THE BOOK--Playing The Percentages In Baseball

Wow, that would be awesome (the previous-pitch thing).

I suppose you’d also want to split it by count, right?  Otherwise, you’ll find a much higher % of fastball being followed by fastball, overall, due to it being the overwhelming occurrence at 2-0, 3-0 and such.

I will guess that would be one way to figure if he meant to throw inside.  How often would a pitcher throw two balls consecutively inside, especially if it’s a FB/CH combination?

I don’t think that is going to have anything much to do with it at all.  For a lot of reasons.  For one thing, pitchers are supposed to randomize their pitches/locations, so that they should be just as likely to throw inside after an inside pitch as an outside pitch.  If a pitcher rarely threw to the same location twice in a row, what would be the point if the batter knew that (which he should it it were true)?  That would simply be telegraphing your pitches.  For example, the idea that when ahead in the count, after a high fastball, a pitcher very often (90% or so) comes back with the breaking pitch away is either nonsense (not true) or bad strategy. If that were true, the pitcher might as well say, “Hey, I am going to throw you a breaking pitch away.” Why would he want to do that?  He wouldn’t.

I am not suggesting a “r = -1.0” correlation, such that an up-and-in is followed by a low-and-away, and neither am I suggesting a “r = +1.0” correlation, such that an inside is followed by an inside.

However, I am certainly suggesting that r is not equal to 0.0 correlation, that certainly there is some relationship from pitch-to-pitch (even after accounting for the count, base/out situation, handedness of batter, and quality/family of batter).

Since Ted is clearly an outlier here, I was suggesting to see how different he was.  What normally precedes a CH that was thrown inside?  Is it really random (after accounting for the noted biases above)?  I doubt it.  So, I’d like to see how does Lilly compares to that.

MGL, would it really surprise you about a pitcher having tendencies like that?  Remember the USS Mariner stuff about Felix “establishing” his fastball?  Essentially, something like 85% of his pitches in the first couple innings were fastballs (I believe).  Pitchers and catchers are not economists.  If a professional poker player can have certain tendencies that can be exploited, I’d certainly bet some of those pitchers and catchers have them too.

If someone has a few hours to kill and an mlb.tv subscription, it shouldn’t be hard to pull up a few of Lilly’s starts and record where the catcher’s mitt is for Lilly’s changeups.  That should be a pretty good proxy for where Lilly was trying to throw.

I’d also be interested to see some cumulative numbers on the success rate of inside changeup vs outside changeup for all pitchers. What percent result in strikes, balls, hits, hrs, etc?

Mike, first of all, randomizing pitches does not mean that everything is 50/50.  It rarely is.  But whatever the mean is (50/50, 80/20, 95/5), pitch selection should be random around that mean.  For example, let’s say that on a 2-2 count in a certain situation against a certain batter, the correct mix of fastball/offspeed is 50/50.  It should be 50/50 no matter what the last pitch was, ball or strike, inside or out, etc.  That is not exactly true, of course, and shouldn’t be.  For example, if I just threw my curve ball to get the count to 2-2 and the batter swung and missed and looked bad, it might be that he is particularly vulnerable to my curve ball for whatever reason. But really all I am doing is changing the mean on the fly.  It might have been 50/50, but after I saw his bad swing, it is now 80/20 or whatever.

But I agree, I would be surprised if almost all pitchers didn’t have “tendencies” that could be exploited by batters.  That is one (very important) thing that dictates how successful pitchers are.

Pitchers can be unsuccessful for two (among other of course) reasons:  One, they simply don’t have enough of a balance in their pitches (ideally, you want to be able to throw any pitch at any time), or two, they have some decent balance, but they are too predictable.

This whole thing is an area that there has been absolutely NO research into thus far.  And one that I find fascinating.

Ask and ye shall receive:
http://www.hardballtimes.com/main/article/doubling-up/

Not only do they try to not double-up, they are right in doing so as the second pitch is now less effective. The doubling-up of changeups is the most effective.

I wrote this on Ballhype:

This is a nice start in terms of looking at sequencing, but it is a great example of how complicated such an analysis is.  There are ALL kinds of problems with this analysis and the “conclusions.” I could not even begin to list them.  A few:

1) Using such small sections to define a “double up” is a bad methodology.

2) I agree with Ike above.  Since there are 25 grids and a bunch of possible pitches, the 10.7% makes no sense.

3) There are all kinds of selection biases, some of which Josh mentioned (the 3-0 and 3-1 count).  This is going to screw up ALL the numbers.  For example, the reason why “doubling up” produces a pitch that is not that great (.99 in n100) may be because the sample is heavily weighted by 3-0 and 3-1 fastballs, as Josh mentions.  That does NOT mean that doubling up is a “bad” idea in general.

Again, while this is a nice start and I am thrilled that someone is finally starting to look at sequencing, I caution Josh and others to be really careful in their methodologies, and especially in their conclusions.  This is a really tricky area.  To say that because the n100 is .99 on a double up pitch, that it is correct to use it infrequently, is just not a sound inference for the reasons stated above (and other reasons - there are massive game theory considerations).

Also, since you are using such small grid sections to define a doube up, is 7.5% really that infrquent?  Even if a pitcher TRIES to throw the same pitch in the same location 100% of the time, my guess is that it will only show up in the same section maybe 30% of the time (maybe less).  So that 7.5% may really be a pitcher trying to throw the same pitch in the same location 20 or 25% of the time.

We really want to break this thing down by count.  Again, the 3-0 and 3-1 counts is really going to screw things up.  Even breaking it up by game situation (bases/outs/score) would be helpful. For example, with a base open in a close game, especially with 2 outs, the pitcher is much more likely to keep throwing balls out of the zone and/or off-speed pitches no matter what the count is .

BTW, does pitch100 include balls and strikes (and fouls), or only balls in play?  I assume the former.  If it does (include balls and strikes) is the value of the ball and strike properly weighted for the count?

I replied at ballhype:

The 10.7% makes perfect sense to me.  Suppose that a pitcher throws a fastball either low and away, hi and away, low and in, and hi and in… that’s 4 zones.  But, the low and away happens 70% of the time, while the other 3 zones occurs 10% of the time.

GIVEN that he has thrown the ball low and away, then we expect the “double up” to occur 70% of the time.  GIVEN that he has thrown the ball low and in, then we expect the double up to occur 10% of the time.  And so on.  In this example, the double up occurs 52% of the time.

If on the other hand each zone had 25% of the pitches, then the double-up rate would be .25*.25 times 4 which is 25%.

***

In Josh’s case, he may be treating as a given that you have thrown 2 fastballs, and therefore, the doubling-up presumes that.  Otherwise, you would never see a “double-up” of a changeup.  If CH are thrown say 10% of the time, then just having back-to-back changeups would be 1% of the time.  To have two changeups thrown at the same spot back-to-back would likely happen, say, never.

Unless you are Trevor Hoffman, or whoever, that throws it all the time to begin with.

Tango, your numbers make sense if there are 4 zones.  But Josh is using 25 zones and I think a double up only occurs if the same pitch is thrown to the same section (of those 25 sections).  Unless most of those sections are rarely thrown to, there is no way you can get 10.7%.  Unless I am mis-reading what he considers a “double up.”

In any case, he needs to explain how he computed the 10.7% so we don’t keep arguing about it.

The only real way to look at whether one pitch suggests the same pitch or another pitch, is to control for the count, batter, and game situation.  If you don’t do that, your data is worthless. 

What we really want to know is whether with Jeff Francouer at the plate (or anyone), in the first inning with no one on and 2 outs (and him batting third, pity the Braves), will a pitcher (all pitchers on the average) throw a second pitch fastball to a certain location the same percentage of time regardless of what the first pitch was, assuming the same count (0-1 or 1-0 of course)?

We can’t answer that with aggregate data.  For example, it might look like a fastball means another fastball is likely coming, but they may only be because our samples have a lot of 3-0 fastballs (probably followed by a lot of 3-1 fastballs).  Or that we throw more fastballs followed by fastballs to hitters who are vulnerable to the fastball.  Etc.