Predicting NBA Playoffs Using Machine Learning

Full text

Predicting NBA Playoffs using Machine Learning
Sean Liu ∗
February 18, 2021
Abstract
This project attempts to predict the NBA playoff bracket using ma-
chine learning methods. It will consider one self-constructed model and
one machine learning model built from various machine learning algo-
rithms. The project will also determine the most efficient model for pre-
dicting NBA results and which way to select data gives an accurate and
consistent prediction. Finally, the project will investigate the effect of
home and away variables on the teams’ performance and the model’s ac-
curacy.
1 Introduction
The National Basketball Association (NBA) is considered as the premier bas-
ketball league for professional male basketball players in USA. It is made up of
30 teams, split into the Eastern and Western conferences [Aut01].
During the playoff, the top 8 teams from each conference (Eastern and West-
ern) are chosen to compete for the championship. The rankings are decided
based on the teams’ performances during the regular season. Then, the teams
play against each other with the 1st place playing against the 8th place, the
2nd place playing against the 7th place, etc. Each game will be a best-of-seven
match, and teams will rotate between home and away.
Like AlphaGo in the Go Contest [Sil02], machine learning is a well-known
prediction tool for complex process. The question of this research is, can ma-
chine learning be used to predict the NBA playoff bracket? And what is accu-
racy of such prediction compared with the real results? What machine learning
method is the best solution for NBA playoff bracket prediction?
To better analyzing and comparing the performance of the machine learning
in predicting the NBA playoff bracket, firstly we create a self-made prediction
model based on several key game variables that will impact the game result
mostly by our best knowledge about the NBA games. These variables include 1)
effective field goal percentage, 2) free throw percentage, 3) turn over percentage,
4) Offensive rebound percentage, 5) Defensive rebound percentage. And We
∗Advised by: Derek Sorensen
1

focused on working out the probability of Team A winning against Team B,
then applying this to every game in the playoff. Base on the historic game data
of 2014 to 2018, the playoff prediction of 2018,2017 and 2016 are carried out
and comparison with the real playoff brackets are also presented.
As for the machine learning model for the NBA playoff bracket prediction,
here, we are focusing on 5 different machine learning models that have already
been implemented in the Python Machine Learning Library (Scikit-learn), i.e.,
Logistic Regression (LR), Linear Discriminate Analysis (LDA), Support Vector
Machine (SVM), K-Nearest Neighbors (KNN) and Classification and Regression
Tree (CART) [Dhi03] - [TA11]. For comparison with the self-made model, the
same playoff prediction of 2018, 2017 and 2016 are carried out and comparison
among different machine learning models are given accordingly.
2 Result
2.1 Exposition of self-made prediction model
Based on the testing results for our self-made prediction model, we have the
following prediction results (Table 1). And the predicted playoff bracket with
the original ones are shown in Figure 1 and 2, with the prediction difference
highlighted in red color.
Figure 1: 2018 NBA Playoff (Prediction)
In summary, the self-built prediction model performed best when predicting
the 2018 playoff, getting an accuracy of 80%. The second-best prediction was
the 2017 prediction, obtaining an accuracy of 66.7%. The worst prediction was
for the 2016 bracket, only getting an accuracy of 53.3%. For the 2018 playoff
2

Figure 2: 2018 NBA Playoff (Original)
Figure 3: Playoff Prediction Accuracy of Self-made Prediction Model
prediction, we used teams’ statistics over 4 years from 2014-2018. For the 2017
playoff prediction, we used statistics of teams over a time of 3 years from 2014
- 2017. Finally, for the 2016 playoff prediction, we only used statistics of teams
over two years from 2014-2016.
2.2 Exposition of the Machine Learning Models
In order to evaluate the performance of machine learning in NBA playoff bracket
prediction, 5 different machine learning models are employed, which have been
implemented by Python Machine Learning Library (Scikit-learn), i.e., 1) Logis-
tic Regression (LR), 2) Linear Discriminate Analysis (LDA), 3) Support Vector
Machine (SVM), 4) K-Nearest Neighbors (KNN) and 5) Classification and Re-
gression Tree (CART) [Dhi03] - [TA11]. With two different training data selec-
tion methods and home/away investigation, the most accurate machine learning
model for NBA playoff bracket prediction is finally presented.
There are two ways in which we chose to select the data to train the al-
gorithm. 1. The first method was to consider each Team’s performance when
playing against all other teams 2. The second method was to evaluate a team’s
performance with one other specific Team to determine its win rate against that
3

specific Team
2.2.1 Method 1 – Selecting all data
In this model, we trained the different machine learning algorithms with all the
statistics of every Team. Like the self-constructed model, the 2018 prediction
used data over four years, the 2017 prediction used data over three years, and
the 2016 prediction used data only over two years.
From the above data, algorithms generally performed relatively well and
consistent in 2018 and 2017, except for the SVM model and the LR model
(Figure 3). The SVM model generally had a low and consistent prediction
accuracy in the three years, and the LR did significantly better in 2017 than in
2018.
Overall, LDA had the highest mean accuracy of 71.2%, followed by the
CART model, with a mean accuracy of 69.2%. The worst performing model is
the SVM algorithm with an accuracy of 0.436 only.
Figure 4: Playoff Prediction Accuracy of Different Machine Learning Models
with all data selected and average 30 trial runs
2.2.2 Method 2 – Selecting Partial Data
In this model, we trained the machine learning models with a partial amount
of data, which is only based on one Team’s performance against a specific op-
ponent. In other words, it is the data between two teams that we are trying to
predict.
Based on the data collected (Figure 4), there are no clear trends or patterns
available. All the model predictions have significantly large variations in each
year. It also doesn’t have a clear correlation to data size. Although 2016 was
again the worst year of prediction, the 2017 prediction did significantly better
than the 2018 prediction, proving that there are no trends.
4

In summary, selecting only the data where two teams played against each
other resulted in inaccurate and inconsistent predictions. It also means that
the models’ accuracy in this data selection method will not be considered when
calculating the best performing model due to the inconsistency and inaccuracy.
Therefore, it can be concluded that selecting all data is a better data selection
method.
Figure 5: Playoff Prediction Accuracy of Different Machine Learning Models
with partial data selected and average 30 trial runs
2.2.3 Home and Away Investigation
The home and away variables are widely considered an essential variable on a
team’s performance and players. The home-court will have more fans, and the
positive atmosphere will give the home team a spiritual boost, which may result
in a better performance.
An experiment is conducted by training and running the program three times
with different data. One will have all data from home and away games, another
will have only the data from home games, and the final one will have only the
data from away games. Differences between the predicted results are analyzed
and evaluated.
Based on the above data (Table 2), generally, teams had much better per-
formance and a higher win percentage when playing as the home team.
To summarize, the home and away variable greatly influenced teams’ per-
formance level and win percentage in general. In theory, the variable should not
affect model accuracy to a significant extent. But in this case, the model did
impact the model accuracy, which can be caused by other factors in real life.
5

Figure 6: 2018 LR Prediction Results
2.2.4 Most Accurate Machine Learning Model for NBA Playoff Pre-
diction
To conclude, the most accurate machine learning model at predicting the NBA
playoffs is LDA, which reached an accuracy of 71.2%. The performance of
the models at predicting with a partial amount of data is neglected since it is
considered that the data selection did not give useful information.
3 Discussion
3.1 Self-made prediction model
Based on the model’s accuracy and the size of the data, we see a trend between
the two variables, with 2018 having the largest dataset and the highest model
accuracy and 2016 having the smallest dataset and the lowest model accuracy
(Table 1). One possible explanation for the model’s changing performance is
that it works well with larger datasets while having lower performances when
working with smaller datasets.
Another possible reason is that the model is simply not consistent in predic-
tion. It might be a coincidence that there is a correlation between data size and
model accuracy since we only have data for three years of prediction. Further
investigation can be carried out to confirm the effect of the data size on the
model’s accuracy. This can be done by running the prediction model for more
years with different data sizes to understand the correlation between the two
variables better.
6

3.2 Machine learning prediction model
Similar as the results for the self-made prediction model, the prediction accuracy
of 2018 is the highest while the one of 2016 is lowest when all data are selected
to train the machine learning models (Figure 3). All algorithms except the SVM
model performed significantly worse in 2016. One hypothesis is that the models
reached a tipping point in 2016 when the data size is not big enough to support
accurate predictions.
Figure 7: 2016 Team Performance Prediction with All Data Selected
Another hypothesis is that there is an error in the program itself that is
causing the 2016 prediction to deviate. This can be seen through the models’
accuracy in 2016 (Table 3), except for SVM, all had an accuracy of 53.3%.
Here, it demonstrated that each model except for SVM had the same playoff
prediction for every round. Although teams have slightly shifting percentages
in different models, which may symbolize that there isn’t an error and that all
models are independent of each other, it is still doubtful that each model had
the same win percentage and same prediction. This will be a research question
for future investigations to confirm if there is an error in the program causing
the deviation in 2016, or the model reached a tipping point in data size that is
causing the variation to occur.
7

For model training with partial data selection (Figure 4), it can be concluded
that partial data selection method gives inaccurate and inconsistent predictions.
This is likely because there is very little data for a given pair of teams. To be
specific, two teams only play against each other ten times a year, with Team1
playing as the home team for five matches and Team2 playing as the home team
for another five games. Additionally, only 80% of the data are used to train,
meaning that only eight sets of data are provided for training each year. This
resulted in inaccurate predictions with inadequate dataset. It also means that
the prediction models will be more likely to give the two teams a 50 percent
win rate each due to the small amount of data for testing and training. This
will result in the program randomly selecting a winner between the two teams,
making the prediction model inconsistent.
Additionally, the two sets of predictions namely home and away should have
similar accuracy theoretically. This is because teams typically have a higher
win percentage when playing as the home team and a lower win percentage
when playing as the away team. If all teams perform better when playing as the
home team, they should get roughly the same increase in performance level, so
it should not affect the accuracy to a significant extent. This is the same when
teams are playing as the away team. They should all perform relatively worse,
so the models’ accuracy should not shift by a significant amount.
However, in this case (Table 2), the model accuracy did shift significantly, at
13%. This is due to outliers like the team MIN, which had a better performance
when playing as the away Team than as the home team. It is also because
different teams had different performance levels when playing as the home team.
For example, GSW had an increase in a win percentage of 30% when playing as
the home team. On the other hand, team PHI only had a 9% increase in win
percentage when playing as the home team. One hypothesis is that GSW has
more fans than other teams, so they have a better atmosphere when playing as
the home team. However, many other factors can decide a team’s performance
when playing as the home team and the away team. These factors can be further
investigated in the future.
4 Methods
To test the effectiveness of our self-made NBA playoff prediction model and all
the related machine learning algorithms, certain NBA historic statistics data
from 2014 - 2018 are needed, which can be access from many open source NBA
statistics. And these historic NBA statistics are usually saved as .csv file format,
which we can use the Python pandas library read-csv module to load the dataset
from the corresponding csv URL link, the format and header of the dataset is
of the following form (in Figure 5.)
8

Figure 8: NBA historic statistics dataset format and headers
4.1 Self-made prediction model
To start, we first created our own prediction model to predict the NBA bracket.
We focused on working out the probability of Team A winning against Team B,
then applying this to every game in the playoff.
We have to narrow our focus on specific game variables, which significantly
impact the game result. After some research, we decided to use the following
variables:
1)EFG% effective field goal percentage [Aut12], considers both 2pts field
goals and 3pts field goals in one variable and considered their weight with three-
pointers worth 1.5 times of a two-pointer.
2)FT% free throw percentage [Aut12], calculates the percentage of free-throw
makes for a specific team.
3)TOV% turn over percentage [Aut12], is an estimate of turnovers by a team
per 100 possessions.
4)ORB% Offensive rebound percentage [Aut12], is an estimate of the per-
centage of offensive rebound that a team gets.
5)DRB% Defensive rebound percentage [Aut12], is an estimate of the per-
centage of defensive rebounds taken by a team.
EF G% = (2 point field goals made + 1.5∗3point field goals made)∗100
T otal f ield goals made
F T % = f ree throws made ∗100
free throws attempted
T OV % = number of turnovers ∗100
field goal attempted + 0.44 ∗free throws attempted +number of turnovers
ORB% = off ensive rebounds ∗100
off ensive rebounds +opponent defensive rebounds
DRB% = def ensive rebounds ∗100
defensive rebounds +opponent offensive rebounds
9

4.1.1 Algorithms
The five variables that were chosen are considered the most impactful factors
in the game. The second step of our model is to decide on the algorithm we are
going to use to calculate the probability of Team A beating Team B; the chosen
algorithm was:
Pwi n =c1P1+c2P2+c3P3+c4P4+c5P5
Here, ciis the proportional correlation of the variable viwith winning. In
other words, the larger the value of ci, the more variable viwill contribute to
the winning of a game. Pi is the probability that Team A will have a higher
score than Team B on variable vi. By multiplying the probability of the two
factors together and adding all the numbers up for all five different variables,
we predict Team A beating Team B in a match.
4.1.2 cicalculation
The formula for ciis:
ci=ri
r1+r2+r3+r4+r5
Here, rirepresents the Pearson correlation coefficient of the variable viwith
winning. However, winning is a categorical value that cannot be used in the
Pearson correlation. Therefore we decided to represent winning with the point
difference between the two teams.
4.1.3 Picalculation
To calculate the value for Pi, we used the principle of confidence intervals, which
is defined to be the probability that a parameter will fall between two sets of
values with a specific confidence level. [Wil13]
1 - Calculate a 95% confidence interval of for both teams
2 - We defined the confidence intervals for Team A as [xA,yA] and Team B
as [xB,yB]
3 – The first case is when the intervals don’t overlap. In this situation, the
Team with the higher interval has a 95% chance of scoring higher. (Note: The
percentage might be slightly higher than 95%, but in this case, we consider it
as 95%.)
4 - The second case is when the two intervals overlap, and Team A has a
higher upper limit (yA¿yB). Here, the formula to calculate Pi is:
Pi= 0.95 yA−yB
yA−xB
5 - The third case is when Team B has a higher upper limit (yB¿yA). Here,
the formula to calculate Piis:
Pi= 1 −0.95 yB−yA
yA−xB
10

6 – The last case is when the two sets of data have the same upper limit
(yB=yA). Then Pi = 0.5 in this case.
4.2 Predicting the playoff bracket
In order to predict the playoff bracket, we created Python’s function to calculate
the probability of Team A defeating Team B, and we applied it to predict the
playoff bracket for 2018.
Figure 9: Python Code Snippet for Self-made Prediction Model
In the python code snippet (Figure 6), the function select team(), which
predicts the winner between Team A and Team B, is called many times. This
calculates the winners for the quarter-final, the semi-final, the finals, and in
the end, it calculates the winner of the year. This predicted playoff is then
appended to a list and compared to the actual result of the 2018 playoff to cal-
culate a prediction accuracy. The original playoff is pre-loaded into the program
beforehand.
4.3 Machine learning prediction models
In order to test if machine learning algorithm can be used to predict the NBA
playoff bracket and evaluate which machine learning model has the best predic-
tion accuracy, 5 different machine learning models that have been implemented
in Python Scikit-learn machine learning library are employed with two linear
(LR and LDA) and three nonlinear (KNN, CART and SVM) ones (Figure 6).
11

After loading the dataset from the historic NBA statistics CSV file, depend-
ing on the two different data selection methods, all data or partial data, together
with home or away analysis, data related to year of 2016, 2017 and 2018 for NBA
playoff prediction can be split into different data arrays, so that the prediction
accuracy of different machine learning models can be analyzed accordingly.
To test the prediction accuracy for each different machine learning model,
the dataset needs to be split into two sets, one for the model training and one
for the model prediction on 8:2 randomly selection basis, which means 80% of
the data will be used as training data and 20% will be used to evaluate the
prediction accuracy and the data is randomly selected.
After the dataset is split for training and validation, the fit function for each
machine learning model will be called to train each individual models. After
model training, the predict function for each machine learning model will be
called to make the final prediction based on the validation dataset generated
before and the prediction accuracy for each models will also be calculated by
the accuracyscoref unction(F ig ure7).
Note that in Python, loc function is a frequently used function to retrieve
partial data in the dataset related to certain variable value, like certain year,
certain team, etc.
Figure 10: Python code snippet for importing modules, functions and models
5 Acknowledgement
I would like to thank Mr. Derek Sorensen of Horizon Academic Research Pro-
gram (HARP), together with my parents, my sister for the support.
12

Figure 11: Python code snippet for dataset split and model cross-evaluation
References
[Anu06] Mehta Anukrati. A beginner’s guide to classification and regression
trees. https://www.digitalvidya.com/blog/classification-and-regression-
trees/, 0006.
[Aut01] No Author. National basketball association.
https://en.wikipedia.org/wiki/NationalBasketballAssociation, 0001.
[Aut09] No Author. Machine learning - logistic regression.
https://www.tutorialspoint.com/machinelearningwithpython/machine-
learningwithpythonclassificationalgorithmslogisticregression.htm, 0009.
[Aut10] No Author. Advantages and disadvantages of logistic re-
gression. https://iq.opengenus.org/advantages-and-disadvantages-of-
logistic-regression/, 0010.
[Aut12] No Author. Glossary. Basketball Reference, No Date,
https://www.basketballreference.com/about/glossary.htmltovpct, 0012.
[Cor05] Maklin Cory. Linear discriminant analysis in python.
https://towardsdatascience.com/linear-discriminant-analysis-in-
python-76b8b17817c2, 0005.
[Dhi03] K Dhiraj. Top 5 advantages and disadvantages of decision tree
algorithm. https://medium.com/@dhiraj8899/top-5-advantages-and-
disadvantages-of-decision-tree-algorithm-428ebd199d9a, 0003.
[Jos08] Starmer Josh. Linear discriminant analysis (lda) clearly explained.
https://www.youtube.com/watch?v=azXCzI57Yfct=516s, 0008.
13

[Nar04] Kumar Naresh. Advantages and disadvan-
tages of knn algorithm in machine learning.
http://theprofessionalspoint.blogspot.com/2019/02/advantages-and-
disadvantages-of-knn.html, 0004.
[Rus07] Pupale Rushikesh. Support vector machine (svm) – an overview.
https://towardsdatascience.com/https-medium-com-pupalerushikesh-
svmf4b42800e989, 0007.
[Sil02] David Silver. Mastering the game of go without human knowledge.
ARTICLE, doi:10.1038/nature24270, 0002.
[TA11] Ibrahem Abdlehameed Hassanien Aboul Ella Tharwatt Alaa,
Gaber Tarek. Linear discriminant analysis: A detailed tu-
torial. https://www.researchgate.net/publication/316994943Linear-
discriminantanalysisAdetailedtutorial, 0011.
[Wil13] Kenton Wil. Confidence interval.
https://www.investopedia.com/terms/c/confidenceinterval.asp, 0013.
14