THE BOOK--Playing The Percentages In Baseball

Very good stuff by Jeremy as usual.  This is the type of analysis that managers need to do (or have access to) in order to figure out “matchups” rather than the ole’ index cards with, “so-and-so is 3 for 11 versus so-and-so.”

My only quibble (and it is not really part of Jeremy’s analysis) is this:

“The Book says to next look at platoon splits. Fittingly, Hall and Fielder have identical .348 wOBAs against southpaws.”

He talks (correctly) about using projection rather than career numbers, yet he quotes Hall’s and Fielder’s actual numbers against LHP without doing any regression or incorporating their numbers against RHP.  IOW, I would like to see Hall’s and Fielder’s projections against Maholm using only each player’s projected platoon ratio and overall wOBA (or wOBA against, of course).

Also, the trouble I have with the hypothesis that batters and pitchers have true matchup anomalies based on the batter’s weaknesses and the pitcher’s pitch repertoire, as Jeremy suggests is the case with these guys, is that if that is the case, we would expect to see SOME predictive value from matchup numbers involving 20 or 30 PA (even though that is a small sample), I would think.  According to Tango’s research in The Book, we don’t.  That makes me skeptical.

Bill Hall (RHH) has a career split of 30 wOBA points.  He faced 1030 LHP.

The league average split is 17 points, and for a RHH you regress 50% toward mean when PA=2200.  So, we are going to heavily regress Bill Hall.  Regression rate is 2200/(2200+1030) = 68%.  That 30 split regresses 68% toward the 17 league mean, to give us a true estimated 21 point split.

So, instead of being .346/.316, he’s probably .340/.319.  In EITHER case, the overall weighted average is .325.

(You can further regress his .325 toward the league mean, but him being so close to league average, and with over 3000 PA, well, let’s not bother.)

***

Prince Fielder (LHH) has a 60 point split.  He faced 1005 LHP.

The league average split is 27 points, and for a LHH you regress 50% toward mean when PA=1000.  So, we are going to regress Prince Fielder about halfway.  Regression rate is 1000/(1000+1005) = 50%.  That 60 split regresses 50% toward the 27 league mean, to give us a true estimated 44 point split.

So, instead of being .348/.408, he’s probably .359/.403.  In EITHER case, the overall weighted average is .390.

***

Given NO OTHER INFORMATION, we’d expect Hall to be .340 against LHP and Fielder to be .359.

Whether Hall is better than his career today, or whether he hits pitchers like Malholm better, that’s where Jeremy comes in.

MGL: as for “some” predictive value, sure there is going to be some.

But, how much could there be?  Melvin Mora has 50 PA against Andy Pettitte with say a .440 wOBA, and he’s got 5000 other PA against all other pitchers with a .340 wOBA.

Had I done the study late in The Book cycle, I would have followed the standard pattern to figure out what the regression equation should be.  Unfortunately, it was the first study I did, so I didn’t think to do that.

In any case, I’m sure we’d need to add at least 1000 Mora-average PA, to get his true talent against Pettitte.  So, 50 parts .440 and 1000 parts .340 and we get .345.

(Why at least 1000?  I figured based on what we need to do with LHH sets a minimum baseline.  I could be wrong.)

Thanks for the comments.

MGL, I tried to keep everything kosher by The Book. I don’t have my own projection system, so I thought it would be fine to quote the readily available ZiPS projection and then just say their career platoon split narrows the gap. I don’t think I came across as saying they project equally against LHPs, which they obviously don’t. I probably should have done the work that Tango does in comment 2.

As Tango says, there has to be some value to matchups. Just the 50/1000 split Tango suggests isn’t powerful. So what I’m trying to do is say maybe 50 Pettitte/25 Joe Saunders/15 Cliff Lee or something. That might add a bit more power.

"As Tango says, there has to be some value to matchups.”

It could be deminimus (essentially non-existent) or it could be fairly significant.

That is the task at hand - to determine how much predictive value there is.

According to the research in The Book, it appears to be very little - almost non-existent. Otherwise something would have shown up even in those 30 or 40 PA samples.

Now, even if overall, it is almost non-existent, that does not mean that certain matchups are not significant.  Perhaps the Maholm/Fielder/Hall are and perhaps they are not.

Using pitch f/x is definitely on the right track and should be one of the breakthroughs that teams can use to get an edge.

Right, I agree.  Focusing on performance outcomes (H, HR, BB) is not going to get you anywhere.  After all, we took the 30 most extreme cases each with 20 or whatever PA, and we got nothing there.

The focus has to be at the scouting level.

We will have reached a breakthrough if it ever happens that Angel Pagan is 2 for 19 against Barry Zito, but PITCHf/x says that Pagan FEASTS on curveball pitchers, and so the manager insists that Pagan be in the lineup against Zito.  That’s what PITCHf/x has to deliver on in terms of matchups.

Of course, just because pitch f/x tells us that batter A feasts on curveballs and pitcher B throws a lot of curveballs (not that we need pitch f/x for that), doesn’t mean that it is a good matchup.  If said pitcher throws 30% curveballs, he is likely to only throw 15% against said batter.

We can, of course, test the “good high ball hitter should do well against pitcher who throws high factballs, etc.” theory by using out of sample data.  And we don’t need pitch f/x for that.  Actually, you sort of did that in The Book, by looking at “families” of pitchers.

Basically, I would start with this:

Take the top batters against curve balls in year X and the top batters against fastballs also in year X.  Then look at how each group did in year X+1 (or X-1) against pitchers who throw a lot of fastballs or curveballs.  Compare that performance with what is expected given the overall true talent of the batters and pitchers in each group adjusted for platoon.

You can do the same thing with location (using pitch f/x data of course).  For example, do the same thing with good high ball and low ball hitters against high and low ball pitchers.

Managers will say things like, “I like (or don’t like) this matchup because so-and-so batter is a good low ball hitter and so-and-so pitcher is a sinkerballer.” Now, other than the G/F platoon matchup which we already know about, I have always been skeptical of these kinds of “matchup advantages and disadvantages.” Again, that pitcher is probably going to throw fewer sinkers against this particular batter.  In any case, we can certainly test this basic premise, as I briefly explain above.