Sports ranking methods, 4

Posted on 2017/08/10 by wdjoyner

This is the fourth of a series of expository posts on matrix-theoretic sports ranking methods. This post discusses the Elo rating.

This system was originally developed by Arpad Elo (Elo (1903-1992) was a physics professor at Marquette University in Milwaukee and a chess master, eight-time winner of the Wisconsin State Chess Championships.) Originally, it was developed for rating chess players in the 1950s and 1960s. Now it is used for table tennis, basketball, and other sports.

We use the following version of his rating system.

As above, assume all the $n$ teams play each other (ties allowed)
and let $r_i$ denote the rating of Team $i$ , $i=1,2,\dots,n$ .

Let $A=(A_{ij})$ denote an $n\times n$ matrix of score results:

$A_{ij}= \left\{ \begin{array}{rr} -1,& {\rm if\ team\ } i {\rm \ lost\ to\ team\ } j,\\ +1,& {\rm if\ team\ } i {\rm\ beat\ team\ } j,\\ 0, & {\rm if}\ i=j. \end{array} \right.$

Let $S_{ij}=(A_{ij}+1)/2$ .

As in the previous post, the matrix $A$ associated to the example of the Patriot league is the adjacency matrix of a diagraph.

Initialize all the ratings to be $100$ : ${\bf r}=(r_1,\dots,r_n) = (100,\dots,100)$ .
After Team $i$ plays Team $j$ , update their rating using the formula
$r_i = r_i+K(S_{ij}-mu_{ij}),$

where $K=10$ and

$\mu_{ij} = (1+e^{-(r_i-r_j)/400})^{-1}.$

In the example of the Patriot league, the ratings vector is

${\bf r}=(85.124, 104.79, 104.88, 85.032, 94.876, 124.53).$

This gives the ranking

Lafayette $<$ Army $<$ Lehigh $<$ Bucknell $<$ Holy Cross $<$ Navy.

This gives a prediction failure rate of $13.3\%$ .

Some SageMath code for this:

def elo_rating(A):
    """
    A is a signed adjacency matrix for a directed graph.

    Returns elo ratings of the vertices of Gamma = Graph(A) 
        
    EXAMPLES:
        sage: A = matrix(QQ,[
        [0 , -1 , 1  , -1 , -1 , -1 ],
        [1,   0 ,  -1,  1,  1,   -1  ],
        [-1 , 1 ,  0 ,  1 , 1  , -1  ],
        [1 , -1 , -1,  0 ,  -1 , -1  ],
        [1 , - 1 , - 1 , 1 , 0 , - 1  ],
        [1 ,  1  ,  1  , 1  , 1  , 0 ]
        ])
        sage: elo_rating(A)
        (85.124, 104.79, 104.88, 85.032, 94.876, 124.53)

    """
    n = len(A.rows())
    RR = RealField(prec=20)
    V = RR^n
    K = 10
    r0 = 100 # initial rating
    r = n*[r0]
    for i in range(n):
        for j in range(n):
            if ij and A[i][j]==1:
                S = 1
            elif ij and A[i][j]==-1:
                S = 0
            else:
                S = 1/2
            mu = 1/(1+e^(-(r[i]-r[j])/400))
            r[i] = r[i]+K*(S-mu)
    return V(r)

Sports ranking methods, 3

Posted on 2017/01/26 by wdjoyner

This is the third of a series of expository posts on matrix-theoretic sports ranking methods. This post discusses the random walker ranking.

We follow the presentation in the paper by Govan and Meyer (Ranking National Football League teams using Google’s PageRank). The table of “score differentials” based on the table in a previous post is:

$\begin{tabular}{c|cccccc} \verb+x\y+ & Army & Bucknell & Holy Cross & Lafayette & Lehigh & Navy \\ \hline Army & 0 & 0 & 1 & 0 & 0 & 0 \\ Bucknell & 2 & 0 & 0 & 2 & 3 & 0 \\ Holy Cross & 0 & 3 & 0 & 4 & 14 & 0 \\ Lafayette & 10 & 0 & 0 & 0 & 0 & 0 \\ Lehigh & 2 & 0 & 0 & 11 & 0 & 0 \\ Navy & 11 & 14 & 8 & 22 & 6 & 0 \\ \end{tabular}$
This leads to the following matrix:

$M_0=\left(\begin{array}{cccccc} 0 & 0 & 1 & 0 & 0 & 0 \\ 2 & 0 & 0 & 2 & 3 & 0 \\ 0 & 3 & 0 & 4 & 14 & 0 \\ 10 & 0 & 0 & 0 & 0 & 0 \\ 2 & 0 & 0 & 11 & 0 & 0 \\ 11 & 14 & 8 & 22 & 6 & 0 \\ \end{array}\right) .$

The edge-weighted score-differential graph associated to $M_0$ (regarded as a weighted adjacency matrix) is in the figure below.

sm261_baseball-ranking-application_teams-digraph2

This matrix $M_0$ must be normalized to create a (row) stochastic matrix:

$M = \left(\begin{array}{cccccc} 0 & 0 & 1 & 0 & 0 & 0 \\ {2}/{7} & 0 & 0 /{7} /{7} & 0 \\ 0 /{7} & 0 /{21} /{3} & 0 \\ 1 & 0 & 0 & 0 & 0 & 0 \\ {2}/{13} & 0 & 0 /{13} & 0 & 0 \\ {11}/{61} /{61} /{61} /{61} /{61} & 0 \\ \end{array}\right) .$

Next, to insure it is irreducible, we replace $M$ by $A=(M+J)/2$ , where $J$ is the $6\times 6$ doubly stochastic matrix with every entry equal to $1/6$ :

$A=\left(\begin{array}{cccccc} {1}/{12} & 1/{12} & 7/{12} & 1/{12} & 1/{12} & 1/{12} \\ {19}/{84} & 1/{12} & 1/{12} & 19/{84} & 25/{84} & 1/{12} \\ {1}/{12} & 13/{84} & 1/{12} & 5/{28} & 5/{12} & 1/{12} \\ {7}/{12} & 1/{12} & 1/{12} & 1/{12} & 1/{12} & 1/{12} \\ {25}/{156} & 1/{12} & 1/{12} & 79/{156} & 1/{12} & 1/{12} \\ {127}/{732} & 145/{732} & 109/{732} & 193/{732} & 97/{732} & 1/{12} \end{array}\right).$

Let

${\bf v}_0 = \left( \frac{1}{6} , \frac{1}{6} , \frac{1}{6} , \frac{1}{6} , \frac{1}{6} , \frac{1}{6}\right).$

The ranking determined by the random walker method is the reverse of the left eigenvector of $A$ associated to the largest eigenvalue $\lambda_{max}=1$ (by reverse, I mean that the vector ranks the teams from worst-to-best, not from best-to-worst, as we have seen in previous ranking methods).
In other words, the vector

${\bf r}^*=\lim_{n\to \infty}{\bf v}_0A^n.$

This is approximately

${\bf r}^* \cong \left(0.2237\dots ,\,0.1072\dots ,\,0.2006\dots ,\,0.2077\dots ,\,0.1772\dots ,\,0.0833\dots \right).$

Its reverse gives the ranking:

Army $<$ Lafayette $<$ Bucknell $<$ Lehigh $<$ Holy Cross $<$ Navy.

This gives a prediction failure rate of $13.3\%$ .

Sports ranking methods, 2

Posted on 2016/12/18 by wdjoyner

This is the second of a series of expository posts on matrix-theoretic sports ranking methods. This post discusses Keener’s method (see J.P. Keener, The Perron-Frobenius theorem and the ranking of football, SIAM Review 35 (1993)80-93 for details).

See the first post in the series for a discussion of the data we’re using to explain this method. We recall the table of results:

X\Y	Army	Bucknell	Holy Cross	Lafayette	Lehigh	Navy
Army	x	14-16	14-13	14-24	10-12	8-19
Bucknell	16-14	x	27-30	18-16	23-20	10-22
Holy Cross	13-14	30-27	x	19-15	17-13	9-16
Lafayette	24-14	16-18	15-19	x	12-23	17-39
Lehigh	12-10	20-23	13-17	23-12	x	12-18
Navy	19-8	22-10	16-9	39-17	18-12	x

sm261_baseball-ranking-application_teams-digraph

Win-loss digraph of the Patriot league mens baseball from 2015

Suppose T teams play each other. Let $A=(a_{ij})_{1\leq i,j\leq T}$ be a non-negative square matrix determined by the results of their games, called the preference matrix. In his 1993 paper, Keener defined the score of the $i$ th team to be given by

$s_i = \frac{1}{n_i}\sum_{j=1}^T a_{ij}r_j,$

where $n_i$ denotes the total number of games played by team $i$ and ${\bf r}=(r_1,r_2,\dots ,r_T)$ is the rating vector (where $r_i\geq 0$ denotes the rating of team $i$ ).

One possible preference matrix the matrix $A$ of total scores obtained from the pre-tournament table below:

$A = \left(\begin{array}{rrrrrr} 0 & 14 & 14 & 14 & 10 & 8 \\ 16 & 0 & 27 & 18 & 23 & 28 \\ 13 & 30 & 0 & 19 & 27 & 43 \\ 24 & 16 & 15 & 0 & 12 & 17 \\ 12 & 20 & 43 & 23 & 0 & 12 \\ 19 & 42 & 30 & 39 & 18 & 0 \end{array}\right),$

(In this case, $n_i=4$ so we ignore the $1/n_i$ factor.)

In his paper, Keener proposed a ranking method where the ranking vector ${\bf r}$ is proportional to its score. The score is expressed as a matrix product $A{\bf r}$ , where $A$ is a square preference matrix. In other words, there is a constant $\rho>0$ such that $s_i=\rho r_i$ , for each $i$ . This is the same as saying $A {\bf r} = \rho {\bf r}$ .

The Frobenius-Perron theorem implies that $S$ has an eigenvector ${\bf r}=(r_1,r_2,r_3,r_4,r_5,r_6)$ having positive entries associated to the largest eigenvalue $\lambda_{max}$ of $A$ , which has (geometric) multiplicity $1$ . Indeed, $A$ has maximum eigenvalue $\lambda_{max}= 110.0385...$ , of multiplicity $1$ , with eigenvector

${\bf r}=(1, 1.8313\dots , 2.1548\dots , 1.3177\dots , 1.8015\dots , 2.2208\dots ).$

Therefore the teams, according to Kenner’s method, are ranked,

Army $<$ Lafayette $<$ Lehigh $<$ Bucknell $<$ Holy Cross $<$ Navy.

This gives a prediction failure rate of just $6.7\%$ .

Sports ranking methods, 1

Posted on 2016/12/03 by wdjoyner

This is the first of a series of expository posts on matrix-theoretic sports ranking methods. This post, which owes much to discussions with TS Michael, discusses Massey’s method.

Massey’s method, currently in use by the NCAA (for football, where teams typically play each other once), was developed by Kenneth P. Massey
while an undergraduate math major in the late 1990s. We present a possible variation of Massey’s method adapted to baseball, where teams typically play each other multiple times.

There are exactly 15 pairing between these teams. These pairs are sorted lexicographically, as follows:

(1,2),(1,3),(1,4), …, (5,6).

In other words, sorted as

Army vs Bucknell, Army vs Holy Cross, Army vs Lafayette, …, Lehigh vs Navy.

The cumulative results of the 2016 regular season are given in the table below. We count only the games played in the Patriot league, but not including the Patriot league post-season tournament (see eg, the Patriot League site for details). In the table, the total score (since the teams play multiple games against each other) of the team in the vertical column on the left is listed first. In other words, ”a – b” in row $i$ and column $j$ means the total runs scored by team $i$ against team $j$ is $a$ , and the total runs allowed by team $i$ against team $j$ is $b$ . Here, we order the six teams as above (team $1$ is Army (USMI at Westpoint), team $2$ is Bucknell, and so on). For instance if X played Y and the scores were $10-0$ , $0-1$ , $0-1$ , $0-1$ , $0-1$ , $0-1$ , then the table would read $10-5$ in the position of row X and column Y.

X\Y	Army	Bucknell	Holy Cross	Lafayette	Lehigh	Navy
Army	x	14-16	14-13	14-24	10-12	8-19
Bucknell	16-14	x	27-30	18-16	23-20	10-22
Holy Cross	13-14	30-27	x	19-15	17-13	9-16
Lafayette	24-14	16-18	15-19	x	12-23	17-39
Lehigh	12-10	20-23	13-17	23-12	x	12-18
Navy	19-8	22-10	16-9	39-17	18-12	x

Win-loss digraph of the Patriot league mens baseball from 2015

In this ordering, we record their (sum total) win-loss record (a 1 for a win, -1 for a loss) in a $15\times 6$ matrix:

$M = \left(\begin{array}{cccccc} -1 & 1 & 0 & 0 & 0 & 0 \\ 1 & 0 & -1 & 0 & 0 & 0 \\ -1 & 0 & 0 & 1 & 0 & 0 \\ -1 & 0 & 0 & 0 & 1 & 0 \\ -1 & 0 & 0 & 0 & 0 & 1 \\ 0 & -1 & 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & -1 & 0 & 0 \\ 0 & 1 & 0 & 0 & -1 & 0 \\ 0 & -1 & 0 & 0 & 0 & 1 \\ 0 & 0 & 1 & -1 & 0 & 0 \\ 0 & 0 & 1 & 0 & -1 & 0 \\ 0 & 0 & -1 & 0 & 0 & 1 \\ 0 & 0 & 0 & -1 & 1 & 0 \\ 0 & 0 & 0 & -1 & 0 & 1 \\ 0 & 0 & 0 & 0 & -1 & 1 \end{array}\right).$

We also record their total losses in a column vector:

${\bf b}= \left(\begin{array}{c} 2 \\ 1 \\ 10 \\ 2 \\ 11 \\ 3 \\ 2 \\ 3 \\ 14 \\ 4 \\ 14 \\ 10 \\ 11 \\ 22 \\ 6 \\ \end{array}\right).$

The Massey ranking of these teams is a vector ${\bf r}$ which best fits the equation

$M{\bf r}={\bf b}.$

While the corresponding linear system is over-determined, we can look for a best (in the least squares sense) approximate solution using the orthogonal projection formula

$P_V = B(B^tB)^{-1}B^t,$

valid for matrices $B$ with linearly independent columns. Unfortunately, in this case $B=M$ does not have linearly independent columns, so the formula doesn’t apply.

Massey’s clever idea is to solve

$M^tM{\bf r}=M^t{\bf b}$

by row-reduction and determine the rankings from the parameterized form of the solution. To this end, we compute

$M^tM= \left(\begin{array}{rrrrrr} 5 & -1 & -1 & -1 & -1 & -1 \\ -1 & 5 & -1 & -1 & -1 & -1 \\ -1 & -1 & 5 & -1 & -1 & -1 \\ -1 & -1 & -1 & 5 & -1 & -1 \\ -1 & -1 & -1 & -1 & 5 & -1 \\ -1 & -1 & -1 & -1 & -1 & 5 \end{array}\right)$

and

$M^t{\bf b}= \left(\begin{array}{r} -24 \\ -10 \\ 10 \\ -29 \\ -10 \\ 63 \\ \end{array}\right).$

Then we compute the rref of

$A= (M^tM,M^t{\bf b}) = \left(\begin{array}{rrrrrr|r} 5 & -1 & -1 & -1 & -1 & -1 & -24 \\ -1 & 5 & -1 & -1 & -1 & -1 & -10 \\ -1 & -1 & 5 & -1 & -1 & -1 & 10 \\ -1 & -1 & -1 & 5 & -1 & -1 & -29 \\ -1 & -1 & -1 & -1 & 5 & -1 & -10 \\ -1 & -1 & -1 & -1 & -1 & 5 & 63 \end{array}\right),$

which is

$rref(M^tM,M^t{\bf b})= \left(\begin{array}{rrrrrr|r} 1 & 0 & 0 & 0 & 0 & -1 & -\frac{87}{6} \\ 0 & 1 & 0 & 0 & 0 & -1 & -\frac{73}{6} \\ 0 & 0 & 1 & 0 & 0 & -1 & -\frac{53}{6} \\ 0 & 0 & 0 & 1 & 0 & -1 & -\frac{92}{3} \\ 0 & 0 & 0 & 0 & 1 & -1 & -\frac{73}{6} \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 \end{array}\right).$

If ${\bf r}=(r_1,r_2,r_3,r_4,r_5,r_6)$ denotes the rankings of Army, Bucknell, Holy Cross, Lafayette, Lehigh, Navy, in that order, then

$r_1=r_6-\frac{87}{6},\ \ r_2=r_6-\frac{73}{6},\ \ r_3=r_6-\frac{53}{6},\ \ r_4=r_6-\frac{92}{6},\ \ r_5=r_6-\frac{73}{6}.$

Therefore

Lafayette $<$ Army = Bucknell = Lehigh $<$ Holy Cross $<$ Navy.

If we use this ranking to predict win/losses over the season, it would fail to correctly predict Army vs Holy Cross (Army won), Bucknell vs Lehigh, and Lafayette vs Army. This gives a prediction failure rate of $20\%$ .

Yet Another Mathblog

Tag Archives: sports ranking

Sports ranking methods, 4

Sports ranking methods, 3

Sports ranking methods, 2

Sports ranking methods, 1