Mathematics Invades Competitive Sports
Kenneth Massey
Virginia Tech

Goals
Objectively measure the performance of a team
       relative to the schedule faced
Correct for disparate schedules (esp. college sports)
Predictive vs. Retrodictive
       Wins, scores, date, stats, homefield, preseason, other ?
Seed playoffs

Slide 3

Challenges
There is no property of transitivity!
Disparate schedules
   “Mt. Union Syndrome”
     Separate Divisions
Strength vs. Performance vs. Results
     Environment
            (venue, homefield, weather, day/night, crowd)
     Teams don’t always play at full potential
            (injury, unfavorable matchups, intangible, psychological)
     The score isn’t always a good indicator
            (coaching philosophy, chaos “bounce of ball”)
Lack of data
   Connectedness
Undefeated / winless teams

Slide 5

Slide 6

Slide 7

Slide 8

Types of Ratings
Standings / WL% / Points
Polls  Tabulated votes, subjective, time sensitive,
                  no corrections, incomplete analysis
Formula (RPI)
Update (Elo chess)
Least Squares
MLE
Matrix (Markov)
Other (Neural nets)

Bowl Championship Series (BCS)
Polls
Computers
Schedule
Losses
Quality Wins

Schedule Ratings
Average rating of opponents (corrected for homefield)
A good team prefers a less distributed schedule;
    a bad team prefers a more distributed schedule.
    For example (Florida, Vanderbilt) vs. (Alabama, Arkansas)

BCS Computers
Anderson / Hester  formula
Billingsley            update
Colley                  matrix
Massey              MLE (Gaussian)
Matthews             matrix
Rothman              MLE (logistic)
Sagarin                MLE (logistic)
Wolfe / Baker least squares

Least Squares Model

Least Squares Model

Rank Deficiency

LS Example

LS Example

LS Example

More LS Model Features
Preseason ratings
Weighting
Choice of GOF

LS Notes

Maximum Likelihood Estimator (MLE) Method
Optimization Problem
    Choose ratings to maximize the probability of
       reproducing the observed results
Game Outcome Function
    Measures the result of a particular game
Game Likelihood Function
    The probability of a given result given a set of ratings

Game Outcome Function
Win Indicator Function
Score Ratio
Rothman
Sagarin ?
Massey

GOF Values

GOF Bias

Proof

Game Likelihood Functions
Gaussian
Logistic
 Equivalent to
Arctan

Slide 27

Slide 28

MLE Function

MLE Optimization (Logistic Model)

MLE Issues
Non-uniqueness
Ideally, wins help and losses hurt
Undefeated / Winless Teams
Update ratings based on linearization
    (time dependent, n large)

Ratings on the Web
Massey Ratings
    http://www.masseyratings.com
College Football Rankings http://www.cae.wisc.edu/~dwilson/rsfc/rate/index.html
Bowl Championship Series
    http://www.collegebcs.com