A CLUSTER ANALYSIS OF NBA PLAYERS

DWIGHT LUTZ
[email protected]
303-229-8959

Abstract. To revolutionize the concept of basketball positions, we use a multivariate cluster analysis to
group NBA players based on several statistics. Once these clusters have been formed, we analyze how each
cluster affects winning in the regular season. Along with the main effect of each cluster, we analyze the 2 and
3-way interactions of the clusters. The ultimate goal is to determine which types of players are most relevant
to winning, and also what combinations of players affect winning. This will give us a way to scientifically
measure team chemistry.

1. Introduction
Every NBA player is given a label that is meant to describe what that player should do on the court. There
are 5 of these labels and we call them positions. These labels are decided first by the players physical size. If
hes big and strong, then he is going to be a power forward or center. If hes small and fast then he is labeled
as some sort of guard. From there, a general view of the players skills helps to sort him into one of the 5
specific positions. If hes a guard and shoots it well, hes a shooting guard. If he handles the ball and passes
well, hes a point guard.
The point is, this process is far from scientific even though it has a big impact on coaches decisions with
regards to playing time, and general managers approaches to acquiring new players. This paper aims to use a
multivariate cluster analysis to redefine the positional labels that are given to NBA players. We will abandon
the traditional method of using a players height and weight, since these factors can be terribly misleading.
Certain small players are able to be effective inside the paint on both ends of the floor. Similarly, plenty of
big men have success on the perimeter. This paper will focus less on who the player is and more on what
that player actually does on a basketball court. Several statistics are used in an effort to capture the style of
play for each individual, and we put very little emphasis on how effective the player is. For example, you will
find Dwight Howard in the same cluster as Spencer Hawes. The two players have similar styles, not similar
effectiveness.
We will develop the process that was used to form the clusters and give a specific description for each
cluster. A specific number of clusters was not forced. However, intuition can quickly tell you that 5 positions
seems to be far too few to describe all the different styles of play that exist on a basketball court. Conversely,
using a large number of clusters will overfit the clusters to individuals, and would make analyzing differences
between clusters difficult due to small sample sizes. One of the key focuses of this paper will be determining
which types of players are targeted by winning teams and which types of players are more often found on
losing teams. It is quite possible that a team should choose an inferior player in order to meet a certain need.
Of course, this thought process is nothing new, but this paper will use a much more scientific method to
determine the types of players that exist, and the relevance of those players to a team.
The first part the paper will focus on individual positions and their effect on winning. The second part
will consider different combinations of the clusters that we will develop. Basketball may be the hardest game
to evaluate individual players since each players performance is greatly affected by the other players on the
court. Just as we will determine which individual positions relate to winning, we will try to understand what
combinations of positions are most often found on winning teams. This type of analysis can give a general
manager a tangible idea of how well a certain player will fit on a new team. Obviously, team chemistry goes
far beyond the statistics that will be considered in this paper, but our methods will be an evolution of the

current thought process that NBA executives use to evaluate the compatibility of an available player with the
team they are operating.
2. Cluster Analysis
In order to construct clusters that represent playing style we needed to use variables of the same variety.
The following variables were used: games played, minutes played per game, percent of made field goals that
are assisted, assist rate, turnover rate, offensive rebound rate, defensive rebound rate, steals per 40 minutes,
blocks per 40 minutes, and the number of shots attempted per 40 minutes at each of the following locations:
at the rim, from 3-9 feet, from 10-15 feet, from 16-23 feet, and beyond the 3-point line (see Table 1).
Since many of these variables are measured on different scales, we standardized all the variables using
z-scores in order to put them all on the same scale, and thus gave an equal weight to each variable. Next,
we experimented with several different clustering algorithms including the classical k-means [1] method and
Wards hierarchical clustering [2]. The method that we ultimately decided to use was the EM algorithm for
Gaussian mixture models [1]. The EM algorithm has proven to be extremely effective in maximizing the
separation of the clusters and has become the modern form of cluster analysis [1] [3].
The cluster analysis was performed on all players from the 2010-2011 season having played in at least 30
games and averaging at least 10 minutes a game. This left us with 329 players with which to build the clusters.
All data was collected from hoopdata.com [4]. The Mclust function was used in the R Project for Statistical
Computing, and this function uses the Bayesian Information Criterion to determine the parameters of the
model and how many clusters to use [3]. This is how we settled on using 10 clusters. The most important
trait of the cluster solution is the separation of the clusters, which we examine graphically in Figure 1. All
variables seem to be showing strong variability from cluster to cluster, which mathematically validates the
use of each variable. If there was a variable that remained relatively constant across clusters, then we would
consider removing it. In order to increase the sample size, we also considered players from previous seasons,
but we needed put these players into one of the clusters that we had just created. For this we used Fishers
Linear Discriminant [2] to place the players from the 2008-2009 and 2009-2010 seasons into one of the 10
clusters. Discriminant analysis is a classification technique used to determine group membership of a new
set of observations. From this group, there were 646 players that met the games and minutes played criteria
described above, giving us a total of 975 observations. When using Fishers Linear Discriminant to predict
cluster membership, we had a misclassification rate of about 16%. Moving forward, we can expect 84% of
players from the 08-09 and 09-10 seasons to be classified into the correct cluster. Although there is a bit of
error here which could potentially be improved upon with techniques such as the Support Vector Machine, we
are rather pleased with accuracy of our cluster predictions.
We left out many details regarding clustering with EM, Fishers Linear Discriminant, and the Mclust function. For more insight on the mathematical theory behind these topics, see Bishop [1], Izenman [2], and Fraley
and Raferty [3] respectively.
Before we begin to talk about which clusters affect winning, it is important to realize the relevance of the
clusters that we have created. One could look at the names of the clusters that I have given in Table 2 and
begin relate them to the 5 traditional positions. When doing this, you might be surprised by which clusters
certain players fall into. For instance, I labeled one cluster as Ball Handlers since that cluster had high
assists, high turnovers, very few assisted field goals, high steals, low rebounds, and most of their shots are from
the perimeter. Those characteristics describe a typical point guard. Uninterestingly, many of the leagues point
guards find themselves in that cluster. However, Hedo Turkoglu (on the Magic only) is also in that cluster.
All of Hedos stats agree with the description of this cluster given above (except for the steals). Hedo was constantly creating his own shot, and doing so for others. Whether it was effective or not, Turkoglu was playing
point guard, or as weve labeled it, Ball Handler, for the Orlando Magic last season. Another interesting
result is that neither of the two Laker point guards, Steve Blake or Derek Fisher, find themselves in the Ball
Handler cluster, they are both in the Combo Guard cluster. The members of this cluster score mostly on
assisted field goals, but rarely turn the ball over. It has been well documented that Phil Jacksons triangle
offense requires less need for a traditional point guard than a usual NBA offense. It is quite possible that Steve
2

Blake, a more traditional point guard for most of his career, struggled last season because he was being forced
to play a position he was not used to playing. It would be hard to see this phenomenon from the traditional
perspective we have of NBA positions, since he played point guard both before the Lakers and with the Lakers.
These are just two examples of many situations where this analysis allows GMs and coaches to more efficiently acquire and play the correct players for their team. Could the Magic have gone stretches of games
or the season with the 610 Turkoglu at point guard? Coach Stan Van Gundy would have been called crazy
for trying it, but Hedo was already doing it. Turkoglus style of play is evidence that there would have been
a smooth transition if the Magic had no other point guard on the floor. This does not imply that the Magic
should have dumped all other point guards to let Turkoglu play the position, it simply means that Turkoglus
effect on the game most likely would not have changed if he was the only point guard. Recall, the result is
not that Hedo is an effective point guard, it is that his style of play resembles that of point guard, which is
why he was categorized as one.
For more detail on how each cluster is defined, see Table 3 which gives the average z-score for each statistic
for each cluster. Also note some typical members of each cluster in Table 2.
3. The individual effects of the clusters on winning and losing
This section of the paper is our best effort to determine which types of players on a basketball court are
most important to a teams overall success. Before we begin, I want to reemphasize that we are not accounting for the quality of the players. Two players from the same cluster might be vastly different in terms of
effectiveness, but they will always be similar in style. When we label a player as a shooter that means he
shoots a lot; it does not necessarily mean that he makes a lot of shots. The better players will always produce
more wins, but we will show you which types players seem to be found more often on the winning franchises
on average.
First we must define winning and losing. We did not use wins, but instead used average point differential
per game for all 90 teams from the last three regular seasons. The first method was to analyze the effect of
the clusters on point differential as a one-way Analysis of Variance (ANOVA) [5] with ten levels of the cluster
factor and each of the teams being the 90 observations. From this we got a p-value of .0111 for the F-test. This
tells us that there is a significant cluster effect on team point differential. Now we investigate which clusters
are the most relevant. We calculated all pairwise 95% confidence intervals for the differences between the
cluster means using Bonferronis correction. Unfortunately, zero was in all of these 95% confidence intervals,
so we could not determine that point differential was significantly greater for any one cluster than another.
However, we did compute the mean point differential for each cluster and those results agree with the more
interesting analysis that we give below.
Next, we considered only the winning teams, or those with positive point differential. From the pool of
winning teams, we wanted to find out which types of players were most common. Note that each cluster does
not have the same number of players; the number of players per cluster is given in Table 4. Many inferences
can be made from the varying sizes of the clusters. For example, there is an abundance of Ball Handlers (8)
but the Durable Shooters are quite scarce. The Durable Shooters (10) shoot many 3s and some long 2s,
but also have high steals and play an incredibly large amount of games, thus we describe this group as durable.
This is an area where general managers can find potential value. If a team understands how abundant Ball
Handlers are, that team is less likely to overpay for one. However, it is important to understand how much
each position effects winning in order to understand value. It is quite possible that Ball Handlers are a critical
piece to the success of a franchise, in which case it makes sense to pay a high price for one.
So the question of interest is, of the players in each cluster, what percentage of them are on winning teams?
Table 4 gives these percentages ordered from highest to lowest. Durable Shooters (10) are the type of player
that most often are members of winning organizations as 66.7% of these players are a winning team. So the
Durable Shooters are the most rare breed of player and they are the most owned player by winning teams.
These two aspects together make Durable Shooters extremely valuable. How about the Ball Handlers we
mentioned earlier? Only 44% of these players find themselves on winning teams. When you combine that
3

with how large of a group they are, it seems that Ball Handlers are rather expendable.
If a certain type of player had no effect on a teams point differential, then we would expect 50% of the
players in that cluster to be on winning teams. So we should compare each percentage in Table 4 to 50%.
Actually, since only 48% of all players were on winning teams, we should compare each percentage to 48%.
For whatever reason, there were more qualifying players on the losing teams. An important question we need
to answer is, are the percentages of players on winning teams for each cluster significantly different than 50%?
By significantly, we mean in the statistical sense of the term. To determine this, we did one-sample T-tests
for each cluster and the two-sided p-values for these tests are given in Table 4.
We can conclude that there is strong evidence (p-value < .1) that the following clusters have a positive
impact on point differential: Durable Shooters (10) and Combo Guards (1), whereas these clusters have a
negative impact on point differential: Aggressive Bigs (7) and Perimeter Scorers (9). Also, there is some evidence (p-value < .15) that Ball Handlers (8) and Big Bodies (6) have a negative impact on point differential as
well. Although the other 4 clusters have a percentage of their players on winning teams that is a bit different
from 48%, these results are not statistically significant, and thus inconclusive. Collecting more data could
potentially allow us to determine if the results from these remaining clusters are meaningful.
Now let us consider these results from a basketball point of view. Even though we have strong mathematical
evidence, does it make sense that Durable Shooters (10) are stockpiled by winning teams? Lets recall how this
cluster is defined. This group plays a high amount of minutes and plays more games than any other cluster.
This type of consistency is sure to help a team to stay on balance throughout the course of a season. There
have been many cases of teams struggling for extended stretches because of injuries, but the Durable Shooters
simply do not get hurt. There is also something to be said about this clusters high amount of minutes played
per game. With long-distance shooters, coaches often find it difficult to play these types of players extended
minutes because of their lack of contribution on the defensive end and in other areas. However, this group
has convinced its coaches to leave them on the floor despite the fact that their most relevant trait is shooting
without creating their own shot (they have a high percent of assisted field goals made). This group also collects
a large number of steals which allows us to believe that these players have at least somewhat of an impact on
the defensive end of the floor. Most importantly, this groups turns the ball over less than any other cluster
but one. These are all traits common to a team that has an efficient offensive flow, and since these players are
competent defensively, it seems of little surprise that winning teams collect these types of players.
Now lets look at a cluster that is found more often on losing teams. The Aggressive Bigs (7) cluster has
the lowest percentage of its players on winning teams, a result that is quite significant given the small p-value.
This is group is mostly made up of power forwards and centers who rebound on both ends and block shots
extremely well. This group also like to be aggressive offensively. They take well more shots at the rim, from
3-9 feet, and from 10-15 feet than average. They even shoot some from 16-23 feet. So why do we find this type
of player on losing teams? The answer is not quite as clear as it was for Durable Shooter group. Consider this,
the Aggressive Bigs cluster is very similar to the Defensive Bigs (2) cluster, which has a positive (although statistically insignificant) percentage of players on winning teams. The main difference between the two clusters
is that the Aggressive Bigs take a lot more shots from 3-23 feet. Both clusters are composed mostly of bench
players as they both have low minutes played per game. Perhaps the members of the Aggressive Bigs cluster
are too aggressive offensively and tend to disrupt the teams offensive flow. Furthermore, the Big Bodies (6)
cluster is another similar group that also has a negative impact on point differential. Players in the Big Bodies
cluster play extremely few minutes per game, but tend to shoot many shots from 16-23 feet. This is even more
evidence that big men who come off the bench and are overaggressive offensively tend to find themselves on
losing teams. Whether they are losing because they are shooting too much, or they are shooting too much
because they are losing is an entirely different question, but the two traits are undoubtedly intertwined.

4. The effects of pairs and triples

As mentioned earlier, every player on a basketball court is greatly affected by the other players on his team.
Many people call basketball the ultimate team game since all 5 players need to have a certain basketball
4

chemistry in order to be successful. In the previous section, we discussed how individual types of players
impact a teams success, which is essentially the main effect of the clusters on winning. In this section, we
will focus on which combinations of clusters are most often seen on winning teams. In other words, we will
consider the interaction effects of the clusters on point differential. Ideally, we would not only consider what
pairs of clusters affect a teams point differential, but also which triples, groups of 4, 5, and even 6 clusters
are most often found on winning teams. For the sake of brevity, we will only cover pairs and triples in this
paper. After all, we saw 218 different three-way combinations on the different teams.
Although there are many possible strategies to analyze the interactions between clusters, we chose to maintain consistency and analyze the combinations in a similar fashion as the individual clusters, which is to
determine which combinations show up far more often on the winning teams than on losing teams. From
there, we did one-sample T-tests to determine if the percentage of each combination that was found on winning teams was significantly different from 48%.
Table 5 gives the ten 2-way combinations with the highest percentage of players on winning teams. Included
in the Table 5 are the p-values for the two-sided T-tests. Clearly, the results for the combinations are much
stronger than those for the individual clusters. Mathematically, this result is biased towards the performance
of the individual clusters. When considering a combination of two individual clusters that are very common on
winning teams, a few occurrences of that combination will appear on winning teams even if no true interaction
between the two clusters exists. However, from a basketball standpoint, it seems rather intuitive that two
players can have a greater marginal impact on a game than one player can. This concept is one of the reasons
that basketball analysts have begun using five-man performances alongside a players individual output. In
order to account for the bias, we included a column in the table called Increase. This column describes the
impact (positive or negative) that adding the worse performing cluster to a team (that already has a player
from the better performing cluster) has on the teams point differential. So adding a 10 to another 10 increases
a teams chance of having a positive point differential by 78.95 66.67 = 12.3%.
With regards to specific clusters, there are many conclusions to be had. The Durable Shooters (10) cluster
seems to interact well with several other clusters. We find this as no surprise given the breakdown of the
characteristics of that cluster above. The other cluster that is prominently featured in Table 5 is the Combo
Guards (1). This group spreads the floor by shooting from range and rarely creating their own shot in a similar
fashion to the Durable Guards, but they handle the ball more as they have high assists and high turnovers.
These two groups seem to interact well with each other. Both of these clusters play a high amount of games,
spread the floor, and rarely create their own shot. In other words, they can easily complement players of
any other style of play. The same is simply not true for other types of players. For example, the Perimeter
Scorers (9) group does not coexist well with many other clusters. In fact, this group is peppered throughout
Table 6, which are the 10 combinations most often found on losing teams. Furthermore, the combination of
two Perimeter Scorers has the lowest percentage of appearances on winning teams of any possible pair. Also,
the negative impact of adding a second Perimeter Scorer to a team is the greatest of any addition of a single
cluster to a team that already has one type of player. As seen in the Increase column of Table 6, adding a
second perimeter scorer (assuming the team already has one) drags the teams probability of having a positive
point differential down by about 19%.
Why is the Perimeter Scorers cluster so incompatible with others? This cluster takes a ton of shots. Note
in Table 3 that they take more shots than average from every distance. They rarely score assisted field goals
and essentially are one-on-one players. The argument has been made for years that a team needs to move
the ball on offense and cannot have players who stop the ball to go one-on-one. All of our research supports
this idea, and suggests that having two players like this is exponentially worse. A case in point of this is the
situation with Dwyane Wade and LeBron James of the Miami Heat. The struggle for these two to coexist has
been well documented, and it so happens that both players are in the Perimeter Scorers cluster. If Lebron
and Wade were equally talented to the average player in their cluster, we would have given them less than a
23% chance to have a positive point differential, holding the skill level of all other teammates constant. Since
LeBron and Wade are far above average, they were able to excel from the standpoint of winning, but their

early struggles would have been no surprise to this research.

Although there are many other conclusions we could make about the 2-way combinations, in fact there
were 36 combinations significantly different from 48% at the = 0.1 level, we will move on to the three-way
combinations, or triples. Table 7 gives the top ten triples that are most found on winning teams along with
their P-values for two-sided T-tests. Again, theres no surprise that Durable Shooters (10) and Combo Guards
(1) play a prominent role. However, notice that the Perimeter Scorers (9) are in every single one of these top
triples! We spent the last paragraph talking about how poorly they mesh with other clusters, but now they
seem more valuable than ever. The point is, when a team is able to put the correct players around a Perimeter
Scorer, the chances of having a positive point differential increase dramatically. The 3-way combination of
Perimeter Scorer (9), Durable Shooter (10), and Combo Guard (1) is found on winning teams 24 out of 25
times (Table 7), which seems far from an accident. From a basketball standpoint, it seems that you have
a guard willing to give the Perimeter Scorer the ball and then spot up. Then the Perimeter Scorer gets to
play one-on-one and dish off to the Combo Guard or Durable Shooter if the defense puts too much focus
on his one-on-one move. Remember, the Perimeter Scorers have the 4th highest assist rate of any cluster.
Another cluster thats worth mentioning here is the Defensive Bigs (2). This group performs consistently well
individually and in 2 and 3-way combinations. The interesting part of this is, this group shoots far less than
any other of the big men clusters. They interact extremely well with the offense-oriented clusters (Perimeter Scorers, Durable Shooters, even the Floor Spacers (4)). This result suggests that it makes sense to have
specialty players, or certain players on the team that excel on offense and other players that excel on defense.
Table 8 gives us the other end of the spectrum, or the triples most often found on teams with a negative
point differential. Its clear that the Aggressive Bigs (7) performs extremely poorly, and we discussed earlier
why this might be the case. Let us consider some of the clusters that are interacting with the Aggressive Bigs.
Notice that several of these clusters are comprised of big men. The Defensive Bigs (2), Elite Bigs (5), and Big
Bodies (6) are clearly all big men of different varieties. These results question the age old basketball theory
that a team can never have too much height. Several 3-way combinations made up of multiple big men often
find themselves on losing teams.
The other cluster that is prevalent in Table 8 is the Versatile Swingmen (3) group. This the epitome of a
the type of player that does many things well, but does not do one thing great. They rebound well from the
perimeter, they get a few assists and steals, even a few blocks. Most of their shots are assisted and happen
at the rim or outside the 3-point line. NBA analysts often call these types of players glue guys, and claim
that every winning team needs a player or two like this. However, having three Versatile Swingmen seems
to be rather detrimental, as only 4% of these triples are on teams with positive point differential. Also, this
group does not seem to combine well with big men that are aggressive offensively, as we see the (3,7) and (3,6)
popping up in both the losing triples (Table 8) and losing doubles (Table 6). These results show that we may
have a tendency to overrate the intangible qualities and perimeter defense that this type of player brings to
the table.
5. Conclusion
The results that have been described in this paper are only a subset of the information that can be gained
from this analysis. However, there are potential criticisms to this analysis, such as the thought that the quality
of the player is more important than the type of player. After all, we can all agree that the Miami Heat did
not have compatible parts, but they still won 58 games and had a +7.5 point differential. Also we recognize a
lack defensive statistics used in the cluster analysis. This was mainly due to data limitations. Also, our results
would be more accurate if we had not used a binary dependent variable to represent winning. Despite all of
this, we have been able find some very powerful trends with regards to how different types of players affect
winning. We cannot say that acquiring Durable Shooters is the undisputed way to win NBA championships,
but we cannot deny that their presence helps teams win games either. Our analysis of the combinations of
the clusters gives a new way to think about team chemistry and how the compatibility of a teams pieces will
affect winning on average. Overall, we feel that the development of these clusters produces a more scientific
way to think about the positions of individual players, and the overall structure of a teams roster.

0.5
0.0
1.5

1.0

0.5

Cluster Mean

1.0

1.5

2.0

6. Appendix: Figures and Tables

10

12

14

Variable

Figure 1. We are interested in the separation of the clusters for each variable. Each line
represents one of the 10 clusters, and we hope to see each cluster have a unique line such that
no two lines fall on top of each other.

Table 1. Variable Abbreviations and Formulas

Variable

Abbreviation

Formula (if needed)

Games Played (minimum of 30)

GP

Minutes played per game (minimum of 10)

Min

% Ast

made field goals that are assisted

total made field goals

AR

Assists100
FGA+(FTA.44)+Turnovers

Turnover Ratio

TOR

Turnovers100
FGA+(FTA.44)+Turnovers

Offensive Rebound Rate

ORR

100(Player ORebs(Team Min/5))

(Player Min(Team ORebs+Opp DRebs))

Defensive Rebound Rate

DRR

100(Player DRebs(Team Min/5))

(Player Min(Team DRebs+Opp ORebs))

Rim

3-9

Attempted field goals from 10-15 feet per 40 minutes

10-15

Attempted field goals from 16-23 feet per 40 minutes

16-23

3s

Steals per 40 minutes

Stls

Blocks per 40 minutes

Blks

Percent of made field goals that are assisted

Assist Ratio

Attempted field goals at the rim per 40 minutes

Attempted field goals from 3-9 feet per 40 minutes

3-point field goals attempted per 40 minutes

Table 2. Cluster labels and typical members

Cluster

Cluster Label

Combo Guards
Backup Bigs
Skilled Swingmen
Floor Spacers
Elite Bigs
Big Bodies
Aggressive Bigs
Ball Handlers
Perimeter Scorers
Versatile Shooters

1
2
3
4
5
6
7
8
9
10

Typical Members
Steve Blake / Mario Chalmers / Rudy Fernandez
DeAndre Jordan / Ben Wallace / Brendan Haywood
Shane Battier / Lamar Odom / Paul George
Channing Fry / Matt Bonner / Mike Miller
Amare Stoudamire / Elton Brand / Brook Lopez
Jason Collins / Antonio McDyess / Kurt Thomas
Brandon Bass / Carlos Boozer / Tyler Hansbrough
Chris Paul / Kyle Lowrie / Devin Harris
Rudy Gay / Dwyane Wade / Eric Gordeon
James Harden / Ray Allen / Nicolas Batum

Table 3. The values in the table represent the average z-score for each variable for the given
cluster. Recall, positive z-scores refer to above average values and negative z-scores refer to
below average values. The further the z-score is from 0, the further from the mean that
statistic is.
Cluster (#)
Combo Guards (1)
Defensive Bigs (2)
Versatile Swingmen (3)
Floor Spacers (4)
Elite Bigs (5)
Big Bodies (6)
Active Bigs (7)
Ball Handlers (8)
Perimeter Scorers (9)
Durable Shooters (10)

GP

Min

% Ast

AR

TOR

ORR

DRR

Rim

3-9

10-15

16-23

3s

Stls

Blks

0.38
-0.41
-0.49
-0.48
0.83
-0.55
-0.08
0.04
0.04
1.09

-0.02
-0.62
-0.41
-0.39
1.04
-1.06
-0.29
0.34
0.59
0.38

0.61
0.23
0.26
0.99
0.02
0.67
0.29
-1.53
-0.56
0.32

0.20
-0.45
0.09
-0.52
-0.47
-0.60
-0.90
1.69
-0.10
-0.21

-0.47
0.88
0.23
-0.82
-0.31
0.33
-0.19
0.94
-0.42
-0.82

-1.01
1.57
-0.08
-0.48
0.86
0.85
1.25
-0.85
-0.60
-0.58

-1.01
0.99
-0.16
-0.07
1.23
0.52
0.98
-0.95
-0.23
-0.53

-1.32
0.82
0.15
-0.89
0.85
-0.21
0.82
-0.12
0.16
-0.16

-0.84
-0.06
-0.30
-0.81
1.93
-0.30
0.67
-0.14
0.36
-0.39

-0.34
-0.94
-0.66
-0.73
1.03
-0.03
0.65
0.09
1.09
-0.28

0.17
-1.34
-0.66
-0.25
0.40
0.72
0.08
0.15
0.90
-0.02

1.05
-1.09
0.00
1.06
-0.95
-1.03
-1.08
0.37
0.21
0.88

-0.21
-0.36
0.67
-0.37
-0.20
-0.55
-0.77
0.75
-0.03
0.64

-0.78
1.11
-0.02
-0.31
0.80
0.22
1.29
-0.81
-0.34
-0.52

Table 4. Cluster counts and percentage of each cluster on winning teams

Cluster
# of Players
% on winning teams
p-value

10

57
66.67%
0.011

85
62.35%
0.022

110
54.55%
0.343

86
52.33%
0.669

120
45.83%
0.363

172
44.19%
0.128

84
44.05%
0.278

120
41.67%
0.068

64
40.63%
0.135

77
37.66%
0.029

Table 5. Top ten pairs most found on winning teams

Combination
10
1
2
3
1
1
1
9
5
1

Winning Count

10
10
10
10
1
2
5
10
10
4

15
32
42
37
22
77
57
32
38
62

Total

Percentage

Increase

P-value

19
42
56
52
31
113
87
49
59
100

78.95%
76.19%
75.00%
71.15%
70.97%
68.14%
65.52%
65.31%
64.41%
62.00%

12.3%
9.5%
8.3%
4.5%
8.6%
5.8%
3.2%
-1.4%
-2.3%
-0.4%

0.007
0.000
0.000
0.002
0.017
0.000
0.003
0.031
0.026
0.016

Table 6. 10 pairs most found on losing teams

Combination
3
4
3
6
3
7
6
7
6
9

9
7
7
9
6
9
7
7
6
9

Winning Count
37
34
28
28
15
28
13
6
3
15

Total

Percentage

Increase

P-value

106
100
84
86
49
108
51
24
13
66

34.91%
34.00%
33.33%
32.56%
30.61%
25.93%
25.49%
25.00%
23.08%
22.73%

-9.1%
-11.8%
-10.7%
-9.1%
-13.4%
-15.7%
-15.1%
-12.7%
-17.5%
-18.9%

0.002
0.001
0.002
0.001
0.005
0.000
0.000
0.011
0.047
0.000

Table 7. Top 10 3-way combinations most often found on winning teams

Combination
1
1
2
3
4
2
1
5
1
2

9
1
9
9
9
2
2
9
5
4

10
9
10
10
10
9
9
10
9
9

Winning Count

Total.Count

24
28
41
27
23
32
90
29
66
76

25
30
44
29
25
35
99
33
76
88

Percentage

P-value

96.00%
93.33%
93.18%
93.10%
92.00%
91.43%
90.91%
87.88%
86.84%
86.36%

< .001
< .001
< .001
< .001
< .001
< .001
< .001
< .001
< .001
< .001

Table 8. 10 3-way combinations most often on losing teams

Combination
7
1
2
3
5
2
5
3
4
3

7
7
6
7
5
7
7
3
7
6

9
7
6
7
7
7
7
3
7
6

Winning Count

Total Count

Percentage

P-value

3
6
3
7
2
6
4
1
5
1

35
73
39
105
31
103
75
25
138
36

8.57%
8.22%
7.69%
6.67%
6.45%
5.83%
5.33%
4.00%
3.62%
2.78%

< .001
< .001
< .001
< .001
< .001
< .001
< .001
< .001
< .001
< .001

7. References

[1] Christopher M. Bishop, Pattern Recognition and Machine Learning, New York, NY, 2006.
[2] Alan J. Izenman, Modern Multivariate Statitsical Techniques, New York, NY, 2008.
[3] Chris Fraley and A.E. Rafterty, Model-based Methods of Classifcation: Using the mclust Software in Chemometrics, Journal
of Statistical Software, vol. 18, issue 6, Jan. 2007.
[4] Hoopdata - NBA Statistics and Analysis. Oct. 2009. Web. 13 Jan. 2012. <http://www.hoopdata.com/>.
[5] Gary W. Oehlert, A First Course in Design and Analysis of Experiments, New York, NY, 2000.

10