REGRESSION PLANES TO

IMPROVE THE PYTHAGOREAN

PERCENTAGE
A regression model using common baseball statistics to project offensive and defensive efficiency

by
Dennis Moy
A thesis submitted in fulfillment of the
requirements for the degree of honors in
Statistics
University of California - Berkeley
2006

UNIVERSITY OF CALIFORNIA - BERKELEY

ABSTRACT
Prediction Planes for the Pythagorean Percentage
by Dennis Moy
Advisor:

Professor David Aldous

Department of Statistics

In 1985, Bill James, arguably the most renowned analytical baseball statistician, devised
a very simple, but effective formula that predicted a teams winning percentage given
its runs scored and runs allowed. Despite its remarkable accuracy, this model, coined
Pythagorean expectation, was used primarily on seasons of the past rather than
performance forecasts. This thesis develops prediction models for runs scored and
runs allowed that will be converted by Pythagorean expectation to winning
percentages. Data from the past twenty years were taken from four different sources
of baseball statistics via the internet to produce 562 arrays that underwent
computations through GRETL to create two different ordinary least-squares
regression planes (offense and defense). The GRETL outputs yielded robust models
that had strong positive R2 results with significant F-statistics from the Wald test that
evaluated the planes goodness of fit, which with a potentially adjusted Pythagorean
expectation, can now forecast future winning percentages. Armed with this knowledge
and a little calculus, baseball executives can determine which talent is more valuable
when building a successful team to maximize winning percentage.

TABLE OF CONTENTS

List of Figures and Tables...................................................................................................... ii

Acknowledgements................................................................................................................ iii
Glossary................................................................................................................................... iv
Introduction and Background ............................................................................................... 1
Materials and Method............................................................................................................. 8
Data Analysis and Findings.................................................................................................. 10
Discussion of Results............................................................................................................ 20
Conclusion and Extensions ................................................................................................. 25
Appendix (Derivations)........................................................................................................ 28
Bibliography........................................................................................................................... 31

LIST OF FIGURES AND TABLES

Number

Page

Table 1: Summary Statistics from GRETL........................................................................ 10

Figure 1: Scatterplot of Runs (Adjusted) versus Year....................................................... 11
Figure 2: Scatterplot of OBP versus Year .......................................................................... 13
Figure 3: Scatterplot of SLG versus Year........................................................................... 13
Figure 4: Scatterplot of WHIP versus Year ....................................................................... 14
Figure 5: Scatterplot of DER versus Year.......................................................................... 14
Table 2: OLS Estimates of Runs Scored (Adjusted) versus OBP and SLG .................. 15
Figure 6: Fitted, Actual Plot of Runs Scored (Adjusted) versus OBP and SLG......... 16
Figure 7: Residuals for Runs Scored (Adjusted) Regression Model ............................. 17
Table 3: OLS Estimates of Runs Allowed (Adjusted) versus WHIP and DER............ 17
Figure 8: Fitted, Actual Plot of Runs Allowed (Adjusted) versus WHIP and DER ..... 18
Figure 9: Residuals for Run Allowed (Adjusted) Regression Model............................... 19

ii

ACKNOWLEDGMENTS

I would like to express my sincere appreciation to my parents first and foremost for
sending me through school for the past 19 years and always making sure I was aiming
high and giving my best efforts. Also, I want to thank Professor David Aldous for
being a helpful advisor who allowed me to apply all that I have learned to a topic I
love. Thanks to the immortal Team Savage, who provided a gateway each week for me
to chase and realize the dream of an intramural softball championship. Also, I need to
thank my two roommates, Kevin and Herman, for dealing with my complaints about
doing this for a whole semester and the rest of my friends, for using my thesis as an
excuse for not going out and enjoying my last semester in college. Thank you to Mr.
Spellicy for being a great mentor and telling me to pursue what I enjoy. Most
importantly, I need to thank Ms. Delfino for suggesting me to take Advanced
Placement Statistics as a sophomore at Lowell High School, as that paved the road for
my interest and major in statistics. And last but not least, I want to state my deep
gratitude to Eva for keeping me on top of my thesis and helping me edit until the
perfect version came into fruition.

iii

GLOSSARY

Batting Average (AVG or BA). Total Hits / Total At-Bats.

On-Base Percentage (OBP). (Hits + Walks + Hit-By-Pitches) / (At-Bats + Walks +
Sacrifice Flies + Hit-By Pitches).
Slugging Percentage (SLG). Total Bases / Total At-Bats.
OPS. On-Base Percentage + Slugging Percentage.
WHIP. (Total Walks + Total Hits) / Total Innings Pitched. Commonly understood as
average number of base runners per inning.
Defensive Efficiency Rating (DER). The ratio of team defensive outs recorded in
defensive opportunities. 1 [(Total number of hits home runs allowed) / (All
balls hit into play home runs allowed)].
Earned Run Average. (9 * Total earned runs allowed) / (Innings pitched). The most
common measure of a pitchers ability because of its simplicity, despite being
inherently biased and flawed.
Pythagorean Win-Loss Percentage. Bill James stumbled upon this model in The Bill
James Historical Abstract which has no true baseball reasoning. However, this model
predicted winning percentage remarkably accurately with only two simple factors
runs scored and runs allowed. Projected winning percentage = (Runs scored^x)/(Runs
Scored^x + Runs Allowed^x) Upon publication of his Abstract, James allowed x to be
2, thus resembling the mathematical/geometric Pythagorean theorem associated with
right triangles. Over time, the value of x became closer and closer to 1.8.

iv

 Introduction and Background

Open up the newspaper. Which sections use numbers or quantifiable data to transmit their
news? Sure, the business section has a plethora of figures for trade activity in stocks, bonds and other
markets. People interpret these numbers to gauge the value of their holdings and their holdings
competition. A reader can quickly see if Google outperformed Yahoo during the previous days
trading. But where else in the newspaper do you see numbers? Flip to the sports page. The March 22,
2006 version of a San Francisco newspaper has these results scattered among its sports section:
San Antonio Spurs 107, Golden State Warriors 96
San Jose Sharks 6, St. Louis Blues 0
Japan 10, Cuba 6
Oakland Athletics 6, San Francisco 4

However, along with these straightforward scores, there are boxscores that tell the story of
each game. Similar to the business section, these numbers in the boxscores gauge the value of each
player that participated in the game and his or her teammates and opponents. A reader can quickly
look up whether Tim Duncan performed better than Jason Richardson and get a good idea of how the
Oakland Athletics mustered enough runs to defeat their cross-town rivals, the San Francisco Giants, in
a spring training exhibition game. Every major American sports game played has a resulting boxscore,
but baseball is different among the four major sportsfootball, basketball, and hockey. Its regular
season of 162 games is significantly longer than other sportshockey and basketball both have 82
games while football has only 16. During their seasons, football players play once a week on the
gridiron; hockey players rarely skate on back-to-back days for games on the rink; basketball players
barely spend the majority of a week in front of a crowd on a court; however, baseball players
frequently play every day or night for two straight weeks, sometimes even twice a day. With all of this
action, it is fast and easy for every baseball player to build up statistics with significant repetition (or

-1-

sample size). The best position players on each team easily accumulate 500 at-bats, the healthiest
starting pitchers face at least 800 batters, and the best relief pitchers enter the game 60 or more times
during the regular season. Above the other three major sports, baseball stands out as the most
quantifiable and most statistically developed; incidentally, this started before computers were
commonplace in every home and office.
The origins of baseball have been widely disputed; the most common myth credits a Civil War
general, Abner Doubleday, with the invention of baseball in 1839 even though there has been
documented proof that versions of the game existed prior to 1839. In 1845 though, Alexander
Cartwright published the first formal rulebook for the game of baseball; the first resemblance of a
boxscore, or abstract as it was called then, appeared in the October 22, 1845, New York Morning News
issue. Soon after, Henry Chadwick devised the most primitive statisticsbatting average and earned
run averageto valuate different players on different teams. However, it would take the greater part
of a century before baseball statistics reached a new echelon. As the game grew more popular, statfilled books such as Whos Who in Baseball and Balldom: The Britannica of Baseball were published in the
1910s by trailblazing statisticians. During this time, the value of a player was judged by the eyes of old,
leather-skinned scouts who would spend many summer days under the sun scouring local baseball
games to discover talent; a players past statistics merely served as tools for banter among the crowds
of fans. Signs of progressive change surfaced in the late 1940s for the world of baseball and the world
of baseball statistics as well. Branch Rickey, one of the most innovative men among the ranks of a
baseball teams front ranks, spearheaded what seemed to be a radical baseball evolution. After
developing the first full-time spring training facilities which are a staple of March in Arizona and
Florida today, Branch Rickey shook Major League Baseball upside down when he signed Jackie
Robinson in 1947 to the Brooklyn Dodgers to break the color barrier. Below the hoopla of this

-2-

transaction, Rickey welcomed a Canadian, Allan Roth, to his team hours before the Robinson signing.
No, Allen Roth was not the first Canadian baseball player, but Roth was a statistician who had groundbreaking theories on a new statistic to evaluate offensive performance. Batting average conspicuously
omits bases on balls in its computation; however, Allan Roth compensated for this omission when he
helped Rickey introduce the new statistic of on-base percentage or OBP. Logically, a walk was as good
as a base hit, since the batter did not cause an out and reached base safely after judging four balls
outside of the strike zone or being struck by a pitched ball; batting average fails to acknowledge this
skillful performance as on-base percentage does, but it would take until the 1960s for more abstract
analysis of baseball statistics.
Two publications during this decade invigorated younger, mathematical minds to pursue the
unexplored realm of statistics in baseball. Earnshaw Cook penned a theory-intensive book titled
Percentage Baseball which nebulously disproved the effects of many common, traditional baseball
strategies such as stealing bases and using sacrificing bunts. This book was written too early for its
time, but it did pave the way for researchers across the nation to collaborate for a tour de force that
was published in 1969. The Baseball Encyclopedia, a bestseller published by Macmillan, was the collective
product of many man-hours of perusing old boxscores to build a source of reference on every baseball
player and how he performed during his career. Not only did this book introduce the use of
technology to compute new baseball statistics, but it brought together a legion of baseball statisticians
for the first ever meeting of the Society for American Baseball Research.
The Society for American Baseball Research, or SABR for short, was born in 1971 when 16
mean gathered in Cooperstown to discuss the importance of preserving baseball history; membership
swelled to 6,000 members in 15 years and over time, baseball statisticians became known as

-3-

sabermetricians because of this organized group. Ironically, the man who would contribute most to
sabermetrics after the societys inception could not afford the $25 cost of The Baseball Encyclopedia.
In 1977, Bill James finished his masterpiece, The Bill James Historical Baseball Abstract, which
brought baseball research to a new level. Unlike books in the past, James wrote about topics that were
never thought about; for instance, he went out of the proverbial box and hypothesized how to valuate
a player by the number of runs he contributed to the team. Aside from merely reporting mundane
statistics like batting average or on-base percentage, this book identified that runs were the key
ingredient in allowing a team win. Runs were the most valued asset in baseballscore more, and your
team has a better chance of winning; prevent your opponent from scoring, and once again your team
has a better chance of winning. Branching out even farther, James essentially stumbled upon a key
model that estimated a teams winning percentage using runs scored and runs allowed that mystically
resembled the Pythagorean theorem in the mathematical and geometric world that related the sides of
a right triangle. The model was precise enough that teams began to use it to assist them in evaluating
which players to add or subtract to their team in order to judge the transactions effect on winning
percentage. He appropriately coined the model as the Pythagorean Win-Loss Percentage. The glory of
this model was that the two inputsruns scored and runs allowedcreated a projected winning
percentage which was independent of which league the team played in, the length of the season, and
most importantly, during what era the team played. Initially, James used this formula to attribute
whether teams were lucky or unlucky in seasons past; because of the abovementioned independences,
deviation from the estimated winning percentages were simply blamed on good or bad luck.
Following James footsteps by tinkering with the exponent in the model, many sabermetricians
used historical statistics for every team in every season to minimize the residuals to create the best
fitting Pythagorean Win-Loss Percentage model (with an exponent that floated around 1.8). But these

-4-

sabermetricians seemingly never used this model to forecast future winning percentages, for there were
only primitive (and complex) models that estimated runs scored and runs allowed. Bill James runs
created was such a model that used eleven input variables, which aside from being too difficult to
compute (as it needed the values for hits, walks, times caught stealing, times hit by a pitch, double
plays, total bases, stolen bases, intentional walks, sacrifice bunts, sacrifice flies, and at bats), they
artificially inflated the correlation coefficients, thus deceptively implying importance. Also, there was
very little research into creating a model for runs allowed; despite the development of many nontraditional pitching statistics, measuring a teams fielding abilities was still unrealized. As computers
became more and more popular, baseball research began to accelerate.
By the 1990s, sabermetrics reached a new level of popularity with the advent of the internet
and the World Wide Web. A faction of the Society for American Baseball Research convened in 1994
with the mission of digitizing the boxscore of every single baseball game ever played. This group,
Retrosheet, Inc., is still at work today, trying to scour old newspapers for hardcopies of these
boxscores. As these records were digitized, they were made available via their website which gave birth
to a new breed of baseball techies. Sean Forman is one of these techies and he manipulated the
digitized boxscores to create baseball reference websitesdutifully titled Baseball-Reference.com and The
Baseball Archivewhich had pages for every player and his statistics who played in the major leagues
and for every team as well. He even made his work free and accessible through spreadsheets and
retrievable through databases using Structured Query Language (SQL). Currently, Gary Cohen is
taking this research one step further by perusing for minor league statistics and posting them through
his website, TheBaseballCube.com. As these resources became so readily available, websites that require
membership such as BaseballProspectus.com performed very detailed analysis and consequently new

-5-

abbreviations for baseball statistics sprang up just as Franklin Roosevelts New Deal alphabet soup
deluged America during the Great Depression.
Ironically, it took a book not about baseball to put some of these abbreviations into the
mainstream. Michael Lewis, a well-known author because of his most prestigious bestseller Liars
Poker, finished Moneyball in 2003, which only used the context of baseball to emphasize the importance
of capitalizing on undervalued resources in a competitive market. In 2002, Lewis shadowed Billy
Beane, the general manager for the Oakland Athletics, to gain insight from the teams front office.
Beane was a huge proponent of players with high on-base percentages, for his theory revolved around
the fact that a 9-inning game had a finite number of outs27; on-base percentage was a measure of
players ability to avoid making an out by reaching base safely. Thus, by fielding a team full of players
who produced fewer outs on the average compared to other players, Beane theorized that the resulting
team would maximize offensive production. As a small-market team, the Oakland Athletics were
forced to dig deeper and investigate the subtle statistics of players, rather than the glowing statistics
such as batting average and home runs, since the large-market teams could spend more money and
sign these superstar players. Lewis book revolved around this concept and how the Oakland Athletics
were able to discover an undervalued resource and capitalize on its under-appreciation.
As The Bill James Historical Abstract became more popular in the baseball community, some
readers became skeptical of some James theoriesone of which was Beanes right-hand man, Paul
DePodesta, who Lewis relates about thusly on page 128 in Moneyball:
Baseball fans and announcers were just then getting around to the Jamesean obsession with
on-base and slugging percentages. The game, slowly, was turning its attention to the new statistic, OPS
(on-base plus slugging)Crude as it was, it was a much better indicator than any other offensive statistic
of the number of runs a team would score. Simply adding the two statistics together, however, implied
they were of equal importanceAn extra percentage point of on-base was as good as an extra
percentage point of slugging.

-6-

Before his thought experiment Paul had felt uneasy with this crude assumption; now he saw
that the assumption was absurd. An extra point of on-base percentage was clearly more valuable than an
extra point of slugging percentagebut by how much? He proceeded to tinker with his own version of
Bill Jamess Runs Created formula. When he was finished, he had a model for predicting run
production that was more accurate than any he knew of. In his model an extra point of on-base
percentage was worth three times an extra point of slugging percentage.
Pauls argument was radical even by sabermetric standards. Bill James and others had stressed
the importance of on-base percentage, but even they didnt think it was worth three times as much as
slugging. Most offensive models assumed that an extra point of on-base percentage was worth, at most,
one and a half times an extra point of slugging percentagePauls argument was basically heresy.

Was Paul DePodestas argument and model heresy? Or were the original sabermetricians and their
models for offensive production erroneous? This thesis will take the opportunity to settle this
argument. Indeed, the Oakland Athletics succeeded in advancing to the playoffs for four straight years
(2000 to 2003), but they never advanced past the first round of the postseason. This may occur
because of a horrible string of bad luck, especially since the Athletics were eliminated in the do-or-die
game each year. But, despite the success the Athletics experienced in the regular season, were they
doomed for failure because of the overemphasis Billy Beane and Paul DePodesta had on on-base
percentage? Did these Athletics teams of the new millennium need more power (a higher team
slugging percentage) to advance deeper into the playoffs? And defensively, did Billy Beane focus too
much on building a strong pitching rotation and ignore fielding prowess? To answer these questions,
the following pages will take 20 years of team data and create a new offensive production model; they
will create a defensive ability model, which has been relatively unexplored by sabermetricians.
Consequently, these two models will result in runs scored and runs allowed projections for every
team in the past twenty years, which will then be plugged into James tried, tested, and adjusted
Pythagorean Win-Loss Percentage formula to conclude whether the new regression equations are
valid. If these findings are validated, front offices of Major League baseball teams will have a better
grasp on a players value upon a decision-making transaction by seeing how he contributes or detracts
from their teams ability to score or prevent runs from scoring.

-7-

 Materials and Methods

As mentioned before, there is an abundance of sortable baseball statistics from various internet
sources; aside from the more private-run websites such as Baseball-Reference.com, The Baseball Archive, and
TheBaseballCube.com, more public websites such as MLB.com and ESPN.com have plentiful statistics that
are free and easy to obtain. Since Bill James devised the Pythagorean prediction formula in 1985, the
data were taken starting from the 1986 season from these various sources.
Initially, Statement Query Language was used to sort the data, but that process became drawn
out and eventually fruitless. Sean Lahman, put up on his free access webpage, The Baseball Archive,
many zipped archives of text files with conveniently comma-separated values that had statistics of all
sorts drawn from Retrosheet.orgs digitized boxscores dating from the 19th century. Jim Albert and Jay
Bennett, in Curve Ball, suggested a model to project offensive productionassuming that runs were
literally a product of on-base and slugging percentage (OBP x SLG). However, by multiplying the two
statistics together, it becomes impossible to distinguish which had more weight in run production.
Also, in order to stray away from Bill James complex runs created model, it was decided to limit the
number of independent variables to two (for both offensive and defensive models) to maintain
simplicity and allow the graphing of the model with a regression plane in three dimensions. Therefore,
compared to the defensive model, the offensive model was simpler to put together:

Runs Scored = !0OBP + !1SLG + !2

As for the defensive model, earned-run average could not be used since runs would be on
both sides of the regression equation, resulting in unwarranted and incorrect correlation. Instead, runs
allowed were regressed on WHIP (walks and hits per inning pitched) and DER (defensive efficiency),
which was based solely on balls put into play excluding homeruns, thus gauging fielding ability. In

-9-

other words, the ability to limit the opponents scoring is based on keeping runners off base and
making outs on playable balls:

Runs Allowed = "0WHIP + "1DEF + "2

Collection of data was both easy and difficult. Initially, ESPN.coms statistics were observed
but it was impossible to gather large chunks of data without avoiding manual input. Instead, the
CSV files from Sean Lahmans archive were unzipped and converted into Microsoft Excel
spreadsheet format; the spreadsheet was filtered to leave only the pertinent data (games played, wins,
losses, runs scored, runs allowed, OBP, SLG, walks, hits, and innings pitched) for the appropriate
time period (after 1986). A simple formula was used to convert walks, hits and innings pitched into
the WHIP statistic, but only MLB.com had historical DER statistics. Therefore, those numbers were
manually appended to the spreadsheet. A careless error was avoided by adjusting the runs scored
and runs allowed statistics to a full 162 game season, since 1994 and 1995 had shortened seasons
because of the players strike. With data in proper form, the regression models could be built
through an open-source program primarily used for econometric research. This laid the groundwork
to construct the regression planes which predicted runs scored and runs allowed and in turn,
winning percentage via the Pythagorean expectation.

- 10 -

 Data Analysis and Findings

Gnu Regression, Econometrics and Time-series Library, better-known as GRETL, is an open
source program used to compile and analyze statistics via least-squares regression models. After
GRETL imported the data, it provided straightforward summary statistics:

Summary Statistics, using the observations 1 562

Variable
RS_a
RA_a
OBP
SLG
WHIP
DER

Mean
753.815
753.833
0.330665
0.411069
1.38796
0.710258

Median
747.500
748.000
0.329000
0.409000
1.37869
0.710700

Minimum
548.000
539.000
0.296000
0.327000
1.15149
0.675500

Maximum
1009.00
1103.00
0.373000
0.491000
1.73394
0.744000

Variable
RS_a
RA_a
OBP
SLG
WHIP
DER

Std. Dev.
89.1643
94.7940
0.0144796
0.0306140
0.0941603
0.0122506

C.V.
0.118284
0.125749
0.0437894
0.0744742
0.0678407
0.0172480

Skewness
0.261585
0.281047
0.239716
0.0844287
0.282385
0.0101061

Ex. kurtosis
-0.304824
-0.235518
-0.238634
-0.310604
-0.102395
-0.0983793

Table 1: Summary Statistics from GRETL.

RS_a = runs scored adjusted to 162 games
RA_a = runs allowed adjusted to 162 games.
Aside from summary statistics, GRETL quickly created scatterplots with respect to the year,
thus demonstrating the trend of the statistic over time:

- 11 -

Figure 1: Scatterplot of Runs (Adjusted) versus Year

As one can see, there has been a gradual increase in run production from 1986 to 2005.
However, it is uncertain whether this increase is due to the improved abilities of hitters or the
decreased abilities of pitchers and fielders. To visually check this, scatterplots of OBP, SLG, WHIP
and DER can be constructed to view the trend:

- 12 -

Figure 2: Scatterplot of OBP versus Year

Figure 3: Scatterplot of SLG versus Year

Offensive output has seemingly had a significant gradual increase, with a jump between the
years of 1993 and 1994, especially with power (slugging percentage).

- 13 -

Figure 4: Scatterplot of WHIP versus Year

Figure 5: Scatterplot of DER versus Year

- 14 -

Pitching seems to have remained stable, as WHIP has increased only very little, but fielding
ability has dipped during the past twenty years. Despite the number of homeruns increasing since the
late 1990s, possibly due to the Steroid Era in baseball (which is being resolved), DER has decreased,
where an increase in homers should cause the DER to increase as well.
To truly test the hypothesis of this thesis, GRETL can construct ordinary least-squares
regression planes and their corresponding equations, correlation coefficients, and other helpful
statistical tests to verify the offensive model:
Model 1: OLS estimates using the 562 observations 1-562
Dependent variable: RS_a
Variable
const
OBP
SLG

Coefficient
-849.805
2918.72
1553.26

Std. Error
27.448
128.515
60.7843

t-statistic
-30.9605
22.7111
25.5536

p-value
< 0.00001
< 0.00001
< 0.00001

significance
***
***
***

Mean of dependent variable = 753.815

Standard deviation of dep. var. = 89.1643
Sum of squared residuals = 407849
Standard error of residuals = 27.0112
Unadjusted R2 = 0.908556
Adjusted R2 = 0.908229
F-statistic (2, 559) = 2777.02 (p-value < 0.00001)
Log-likelihood = -2648.43
Akaike information criterion = 5302.87
Schwarz Bayesian criterion = 5315.86
Table 2: OLS Estimates of Runs Scored (Adjusted) versus OBP and SLG

The best-fitting model results in a very good fit, according to the significance of the variables
and the strong, positive correlation coefficient. As expected, the model suggests that on-base
percentage and slugging percentage have direct positive effects on runs scored; the higher the OBP
or SLG, the higher the runs scored. In addition, an F-statistic with a low p-value suggests that none
of the variables (OBP and SLG) can be omitted when computing the regression plane:

- 15 -

Figure 6: Fitted, Actual Plot of Runs Scored (Adjusted) versus OBP and SLG
Visually, regression planes, in three-dimensions, are much easier to grasp, whereas regression
models with more than two input variables cannot be graphed. GRETL allows the threedimensional figure to be rotated, and the image above optimally portrays the plane in space such
that the axes are easily visible while the actual values can be seen relative to the regression plane.
Also, the residuals appear to be randomly scattered, thus discounting heteroskedasticity (unequal
scatter) or autocorrelation:

- 16 -

Figure 7: Residuals for Runs Scored (Adjusted) Regression Model

And for the regression results of the defensive model:
Model 2: OLS estimates using the 562 observations 1-562
Dependent variable: RA_a
Variable
const
WHIP
DER

Coefficient
43.9101
862.697
-686.326

Std. Error
159.474
24.3072
186.83

t-statistic
0.2753
35.4914
-3.6735

p-value
0.783155
< 0.00001
0.000262

Mean of dependent variable = 753.833

Standard deviation of dep. var. = 94.794
Sum of squared residuals = 727668
Standard error of residuals = 36.0795
Unadjusted R2 = 0.855652
Adjusted R2 = 0.855136
F-statistic (2, 559) = 1656.8 (p-value < 0.00001)
Log-likelihood = -2811.12
Akaike information criterion = 5628.23
Schwarz Bayesian criterion = 5641.23
Table 3: OLS Estimates of Runs Allowed (Adjusted) versus WHIP and DER

- 17 -

significance
***
***

Despite having a lower correlation coefficient than the offensive projection plane, the
defensive model still produces significant coefficients and F-statistic. The model also confirms the
hypothesis that a higher WHIP and lower DER correlates to more runs allowed, whereas a lower
WHIP and higher DER correlates to fewer runs allowed.

Figure 8: Fitted, Actual Plot of Runs Allowed (Adjusted) versus WHIP and DER
The plane seems to be a good fit, graphically as the difference between the fitted and actual
values are minimized. Just like the offensive regression plane, the residuals appear to be scattered
randomly (aside from a couple extreme outliers), avoiding heteroskedasticity and autocorrelation.
For both models, the regression model seems appropriate given the proximity of the actual values to
the fitted plane.

- 18 -

Figure 9: Residuals for Runs Allowed (Adjusted) Regression Model

With all of these graphs and tables computed by GRETL, the results can be contextualized to
answer the questions mentioned before, and consequently, teams can use these answers to make
personnel decisions when building a winning baseball team.

- 19 -

 Discussion of Results
Without a doubt, run production has increased steadily in the past twenty years. As the
previous scatterplots highlighted, OBP and SLG (the two clear gauges for offensive efficiency) both
have climbed in the recent past while WHIP has correspondingly increased as well. However, this can
be due to the fact that OBP and WHIP are inversesOBP is the ability to become a base runner,
while WHIP is the number of base runners allowed per inning. The gradual decrease in DER seems
completely independent of the other three statistics.
There have been several, popular theories that attempt to explain these changes in baseball.
The most logical proposal attributes the increase in offensive production to baseballs expansion from
26 to 30 teams between 1986 and 2005. As there are 25 players per team roster, this means that there
were 100 more fringe major league-level players in 2005 than 1986. Also, weak hitters can be hidden
on the teams bench, used in fill-in situations or when the game is already decided. In contrast
however, the better pitchers are physically unable to pitch every inning of every game, so in order to
compensate for the lost innings, the pitchers among the 100 extra fringe players would have filled
those innings. Consequently, worse pitchers facing better hitters naturally inflates run output, which
legitimizes the upward trends of OBP, SLG and WHIP. As for the decrease in DER, an explanation
for this can be rooted on the more recent emphasis on offensive skills in lieu of defensive ability. In
other words, gone are the days of the slick-fielding middle infielders (shortstops and second basemen)
who have sub-par batting averages. General managers are favoring such middle infielders that can hit
as well as the more traditional power positions (first basemen, third basemen, and corner outfielders)
who sacrifice fielding ability for offensive prowess.

- 20 -

Aside from the expansion theory, other suggestions for offensive inflation include rumors of
Major League Baseball executives authorizing juiced ballsbaseballs that are more tightly wound at
their seams, which would cause them to carry a longer distance. Also, pitchers have opined that the
strike zone, which is interpreted independently by different umpires, has generally shrunken; this
forces pitchers to pitch around the plate more, thus providing batters with better chances to make
solid contact. The most recent and popular theory attributes offensive inflation to steroids and
performance-enhancing drugs. For years, the players unions collective bargaining agreement
prohibited drug testing by the league, which essentially allowed the players to use and/or abuse any
drug without the threat of punishment. With this in mind, there is almost no doubt that a significant
percentage of players abused steroids in order to improve their performance on the field, including allstars that have been subjugated to a grand jury investigation in 2003 whose after-effects have forced
Congress to intervene. For example, Barry Bonds, who has been under the most scrutiny recently, saw
his yearly homerun count jump in 2000 and peak at a record-setting 73 homeruns in 2002, despite
arguably playing after his prime into his late 30s and early 40s. As steroid distributors like Victor Conte
and Greg Anderson underwent prison time for supplying steroids to athletes including baseball
players, consequently in 2005, Major League Baseball imposed progressively harsh suspension
penalties upon failure of random drug tests to eradicate steroids from the sport. No matter what
theory is proposed, the data do not lie, since the trends have suggested that offense has increased in
the past twenty years.
The fitted model for offensive production refutes both Paul DePodesta and traditional
sabermetricians to certain degrees. As mentioned before, Paul DePodesta argued in an interview in
Moneyball that OBP was three times as important as SLG, and traditional sabermetricians, by simply
creating the OPS statistic (OBP + SLG) inherently value OBP and SLG as equals. But the equation

- 21 -

for the regression plane, as shown in Table 2, yields coefficients of 2918.72 and 1553.26 for OBP and
SLG, respectively. This implies that OBP is roughly twice as important as SLG, when used to measure
offensive production through runs. The plane projects that a .001 increase in OBP adds approximately
3 runs of offense, whereas a .001 increase in SLG only adds 1.5 runs. The model falls in between Paul
DePodestas argument and the proposal of more traditional sabermetricians. In addition, the high
correlation coefficient, meaning that the proportion of the amount of variance in runs scored that can
be explained by OBP and SLG, is 0.908. A correlation coefficient of 1 suggests strict dependence.
With this in mind, the regression plane lends itself as a very good fit in estimating runs scored. The
more often a teams players can get on-base (high OBP), coupled with the ability to drive them in with
power (SLG) should provide that team with a healthy amount of runs scored. Logically, OBP should
have more weight than SLG. For instance, if a team is able to load the bases with runners and then hit
a homerun, the grand slam would score four runs. But if a team cannot get runners on base before the
homerun, the solo homerun would score only one run; in other words, OBP precedes SLG when
scoring.
Obviously, there are other facets to offense in baseball that have been omitted in this relatively
simple three-dimensional model. Statistics such as stolen bases and sacrifice hits should not be ignored
in an offensive strategy. However, they were left out of the presented offensive model in order to
restrict the number of independent variables to two. By doing this, the model remains graphical (in
three dimensions with a plane) and thus easier to grasp for the less statistically-inclined. In addition, by
adding more variables, the correlation coefficient becomes artificially inflated in a regression model. As
for removing either of the two variables (OBP and SLG) from the proposed model, the Wald test and
consequent F-statistic argue against omitting either variable. The residual plot appears random enough

- 22 -

to discount serial correlation or heteroskedasticitythis result pushes for more support of the
proposed offensive model.
The second model was initially built based on earned-run average (ERA) and fielding
percentage, both of which are common measurements of pitching ability and defensive prowess;
however, this preliminary hypothesis was dropped immediately as the dependent variable, runs
allowed, is directly influenced by ERA. Also, fielding percentage only measures the proportion of
errors to assists and putouts. In lieu of ERA and fielding percentage, WHIP and defensive efficiency
rating (DER) provide accurate gauges of pitching and defensive abilities, respectively. Another reason
why these two measures are better suited is that they are independent of each other. WHIP is rooted
into pitching aloneit is the measure of a pitchers ability to keep runners off the basepaths, and if
there are no base runners, no runs can score. DER is a measure of the teams defensive ability to
convert any balls put into play into outs, which is not influenced by the pitchers ability to prevent
runners from reaching base. Accordingly, one would expect runs allowed to increase as WHIP
increases and as DER decreases.
The model confirms this hypothesis as detailed in Table 2. The coefficient for WHIP is about
863, meaning a one-hundredth increase in WHIP allows 8.6 more runs over the course of the season,
while the coefficient for DER is -686, which suggests that a one-hundredth increase in DER takes
away 6.7 more runs over the course of the season. It is evident that keeping a low WHIP and a high
DER would limit the number of runs allowed. The models correlation coefficient of 0.855, despite
not being as strong as the offensive models correlation coefficient, still suggests that it can predict
runs allowed accurately. The less often a teams pitchers allow the opponent to reach base, coupled
with a high proportion of turning balls put in play into outs would impair the opponents offensive
prospects.

- 23 -

The defensive regression planes residuals appear to be random as well, again discounting
heteroskedasticity and autocorrelation. The F-statistic through the Wald test does not suggest omitting
either WHIP or DER in the model. Unlike the offensive model, there are only a few more statistics,
mostly pitching ones that can be considered in constructing the plane. Strikeouts are conspicuously
absent though, as this model assumes strikeouts are equivalent to any other kind of outgroundout or
flyout. At times, a groundout or flyout can be more productive in scoring runs by advancing runners,
by way of a fielders choice or sacrifice bunts or fly balls. However, this factor may be offset by the
volume of double plays turnedin other words, a strikeout in this situation only results in one out,
whereas a double play yields two outs, further dampening the chance to score in that inning.
Although both models are reasonably good fits to predict runs scored and allowed, it is
impossible to completely find the perfect combination of variables. There is a certain luck factor
involved in the gamein a 162 game season, anything can happen; the leagues worst team can
surprise the leagues best team and win a three-game series. Also, Bill James argued that almost all onerun games rely on luck. For instance, if a team strings together three hits in one inning and scores one
run, while its opponent racks up ten hits over the duration of the game but fail to score, the team that
had worse offensive output and defensive ability for that game wins 1-0. Because of this, the
correlation coefficients that were borne from both models, since they are reasonably close to 1, should
be adequate in suggesting the sufficiency of the two regression planes.

- 24 -

 Conclusion and Extensions

Some might say that the models above are common sensethat it is logical to predict
offensive production from a teams capacity to reach base and to drive runners home to score and to
predict defensive aptitude from the ability to prevent runners from getting on base and recording outs
when the ball is hit into play. However, many baseball fans and pundits still heavily lean on statistics
that fail to accurately describe a player or teams abilitiesbatting average, runs batted in, earned-run
average, and fielding percentage. These statistics are too basic albeit very simple to compute, which is
probably why they are so frequently broadcasted. The significance of the high correlation coefficients
will hopefully convince a portion of the baseball masses that there are baseball statistics out there aside
from the mundane that are very useful and powerful in describing a player or his team. Perchance,
statistics such as OBP, SLG, WHIP, and DER will someday become mainstreamOBP has grown
popular because of Moneyball while SLG and WHIP are making headway in the media. However, DER
is still a mystery to many, despite being a great measure of a teams skill in converting balls put into
play into outs, thus prohibiting the opponent to score.
This model has already established the weights of importance to the four above-mentioned
statistics in offensive and defensive contexts, but what can this model extend to? Returning to Bill
James Pythagorean Percentage, where the formula is:

Win% =

RunsScored
( RunsScored + RunsAllowed )

Originally, the exponent that was used was two, which resembled the traditional Pythagorean
equivalence for right triangles. However, as mentioned before, this exponent has been adjusted to

- 25 -

approximately 1.8. Now, if the regression models from before are plugged into the Pythagorean
model:

Win% =

(2918.72*OBP+1553.26*SLG-849.91)
(2918.72*OBP+1553.26*SLG-849.91) + (862.70*WHIP-686.33*DER+43.91)

Using this model and its corresponding partial derivatives:

Win% Win% Win%

Win%
,
,
, and
can determine the precise value of each of the
OBP
SLG WHIP
DER
variables and how their changes would affect team winning percentage. The quotient rule for simple
calculus can be used to find the partial derivatives and its numerical value can actually be computed.
Thus, one can observe the change in winning percentage due to an incremental change in any of the
four other statistics. In addition, OBP, SLG, and WHIP are statistics that can be individualized; that is,
individual players have these measures themselves. Because of this, a general manager for a baseball
team, upon deciding whether to add or subtract a certain player from the team, can gauge the change
in the teams OBP, SLG, or WHIP and its corresponding effect on winning percentage. Unlike the
three other statistics, DER fails to trickle to the individual levela certain players defensive ability is
usually measured through errors, assists, put-outs and the consequential fielding percentage. However,
there are new statistics such as zone ratio (ZR) and range factor (RF) that can describe a defensive
player. These measures are still relatively subjective, but are definitely more objective than a scouts eye
which is still useful for instances such as gauging players arm strength, which is immeasurable for nonpitchers when playing defense. Perhaps this is what the future holds for the ongoing battle between
traditional scouts and statistical analysis; statistical analysis can confirm a scouts opinions or a scouts
observations can confirm statistical analysis. There can be a fine line or balance such that both sides
work together to optimally judge a player and his impending value for his team.
- 26 -

The two models above are not meant to be revolutionary in the world of sabermetrics; but
coupled with the Pythagorean percentage, they should be able to influence player personnel decisions.
The regression models are restricted to two inputs each, so that the results can be graphed and more
easily interpreted, even for the non-statistically inclined. The planes are accurate fits that should not be
ignored when evaluating talent with statistical analysis. Hopefully, OBP, SLG, WHIP, and DER will
not only become mainstream statistics, replacing those that are just too basic to accurately measure
ability, but become the statistics of choice when judging a player or teams offensive and defensive
skills.

- 27 -

APPENDIX

Derivation of the Three-Variable Regression Model (from Ramanathan)

A three-variable regression model (a plane in three dimensions) projects the dependent variable Y to a
constant and to two independent variables X 2 and X 3 . The formal model is:
(Eq. 1)

Yi = 1 + 2 X t 2 + 3 X t 3 + ut
After taking the mean of each term in the model, the equation becomes:

(Eq. 2)

Y = 1 + 2 X 2 + 3 X 3 + u
Subtracting Eq. 2 from Eq. 1, the model is in deviation form:
yi = 2 xt 2 + 3 xt 3 + et

Where yt = Yt Y , xt 2 = X t 2 X 2 , xt 3 = X t 3 X 3 , and et = ut u . The lowercase letters mean that the

variables are deviations from the respective averages. By doing this, the model in deviation form only
requires the estimation of two parameters ( 2 and 3 ). If 1 , 2 and 3 are the estimates of the
regression coefficients, 1 is estimated as:

1 = Y 2 X 2 3 X 3

And an estimate of the residual term is:

ut = Yt 1 2 X t 2 3 X t 3

- 28 -

The ordinary least-squares procedure minimizes the error sum of squares (ESS= ut2 ) with respect
to 1 , 2 and 3 , which is equivalent to minimizing et2 = (Yt 1 2 X t 2 3 X t 3 ) 2 . Setting the
partial derivatives of this with respect to 2 and 3 , it is simple to verify that the conditions are:

xt 2et = 0 = xt 2 ( yt 2 X t 2 3 X t 3 )

xt 3et = 0 = xt 3 ( yt 2 X t 2 3 X t 3 )
This leads to the following two normal equations (omitting the t-subscript):
2 x22 +3 x2 x3 = yx2

3 x32 +2 x2 x3 = yx3
With simplified notation, Eq. 4 and Eq. 5 can be rewritten as:
2 S 22 + 3 S23 = S y 2

(Eq. 3)

2 S 23 + 3 S33 = S y 3

(Eq. 4)

Where

S 22 = xt22 = ( X t 2 X 2 )2

S 23 = xt 2 xt 3 = [( X t 2 X 2 )( X t 3 X 3 )]

S33 = xt23 = ( X t 3 X 3 )2

- 29 -

S y 2 = xyt xt 2 = [(Yt Y )( X t 2 X 2 )]

S y 3 = xyt xt 3 = [(Yt Y )( X t 3 X 3 )]
The solutions to Eq. 6 and Eq. 7 yield:
2 = (S y 2 S33 S y 3 S 23 ) /( S22 S33 S232 )

3 = ( S y 3 S 22 S y 2 S 23 ) /( S22 S33 S232 )

- 30 -

REFERENCES

JAMES, BILL (1985): The Bill James Historical Abstract. Villard Books.
ALBERT, JIM, AND JAY BENNETT (2003): Curve Ball. Copernicus Books.
SCHWARZ, ALAN (2004): The Numbers Game: Baseballs Lifelong Fascination with Statistics. Thomas Dunne
Books.
LEWIS, MICHAEL (2003): Moneyball: The Art of Winning an Unfair Game. W. W. Norton & Co.
GOLDMAN, STEVEN (2006): Can a Team Have Too Much Pitching? Baseball Between the Numbers. Ed.
Jonah Keri. Basic Books / Perseus Books Group. 272-291.
RICE, JOHN A. (1995): Mathematical Statistics and Data Analysis, Second Edition. Duxbury Press /
Wadsworth Publishing Company.
RAMANATHAN, RAMU (2002): Introductory Econometrics with Applications, Fifth Edition. South-Western /
Thomson Learning.
FORMAN, SEAN L.: Baseball-Reference.com Major League Statistics and Information. http://www.baseballreference.com/. (January 31, 2006).
FORMAN, SEAN L.: The Baseball Archive. http://www.baseball1.com/statistics/. (January 31, 2006).
COHEN, GARY: The Baseball Cube. http://www.thebaseballcube.com/. (January 31, 2006).
MAJOR LEAGUE BASEBALL: Major League Baseball: The Official Site.
http://mlb.mlb.com/NASApp/mlb/index.jsp. (January 31, 2006).
ESPN.COM: Baseball Index. http://sports-att.espn.go.com/mlb/index. (January 31, 2006).
RETROSHEET, INC.: Retrosheet. http://www.retrosheet.org/. (January 31, 2006).

- 31 -