Вы находитесь на странице: 1из 4

Improving NCAA Football Bowl Game Prediction

Using a Neural Network


James Kardatzke Christopher Kardatzke
River Valley High School River Valley High School
Email: james.kardatzke@rvschools.org Email: christopher.kardat@rvschools.org

II. DATASET
AbstractIn this report, we attempt to predict the results of Our training dataset contained the results of 5 seasons (from
FBS college football bowl matches using machine learning
algorithms. We created a dataset of historical bowl games that 2011-2015) of bowl matches, and our testing dataset included
includes a variety of metrics for each competing team. We used the results of 1 season (2016) of bowl matches. From 2011 to
this data to create three different models: one which picks the 2014, there were 35 bowl matches per season, while 2015 and
team with the higher Colley Rating to win each game, one using 2016 both had 41 bowl matches. This allowed us to have 181
linear regression, and another using neural networking. We games in our training set and 41 games in our testing set. We
compared the performance of these models predictions with
used a dataset of all 4746 regular season games from
those of sports experts. Each of these models uses binary
classification: predicting whether each match will be a win or a 2011-2016 to compute the week-by-week Colley Rankings of
loss for the home team. After training, the neural network was each team. For each bowl game, our dataset included these
the model which gave the lowest error rate (.243) on our test set. Colley Rankings (one per team per week of the regular
Our error analysis focused on improving hyperparameters and season), boolean values indicating whether each team was a
the parameters included in the dataset. member of a Power Five conference, AP rankings at the time
I. INTRODUCTION of the game, and the winner. We obtained the weekly game
data from a cleaned ESPN dataset found on GitHub [4]. We
At the end of every season of NCAA college football, teams obtained the historical AP ranking data from the Sports
are selected to compete in around forty postseason matches. Reference website [5]. We preprocessed and compiled this
These matches attract hundreds of millions of viewers, and data with python.
being able to predict their outcomes has importance for
journalism, coaching, and gambling. Recently, several III. M ETHODOLOGY
organizations such as FiveThirtyEight and ESPN have A. Initial Challenges
attempted to predicted these matches using analytical
approaches. In our project, we will discuss prior works before We ran into several challenges over the course of this project.
analyzing feature selection, discussing performance of various The first was difficulty obtaining game data. Although countless
models, and analyzing our results. websites list historical college football scores, very few had them in
formats which were conducive for use in our project. Ultimately, we
A. Related Literature were able to find a data source with data from every game since 2011.
Although mainstream sports prediction has traditionally While this source contained many of the features we were looking for,
been dominated by subjective analysis, in the past two such as final scores and game location, it had the drawback of
decades there has been a surge in quantitative analytics. For containing only six years of data. A second challenge which we
the past several years, Nate Silvers FiveThirtyEight website encountered was variations on team names between our data sources.
has published win probabilities for upcoming matches in a For example, the game data listed the University of South Florida as
wide array of sports, including college football. He uses an S. Florida, while the AP data listed them as USF. This led to some
Elo model, which computes win probabilities primarily based confusing results at first, and we had to go back and manually alter the
on teams relative strengths [1]. While this model did give us datasets to match. Another issue we had was that, unlike many other
some insight into the importance of a team rating system, we sports leagues, FBS football is not a closed-system. This means that
found that it did not work well when applied to college there are some matches between FBS and non-FBS teams. In order to
football. This is likely due to the large number of teams (over account for this, we generalized all of the non-FBS teams, and
129 in 2017), and the small number of games. We had greater classified them together for the purposes of Colley Rankings.
success using a ranking system devised by Colley et al [2]. Although this doesnt completely account for potential differences in
The Colleys Bias Free College Football Ranking Method is a quality between non-FBS teams, we assumed that it would not impact
very effective tool for computing the relatives strengths of our model too much, given that they are typically of a significantly
teams with no history of competition. Finally, we were lower quality than teams in the FBS, rarely play more than one FBS
influenced by a 2015 Stanford research paper by Chesire et al. opponent a year, and do not play in bowl games.
which used logistic regression, a Naive Bayes model, and B. Feature Selection
Support Vector Machines to attempt to predict College
College football is unique from many other sports in that
Football Bowl matches [3].
teams are allowed to arrange their own non-conference
matchups. Different college football programs often take very
Improving NCAA Football Bowl Game Prediction Using a Neural Network Page 2

different approaches to regular season match scheduling, with with over-fitting, with a training error of .01 and a test
some setting up much easier matchups than others. Due to this, error of .35. After graphing the models performance on
we didnt think it would be wise to include features which different numbers of iterations of gradient descent, as
were heavily influenced by strength of schedule, such as shown on Figure 1, we determined that 14,500 iterations
Offensive Yards per Game. We determined that features such was optimal for avoiding both over- and under-fitting.
as whether a team is home or away, Colley rankings (which
give a teams relative strength adjusted for strength of
schedule), AP rankings (which are the results of a subjective
survey of top sports journalists), and whether or not a team is a
member of a Power Five conference could provide insight.
Initially, we only used a teams Colley Ranking at the time of
the bowl game. However, we discovered that including
features for teams rankings after each week of the season
allowed the model to glean more information about a teams
season and recent performance, and provide a more accurate
prediction.

C. Models
We tried several different methods to address our
classification problem: Figure 1. Neural Networks performance with varying iterations
1) Colley Model: For a baseline for comparison, we used a
model which predicted that the team with the higher Colley
Ranking at the time of each match would win. This model had Another hyperparameter which we wanted to optimize was the
an error of .487 on the training set. This shows that solely number of nodes in our hidden layer. We tested the models
comparing two teams Colley Rankings is very ineffective, performance with between two and seven nodes and, as shown
likely due to the fact that teams are generally matched up against in Table 2, determined that three nodes was optimal.
teams with very similar rankings.
2) Expert Predictions: Another, more realistic,
Number of
comparison method which we used was a compilation of Nodes 2 3 4 5 6 7
expert predictions. We looked at bowl game predictions Test Set Error 0.269 0.243 0.293 0.319 0.293 0.293
from a number of sports websites and blogs [6, 7, 8], and
average the accuracy of these predictions. See Table 1 for Table 2. Neural Networks performance with varying nodes
more information on this model. This model had a
training set error of .382. Using these hyperparameters, the model had a training error
of .243. Figure 2 shows the confusion matrix for this model.
Bleacher USA
Expert ESPN Average
Report Today
Test Set Error 0.341 0.439 0.366 0.382 Home Win Away Win

Table 1. Expert predictions accuracy Predicted Home Win 0.341 0.122

Predicted Away Win 0.146 0.390


3) Logistic Regression: We used logistic regression to
compare the accuracy of our neural network with another Figure 2. Neural Networks confusion matrix
machine learning model. Although bowl games are at at
neutral site, a coin toss decides which team is designated IV. RESULTS
the home team. We trained our model with binary
We tested 4 different models. The test set error rates for each of
classification to predict whether or not the home team
these models can be found in Table 3.
would win. We used a sigmoid activation function. After
this model was implemented, and its hyperparameters Logistic Neural
Model Naive Experts
were tuned, the test set had a .317 error rate and a .074 Regression Network
standard error. The model regressed to the coefficients Test Set 0.487 0.382 0.317 0.243
Error
shown in Appendix 1. These coefficients show that the
most significant parameters were each of the teams AP Table 3. Performance of each of the four models
ranking. Interestingly, these coefficients also seem to
As shown, the neural network performed best by a significant
indicate that being a member of a Power Five conference
margin. the expert predictions had an error rate more than 55%
is correlated with a higher chance of winning a bowl that of the neural network. The neural network also outperformed
game. the logistic regression model, showing that
2 there
2 may be some
insight to be gained from looking at nonlinear relationships
4) Neural Network: Lastly, we implemented a neural between our variables (such as whether a teams Colley Rankings
network with one hidden layer. Initially, we problems have increased or decreased in the last few weeks of the season).
Improving NCAA Football Bowl Game Prediction Using a Neural Network Page 3

V. CONCLUSION ACKNOWLEDGMENT
We would like to thank Professor Laura Albert for her
Our project demonstrates that it is possible to train a neural
network to predict college football bowl matches to a guidance and insight on this problem. We would also like to
significantly higher degree of accuracy than experts subjective thank Ben Ulmer, Matthew Fernandez, and Matt Mills
analysis. Using only a handful of parameters, our model was REFERENCES
able to predict these game with over 75% accuracy. One of the [1] (2014) Introducing NFL Elo Ratings [Online] Available: https://
biggest areas of improvement for our neural network is the fivethirtyeight.com/features/introducing-nfl-elo-ratings/
[2] W. N. Colley, Colley's Bias Free College Football Ranking Method: The
quantity of data. The data source which we used only had six Colley Matrix Explained, 2002.
seasons worth of data, which meant that we had less than 200 [3] E. Chesire, A. Childs, T. Leung, College Football Game Predictor,
cases in our training set. By making a new model with a larger 2015.
[4] NCAA Game Data. Scraped from ESPN data. [RDS files]. Available:
dataset, we could likely improve performance. Another issue https://github.com/mattmills49/CFB_Analysis/tree/master/pbp_cleaning
with our model is that it doesnt evaluate potential changes [5] AP Poll Data. Sports Reference. [CSV file]. Available: https://
www.sports-reference.com/cfb/years/2016-polls.html
between the final week of the regular season and bowl games [6] (2016) "Predictions for Every 2016-17 College Football Bowl
(For example, in 2012 Wisconsin Badgers coach Bret Bielema Game" [Online]. Available: http://bleacherreport.com/articles/2678675-
predictions-for-every-2016-17-college-football-bowl-game
left the team during this time-span). Although radical changes [7] (2016) "41 College Football Games, 42 Predictions" [Online]. Available:
are rare, they could be accounted for by including parameters http://www.espn.com/college-football/story/_/id/18202412/41-games-42-
prediction
which analyze news sentiment during this time period. These [8] (2016) "USA Today Sports College Football Staff Picks for Every Bowl
improvements deserve further investigation, and could help the Game" [Online]. Available: https://www.usatoday.com/story/sports/ncaaf/
2016/12/15/staff-picks-college-football-bowl-games-fiesta-bowl-orange-
neural network become an even more accurate model for bowl/95459490/
predicting college football bowl games.
Appendix

Away Home
Away Home Away Home
Parameter Intercept Power Power
AP AP Colley Colley
Five Five

Coefficient 0.025 -0.899 1.41 -0.631 0.139 -0.003 0.107

Appendix 1. Logistic Regression Coefficients

Вам также может понравиться