multiply by BIP times runs per hit avoided).  This makes me wonder why we’re ever talking about k/bb data or FIP for HOF candidates.  I.e. we know exactly how good the guy was, so why use k/bb as a proxy?

Excellent point.  His RA9 would also have to be adjusted for quality of fielders (in addition to park).

Tango, what is the state of the art for adjusting for fielders?  I looked at team DER, but I think think there must be some overlap between team DER and park factor.  E.g. double-counting of foul territory in Oakland.

Team DER can be adjusted for park here:

http://www.tangotiger.net/BIPteam7490.html

You can adjust for fielding using UZR or TotalZone.

thanks.  Is team UZR something that’s calculated and published?

If you go to Fangraphs, it’s there.  Back to 2002 I think.

If you are adjusting for actual UZR (rather than some regressed version or some talent estimate of all the fielders), then you might as well just use league average BABIP - no need to use UZR.  The way I adjust pitcher stats to tease out the defense is to estimate team fielding talent for that year using individual fielder (UZR) projections, and then add them up and prorate according to the number of GB and FB the pitcher allows and his handedness.  The proper way is to use actual UZR for each player regressed to their talent estimate.

IOW, let’s say that behind a pitcher in one year, his fielder’s had a UZR of +.25 rpg, after adjusting for that pitchers number of BIP in the OF/IF and his handedness.  But, let’s say that the talent estimate of those fielders was only 0m using data from several years.  IOW, it looks like that may have played exceptionally well in the field that year AND they probably had easier to field balls than the UZR engine thought. So, I would NOT adjust my pitchers by .25 rpg. I might use half that, .125.  Just a guess though.  Remember, UZR does not tell us what happened, only what the very rough data and engine THINK happened.  If we want to better know what likely happened, we can look at the estimates of the fielder’s true talent for that year and adjust our raw UZR numbers.  Again, for example, if Cecil Fielder has a +10 UZR year, it is not likely that he had a great fielding year.  It is likely that he had a decent fielding year, for him, and that the data and engine contained and made lots of mistakes…

Make that Prince Fielder, although I suppose Cecil would work too…

I remember Codiroli’s baseball card, and that might be why you DON’T remember.  It was one of those ugly Topps cards where they pulled the guy to the side, he’s sweating, he puts his cap up on his head a little ways, and he looks miserable.

Or maybe that was Joe Rudi. 

http://www.checkoutmycards.com/Cards/Baseball/1984/Donruss/345/Chris_Codiroli/1890987

Some quick and dirty research:

I formed 3 groups of SPs who pitched from 1973 up. And, compared their first 5 years to the next 5 years. I then broke them down into under 500 IP, 500-1000 IP, and 1000-1500.

The groups averaged

YEARS 1-5 YEARS 6-10
IP FIP RA9 RA9
377 5.01 5.07 4.98
762 4.45 4.29 4.47
1117 4.01 3.67 3.49

The FIP (5.01) of the lowest group predicts the RA9 (4.98) better then the RA (5.07).

The FIP (4.45) of the middle group predicts the RA9 (4.47) better than the RA9 (4.29).

The FIP (4.01) of the higher group is worse then RA9 (3.67) at predicting by their RA9 (3.49).

The trend is pretty obvious. I know there is stuff missing.

So, somewhere around 1000 IP, we’re talking about RA9 might better predict future RA9 than FIP does?  That’s around 4300 PA.

Great work!  Definitely worth pursuing further.