See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/315766823

A PageRank Model for Player Performance Assessment in Basketball, Soccer and

Hockey

Article · March 2017

CITATIONS READS

8 206

1 author:

Shael Brown
McGill University
1 PUBLICATION   8 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Shael Brown on 22 April 2021.

The user has requested enhancement of the downloaded file.

 A PageRank Model for Player Performance
arXiv:1704.00583v1 [stat.AP] 31 Mar 2017

Assessment in Basketball, Soccer and Hockey

Shael Brown
Department of Mathematics and Statistics
Dalhousie University, Halifax, Nova Scotia, Canada B3H 3J5

Abstract
In the sports of soccer, hockey and basketball the most commonly
used statistics for player performance assessment are divided into two
categories: offensive statistics and defensive statistics. However, quali-
tative assessments of playmaking (for example making “smart” passes)
are difficult to quantify. It would be advantageous to have available
a single statistic that can emphasize the flow of a game, rewarding
those players who initiate and contribute to successful plays more.
In this paper we will examine a model based on Google’s PageRank.
Other papers have explored ranking teams, coaches, and captains but
here we construct ratings and rankings for individual members on both
teams that emphasizes initiating and partaking in successful plays and
forcing defensive turnovers.
For a soccer/hockey/basketball game, our model assigns a node for
each of the n players who play in the game and a “goal node”. Arcs
between player nodes indicate a pass in the reverse order (turnovers
are dealt with separately). Every sport-specific situation (fouls, out-
of-bounds, play-stoppages, turnovers, missed shots, defensive plays)
is addressed, tailored for each sport. As well, some additional arcs
are added in to ensure that the associated matrix is primitive (some
power of the matrix has all positive entries) and hence there is a unique
PageRank vector. The PageRank vector of the associated matrix is
used to rate and rank the players of the game.
To illustrate the model, data was taken from nine NBA games
played between 2014 and 2016. The model applied to the data showed
that this model did indeed provide the type of comprehensive statistic

1
 described in the introductory paragraph. Many of the top-ranked
players (in the model) in a given game had some of the most impressive
traditional stat-lines. However, from the model there were surprises
where some players who had impressive stat-lines had lower ranks, and
others who had less impressive stat-lines had higher ranks. Overall,
the model provides an alternate tool for player assessment in soccer,
basketball and hockey. The model’s ranking and ratings reflect more
the flow of the game compared to traditional sports statistics.

1 Background
Google PageRank was created as the backbone to what is now the most in-
fluential search engine ever created [4]. Its purpose is to rank the importance
of web pages when a user makes a query. The foundations of PageRank lie
in Markov chain theory: given a finite set of states S = {s1 , . . . , sn }, let ti,j
be the probability of moving from state si to state sj from time k to k + 1,
which is independent of k. Let T = (ti,j ) denote the transition matrix of the
Markov chain. Provided that the matrix T is primitive (i.e.., T m > 0 for
some positive integer m), there is a unique stationary vector, v, such that
vt 1 = 1 and T t v = v, where 1 is the vector in Rn consisting of all 1’s (that
is, v is an eigenvector of T t with eigenvalue 1 whose nonnegative entries sum
to 1) [1]. We call such a vector the PageRank vector of the Markov chain.
(The lack of primitivity in Google’s Markov model in general requires some
alteration to the transition matrix in that case.) Each component of the
PageRank vector is thought of as the rank of that state (and an ordering of
the states is derived from these values).
There is a natural way to construct a Markov chain from a (finite) directed
graph. The states of the chain are the nodes of the graph. If there are ni,j
arcs from node i to node j and node i has a total of ni outgoing arcs, then
n
Ti,j = ni,ji . For a Markov chain derived from a directed graph, primitivity of
the Markov chain corresponds to the existence of a positive integer k such
that there is a walk of length k between any two nodes. In such a case there
is a unique stationary vector for the associated Markov chain, which can be
calculated from a linear system (see [4] for more details). We remark that
it has been observed that the PageRank vector is fairly insensitive to small
changes in the network involving only lowly-ranked nodes [12].
Previous applications of PageRank to sports metrics usually address rank-

2
ing either teams [5, 3, 10], coaches [13], or individual players on various teams
[10, 11]. A number of models have relied on underlying graph networks of
games in their respective sports. In [7] a directed graph representing the
passes between the starting players on an individual soccer team was con-
structed and a PageRank vector was computed to highlight who were the
most important players on the team based on ball reception; the network as
well was analyzed to determine the strategies and weak-points of each team.
In [6] every team (in some team sport involving passing) has a correspond-
ing weighted directed graph representing passes and two additional nodes
representing the other team’s goal and missed shots. Batsmen and bowlers
that face each other on separate teams are compared using a basic model
much like the one that compares teams in the “win-loss” PageRank method
for ranking teams [11]. A weighted directed graph for players across many
basketball teams is created in [8] where arcs exist only between players who
played on the court together on the same team at some point and the weight
of these arcs corresponds to how effective they were in playing together. Fi-
nally, in [9] a PageRank network is created for each individual soccer team
where arcs indicate passes between players.
None of these models allow for the effective comparison of any two players
playing in the same game together (or even in different games), possibly of
different positions or teams, using a PageRank method. More importantly,
they don’t emphasize playmaking ability, as opposed to pure offensive statis-
tics, and that is exactly what we plan to do.

2 The Model
We create a directed graph (which we abbreviate as a digraph) representing
the progression of play during a particular game – whether soccer, basketball
or hockey. For each of the n players in the game there is a node, and the di-
graph contains one additional goal node (while our implementation of a goal
node is not unique, the use of a “missed shot” node in [6] is redundant when
we have both teams that are competing against each other in a game repre-
sented on the same graph). Since we desire a model that values playmaking,
we must reverse the direction of most arcs that would result from a straight-
forward progression of play (such an idea was raised but not deeply pursued
in a multi-team setting in [9]). For player i and player j (represented by
node i and node j, respectively) on the same team, whenever player i passes

3
to player j we draw an arc from node j to node i. However, if players i and
j are on separate teams and player i loses the ball/puck to player j we draw
an arc from node i to node j. If player i scores, the sport specific value of
the score will be the number of arcs drawn from the goal node to node i (for
example an NBA 3-pointer would result in three arcs). All of our choices
for arc direction ensures the flow of rank rewards playmaking. There will be
more game-specific arcs, to be discussed below.
We initialize the digraph for a given game as follows. We draw arcs in
both directions between each player node and the goal node, and in addition,
we draw a loop from the goal node to itself. Outside of goal scoring, no
further arc will be drawn to or from the goal node. If team 1 has n1 players
and team 2 has n2 players (n1 + n2 = n) then in the corresponding game
transition matrix T ∈ Mn+1,n+1 (R) the first n1 columns (and rows) of T
represent players on team 1, columns (and rows) n1 + 1 to n representing
players on team 2, and the (n + 1)th column and row representing the goal
1
node. Thus, before the game starts, Tn+1,i = n+1 for 1 ≤ i ≤ n + 1 and
Ti,j = 0 otherwise. The method of construction of the initial digraph ensures
that any two nodes are connected by a path of length exactly two, and hence
the corresponding Markov chain transition matrix will have each entry of T 2
nonnegative, making T primitive, and therefore the Markov chain associated
with the digraph has a unique PageRank vector.
We let r = (r1 , . . . , rn , rg ) be the PageRank vector of the game transition
matrix. For a number of reasons, to be listed, we shall rescale the values in
the PageRank vector (such a process does not change the induced ordering of
the players’ ranks). One immediate issue with the model is the fact that the
goal node may have different rank in each game depending on the number
of players in the game, thus making the comparison of ranks of players in
different games dependent on the rank of the goal node. We can scale the
computed rank vector r = (r1 , r2 , ..., rg ) by any scalar (the eigenspace corre-
sponding to eigenvalue 1 has dimension 1, as the digraph, being primitive, is
strongly connected). We standardize r by defining the integrated playmaking
metric (IPM) of player i in the game by
ri 50n · ri
IP Mi = 50n · Pn = .
j=1 rj 1 − rg
The choice of scaling is as follows: the denominator removes the effect of the
goal node’s rank, and the numerator ensures (a) that the IPM is insensitive
to the number of players in the game and (b) provides values on a reasonable

4
scale, between 0 and 1000 (it is not hard to see that the average IPMs of
all players in a game is 50). The IPMs can thus be used for meaningful
comparison of players in different games.
We now return to how the digraph itself is built up in the three sport spe-
cific situations. We illustrate the process with basketball (the rules for soccer
and hockey can be found in Appendices A and B respectively). Each bul-
let point is a play “event”, with its subsequent descriptor the corresponding
arc(s) to add to the digraph.
Basketball rules of implementation:

• Pass from player i to player j.

− An arc from node j to node i

• Player i dispossesses player j. [This could include cases where player

i does not gain possession of the ball after dispossessing player j –
for example a defensive touch leading to the ball being out of play or
deflecting a pass still into play but away from its intended target.]

− An arc from node j to node i

• Player i scores n points where 1 ≤ n ≤ 4.

− n arcs from the goal node to node i

• Player i shoots when being contested and defended by player j and

misses the net. [Same as player j dispossessing player i, play resum-
ing with the rebounding/inbounding player. This case includes the
situation where player j blocks player i.]

− An arc from node i to node j

• Player i shoots and misses the net under no pressure and the ball is
rebounded by player j. [Same as player j dispossessing player i.]

− An arc from node i to player j

• Player i fouls player j and player j makes at least one free throw.

− If player j makes n > 0 free throws then n arcs are created from
the goal node to node j

5
 • Player i fouls player j and player j makes zero free throws. [Same as
player j dispossessing player i – it was a “smart” foul.]

− An arc from node j to node i

• Any stoppage of play that does not have to do with the game (i.e. a
technical foul, fan interference, injury, altercation etc.). [Play is dead.]

− No arc drawn

• Player i intercepts a pass from player j. [Same as player i dispossessing

player j.]

− An arc from node j to node i

• Player i touches the ball without having possession (for example the
ball hits player i, or a “pinball” play).

− No arc drawn

• Any unforced turnover by player i.

− No arcs drawn

We illustrate the process with a small example. Suppose that two bas-
ketball teams, the Reds and the Blues, are playing against each other in a
“3-on-3” match (where all baskets are worth one point). We will denote the
players on the Reds by A, B and C and those on the Blues by D, E and F .
Before the game begins we have the setup of the digraph shown in Figure 1
(where G stands for the goal node). In this and subsequent diagrams of di-
graphs, if there exists more than one arc between two nodes we will draw one
arc but label it with the number of arcs that exist between the two nodes.

6
 A F

G
B E

C D

Figure 1: Small example’s initial digraph (before play).

Now suppose we have the following sequence of plays in the game: (where
“→” represents the movement of the ball between nodes and “0” is the end
of play sequence symbol):
A→B→A→F →G
D→F →E→F →D→C→B→C→A→C→B→A→G
D→C →A→C →B→A→G
D→F →0→B→C→A→G
D→F →E→F →D→G
A→B→F →G

After these sequences of plays our updated network becomes:

7
 A F

3 2
2 3 2
G
B E
2
2 3
3

C D
2

Figure 2: Small example updated

The adjacency matrix of the digraph (whose (i, j)–th entry is the number of
arcs in the digraph from node i to node j) is
 
0 3 3 0 0 1 1
2 0 3 0 0 1 1
 
2 2 0 0 0 0 1
 
0 0 2 0 0 2 1 ,
 
0 0 0 0 0 2 1
 
0 0 0 3 2 0 1
4 1 1 2 1 3 1

and so the transition matrix for this game is

 
0 2/7 2/5 0 0 0 4/13
3/8 0 2/5 0 0 0 1/13
 
3/8 3/7 0 2/5 0 0 1/13
 
T =  0 0 0 0 0 1/2 2/13
.
 0 0 0 0 0 1/3 1/13
 
1/8 1/7 0 2/5 2/3 0 3/13
1/8 1/7 1/5 1/5 1/3 1/6 1/13

The calculated IPMs of the players in the system are as follows:

While player A scored the most goals in the game and player F is tied
with player A for the most points (sum of goals and assists), we see that

8
 Player Team IPM
C Reds 64.66
F Blues 60.38
A Reds 58.79
B Reds 52.39
D Blues 39.17
E Blues 24.61

in fact player C, who had only 1 assist and no goals, had the highest IPM.
However, a cursory examination of the plays clearly shows how integral player
C was in the game as a playmaker. The example shows how a player whose
contribution to the game might be ignored under the usual stats lines receives
their well deserved acknowledgement under the proposed model.
Before we continue with some experimental results it is natural to ask how
this model fits with our intuition of evaluating the performance of athletes.
We observe first that the lowest possible IPM of any player in a game of
n players is no more than 50, since if the smallest IPM of any player in the
game is more than 50 we would contradict the fact that the average IPM of
all players in the game is always 50. The following propositions consider the
spacing of IPMs in a game.
Proposition 2.1. If there are k starters (out of n players) in a given game
with a cumulative starter IPM of R then there must exist a bench player with
a IPM that is at most n(R−50k)
k(n−k)
distance from the IPM of some starter.

Proof. We first note that the average bench player IPM must be 50n−R n−k
as
there are n−k bench players and the total sum of all the IPMs is 50n. Clearly
the worst case scenario is if all starters have a IPM of Rk and all bench players
have a IPM of 50n−R
n−k
(as otherwise there would have to be one bench player
with IPM greater than 50n−Rn−k
or one starter with IPM less than Rk ). Thus,
in this worst case scenario the distance between any starter and any bench
player is Rk − 50n−R
n−k
= Rn−Rk−50nk+Rk
k(n−k)
= n(R−50k)
k(n−k)
.
Proposition 2.2. If there are n players in a game then there must exist at
25n
least two players who have IPMs within n−2 of each other.
Proof. Clearly 0 is a lower bound on the IPM of a player in a given game
and upper bound is 50n (the sum of all the IPMs). Thus, we may partition

9
 50n
this range into n − 2 equal length intervals, each of length n−2 using n − 1 of
the n total players. We must have that at least three players must be within
(or on the boundary of) one of the n − 2 intervals, meaning that two of their
IPMs must be in the same half interval, making the difference of their IPMs
25n
no more than n−2 .
The above two results show that there has to be some bench player of
non-negligible importance to the game, and that some players have to have
IPMs that are “somewhat” close to each other. The first result fits with the
widely accepted notion that bench play is a component to the success of a
basketball team.
Finally, we may be interested in how the number of goals scored by players
on different teams affects their IPMs. As an illustration, suppose we have
only two players, P1 and P2 in the system, each on separate teams, with no
interaction between them and P1 scores g1 goals and P2 scores g2 goals. Then
if g1 > g2 then the IPM of P1 is greater than that of P2 . This follows as it is
clear that the IPM of P1 is (g150nr g (g1 +1)
+g2 +2)(1−rg )
and the IPM of P2 is (g150nr g (g2 +1)
+g2 +2)(1−rg )
where rg is the rank of the goal node.

3 Sample Data Analysis

We now apply our model to some real-life basketball games. The specific
games used were
• Chicago Bulls vs. San Antonio Spurs (November 30th 2015),
• Golden State Warriors vs. Cleveland Cavaliers (January 18th 2016),
• San Antonio Spurs vs. Brooklyn Nets (December 3rd 2014),
• Chicago Bulls vs. Charlotte Hornets (December 3rd 2014),
• Los Angeles Lakers vs. Golden State Warriors (November 1st 2014),
• Toronto Raptors vs. Cleveland Cavaliers (December 9th 2014),
• Golden State Warriors vs. Cleveland Cavaliers (December 25th 2015),
• Chicago Bulls vs. Oklahoma City Thunder (December 25th 2015), and
• Washington Wizards vs. Cleveland Cavaliers (November 26th 2014).

10
The games are identified by the teams playing, date, score and winning team.
In each game, plays were manually transcribed, and the resulting transition
matrices and PageRank vectors were calculated. The nine tables list, in
decreasing order, the IPMs of each of the players (rounded to the nearest
hundredth) in all nine NBA basketball games from which data was taken,
followed by each player’s points (P), assists (A), rebounds (R), steals (S),
turnovers (T) and field goal percentage (FG%), all accessed from nba.com.

Chicago Bulls vs. San Antonio Spurs, November 30th 2015 (Bulls win 92-89)

Player name Team IPM P A R S T FG%

Parker Spurs 103.66 13 9 1 0 0 50
Rose Bulls 83.68 11 6 4 1 1 29
Duncan Spurs 74.05 6 3 12 0 2 43
Gasol Bulls 72.30 18 4 13 1 1 33
Leonard Spurs 63.62 25 3 8 2 2 77
Aldridge Spurs 59.99 21 0 12 0 2 55
Noah Bulls 58.32 8 7 11 0 0 67
Green Spurs 57.69 9 1 4 2 1 30
Butler Bulls 49.23 14 3 3 1 5 55
Mirotic Bulls 46.55 8 2 5 0 2 38
Ginobli Spurs 44.89 4 2 1 1 1 25
Diaw Spurs 44.82 5 0 6 0 1 40
Mills Spurs 40.31 4 1 0 0 0 25
Moore Bulls 33.42 6 1 2 0 1 50
West Spurs 31.02 2 1 3 0 0 20
Snell Bulls 27.48 11 1 6 0 0 80
McDermott Bulls 24.97 12 0 3 0 0 42
Gibson Bulls 20.86 4 1 4 1 0 40
Anderson Spurs 13.13 0 0 0 0 0 0

11
Golden State Warriors vs. Cleveland Cavaliers, January 18th 2016 (Warriors
win 132-98)

Player name Team IPM P A R S T FG%

Curry Warriors 122.91 35 4 5 3 1 67
Green Warriors 109.34 16 10 7 0 1 50
Dellavedova Cavaliers 109.26 11 6 1 0 1 50
James Cavaliers 91.43 16 5 5 1 3 44
Irving Cavaliers 78.75 8 3 5 0 2 27
Barnes Warriors 66.94 12 0 2 0 2 50
Livingston Warriors 62.80 4 0 2 0 2 67
Love Cavaliers 60.54 3 2 6 0 1 20
Iguodala Warriors 60.24 20 5 3 0 1 88
Bogut Warriors 58.61 4 0 6 0 0 67
Varejao Cavaliers 50.63 5 3 4 1 1 50
Shumpert Cavaliers 39.84 10 0 2 0 3 67
Barbosa Warriors 39.52 8 4 1 2 0 50
Thompson Warriors 39.05 15 2 1 1 1 45
Clark Warriors 38.24 6 2 2 0 0 29
Ezeli Warriors 29.79 4 0 2 0 2 67
Mozgov Cavaliers 28.84 6 3 0 0 1 50
Smith Cavaliers 26.99 14 1 2 1 1 67
Thompson Cavaliers 23.73 2 0 2 0 0 0
J Thompson Warriors 21.89 1 1 3 0 0 1
Cunningham Cavaliers 20.39 9 1 3 0 0 60
Jefferson Cavaliers 18.47 6 0 3 0 1 100
Jones Cavaliers 17.43 8 1 0 0 1 60
Speights Warriors 17.22 4 0 1 0 0 25
Rush Warriors 17.17 3 1 3 0 0 33

12
San Antonio Spurs vs. Brooklyn Nets, December 3rd 2014 (Nets win 95-93)

Player name Team IPM P A R S T FG%

Williams Nets 111.40 17 9 3 0 2 40
Teletovic Nets 90.99 26 2 15 0 0 69
Duncan Spurs 81.11 14 1 17 0 2 28
Parker Spurs 79.92 9 6 1 0 3 50
Ginobli Spurs 68.23 15 5 6 1 0 46
Green Spurs 65.77 20 2 10 2 0 50
Lopez Nets 62.01 16 3 16 0 0 35
Leonard Spurs 56.59 12 1 13 1 0 25
Joseph Spurs 55.32 7 3 3 0 0 38
Johnson Nets 53.48 8 2 5 1 1 25
Jack Nets 47.11 8 3 1 0 2 40
Diaw Spurs 45.72 0 3 2 0 2 0
Bonner Spurs 35.93 7 0 1 0 0 30
Bogdanovic Nets 32.72 14 0 8 0 2 50
Baynes Spurs 17.78 4 1 4 1 0 40
Anderson Nets 14.49 2 1 1 0 2 20
Belinelli Spurs 11.89 5 1 1 0 1 67
Jordan Nets 9.83 2 0 2 1 0 100
Plumlee Nets 9.73 2 0 3 0 1 33

13
Chicago Bulls vs. Charlotte Hornets, December 3rd 2014 (Bulls win 102-95)

Player name Team IPM P A R S T FG%

Walker Hornets 110.08 23 4 5 1 0 39
Rose Bulls 85.77 15 5 2 0 2 42
Gasol Bulls 84.94 19 3 15 0 2 37
Noah Bulls 77.80 14 7 10 1 2 67
Mirotic Bulls 73.59 11 1 2 0 1 50
Stephenson Hornets 64.26 20 4 8 1 4 50
Zeller Hornets 63.21 12 2 8 0 0 45
Williams Hornets 59.65 6 0 3 1 0 40
Butler Bulls 52.88 15 5 2 2 1 45
Brooks Bulls 51.90 7 3 3 0 2 43
Hinrich Bulls 48.90 12 2 3 0 1 44
Roberts Hornets 37.93 3 3 1 0 0 13
Jefferson Hornets 35.36 13 2 7 0 0 38
Dunleavy Bulls 32.20 9 0 1 1 0 60
Henderson Hornets 30.55 10 1 4 0 1 50
Hairston Hornets 14.67 4 1 2 2 0 14
Snell Bulls 13.54 0 1 1 0 0 0
Biyombo Hornets 12.64 4 0 4 0 0 50
Pargo Hornets 0.14 0 0 0 0 0 0

14
Los Angeles Lakers vs. Golden State Warriors, November 1st 2014 (Warriors
win 127-104)

Player name Team IPM P A R S T FG%

Curry Warriors 132.74 31 10 5 3 2 53
Lin Lakers 94.24 6 6 4 1 5 0
Green Warriors 89.61 9 1 5 1 1 33
Iguodala Warriors 84.83 9 6 4 2 4 50
Bogut Warriors 79.29 6 3 10 1 5 30
Bryant Lakes 76.87 28 1 6 2 7 43
Hill Lakers 74.27 23 4 5 0 2 71
Price Lakers 64.48 1 6 4 2 2 0
Davis Lakers 59.39 13 2 6 1 1 71
Barnes Warriors 53.12 15 3 4 1 1 83
Thompson Warriors 50.27 41 2 5 0 1 78
Livingston Warriors 42.72 2 1 2 1 1 50
Boozer Lakers 38.93 9 1 4 0 0 44
Ezeli Warriors 37.05 3 1 4 0 2 100
Barbosa Warriors 34.37 9 3 1 1 3 50
Johnson Lakers 33.40 15 0 4 0 1 67
Ellington Lakers 20.53 2 1 4 1 1 50
Speights Warriors 14.58 2 0 3 0 0 50
Sacre Lakers 7.69 4 0 1 0 2 50
Clarkson Lakers 5.52 3 0 1 2 1 20
Henry Lakers 4.95 0 0 0 0 0 0
Holiday Warriors 1.18 0 0 0 0 0 0

15
Toronto Raptors vs. Cleveland Cavaliers, December 9th 2014 (Cavaliers win
105-101)

Player name Team IPM P A R S T FG%

Lowry Raptors 123.55 16 14 4 1 0 33
Irving Cavaliers 118.30 13 10 1 2 2 42
James Cavaliers 95.39 35 4 2 2 2 57
Love Cavaliers 70.70 17 4 9 0 3 40
Valanciunas Raptors 65.34 18 0 15 0 3 86
Dellavedova Cavaliers 59.32 6 5 3 0 0 50
Patterson Raptors 47.36 12 1 4 0 1 71
Thompson Cavaliers 44.27 8 0 8 0 1 60
Williams Raptors 43.97 6 4 1 0 1 25
A. Johnson Raptors 41.76 10 2 2 0 2 50
Ross Raptors 37.68 18 1 3 0 5 62
Vasquez Raptors 33.11 3 2 0 0 1 33
Fields Raptors 32.38 4 2 1 2 1 100
Varejao Cavaliers 32.15 8 1 6 0 1 40
Waiters Cavaliers 29.92 18 2 1 0 1 70
J. Johnson Raptors 24.98 12 0 4 1 1 46
Marion Cavaliers 23.12 0 0 1 0 1 0
Jones Cavaliers 20.41 0 1 0 0 0 0
Hayes Raptors 6.30 2 0 1 0 0 100

16
Golden State Warriors vs. Cleveland Cavaliers, December 25th 2015 (War-
riors win 89-83)

Player name Team IPM P A R S T FG%

Green Warriors 140.89 22 7 15 0 4 47
Curry Warriors 117.29 19 7 7 2 3 40
Love Cavaliers 106.85 10 4 18 0 1 31
Dellavedova Cavaliers 98.70 10 1 5 1 1 36
James Cavaliers 97.45 25 2 9 1 4 38
Iguodala Warriors 74.47 7 3 2 1 0 17
Irving Cavaliers 65.18 13 2 3 1 2 27
Thompson Cavaliers 57.09 8 1 10 1 0 50
Livingston Warriors 54.35 16 2 3 1 4 89
Thompson Warriors 47.08 18 1 6 0 1 38
Bogut Warriors 47.03 4 1 7 0 0 100
Ezeli Warriors 36.77 3 0 4 0 2 25
Shumpert Cavaliers 30.38 0 1 4 1 0 0
Smith Cavaliers 28.77 14 0 1 1 2 44
Rush Warriors 18.84 0 0 3 1 1 0
Mozgov Cavaliers 15.40 0 0 3 0 1 0
Clark Warriors 14.20 0 0 0 1 0 0
Barbosa Warriors 14.04 0 0 1 0 0 0
McAdoo Warriors 13.33 0 0 1 0 0 0
Speights Warriors 10.75 0 0 0 1 1 0
Jones Cavaliers 5.78 0 0 2 0 0 0
Williams Cavaliers 5.35 3 1 0 0 0 0

17
Chicago Bulls vs. Oklahoma City Thunder, December 25th 2015 (Bulls win
105-96) *last 36.4 seconds of second quarter and first 17 seconds of 3rd quar-
ter were not able to be seen from the source.

Player name Team IPM P A R S T FG%

Westbrook Thunder 125.96 26 8 7 6 6 39
Gasol Bulls 104.06 21 6 13 0 4 50
Butler Bulls 94.85 23 4 6 4 3 45
Durant Thunder 79.36 29 7 9 1 2 52
Rose Bulls 79.34 19 1 4 0 4 39
Kanter Thunder 70.19 14 1 13 0 0 50
Gibson Bulls 62.38 13 2 10 1 1 75
Ibaka Thunder 58.25 6 0 7 2 2 25
Portis Bulls 49.98 7 3 5 1 1 38
Hinrich Bulls 41.15 2 2 0 0 0 50
Adams Thunder 38.62 3 0 4 0 0 25
Brooks Bulls 35.67 6 1 4 0 0 50
Mirotic Bulls 32.04 6 2 7 1 1 20
Augustin Thunder 28.25 3 1 1 1 2 25
Roberson Thunder 23.22 2 1 4 0 0 17
McDermott Bulls 19.31 5 1 3 1 1 29
Morrow Thunder 18.27 9 0 1 1 0 50
Snell Bulls 17.27 3 0 1 0 1 25
Waiters Thunder 12.41 2 2 0 1 1 17
Collison Thunder 9.42 2 0 2 0 0 50

18
Washington Wizards vs. Cleveland Cavaliers, November 26th 2014 (Cavaliers
win 113-87)

Player name Team IPM P A R S T FG%

James Cavaliers 117.45 29 8 10 3 4 50
Irving Cavaliers 117.00 18 5 1 3 1 47
Wall Wizards 113.96 6 7 4 0 5 33
Beal Wizards 62.07 10 2 2 3 1 40
Love Cavaliers 61.40 21 0 5 0 2 70
Gortat Wizards 52.63 12 1 2 1 3 50
Thompson Cavliers 50.89 10 0 1 0 0 100
Waiters Cavaliers 49.51 15 6 3 2 1 35
Miller Wizards 48.21 7 6 2 0 0 75
Seraphin Wizards 47.26 7 3 3 0 2 38
Varejao Cavaliers 47.03 10 0 7 0 1 100
Humphries Wizards 45.24 3 1 3 0 1 14
Pierce Wizards 42.62 15 3 3 0 3 80
Marion Cavaliers 40.00 6 2 4 2 0 25
Porter Jr. Wizards 30.72 2 1 2 0 0 25
Cherry Cavaliers 28.63 2 0 0 2 0 0
Blair Wizards 23.79 0 0 1 0 1 0
Butler Wizards 23.11 23 0 1 1 2 60
Amundson Cavaliers 20.52 0 1 1 0 0 0
Gooden Wizards 16.66 2 0 3 0 0 50
Harris Cavaliers 11.30 2 0 2 0 0 50

19
 There are several general trends that are reflected in the sample data that
intuitively matches what we would expect the trend to be. For example, out
48
of all the players with IPMs under 30, 55 ≈ 87% of them had their sum of
points, assists, rebounds and steals no greater than 10, corresponding to a
“small” stat-line. On the other hand out of all players who had IPMs greater
than 70, 28
44
≈ 64% had their sum of points, assists, rebounds and steals be
at least 25, corresponding to a “large” stat-line. In terms of highest ranked
positions, 7 of the games had a point guard ranked the highest; however,
a forward was ranked in the top five IPMs in all nine games as well. On
average, the starters of both teams owned approximately 7.1 out of the first
10 highest IPMs, which fits with our knowledge that at least one bench player
must have some significant importance to the game. Out of the nine games
sampled, the highest ranked player was on the losing team four times.
Now let us consider some basic trends in certain averages of the IPMs in
the nine games. In the table below all values were rounded to the nearest
hundredth, WT stands for winning team, LT stands for losing team and
AIPM stands for average IPM.

Game WT AIPM LT AIPM WT starter AIPM LT starter AIPM

1 46.31 53.32 55.85 71.80
2 52.59 47.19 79.37 57.31
3 47.97 51.82 70.11 65.82
4 57.95 42.85 66.72 60.69
5 56.34 43.66 81.00 63.54
6 54.84 45.64 67.93 60.14
7 49.09 51.10 74.23 62.73
8 53.60 46.40 74.53 65.08
9 54.37 46.03 76.58 63.31

Out of the nine games sampled, the winning team had a higher average
IPM than that of the losing team six times. However, the average IPMs of
starters on the winning team was greater than that for the losing team in
eight out of the nine games.
In many cases our intuition of how a strong offensive/overall performance
by a player should be ranked agreed with the generated IPMs. For example,
consider the cases of LeBron James (Cleveland Cavaliers), Stephen Curry
(Golden State Warriors) and Pau Gasol (Chicago Bulls). Each are recognized

20
as being very talented players in the NBA and in each game surveyed in which
they played they each had a “strong” stat-line, and also a very high IPM.
However, there were certainly also some surprises where players with “strong”
stat-lines did not have large IPMs. In the Lakers vs. Warriors game, Klay
Thompson scored 41 points yet had a very average IPM of 50.27. In the
Wizards vs. Cavaliers game, Butler had 23 points and a IPM of about 23.
A general trend that we observe is that out of all the players who had
at least 15 points and had a IPM less than 50, 89% had the sum of their
assists, rebounds and steals be under 10, i.e. scoring alone was not generally
highly valued by the model. In contrast, out of the 30 players in all games
who had at least 5 assists, 28 of them had a IPM of at least 50 (about 93%)
showing a general trend of rewarding passing compared to scoring. There
were of course also some overachievers in the sample data; of the top 10
IPMs in each game, on average 36/910
= 40% had a sum of their points, assists,
rebounds and steals be at most 20, corresponding to what one could call an at
most “standard/average” stat-line. Specific examples of high ranking players
with low stat-lines in this model were John Wall (Washington Wizards) vs.
the Cavaliers, Ronnie Price (Los Angeles Lakers) vs. the Warriors, Jeremy
Lin (Los Angeles Lakers) vs. the Warriors, Tony Parker (Spurs) vs. the Nets
and Matthew Dellavedova (Cavaliers) vs. the Warriors (both games) and vs.
the Raptors.

4 Conclusion
While the IPMs of players are comparable between games, in future work we
may consider, over a whole season in one sports league, having all players
and teams represented by one large weighted directed graph with one goal
node. This new statistic could be used as another measure of performance
between any athletes in the same league. Note that adjusting that larger
model for dealing with such events as player trades would not be difficult.
In any case, automating processing input is a necessary step in scaling the
model. Moreover, adjusting the model to encompass more sports that follow
similar models to soccer, basketball and hockey, for instance volleyball or
water polo, could find useful applications as well.
The PageRank-based model which we have constructed appears to be the
first of its kind to give a quantifiable measure in which players on different
teams and even different games can be ranked and compared inclusive of

21
their offensive and defensive skills. While this model could serve as a useful
mainstream statistic for scouts, coaches, managers and fans, more data anal-
ysis is required to see if the model can provide accurate outcome predictions
for games (comparing the average IPM of the starters of each team, prior to
that game, may be a useful tool for game outcome prediction). The statis-
tics from our proposed model could be used in conjunction with the standard
player performance metrics in each sport to help deepen our understanding
of who is really affecting the game the most.

References
[1] A. Langville, and C. Meyer, “Google’s PageRank and Beyond: The
Science of Search Engine Rankings”, Princeton University Press, (2011).

[2] C. de Kerchove , L. Ninove and P. Van Dooren, Maximizing PageRank

via outlinks, Linear Algebra and its Applications 429, (2008), 1254–1276.

[3] D. Barc, I. Drayer, P. Elliott, G. Gaut, and B. Osting, Ranking rank-

ings: an empirical comparison of the predictive power of sports ranking
methods, Journal of Quantitative Analysis in Sports 9, (2013), 187–202.

[4] D. Glelch, PageRank Beyond the Web, SIAM Review 57, (2015), 321–
363.

[5] E. Balreira, B. Miceli and T. Tegtmeyer, An Oracle Method to Predict

NFL Games, Journal of Quantitative Analysis in Sports 10, (2014), 183–
196.

[6] J. Duch, J. Waitzman and L. Amaral, Quantifying the Performance of

Individual Players in a Team Activity, PLoS ONE 5 (2010).

[7] J. Pea and H. Touchette, A network theory analysis of football strategies,

Sports Physics: Proc. 2012 Euromech Physics of Sports Conference,
(2013), 517–528.

[8] J. Piette, L. Pham and S. Anand, Evaluating Basketball Player Perfor-

mance via Statistical Network Modeling, MIT Sloan Sports Conference
(2011).

22
 [9] M. Brandt and U. Brefeld, Graph-based Approaches for Analyzing Team
Interaction on the Example of Soccer, Proceedings of the ECML/PKDD
Workshop on Machine Learning and Data Mining for Sports Analytics
8, (2015).

[10] S. Mukherjee, Identifying the greatest team and captain - A complex

network approach to cricket matches, Physica A: Statistical Mechanics
and its Applications 391, (2012), 6066–6076.

[11] S. Mukherjee, Quantifying individual performance in Cricket - A net-

work analysis of batsmen and bowlers, Physica A: Statistical Mechanics
and its Applications 393, (2014), 624–637.

[12] T. Chartier, E. Kreutzer, A. Langville and K. Pedings, Sensitivity and

stability of ranking vectors, Siam Journal of Scientific Computing 33,
(2011), 1077–1102.

[13] Z. Hu, J. Zhou, M. Zhang and Y. Zhao, Methods for ranking college
sports coaches based on data envelopment analysis and PageRank, Ex-
pert Systems: The Journal of Knowledge Engineering 32, (2015), 652–
673.

A Appendix A
Soccer rules of implementation:

• Pass from player i to player j.

− An arc from node j to node i

• Player i dispossesses player j. [This could include cases where player

i does not gain possession of the ball after dispossessing player j, for
example a defensive tackle leading to the ball being out of play or
deflecting a pass still into play but away from its intended target.]

− An arc from node j to node i

• Player i scores.

− An arc from the goal node to node i

23
• Player i shoots when being pressed by player j and misses the net.
[Same as player j dispossessing player i. This case includes the situation
where player j blocks player i.]

− An arc from node i to node j

• Player i shoots and misses the net under no pressure. [Play is dead.]

− No arcs drawn

• Player i shoots and shot is saved by the goalkeeper, player j. [Same as

player j dispossessing player i.]

− An arc from player i to player j

• Player i fouls player j, not resulting in a goal. [Play is dead.]

− No arcs drawn

• Player i fouls player j resulting in a penalty or a goal from a free kick.

[Smart drawing of a foul by player j, scoring from a free kick could
include a direct shot, a header or volley from the free kick or a rebound
inside the box following the free kick.]

− Arc from node i to node j

• Any stoppage of play that does not have to do with the game (i.e.
weather, fan interference, injury, altercation etc.). [Play is dead.]

− No arc drawn

• Player i intercepts a pass from player j. [Same as player i dispossessing

player j.]

− An arc from node j to node i

• Player i touches the ball without having possession (for example the
ball hits player i, or a “pinball” play). [Player i did not have possession.]

− No arc drawn

• Any unforced turnover by player i.

24
 − No arcs drawn

• Player i is offside when player j passes the ball. [Player j passes the
ball to player i who is in an offside position so the play ends at player
i.]

− An arc from node j to node i

B Appendix B
Hockey rules of implementation:
• Pass from player i to player j.

− An arc from node j to node i

• Player i dispossesses player j. [This could include cases where player

i does not gain possession of the puck after dispossessing player j, for
example a defensive touch leading to the puck being out of play or
deflecting a pass still into play but away from its intended target.]

− An arc from node j to node i

• Player i scores.

− An arc from the goal node to node i

• Player i shoots when being defended by player j and misses the net.
[Same as player j dispossessing player i, play resumes with the player
that collects the puck after the shot. This case includes the situation
where player j blocks the shot of player i.]

− An arc from node i to node j

• Player i shoots and the shot is saved by the goalkeeper, player j. [Same
as player j dispossessing player i.]

− An arc from node i to player j

• Player i shoots and misses the net under no pressure. [Play is dead.]

− No arcs drawn

25
 • Player i draws a penalty from player j during which no power-play goal
is scored. [A “smart” penalty.]

− An arc from node i to node j

• Player i draws a penalty from player j during which a power-play goal

is scored. [A “smart” drawing of a penalty.]

− An arc from node j to node i

• Any stoppage of play that does not have to do with the game (i.e. a
penalty, fan interference, injury, altercation etc.). [Play is dead.]

− No arc drawn

• Player i intercepts a pass from player j. [Same as player i dispossessing

player j.]

− An arc from node j to node i

• Player i touches the puck without having possession (for example the
puck hits player i, or a “pinball” play). [Player i did not have posses-
sion.]

− No arc drawn

• Any unforced turnover by player i.

− No arcs drawn

• Player i is offside when player j passes the puck. [Player j still passed
player i the puck, the play ended by player i being in an offside position.]

− An arc from node i to node j

• Player i ices the puck which is touched by player j. [Same as player i

turning the puck over to player j.]

− An arc from node i to node j

26

View publication stats