Professional Documents
Culture Documents
Scores Ranking of Sports Teams Including
Scores Ranking of Sports Teams Including
Abstract
Examination of play in college football and in professional basketball
finds pronounced differences in the impact of home-field, or home-court,
advantage upon the ranking of teams within the two sports. This difference
results from that in the extent to which team pairings are sampled in the two
cases; the sampling is exhaustive for professional basketball, but extremely
sparse for college football.
2017
c Gary H. Price
1 Introduction
Playing at home generally is agreed to confer an advantage on a sports team. In the
following the significance of this advantage is examined, with particular attention
paid to its impact on the ranking of teams within a league or conference of com-
peting teams. A number of properties of the least-squares ranking methodology
utilized to this end are also pointed out along the way.
Examples from two sports for which the particulars of play differ considerably
are examined. One of these is collegiate football, as exemplified by play in the
Football Bowl Subdivision (FBS) of the National Collegiate Athletic Association
(NCAA) during its 2015 regular season. The other is professional basketball,
as exemplified by play in the National Basketball Association (NBA) during its
2016-2017 regular season. These examples differ greatly both in the number of
games played by each team and in the number of teams participating, as well as in
a consequent difference in the thoroughness with which the games played during
the course of a season sample all pairings of teams.
1
Scores Including HF 2
2 Background
Price (2008) describes using the method of least squares (Whittaker and Robin-
son 1967; Press, Teukolsky, Vetterling, and Flannery 1999; Guthrie, Filliben, and
Heckert 2016) to rank sports teams by best-fit matching of calculated scores to
those recorded by teams in actual games. The teams are then ranked using their
relative strengths, which are defined as the sum for each team of an offensive
strength and a defensive strength. The fit optimizes the values of these quantities
by minimizing the sum of the squared differences between predicted scores and
those recorded for each game, with the predicted score for each team defined to
be the mean of all scores plus the difference between the team’s offensive strength
and its opponent’s defensive strength. Price’s analysis yields a set of linear equa-
tions equivalent to those obtained by others using similar methods (Massey 1997;
Harville 2002; Langville and Meyer 2012).
This model is elaborated here to include in the formulas adopted to calcu-
late scores additional fit parameters representing the home-field advantage (HFA).
Three versions of the elaborated model are described in detail in Appendices A
through C. These range from a single parameter taken to be applicable to all
teams, through defining a separate parameter for each team, to splitting these last
into offensive and defensive components.
3 Models
3.1 General
To rank teams, a fit is made as previously (Price 2008) of predicted scores Pijn to
the scores Sijn achieved by the ith team playing the j th team for the nth time, with
the slight elaboration that scores from games played between the same two teams
here are made explicit rather than being merged into averages over these games.
This elaboration facilitates accommodation of the sums of products that now ap-
pear in the analysis in that these cannot be calculated from the sums involved in
averaging the scores themselves.
The fit parameters again are optimized using the Method of Least Squares
(Whittaker and Robinson 1967; Press, Teukolsky, Vetterling, and Flannery 1999;
Guthrie, Filliben, and Heckert 2016) to minimize the sum of the squares of the
differences between these quantities, denoted the cost C, for games played be-
tween the teams to be ranked. This quantity is written as a sum over all possible
Scores Including HF 3
The insertion of Eq. (2) into Eq. (1) then yields for C
M X Nij h
M X i2
X
C= Sijn − S̄ + Oi − Dj + σijn A . (5)
i=1 j=1 n=1
The complete analysis for this model, with its single HFA parameter, is described
in Appendix A.
The complete analysis for this model, with its HFA parameter for each team, is
described in Appendix B.
The insertion of this Pijn into Eq. (1) then yields for the cost
M X Nij h
M X i2
Aoi − σjin Adj
X
′ ′
C= Sijn − S̄ + Oi − Dj + σijn . (10)
i=1 j=1 n=1
The complete analysis for this model, with pair of offensive and defensive HFA
parameters for each team, is described in Appendix C.
4 Results
4.1 2015 NCAA/FBS Play
A total of 128 teams participated in NCAA/FBS football competition in 2015.
Each team played 12 games during the regular season. Most, but usually not
all, of these games were played between two FBS teams. With the exception of
three independent teams, competition between the teams is concentrated in intra-
conference play, Table 1, with a further focus on intra-division play within each
conference.
Scores Including HF 5
# Conference Games
Conference # Divisions # Teams Total Inter-division
American Athletic 2 12 8 3
Atlantic Coast 2 14 8 2
Big 12 1 10 9 –
Big Ten 2 14 8 2
Conference USA 2 13 8 2 or 3
Mid-America 2 13 8 2 or 3
Mountain West 2 12 8 3
Pac-12 2 12 9 4
Southeastern 2 14 8 2
Sun Belt 1 11 8 –
Table 1: NCAA/FBS Conferences in 2015
Only games played between teams which were both in the FBS, which totaled
716, were included in the analysis. The teams were ranked using each of the
different models described in the appendices together with one lacking the home-
field parameter HFA altogether that had been developed previously (Price 2008).
Scores at the end of regular play were used for games tied at that time, rather than
those reflecting the outcome after overtime play. This choice was made as likely
to yield a more uniform data set across the suite of games played.
The single-HFA model, Appendix A, yields a value of 1.1002 for the home-
field advantage. The individual HFA values calculated for each team by the model
described in Appendix B are distributed as shown in Figure 1.2 They appear
roughly to be centered on the HFA calculated by the single-HFA model; calcu-
lation of the mean of the distribution yields a slightly smaller value of 1.0656,
with a standard deviation of 4.6774. The HFA values calculated for each team
when split into offensive and defensive components are, as is noted in Appendix
C, identical to those calculated without their having been split, Appendix B, if the
overall HFA for each team is taken to be the sum of these two components. The
optimized costs C produced by each model are shown in Table 2.
Even the introduction of a single HFA parameter somewhat affects the rank-
ings in FBS play. As shown by the distribution of the changes plotted in Figure
2, the rankings for about 1/3 of the teams remain unchanged and those for about
another 1/3 are changed by only 1. The remaining larger changes tail off to zero
2
Figures follow the appendices.
Scores Including HF 6
Model Cost
No HFA 1.3552e+05
Single HFA 1.3391e+05
By Team HFA 1.2141e+05
Split By Team HFA 1.0972e+05
Table 2: Optimized Costs for FBS Play in 2015
count quickly, with the extremes of the distribution lying at -6 and at +4. A plot of
the change in rank versus initial3 rank shows changes to be scattered throughout
the rankings, Figure 3, but possibly with some tendency to be less likely for the
teams with poorer rankings.
The changes in rank become pronounced with the introduction of an HFA
parameter for each team, Figure 4. The team rankings for only about 1/10 of the
teams now remain unchanged, and the extremes of the distribution expand to span
a range from -34 to 48. A scatter plot of the change in rank versus initial rank
for this case, Figure 5, shows no obvious dependence of change on the sans-HFA
initial ranking of a team.
again appear roughly to be centered on the HFA value calculated by the single-
HFA model, which is found in this case also to be their mean. The standard
deviation of the distribution is 1.3645. As noted above with regard to play in the
FBS, the individual values of HFA calculated for each team with its HFA split into
offensive and defensive components are identical to those calculated without their
having been split if the overall HFA for each team is taken to be the sum of these
two components. The optimized costs C produced by each model are shown in
Table 3.
Model Cost
No HFA 2.8346e+05
Single HFA 2.7736e+05
By Team HFA 2.7522e+05
Split By Team HFA 2.7097e+05
Table 3: Optimized Costs for NBA Play in 2016–17
5 Discussion
5.1 Home-field Impact
The small number of football games that are played by a team during a season
limits the conclusions that can be drawn from the NCAA/FBS results presented
above. Nonetheless, they do support the existence of a home-field advantage in
football, as also has been inferred by, for example,4 Langville and Meyer (2012,
pp. 120–121). Furthermore, notwithstanding the relatively small magnitude of
this advantage found here, it is seen to be capable of significantly impacting team
rankings.
The results for by-team HFAs in the FBS are even less robust given the sub-
stantial increase in the number of parameters being optimized that accompanies
4
Although in the context of NFL play.
Scores Including HF 8
Particularly remarkable is that the the equations obtained for the sums Ok +Dk
and Aok + Adk in the model for which the team home-field advantages are split into
offensive and defensive components, Appendix C are identical to those obtained
for the Ok + Dk and the Ak in the model, Appendix B, that lacks this splitting.
Consequently the overall team strengths and home-field advantages are seen to be
identical for these two models, with the further consequence that the team rank-
ings, which depend only on the overall team strengths, also are identical. This
circumstance closely parallels one found in the absence of any home-field param-
eter whatsoever in the model, namely the insensitivity of overall team strengths to
the splitting of these strengths into offensive and defensive components discussed
by Price (2017a).
With the separation of the individual HFAs into offensive and defensive com-
ponents, Appendix C, the solution for the differences between these components,
together with those between offensive and defensive team strengths, Ok − Dk ,
proceeds in part from a set of equations, Eqs. (89), that again are closely related to
those for the Ok −Dk , Eqs. (22), obtained in the absence of this splitting, Appendix
B. The difference consists only of the addition to the latter of terms involving the
Aok −Adk . A second set of equations, Eqs. (94), that have no counterpart absent the
splitting of the home-field advantage completes the now necessary complement of
2M equations of constraint that are required to solve for both the Aok − Adk and
the Ok − Dk .
Implementation of the solution to this last set of equations found them also
to be rank deficient, as noted in Appendix C. This circumstance is reminiscent
of that underlying the insensitivity of overall team strengths to their being split
into offensive and defensive components noted above. It is addressed in much the
same way as the latter, namely by replacing one of the rank-deficient equations of
constraint with an independent constraint. That adopted here is to set to zero the
sum of the Aok − Adk across all teams.
One further property of note in the single-HFA parameter model, Appendix
A, is that the organization of play within a competing group of teams so that each
team plays an equal number of games at home and away serves6 to decouple the
team strengths, and hence the team rankings, from the home-field advantage.
6
Cf. Eq. (23).
Scores Including HF 10
6 Conclusions
The relative insensitivity of team rankings to home-field advantage seen in NBA
play are consistent with scheduling there limiting its impact. Conversely, although
somewhat tentatively given the nature of play in football, the significant impact
seen in NCAA/FBS play suggests that the sensitivity of FBS conference play to
the vagaries of schedule explored by Price (2017b) may well be exacerbated by
home-field advantage. The suppression of home-field bias in team rankings pro-
vided by simply insuring that all teams play an equal number of games at home
and away, noted in Section 5.2 above, could usefully inform a rethinking of the
scheduling of play within FBS conferences advocated there.
The separation of home-field advantage into offensive and defensive compo-
nents, at least insofar as least-squares modeling is concerned, serves little purpose
in that its impact upon derived team strengths, and hence rankings, remains un-
changed from that absent such separation.
References
Guthrie, W., J. Filliben, and A. Heckert (2016, 6 April). Least squares.
In C. Croarkin, P. Tobias, and W. Guthrie (Eds.), NIST/SEMATECH e-
Handbook of Statistical Methods. U.S. Department of Commerce, NIST.
http://www.itl.nist.gov/div898/handbook/; Sec. 4.4.3.1.
Harville, D. A. (2002). College football: A modified least squares approach to
rating and predicition. In Proc. Am. Statistical Association 2001 Joint Sta-
tistical Meeting, Section on Statistics in Sports, Alexandria, VA, pp. 1383–
1390. American Statistical Association: American Statistical Association.
CD-ROM.
Harville, D. A. and M. H. Smith (1994, Feb.). The home-court advantage: How
large is it, and does it vary from team to team? American Statistician 48(1),
22–28.
Langville, A. N. and C. D. Meyer (2012). Who’s # 1? The Science of Rating
and Ranking. Princeton, New Jersey 08540: Princeton Univ. Press.
Massey, K. (1997, Spring). Statistical models applied to the rating of sports
teams. Honors project: Mathematics, Bluefield College.
Press, W. H., S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery (1999).
Numerical Recipes in Fortran 77 — The Art of Scientific Computing (Sec-
Scores Including HF 11
ond ed.), Volume 1, Chapter 15. Cambridge, New York, and Melbourne:
Cambridge Univ. Press. Sec. 15.4.
Price, G. H. (2008). Ranking sports teams by comparing scores. J. Recreational
Math. 37(2), 132–143. N.b., actual year of publication of this journal issue,
per copyright notice thereon, is 2012.
Price, G. H. (2017a, July). A comparison of sport team least-square ranking
procedures. Available on the web at Academia.
Price, G. H. (2017b, July). On the ranking of teams at the end of regular-season
play within NCAA FBS conferences. Available on the web at Academia.
Whittaker, E. and G. Robinson (1967). The Calculus of Observations (Fourth
ed.)., Chapter IX. New York: Dover Publications, Inc.
Appendices
where M is the number of teams, and Nij is the number of games played between
the ith and the j th teams.
The predicted scores Pijn are modeled as departures from the average score
calculated for all games played, S̄, using the fit parameters Oi , representing the
7
With terms of the i and j sums for which Nij = 0 taken to be zero.
Scores Including HF 12
offensive ability of the ith team, Dj , representing the defensive ability of the j th
team, and A, the scoring advantage enjoyed by the home team; to wit,
Pijn = S̄ + Oi − Dj + σijn A, (12)
where PM PM PNij
i=1 j=1 n=1 Sijn
S̄ = PM PM (13)
i=1 j=1 Nij
and
−1, team j home field for game n
σijn = 0, neutral field . (14)
1, team i home field for game n
and
M X M X ij N
∂C X h i
= −2 σijn Sijn − S̄ + Oi − Dj + σijn A = 0. (18)
∂A i=1 j=1 n=1
The retention of only selected terms in the summations in Eqs. (16) and (17)
as indicated by the δ functions therein, together with the adoption of a common
summation index for their remaining sums and the elimination of constant factors
throughout, reduces these equations to
M N
X X kj h i
Skjn − S̄ + Ok − Dj + σkjn A = 0, k = 1 : M, (19)
j=1 n=1
Scores Including HF 13
M N
X X jk h i
Sjkn − S̄ + Oj − Dk + σjkn A = 0, k = 1 : M, (20)
j=1 n=1
We next note that Eqs. (19) and (20) can be summed and differenced to yield
two independent sets of equations in Ok + Dk and Ok − Dk in that Nkj = Njk
and σkjn = −σjkn . After some rearrangement, these become
M N
X X kj h i
(Skjn + Sjkn ) − 2S̄ − (Ok − Dk ) − (Oj − Dj ) = 0, k = 1 : M,
j=1 n=1
(22)
and
M N
X X kj h i
(Skjn − Sjkn ) − (Ok + Dk ) + (Oj + Dj ) − 2σkjn A = 0, k = 1 : M.
j=1 n=1
(23)
As the further reduction of Eq. (21) that follows will make evident, its solution
together with Eqs. (23) suffices to rank the teams on the basis of their overall
strengths Ok +Dk and determine the magnitude of the home-field advantage A. In
order to calculate predicted scores using Eq. (12), however, Eqs. (22) also need to
be solved in order to be able to determine the values of the Ok and Dk separately.
To this end, Eq. (21) is recast in a form that involves only O + D sums. This is
accomplished by rewriting the double summation over the square grid that delin-
eates possible team pairings in games played, indexed by i and j, to be taken over
the upper right subdiagonal of this grid, with the lower left subdiagonal elements
of the matrices needed to complete the summation included by pairing terms; i.e.,
M −1 Nij n
M X h i
X X
σijn Sijn − S̄ + Oi − Dj + σijn A
i=1 j=i+1 n=1
h io
+σjin Sjin − S̄ + Oj − Di + σjin A = 0, (24)
where the diagonal terms have been dropped from the summation by virtue of
Nii = 0.
Scores Including HF 14
Attempts to solve Eq. (26) together with Eqs. (23) as they stand for A and the
Ok +Dk will be found to fail, however, in that, notwithstanding the addition of the
new parameter A, they remain as previously (Harville 2002; Price 2008) nonfull
rank in the Ok + Dk . This difficulty can be addressed by replacing one of these
equations with the condition that
M X Nij
M X M M
X X X
(Oi + Di ) = (Oi + Di ) Nij = 0, (27)
i=1 j=1 n=1 i=i j=1
a constraint chosen in light of our having defined the Ok and Dk as describing the
departures of the individual scores from the mean score S̄ (Price 2008). Which
equation is selected to be replaced will be found not to matter, as the problem is
that Eqs. (23) are sufficient only to determine the differences between overall team
strengths — a circumstance here reflected in the appearance within them of the
Ok + Dk only in pairs delineating such differences. The substitution of Eq. (27)
for any one of these equations selects, from the infinity of possible solutions that
are differentiated from one another only by uniform shifts of all the Ok + Dk , the
one that satisfies this additional constraint.
It is convenient for computational purposes to rewrite Eqs. (23) and (26 in
a manner that makes more directly evident the matrix elements when the set of
equations to be solved is written in matrix form as Ax = b. The solution then
follows immediately as x = A−1 b provided A is not singular.
Reorganizing our equations to this end, we have for Eqs. (23)
M N
X X kj h i M N
X X kj
which, with reordering of the summation over the Oj + Dj and some rearrange-
ment, becomes
M −1 Nij
M X M j−1 Nij
X X X XX
(Oi + Di ) σijn − (Oj + Dj ) σijn
i=1 j=i+1 n=1 j=2 i=1 n=1
M −1 Nij
M X M −1 Nij
M X h i
2
X X X X
+2A σijn = σijn Sijn − Sjin . (30)
i=1 j=i+1 n=1 i=1 j=i+1 n=1
M N
X X kj M
X
Ak,k = (1) = Nkj , k = 1 : M; (33)
j=1 n=1 j=1
Nkj
X
Akj = − (1) = −Nkj , k = 1 : M, j (6= k) = 1 : M ; (34)
n=1
M N
X X kj
and
M N
X X kj
Nij
M X Nji
i−1 X
X X
AM +1,i = σijn − σjin , i = 2 : M − 1; (38)
j=i+1 n=1 j=1 n=1
M
X −1 NX
j,M
M −1 Nij
M X
2
X X
AM +1,M +1 = 2 σijn ; (40)
i=1 j=i+1 n=1
and
M −1 Nij
M X h i
X X
bM +1 = σijn Sijn − Sjin . (41)
i=1 j=i+1 n=1
Ac,M +1 = 0, (43)
and
bc = 0, (44)
where the row index c indicates which of Eqs. (28) is replaced by Eq. (27).
As noted above, Eqs. (22) also need to be solved for the Ok − Dk in order to
calculate game scores. Upon rewriting these equations to facilitate identification
of the elements in the corresponding matrix equation Ax = b one obtains
M N
X X kj h i M N
X X kj h i
(Ok − Dk ) + (Oj − Dj ) = (Skjn + Sjkn ) − 2S̄ ,
j=1 n=1 j=1 n=1
k = 1 : M. (45)
Scores Including HF 17
Setting to zero the partial derivative of C with respect to each of the fit parameters
parameters now yields for our equations of condition
M X M X ij N
∂C X h
′ ′
i
= −2 δik Sijn − S̄ + Oi − Dj + σijn Ai − σjin Aj
∂Ok k=1:M i=1 j=1 n=1
= 0, (53)
M X M X ij N
∂C X h
′ ′
i
=2 δjk Sijn − S̄ + Oi − Dj + σijn Ai − σjin Aj = 0,
∂Dk k=1:M i=1 j=1 n=1
(54)
and
∂C
=
∂Ak k=1:M
M X Nij
M X h i
X
′ ′ ′
−2 δik σijn Sijn − S̄ + Oi − Dj + σijn Ai − σjin Aj
i=1 j=1 n=1
M X Nij
M X h i
X
′ ′ ′
+2 δjk σjin Sijn − S̄ + Oi − Dj + σijn Ai − σjin Aj
i=1 j=1 n=1
=0, (55)
in the last of which the sums have been written separately for the two derivative
terms in order to facilitate their manipulation. Upon further noting, from Eq. (51),
′
that σijn ′
σjin = 0, this last equation becomes
M X M X ij N
∂C X
′
h
′
i
= −2 δik σijn Sijn − S̄ + Oi − Dj + σijn Ai
∂Ak k=1:M i=1 j=1 n=1
M X Nij
M X h i
X
′ ′
+2 δjk σjin Sijn − S̄ + Oi − Dj − σjin Aj
i=1 j=1 n=1
=0. (56)
The retention of only selected terms in the summations in Eqs. (53, 54 & 56)
as indicated by the δ functions therein, together with the adoption of a common
summation index for their remaining sums and the elimination of constant factors
throughout, reduces these equations to
M N
X X kj h i
′ ′
Skjn − S̄ + Ok − Dj + σkjn Ak − σjkn Aj = 0, k = 1 : M, (57)
j=1 n=1
Scores Including HF 19
M N
X X jk h i
′ ′
Sjkn − S̄ + Oj − Dk + σjkn Aj − σkjn Ak = 0, k = 1 : M, (58)
j=1 n=1
and
M N
X X kj h i
′ ′
σkjn Skjn − S̄ + Ok − Dj + σkjn Ak
j=1 n=1
M N
X X jk h i
′ ′
− σkjn Sjkn − S̄ + Oj − Dk − σkjn Ak = 0,
j=1 n=1
k = 1 : M. (59)
We next note that Eqs. (57) and (58) can be summed and differenced to yield
two independent sets of equations in Ok + Dk and Ok − Dk in that Nkj = Njk .
After some rearrangement, these become
M N
X X kj h i
(Skjn + Sjkn ) − 2S̄ − (Ok − Dk ) − (Oj − Dj ) = 0, k = 1 : M,
j=1 n=1
(60)
and
M N
X X kj h i
′ ′
(Skjn − Sjkn ) − (Ok + Dk ) + (Oj + Dj ) − 2σkjn Ak + 2σjkn Aj
j=1 n=1
= 0, k = 1 : M. (61)
The solution of Eqs. (61) together with Eqs. (62) suffices to rank the teams on
the basis of their overall strengths Ok + Dk and determine the magnitudes of the
home-field advantages Ak . In order to calculate predicted scores using Eq. (50),
however, Eqs. (60) also need to be solved for the Ok − Dk so that the values of
the Ok and Dk can be determined separately.
Attempts to solve Eqs. (61) and Eqs. (62) as they stand for the Ak and the
Ok +Dk will again be found to fail in that they remain nonfull rank in the Ok +Dk .
Scores Including HF 20
M N
X X kj M
X
Akk = (1) = Nkj , k = 1 : M; (65)
j=1 n=1 j=1
Nkj
X
Akj = − (1) = −Nkj , k = 1 : M, j (6= k) = 1 : M; (66)
n=1
M N
X X kj
′
Ak,M +k = 2 σkjn , k = 1 : M; (67)
j=1 n=1
and
Nkj
X
′
Ak,M +j = −2 σjkn , k = 1 : M, j (6= k) = 1 : M; (68)
n=1
where j 6= k follows from Nkk = 0 for Eqs. (66) and (68). Finally, Eqs. (63) yield
for b
M N
X X kj
Nkj
X
′
AM +k,j = − σkjn , k = 1 : M, j (6= k) = 1 : M; (71)
n=1
M N kj
)2 ,
X X
′
AM +k,M +k = 2 (σkjn k = 1 : M; (72)
j=1 n=1
Ac,M +j = 0, j = 1 : M; (76)
and
bc = 0, (77)
where the row index c again indicates which of Eqs. (63) is replaced by Eq. (27).
As noted above, Eqs. (60) also need to be solved for the Ok − Dk in order to
calculate game scores. Upon rewriting these equations to facilitate identification
of the elements in the corresponding matrix equation A− x− = b− one obtains
M N
X X kj h i M N
X X kj h i
(Ok − Dk ) + (Oj − Dj ) = (Skjn + Sjkn ) − 2S̄ ,
j=1 n=1 j=1 n=1
k = 1 : M. (78)
M N
X X kj M
X
A−
kk = (1) = Nkj , k = 1 : M; (80)
j=1 n=1 j=1
Nkj
X
A−
kj = (1) = Nkj , k = 1 : M, j (6= k) = 1 : M; (81)
n=1
and
M N
X X kj h i
b−
k = (Skjn + Sjkn ) − 2S̄ , k = 1 : M. (82)
j=1 n=1
With this elaboration, the derivatives of C with respect to the offensive (O) and
defensive (D) components of the team strength, Eqs. (53) and (54), respectively,
become
M X M X ij N
∂C h i
Aoi − σjin Adj
X
′ ′
= −2 δik Sijn − S̄ + Oi − Dj + σijn
∂Ok k=1:M i=1 j=1 n=1
= 0 (85)
and
M X M X ij N
∂C h i
Aoi − σjin Adj = 0.
X
′ ′
=2 δjk Sijn − S̄ + Oi − Dj + σijn
∂Dk k=1:M i=1 j=1 n=1
(86)
Scores Including HF 23
and
M N jk h i
Aoj − σkjn Adk
X X ′ ′
Sjkn − S̄ + Oj − Dk + σjkn = 0, k = 1 : M. (88)
j=1 n=1
and, differencing,
M N
X X kj h
The derivatives of C with respect to the Aok and Adk involve a bit more manip-
ulation than do those of C with respect to the Ak , Eqs. (55), in that they produce
two sets of equations in place of the single set with two terms that results for the
latter. Namely,
M N kj h i
Aok
X X
′ ′
σkjn Skjn − S̄ + Ok − Dj + σkjn = 0, k = 1 : M, (91)
j=1 n=1
and
M N jk h i
Adk
X X
′ ′
σkjn Sjkn − S̄ + Oj − Dk − σkjn = 0, k = 1 : M. (92)
j=1 n=1
after the usual reductions and the additional simplification that is provided by
′
noting that σijn ′
σjin = 0 in that one or the other of them must be zero (cf. Eq. 51).
Scores Including HF 24
These equations can, as with those resulting from taking derivatives of C with
respect to Ok and to Dk , be summed and differenced to yield two sets of equations
of which one involves only sums of the Aok and Adk as well as of the Ok and Dk ,
and the other only differences. Namely, differencing,
M N
X X kj h
′
σkjn (Skjn − Sjkn ) − (Ok + Dk ) + (Oj + Dj )
j=1 n=1
i
′
−σkjn Aok + Adk = 0, k = 1 : M, (93)
and, summing,
M N
X X kj h
′
σkjn (Skjn + Sjkn ) − 2S̄ − (Ok − Dk ) − (Oj − Dj )
j=1 n=1
i
′
−σkjn Aok − Adk = 0, k = 1 : M. (94)
One has at this point two independent sets of equations to be solved, one for
the sums Ok +Dk and the Aok +Adk , namely Eqs. (90) & (93), and the other, namely
Eqs. (89) & (94), for the differences Ok − Dk and Aok − Adk . For the former, we
further note that, upon taking Aok + Adk ≡ 2Ak , the equations are identical to those
obtained in Appendix B for the Ok + Dk and the Ak in the absence of the splitting
of the home-field advantages into offensive and defensive components, namely
Eqs. (23) & (24). Consequently, the solution for these sums remains unchanged
from that described in Appendix B.
For the differences, upon reorganizing Eqs. (89) & (94) to facilitate their re-
statement in the matrix form Ax = b, one has for the former
M N kj h i
Aok − Adk + σjkn Aoj − Adj
X X
′ ′
(Ok − Dk ) + (Oj − Dj ) + σkjn
j=1 n=1
M N
X X kj h i
= (Skjn + Sjkn) − 2S̄ , k = 1 : M, (95)
j=1 n=1
Nkj
X
Akj = (1) = Nkj , k = 1 : M, j (6= k) = 1 : M; (98)
n=1
M N
X X kj
′
Ak,k+M = σkjn , k = 1 : M; (99)
j=1 n=1
Nkj
X
′
Ak,j+M = σjkn , k = 1 : M, j (6= k) = 1 : M; (100)
n=1
and
M N
X X kj h i
bk = (Skjn + Sjkn) − 2S̄ , k = 1 : M. (101)
j=1 n=1
Nkj
X
′
Ak+M,j = σkjn , k = 1 : M, j (6= k) = 1 : M; (103)
n=1
M N kj
)2 , k = 1 : M;
X X
′
Ak+M,k+M = (σkjn (104)
j=1 n=1
These last equations, however, were found to be rank deficient upon attempt-
ing to calculate A−1 using them as they stand. Removal of this singularity can
again, as with the team strengths, be accomplished by replacement of one of the
equations with an independent equation of condition. That chosen is to set to zero
Scores Including HF 26
the sum over teams of the differences between offensive and defensive compo-
nents of the home-field advantages; i.e.,
M
Aocj − Adcj ,
X
0= (107)
j=1
where the row subscript c denotes which of the equations of condition is replaced.
From Eq. (107) follows
Acj = 0, j = 1 : M; (108)
Ac,j+M = 1, j = 1 : M; (109)
and
bc = 0. (110)
Scores Including HF 27
Figures
25
20
15
count
10
0
-11 -9 -7 -5 -3 -1 1 3 5 7 9 11 13
HFA
50
40
30
count
20
10
0
-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
change in rank
Figure 2: Distribution of changes in rank going from sans HFA to single HFA for
2015 FBS play.
Scores Including HF 29
4
change in rank for single HFA
-2
-4
-6
0 20 40 60 80 100 120
initial rank -- sans HFA
Figure 3: Changes in rank vs. initial rank going from sans HFA to single HFA for
2015 FBS play.
Scores Including HF 30
14
12
10
8
count
0
-50 -40 -30 -20 -10 0 10 20 30 40 50
change in rank
Figure 4: Distribution of changes in rank going from sans HFA to by-team HFAs
for 2015 FBS play.
Scores Including HF 31
40
20
change in rank for by-team HFA
-20
-40
0 20 40 60 80 100 120
initial rank -- sans HFA
Figure 5: Changes in rank vs. initial rank going from sans HFA to by-team HFA
for 2015 FBS play.
Scores Including HF 32
10
6
count
0
-3 -2 -1 0 1 2 3 4 5
HFA
10
6
count
0
-10 -8 -6 -4 -2 0 2 4 6 8 10
change in rank
Figure 7: Changes in rank going from sans HFA to by-team HFAs for 2016–17
NBA play.