Analyzing Factors that Lead to NBA Regular Season Success
Mohamad El-Hajj, Jackson Steed, Victor Gore, Craeg Jethro Infante, Raniel Flores, Danindu Wakista
and Mohammed Elmorsy
Computer Science Department, MacEwan University, Edmonton, Canada
Keywords:
Clustering, K-Means, Decision Trees, Sports Analytics, National Basketball Association.
Abstract:
The National Basketball Association (NBA) values regular-season success and acknowledges the crucial role
of a team’s roster composition in determining overall performance. This study uses machine learning tech-
niques, specifically unsupervised learning clustering and decision tree models, to predict the composition of
a winning roster. Our research identified three distinct clusters based on win percentage and the distribution
of players across different skill levels. Successful teams typically have more top-tier players and a significant
representation of players in the lowest skill level. In contrast, teams that spread their talent across the entire
roster are less successful. We have noticed that players with average to above-average skills are notably af-
fected by excessive playing time in the previous game, which leads to decreased performance and potential
losses for the team in the next game. Considering the time of year and the gap between games, we recom-
mend prioritizing the rest and recovery of top players, especially in the latter half of the season. It’s crucial to
ensure that players who are not as skilled as the top players but still make significant contributions to the team
maintain consistent performance, especially during the first half of the season. Analyzing height’s impact on
basketball player performance has revealed practical insights that can empower coaches and management. We
found that the shortest and tallest players often perform less than those of average height. Most top performers
in the NBA tend to have heights closer to the average. However, for players who frequently operate near
the net and encounter numerous rebound opportunities, it is generally preferable to have an average or taller
player for slightly enhanced overall performance compared to below-average height players. Teams can use
these insights to improve their roster construction and maximize player utilization by coaches from one game
to the next. This research provides practical strategies that can be immediately implemented to enhance team
performance.
1 INTRODUCTION
The NBA is considered the top professional basket-
ball league globally, with 30 teams, each with 15 tal-
ented players. Teams range from smaller market ones
like the New Orleans Pelicans to globally celebrated
franchises like the Los Angeles Lakers (NBA, 2023),
(Burns, 2023).
NBA teams increasingly use advanced analytics
to improve their operations and overall performance.
This includes identifying players with long-term po-
tential, reducing injury risks, and maintaining consis-
tent performance. Analytics have led to increased rev-
enue and a stronger winning record (Bishop, 2023).
Teams analyze player longevity, injury susceptibility,
performance at different stages of the season, play
styles, and more to evaluate a player’s suitability and
alignment with the team’s vision. This work suggests
going beyond standard statistics to using data mining
to ensure the selection of effective players and the for-
mation of successful team combinations.
NBA teams must balance building competitive
rosters with financial stability. The league enforces
a salary cap to create fairness and equal opportunities
for all teams. Teams aim to optimize player combi-
nations while managing expenses. Identifying under-
valued players through data analysis can help teams
secure talented players at lower costs.
Teams closely guard their proprietary advanced
analytics systems and methods to gain a competi-
tive edge. Despite this secrecy, certain statistics are
widely used by sports media, basketball aficionados,
and team management. One such advanced statistic
is win-shares, which aims to estimate the number of
wins each player contributes to their team over a sea-
son. It serves as a valuable metric that essentially
functions as a scorecard, measuring a player’s overall
impact on their team’s success. Another crucial statis-
El-Hajj, M., Steed, J., Gore, V., Infante, C., Flores, R., Wakista, D. and Elmorsy, M.
Analyzing Factors that Lead to NBA Regular Season Success.
DOI: 10.5220/0013041500003828
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 12th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2024), pages 83-94
ISBN: 978-989-758-719-1; ISSN: 2184-3201
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
83

tical measure is a player’s offensive rating, which as-
sesses their scoring effectiveness by evaluating their
scoring efficiency while considering the number of
possessions they utilize. Additionally, (Basketball
Reference, 2023a), the plus/minus statistic helps in
gauging whether a team outscores their opponent or
is outscored when a specific player is on the floor. In
this work, we use data mining methods to answer four
research questions were as follows:
1. What is the optimal strategy of a roster: Should
a roster concentrate on the best players and sup-
plement with weaker ones, or distribute talent
evenly across the team?
2. Can the timing of the year or the duration be-
tween games impact the performance?
3. Could a player’s performance in the current
game be influenced by their duration of participa-
tion in the previous game? What are the different
ways in which a player’s height may affect their
performance?
4. Can we explore the elements that led to the team’s
exceptional performance in a particular season?
Our primary objective is to have a positive impact on
coaches and management by offering them pioneering
and effective strategies for the development and man-
agement of their rosters. We are dedicated to provid-
ing valuable insights and implementing practical solu-
tions that will significantly improve the efficiency and
effectiveness of their roster management processes.
2 RELATED WORKS
Franks et al. (Franks et al., 2015). evaluated the de-
fensive metrics that influence the outcome of an NBA
game. They used a matchup matrix and a spatial re-
gression model to create a new metric. Their analy-
sis involved closely evaluating the number of points a
defensive player prevents an opponent from scoring.
They then identified the specific location on the court
and developed a disruption score to determine where
a defender is most likely to stop a shot.
McIntyre et al. (McIntyre et al., 2016) conducted
a study on defensive strategies used in response to of-
fensive screens in basketball. They used data mainly
from SportVU and carefully analyzed how the defen-
sive team reacts when the offensive team sets a screen.
The researchers categorized the screens based on their
location on the court and developed four distinct clas-
sifications for the defensive tactics used in response
to these screens. This detailed analysis better ex-
plains how offensive and defensive strategies interact
in NBA games.
Gonzalez et al. (Gonzalez et al., 2013) studied
how a player’s performance changes throughout a
season. They examined two main factors: the number
of minutes a player spent in a game and their Verti-
cal Jump Power (VJP). The researchers compared the
VJP of players in the starting lineup with those who
were nonstarters. Their analysis found that starters
who played an average of 27.8 ± 6.9 minutes per
game tended to increase their VJP compared to non-
starters, who played an average of 11.3 ± 7.0 minutes
per game. Specifically, starters increased their VJP by
77.3 ± 78.1 W, while nonstarters increased by 2160.0
± 151.0 W.
In their study, Drakos et al. (Drakos et al., 2010)
thoroughly analyzed NBA injuries over 17 years.
They examined the total number of injuries related to
the number of games played and calculated the in-
jury rate per thousand athletes. The researchers also
looked into the specific body areas affected by these
injuries. They attempted to identify potential correla-
tions between injuries and demographic factors such
as weight, height, player age, and NBA experience.
However, they did not find any significant correlations
between these variables.
Berri et al. (Berri et al., 2011) critically exam-
ined the reverse-order draft system for amateur play-
ers. This system is designed to give weaker teams an
advantage by allowing them to secure the first draft
picks. The researchers evaluated various factors such
as the players’ college performance, draft age, years
of college basketball experience, player height, and
position played. The study focused on college bas-
ketball players’ performance metrics and their influ-
ence on draft day. It found that the number of points
scored in college was a significant factor in the draft,
but it had minimal correlation with a player’s scoring
potential in the NBA. This suggests that the current
draft system may overlook crucial performance met-
rics when selecting future star players.
Fearnhead and Taylor (Fearnhead and Taylor,
2011) critically examine the prevalent rating systems
used to evaluate an NBA player’s performance. They
start by looking at the conventional regression model
that correlates a player’s performance with the num-
ber of wins their team achieves. They argue that
this model, while helpful, falls short in capturing the
player’s complete individual performance as it tends
to diminish the player’s contribution to the team’s suc-
cess. Fearnhead and Taylor have developed a new
model that provides a more accurate assessment of a
player’s abilities by separating their performance into
offensive and defensive ratings. This approach allows
for a more comprehensive evaluation of a player’s
skill set, taking into account their contribution to the
icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support
84

team beyond just the number of wins. The method
leverages data from multiple seasons to estimate a
player’s ability in a specific season and measures de-
fensive and offensive ratings separately, combining
them to give an overall rating.
Most literature used statistical models to obtain
their results. In our work, we introduced machine
learning models, such as classifications and cluster-
ing, to predict the answers to our questions.
3 DATA
For our analysis, we utilized two main datasets gath-
ered from Kaggle (Kaggle, 2022) and the NBA open
data (ESPN, 2023). The first dataset contains 26,652
rows and 21 columns, offering a comprehensive
overview of overall NBA game statistics. The second
dataset consists of 668,629 rows and 29 columns, pro-
viding detailed individual NBA player statistics per
game. These available datasets cover the period from
2003 to 2020 and were merged to create our primary
dataset. To enhance our analysis, we incorporated
win-share, offensive win-share, defensive win-share,
season team wins, season team losses from basket-
ball (Basketball Reference, 2023b), and NBA player
height data from ESPN (ESPN, 2023).
We analyzed individual player game perfor-
mances from 2003 to 2020, focusing on the more re-
cent style of play. The dataset contained over 600,000
rows. After data cleaning, around 550,000 rows were
left. We used feature selection and creation to retain
relevant columns such as season, plus-minus, height,
points, assists, rebounds, steals, turnovers, and more.
This helped prevent the curse of dimensionality. We
excluded irrelevant features like players’ nicknames
and team abbreviations from our analysis.
As per Basketball Reference, the win-share met-
ric is a player statistic designed to apportion credit
for team success among team members (NBA Stuffer,
2023a). Win-shares estimate the number of wins a
player contributes to their team through offensive and
defensive performances. Offensive win-shares cen-
ter on a player’s offensive contributions, such as scor-
ing points, creating team opportunities, and efficient
shooting (Sporting Charts, 2023). Defensive win-
shares isolate a player’s defensive impact, including
blocking shots, stealing the ball, and overall defen-
sive prowess (Sports Lingo, 2023). These metrics as-
sess an individual’s collective offensive and defensive
performance in a season.
We obtained the team’s season wins and losses
data from Basketball (Basketball Reference, 2023b),
which included the number of wins, losses, team
name, and season. Combining the wins and losses
provided the team’s total games for that season. We
then calculated the win percentage by dividing the
number of games won by the total games played that
season.
Win Percentage =
Number of Games Won
Total Games
We introduced game dates and minutes played as
key components to create multiple new dimensions.
The game date was crucial for identifying the date of
the previous game and calculating the gap between
each game. Additionally, we leveraged the game
date to categorize the season into early, mid, and late
stages, with each stage representing a three-month pe-
riod. By accessing the previous game date, we were
able to extract the minutes played in the preceding
game. These additional dimensions enable us to ex-
amine how the stage of the season, the duration be-
tween games, and the minutes played in the previous
game influence an individual’s current game perfor-
mance.
In the world of the NBA, plus-minus (+/-) is a sta-
tistical tool used to gauge the point differential when
a player is on the court (NBA Stuffer, 2023b). It
provides valuable insights into a team’s performance
with a specific player on the floor. A positive plus-
minus value indicates that the player’s team outscored
the opponents while they were on the court. Con-
versely, a negative plus-minus value suggests that the
opposing team outscored the player’s team during
their time on the court. Win shares are crucial met-
rics for a player’s season performance, making plus-
minus an important measure of a player’s game per-
formance. Additionally, points, field goals attempted,
free throws attempted, and turnovers are factored in
to create an offensive rating metric, offering a com-
prehensive analysis of a player’s offensive game per-
formance. The offensive rating is designed to quan-
tify a player’s offensive efficiency and contribution to
their team’s scoring, often expressed as the number of
points a player produces per 100 possessions (Fromal,
2023).
Player’s Possessions = Field Goals Attempted + 0.44×
Free Throws Attempted + Turnovers
Offensive Rating =
Points
Player’s Possessions
× 100
In our data analysis, we observed a wide range
of values within each category. To address this, we
opted to use a straightforward discretization method
called equal frequency binning for the mentioned val-
ues. Equal frequency binning involves dividing a di-
mension into bins to ensure that each bin contains a
Analyzing Factors that Lead to NBA Regular Season Success
85

Figure 1: Optimal number of clusters.
Figure 2: K-means clustering depicting heights and win-shares.
similar frequency of values. In this case, we created
six bins for each dimension. This approach ensures
that each category or bin will have an equally dis-
tributed representation when we run algorithms on the
data, thereby enhancing the resilience and accuracy of
our analysis.
4 ANALYSIS AND
METHODOLOGY
In the next section of our study, we will carefully an-
alyze and discuss the particular techniques utilized to
address each of the four research questions we have
identified. Subsequently, we will provide an in-depth
explanation of the results and valuable insights ob-
tained from our extensive analysis of each of these
research inquiries.
4.1 Question One: Optimal Strategy
When deciding whether teams should focus on hav-
ing a few standout players or distributing talent across
various positions, we utilized the k-means clustering
technique. This method helps identify data points that
are more similar to each other than to others. It in-
volves randomly placing centroids, which represent
the center of a cluster, and then assigning data points
to clusters. The algorithm then calculates the distance
between each cluster’s centroid and the specific data
point and assigns the point to the nearest centroid.
New centroids are computed based on the points be-
longing to each cluster, and this process is repeated
until the centroids stop changing significantly.
We also used the elbow method to determine that
three clusters, as in Figure 1, are optimal for analyz-
ing the distribution of player types from each team
and their respective win percentages. This strategic
insight allows us to understand the balance between
icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support
86

superstar players and talent spread across many po-
sitions. Before this analysis, all player features had
been categorized into bins, with bin 1 denoting the
lowest rating and bin 6 representing the highest. Sub-
sequently, we tabulated the number of players at each
level for every team and year, using these figures as
the basis for the clustering features.
After running the k-means algorithm on the
dataset, we observed the formation of three well-
defined clusters, as illustrated in Figure 2. These clus-
ters exhibit noticeable similarities in their features,
providing us with crucial insights into the character-
istics required for constructing a successful roster.
4.2 Question Two: Timing of the Year
Decision trees are used to evaluate how specific fac-
tors impact an NBA player’s performance. This in-
volves categorizing input data into different classes
using a classification method. Classification decision
trees make decisions based on the features of the data
and create rules to assign each instance to a specific
class (Raj, 2023). The key factors being examined in-
clude parts of the season, the number of days between
games, and the player’s previous and current game
minutes. Other features, such as the winning team,
were included to provide additional insight. The de-
cision trees use plus-minus and offensive ratings as
target variables to determine if the specified features
influence the player’s performance. Since this is a
classification problem, each feature was divided into
six equal-frequency bins to ensure the readability of
the decision tree outcomes. Before running the algo-
rithm, the dataset is split into a training set for build-
ing the tree and a testing set for evaluating perfor-
mance, with a random seeding and a 0.7 ratio. The
classification decision tree is applied to different as-
pects of the data. First, the data is examined as a
whole, and then each of the six bins in win-share,
offensive win-share, and defensive win-share are ana-
lyzed to compare differences between different player
levels. This results in 38 decision trees: 19 targeting
plus-minus and 19 targeting offensive ratings.
The study aimed to analyze how a gap influences a
player’s performance across different points in a sea-
son and to determine if this influence varies based on
the player’s skill level. Players in consecutive win
share brackets were combined to form three distinct
categories to simplify the findings.
The analysis utilized the f-regression function (Pe-
dregosa et al., 2011), incorporating the target plus-
minus to refine the assessment. This function eval-
uates the correlation between each regressor and the
target variable, converting these correlations into F-
scores. These scores measure the degree of linear de-
pendency between each regressor and the target, thus
aiding in identifying the most predictive features of
the outcome.
4.3 Question Three: Player’s
Performance
We conducted an in-depth analysis of various perfor-
mance metrics to investigate the connection between
a player’s height and their performance on the basket-
ball court. One of the metrics we found to be particu-
larly useful was offensive win-shares, which take into
account a variety of offensive statistics. As we sought
to develop a comprehensive performance statistic, we
initially considered multiple factors but eventually
honed in on the relationship between player height
and win-shares as the key components.
In narrowing our focus to player heights and of-
fensive win-shares, we employed the elbow method
as a crucial step in our analysis. This meticulous
approach allowed us to determine that the optimal
number of clusters was 3, laying the groundwork for
implementing a k-means clustering algorithm on the
data. This reaffirmed the precision and rigour of our
analysis.
Upon applying the k-means clustering algorithm,
we uncovered three distinct clusters that encapsulated
player heights ranging from 165cm to 231cm. By di-
viding win-shares into six bins, we could visually de-
pict player performance across different height cate-
gories described later
This led us to delve deeper into the relationship
between a player’s height and their ability to secure
rebounds. To analyze this, we sorted the players
into six performance tiers and then further catego-
rized them into three height groups. H1 represents the
tallest players, H2 consists of players of average NBA
height, and H3 encompasses the shortest players.
4.4 Question Four: Team’s Exceptional
Performance
During our analysis, we delved into the factors con-
tributing to exceptional athletic performances, focus-
ing on the remarkable success of the 2015-2016 War-
riors team. Notably, the Warriors concluded the
regular season with a historic 73–9 record, surpass-
ing the previous record of 72–10 established by the
1995–1996 Chicago Bulls. This achievement solidi-
fied their position with the best regular-season record
in the history of the NBA.
To conduct a comprehensive analysis of the
Golden State Warriors’ performance during the 2015-
Analyzing Factors that Lead to NBA Regular Season Success
87

2016 NBA season, we implemented a thorough
methodology. Our approach commenced by meticu-
lously filtering the dataset to exclusively focus on the
2015-2016 season. Subsequently, we meticulously
gathered and organized detailed statistical informa-
tion, including but not limited to points per game
and three-point shooting percentages, for each indi-
vidual player on the team. Given that each entry in
the dataset corresponded to a player’s statistical con-
tribution to a specific game on a specific date, we con-
scientiously compiled multiple entries for each player
to ensure an accurate representation of their perfor-
mance throughout the season.
We gathered detailed data for each player, includ-
ing their points per game and three-point percentages.
This allowed us to analyze their performance through-
out the season thoroughly. We expanded our analysis
to include the Cleveland Cavaliers, also known as the
Cavs, a professional basketball team based in Cleve-
land. Between 2015 and 2018, the Cavaliers faced the
Golden State Warriors in four consecutive NBA Fi-
nals, igniting a fierce rivalry and creating one of the
most memorable matchups in modern NBA history.
For the Cavaliers, we collected and examined their
corresponding performance metrics. By using a sim-
ilar methodology, we calculated the Cavaliers’ mean
points per game and three-point percentages, enabling
a comprehensive comparison between the two teams’
performance. This comprehensive approach provided
valuable insights into the factors contributing to each
team’s success.
5 RESULTS AND DISCUSSION
In this section, we will thoroughly examine the results
for each research question.
5.1 Question 1: Optimal Strategy
In our upcoming discussion, we will thoroughly an-
alyze the three primary clusters identified during the
initial cluster analysis. Our focus will be primarily on
examining the mean values for each feature. We will
also provide some insights based on the analysis of
median values, albeit to a lesser extent.
In Cluster 1 in Figure 3, teams experienced the
least success, with an average win percentage of 1.9,
equivalent to roughly 27-55 in the regular season.
These teams had the lowest number of top-end play-
ers, averaging 1.02, and the highest number of low-
end players, averaging 5. Their struggle to win is
attributed to a need for more high-end talent and an
abundance of low-end players.
The most compelling analysis arises from the
comparison of clusters 2 and 3. Cluster 2 has a diverse
distribution of talent, with teams possessing between
2.06 and 3.50 players in bins 1-5 and 1.53 players in
bin 6. Their win percentage is approximately 0.379,
translating to a 31-51 record. This group displays
marginal improvement over Cluster 1. Conversely,
Cluster 3 showcases a top-heavy lineup, with teams
having 3.21 players in bin 6, 2.45 in bin 5, fewer than
two players in bins 2-4, and 2.71 in bin 1. Their reg-
ular season record is roughly 0.622, equivalent to 51-
31. Although teams in Cluster 3 have more players in
bins 5 and 6, they also have more in Bin 1 compared
to the Cluster 2 teams. An analysis of win percent-
age makes it clear that the more successful teams are
those in Cluster 3.
A particularly interesting statistic reveals that
when there are three players in bin 6 and 2 players
in bin 5, successful teams in cluster 3 can assemble a
lineup of 5 players who are well above average, ensur-
ing a cohesive team with no weak links on the floor.
On the other hand, teams in cluster 2 cannot achieve
this and are more likely to field a lineup where at least
one of the five players on the floor is only slightly
above average or even below average. Additionally,
teams with lower win percentages need more super-
star players to rely on during crucial game moments.
At times, all a team needs is a brief period where
their best players completely take over a game, and
the more superstar players a team has, the more likely
they are to accomplish this.
Based on our research, it is recommended that
teams prioritize acquiring top-tier talent rather than
focusing on the depth of their roster or evenly dis-
tributing talent. This is particularly relevant in the
NBA, which is characterized as a superstar-driven
league. Our findings indicate that teams with a few
dominant superstars tend to achieve greater success
and have a higher regular-season win percentage. In
contrast, those lacking a superstar player often need
help to keep up with the competition.
5.2 Question 2: Timing of the Year
In this study, we analyzed three statistics: Defensive
Rebounds (DREB), Rebounds (REB), and Offensive
Rebounds (OREB). We categorized their values into 6
bins, as shown in Figure 4. In our analysis, regardless
of the win-share category, each bin displayed a con-
sistent pattern in the decision trees we later discuss,
with offensive rating as the target classifier.
We found that players who were on the court for
bin 1 or bin 2 minutes (equivalent to 20 minutes or
less, considered a low amount of time) tended to have
icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support
88

Figure 3: Number of players of each type and win percentage for teams in each cluster.
Figure 4: Defensive Rebounds, Rebounds, and Offensive rebounds used with Win share bins.
a bin 1 or bin 2 offensive rating, ranging from 0 to
70. This outcome was expected, as offensive rating
is heavily influenced by the points scored by an NBA
player. Limited time on the court leads to fewer op-
portunities to contribute offensively.
When we change the target classifier to plus-
minus, we observe different outcomes. The results
are affected by whether the team wins or loses. Gen-
erally, when a player’s team loses, their plus-minus
is between bins 1 to 3, regardless of their win-share
variation. On the other hand, if the team wins, the
plus-minus is either bin 5 or bin 6, indicating a high
overall performance if the player played 20 minutes
or more. These patterns align with our expectations,
Analyzing Factors that Lead to NBA Regular Season Success
89

suggesting that a player’s overall performance is posi-
tively influenced by the team’s victory and significant
playing time, showcasing their consistent contribution
to the team’s success.
These trends are generally expected. When we
look at plus-minus, we focus on players who fall
into the fourth category for each win-share category.
These players are average or above average and have
significantly contributed to their team’s success dur-
ing the season. They can be described as ”role play-
ers” or good players but are not considered ”all-star”
caliber players. These players either support high-
level talent in a roster by assisting them when they
are on the court together or by holding down the team
when their all-star players are not on the court.
When analyzing their decision trees, the number
of minutes played in the previous game becomes an
important feature in the player’s plus-minus. If the
team lost and the player’s previous game minutes
were over 25, the player’s current game performance
had a plus-minus in the first category, which is ex-
tremely low. However, if their minutes were less than
25, their current game performance had a plus-minus
in the second category, which is still low but not as
bad. The difference in plus-minus is small; however,
we can see that the excessive use of role players in
the previous game not only negatively affects their
overall performance in the current game but may also
cause their team’s loss. This suggests that role players
should be used cautiously if their previous game was
strenuous.
In the early part of the season, players in groups
1 and 2 are most affected by the gap between games,
while the impact on the rest of the players is mini-
mal. However, as we move into the later part of the
season, players in groups 1 and 2 still experience an
impact on their +/-. This impact is reduced compared
to the early part of the season. Once again, players
in the middle of the pack do not experience a major
impact, while top players see a significant impact on
their performance due to the gaps between games.
The results are shown in Figure, 5, it is worth not-
ing that the accuracy of these decision trees ranges
from 0.3 to 0.5, representing extremely low accuracy.
This could be attributed to the extensive use of equal-
frequency binning. Increasing the number of bins
from 6 might lead to more accurate results.
As the season begins, players in the lower skill
level groups, specifically bins 1 and 2, are most af-
fected by the extended breaks between games. This
can have a significant impact on their performance
and readiness. However, as the season advances, the
influence on players in these skill-level groups grad-
ually decreases. Meanwhile, players in the interme-
diate skill level range continue to encounter mini-
mal impact from the gaps between games, while the
top-tier players are notably affected by the extended
breaks, potentially impacting their momentum and
form.
Upon analyzing the impact of the timing of the
season and the duration between games on player per-
formance, the findings are as follows: Figure 6. In the
early stages of the season, underperforming players
are notably more affected by longer breaks between
games. As the season progresses, top-performing
players are increasingly impacted by the duration of
breaks between games, while mid-range players tend
to maintain consistent performance regardless of the
gap. From a strategic perspective, top players should
prioritize rest at the season’s commencement and re-
duce rest as the season advances. Moreover, due to
their versatility, role players can be utilized more fre-
quently. Lastly, players ranked at the bottom in terms
of performance would benefit from consistent game
time, particularly at the season’s onset.
5.3 Question 3: Player’s Performance
After analyzing the k-means clustering plot, we dis-
covered some fascinating results. Previous studies,
such as the one by Berri et al. (Berri et al., 2011),
have indicated that height plays a significant role in
the selection of amateur players in drafts. However,
our plot unveiled an intriguing pattern. The players
were effectively categorized into three clusters based
on height, with distinct groups for shorter, medium-
height, and taller players.
The analysis of the players’ heights in relation to
their performance yielded intriguing findings. Upon
closer examination, it was noted that the shortest play-
ers tended to exhibit lower offensive win-shares, sug-
gesting a diminished level of performance. How-
ever, this trend was not confined to shorter players,
as taller players also displayed a similar pattern. Inter-
estingly, a performance peak was identified within the
medium-height range, followed by a decline among
the taller players. These observations point to the pos-
sibility that optimal performance may not necessarily
correlate with extreme heights, but rather lie some-
where within the middle range of heights.
Upon analyzing the height-rebounds graph, a clear
pattern emerges indicating that players with higher
overall performance also excel in rebounds, in line
with expectations. Notably, players in the average
height category (H2) exhibit the most impressive re-
bound performance within each group. Consistently,
a trend is evident wherein shorter players tend to un-
derperform compared to their average or tall counter-
icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support
90

Figure 5: Decision tree with plus-minus as target classifiers in bin 4 of the win-share categories.
parts across offensive and defensive rebounds within
the same or lower performance groups. In summary, it
can be inferred that an average-height player is well-
suited for positions such as Center or Power Forward,
although locating such players may not always be fea-
sible. In such instances, sacrificing some performance
in favour of added height can lead to favourable out-
comes. However, it is imperative to stress that the pri-
mary focus should not solely revolve around acquir-
ing the tallest players; prioritizing the best performers
remains crucial. In scenarios where a choice must be
made between a shorter top performer and a slightly
less skilled player of average or above-average height,
the latter should take precedence.
5.4 Question 4: Team’s Exceptional
Performance
In the 2015-2016 season, the Golden State War-
riors finished first in the Western Conference with an
unprecedented 73-9 record, surpassing the previous
record set by the 72-10 Chicago Bulls led by Michael
Jordan. The Cleveland Cavaliers finished first in their
Eastern Conference with a 57-25 record. In our anal-
ysis of the most correlated factors contributing to ex-
ceptional achievements, we found that two statistics
played significant roles. These factors are the abil-
ity to score high points per game and the ability to
score many 3-pointers. These two aspects emerged
as strong indicators of exceptional performance and
were closely linked to achieving outstanding results.
When examining the statistical data of the Cleve-
land Cavaliers and the Golden State Warriors, it is ev-
ident from Figure 7 that the points per game graph
illustrates the significant scoring advantage of the
Golden State stars, Stephen Curry and Klay Thomp-
son, over the Cavaliers’ stars, LeBron James and
Kyrie Irving. While the remaining players on both
teams make valuable contributions to their respective
performances, it is noteworthy that Stephen Curry’s
exceptional average of 30.1 points per game stands
out as a rare achievement in the NBA, playing a piv-
otal role in the Warriors’ historic season. Our analy-
sis indicates a strong correlation between the presence
of high-scoring players and the attainment of this re-
markable achievement.
The second highly influential factor contributing
to exceptional achievements is the 3-point score per-
centage. In Figure 7.B, a significant disparity between
the two teams is evident, emphasizing one of the pri-
mary reasons for Golden State’s formidable perfor-
mance. Steph Curry and Klay Thompson’s shooting
percentages far surpass those of the Cavaliers’ start-
ing lineup (excluding centers due to limited data near
the basket). Furthermore, Harrison Barnes and Dray-
mond Green demonstrate superior 3-point shooting
percentages compared to Lebron James, Kyrie Irv-
ing, and Kevin Love. It’s noteworthy that although
the Cavaliers’ 3-point percentages were considered
good, the Warriors’ dominance completely overshad-
owed them.
6 LIMITATIONS AND FUTURE
WORK
As we move forward, we will focus on a comprehen-
sive analysis of NBA postseason games. It’s impor-
tant to note that our dataset is primarily focused on
the 82-game regular season, which means our insights
Analyzing Factors that Lead to NBA Regular Season Success
91

Figure 6: Fregression function visualization with specific bins correlated together.
into team performance during the playoffs will be
somewhat limited. However, the best-of-seven play-
off series format provides a unique opportunity for in-
depth analysis, offering a more streamlined approach
compared to the extensive data entries from the reg-
ular season. Moreover, we can delve into a thorough
examination of individual player performances, aim-
ing to gain insights into their influence on their re-
spective teams, especially those who made significant
contributions to their teams’ advancement in the play-
offs or those who were eliminated early.
7 CONCLUSION
Our analysis of the factors influencing NBA regular
season performance shows that a team’s roster com-
position significantly affects its success. We’ve iden-
icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support
92

Figure 7: A. Points per game and B. three-point percentage for Cleveland. Cavaliers (brown) and the Golden State Warriors
(blue).
tified three clusters based on win percentage and the
number of players at different skill levels. The most
successful teams tend to have a higher number of top-
end players and a significant number of players in
the lowest skill level. On the other hand, teams that
evenly distribute their talent across the roster tend to
be less successful. We’ve also discovered that play-
ers of average to above-average skill levels are most
affected by excessive playing time in the previous
game. If these players have logged significant min-
utes in the previous game, they are more likely to
show a decline in performance and potentially lead
the team to a loss in the next game. Taking into ac-
count the time of year and the gap between games,
we recommend giving priority to the rest and recov-
ery of the top players, particularly in the latter half of
the season. It is crucial to ensure that below-average
players maintain consistent performance, especially
during the first half of the season.
When we consider the impact of height on player
performance, we find that the shortest and tallest play-
ers tend to underperform compared to those closer to
average height. The majority of top performers in the
NBA have an average height compared to other play-
ers. However, for players who operate near the net
and encounter many rebound opportunities, an aver-
age or taller player is preferable to a below-average
height player for slightly better overall performance.
Analyzing Factors that Lead to NBA Regular Season Success
93

Coaches and management could use this information
to construct and deploy teams more effectively, lead-
ing to an increased win percentage in regular sea-
son games. Additionally, coaches could analyze suc-
cessful seasons, such as the Golden State Warriors
in 2015-2016, to identify important factors leading to
these achievements, such as having players who can
effectively score three-pointers.
REFERENCES
Basketball Reference (2023a). Glossary. https://www.
basketball-reference.com/about/glossary.html. Ac-
cessed: 10-15-2023.
Basketball Reference (2023b). Nba player stats:
Advanced. https://www.basketball-reference.com/
leagues/NBA 2022 advanced.html. Accessed: 12-10-
2023.
Berri, D. J., Brook, S. L., and Fenn, A. J. (2011). From
college to the pros: Predicting the nba amateur player
draft. Journal of Productivity Analysis, 35:25–35.
Bishop, E. (2023). 5 ways nba teams use an-
alytics to gain a competitive edge. https:
//www.sportskeeda.com/basketball/5-ways-nba-
teams-use-analytics-gain-competitive-edge. Ac-
cessed: 12-10-2023.
Burns, M. (2023). Nba team market size rank-
ings. https://hoop-social.com/nba-team-market-size-
rankings/. Accessed: 12-10-2023.
Drakos, M. C., Domb, B., Starkey, C., Callahan, L., and
Allen, A. A. (2010). Injury in the national basket-
ball association: a 17-year overview. Sports health,
2(4):284–290.
ESPN (2023). Nba player stats. https://www.espn.com/nba/
stats/player/\ /season/2020/seasontype/2. Accessed:
12-10-2023.
Fearnhead, P. and Taylor, B. M. (2011). On estimating the
ability of nba players. Journal of Quantitative analysis
in sports, 7(3).
Franks, A., Miller, A., Bornn, L., and Goldsberry, K.
(2015). Counterpoints: Advanced defensive metrics
for nba basketball. In MIT Sloan Sports Analytics
Conference, Boston, MA. Presented at the MIT Sloan
Sports Analytics Conference.
Fromal, A. (2023). Understanding the nba: Ex-
plaining advanced offensive stats and metrics.
https://bleacherreport.com/articles/1039116-
understanding-the-nba-. Accessed: 12-14-2023.
Gonzalez, A. M., Hoffman, J. R., Rogowski, J. P., Burgos,
W., Manalo, E., Weise, K., Fragala, M. S., and Stout,
J. R. (2013). Performance changes in nba basketball
players vary in starters vs. nonstarters over a compet-
itive season. The Journal of Strength & Conditioning
Research, 27(3):611–615.
Kaggle (2022). Nba games data. https:
//www.kaggle.com/datasets/nathanlauga/nba-
games?select=games details.csv. Accessed: 12-
10-2023.
McIntyre, A., Brooks, J., Guttag, J., and Wiens, J. (2016).
Recognizing and analyzing ball screen defense in the
nba. In Proceedings of the MIT sloan sports analytics
conference, Boston, MA, USA, pages 11–12.
NBA (2023). About the nba. https://www.nba.com/news/
about. Accessed: 12-10-2023.
NBA Stuffer (2023a). Analytics movement in the
nba. https://www.nbastuffer.com/analytics101/nba-
analytics-movement/. Accessed: 12-10-2023.
NBA Stuffer (2023b). Nba plus-minus and impact metrics
in basketball explained. https://www.nbastuffer.com/
analytics101/plus-minus/. Accessed: 12-14-2023.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,
Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P.,
Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Cournapeau, D., Brucher, M., Perrot, M., and Duch-
esnay,
´
E. (2011). Scikit-learn: Machine learning in
python.
Raj, A. (2023). An exhaustive guide to decision tree clas-
sification in python 3.x. https://towardsdatascience.
com/an-exhaustive-guide-to-\newlineclassification-
using-decision-trees-8d472e77223f. Accessed:
12-14-2023.
Sporting Charts (2023). What is offensive win
shares. https://www.sportingcharts.com/dictionary/
nba/offensive-win-shares-ows.aspx. Accessed: 12-
14-2023.
Sports Lingo (2023). Defensive win shares
(dws). https://www.sportslingo.com/sports-
glossary/d/defensive-win-shares-dws/. Accessed:
12-14-2023.
icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support
94