STAT 410 Final Project Hongyu Zhang

Finding the Next NBA Superstar:

How to Predict a Player’s Performance?

Hongyu Zhang

What factors determine an excellent basketball player? This is an interesting and

significant question that concerns coaches, basketball managers, as well as many basketball
fans. For instance, Michael Jordan, the well-known NBA superstar, had 6 NBA
Championships and was selected to All Star 14 times1. In his career, he scored average 30.1
points per game2 and was considered to be one of the greatest players in NBA history3.
Similarly, Kobe Bryant, with 25.0 average points per game, won 5 NBA Championships and
was selected to All Star 18 times4. Can we find any similarities between these two players
that made them superstars in NBA history? Inspired by this intriguing problem, I want to find
out whether there is a way to predict a player’s performance using a variety of factors. The
goal of this study is to construct a statistical model that uses players’ statistics to explain his
performance in game. Solving this particular problem will give us a clue about the question
stated at the beginning: for basketball players, what factors influence their performance?

Introduction
The easiest way to evaluate a player’s performance is to see how many points in
average he scored per game. The higher his average points are, the better the player is. There
are only 2 ways to score points in a game: free throw and field goal. Free throws result from a
foul in opposing team. In a free throw, a player attempts to score points behind the free throw
line without opposed interference5. A field goal is “a basket scored on any shot or tap other
than a free throw”6. There are also many other moves in a basketball game, like rebounds,
steals, blocks and so on. All of these data are purely numerical and objective, and they can be
easily found from the official website of NBA or other sources. My plan is to construct a
linear regression model to explain players’ average points scored by using players’ data like
height, weight, field goal scored, free throw scored and so on. I plan to first construct a
relatively simple model that fits using few regressors. Then I hope to find the best model
from many variables using stepwise search.

Data Collection
I was able to find this dataset online from Cengage7. This dataset referred The official
NBA basketball Encyclopedia, Villard Books, which is an official book that records all NBA
players’ statistics. The dataset contains 54 players’ data in 5 categories: column 1 is height
(in feet), column 2 is weight (in pounds), column 3 is field (percent of successful field goals
out of 100 attempted), column 4 is free (percent of successful free throws out of 100
attempted) and column 5 is points (average points scored per game). Here, the response
variable is points, since it is the variable that I want to predict. I can read the data and extract
some useful information about this dataset. Below are the first 6 rows of this dataset and
basic summary statistics of these 5 variables.

1
STAT 410 Final Project Hongyu Zhang

To better understand this dataset, I can also visualize it by constructing scatter plot matrix of
the average points scored and each of the regressors.

scatter plot matrix

From this dataset, we can see that there might exist some relationship between average points
scored and the other 4 regressors, although the variation is relatively high. Next, I am going
to build linear regression models over this dataset and analyze this data.

Simple Approach
From the scatter plot, we can see that there probably exists some relationship between
average points scored and percent of successful field goals and free throws. Based on our
intuition, this also makes sense: if a player has high successful field goals rates and free
throw rates, then he will probably have high average points per game because of his
accuracy, and vice versa.
First, I need to make sure this model does not have multicollinearity issue. In the
design matrix X, the first column is column vector 1, and the second column is the percent of
successful field goals, and the third column is the percent of successful free throws. Using the
eigen function in R, I calculate eigenvalues of XTX:
94.7953878 0.3816589 0.1344273
T
The condition number of X X is:
!"#$ ().+(,-.+.
K = !"&' = /.0-))1+- = 705.1798

2
STAT 410 Final Project Hongyu Zhang

This condition number is not very large, which indicates that it is numerically stable to
calculate the inverse of XTX. I also calculate the result of XTX * (XTX)-1:
1.000000e6// 1.199041e90) 0
27.105427e90, 1.000000e6// 0?
7.105427e90, −7.327472e90, 1
We can see that X X * (X X)-1 is very close to the identity matrix. These evidences show that
T T

our design matrix does not have multicollinearity issue. We can then start build the model.
The model can be described as y = b/ + b0 x1 + b1 x2 + ε, where y is the average points
scored, x1 is the percent of successful field goals, and x2 is the percent of successful free
throws. In matrix notation, we have: y = Xb + ε, where y has dimension 54 × 1, X has
dimension 54 × 3, b has dimension 3 × 1, and ε has dimension 54 × 1. Using R, I obtain bB :
−15.27738
2 35.82503 ?
14.79905
and SSres is 1516.422 and 𝜎E1 (also MSres) is 29.16196. The fitting model becomes:
yG = -15.277 + 35.825x1 + 14.799x2.
Next, I need to investigate the normal assumption and homogeneous variance
assumption. I plot the normal QQ plot of standardized residues:

From this graph, we observe that the blue line (straight line of standardized residuals) is close
to the red line (reference line). This indicates that our normal assumption is good.
Also, I plot the residuals vs. fitted values graph:

In this graph, we observe roughly a random scatter of points around the horizontal axis.
Therefore, our homogeneity assumption is good. Since the normal assumption and the
homogeneity assumption both hold, our model is valid.

3
STAT 410 Final Project Hongyu Zhang

We have obtained the fitting model, and the next step is to conduct hypothesis tests
about our estimated coefficients. Using R, the p-value for the hypothesis test of b1 being 0 is
0.008606811, and the p-value for the hypothesis test of b2 being 0 is 0.0509969. Because
both p-values are relatively small, we do not have enough statistical evidence to support the
null hypotheses. Therefore, we reject the claims that b1 is 0 and b2 is 0, and these two
estimated coefficients are statistically significant.
I can verify my results by comparing them to the results obtained by lm function in R:

My results and the results computed by lm function are almost exactly the same (considering
rounding errors). Also, the overall F-test for this model is highly statistically significant
because of the small p-value 0.00678.
Currently, my model is yG = -15.277 + 35.825x1 + 14.799x2. To find out whether this
model fits the dataset well, we can plot the actual values vs. fitted values graph and add a
straight fit line for reference:

In this plot, we observe that roughly most of the points are near around the straight fit line,
which indicates that our model fits the dataset well in general. However, some points deviate
from the straight fit line very much, and this provides evidence that this model is probably
not the best model to fit our data. In the next step, I plan to use the stepwise approach to
obtain the best model from these variables.

Stepwise Search to Find the Best Model

Given our dataset, there are 4 possible regressors: height, weight, field (percent of
successful field goals) and free (percent of successful free throws). We can also add one
interaction term between percent of successful field goals and percent of successful free

4
STAT 410 Final Project Hongyu Zhang

throws to discover the association between these 2 variables. Our goal is to find the best
fitting model built from these regressors. We need to be cautious about what variables we
want to include in the model, because some of the variables may be highly correlated. For
example, there might exist a positive correlation between height and weight, since in general
taller people tend to also have higher weights. We can examine this correlation using R and
the correlation between height and weight is 0.834324. This result verifies my conjecture:
height and weight in this dataset are highly correlated. They carry similar information about
the average points scored, and our model might not differentiate their separate effects on the
response variable. Because of the high correlation, we cannot have both height and weight as
regressors in our model, otherwise this will cause multicollinearity issue. In order to find the
best set of variables, we can use the stepwise search approach.
To obtain the best set of variables, we can use forward selection based on Bayesian
Information Criterion (BIC). The reason I use BIC is that I want to obtain the best fitting
model using fewest variables. BIC is defined by:
BIC = -2logL(θB) + Klogn.
Compared to Akaike’s Information Criterion (AIC) which penalizes the complexity of model
by a factor of 2, BIC penalizes the complexity of the model by a factor of logn, where n is the
sample size. Using R, I do the forward selection process and graph the bar plot of the BICs of
different models:

From the forward selection result on the left and the bar plot on the right, we can see that first
I start with regressor 1 and the model has BIC 194.66. Then I find out that adding the
variable field can reduce the BIC to 192.07, so I add field to the model. Then I find out that
adding the variable height can reduce the BIC to 191.32, so I add height to the model. In the
end, I find out that there is no way to reduce the BIC further by adding more variables, so
forward selection ends and the final model I obtain has 2 regressors: height and field.
Besides forward selection, I also use backward elimination based on BIC. I start with
all potential regressors in the model, height, weight, field, free, and the interaction term
between field and free. In each step, I remove the regressor that will make the new model

5
STAT 410 Final Project Hongyu Zhang

have the lowest BIC. I continue this process until BIC increases or all variables have been
removed. Using R, I obtain the following results:

I first start with the model with all potential regressors, which has BIC 199.87. I find out that
removing weight will give me the model with the lowest BIC 195.96, so I remove weight
from the model. Then I find out that removing the interaction term between field and free
will give me the model with the lowest BIC 193.10, so I remove this interaction term. In the
next step, removing free will give me the lowest BIC 191.32, so I remove free. In the end,
there is no way to reduce the BIC further by removing regressors, so backward elimination
ends, and I obtain the final model with 2 regressors: height and field, which is the same
model I obtain in forward selection process.
In both forward selection and backward elimination, I get the same results. This
provides further evidence that the model with 2 regressors height and field is our final
model. Our final model is: y = b/ + b0 x1 + b1 x2 + ε, where y is points, x1 is height, and x2 is
field.

Analysis of Final Model

In matrix notation, our final model is: y = Xb + ε, where y has dimension 54 × 1, X
has dimension 54 × 3, b has dimension 3 × 1, and ε has dimension 54 × 1.
First, I check that XTX is not a singular matrix. The result of XTX * (XTX)-1:
1.000000e6// 7.105427e90, −2.842171e90)
2−2.842171e90) 1.000000e6// 0.000000e6// ?
90, 90,
−6.217249e −5.329071e 1.000000e6//
T T -1
We can see that X X * (X X) is very close to the identity matrix. This shows that our
design matrix does not have multicollinearity issue. We can start to construct the model.
Using R, I obtain bB :
15.209793
2−4.034628?
51.562276
E 1
and SSres is 1495.732 and 𝜎 (also MSres) is 28.76407. The fitting model becomes:

6
STAT 410 Final Project Hongyu Zhang

yG = 15.210 −4.035x1 + 51.562x2.

In order to investigate the normal assumption and homogeneous variance assumption,
I plot the normal QQ plot of standardized residues and the residuals vs. fitted values graph:

From the QQ plot, we observe that the blue line (straight line of standardized residuals) is
relatively close to the red line (reference line). This indicates that our normal assumption is
good. In the residuals vs. fitted values graph, we observe a random scatter of points around
the horizontal axis. Therefore, our homogeneity assumption is good. Since the normal
assumption and the homogeneity assumption both hold, our final model is valid.
After obtaining the final model, two questions arise: 1. Are there any differences in
average points scored for players who have the same percent of successful field goals?
2. Are there any differences in average points scored for players who have the same height?
To answer these questions, we can conduct hypothesis tests about our estimated
coefficients. Question 1 is equivalent to test whether b1 is 0. Hypothesis testing: H0: b1 = 0,
H1: b1 ≠ 0. Using R, the p-value for the hypothesis test of b1 being 0 is 0.03357903. Because
the p-value is small (< 0.05), we reject the null hypothesis that b1 = 0. Therefore, we have
statistical evidence to support that there are differences in average points scored for players
with the same successful field goal rates.
Question 2 is equivalent to test whether b2 is 0. Hypothesis testing: H0: b2 = 0,
H1: b2 ≠ 0. Using R, the p-value for the hypothesis test of b2 being 0 is 0.001161757. Because
the p-value is very small, we reject the null hypothesis that b2 = 0. Thus, we have statistical
evidence for differences in average points scored for players who have the same height. Both
estimated coefficients are statistically significant.
Using lm function, I can verify my results:

7
STAT 410 Final Project Hongyu Zhang

My results and the results obtained by lm function are almost the same, considering rounding
errors. In the summary, we see that the overall F-test for this model is highly statistically
significant with p-value 0.004776.
In order to see if the final model fits the dataset well, we can plot the actual values vs.
fitted values graph and add a straight fit reference line:

In this plot, we observe that roughly most of the points are near around the straight fit line.
This provides further evidence that the final model fits the dataset well.

Summary & Discussion

In this project, I first construct a multiple linear regression model of points and field
and free. Then, in order to find the best fitting model, I use stepwise search technique to find
the final model, which has 2 regressors: height and field. The final model is:
yG = 15.210 −4.035x1 + 51.562x2. The mean square error of the final model is 28.76407,
which is smaller than the previous model. And the sum of square residuals is 1495.732,
which is also smaller than the previous model. These evidences show that the final model is
better than the previous model. The forward selection and backward elimination based on
BIC also show that this final model is the best fitting model. According to the final model, if
player A is 6.3 feet tall and has 45% successful field goals, then his predicted average points
scored is approximately 12.99466. If player B is 6.4 feet tall and has 46% successful field
goals, then his predicted average points scored is approximately 13.10682. Basketball
manager and team coach will probably choose player B instead of player A. This final model
provides predictive power so that given a player’s statistics, we can predict his average points
scored per game.
Although this model fits the dataset well, future work can be done to improve it. First,
I want to include more players’ statistics in the current year, and include more variables in the
dataset, like assists, rebounds, steals, blocks and so on. I think if I enlarge the sample size and
include more variables in the dataset, I can improve the model to be more comprehensive and
accurate. Also, currently this project is about NBA players, but in the future, I want to put
NCAA (National Collegiate Athletic Association) players into consideration. I want to
improve the model so that it has predictive power for both NBA players and NCAA
basketball players. I can consider NBA player and NCAA player as a categorical variable and
find out whether there are differences between these 2 types of basketball players.

8
STAT 410 Final Project Hongyu Zhang

References
1: Michael Jordan career statistics:
[Link]
2: Michael Jordan career statistics:
[Link]
3: Fox sports ranking: [Link]
ranking-the-25-greatest-players-in-nba-history-100716
4: Kobe Bryant career statistics:
[Link]
5: Free throw definition: [Link]
6: Field goal definition: [Link]
7: Dataset source: [Link]
7e/students/datasets/mlr/frames/[Link]