Professional Documents
Culture Documents
Basketball Statistics Starting Point
Basketball Statistics Starting Point
Sports
Volume 3, Issue 3 2007 Article 1
Recommended Citation:
Kubatko, Justin; Oliver, Dean; Pelton, Kevin; and Rosenbaum, Dan T. (2007) "A Starting Point
for Analyzing Basketball Statistics," Journal of Quantitative Analysis in Sports: Vol. 3: Iss. 3,
Article 1.
DOI: 10.2202/1559-0410.1070
©2007 American Statistical Association. All rights reserved.
Brought to you by | University of Toronto-Ocul
Authenticated
Download Date | 3/2/16 11:03 PM
A Starting Point for Analyzing Basketball
Statistics
Justin Kubatko, Dean Oliver, Kevin Pelton, and Dan T. Rosenbaum
Abstract
The quantitative analysis of sports is a growing branch of science and, in many ways one that
has developed through non-academic and non-traditionally peer-reviewed work. The aim of this
paper is to bring to a peer-reviewed journal the generally accepted basics of the analysis of
basketball, thereby providing a common starting point for future research in basketball. The
possession concept, in particular the concept of equal possessions for opponents in a game, is
central to basketball analysis. Estimates of possessions have existed for approximately two
decades, but the various formulas have sometimes created confusion. We hope that by showing
how most previous formulas are special cases of our more general formulation, we shed light on
the relationship between possessions and various statistics. Also, we hope that our new estimates
can provide a common basis for future possession estimation. In addition to listing data sources
for statistical research on basketball, we also discuss other concepts and methods, including
offensive and defensive ratings, plays, per-minute statistics, pace adjustments, true shooting
percentage, effective field goal percentage, rebound rates, Four Factors, plus/minus statistics,
counterpart statistics, linear weights metrics, individual possession usage, individual efficiency,
Pythagorean method, and Bell Curve method. This list is not an exhaustive list of methodologies
used in the field, but we believe that they provide a set of tools that fit within the possession
framework and form the basis of common conversations on statistical research in basketball.
Author Notes: We would like to thank two anonymous referees, Javier Gonzalez, Thomas Ryan,
Steven Schran, and the APBRmetrics community for years of commentary and critique of the
ideas contained in this paper. We also would like to make it clear that this paper represents the
views of the authors alone and not the organizations that the authors are associated with.
1. Introduction
1
Oliver (2004), Hollinger (2003, 2004, and 2005), and Berri et al. (2006) all use possessions as
the foundation for their analyses of basketball statistics.
A possession starts when one team gains control (or possession) of the basketball
and ends when that team gives up control of the basketball. Teams can give up
possession of the basketball in several ways, including (1) made field goals or free
throws that lead to the other team taking the ball out of bounds, (2) defensive
rebounds, and (3) turnovers. Note that under this definition of a possession, an
offensive rebound does not start a new possession; an offensive rebound starts a
new play.
Possessions are guaranteed to be approximately the same for two teams in a
game (within two for a non-overtime game), so possessions provide a useful basis
for evaluating the efficiency of teams and individuals. To win, teams and
individuals try to score more points per possession than their opponents.
Possessions are analogous to outs in baseball, where baseball teams typically have
27 outs to outscore their opponents. Given the centrality of the concept of a
possession, it is surprising that this is not an officially tracked statistic in most
basketball games.2 However, possessions can be counted using play-by-play
game logs.3
Possessions can also be estimated using commonly available box score data.
A general formula to estimate possessions for team t (POSSt) is:
2
Possessions have been officially tracked in the Women’s National Basketball Association
(WNBA) since 2004.
3
No possession is counted at the end of a period when there are less than or equal to four seconds
left and there are no field goal attempts, free throw attempts, or turnovers.
4
TOt includes team turnovers, such as five second or 24 second violations that are not credited to
any individual. These are not always reported and average about 0.666 per team per game. For
readers using data from Doug Steele’s site (see Section 4), where team turnovers are not included,
0.666 should be added to turnovers before applying the formulas in Tables 1 and 2.
5
Free throws that end possessions do not include first free throws of two, first and second free
throws of three, or free throws due to technicals, flagrant fouls, and clear path fouls. Also, free
DOI: 10.2202/1559-0410.1070 2
Equation (1) recognizes that each turnover, made field goal, and made possession-
ending free throw constitutes a possession, i.e. has a possession value of one.
Missed field goal attempts and missed possession-ending free throw attempts
share credit for the possession with defensive rebounds. Missed field goal
attempts and missed possession-ending free throw attempts get an D share of the
possession, while the defensive rebound gets a 1 – D share. Offensive rebounds
undo missed field goal attempts and missed possession-ending free throw
attempts, so their possession value is -D.
One of the most common (and simplest) formulations of (1) is to assume that D
= 1 and O = 0.44, which results in the following formula for possessions:
It is not possible to perfectly predict possessions using box score data, because,
among other reasons, possession-ending free throws are not identified in box
scores, some quarters end with offensive rebounds without follow-up shots, and
not all rebounds are attributed to individuals in box scores.7 Some rebounds, in
particular missed shots that go out of bounds (often blocked shots), are recorded
throws after made field goals are not (double) counted, since the made field goal already has
counted that possession. In the 2002-03 through 2005-06 seasons, 43.8% of free throws were
possession ending free throws.
6
The assumption of D = 0 results in POSSt = FGMt + 0.44 × FTMt + DREBo + TOt.
7
In addition, possession-ending free throws are not identified in box scores, so a fraction of these
free throws (O) are assumed to be possession-ending. Finally, some end of quarter situations, such
as an offensive rebound right as the quarter ends, results in no possession being estimated. In a
way this is a team turnover, but that is not how this is recorded in play-by-play logs.
as team rebounds.8 For this reason we estimate (3) without imposing all of the
restrictions in (1); this more flexible form allows us to capture relationships
between these variables and other factors, such as team rebounds, that we do not
include in the model.
We estimate (3) for both teams’ possessions in a game averaged together,
because we find that the best predictor of a particular team’s number of
possessions is the average estimated using both teams rather than the estimated
number for just that specific team.9
Table 1
OLS Regressions Predicting Possessions using Box Score Statistics
(1) (2)
Std Std
Variable Coeff Error Coeff Error
Field goals attempted 0.9640 0.0039 0.9492 0.0036
Field goals missed -0.3452 0.0086 -- --
Free throws attempted 0.4637 0.0034 0.4437 0.0030
Free throws missed -0.2073 0.0098 -- --
Offensive rebounds -0.6227 0.0100 -0.9599 0.0074
Defensive rebounds (for opponent) 0.3643 0.0086 -- --
Turnovers 0.9767 0.0053 0.9550 0.0061
Constant 3.2258 0.5200 3.8810 0.6070
R2 0.9615 0.9473
Number of observations 5,178
Data are from the 2002-03 through 2005-06 seasons.
In specification (1) of Table 1, each field goal made is a field goal attempted,
so the possession value is estimated to be about 0.96 possessions. Each missed
field goal is both a field goal attempt and a field goal miss, so we add those two
coefficients to estimate that each missed field goal has a possession value of about
0.62 possessions. Missed field goals have a lower possession value than made
field goals, because made field goals almost always lead to a change of
8
A team rebound is recorded when a player misses a non-possession-ending free throw, i.e. one
that cannot lead to an individual rebound. These rebounds are not differentiated in box scores
from other team rebounds; also offensive and defensive team rebounds are not differentiated.
9
Using equation (2), the correlation of a team’s actual possessions with its own estimated
possessions is 0.9493. Using the averaged possessions from both teams, the correlation rises to
0.9726.
DOI: 10.2202/1559-0410.1070 4
Table 2
Correlations and Means for Various Possession Formulas
Possession Formula Correlation Mean
(1) Actual possessions 1.0000 91.67
(2) Possessions from specification (1) of Table 1 0.9806 91.67
(3) Possessions from specification (2) of Table 1 0.9733 91.67
(4) FGAt + 0.44 × FTAt – OREBt + TOt 0.9729 93.88
(5) FGAt + 0.44 × FTAt – OREBt + TOt, own team 0.9488 93.88
(6) FGAt + 0.5 × FTAt – OREBt + TOt 0.9727 95.40
(7) FGAt + 0.4 × FTAt – 1.07 × OREBMisst + TOt 0.9766 91.28
Data are from 5,178 games from the 2002-03 through 2005-06 seasons. Possessions are averaged
across both teams in a game, except for row (5). Correlation gives the correlation with actual
possessions. OREBMisst = OREBt × (FGAt – FGMt) ÷ (OREBt + DREBo).
Given the definition of possessions, the next step is to use them to evaluate
efficiency. The common term for per-possession efficiency is a “rating.”
Specifically, points scored and allowed per 100 possessions have been called
offensive rating and defensive rating, respectively, for team t.15
The traditional measures of team offense and defense – points scored and
allowed per game, respectively – reflect both how efficient a team is and the pace
at which it plays. Given that, in any game, the number of possessions is dictated
13
One exception is the formula using the offensive rebounding percentage of missed shots from
Oliver (2004), which slightly underestimates possessions. That formula also is more highly
correlated with actual possessions than the “simpler” formulas, which should not be that surprising
since it uses more information, in particular information of missed shots and defensive rebounds.
14
Note that this formula is calibrated for the current NBA. Adjustments would likely have to be
made for earlier periods in the NBA and for basketball at other levels. Also, note that a “really
simple” formula (for the current NBA) is POSSt = FGAt + 0.5 × FTAt – OREBt + TOt – 4. This
formula (that can be computed without a calculator) predicts possessions nearly as well as the
more complicated formulas.
15
The difference between offensive ratings and defensive ratings is often referred to as net
efficiency ratings.
DOI: 10.2202/1559-0410.1070 6
Figure 1
Offensive and Defensive Ratings in 2005-06
Figure 2 shows how the league average ratings have changed historically.
League ratings were quite low in the late 1970s, but increased steadily. In the late
1980s and early 1990s, league ratings were at their highest point, but between
1994-95 and 1998-99 (a lockout shortened season), league ratings dropped by
about six points per 100 possessions. Since then the league rating has recovered
about two thirds of the previous decline.
Figure 2
Average League Ratings from 1973-74 to 2005-06 Seasons
In Oliver (2004), the term floor percentage is also introduced as the percentage
of possessions on which at least one point is scored. This floor percentage then
can serve as the characteristic probability in a binomial distribution describing the
sequence of scores on a court. There can be issues of independence, as mentioned
in Oliver (1991).
Plays
Plays are similar to possessions, except that offensive rebounds constitute a new
play. A team can shoot, miss, rebound its own shot many times, resulting in
several plays for one team without a corresponding play for their opponent.
Hence, plays are not approximately equal for two teams in a game and thus are
not as useful as a basis for evaluating team efficiency.
Note that because of the lack of standardization of terminology, sometimes
plays are referred to as possessions or minor possessions (with possessions also
called major possessions). In fact, early work by longtime University of North
Carolina basketball coach Dean Smith (also working with Frank McGuire, his
predecessor there) about 50 years ago appeared to refer to plays as possessions.
DOI: 10.2202/1559-0410.1070 8
Due to the obvious potential for confusion, the authors do not recommend using
the term possessions when speaking about plays.
The number of plays can be counted, but it is often estimated as:
Per-Minute Statistics
Another important breakthrough for analysis of the NBA was finding that
statistics calculated on a per-minute basis tend to be fairly consistent even when a
player's minutes played are variable. This allows for direct comparisons of starters
and reserves who play fewer minutes (per-minute statistics become unreliable for
players who have played very few minutes; generally, 500 or 1,000 minutes
played in an NBA season is used as a cut-off point). Sometimes, per-minute
statistics are referred to as rates; the scoring rate for player p, for example, is
points scored per 40 minutes (PTS40p):
where PTSp is points for player p and MINp is minutes for player p.
The use of per-minute statistics allowed analysts to identify young players like
Andrei Kirilenko, Michael Redd and Zach Randolph as candidates to break out
when they had the opportunity to play more minutes. Hollinger (2003, 2004,
2005) has used a basis of 40 minutes despite NBA games being 48 minutes long.
Because college and international games are 40 minutes long and NBA players
rarely play more than 40 minutes per game, this basis can have its benefits for
consistency across different leagues. For that reason we have a slight preference
for using 40 minutes as a basis, although the number of minutes used as a basis
does not alter the effectiveness of per-minute statistics. We will use a 40 minute
basis for consistency throughout the rest of this article.
Pace Adjustments
A team that averages 100 possessions per game gives their team 25 percent more
chances to shoot, assist, rebound, etc. than a team that averages 80 possessions
per game. Hence, many basketball analysts pace adjust their team and player
statistics.
Field goal percentage (FG%) does not account for three pointers or free throws,
so two common alternatives have been developed: effective field goal percentage
(eFG%) and true shooting percentage (TS%). These can be measured at the
individual or team level.
Effective field goal percentage accounts for made three pointers (3PM), whereas
true shooting percentage accounts for both three pointers and free throws. True
shooting percentage provides a measure of total efficiency in scoring attempts,
while effective field goal percentage isolates a player’s (or team’s) shooting
efficiency from the field. Both measures are appropriate for different situations;
the authors would not advocate one or the other for exclusive use. Over the 1996-
97 through 2005-06 seasons, means (and standard deviations) for the three
measures, measured at the player level, are as follows.17
16
Points per shot is used at times and is analogous to TS%, with a formula of PTS/(FGA + 0.44 ×
FTA).
17
These shooting percentage measures are weighted by field goal attempts.
DOI: 10.2202/1559-0410.1070 10
Rebound Rate
Per-minute rebound statistics are affected not only by the pace at which different
teams play, but also their ability to make shots and force opponents to miss. As a
result, rebounding is best evaluated by the percentage of all missed shots a player
rebounds when they are in the game. (This is usually estimated by their team’s
rebounds per minute added to opponent rebounds per minute and multiplied by
the player's minutes.) This is known as rebound rate or rebound percentage for
player p (REB%p).
§ REB p · § MIN p ·
(13) REB% p ¨ ¸ ¨ ¸ ,
© REBt REBo ¹ © MIN t ¹
where REBp is rebounds for player p, REBt is rebounds for team t, and REBo is
rebounds for the opponents o of team t, MINp is minutes for player p, and MINt is
minutes for team t. Player rebounding percentage can also be split into offensive
rebounding percentage (OREB%p) and defensive rebounding percentage
(DREB%p), which can prove insightful because few players are equally adept at
rebounding on both ends of the court.
§ OREB p · § MIN p ·
14 OREB% p ¨ ¸ ¨ ¸
© OREBt DREBo ¹ © MIN t ¹
§ DREB p · § MIN p ·
15 DREB% p ¨ ¸ ¨ ¸
© OREBo DREBt ¹ © MIN t ¹
11
OREBt
16 OREB%t
OREBt DREBo
DREBt
17 DREB%t
OREBo DREBt
OREB%t DREB%t
18 REB%t
2
At the team level, there is a negative relationship between offensive and defensive
rebounding percentages (over the 1996-97 through 2005-06 seasons the
correlation is -0.31), probably because offensive rebounding depends heavily on
whether the coach chooses to crash the boards or play back to prevent fast breaks.
The average offensive and defensive rebounding percentages are 31.2 and 68.8
percent, respectively, over the 1996-97 through 2005-06 seasons. Notice this is
not too different from the reverse of the coefficients on offensive and defensive
rebounds in specification (1) of Table 2. This is likely not a coincidence. A
missed shot leads to a loose ball that has about a 69 percent chance of becoming a
defensive rebound and thus ending the possession. Hence, the possession value of
0.62 for an offensive rebound is roughly consistent with the 69 percent chance
that it “saves” a possession. Similarly, the loose ball after the missed shot has a
31 percent chance of becoming an offensive rebound. So the possession value of
0.36 is roughly consistent with the 31 percent of the time that missed shots turn
into offensive rebounds. Defensive rebounds deny the offense the chance to
“save” a possession, so it is reasonable for the possession value of a defensive
rebound to be approximately the same as the probability of getting an offensive
rebound.
Four Factors
DOI: 10.2202/1559-0410.1070 12
negative factor because total offensive rebounds are highly correlated with
missed shots.
x Free throw rate (FTMt/FGAt): Dividing a team’s free throws made by field
goal attempts represents simultaneously the team’s ability to get to the foul
line and ability to make foul shots. Strictly, this term could be divided into
two terms, one representing how often a team gets to the foul line (relative to
shooting from the field) and the other representing how well they shoot from
the foul line. This would imply Five Factors, but this one term tends to
capture the most important elements of both.
Figure 3
Average League Four Factor Values from 1979-80 to 2005-06 Seasons
13
Plus/Minus Statistics
The National Hockey League has long tracked how well a player’s team does
while he is on the ice. This is called a “plus/minus statistic.” In the NBA,
tracking this statistic has been made possible only recently thanks to the
availability of play-by-play data on the Internet. In particular, 82games.com
began collecting and posting this information for many players beginning in 2003.
Prior to 82games.com, Harvey Pollack of the Philadelphia 76ers began collecting
this information in 1993-94 and published total plus/minus for each player in his
annual statistical handbook (Pollack, various years). Note that “plus/minus” is
often used interchangeably for somewhat different concepts, as described below.
DOI: 10.2202/1559-0410.1070 14
net plus/minus statistics. Starting lineups often play a large number of minutes
together, as do deep reserves.
Adjusted plus/minus statistics use regression techniques to account for the
players a given player plays with and against. First developed as WINVAL by
Jeff Sagarin and Wayne Winston, Rosenbaum (2004) outlines the general method
that relates team point differentials (or net efficiency ratings) to variables
indicating whether a player is in the game (and on the home or away team). The
coefficients on the player variables isolate the marginal contributions of those
players, accounting for the contributions of their teammates and opponents.
However, in general, it takes many games to get reasonably precise estimates for
these coefficients.
Counterpart Statistics
Counterpart statistics for player A represent the statistics posted by the opposing
team’s player at player A’s position. 82games.com has posted these statistics
since 2003-04. Because positions are often not clearly defined and because cross-
match-ups can occur, counterpart statistics have been considered difficult to
interpret. These dangers are greatest when counterpart statistics are estimated
from box score data. Estimating from play-by-play data is better, while using
observers to chart games is the preferred way to collect these statistics.
15
assessing the size of offensive roles played by players and how efficient they are
in those roles.
In addition, most basketball analysts argue that there is typically a tradeoff
between possession usage and efficiency. In fact, many analysts would argue that
this is the central optimization problem for a basketball team; Oliver (2004)
provides an empirical example of such an optimization. Players use up their easy
opportunities to score on dunks, lay-ups, and wide open shots. However, as they
increase their possession usage beyond those shots (and assists), the quality of the
opportunities fall. But they fall at different rates for different players. Moreover,
teams adjust their defensive strategies to reduce the efficiency of the players most
likely to be able to shoulder a higher possession rate. All of these factors have
resulted in it being difficult to obtain conclusive evidence on the negative
relationship between possession usage and efficiency (at least in publicly
available studies). What is necessary is a credible instrument (something related
to possession usage, but with no independent effect on efficiency) to isolate the
relationship between possession usage and efficiency.
Linear Weights
The term linear weights represents a class of player valuation methods that are a
weighted sum of player statistics. Many different forms exist, most of which
were developed following logical approaches. The most basic form is that of the
NBA’s efficiency statistic (NBA_EFFp), used on the NBA Web site (NBA.com,
2007), but also used by other practicioners:
Other linear weights measures apply different weights to each of these terms;
also, a player’s rebounds (REBp) can be broken down into offensive and defensive
rebounds, each with different weights. A summary of many different linear
weights formulations can be found in Oliver (2004).
As pointed out by several authors, including Oliver (2004) and Berri et al.
(2006), linear weights have numerous faults, including the frequently subjective
weights applied to the statistics, the lack of defensive statistics available for such
systems, the lack of correlation with winning at the team level, the theoretical
difficulty in incorporating new statistics that may be developed, and the general
lack of a measuring stick to calibrate their accuracy.
Recent work by Berri et al. (2006) assumes the externalities that players
provide their teammates are not important, which allows them to develop a linear
DOI: 10.2202/1559-0410.1070 16
weights formulation by first determining the relationship between wins and team
statistics and then assuming a similar relationship between wins and individual
statistics. Rosenbaum (2004) uses regression to develop a linear weights
approximation for his adjusted plus/minus estimates of player value.
Besides these methods, the most prominent linear weights techniques include
Player Efficiency Rating (PER) from John Hollinger (2003, 2004, and 2005),
TENDEX from Dave Heeren (1988), and Points Created from Bob Bellotti
(1994).
PTStx
20 PYTH t .
PTStx PTSox
where the subscript t still indicates team, the subscript o indicates opponent, and
the superscript x is an exponent that is empirically determined. In baseball, James
(1985, e.g.) found empirically that x = 2, leading to the use of the term
Pythagorean. Subsequent rigorous distributional work, especially by Miller
(2006), found that a slightly lower exponent of around 1.8 works best. In NBA
basketball, the value of x has been estimated empirically from different works to
be between about 13 and 17 (see Oliver, 1996, CoolStandings, 2006, for
example). The value varies both by the era and by how important the estimator
has seen capturing the extremes. In particular, margins of victory have changed
little over the last 30 years – even as the pace of the NBA slowed. This has
resulted in smaller exponents being necessary to correlate points to winning
percentage. But capturing the very good and very bad teams, of which there are
only a few each season, is done much better with a larger exponent. Typical
least-squares-type error estimates of the exponent tend to yield a smaller exponent
because so many teams are bunched between about 30 and 50 wins (winning
percentages between about 38% and 62%). But increasing the exponent to better
capture very good and very bad teams does not significantly compromise those
teams in the middle and, as such, Oliver has been joined by ESPN (2007) in the
use of an exponent of 16.5.
In other leagues, including college, high school and the WNBA, smaller
exponents have been found to work better because there are fewer possessions in
17
the typical game in these leagues (see Pomeroy, 2006, for example). Also, note
that because of the equality of possessions for a team and its opponents, PTSt and
PTSo can be replaced in the above equation by ORtgt and DRtgt.
19
This is the NORMSDIST function in MS Excel. Note that individual game data on net points is
necessary to apply the Bell Curve method. Note also that, because ties are not allowed in
basketball, the denominator of the function above (representing the standard deviation of the
team’s net points) incorporates extra noise as additional variance. This causes the method to be
slightly biased towards 0.500 a little bit (about one to two percentage points) for teams that are far
from 0.500.
DOI: 10.2202/1559-0410.1070 18
It also has additional accuracy, but this difference is very slight and the simplicity
of the Pythagorean method can often make it preferred.
The Web provides an array of sites with basketball data. Most people know about
Web sites such as NBA.com (http://www.nba.com), ESPN.com
(http://www.espn.com), and Yahoo! Sports (http://sports.yahoo.com). These sites
provide daily coverage of the NBA, and their sites are constantly updated during
the season with contemporary statistics, rosters, and schedules. Other than the big
media sites, here are some other sites that we frequently use for historical and
other contemporary information:
82games.com (http://www.82games.com): This site, which is run by Roland
Beech, provides both modern analysis of the NBA using some of the techniques
above, but also some data resources. The site uses detailed play-by-play data
from 2002-03 through the (regularly updated) current season to examine questions
such as (a) how the Lakers perform when Kobe Bryant is on the court as opposed
to when he is off the court, (b) how opposing centers have performed when Yao
Ming is on the court, (c) how teams shoot from different parts of the court or at
different stages of the shot clock, (d) how teams have fared with different line-ups
on the court, (e) how players have performed in clutch time, and much, much
more. These play-by-play data fill in many of the gaps left by traditional box
score statistics. The original play-by-play data is not available publicly on the
82games site, but in the past sufficiently original and interesting research plans
have resulted in 82games providing some of its data to researchers. Beech takes
requests from readers for “research projects,” providing them with data in return
for articles written for the site. This can be a useful way to start research projects.
APBRmetrics Forum (http://sonicscentral.com/apbrmetrics/): This discussion
board is not a data source, per se, but it is an excellent resource for anyone
interested in serious discussions about the statistical analysis of basketball. The
site is run by Kevin Pelton, and is a descendant of the APBR_analysis group on
Yahoo! Groups.
Basketball-Reference.com (http://www.basketball-reference.com): An on-line
basketball encyclopedia run by Justin Kubatko, this site is practically a one-stop
shop for historical basketball statistics. This site includes regular season data for
players, coaches and teams for the entire history of the NBA. Data from the
playoffs and individual game box score data are available for an increasing
number of years. The site is also rich with biographical data and even includes
college statistics and salary information for many players. Also, much of the
work by analysts such as John Hollinger and Dean Oliver has been incorporated
19
into the site. Finally, the site allows for queries of specific information, which
can be very useful for research purposes.
DatabaseBasketball.com (http://www.databasebasketball.com): This bare
bones site provides zip files for download of NBA season stats for players and
teams, including playoff data. It also has information on coaches. Some of this
information is incorrect (so research using these data should be done with care),
but these files provide a reasonable starting point for a personal database.
Doug’s MLB and NBA Statistics (http://www.dougstats.com): Doug Steele’s
site provides current season data in a form that is easily imported into most
spreadsheet applications and is updated on a daily basis. Doug also provides
historical NBA statistics dating back to the 1988-89 season.
Kenpom.com (http://www.kenpom.com): Ken Pomeroy’s site provides data
and analysis for college basketball. The site includes (among many other things)
advanced statistics for both teams and players.
Rodney Fort’s Sport Business Data page
(http://www.rodneyfort.com/SportsData/BizFrame.htm): Fort has collected a
treasure trove of historical financial, attendance, and winning percentage data for
the NFL, NBA, NBA, and MLB.
5. Conclusion
The quantitative analysis of sports is a new branch of science and, in many ways
one that has grown through non-academic and non-traditionally peer-reviewed
work. The aim of this paper is to bring to a peer-reviewed journal the generally
accepted basics of the analysis of basketball, thereby providing a common starting
point for future research in basketball.
The possession concept is central to this work. The concept of equal
possessions for opponents in a game has been in the statistical community for
approximately 20 years and has withstood substantial review and commentary.
Possession estimates have existed for approximately two decades, but the various
formulas have sometimes created confusion. We hope that by showing how most
previous formulas are special cases of our more general formulation, we shed
light on the relationship between possessions and various statistics. Also, we
hope that our new estimates can provide a common basis for future possession
estimation. Regardless, we suggest that the possession framework be considered
in future basketball research, as it continues to get review from “peers” now
working in the NBA.
The other concepts and methods discussed above are by no means a
comprehensive list of methodologies used in the field or even by the authors, but
we believe that they provide a set of tools that fit within the possession framework
and form the basis of common conversations on statistical research in basketball.
DOI: 10.2202/1559-0410.1070 20
References
Bellotti, Bob. The Points Created Pro Basketball Book of 1993-94. 1994.
Nightwork Publishing, New Brunswick, NJ.
Berri, David J., Martin B. Schmidt, and Stacey L. Brook. The Wages of Wins:
Taking Measure of the Many Myths in Modern Sport. 2006. Stanford
University Press, Stanford, CA: Stanford University Press,.
CoolStandings. “Do You Use the Bill James Pythagorean Theorem for Sports
other than Baseball?” 2007.
http://www.coolstandings.com/football/faq.asp?sn=2006#faq14, available
as recently as 1/22/07.
ESPN.com. “2006-07 NBA Expected Winning Percentage.” 2007.
http://sports.espn.go.com/nba/stats/rpi, available as recently as January 15,
2007.
Heeren, Dave. The Basketball Abstract. 1988. Prentice Hall, Englewood Cliffs,
NJ, Prentice Hall.
Hollinger, John. Pro Basketball Forecast: 2004-05 Edition. 2004. Brassey’s,
Washington, DC: Brassey’s.
-----. Pro Basketball Forecast: 2005-06 Edition. 2005. Brassey’s, Washington,
DC: Brassey’s, 2005.
-----. Pro Basketball Prospectus: All-New 2003-04 Edition. 2003. Brassey’s,
Washington, DC: Brassey’s.
James, Bill. The 1985 Baseball Abstract. Ballantine Books, 1985.
Küpfer, Ed. “Team Similarity.” 2005. APBRMetrics forum,
http://www.sonicscentral.com/apbrmetrics/viewtopic.php?p=118&sid=ff9
25dbe120f6c6f86e596d0d6dabbb8.
McGuire, Frank. Defensive Basketball. 1959. Prentice Hall, Englewood Cliffs,
NJ.
Miller, Stephen J. “A Derivation of the Pythagorean Won-Loss formula in
Baseball,” Chance Magazine, forthcoming.
NBA.com. “Efficiency Statistics.”
http://www.nba.com/statistics/efficiency.html, available as recently as
January 19, 2007.
Oliver, Dean. Basketball on Paper. 2004. Brassey’s, Washington, DC.
-----. “Established Methods.” 1996. Journal of Basketball Studies,
http://www.rawbw.com/~deano/methdesc.html#pyth.
-----. “New Measurement Techniques and a Binomial Model of the Game of
Basketball.” 1991. Journal of Basketball Studies,
http://www.rawbw.com/~deano/articles/bbalpyth.html.
Pollack, Harvey. Harvey Pollack’s NBA Statistical Yearbook. Various years.
21
DOI: 10.2202/1559-0410.1070 22