Google Cloud and NCAA ML Competition 2019-

Women’s
Zijing Wang, Xuanyu Liang, Kevin Xue, Xiao Wang, Jiacheng
Shi, Minsheng Liu, and Junxian Tan.
March 25, 2019

1 Introduction

NCAA Division 1 Women Tournament is a basketball tournament that

is played between 64 collegiate basketball games. The tournament
is usually played in March of each season and the tournament has
been popular across the country. Teams that enter the tournament are
selected by their overall performance and ranking in different con-
ference games. The first round of the tournament will commence in
the late March and the championship game will usually take place in
the early April. In the project, we are given data of both regular sea-
sons result and tournament results from the last few years and asked
to predict the result of the 2019 NCAA Division 1 Women Basketball
Tournament.
For this project, we will first explore the data and discover some
potential hypothesis, which we will use statistical methods for valida-
tion. After the conclusion of the hypothesis is drawn, we will incorpo-
rate them in predicting the 2019 tournament results. In the project, we
will first analyze if there is a difference between playing at home and
away; or equivalently, testing if there exist home field advantages. For
this hypothesis, a numerical summary of the data will be used and a
histogram will be drawn before we conduct a hypothesis test. Then we
will analyze the difference between a team’s performance in regional
conferences and the tournament by using graphical and testing meth-
ods. Furthermore, we will use different statistical methods, graphical
methods, and testings to see if there exist a difference in a team per-
formance due to regions, break-time between consecutive games, and
competitors. After the result from these hypotheses is drawn, we will
incorporate the results to predict tournament outcomes.

2 Background

The following sections will provide information about the background

of the study and will include additional research for this project.
 google cloud and ncaa ml competition 2019-women’s 2

NCAA Division 1 Women Basketball Tournament

NCAA Division 1 Women Tournament is a basketball tournament that
usually takes place in March of each season. The Women’s NCAA
Division 1 has gained people’s interest and attention over time; it is
considered one of the most influential basketball games in the country.
NCAA Division 1 Women Tournament is played between 64 collegiate
basketball teams. Due to a large number of teams participating, the
tournament will be in the format of a single elimination tournament.
In Women’s Basketball: Road to the Championship, NCAA has explained
their selection process: 32 teams (Automatic Qualifiers) will receive
automatic qualification determined by their conference tournament
results and the rest 32 teams (At-Large) will be selected by the NCAA
selection committee with some criterions, including their conference
records, overall records, strength of the conference, overall record and
so on. In the webpage Selection Criteria, NCAA has also mentioned
how ”seeding” works. The NCAA selection committee will create
a seed list of 64 teams to offer audience information about teams’
competitive ranking on a national scale. The tournament is usually
played in March of each season and the tournament has been popular
across the country. The first round of the competition will commence
in the late March and the championship game will usually take place
in the early April.

Game Result Prediction

Researchers have been using past games’ statistics to analyze the cor-
relation between some statistics and game output. Multiple studies
have shown that the seedings provided by the NCAA are strong in-
dicators of tournament championship since the seeds are assigned
based on a team’s performance in regional conferences. In A Simple
and Flexible Rating Method for Predicting Success in the NCAA Basketball
Tournament: Updated Results from 2007, West points out that there may
be bias in the seeds assignment as selection committee may have sub-
jective opinions and hence it is beneficial to use statistical methods to
develop rating methods to conduct game prediction (2). In the article,
West has presented some potential rating methods for predicting, in-
cluding Rating Percentage Index, a rating that compares team based
on its winning percentages and its opponents’ and Jeff Sagarin’s Com-
puter Ratings (3). Furthermore, West suggests that an ordinal linear
regressing modeling and expectation (OLRE) could be an efficient
method since it is simple for calculation and flexible in terms of dating
usage. In Heterogeneous Skewness in Binary Choice Models: Predicting
Outcomes in the Men’s NCAA Basketball Tournament, Caudill and God-
win have derived a method that is using the skewed logit model that
 google cloud and ncaa ml competition 2019-women’s 3

allows heterogeneity in the skewness parameter and find their model

have better fitting when predicting tournament outcomes.

Advanced Statistics in Basketball

In this project, we will use some basketball-specific ratings to analyze
the performance of individual teams.

Possessions = 0.96 × [FGA + Turnovers + (0.475 × FTA)] − Off Rebounds

Opponent’s Score
Defensive Rating = 100 × ( )
Possessions
Score
Offensive Rating = 100 × ( )
Possessions
Net Rating = 100 × (Offensive Rating − Defensive Rating)
FTA(Free Throws Attempted)
Free Throw Rate =
FGA(Field Goal Attempted)

3 Data

The data of this study is composed of the games boxscores from 1998
to 2018, including boxscores results from both regular seasons and
tournaments. Detailed box scores are provided for the tournaments
between 2008 and 2018 and regular seasons between 1998 and 2018.
There are 10 files in the data folder that include information about the
season, location, TeamID, seeds, and specific team box scores. Table 1
lists all fields and their descriptions.

4 Analysis

Does there exist home-field advantage?

Multiple studies and researches have shown that there are home-
advantages in competitive sports.In this section, we are going to exam-
ine whether there are home advantage in NCAA Women Basketball
games over the last decade including both regular seasons and tour-
neys.
In figure 1, we observe that it is more likely to win at home than
away. In figure 2, this is also true for tourney matches. In fact, we see
the difference is even bigger. Since the performance of WNCAA teams
varies within the home and away games, home advantage by teams is
hereby investigated by analyzing home-win rate and away-win rate.
Examining the difference between home and away games is more
optimal than just analyzing home-win rate solely because we could
have more intuition on home field advantage.
 google cloud and ncaa ml competition 2019-women’s 4

Field Description
TeamID Identification number (4-digit) of each NCAA Women team
TeamName Name of the team
Season Year of which the tournament is played
DayZero The date corresponding to daynum=0 of the season
Region in W, X, Y, Z; identifier of the region
Seed The team’s seed in one of the corresponding region
DayNum The number tells what day the games took place with respect to DayZero
WScore Points Scored by the winning team
LScore Points Scored by the losing team
NumOT The number of overtime in the game
WLoc The location of the winning team (H: home, A: away N:neutral court)
WFGM Winning Team’s Field Goals Made
WFGA Winning Team’s Field Goals Attempted
WFGM3 Winning Team’s Three Pointers Made
WFGA3 Winning Team’s Three Pointers Attempted
WFTM Winning Team’s Free Throws Made
WFTA Winning Team’s Free Throws Attempted
WOR Winning Team’s Offensice Rebounds
WDR Winning Team’s Defensive Rebounds
WAst Winning Team’s Assists
WTO Winning Team’s Turnovers Commiteed
WStl Winning Team’s Steals
WBlk Winning Team’s Blocks
WPF Winning Team’s Personal Fouls Committed
Table 1: All fields and their descriptions.
Figure 1: Histogram of number of games
win at home or away during regular
seasons
 google cloud and ncaa ml competition 2019-women’s 5

Figure 2: Histogram of number of

games win at home or away during
tourney

In order to ensure the accuracy of the experiment, we eliminate the

influence of competition in a neutral court. Then, we will cover basic
results regarding confidence intervals under binomial distribution.

Hypothesis 4.1. Let p be the home-win rate. H0 : p = 0.5; There does

not exist home-field advantage. H1 : p > 0.5 There exist home-field
advantage, that is the home team have higher chance of wining.

From Table 1 and Table 2, we observe that the home-winning per-

centages are higher than 0.50 in both regular seasons and tourneys in
2010-2018 seasons, indicating that home teams have higher chance of
winning the game than away teams.
Table 2 records the home/away performance discrepancy over the
last decade, done at p = 0.5.

Table 2: Results of the binomial test

Season Away Home Lower Estimate Upper p-value
done for home/away performance
2010 1770 2912 0.608 0.622 0.636 4.851E-63 discrepancy in the regular seasons over
the last decade.
2011 1748 2836 0.604 0.619 0.633 1.535E-58
2012 1755 2738 0.595 0.609 0.624 5.622E-49
2013 1814 2743 0.588 0.602 0.616 2.732E-43
2014 1789 2787 0.595 0.609 0.623 1.520E-49
2015 1844 2981 0.604 0.618 0.632 1.147E-60
2016 1858 2855 0.592 0.606 0.620 4.849E-48
2017 1864 2984 0.602 0.616 0.629 1.258E-58
2018 1878 2835 0.587 0.602 0.616 2.257E-44
Table 3 gives the result for tourney.
 google cloud and ncaa ml competition 2019-women’s 6

Table 3: Results of the binomial test

Season Away Home Lower Estimate Upper p-value
done for home/away performance
2010 2 21 0.720 0.913 0.989 6.604E-05 discrepancy in the tourneys over the last
decade.
2011 6 19 0.549 0.760 0.906 1.463E-02
2012 1 5 0.359 0.833 0.996 2.188E-01
2013 5 22 0.619 0.815 0.937 1.514E-03
2014 2 14 0.617 0.875 0.984 4.181E-03
2016 6 27 0.645 0.818 0.930 3.241E-04
2017 4 30 0.725 0.882 0.967 6.165E-06
2018 4 29 0.718 0.879 0.966 1.093E-05

In Table 1 and Table 2, the p-value are all less than 0.5. With sig-
nificance level α=0.05, we reject the null hypothesis and accept the
alternative. By conducting a 95% confidence interval, we are 95%
confident that the true home-win rate p is higher than 0.50. Hence,
from the one-sided hypothesis test, we conclude that there exist home
advantage over the last decade during both regular seasons and tour-
neys. The results confirmed our assumption, which is competitive
sports have home-advantage. Since the home team have psychological
advantages and more familiar with the environment at home, they can
pay more attention on games rather than struggling on their travels.

Are Tourney Matches More Intense?

Our next question is whether matches in the tourney are more intense
than those in the regular reason. Since teams who made into the tour-
ney will be stronger than their not so fortunate peers, it is reasonable
to expect that matches between those stronger teams are more com-
petitive. To measure competitiveness, we look at the score difference
between the two teams in a match: the closer the difference, the more
competitive the match is.
If neither team in a match made into the tourney, both are not as
strong and hence their score difference could still be small. Hence,
for the regular season we only look at matches where one of team
made into the tourney. Table 4 shows statistics of score differences in
matches from 2010 to 2018.

Table 4: Mean, variance, and skewness

Stage Mean Variance Skewness
of score differences in matches from
Regular 17.237 166.534 1.290 2010 to 2018.

Tourney 17.139 175.403 1.517

As stated above, we believe that matches in the tourney should be

more competitive. Formally, our null hypothesis is

Hypothesis 4.2. The mean of score differences in the selected matches

 google cloud and ncaa ml competition 2019-women’s 7

from the regular season follow the same distribution as that from the
tourney.

Before the test, we use a bar plot to visually compare the mean of
score differences, where our data are divided by year, shown in figure
3. Surprisingly, we do not see much discrepancy between the mean
from the regular season and that from the tourney. In year 2010 to
2013 as well as 2016, the mean from the regular season is higher than
that from the tourney, but the situation is reversed in other years.

20 Figure 3: Discrepancy of the mean

difference between the regular season
and the tourney, separated by year.

15
Mean Diff

Stage
10 Regular
Tourney

2010 2011 2012 2013 2014 2015 2016 2017 2018

Season

In any case, there is difference between the regular season and the
tourney. We perform the chi-squared test to see if such difference is
significant. The test is done separately for each year. Table 5 shows the
test results.

Table 5: Results of the chi-squared tests.

Season Regular Tourney χ2 p-value
2010 16.886 16.460 0.005 0.941
2011 17.321 15.524 0.098 0.754
2012 17.217 15.794 0.061 0.804
2013 17.623 16.429 0.042 0.838
2014 17.149 17.508 0.004 0.951
2015 16.823 15.159 0.087 0.769
2016 16.966 18.524 0.068 0.794
2017 17.220 19.238 0.112 0.738
2018 17.931 19.619 0.076 0.783

As we can see, for all years p > 0.05, meaning that we cannot reject
hypothesis 4.2, and that we are 95% confident that the mean score
 google cloud and ncaa ml competition 2019-women’s 8

difference between regular season and the tourney is the same. As a

conclusion and an answer to our section’s question, the competitive-
ness in regular seasons and tourneys is, contrary to our intuition, not
significantly different.

Does a strong team perform better when facing a strong competitor?

In this section, we want to see if there is a difference between match
of two powerful teams and match of not too powerful teams. We used
the total score of two teams to compare. We will see from the next
section that there is a relationship between offensive rating and wins.
Offensive rating is a direct consequence of how many points scored.
We define strong teams as the teams who entered the tourney that
year. Weak teams as the teams that did not enter the tourney that year.
In Figure 4,5 and 6, we can see that in 2016, 2017, and 2018, the
matches between strong teams will have a higher score total than the
matches between weak teams. The next thing we want to do is to see
whether there is a difference between strong teams and weak teams.

Figure 4: Histogram: mean of total

scores in games between two strong
teams and two weak teams in 2016

We draw the distribution of strong vs strong teams and weak vs

weak teams. Figure 7 is strong vs strong, and Figure 8 is weak vs
weak. From these two diagrams, we observe that these two distribu-
tions might be normal. So we performed kurtosis and skewness test to
test for normality and the results are displayed in Table 6. The result
of adjusted kurtosis test (normal will have a result of 0 instead of 3)
showed that we are unsure about the distribution of the second one
being normal. So we used the K-S test to check since K-S test is a more
google cloud and ncaa ml competition 2019-women’s 9

Figure 5: Histogram: mean of total

scores in games between two strong
teams and two weak teams in 2017

Figure 6: Histogram: mean of total

scores in games between two strong
teams and two weak teams in 2018
google cloud and ncaa ml competition 2019-women’s 10

Figure 7: Histogram: distribution of

points total per game between two
strong teams

Figure 8: Histogram: distribution of

points total per game between two weak
teams
 google cloud and ncaa ml competition 2019-women’s 11

powerful test to test for normality in general. We set the null hypothe-
sis of K-S test to be the distribution is normal. The test result gives us a
p-value less than 0.05, so we rejected the null hypothesis. The second
distribution is not normal.

Table 6: Results of the Skewness test and

Skewness Kurtosis
Kurtosis test
Strong vs Strong 0.381942 0.111631
Weak vs Weak 0.396316 0.480225

Table 7: Results of Kolmogorov–

K-S Statistics p-value
Smirnov test
Strong vs Strong 0.042898 0.094649
Weak vs Weak 0.035909 3.98534e-12

Table 8: Results of Welch-t

var(strong vs strong) 0.045319 test(bootstrap)
var(weak vs weak) 0.037764
t-test stat 944.7691
p-val 0.0

Next, we tried to perform Bootstrap algorithm to get a normal

distribution by using the central limit theorem. We calculated the
mean of each bootstrapping result and get 1000 mean from 1000 times
of bootstrapping in order to get a normal distribution. We first want
to use two sample t test, but realized that the variance is different, so
we used Welch t test instead. We get a p-value close to 0, so we being
able to reject the null hypothesis. Therefore to conclude that the mean
of the total score between strong teams are different than the mean of
the total score between weak teams. This might because when a team
is facing a strong team, it is more likely to encourage them to perform
well because they want to prove themselves.

Offensiveness and Winning

We want to investigate the relationship between offensiveness and
the wins. First, we introduce a common statistic in basketball games,
”offensive rating”.

Possessions = 0.96 × [FGA + Turnovers + (0.475 × FTA)] − Off Rebounds

Score
Offensive Rating = 100 × ( )
Possessions
In general, offensive rating is used to define how ”offensive” a team
is. Since some teams tend to focus on defense, and others focus on
offense, we want to know what competitive teams tend to be.
 google cloud and ncaa ml competition 2019-women’s 12

In figure 9, x-axis is each team’s offensive rating and y-axis is its

corresponding wins in a specific year. Although there are plots in
several years, the fitted line and the trend is very similar so we will
take an example of 2013.
In the linear regression, we obtained th following equation: wins =
0.66*Offrtg-45.70 with R-adj = 0.735. This shows a moderately strong
relationship between offensive rating and wins.
In figure 9, the residual plot shows residuals are randomly scat-
tered, no pattern is observed, so the linearity is valid. The regression
line indicates that on average, every increase in offensive rating is as-
sociated with an increase of 0.66 wins. The more offensive the team is,
the more the team wins
offensiveness and the wins. First, we introduce a common statistic
in basketball game, ”offensive rating”. Offensive Rating = 100 x (Score
/ Possessions) Possessions = 0.96 x (FGA + Turnovers + (0.475 x FTA) -
Off rebounds In general, offensive rating is used to define how ”offen-
sive” a team is. Since some teams tend to focus on defense, and others
focus on offense, we want to know what competitive teams tend to be.
In figure 9, x-axis is each team’s offensive rating and y-axis is its
corresponding wins in a specific year. Although there are plots in
several years, the fitted line and the trend is very similar so we will
take an example of 2013.
In the linear regression, we obtained the following equation: wins =
0.66*Offrtg-45.70 with R-adj = 0.735. This shows a moderately strong
relationship between offensive rating and wins.
In figure 9, the residual plot shows residuals are randomly scat-
tered, no pattern is observed, so the linearity is valid. The regression
line indicates that on average, every increase in offensive rating is as-
sociated with an increase of 0.66 wins. The more offensive the team is,
the more the team wins

What affects the overall chance of winning?

In the last section of analysis, we want to investigate what types of
data have a great impact on the chance of winning. The types of data
we are looking into include: offensive rating(efficiency), defensive
rating(efficiency), net rating, and free throw rate.
The first two diagrams are offensive rating and defensive rating. We
can see that the winning team will have a better offensive rating since
the winning team and all team have the same distribution, but the
curve shifts to the right as a whole for winning teams. On the other
head, defensive rating might not have an affect on the outcome of the
game. The distribution and curve on defensive rating are identical.
The net rating is calculated by offensive rating minus defensive rating.
 google cloud and ncaa ml competition 2019-women’s 13

2010 2011 2012

40 40 40

30 30 30
Wins

Wins

Wins
20 20 20

10 10 10

0 0 0
60 80 100 120 60 80 100 120 60 80 100 120
Rating Rating Rating
2013 2014 2015
40 40 40

30 30 30
Wins

Wins

Wins
20 20 20

10 10 10

0 0 0
60 80 100 120 60 80 100 120 60 80 100 120
Rating Rating Rating
2016 2017 2018
40 40 40

30 30 30
Wins

Wins

Wins

20 20 20

10 10 10

0 0 0
60 80 100 120 60 80 100 120 60 80 100 120
Rating Rating Rating

Figure 9: Regression of offensive ratings

and winning counts.
google cloud and ncaa ml competition 2019-women’s 14

Figure 10: Histogram: Comparison

between winning team’s offensive rating
and the losing’s

Figure 11: Histogram: Comparison

between winning team’s defensive
rating and the losing’s
google cloud and ncaa ml competition 2019-women’s 15

Figure 12: Histogram: Comparison

between winning team’s net ratings and
the losing’s

Figure 13: Histogram: Comparison

between winning team’s FTR and the
losing’s
 google cloud and ncaa ml competition 2019-women’s 16

It is interesting that the net rating for all teams do not have a normal
distribution. This might because it is unlikely for a team to have a
zero net rating, since a zero net rating means that this team gets and
losses the same amount of points at the same time. But we can see that
winning team still will have a higher net rating than all the net rating.
The last thing we look at is the free throw. As we can tell from the
graph, there is no significant difference between winning teams and
all the teams. But the losing team do have a larger tail to the left, and a
higher peak.

5 Prediction: Who Would Win in the Tourney?

After our exploratory data analysis, we are now well-quipped for our
ultimate goal: predicting who would win in the tourney. In this sec-
tion, we employ machine learning techniques to solve the problem.
We would prepare a common training set and evaluate three mod-
els: support vector machine with linear kernels, random forest, and
AdaBoost.
We focus on predicting the result for any given match in the tour-
ney, given the two participating teams. The final winner can be pre-
dicted by using our estimated predictors repetitively following the
tourney rules.
The core idea behind our approach is to build a profile for each
team using their performance in the regular season. We assume that
there exists a prediction function, which takes two teams’ profile
as input and output the winner. Our work here is to estimate this
function.
In the first hypothesis, we test whether the location of the match
has impact on the match results, and the conclusion is positive. As a
result, we, in addition to the team profile, also supply location infor-
mation to our predictor.
Which information should be used to build the team profile? Ma-
chine learning algorithms are good at finding important features
among all available variables automatically. For that to work, how-
ever, we need to supply enough information. Moreover, we need to
avoid providing useless information.
Our second and third hypotheses shows that match score could
fluctuate depending on whether a team is strong or weak. However,
any team that made into the tourney is strong, and the absolute value
of score does not determine who wins and who loses. Therefore, we
decide to leave out match score from our team profile.
Our fourth and fifth hypotheses checks whether advanced metrics
give a good insight in the result of a match. While information like of-
fensive rating is correlated with the outcome a match, some others are
 google cloud and ncaa ml competition 2019-women’s 17

not so helpful. Therefore, instead of doing manual feature engineering

and selection, we provide all information required to compute such
ratings, like field goal, three pointer, free throws, and so on. The rest
of job is delegated to machine learning models.
One might question whether machine learning models are pow-
erful enough to derive all advanced metrics. On the one hand, we
need to be cautious regarding the approximation power to avoid over-
fitting. On the other hand, metrics like offensive rating are nothing but
linear combinations of features we use to build the team profile, and
it is reasonable to believe that the three machine learning models can
learn such linear relationships.
Team performance in any single match could be subject to many
variables and have a high variance. It is natural to build a team’s pro-
file using records from all matches it plays during the whole reason.
For each record, we need to summarize its change in all matches dur-
ing the season. We choose to compute the mean, median, standard
deviation, and skewness for each record, which, hopefully, would
provide sufficient information for these models to obtain a good ap-
proximation to the changes of those records over the season.
One pitfall is the attempt to use results across seasons. Since the
team members change from year to year, it is reasonable to believe
that team profiles also change from year to year. While in two consec-
utive season a team’s profile might not change much, it surely does be-
tween, say 2010 and 2018. Hence, for the simplicity of our model, we
treat a single team in two different seasons as essentially two teams.

Data Preparation
We first assign each team in each year a unique identifier, which
means each team would occur nine times. Then, we compute the
mean, median, standard deviation, and skewness for each advanced
metric. There are 13 advanced metrics, and we obtain a 13 × 4 = 52
dimension vector for each team serving as the team profile.
Then, for each match result in the given data, we get the team pro-
file vector for both teams. We append the row with a two-dimension
vector denoting which team is the host. It is similar to the one-hot
vector, with a slight difference that both dimensions can be zero if the
match happens in a neutral place.
Lastly, we generate the labels, naming the match result. Notice that
the data are organized that the first team is the winning one, but this
might allow machine learning models to over-fit. Hence, we randomly
toss a fair coin and decide for each match, whether the winning or the
losing team is the first one.
 google cloud and ncaa ml competition 2019-women’s 18

Results
We reserve 25% of the whole data (about 40k matches) as the vali-
dation set. We evaluate three models: linear support vector machine
with hard margin, random forest, and AdaBoost. Due to the pro-
hibitively high running cost, we only train these models on a small
portion of the training set, but evaluating them on the full validation
set.
We choose to train on 500, 1000, . . . , 7500, 8000 matches for all three
models. Figure 14 summarizes the performance of all three models.

0.750 Figure 14: Evaluation results on the

validation sets for each model with
the given training data size. The y-
axis represents the portion of correct
Prediction Performance

0.725 labels predicted by the trained model,

where 0.5 corresponds to the baseline—
randomly guessing the result—and 1.0
Model the oracle.
0.700 AdaBoost
RandomForest
SVM

0.675

2000 4000 6000 8000

Training Size

The model that performs the best is linear support vector machine,
followed by random forest, and the worst-performing one is Ad-
aBoost. The final performance for the three models, where the training
set contains 8000 matches, are 74.6%, 73.1%, and 72.6% respectively.
The performance for all three models increases steadily with more
training data. Moreover, from the figure we guess that we have not yet
reached the asymptotic line, that more computational power we can
have even better performance.

6 Theory

Binomial Distribution: Binomial distribution is a probability distri-

bution for discrete random variables. Binomial distribution gives the
probability distribution of x success out of n Bernoulli trials with suc-
cess probability p. X denotes the number of successes in n Bernoulli
 google cloud and ncaa ml competition 2019-women’s 19

trials. ( )
n x
P (X = x) = p (1 − p)n−x
x
MLE of Binomial distribution: For a binomial distribution with
parameter n and p, the probability mass function is
( )
n x
P (X = x) = p (1 − p)n−x
x
In order to calculate the MLE, we first derive the likelihood function:
∏n ( )
n xi
L(p) = p (1 − p)n−xi
i=1
x

Second, we calculate the log-likelihood, and denote constant by C:

∑
n ∑
n
ℓ(p) = ln (nC) + xi ln (p) + (n − xi ) ln (1 − p)
i=1 i=1

Then, take the first and second derivative of the log-likelihood func-
tion:
∑n ∑n
dℓ i=1 xi n − i=1 xi
= −
dp p 1−p
∑n
d2 ℓ −x n − i=1 xi
= 2 − <0
dp2 p (1 − p)2
At last, set the first derivative equals to zero and derive the MLE as
the following: ∑n
xi
p̂ = i=1
n
Offensive Efficiency: For calculating the offensive efficiency, we
count the number of points a team scores every 100 possessions.
Defensive Efficiency: For calculating the defensive efficiency, we
count the number of points a team allows every 100 possessions.
Mean: the mean of the sample is taking the sum of all the data in
sample and divide it by sample size, the mean of X is usually denoted
byX̄,
Median: it is a value that indicates the middle of the data sample,
located in the middle of the data sample.
Standard Deviation: Standard deviation is the summation of the
distance between data points and the median of the data.
Variance: The variance is the square of the standard deviation
Quantile: The First Quantile represents the cutoff point between
the first 25 percent of the data and the remaining 75 percent of the
data.The Third Quantile represents the cutoff point between the first
75 percent of the data and the remaining 25 percent of the data.
Linear Regression: Linear regression is a linear approach to
model the relationship between two variables.
 google cloud and ncaa ml competition 2019-women’s 20

Skewness and Kurtosis: Skewness and kurtosis measure the

normality of the probability distribution of a real-valued random
variable. It is defined as follows
( )3
1 n Xi − X̄
skew(X) = ∑i=1
n S
( )4
1 Xi − X̄
kurt(X) = ∑ni=1
n S

where X̄ is the expectation of X and S is the standard deviation:

1 n
X̄ = ∑ Xi
n i=1
1
S2 = ∑n (Xi − X̄)2 .
n − 1 i=1
For a normally distributed random variable X, skew(X) = 0 and
kurt(X) = 3.
Histogram: Histogram is a representation of data by its frequency.
It is usually used as a method to understand the distribution of a
continuous variable. By dividing data into bins, histogram shows the
density of data.
Box plot: Box plot is a graphical method to visualize a set of data
in quantiles. The maximum, minimum, median, first quantile , third
quantile, and outliers will be shown on the boxplot. Two or more
distributions could be graphed in the same plot for comparison.
Quantile-quantile Plot: Quantile-quantile plot (Q-Q Plot) is a
plot that is usually used to compare if the data is normal.
Wilcoxon Rank Sum Test: Wilcoxon Rank Sum Test is a order
test by ranking each set of data in ascending data. Then we compare
the order of the data to test if there exists difference of distribution
between two observation.
Chi-Square Test: χ2 test is used to determine if there exists a dif-
ference between the expected frequencies and the observed frequen-
cies. Assuming the null hypothesis is true, it evaluates the probability
of the observations that are made in the data. The chi-square test is
calculated as the below equation:

∑
k
(xi − mi )
2
X2 =
i=1
mi

Multiple Regression: Multiple regression predicts the outcome of

a response variable using multiple explanatory variables. It explores a
linear relationship between the independent variables and dependent
variable. For i=n observations, the formula for multiple regression is
the following:

yi = β0 + β1 xi1 + β2 xi2 + ... + βp xip + ϵ

 google cloud and ncaa ml competition 2019-women’s 21

where yi is the dependent variable, xi are the explanatory variables,

β1 is the interception point of y axis, βp are the slope coefficients for
each explanatory variable, and ϵ is the model’s error term.
Welch T-test: Welch T-test is a two sample test that is used to
discuss the difference between the two correspond means, with the
assumption that both samples are normally distributed.
Let X̄1 , s21 ,and N1 be the mean, sample variance, and the sample
size of the first sample.
¯ − X2
X1 ¯
Welch’s t = √ 2
S1 S22
N1 + N2

Kolmogorov–Smirnov Test: We use the Kolmogorov–Smirnov

test to determine whether there exist some differences of the given two
distributions. As a result, the KS test gives a p-value which indicates
the probability of the given two distribution come from the same
distribution. The test statistic of KS test for a given cdf is

Dn = sup|Fn (x) − Fn |
x

Central Limit Theorem: When independent random variables

are added into any given probability distribution, the central limit
theorem states the sum of that particular probability distribution
would tends be a normal distribution. This theorem states the fact
that a sum of many independent and identically distributed random
variables can be approximated by a normal distribution.
Bootstrap: Under the condition that the test has a random sample
with replacement, bootstrap allows the data to be more precise using
confidence intervals, prediction error, etc. It estimate the properties of
an estimator by drawing information from the approximated distribu-
tion. A common use of bootstrapping is inference the population from
sample data so that the sample data can be remodeled.
Random Forest: Random Forest is a flexible and popular machine
learning algorithm for classification, regression and other tasks. at
each tree split, a random sample of m features is drawn, and only
√
those m features are considered for splitting. Typically m = p or
log2 (m), where p is the number of features. It corrects for decision
trees’ over-fitting to their training set. Each tree has the same expecta-
tion.
Support-vector machine(SVM): One of the most widely used
clustering algorithms in industrial applications, SVM is one super-
vised learning algorithm with associated learning algorithms defined
by a separating hyper-plane. Given labeled training data, SVM algo-
rithm builds an optimal hyper-plane that assigns new examples to
categories. The kernels of SVM algorithms transform the problem
 google cloud and ncaa ml competition 2019-women’s 22

based on some linear algebra, which is called kernel trick. It has linear,
polynomial, and exponential kernels.
Ada boosting: adaptive boosting is a popular machine learning
meta-algorithm for classification, regression and other tasks, which
can incorporate multiple weak classifiers into a single strong classi-
fier. Basically, Ada boosting can decide how much weight should be
given to each classifier. Also, based on the results of the original clas-
sifier, we can select the training set for new classifier. The following
introduction is given by Trevor Hastie from Stanford University.
1
1. Initialize the observation weights wi = N,i = 1, 2, ..., N.

2. from m = 1 to M repeat the following steps

a. Fit a classifier Cm (x) to the training data using weight wi
b. Compute weighted error of newest tree
∑N
i=1 I(yi ̸= Cm (xi ))
errm = ∑N
i=1 wi

c. Compute
1 − errm
αm = log[ ]
errm
d. Update weights

for i = 1,....,N:wi ← wi ×exp[αm ×I(yi ̸= Cm (xi ))]and renormalize to wi to sum to 1.

∑M
3. Output C(x)= sign [ m=1 αm ×Cm (x)

7 Limitation

Even though we have access to a large database, we still encounter

various limitations during the project.

1. When investigating whether home field advantage exists in both

regular seasons and tournaments, we have simply eliminated the
games that are played in a neutral court, which could result in a
loss of data.

2. The objective of this project is to provide a valid prediction for the

tournament result; however, the data given does not include any
data about the referees. It is clear that referees could have some
impacts on the games results, but we are unable to clarify such
impact.

3. To investigate hypothesis 4, we have employed two metrics: de-

fensive rating and offensive rating. Since some of the teams in the
 google cloud and ncaa ml competition 2019-women’s 23

regular season are unqualified to enter the tournament, errors may

exist in both offensive rating and defensive rating for the teams that
enter tournament.

4. In hypothesis three, we have used bootstrap to randomly generate

more than 11,000 data points out of 800 data points we have since
we need to have same sample size for two samples. There could be
biases in the bootstrap process.

8 Conclusion

In the project, we attempt to make a prediction on the 2019 NCAA

Division 1 Women Basketball Tournament methods. In the beginning,
we have conducted researches on the tournament format as well as
some prevalent methods of predicting tournament results.
Before using advanced machines learning methods, investigation
on factors that affect winning rates are conducted. In the beginning,
we want to see if there exist home field advantage: home teams have
a higher chance of winning than the away team in a game. To accom-
plish this objective, we have first eliminated games that are played in
neutral courts and then conducted a one-sided hypothesis test and
confidence intervals. Results from both regular seasons and tourna-
ment have shown that home teams have a higher chance of winning
than away teams in any basketball game, the home-win rate is signifi-
cantly higher in the tournament.
For the second hypothesis, we have investigated if the tournament
games are more intense than matches in regular seasons. To mea-
sure competitiveness, we have measured the score difference between
two teams in a match: a smaller difference in points implies a more
competitive match. After using numerical representation and con-
ducting hypothesis test, we have shown that there does not exist a
statistically significant points difference between regular seasons and
tournaments.
Then we want to analyze if a strong team perform better when
facing a strong competitor. By drawing histograms and performing
KS test and Welch’s t-test, we conclude that a team will perform better
when facing a strong competitor. In hypothesis 4, we want to see if
there is a correlation between offensiveness and winning rate of a
team. We have used an additional metric named offensive rating to
run linear regression against the overall winning rate. Results have
clearly shown that there is a positive correlation between a team’s
offensive ratings and its winning probability.
For the last hypothesis, we have investigated what factors will af-
fect the overall chance of winning. By drawing distribution plots, we
 google cloud and ncaa ml competition 2019-women’s 24

conclude that offensive rating and net rating could affect the chance of
winning while free throw rates and defensive ratings do not seem to
affect the winning chance.
Built upon our knowledge, we use advanced machine learning to
provide a prediction for the 2019 tournament result.

9 Works Cited

1. Caudill, Steven B., and Norman H. Godwin. ”Heterogeneous skew-

ness in binary choice models: Predicting outcomes in the men’s
NCAA basketball tournament.” *Journal of Applied Statistics* 29.7
(2002): 991-1001

2. Hastie, Trevor, Trees, Bagging, Random Forests and Boosting, Lecture

3. “Road to the Championship.” *Women’s Basketball Road to the

Championship*, NCAA, https://www.ncaa.com/womens-final-
four/road-to-the-championship.

4. ”Selection Criteria” *Women’s Baksterball: Selection 101*, NCAA,

http://www.ncaa.org/about/resources/media-center/womens-
basketball-selections-101-selections

5. West, Brady T. ”A simple and flexible rating method for predicting

success in the NCAA basketball tournament: Updated results from
2007.” *Journal of Quantitative Analysis in Sports* 4.2 (2008).