Team #30680

Summary
The objective of our team was to find the five best coaches of the last 100 years for three
different college sports. Our team decided to look at men’s basketball, football, and baseball. We
wanted to be able to definitively determine team skill from the games played, and then use a machine-
learning algorithm to calculate the correct coach skills for each team in a given year. We created a
networks-based model to calculate team skill from historical game data. For basketball, we were able
to obtain the final scores of every single Div. I game played from 1939 to 2013. The range for football
was even larger: 1869 to 2013. The data for baseball was sparse. We were only able to obtain the final
scores of tournament games from 1947 to 2013. A digraph was created for each year in each sport.
Nodes represented teams, and edges represented a game played between two teams. The arrowhead
pointed towards the losing team.
We calculated the team skill of each graph using a right-hand eigenvector centrality measure.
Eigenvector centrality calculates the relative ‘importance’ of a node based on both the number of
connections and the importance of the nodes that it is connected to. In this way, teams that beat good
teams will be ranked higher than teams that beat mediocre teams. The eigenvector centrality rankings
for most years were well correlated with tournament performance and poll-based rankings.
We assumed that the relationship between coach skill (𝐶! ), player skill (𝑃! ), and team skill (𝑇! )
was this: 𝐶! ∗ 𝑃! = 𝑇! . Our team then created a function to describe the probability that a given score
difference would occur based on player skill and coach skill. Our rationale was this: if two teams have
unevenly matched players, coach skill will likely not influence the outcome of the game. However, if
two teams have evenly matched players, coach skill will manifest itself in player substitutions, time-
outs, etc. and will determine the team that won and the score difference. We multiplied the probabilities
of all edges in the network together to find the probability that the correct network would occur with any
given player skill and coach skill matrix. Our team was able to determine player skill as a function of
team skill and coach skill, eliminating the need to optimize two unknown matrices. The top five
coaches in each year were noted, and the top coach of all time was calculated by dividing the number
of times that coach ranked in the yearly top five by the years said coach had been active. The top five
coaches in the last century are: Basketball: John Wooden (0.28), Lute Olson (0.26), Jim
Boeheim (0.24), Gregg Marshall (0.23), and Jamie Dixon (0.21). Football: Glenn Scobey Warner
(0.24), Bobby Bowden (0.23), Jim Grobe (0.18), Bob Stoops (0.17), and Bill Peterson (0.16).
Baseball: Mark Marquess (0.27), Augie Garrido (0.24), Tom Chandler (0.22), Richard Jones
(0.19), and Bill Walkenbach (0.16).

更多数学建模资料请关注微店店铺“数学建模学习交流”
https://k.weidian.com/RHO6PSpA
Who’s  the  Best  Coach?  
By:  MCM  Team  #30680  
 
As  football  and  basketball  legends  like  Nick  Saban  and  Mike  Krzyzewski  fight  hard  
on  the  court  to  keep  their  reputations  high,  those  of  us  at  home  would  like  to  ask:  
are  these  reputations  justified?  Does  being  well-­‐known  necessarily  make  a  coach  
better?  Is  the  best  coach,  contrary  to  popular  belief,  someone  who  is  less  well-­‐
known  than  some  of  their  colleagues?  We  are  going  to  set  out  to  find  the  best  coach  
of  the  last  one  hundred  years  in  the  college  sports  of  basketball,  football,  and  
baseball.  We  will  utilize  some  cool  mathematics  during  our  journey  to  make  our  
rankings  less  biased  than  other  rankings  like  the  AP  Poll  or  USA  Today.  
 
We  will  start  by,  for  every  year  in  the  last  century,  ranking  the  college  sports  teams  
that  play  in  the  top  college  division.  We  do  this  in  a  really  interesting  way  that  is  
quite  simple  to  understand.  Let’s  say  you  have  two  teams  that  are  playing  a  game  
against  one  another.  On  a  piece  of  paper,  draw  two  circles  representing  the  two  
teams.  Then,  draw  an  arrow  from  one  circle  to  the  other.  The  arrow  head  will  point  
to  the  team  that  lost  the  game.  Now,  we  have  created  a  very  simple  drawing  that  can  
represent  this  game.  
 
 
 
Now  imagine  that  this  process  is  repeated  for  all  5000  games  played  in  the  2012-­‐
2013  basketball  season.  The  resulting  drawing  will  look  something  like  this:  

 
 
From  this  drawing,  we  can  see  exactly  which  teams  beat  which  teams.  From  this,  we  
can  use  various  different  graph  metrics  to  calculate  the  best  teams.  In  the  drawing  
above,  the  best  teams  have  the  biggest  circles  drawn.    
 
So  what  does  the  best  coach  have  to  do  with  any  of  this?  Well,  we  think  that  coaches  
that  lead  the  highest  ranked  teams  must  be  doing  something  right.  However,  we  
aren’t  just  going  to  give  the  highest  ranked  teams  the  distinction  of  having  the  best  
coaches.  The  highest  ranked  teams  will  on  average  have  better  coaches  –  sure,  but  
there’s  no  reason  that  the  #2  team  won’t  have  a  better  coach  than  the  #1  team.    
 
We  think  that  it  is  pretty  safe  to  say  that,  if  the  players  of  one  team  are  much  worse  
than  the  players  of  one  of  their  opponents,  even  the  best  coaching  won’t  help  the  
team  pull  off  a  win.  However,  when  two  teams  are  evenly  matched  and  they  face  
each  other,  we  think  that  coaching  actually  has  a  big  impact  on  who  wins  the  game.  
Think  about  it  this  way:  If  the  players  of  two  teams  are  evenly  matched,  then  the  
game  is  probably  going  to  be  close  all  the  way  from  the  start  to  the  end.  The  coaches  
of  either  team  will  be  the  ones  determining  who  wins  the  game  through  things  like  
the  the  timing  of  player  substitutions  and  time-­‐outs,  as  well  as  effectively  raising  
team  morale  and  calling  good  plays.    
 
Therefore,  when  two  teams  that  are  evenly  matched  (we  determine  ‘evenly  
matched’  based  upon  the  ranking  system  described  above),  the  team  that  wins  will  
have  a  better  coach  than  the  team  that  loses.  If  we  calculate  this  for  any  year  in  a  
college  sport,  we  will  get  a  list  of  the  top  coaches.  We  can  do  this  up  to  100  years  in  
the  past  and  see  which  coaches  show  up  highest  on  the  list.  These  are  then  
determined,  through  the  power  of  mathematics,  to  be  the  best  coaches  in  said  sport  
for  the  last  100  years.  
 
Everything  you  just  read  wasn’t  just  an  exercise  in  thought.  We  have  actually  gone  
through  this  entire  process,  just  as  it  was  described  to  you  (it  is  actually  a  little  more  
complicated  than  this).  We  were  able  to  come  up  with  the  best  coaches  of  the  last  
100  years  for  the  college  sports  of  mens  basketball,  football,  and  baseball.  These  
coaches  are:  
 
Rank   Basketball   Football   Baseball  
1   John  Wooden   Glenn  Scobey  Warner   Mark  Marquess  
2   Lute  Olson   Bobby  Bowden   Augie  Garrido  
3   Jim  Boeheim   Jim  Grobe   Tom  Chandler  
4   Gregg  Marshall   Bob  Stoops   Richard  Jones  
5   Jamie  Dixon   Bill  Peterson   Bill  Walkenbach  
 
Thanks  so  much  for  reading!  Agree  with  our  results?  Disagree?  Feel  free  to  send  us  
an  email  at:  [email protected].  We’re  always  happy  to  hear  from  our  readers.  
 
 
A Networks and Machine Learning Approach to
Determine the Best College Coaches of the
20th-21st Centuries
Team #30680
February 10, 2014

1
Team #30680 Page 2 of 18

Contents
1 Problem Statement 3

2 Planned Approach 3

3 Assumptions 3

4 Data Sources and Collection 4

4.1 College Football . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
4.2 Men’s College Basketball . . . . . . . . . . . . . . . . . . . . . . . 5
4.3 College Baseball . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

5 Network-based Model for Team Ranking 6

5.1 Building the Network . . . . . . . . . . . . . . . . . . . . . . . . 6
5.2 Analyzing the Network . . . . . . . . . . . . . . . . . . . . . . . . 6
5.2.1 Degree Centrality . . . . . . . . . . . . . . . . . . . . . . . 6
5.2.2 Betweenness and Closeness Centrality . . . . . . . . . . . 7
5.2.3 Eigenvector Centrality . . . . . . . . . . . . . . . . . . . . 8

6 Separating the Coach Effect 10

6.1 When is Coach Skill Important? . . . . . . . . . . . . . . . . . . 11
6.2 Margin of Win Probability . . . . . . . . . . . . . . . . . . . . . . 12
6.3 Optimizing the Probability Function . . . . . . . . . . . . . . . . 13
6.3.1 Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . 13
6.3.2 Nelder-Mead Method . . . . . . . . . . . . . . . . . . . . 14
6.3.3 Powell’s Method . . . . . . . . . . . . . . . . . . . . . . . 14

7 Ranking Coaches 15
7.1 Top Coaches of the Last 100 Years . . . . . . . . . . . . . . . . . 15

8 Testing our Model 15

8.1 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 15
8.2 Strengths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
8.3 Weaknesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

9 Conclusions 17

2
Team #30680 Page 3 of 18

1 Problem Statement
College sport coaches often achieve widespread recognition. Coaches like Nick
Saban in football and Mike Krzyzewski in basketball repeatedly lead their
schools to national championships. Because coaches influence both the per-
formance and reputation of the teams they lead, a question of great concern to
universities, players, and fans alike is: Who is the best coach in a given sport?
Sports Illustrated, a magazine for sports enthusiasts, has asked us to find the
best all-time college coaches for the previous century. We are tasked with creat-
ing a model that can be applied in general across both genders and all possible
sports at the college-level. The solution proposed within this paper will offer an
insight to these problems and will objectively determine the top five coaches of
all time in the sports of baseball, men’s basketball, and football.

2 Planned Approach
Our objective is to rank the top 5 coaches in each of 3 different college-level
sports. We need to determine which metrics reflect most accurately the ranking
of coaches within the last 100 years. To determine the most effective ranking
system, we will proceed as follows:

1. Create a network-based model to visualize all college sports teams, the

teams won/lost against, and the margin of win/loss. Each network de-
scribes the games of one sport over a single year.
2. Analyze various properties of the network in order to calculate the skill of
each team.

3. Develop a means by which to decouple the effect of the coach from the
team performance.
4. Create a model that, given the player and coach skills for every team, can
predict the probability of the occurrence of a specific network of a) wins
and losses and b) the point margin with which a win or loss occurred.

5. Utilize an optimization algorithm to maximize the probability that the

coach skill matrix, once plugged into our model, generates the network of
wins/losses and margins described in (1).
6. Analyze the results of the optimization algorithm for each year to deter-
mine an overall ranking for all coaches across history.

3 Assumptions
Due to limited data about the coaching habits of all coaches at all teams over
the last century in various collegiate sports, we use the following assumptions to

3
Team #30680 Page 4 of 18

complete our model. These simplifying assumptions will be used in our report
and can be replaced with more reliable data when it becomes available.

• The skill level of a coach is ultimately expressed through his/her team’s

wins over another and the margin by which they win. This assumes that
a team must win to a certain degree for their coach to be good. Even if
the coach significantly amplifies the skills of his/her players, he/she still
cannot be considered “good” if the team wins no games.

• The skills of teams are constant throughout any given year (ex: No players
are injured in the middle of a season). This assumption will allow us to
compare a team’s games from any point in the season to any other point in
the season. In reality, changing player skills throughout the season make
it more difficult to determine the effect of the coach on a game.

• Winning k games against a good team improves team skill more than
winning k games against an average team. This assumption is intuitive
and allows us to use the eigenvector centrality metric as a measure of total
team skill.
• The skill of a team is a function of the skill of the players and the skill
of the coach. We assume that the skill of a coach is multiplicative over
the skill of the players. That is: Ts = Cs · Ps where Ts is the skill of
the team, Cs is the skill of the coach, and Ps is a measure of the skill of
the players. Making coach skill multiplicative over player skill assumes
that the coach has the same effect on each player. This assumption is
important because it simplifies the relationship between player and coach
skill to a point where we can easily optimize coach skill vectors.
• The effect of coach skill is only large when the difference between player
skill is small. For example, if team A has the best players in the conference
and team B has the worst, it is likely that even the best coach would not
be able to, in the short run, bring about wins over team A. However,
if two teams are similarly matched in players, a more-skilled coach will
make advantageous plays that lead to his/her team winning more often
than not.
• When player skills between two teams are similarly matched, coach skill
is the only factor that determines the team that wins and the margin by
which they win by. By making this assumption, we do not have to account
for any other factors.

4 Data Sources and Collection

Since our model requires as an input the results of all the games played in a
season of a particular sport, we first set out to collect this data. Since we were
unable to identify a single resource that had all of the data that we required, we

4
Team #30680 Page 5 of 18

found a number of different websites, each with a portion of the requisite data.
For each of these websites, we created a customized program to scrape the data
from the relevant webpages. Once we gathered all the data from our sources, we
processed it to standardize the formatting. We then aimed to merge the data
gathered from each source into a useable format. For example, we gathered
basketball game results from one source, and data identifying team coaches
from another. To merge them and show the game data for a specific coach, we
attempted to match on common fields (ex. “Team Name”). Often, however, the
data from each source did not match exactly (ex. “Florida State” vs “Florida
St.”). In these situations, we had to manually create a matching table that
would allow our program to merge the data sources.
Although we are seeking to identify the best college coach for each sport
of interest for the last century, it should be noted that many current college
sports did not exist a century ago. The National Collegiate Athletic Association
(NCAA), the current managing body for nearly all college athletics, was only
officially established in 1906 and the first NCAA national championship took
place in 1921, 7 years short of a century ago. Although some college sports
were independently managed before being brought into the NCAA, it is often
difficult to gather accurate data for this time.

4.1 College Football

One of the earliest college sports, College Football has been popular since its
inception in the 1800’s. The data that we collected ranges from 1869 to the
present, and includes the results and final scores of every game played between
Division 1 men’s college football teams (or the equivalent before the inception
of NCAA) [2]. Additionally, we have gathered data listing the coach of each
team for every year we have collected game data [4], and combined the data in
order to match the coach with his/her complete game record for every year that
data was available.

4.2 Men’s College Basketball

The data that we gathered for Men’s College Basketball ranges from the sea-
son of the first NCAA Men’s Basketball championship in 1939 to the present.
Similarly to College Football, we gathered data on the result and final scores of
each game in the season and in finals [2]. Combining this with another source of
coach names for each team and year generated the game record for each coach
for each season [4].

4.3 College Baseball

Although College Baseball has historically had limited popularity, interest in
the sport has grown greatly in the past decades with improved media coverage
and collegiate spending on the sport. The game result data that we collected

5
Team #30680 Page 6 of 18

ranges from 1949 to the present, and was merged with coach data for the same
time period.

5 Network-based Model for Team Ranking

Through examination of all games played for a specific year we can accurately
rank teams for that year. By creating a network of teams and games played,
we can not only analyze the number of wins and losses each team had, but can
also break down each win/loss with regard to the opponent’s skill.

5.1 Building the Network

We made use of a weighted digraph to represent all games played in a single
year. Each node in the graph represents a single college sports team. If team
A wins over team B, a directed edge with a weight of 1 will be drawn from
A pointing towards B. Each additional time A wins over B, the weight of the
edge will be increased by 1. If B beats A, an edge with the same information
is drawn in the opposing direction. Additionally, a list containing the margin
of win/loss for each game is associated with the edge. For example, if A beat
B twice with score : 64 − 60, 55 − 40, an edge with weight two is constructed
and the winning margin list 4, 15 is associated with the edge. Since each graph
represents a single season of a specific sport, and we are interested in analyzing
a century of data about three different sports, we have created a program to
automate the creation of the nearly 300 graphs used to model this system. The
program Gephi was used to visualize and manipulate the generated graphs.

5.2 Analyzing the Network

We are next interested in calculating the skill of each team based on the graphs
generated in the previous section. To do this, we will use the concept of central-
ity to investigate the properties of the nodes and their connections. Centrality is
a measure of the relative importance of a specific node on a graph based on the
connections to and from that node. There are a number of ways to calculate
centrality, but the four main measures of centrality are degree, betweenness,
closeness, and eigenvector centrality.

5.2.1 Degree Centrality

Degree centrality is the simplest centrality measure, and is simply the total
number of edges connecting to a specific node. For a directional graph, indegree
is the number of edges directed into the node, while outdegree is the number
of edges directed away from the node. Since in our network, edges directed
inward are losses and edges directed outwards are wins, indegree represents
the total number of losses and outdegree measures the total number of wins.
Logically, therefore, outdegreee win
indegreee represents the loss ratio of the team. This ratio
is often used as a metric of the skill of a team; however, there are several

6
Team #30680 Page 7 of 18

Figure 1: A complete network for the 2009-2010 NCAA Div. I basketball season.
Each node represents a team, and each edge represents a game between the two
teams. Note that, since teams play other teams in their conference most often,
many teams have clustered into one of the 32 NCAA Div.1 Conferences.

weaknesses to this metric. The most prominent of these weaknesses arises from
the fact that, since not every team plays every other team over the course of
the season, some teams will naturally play more difficult teams while others
will play less difficult teams. This is exaggerated by the fact that many college
sports are arranged into conferences, with some conferences containing mostly
highly-ranked teams and others containing mostly low-ranked teams. Therefore,
win/loss percentage often exaggerates the skill of teams in weaker conferences
while failing to highlight teams in more difficult conferences.

5.2.2 Betweenness and Closeness Centrality

Betweenness centrality is defined as a measure of how often a specific node
acts as a bridge along the shortest path between two other nodes in the graph.
Although a very useful metric in, for example, social networks, betweenness
centrality is less relevant in our graphs as the distance between nodes is based
on the game schedule and conference layout, and not on team skill. Similarly,
closeness centrality is a measure of the average distance of a specific node to

7
Team #30680 Page 8 of 18

another node in the graph - also not particularly relevant in our graphs because
distance between nodes is not related to team skills.

5.2.3 Eigenvector Centrality

Eigenvector centrality is a measure of the influence of a node in a network
based on its connections to other nodes. However, instead of each connection
to another node having a fixed contribution to the centrality rating (e.g. de-
gree centrality), the contribution of each connection in eigenvector centrality
is proportional to the eigenvector centrality of the node being connected to.
Therefore, connections to high-ranked nodes will have a greater influence on
the ranking of a node than connections to low-ranking nodes. When applied
to our graph, the metric of eigenvector centrality will assign a higher ranking
to teams that win over other high-ranking teams, while winning over lower-
ranking nodes has a lesser contribution. This is important because it addresses
the main limitation over degree centrality or win/loss percentage, where winning
over many low-ranked teams can give a team a high rank.
If we let G represent a graph with nodes N , and let A = (an,t ) be an
adjacency matrix where an,t = 1 if node n is connected to node t and an,t = 0
otherwise. If we define xa as the eigenvector centrality score of node a, then the
eigenvector centrality score of node n is given by:
1 X 1X
xn = xt = an,t xt (1)
λ λ
t∈M (n) t∈G

where λ represents a constant and M (n) represents the set of neighbors of

node n.
If we convert this equation into vector notation, we find that this equation
is identical to the eigenvector equation:

Ax = λx (2)
If we place the restriction that the ranking of each node must be positive,
we find that there is a unique solution for the eigenvector x, where the nth
component of x represents the ranking of node n. There are multiple different
methods of calculating x; most of them are iterative methods that converge on a
final value of x after numerous iterations. One interesting and intuitive method
of calculating the eigenvector x is highlighted below. It has been shown that
the eigenvector x is proportional to the row sums of a matrix S formed by the
following equation [6, 9]:

S = A + λ−1 A2 + λ−2 A3 + ... + λn−1 An + ... (3)

where A is the adjacency matrix of the network and λ is a constant (the
principle eigenvalue). We know that the powers of an adjacency matrix describe
the number of walks of a certain length from node to node. The power of
the eigenvalue (x) describes some function of length. Therefore, S and the

8
Team #30680 Page 9 of 18

eigenvector centrality matrix both describe the number of walks of all lengths
weighted inversely by the length of the walk. This explanation is an intuitive
way to describe the eigenvector centrality metric. We utilized NetworkX (a
Python library) to calculate the eigenvector centrality measure for our sports
game networks.
We can apply eigenvector centrality in the context of this problem because
it takes into account both the number of wins and losses and whether those
wins and losses were against “good” or “bad” teams. If we have the following
graph: A → B → C and know that C is a good team, it follows that A is also
a good team because they beat a team who then went on to beat C. This is
an example of the kind of interaction that the metric of eigenvector centrality
takes into account. Calculating this metric over the entire yearly graph, we can
create a list of teams ranked by eigenvector centrality that is quite accurate.
Below is a table of top ranks from eigenvector centrality compared to the AP
and USA Today polls for a random sample of our data, the 2009-2010 NCAA
Division I Mens Basketball season. It shows that eigenvector centrality creates
an accurate ranking of college basketball teams. The italicized entries are ones
that appear in the top ten of both eigenvector centrality ranking and one of the
AP and USA Today polls.

Rank Eigenvector Centrality AP Poll USA Today Poll

1 Duke Kansas Kansas
2 West Virginia Michigan St. Michigan St.
3 Kansas Texas Texas
4 Syracuse Kentucky North Carolina
5 Purdue Villanova Kentucky
6 Georgetown North Carolina Villanova
7 Ohio St. Purdue Purdue
8 Washington West Virginia Duke
9 Kentucky Duke West Virginia
10 Kansas St. Tennessee Butler

As seen in the table above, six out of the top ten teams as determined by
eigenvector centrality are also found on the top ten rankings list of popular polls
such as AP and USA Today. We can see that the metric we have created using
a networks-based model creates results that affirms the results of commonly-
accepted rankings. Our team-ranking model has a clear, easy-to-understand
basis in networks-based centrality measures and gives reasonably accurate re-
sults. It should be noted that we chose this approach to ranking teams over
a much simpler approach such as simply gathering the AP rankings for vari-
ous reasons, one of which is that there are not reliable sources of college sport
ranking data that cover the entire history of the sports we are interested in.
Therefore, by calculating the rankings ourselves, we can analyze a wider range
of historical data.
Below is a graph that visualizes the eigenvector centrality values for all
games played in the 2010-2011 NCAA Division I Mens Football tournament.

9
Team #30680 Page 10 of 18

Larger and darker nodes represent teams that have high eigenvector centrality
values, while smaller and lighter nodes represent teams that have low eigenvector
centrality values. The large nodes therefore represent the best teams in the
2010-2011 season.

Figure 2: A complete network for the 2012-2013 NCAA Div. I Men’s Basketball
season. The size and darkness of each nodes represents its relative eigenvector
centrality value. Again, note the clustering of teams into NCAA conferences.

6 Separating the Coach Effect

The model we created in the previous section works well for finding the relative
skills of teams for any given year. However, in order to rank the coaches, it is
necessary to decouple the coach skill from the overall team skill. Let us assume
that the overall team skill is a function of two main factors, coach skill and
player skill. Specifically, if Cs is the coach skill, Ps is the player skill, and Ts is

10
Team #30680 Page 11 of 18

the team skill, we hypothesize that

Ts = Cs · Ps , (4)
as Cs of any particular team could be thought of as a multiplier on the player
skill Ps , which results in team skill Ts .
Although the relationship between these factors may be more complex in
real life, this relationship gives us reasonable results and works well with our
model.

6.1 When is Coach Skill Important?

We will now make a key assumption regarding player skill and coach skill. In
order to separate the effects of these two factors on the overall team skill, we
must define some difference in effect between the two. That is, the player skill
will influence the team skill in some fundamentally different way from the coach
skill.
Think again to a game played between two arbitrary teams A and B. There
are two main cases to be considered:
Case one: Player skills differ significantly: Without loss of generality,
assume that P (A) >> P (B), where P (x) is a function returning the player skills
of any given team x. It is clear that A winning the game is a likely outcome. We
can draw a plot approximating the probability of winning by a certain margin,
which is shown in Figure 3.

Probability

Margin of Win

Figure 3: A has a high chance of winning when its players are more skilled.

Because the player skills are very imbalanced, the coach skill will likely not
change the outcome of the game. Even if B has an excellent coach, the effect of
the coach’s skill will not be enough to make B’s win likely.
Case two: Player skills approximately equal: If the player skills of the
two teams are approximately evenly matched, the coach skill has a much higher
likelihood of impacting the outcome of the game. When the player skills are

11
Team #30680 Page 12 of 18

similar for both teams, the Gaussian curve looks like the one shown in Figure 4.
In this situation, the coach has a much greater influene on the outcome of the
game - crucial calls of time-outs, player substitutions, and strategies can make
or break an otherwise evenly matched game. Therefore, if the coach skills are
unequal, causing the Gaussian curve is shifted even slightly, one team will have
a higher chance of winning (even if the margin of win will likely be small).

Probability

Margin of Win

Figure 4: Neither A nor B are more likely to win when player skills are the
same (if player skill is the only factor considered).

With the assumptions regarding the effect of coach skill given a difference
in player skills, we can say that the effect of a coach can be expressed as:
 
1
(CA − CB ) · (5)
1 + α|PA − PB |
Where CA is the coach skill of team A, CB is the coach skill of team B, PA is
the player skill of team A, PB is the player skill of team B, and α is some scalar
constant. With this expression, the difference in coach skills is diminished if the
difference in player skills is large, and coach effect is fully present when players
have equal skill.

6.2 Margin of Win Probability

Now we wish to use the coach effect expression to create a function giving the
probability that team A will beat team B by a margin of x points. A negative
value of x means that team B beat team A. The probability that A beats B is:
 2
1
− (C · player effect + D · coach effect − margin)
K ·e E (6)
where C, D, E are constant weights, player effect is PA − PB , coach effect is
given by Equation 5, and margin is x.

12
Team #30680 Page 13 of 18

This probability is maximized when

C · player effect + D · coach effect = margin.
This accurately models our situation, as it is more likely that team A wins by
a margin equal to their combined coach and team effects over team B.
Since team skill is comprised of player skill and coach skill, we may calculate
a given team’s player skill using their team skill and coach skill. Thus, the
probability that team A beats team B by margin x can be determined solely
using the coach skills of the respective teams and their eigenvector centrality
measures.

6.3 Optimizing the Probability Function

We want to assign all the coaches various skill levels to maximize the likelihood
that the given historical game data occurred. To do this, we maximize the
probability function described in Equation 6 over all games from historical data
by finding an optimal value for the coach skill vectors CA and CB . Formally,
the probability that the historical data occurred in a given year is

 2
1
− (C · player effect + D · coach effect − margin)
E
Y
K ·e . (7)
all games

After some algebra, we notice that maximizing this value is equalivalent to

minimizing the value of the cost function J, where
X
J(Cs ) = (C · player effect + D · coach effect − margin)2 (8)
all games

Because P (A beats B by x) is a nonlinear function of four variables for each

edge in our network, and because we must iterate over all edges, calculus and
linear algebra techniques are not applicable. We will investigate three techniques
(Genetic Algorithm, Nelder-Mead Search, and Powell Search) to find the global
maximum of our probability function.

6.3.1 Genetic Algorithm

At first, our team set out to implement a Genetic Algorithm to create the coach
skill and player skill vectors that would maximize the probability of the win/loss
margins occurring. We created a program that would initialize 1000 random
coach skill and player skill vectors. The probability function was calculated
for each pair of vectors, and then the steps of the Genetic Algorithm were ran
(carry over the “most fit” solution to the next generation, cross random elements
of the coach skill vectors with each other, and mutate a certain percentage of
the data randomly). However, our genetic algorithm took a very long time to
converge and did not produce the optimal values. Therefore, we decided to forgo
optimization with genetic algorithm methods.

13
Team #30680 Page 14 of 18

6.3.2 Nelder-Mead Method

We wanted to attempt optimization with a technique that would iterate over the
function instead of mutating and crossing over. The Nelder-Mead method starts
with a randomly initialized coach skills vector Cs and uses a simplex to tweak
the values of Cs to improve the value of a function for the next iteration [7].
However, running Nelder-Mead found local extrema which barely increased the
probability of the historical data occurring, so we excluded it from this report.

6.3.3 Powell’s Method

A more efficient method of finding minima is Powell’s Method. This algorithm
works by initalizing a random coach skills vector Cs , and uses bi-directional
search methods along several search vectors to find the optimal coach skills.
A detailed explanation of the mathematical basis for Powell’s method can be
found in Powell’s paper on the algorithm [8]. We found that Powell’s method
was several times faster than the Nelder-Mead Method and produced reasonable
results for the minimization of our probability function. Therefore, our team
decided to use Powell’s method as the main algorithm to determine the coach
skills vector. We implemented this algorithm in Python and ran it across every
edge in our network for each year that we had data. It significantly lowered our
cost function J over several thousand iterations.
Rank 1962 2000 2005
1 John Wooden Lute Olson Jim Boeheim
2 Forrest Twogood John Wooden Roy Williams
3 LaDell Anderson Jerry Dunn Thad Matta

The table above shows the results of running Powell’s method until the
probability function shown in Equation 6 is optimized, for three widely separated
arbitrary years. We have chosen to show the top three coaches per year for the
purposes of conciseness. We will additionally highlight the performance of our
top three three outstanding coaches.
John Wooden - UCLA: John Wooden built one of the ’greatest dynasties
in all of sports at UCLA’, winning 10 NCAA Division I Basketball tournaments
and leading an unmatched streak of seven tournaments in a row from 1967 to
1973 [1]. He won 88 straight games during one stretch
Jim Boeheim - Syracuse: Boeheim has led Syracuse to the NCAA Tour-
nament 28 of the 37 years that he has been coaching the team [3]. He is second
only to Mike Krzyzewsky of Duke in total wins. He consistently performs even
when his players vary - he is the only head coach in NCAA history to lead a
school to four final four appearences in four separate decades.
Roy Williams - North Carolina: Williams is currently the head of the
basketball program at North Carolina where he is sixth all-time in the NCAA
for winning percentage [5]. He performs impressively no matter who his players
are - he is one of two coaches in history to have led two different teams to the
Final Four at least three times each.

14
Team #30680 Page 15 of 18

7 Ranking Coaches
Knowing that we are only concerned with finding the top five coaches per sport,
we decided to only consider the five highest-ranked coaches for each year. To
calculate the overall ranking of a coach over all possible years, we considered
the number of years coached and the frequency which the coach appeared in the
yearly top five list. That is:
Na
Cv = (9)
Nc
Where Cv is the overall value assigned to a certain coach, Na is the number
of times a coach appears in yearly top five coach lists, and Nc is the number of
years that the coach has been active. This method of measuring overall coach
skill is especially strong because we can account for instances where coaches
change teams.

7.1 Top Coaches of the Last 100 Years

After optimizing the coach skill vectors for each year, taking the top five, and
ranking the coaches based on the number of times they appeared in the top five
list, we arrived at the following table. This is our definitive ranking of the top
five coaches for the last 100 years, and their associated career-history ranking:

Rank Mens Basketball Mens Football Mens Baseball

1 John Wooden - 0.28 Glenn Scobey Warner - 0.24 Mark Marquess - 0.27
2 Lute Olson - 0.26 Bobby Bowden - 0.23 Augie Garrido - 0.24
3 Jim Boeheim - 0.24 Jim Grobe - 0.18 Tom Chandler - 0.22
4 Gregg Marshall - .23 Bob Stoops - 0.17 Richard Jones - 0.19
5 Jamie Dixon - .21 Bill Peterson - 0.16 Bill Walkenbach - 0.16

8 Testing our Model

8.1 Sensitivity Analysis
A requirement of any good model is that it must be tolerant to a small amount
of error in its inputs. In our model, possible sources of error could include im-
properly recorded game results, incorrect final scores, or entirely missing games.
These sources of error could cause a badly written algorithm to return incorrect
results. To test the sensitivity of our model to these sources of error, we decided
to create intentional small sources of error in the data and compare the results
to the original, unmodified results.
The first intentional source of error that we incorporated into our model
was the deletion of a game, specifically a regular-season win for Alabama (the
team with the top-ranked coach in 1975) over Providence with a score of 67 to
60. We expected that the skill value of the coach of the Alabama team would

15
Team #30680 Page 16 of 18

decrease slightly with this modification. When we ran and analyzed the results,
we found that the coach skill value did in fact decrease by approximately 1%, as
we expected. However, the Alabama coach maintained his ranking of top coach
for the season.
The second change that we incorporated was to switch the results of the
same game (Alabama 67, Providence 60) to a win for Providence (Providence
67, Alabama 60). We expect this will have a greater negative influence on
the skill value of the Alabama coach, and when we ran the analysis we found
that, indeed, the Alabama coach skill value decreased by approximately 4%.
Although a relatively minor difference, the second-ranked coach originally had
a skill value only very slightly behind the Alabama coach, and the 4% loss in
fact placed the second-ranked coach in the first ranking position.
From this analysis we can see that our model follows our predictions accu-
rately, and that removing factors that add positively to the skill ranking of a
coach is detrimental to their skill value. Although the changes we made were
minor, there is often a lot of competition for the first-place ranking of a coach,
and due to the limited number of games played per season, a change to this
data can have an influence on the final ranking. Although these results indicate
that error in our data can effect the final ranking, the analysis also shows that
our model responds predictably to a variation in the input.

8.2 Strengths
The main strength of our approach is that it is able to separate coach proficiency
from team proficiency by calculating probabilities that the historical game data
occur given coach skills. This allows us to more accurately gauge the skills
of a coach without factoring in the skills of his/her players. Furthermore, our
approach is flexible as many relationships can be modified. For example, if a
study shows that there is a better function to describe the relationship between
coach skill, player skill, and team skill, it can easily be used in our model. Our
model is also able to compare the relative effectiveness of coaches from all time
periods, as long as the average margin of victory is similar across time periods.

8.3 Weaknesses
The main weakness of our model pertains to computational efficiency, as our
computers could not always adequately calculate all necessary values in our
model. For example, the computer could not find the eigenvector centrality
values on a small percentage of the graphs, as the Von Mises iteration failed
to converge. Furthermore, sometimes Powell’s method of minimizing our cost
function yielded high costs relative to other years because the initialized array
of coach skills was close to a local minima. This could be solved by running
Powell’s method from several randomly initialized coach skills arrays, but this
increases computational time. In fact, the overall results of our model could
likely be improved significantly given more time to run the iterative optimization

16
Team #30680 Page 17 of 18

algorithms with a higher accuracy, resulting in better approximations for the

ideal matrix of coach skills.

9 Conclusions
In this report, we have analyzed nearly 300 years of data in order to determine
the most accurate and unbiased ranking of all-time best college sports coaches.
By constructing comprehensive networks with with edges representing each and
every game played in the last century of the college sports that we analyzed,
we were able to create a comprehensive metric of team skill using the concept
of eigenvector centrality. By considering win/loss margins, we were able to
identify patterns that enabled us to separate the team skill measure into its
two components - player skill and coach skill. We then created a probability
function based on player skill and coach skill to determine the likelihood of an
edge in our network occurring. By multiplying this probability across all edges,
we were able to determine the probability of the entire graph occurring given
team skill and coach skill vectors. Using an iterative, multivariable, machine
learning algorithm, we maximized this probability function for coach skill for
each season and for each sport. Using data that mapped the name of a coach
to his/her team for each season, we were able to combine the results of each
individual season and analyze the skill of each individual coach over their entire
coaching history. From this data, we selected the top 5 coaches from every sport
to feature as our all-time best coaches of the century.

17
Team #30680 Page 18 of 18

References
[1] John Wooden. Retrieved from: http://msn.foxsports.com/
collegebasketball/story/John-Wooden-dies-UCLA-coach-99-060410,
2010.
[2] ShrpSports. Retrieved from: http://www.shrpsports.com/, 2011.

[3] Jim Boeheim. Retrieved from: http://cuse.com/coaches.aspx?rc=405&

path=mbasket, 2012.
[4] Sports-Reference. Retrieved from: http://www.sports-reference.com/,
2013.
[5] Roy Williams. Retrieved from: http://www.goheels.com/ViewArticle.
dbml?ATCLID=205497516, 2014.
[6] Stephen P. Borgatti. Centrality and network flow. Social Networks, 27(1):55–
71, January 2005.
[7] JA Nelder and R Mead. A simplex method for function minimization. The
Computer Journal, 1965.
[8] MJD Powell. An efficient method for finding the minimum of a function
of several variables without calculating derivatives. The Computer Journal,
1964.
[9] Leo Spizzirri. Justification and Application of Eigenvector Centrality. 2011.

18