Improving BABIP Projection by Batted Ball Types

Normal
0

false
false
false

MicrosoftInternetExplorer4

/* Style Definitions */
table.MsoNormalTable
{mso-style-name:”Table Normal”;
mso-tstyle-rowband-size:0;
mso-tstyle-colband-size:0;
mso-style-noshow:yes;
mso-style-parent:”";
mso-padding-alt:0in 5.4pt 0in 5.4pt;
mso-para-margin:0in;
mso-para-margin-bottom:.0001pt;
mso-pagination:widow-orphan;
font-size:10.0pt;
font-family:”Times New Roman”;
mso-ansi-language:#0400;
mso-fareast-language:#0400;
mso-bidi-language:#0400;}

In my last article for StatSpeak, I tested the major projections systems’ abilities to accurately project a variety of statistics other than general production level measures such as OPS or wOBA.  One of the statistics that I tested thoroughly was Batting Average on Balls In Play (BABIP).  The projection systems were much worse at projecting BABIP than other statistics, like homeruns, walks, and strikeouts.  Presumably, that is because hitters vary far more in their abilities to achieve/avoid these outcomes, and BABIP is based largely on luck.  However, hitters do vary by their abilities to hit safely on balls in play, and seeing as 70% of plate appearances result in a ball in play (and not a homerun), it is important to project these.  As far as I know, none of the major projection systems use BABIP by batted ball type.  It is my belief that this would greatly improve projection overall.

In this article, I will continue my study of hitters’ BABIP, using more data than I previously had
access to.  Not only are the results
different, but I was able to test more variables than before since I had more
observations.  In my last
article, I used data on hitters who had 300 PA from 2005-2008 to develop my
data.  In this article, I was able to
incorporate detailed data on hitters who had 300 PA from 2003 to 2008, which
allowed me to study far more observations. 
In trying to consider how to use three years of previous data to predict
a hitters’ BABIP, 2005-2008 only allowed me to look at 121 hitters.  However, adding in the four year ranges of
2003-2006 and 2004-2007 allowed me to use 381 hitters (or more accurately,
number of times a hitter got 300 PA in four consecutive seasons, as some
hitters were included more than once).  I
was also able to extend my study of predicting BABIP on groundballs, line
drives, and flyballs by using all of this previous data too. 

 

The regression lines changed in several ways.  For one thing, line drive percentage came out
significant in projecting BABIP again. 
It was surprising that it was so statistically insignificant while
looking at the 2005-2008 data, but now it does seem like this is a persistent
enough skill that it can be used in BABIP projection. 

 

Another particularly nice thing about using more data is
that you can minimize the effect of outliers. 
For instance, the coefficient on the natural log of contact rate (the
percentage of pitches that a hitter swings and either puts in play or fouls
off) was cut in half, and the reason is fascinating.  Coefficients in multiple regression equations
will tell you how to adjust the expected dependent variable (in this case,
BABIP) given what all the other variables were. 
In my regressions, I incorporated groundball percentage, BABIP on
groundballs, flyballs, and correlates with BABIP on line drives.  Additionally, I used the variable of natural
log of contact rate to help say which direction
BABIP should be expected to go given a hitter’s ability to make contact, since
the other coefficients served to regress BABIP back to the mean as far as
historical tendencies indicated it should. 

 

However, the strong negative coefficient on natural log of
contact rate indicated that hitters with poor contact skills would see their
BABIP fall more quickly.  That
coefficient remains significant and negative. 
However, Ryan Howard screwed with my data when I only looked at 121
observations.  His BABIP fell
dramatically after being strong in 2005-2007 (.358, .363, and .336), down to
only .289 in 2008.  The regression wants
to avoid the nasty error term that Howard would have left and the way that
Howard differed from the rest of the league is how low his contact rate is and
how high his homerun rate is.  Since the
natural log formulation served to expand the difference between his contact
rate and the rest of the league and to contract the difference between his
homerun rate and the rest of the league, the model used contact rate as the
cause of low BABIPs.  I realized this
error when I began doing BABIP projection for many other players and found that
guys like Adam Dunn and Mark Reynolds with poor contact rates would have BABIPs
around .240 or lower by my equation, which clearly did not sound right.

 

The reason that Howard’s BABIP fell so much in 2008 is not
that he is a poor contact hitter.  Being
a Phillies fan, I know from observation that defenses became better at shifting
against him.  While teams began to shift
against him towards the end of 2006 and throughout 2007, they had not yet determined
where exactly to position their players. 
They began playing the second baseman even deeper into right field and
closer to the line compared with other infield shifts, and this depressed his
BABIP a lot.  Initially, I only used
contact rate in my equation at all because it is correlated with a high BABIP
on groundballs, but that is probably because hitters who do not center the bat
on the ball often weakly tap balls off the bottom of the bat.  That does not describe Howard, who rips
groundballs into the same predictable locations.  As a result of this and a mixture of bad
luck, Howard saw his BABIP fall in 2008, and the model tried to correct for
that by crediting his low contact rate as the cause.  This is a pretty obvious warning to me to use
more data for regression equations. 
After seeing that the coefficient was much smaller when I removed Howard
from the equation, I thought it was smart to test how the model did with more
data. 

Normal
0

false
false
false

MicrosoftInternetExplorer4

st1\:*{behavior:url(#ieooui) }

/* Style Definitions */
table.MsoNormalTable
{mso-style-name:”Table Normal”;
mso-tstyle-rowband-size:0;
mso-tstyle-colband-size:0;
mso-style-noshow:yes;
mso-style-parent:”";
mso-padding-alt:0in 5.4pt 0in 5.4pt;
mso-para-margin:0in;
mso-para-margin-bottom:.0001pt;
mso-pagination:widow-orphan;
font-size:10.0pt;
font-family:”Times New Roman”;
mso-ansi-language:#0400;
mso-fareast-language:#0400;
mso-bidi-language:#0400;}
table.MsoTableGrid
{mso-style-name:”Table Grid”;
mso-tstyle-rowband-size:0;
mso-tstyle-colband-size:0;
border:solid windowtext 1.0pt;
mso-border-alt:solid windowtext .5pt;
mso-padding-alt:0in 5.4pt 0in 5.4pt;
mso-border-insideh:.5pt solid windowtext;
mso-border-insidev:.5pt solid windowtext;
mso-para-margin:0in;
mso-para-margin-bottom:.0001pt;
mso-pagination:widow-orphan;
font-size:10.0pt;
font-family:”Times New Roman”;
mso-ansi-language:#0400;
mso-fareast-language:#0400;
mso-bidi-language:#0400;}

HOW DOES THE PREVIOUS
MODEL HOLD UP?

 

The table below describes the coefficients for the
regression equation for BABIP using the same independent variables as I did in
my previous article.  Later on, I will
improve this equation, and also provide equations for BABIP by batted ball
types, and projecting BABIP using less than three years of data.

 

The values in the table include the coefficients, with the
standard deviations in parentheses. 
Variables with p-values less than 1% (strongly statistically
significant) are bolded and italicized, variables with p-values between 1% and
5% (statistically significant) are bolded,
and variables with p-values between 5% and 10% (weakly statistically
significant) are italicized.  The first column is for the old data set and
the second column is for the new larger data set.  Note that the variable is the non-weighted
average of the previous three years of data for each of these variables.

 

 

Variable	Small Dataset: Coefficient (Standard Deviation)	Large Dataset: Coefficient (Standard Deviation)
GB%	0.934 (0.333)	0.319 (0.182)
Groundball BABIP (GB-BABIP)	1.48 (0.577)	0.515 (0.311)
GB-HIT% (GB%*GB-BABIP)	-2.89 (1.32)	-0.596 (0.708)
Infield flies/flyball	-0.362 (0.073)	-0.336 (0.418)
OFHIT% (FB%*FB-BABIP)	0.717 (.264)	0.243 (0.0135)
LN(HR/AB)	0.0137 (.00499)	0.0119 (0.00332)
LN(CONTACT%)	0.0161 (.0449)	0.00889 (0.0277)
_CONSTANT	-0.0808 (0.146)	0.178 (0.0802)

 

Look how different these coefficients are!  It seems that the interaction term between
groundball percentage and groundball BABIP is no longer significant at all, and
we will see that I remove it from my later equation.  Leaving it in the equation actually messes
with the coefficients on the two variables themselves.  The interaction term still did come out
negative, but not nearly as much as before. 
Infield flyball percentage remained pretty similar, as did natural log
of homerun rate.  The percentage of balls
contacted that were flyballs that fell in the outfield is only weakly
statistically significant now, and it turns out that it is no longer as good a
measure as before, now less than half its size. 
As I mentioned, the natural log of contact rate remains strongly
statistically significant, but not nearly as much as before.

 

The R-squared fell from a preposterously high .39 last time
down to .28 this time (meaning that actual BABIP correlated with projected
BABIP by 63% last time and only 53% this time). 
That is certainly still respectable, but it seems that the high
correlation before might have come from fitting an equation to match up well
with only 121 observations.

 

Realizing that this equation looked nowhere near as strong
as it did before, I set out to construct an improved equation.  Last time that I began studying BABIP, I
started with looking at team level data to consider the persistence of BABIP by
batted ball type at a team level, and then moved on to work on projecting BABIP
on groundballs, flyballs, and line drives here.  This provided some insight to look at BABIP
overall as a function of some of the regressors used to project BABIP for
batted ball types.  I decided to redo
this analysis with the larger quantity of data.

 

FLYBALLS

 

I began with flyballs. 
Hitters who do well on flyballs do so by avoiding infield flies
primarily.  On top of that, they also
tend to hit their flyballs over a larger area of the field.  This is because they spread the ball around
well from left to right and consistently hit the ball deep enough that they hit
the ball all over the outfield.  In my
previous study using only 2005-2008 data, I only found infield fly rate to be
statistically significant, but this time, I found that outfield flyball BABIP
was strongly statistically significant as well. 
Additionally, switch hitters did incredibly on flyball BABIP.  As before, the variables referred to below
are the non-weighted average of the previous years of data for this
variable.  The regression for projecting
flyball BABIP comes out looking like this:

 

Variable	Coefficient (Standard Deviation)
Outfield flyball BABIP	0.196 (.0629)
Infield flies/flyball	-0.331 (0.0554)
Switch hitter (dummy variable)	0.0136 (.00474)
_Constant	0.130 (0.014)

 

Each of these three variables came out strongly
statistically significant.  It is
surprising that switch hitters improve their BABIP on flyballs by 14 points of
batting average compared to other hitters with similar histories of flyball
BABIP.  The current theory I have about
why this is true is similar to the theory I had in my last article about why
hitters with high groundball percentages tend to see their BABIP on groundballs
fall further over time.  Scouts can get
less information on where hitters are likely to hit the ball if they have hit
fewer balls hit in general.  Switch
hitters have not hit as many flyballs from each side of the plate, and so
positioning fielders is a more difficult task. 
I do not know if this is why, but it could explain some of the effect.

 

GROUNDBALLS

 

Last time I tested groundball BABIP (GB-BABIP), I found that
it could be predicted well using historical GB-BABIP, infield hit rate, and
contact rate.  This time, I was able to
get many other significant regressors as I had more data to work with.  Historical GB-BABIP again explained a large
portion of variance in GB-BABIP in subsequent years.  I refined my infield hit rate statistic as I
have in other research by summing together the infield hits and the number of
times a hitter reached on an error on groundball as this eliminates some of the
noisy measurement error when official scorers decide whether to score something
an infield hit or an error.  I also
included the number of triples per at-bat as a proxy for speed, which came out
quite significant as well. 

 

Faster hitters reach base on groundballs for several
reasons.  One is that they can beat out
infield hits, but another is that fielders must play in to give themselves a
better chance of throwing the runner out in time, so they can squeeze more
balls in the hole and cause more errors.

 

Additionally, the contact rate on pitches in the strike zone
served to be a better correlate with GB-BABIP (though now became not quite
statistically significant), as contact with pitches outside the strike zone
often causes groundballs to be weakly hit. 
I also introduced the rate a hitter swings at pitches in and out of the
strike zone, which have positive and negative effects, respectively.  The coefficient on swinging at pitches out of
the strike zone is not statistically significant, but I include it since its correlation
with swinging at strikes is high enough that it leaves that the effect of
swinging at more pitches in the strike zone statistically insignificant
otherwise.  The results yielded an
R-squared of .22, which means that 47% of the variance in groundballs can be
explained away by these variables.

 

Here are the regression coefficients with standard
deviations.

 

Variable	Coefficient (Standard Deviation)
GB%	0.0739 (0.0357)
GB-BABIP	0.420 (0.0809)
Reach safely/Infield GB	0.194 (0.0936)
Z-Contact%	0.0751 (0.0509)
Triples/At-bat	0.846 (0.471)
Z-Swing%	0.0801 (0.0441)
O-Swing%	-0.0546 (0.0450)
_Constant	-0.0240 (0.0616)

 

It seems that there is an ability to hit balls through the
hole, but a larger component seems to be ability to reach safely on balls hit
in the infield (which is pretty much based on speed).  Good strike zone judgment seems to avoid
chopping balls into the ground.  Hitters
who hit more groundballs actually are more likely to keep up their groundball
BABIP, surprisingly enough.

 

LINE DRIVES

 

As I have mentioned before, BABIP on line drives is mostly
related to luck.  There is clearly a
positive correlation with BABIP on line drives and power.  This is especially true for ability to hit
line drive doubles.  Hitters who hit a
lot of line drive doubles are more likely to hit a lot of homeruns than other
hitters with similar homerun tendencies historically, but this effect also
works the other way.  Hitters who hit a
lot of homeruns have far higher BABIP on line drives.  In my previous research, hitters who had high
BABIPs on line drives in previous seasons were not any more likely than other hitters with similar
homerun rates in previous seasons to have high BABIPs on line drives in subsequent seasons, but there seems to be a
positive (though not statistically significant) effect of previously high
LD-BABIPs as well as previously high LD%. 
Switch hitters also did extraordinary well on line drives.  Here is that regression output:

 

Variable	Coefficient (Standard Deviation)
LN(HR/AB)	0.0220 (0.00526)
LD-BABIP	0.121 (0.0845)
LD%	0.237 (0.149)
Switch hitter	0.0168 (0.00778)
_Constant	0.658 (0.751)

 

 

PUTTING IT ALL
TOGETHER

 

Using all of this information, I began putting together a
new improved equation for projecting BABIP using three previous years of
data.  This equation included line drive
rate, in addition to groundball rate, which now came out as statistically
significant thanks to more data.  It
included GB-BABIP but the interaction term between GB% and GB-BABIP was no
longer statistically significant at all, so I left it out.  Percentage of times reached safely on
groundballs in the infield ({Infield Hits + Groundball Reach on
errors}/{Groundballs – Groundball hits to the outfield}) was strongly
statistically significant as well. 

 

Additionally, while infield flies/flyball remained
significant, strongly negative, and pretty close to the same value as the
previous equation, it seemed much more helpful to use outfield flyball BABIP
(OFFB-BABIP) rather than the percentage of balls in play that were hits to the
outfield.  

 

Natural log of homerun rate and natural log of contact rate
remained strongly statistically significant, but log of contact rate is now
less than half as large due to the neutralization of the Ryan Howard effect.

 

This equation had an R-squared of .32, which means that 57%
of the variance in BABIP can be explained by this equation.  As I have mentioned before, the binomial
distribution suggests that pure chance should explain away at least a third of
it as the average hitter with 300 PA has about 400 balls in play each year.

 

Here is the equation to use with three years of 300 PA or
more available.  Note that all eight of
the independent variables were strongly statistically significant, producing
p-values under 1%.

 

Variable	Coefficient (Standard Deviation)
LD%	0.293 (0.0779)
GB%	0.167 (0.0311)
GB-BABIP	0.183 (0.0556)
IFFB%	-0.268 (0.0440)
OFFB-BABIP	0.111 (0.0473)
LN(HR/AB)	0.0164 (0.00337)
LN(CONTACT%)	0.0767 (0.0270)
Reach Safely/Infield Groundball	0.171 (0.0659)
_Constant	0.192 (0.027)

 

 

This would help you project 155 major leaguers for the 2009
season.  However, as there are many
players who will get a lot of major league at-bats in 2009 who failed to get
300 PA in 2006.  In the next section, I
will provide estimates for the regression equations using two previous years of
data, and using one previous year of data.

 

FEWER YEARS OF DATA
TO WORK WITH

 

Using only two previous years of data left very similar
coefficients, but regressed almost all of the coefficients more towards
zero.  They all remains statistically
significant, and all but natural log of contact rate remain strongly
statistically significant.  Percentage of
infield groundballs that a hitter reaches safely on actually became more larger
(the only variable to do so).  I imagine
that this is because the product of speed from three years ago may provide some
insight into how speedy a player is now, but less so.  Players just beginning their careers may have
larger effects on BABIP of this.  Here is
the output using two previous years of data:

 

 

Variable	Coefficient (Standard Deviation)
LD%	0.259 (0.0563)
GB%	0.106 (0.0241)
GB-BABIP	0.113 (0.0410)
IFFB%	-0.226 (0.0322)
OFFB-BABIP	0.0880 (0.0332)
LN(HR/AB)	0.00975 (0.00248)
LN(CONTACT%)	0.0399 (0.0201)
Reach Safely/Infield Groundball	0.150 (0.0412)
_Constant	0.211 (0.0201)

 

This equation had an R-squared of .24, meaning that 49% of
the variance in BABIP can be explained by just two years of data on these
variables.  This equation used 635
observations.

 

Looking at what to do with just one year of data, I made a
similar equation again.  There were
obviously a lot more observations–1035, to be exact.  But it also was much less reliable, since it
used only one previous year of data.  The
natural log of contact rate was no longer statistically significant at all (the
p-value when I did include was something like 56% or something similarly
useless).  It seems that with one year of
data, it is slightly more advantageous to use the percentage of balls in play
that were flyballs and went for hits, rather than the percentage of flyballs
that went for hits (but the results were similar).  I also included the triples/at-bat, which
captured some of the effect of speed, and therefore canceled out some of the
effect of percentage of infield groundballs that a hitter reached safely
on.  The equation came out as follows:

 

 

Variable	Coefficient (Standard Deviation)
LD%	0.192 (0.0317)
GB%	0.120 (0.0185)
GB-BABIP	0.0828 (0.0242)
IFFB%	-0.166 (0.0216)
(FB%)*(FB-BABIP)	0.127 (0.0535)
LN(HR/AB)	0.00576 (0.00157)
Triples/At-bat	0.527 (0.177)
Reach Safely/Infield Groundball	0.0786 (0.0256)
_Constant	0.209 (0.0140)

 

This equation had an R-squared of .19, meaning that 44% of
the variance in BABIP can be explained from these variables.  Most of these variables are similar to the
ones for the regression equations with more years of historical data, but GB%
was a higher coefficient.  Presumably,
this is due to its correlation with groundball rate.  Once again, all of these variables were strongly
statistically significant, except for the percentage of balls in play that were
flyballs & hits, which was statistically significant but had a p-value of
1.8%.

 

PLAYERS

 

Using the equation for three years of historical data when
the player had 300 PA in 2006-2008, or if not, the equation for two years of
historical data when the player had 300 PA in 2007-2008, and if not, the
equation for one year of historical data when a player had 300 PA in 2008, I
was able to project the BABIP for 277 players. 
The table further down summarizes what all of those are.  Obviously, the 2009 season is already
underway, but I have not incorporated this into my regression results
obviously.  As one might expect, guys
like Derek Jeter, Matt Kemp, Chipper Jones, and Joe Mauer topped the list of
projected BABIP and guys like Mark Ellis, Craig Counsell, Khalil Greene, Joe
Mathis, and Omar Vizquel trailed.

 

Before showing the entire table, here are the top 10 guys
projected to improve their BABIP this year and the top 10 guys projected to
fall:

 

name	babip08	ebabip09	Diff.
Corey Patterson	0.215	0.285	0.071
Jose Vidro	0.243	0.303	0.059
Luis Castillo	0.267	0.324	0.057
Kenji Johjima	0.232	0.287	0.056
Carlos Ruiz	0.237	0.288	0.051
Austin Kearns	0.25	0.299	0.049
Paul Konerko	0.244	0.29	0.046
Geoff Blum	0.242	0.287	0.045
Brandon Inge	0.244	0.284	0.04
Gary Sheffield	0.237	0.276	0.04

 

 

name08	babip08	ebabip09	Diff.
Milton Bradley	0.388	0.322	-0.066
Ian Stewart	0.362	0.298	-0.064
Kelly Shoppach	0.357	0.301	-0.056
Nick Punto	0.335	0.281	-0.054
Mike Aviles	0.357	0.305	-0.052
Ray Durham	0.345	0.294	-0.05
Manny Ramirez	0.37	0.322	-0.048
Reed Johnson	0.36	0.313	-0.047
Shin-Soo Choo	0.367	0.322	-0.045
Ryan Ludwick	0.342	0.297	-0.044

 

Unsurprisingly, most of the guys are ones that you would
expect–guys with very low or very high BABIPs in 2008, but there are still a
few interesting ones.  Luis Castillo
should actually be a high BABIP guy: .324, apparently, but hit only .267 in
2008.  That is presumably due to the
injuries he was suffering from last year. 
In fact, he apparently has .362 BABIP so far this year according to
Fangraphs.  Nick Punto hit .335 on balls
in play in 2008, but apparently he projects to be a low BABIP type and should
hit around .281.  Indeed, his 2009 BABIP
thus far is only .234.

 

Without further ado, the projected BABIPs for 2009:

 

name	Ebabip09
Derek Jeter	0.362
Matt Kemp	0.346
Chipper Jones	0.346
Joe Mauer	0.345
Michael Young	0.342
Fred Lewis	0.341
Matt Holliday	0.341
Denard Span	0.341
Ichiro Suzuki	0.338
Yunel Escobar	0.338
Andre Ethier	0.335
Bobby Abreu	0.332
Josh Hamilton	0.329
Kevin Youkilis	0.329
Edgar Renteria	0.329
Nick Markakis	0.329
Jayson Werth	0.329
Jeff Baker	0.328
Joe Inglett	0.328
Carl Crawford	0.327
Howie Kendrick	0.327
Marlon Byrd	0.327
David Wright	0.326
Fernando Tatis	0.326
Magglio Ordonez	0.325
Placido Polanco	0.324
Luis Castillo	0.324
Miguel Tejada	0.324
Hanley Ramirez	0.324
Jamey Carroll	0.324
Ivan Rodriguez	0.323
Skip Schumaker	0.323
Orlando Hudson	0.323
Manny Ramirez	0.322
Chase Headley	0.322
Milton Bradley	0.322
Shin-Soo Choo	0.322
Robinson Cano	0.321
Jhonny Peralta	0.321
Joey Votto	0.320
Mark DeRosa	0.320
Curtis Granderson	0.320
Kelly Johnson	0.320
Hunter Pence	0.320
Mark Teahen	0.320
Felipe Lopez	0.320
Aaron Rowand	0.319
Mark Grudzielanek	0.319
B.J. Upton	0.319
Paul Bako	0.318
Randy Winn	0.318
Gary Matthews Jr.	0.318
Brian Roberts	0.318
Ryan Braun	0.317
Corey Hart	0.317
Miguel Cabrera	0.316
Kosuke Fukudome	0.316
Elijah Dukes	0.316
Vladimir Guerrero	0.315
Alex Rios	0.315
Freddy Sanchez	0.315
Casey Blake	0.314
Ryan Zimmerman	0.314
Cristian Guzman	0.314
Lyle Overbay	0.314
Maicer Izturis	0.314
Jose Reyes	0.314
Carlos Guillen	0.314
Mark Teixeira	0.313
Darin Erstad	0.313
Justin Upton	0.313
Albert Pujols	0.313
Delmon Young	0.313
Jody Gerut	0.313
Reed Johnson	0.313
Jason Kubel	0.313
Edgar Gonzalez	0.313
Garrett Atkins	0.312
Ramon Vazquez	0.312
Jeremy Hermida	0.312
Michael Bourn	0.311
Chone Figgins	0.311
Jeff Kent	0.311
Jose Guillen	0.311
Juan Pierre	0.311
Chris Davis	0.311
David DeJesus	0.310
Johnny Damon	0.310
Ryan Theriot	0.310
J.D. Drew	0.310
Ryan Howard	0.310
Omar Infante	0.310
Jose Castillo	0.309
Brandon Phillips	0.309
Torii Hunter	0.309
Ryan Sweeney	0.309
James Loney	0.309
Dustin Pedroia	0.309
Brad Hawpe	0.309
Alex Rodriguez	0.309
Ryan Church	0.308
Chase Utley	0.308
Julio Lugo	0.308
Kaz Matsui	0.308
Jack Cust	0.308
Justin Morneau	0.308
Jimmy Rollins	0.308
Ronnie Belliard	0.308
Joey Gathright	0.307
Brendan Harris	0.307
Adrian Gonzalez	0.307
Mark Kotsay	0.307
Adam Jones	0.306
Geovany Soto	0.306
Dan Uggla	0.306
Xavier Nady	0.305
Josh Willingham	0.305
Todd Helton	0.305
Mike Aviles	0.305
Erick Aybar	0.305
Gabe Gross	0.304
Ty Wigginton	0.304
Geoff Jenkins	0.304
Conor Jackson	0.303
Shane Victorino	0.303
Jed Lowrie	0.303
Chris Iannetta	0.303
Grady Sizemore	0.303
Jose Vidro	0.303
Lance Berkman	0.302
Rickie Weeks	0.302
Damion Easley	0.302
Willie Harris	0.302
Jason Bartlett	0.302
Evan Longoria	0.302
Mark Reynolds	0.302
Mike Cameron	0.302
Alfonso Soriano	0.302
Akinori Iwamura	0.302
Russell Martin	0.301
Adam LaRoche	0.301
Ryan Doumit	0.301
Aaron Miles	0.301
Franklin Gutierrez	0.301
Marco Scutaro	0.301
Kelly Shoppach	0.301
Kevin Kouzmanoff	0.300
Aramis Ramirez	0.300
Hideki Matsui	0.300
Carlos Gomez	0.300
Raul Ibanez	0.300
Mike Lowell	0.300
Jacoby Ellsbury	0.300
Ross Gload	0.300
Coco Crisp	0.300
Brandon Boggs	0.299
Austin Kearns	0.299
Aubrey Huff	0.299
Alexei Ramirez	0.299
Clint Barmes	0.299
Adam Lind	0.298
Cesar Izturis	0.298
Cody Ross	0.298
Ian Stewart	0.298
Melvin Mora	0.298
Ryan Garko	0.298
Adam Kennedy	0.298
Ryan Ludwick	0.297
Brian Giles	0.297
Brian McCann	0.297
Lastings Milledge	0.297
Garret Anderson	0.297
Willy Taveras	0.296
Adrian Beltre	0.296
Blake DeWitt	0.296
Jeff Keppinger	0.296
Kurt Suzuki	0.296
David Ortiz	0.296
Jose Lopez	0.295
Jay Payton	0.295
Troy Tulowitzki	0.295
Matt Stairs	0.295
Carlos Beltran	0.295
Vernon Wells	0.295
Billy Butler	0.295
Gregor Blanco	0.295
J.J. Hardy	0.295
Casey Kotchman	0.295
Jay Bruce	0.295
Bengie Molina	0.294
Ray Durham	0.294
Asdrubal Cabrera	0.294
Chris Coste	0.294
David Murphy	0.294
Alex Gordon	0.293
David Eckstein	0.293
Ian Kinsler	0.293
David Dellucci	0.293
Ben Francisco	0.293
Rich Aurilia	0.293
Jeremy Reed	0.293
Tadahito Iguchi	0.293
Melky Cabrera	0.292
John Bowker	0.292
Nate McLouth	0.292
Rick Ankiel	0.292
Carlos Delgado	0.291
Eric Hinske	0.291
Doug Mientkiewicz	0.291
Ramon Hernandez	0.291
Carlos Quentin	0.290
Orlando Cabrera	0.290
Paul Konerko	0.290
Yuniesky Betancourt	0.290
Jim Thome	0.290
Stephen Drew	0.289
Daric Barton	0.289
Bill Hall	0.288
Carlos Ruiz	0.288
Jason Kendall	0.288
Kenji Johjima	0.287
Scott Rolen	0.287
Geoff Blum	0.287
Alexi Casilla	0.287
A.J. Pierzynski	0.287
Jorge Cantu	0.287
Richie Sexson	0.286
Jeff Francoeur	0.286
Prince Fielder	0.286
Luke Scott	0.286
Jack Wilson	0.286
Jim Edmonds	0.286
Jason Michaels	0.285
Corey Patterson	0.285
Carlos Gonzalez	0.285
Gerald Laird	0.285
Alfredo Amezaga	0.285
Scott Hairston	0.285
Brandon Inge	0.284
Ken Griffey Jr.	0.283
Willy Aybar	0.283
Edwin Encarnacion	0.283
Carlos Pena	0.283
Jack Hannahan	0.282
Yadier Molina	0.282
Jose Bautista	0.282
Troy Glaus	0.282
Chris Young	0.282
Nick Punto	0.281
Jason Varitek	0.281
Brian Schneider	0.280
Rod Barajas	0.279
Emil Brown	0.279
Jesus Flores	0.278
Luis Gonzalez	0.278
Adam Dunn	0.278
Dioner Navarro	0.278
Juan Uribe	0.277
Chris Snyder	0.277
Gary Sheffield	0.276
Mike Jacobs	0.276
Miguel Olivo	0.275
Marcus Thames	0.275
John Buck	0.275
Pat Burrell	0.274
Bobby Crosby	0.273
Brad Wilkerson	0.272
Pedro Feliz	0.271
Jason Giambi	0.271
Nick Swisher	0.271
Joe Crede	0.270
Kevin Millar	0.270
Khalil Greene	0.269
Mark Ellis	0.265
Craig Counsell	0.263
Jeff Mathis	0.256
Omar Vizquel	0.253

 

 

In doing my previous article for StatSpeak on projection
systems and their ability to project various statistics, I realized that many
of them were not especially good at projecting BABIP.  ZiPS has a tendency to project hitters to
have very extreme BABIPs that are unlikely to occur.  PECOTA has a tendency to project speedy
hitters to have the same high BABIPs that speedy hitters used to have before
scouting data became some advanced (as PECOTA uses historical comparables), and
CHONE safely projects hitters towards the mean, but all of the systems I
studied had correlations with true BABIP of about .40-.44.  Even for those players with one year of data,
the correlation I found was around .44 and for those hitters with three or more
years of historical data, my projected BABIP had a correlation with true BABIP
of .57.  These systems do incredibly with
projected the three true outcomes, but 70.3% of plate appearances in 2008
resulted in an outcome other than a walk, strikeout, homerun, or hit by
pitch.  As far as I know, none of the
major projection systems use batted ball data for hitters to project
statistics.  These systems are getting
very good, and as Tom Tango has pointed out multiple times, the best systems
only do slightly better than Marcel the Monkey. 
I strongly believe that the way to improve projection is to incorporate these variables that I
have used above to project BABIP in isolation.

About these ads