2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

Using Convolutional Neural Networks to

Forecast Sporting Event Results
Deep Learning: Concepts and Architectures
pp 269-285 |
Cite as

Mu-Yen Chen (1) Email author ([email protected])

Ting-Hsuan Chen (2) 
Shu-Hong Lin (1) 

1. Faculty of Information and Management, National Taichung University of Science

and Technology, , Taichung, Taiwan
2. Faculty of Finance, National Taichung University of Science and Technology, ,
Taichung, Taiwan

Chapter
First Online: 30 October 2019

2
Citations
2.2k
Downloads

Part of the
Studies in Computational Intelligence
book series (SCI, volume 866)

Abstract

Sporting events like the FIFA World Cup and the World Baseball Classic have
increased in popularity, and the enthusiasm with which these competitions are
reported and commented on is evidence of their wide-reaching influence. These games
are popular discussion topics. Many who follow sports are only casual fans, but even
these “temporary” fans have the same expectations that all fans do: the teams they
support should be able to win. This study selects the National Basketball Association
(NBA) to represent competitive ball sports. Predictions herein are based on the
examination and statistical analysis of previous records of NBA games. This research
integrates the field of sports outcome predictions with the methodology of deep
learning via convolutional neural networks. Training and predictions are modelled on
statistics gleaned from a total of 4,235 games over the past three years. We analyze the
training results of a variety of model structures. While previous studies have applied
convolutional neural networks to image or object recognition, our study proposes a
specific encoding method that is integrated with deep learning in order to predict the
results of future games. The prediction accuracy rate of the model herein is
approximately 91%, while the deviation is approximately 0.2. These strong results
confirm the validity of our designated encoding method.

Keywords

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 1/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

Sports prediction  Deep learning  Convolutional neural networks 

National basketball association (NBA) 
Download
chapter PDF

1 Introduction

Whenever the FIFA World Cup or the World Baseball Classic tournaments begin, the
number of devoted or casual fans around the globe rises significantly throughout the
season. Online search trends in connection with the games also surge sharply. Based
on the hot search terms shown in Google’s Trends service, it is obvious that
competitive sports are of great interest to most people, especially during international
tournament seasons. Figure 1 illustrates a Google hot trend for the search term
“soccer” in July 2018 [1], showing how the search trend soared upward in the middle
of June 2018 in connection with the FIFA World Cup.

Fig. 1

Google hot trend (Keyword: Soccer in Chinese)

Business Weekly [2] reported that the Hokkaido Nippon-Ham Fighters, a professional
baseball team in Japan, scouted and identified the most suitable players for each
position by using the Baseball Operation System (BOS) to analyze its player’s
capabilities. Sports vary in terms of types, rules and skills, and are also subject to
various factors that are difficult to quantify (e.g., weather conditions, morale, etc.). In
order to exclude a variety of unstable factors, this study selected the games of the
National Basketball Association (NBA) as the subject of analysis, because they are less
affected by factors that cannot be controlled, and the data are publicly available. Fong
[3] predicted the outcomes and scores of Major League Baseball (MLB) games.
However, since too many variables were taken into consideration, the model results
tended to fail to converge, reducing the accuracy of the predictions. Therefore, we
chose NBA games, which are subject to fewer determinants. The NBA is also the most
prestigious of the professional basketball associations.

The concept of deep learning has gained ground rapidly in recent years. After the
computer Alpha Go defeated South Korean Go grandmaster Lee Sedol, deep learning
became even more well known around the world. However, Alpha Go was soon
superseded by AlphaGo Zero, which defeated Alpha Go when the two programs played
against each other. This shows how rapidly deep learning has been developed and
optimized.

While deep learning has been applied in studies of image/object recognition (e.g.,
board game notations, paintings, people/things/objects in pictures, etc.), deep
learning has also been commonly applied to feature extraction. Rarely have previous
researchers used the deep learning method to predict the outcomes of competitive

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 2/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

sports. Such outcomes are more commonly predicted via the algorithms of statistical
regression analysis, artificial neural networks, or support vector machines. Thus, this
study has adopted the deep learning method to predict the outcomes of competitive
sports.

We use the records of NBA games from the past three years as the subject of analysis
and examination, and we model them in accordance with the deep learning method.
The objectives of this research are as follows:

To reduce the amount of data required for predicting the outcomes.

To convert the format of the original data to make it a proper fit for deep
learning.

To compare the accuracy of the prediction results from models with different
structures.

2 Literature Review

2.1 Convolutional Neural Network Architecture

A convolutional neural network is a class of feed-forward artificial neural networks. Its

neurons can respond to peripheral units within a specific area of coverage. Its
application to visual imagery has yielded excellent results. The structure of the
convolutional neural network mimics the neurons in the biological brain. Hubel and
Wiesel [4] first noticed that when a cat is visually observing different things, the
neurons in the brain react differently. Fukushima [5] proposed an unsupervised
conceptual network, which was able to learn geometrical variations in graphics on its
own.

Indolia et al. [6] presented the architecture of the convolutional neural network in
detail. The convolutional neural network is mainly composed of several structures: the
input layer, convolutional layers, pooling layers, activation function, fully connected
layers and loss function, and the output layer. This architecture is shown in Fig. 2.

Fig. 2

Architecture of convolutional neural network

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 3/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

The main structure in a convolutional neural network consists of one or more

convolutional layers with a fully connected layer at the end. It also includes the
associated weights and the results derived from pooling layers. This structure enables
the convolutional neural network to make use of the two-dimensional structural
characteristics of the data. The convolutional neural network can be trained via back
propagation, and thus needs fewer parameters than other deep networks.

2.2 Related Research Regarding Sports Predictions

Research regarding prediction requires access to historical data. Feature extraction

from existing data allows a future target to be predicted. Sports predictions are no
exception.

Craig et al. [7] proposed the relationship between a team’s offense and the ambient
temperature in American football games, and analyzed it using multi-level regression
analysis. The adopted dataset was provided by the National Football League (NFL).
The results showed that the level of aggression of the football team’s offense has a
positive linear relationship with the ambient temperature. The higher the ambient
temperature, the greater the intensity with which the offense plays. However, this
applied to home teams only.

Maszczyk et al. [8] used four factors to predict javelin throw distances: the first-step
distance in throwing the javelin, the strength of the body and arms, the strength of the
abdominal muscles, and the grip force of the hand. A regression model and an artificial
neural network model were established. Eventually, the differences between their
results were compared. The deviation of the artificial neural network was 16.77 m on
average, and that of the regression model was 29.45 m on average, indicating that the
artificial neural network had much better predictive capability. This also proved that
the prediction results of the artificial neural network were better than those of the
widely used regression model.

Bačić [9] used embedded sensors to test a golf swing application. The golf player’s
current physiological state and the flight path of the golf ball were recorded for each
swing. These data were quantified via HMMA (Human Motion Modeling and
Analysis). The golf ball paths were predicted using the swing lengths and the attack
angles.

Bunker and Thabtah [10] noted that machine learning was a very promising method in
the domains of classification and prediction. The large amounts of money involved in
wagers on sporting events demand highly precise predictions. The researchers
rigorously reviewed the literature on machine learning. The artificial neural network
was then adopted as a novel framework by which machine learning can be used as a
learning strategy.

Kipp et al. [11] established an artificial neural network model to predict the net joint
moments (NJM) of the hip, knees and ankles when a weightlifter is lifting weights.
Data from seven collegiate weightlifters were applied to training and testing. Although

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 4/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

the prediction performance levels were lower than those in previous literature, the
barbell motion and the time series data of NJM were similar to those of past high-level
athletes. This also proved that the prediction results were quite stable.

The extant literature has shown that research into any sort of prediction is heavily
dependent on historical data. Such data may also have a large number of columns. To
predict well, important features must be extracted. This study uses NBA data. After
selecting more important columns from the data, a convolutional neural network is
used to conduct feature extraction and model training.

3 Research Methods

3.1 Development Environment

The hardware for this research was a PC equipped with a 64-bit Win7 operating
system, with a 4-core CPU and 8 GB RAM. We used the Python programming
language to collect and preprocess game data. We also used an Oracle VM VirtualBox
virtual machine for deep learning development and model establishment. The
operating system was Linux and the programming language was also Python, with a
Pytorch framework.

3.2 Research Process

This study’s research process can be divided into five major steps: data collection, data
preprocessing, data encoding, data modeling, and performance evaluation. The
research model and process is illustrated in Fig. 3.

Fig. 3

Research model and process

Step 1: Data Collection

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 5/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

The NBA game data was downloaded for analysis from the NBA Taiwan official
website. Figure 4 shows a sample of the original data [12]. Python was used as a data
crawler, collecting game records from the October 2014 to April 2017 time period,
which comprised the past three NBA seasons for a total of 4,235 games. Home teams
had 2,473 wins and 1,762 losses during this period.

Fig. 4

The NBA data record of the home team in a single game

Step 2: Data Preprocessing

Once the Python data crawler had acquired the data, we found that differences in the
numbers of players in many games, as well as non-starters, resulted in missing values
and different numbers of columns. In the data processing stage, the null values in data
fields were filled with a value of 0 to complete the game data for experimental
purposes. We gathered a total of 20 data items for each player, as shown in Table 1.

Table 1

NBA game data columns

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 6/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

Title of
Title of column Title of column Title of column
column

3 point field goals

Minutes played Blocks Offensive rebounds
attempted

3 point field goals Defensive

Points Field goals made
percentage rebounds

Rebounds Field goals attempted Free throws made Turnovers

Assists Field goal percentage Free throws attempted Fouls

3 point field goals Positive/Negative

Steals Free throw percentage
made (±)

In 1989, Manley [13] proposed the term EFF (efficiency) to calculate NBA player
performance. EFF was calculated based on five additive data items (points, rebounds,
assists, steals and blocks) and three subtractive items (missed field goals, missed free
throws, and turnovers). We chose to include only the items that had more influence on
the overall game; thus, turnovers were not factored into the calculation. Regarding the
more influential items in terms of final points, the point calculation for this study
included free throws and three point field goals, which were calculated as percentages.
The percentages of field goals, three point field goals, and free throws were each
calculated by dividing the number of shots made by the number of shots attempted.
Since the numbers of shots made and attempted were already included in the
percentages, these two items were excluded from the calculation. On the other hand,
since the number of rebounds included offensive and defensive rebounds, these two
items were also excluded from the calculation. Since the efficiency value did not
include the number of fouls, this study also excluded that data.

The positive and negative values (±) were the player’s impact on the score when on the
court. When Player A was on the court, and he or a teammate scored 2 points, the
value was a +2. Lost points were subtracted directly. The positive and negative values
were attached to Player A until he was substituted or the game ended. The positive and
negative values already accounted for the influence of time, so the number of minutes
player was on the court was not included in the calculation. Table 2 shows the data
columns and their calculation methods as used in this research.

Table 2

Game data columns used in this research

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 7/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

Title of Calculation Calculation

Title of columns
columns method method

Points Additive item Field goal percentage Additive item

3 point field goals

Rebounds Additive item Additive item
percentage

Assists Additive item Free throw percentage Additive item

Steals Additive item Turnovers Subtractive item

Positive and negative

Blocks Additive item Additive item
values

Since there was no measuring standard for the final score and statistics in each match,
and overtimes would result in different lengths of time in a typical match, the
maximum values of the statistics might affect the calculated results. Thus, data needed
to be normalized afterwards. Normalization was conducted on a game-by-game basis
by calculating the level of influence of each statistic of each player in a single game.
The minimum value was 0 and the maximum value was 1. The normalization method
is shown in Eq. (1).

x − min (x)
y =
max (x) − min (x)

(1)

Step 3: Data Encoding

Since the convolutional neural network has often been used in studies on image
recognition, the corresponding data types did not match the data type of the present
research. At this stage, the player’s comprehensive skill performance levels calculated
after normalization in Step 2 were encoded to conform to the data type of the
convolutional neural network for the benefit of model construction.

The extent of each player’s influence with respect to each column in a single game was
already calculated in Step 2. Subsequently, those normalized columns were calculated.
The calculated result could be regarded as the comprehensive performance evaluation
of each player in a single game. With that, the home and visiting team players were
compared one by one, and the comparison results were encoded. The comprehensive
performance value of each player was compared separately from the perspective of the
home team. If the influence of the home team player was greater than that of the
visiting team player, the encoded value was 1; otherwise it was −1. If the extent of

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 8/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

player influence was equivalent, it was encoded as 0. The encoding process is shown in
Fig. 5. The NBA allows a maximum of 16 players to be registered in a given match;
hence, the encoded matrix size was 16 × 16.

Fig. 5

Schematic diagram of the matrix encoding process

Step 4: Data Modeling

The input data for modeling was the matrix data encoded in Step 3. It was a two-
dimensional 16 × 16 matrix of the final target field, i.e., win and loss. Win or loss was
determined from the perspective of the home team. The result of the game was
encoded as a 1 for a win, and a 0 for a loss.

We then established a convolutional neural network model with different structures

and parameters for the experiments. We expected to construct a better model by
adjusting parameters. Four models of different architectures were established in this
research. To prevent model overfitting, a dropout layer (setup value 10%) was set after
the pooling layer. The detailed architecture settings of the models are shown in
Table 3. These four models conducted 200 training iterations of 10 cross-validations.
Other detailed parameter settings of the models are shown in Table 4.

Table 3

CNN architectures under different structures

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 9/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

CNN model architecture Model 1 Model 2 Model 3 Model 4

Convolutional layer 1 2 × 2 2 × 2 2 × 2 5 × 5

Pooling layer 2 2 × 2 5 × 5 2 × 2 2 × 2

Dropout setting 10% 10% 10% 10%

Convolutional layer 2 2 × 2 2 × 2 2 × 2 2 × 2

Pooling layer 2 2 × 2 2 × 2 2 × 2 2 × 2

Dropout setting 10% 10% 10% 10%

Convolutional layer 3 – – 2 × 2 –

Pooling layer 3 – – 2 × 2 –

Dropout setting – – 10% –

Fully connected layer 1 1 1 1

Table 4

Parameter settings of the CNN model

  Parameter values

Training iterations 200

Batch size 32

Learning rate 0.0015

Activation function ReLU

Loss value calculation Cross entropy

Optimizer Adam

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 10/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

Step 5: Performance Evaluation

The results of the convolutional neural network were evaluated based on different
architectures, parameters and datasets. The results produced by using different
datasets for experiments were also reviewed and analyzed. We compared the various
results to the artificial neural network to analyze the differences.

3.3 Experiment Design

In addition to the architectural differences, other parameters were also adjusted to

optimize the results of the convolutional neural network. We compared the sizes
between the convolutional layer and the convolution filter in the pooling layer, and we
also adjusted the batch size from 32 to 512. The learning rate was set to 0.0015. The
loss value was calculated via cross entropy. We used the Adam optimizer, which is
more commonly used than the stochastic gradient descent method and can also update
the weights in a more stable manner.

Experiment 1

The data content used in Experiment 1 was a two-dimensional matrix encoded by the
coding method of this research. The experiments were performed using the
aforementioned convolutional neural network model with different architectures. The
data content contained 4,235 records.

Experiment 2

The data content used in Experiment 2 was a two-dimensional matrix created by

horizontally flipping Dataset 1. The horizontal flipping concept is depicted in Fig. 6.
The left side of Fig. 6 is the already encoded two-dimensional matrix, while the right
side of Fig. 6 is the matrix result after reversing horizontally.

Fig. 6

Horizontally flipped matrix

Experiment 3

The data content used in Experiment 3 was composed of Datasets 1 and 2 so that a
large amount of data could be used to train the convolutional neural network. We
expected to be able to improve the performance of the model by increasing the data

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 11/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

amount. Dataset 3 had a total of 8,470 entries, the largest data volume of the three
sets.

Experiment 4

The data used in Experiment 4 was processed by randomly swapping the encoded two-
dimensional matrix by rows and columns. The concept of random swapping is
illustrated in Fig. 7. The left side of Fig. 7 shows the encoded two-dimensional matrix,
in which the red boxes are the columns to be swapped, and the blue boxes are the rows
to be swapped. The right side of Fig. 7 shows the result after random swapping.
Dataset 4 had 4,235 entries.

Fig. 7

Randomly sequenced matrix

After experiments were conducted with convolutional neural network models of

different architectures and parameters, in addition to the comparisons between each
model, the results were compared with those of the back propagation neural network.
Weka ver3.8 was used to establish the back propagation neural network environment.

3.4 Performance Evaluation

This research used the accuracy and loss of the final experimental prediction as the
evaluation criteria. A confusion matrix was used to calculate accuracy. As shown in
Table 5, Eq. (2) was used to compute accuracy. Loss was calculated via cross entropy
[14]. This computation is shown in Eq. (3). When the loss approached 0, the result of
the model prediction was closer to the actual result. Through the loss values, the
model could also perform back propagation to modify the weights and parameters
inside the model. The loss value did not fully represent the quality of the model. It
could only determine whether the prediction result of the model was close to the actual
values.

Table 5

Confusion matrix

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 12/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

Prediction reality Positive Negative

Positive TP (True Positive) FN (False Negative)

Negative FP (False Positive) TN (True Negative)

TP + TN
Accuracy =
TP + FN + FP + TN
(2)
H (p, q) = I ∑ px log qx

(3)

4 Experiment Results

4.1 Dataset Description

This research used four datasets with different contents to conduct the experiments.
The use of four datasets was intended to confirm the feasibility of our proposed
encoding methods. The contents of these four datasets are described respectively,
below. Table 6 shows the combination of the content and the purpose of each dataset.

Table 6

Description of each dataset

  Dataset 1 Dataset 2 Dataset 3 Dataset 4

Expect to
Confirm the Expect to Expect the encoding
retain its
feasibility of the enhance model method of this
Purpose of feature values
encoding method performance research will not be
experiment after the
proposed in this via increasing affected by the
matrix is
research data amount arrangement order
flipped

The result
The matrix data
created by The summation The matrix created by
Data of the encoding
horizontally of Dataset 1 and randomly swapping
composition method proposed
flipping Dataset 2 the sequence
in this research
Dataset 1

Data
4,235 4,235 8,470 4,235
amount

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 13/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

4.2 Results of Experiments 1 and 2

This section shows the experimental results of the four models with different
architectures after using Datasets 1 and 2 under the conditions of different batch sizes.
The data contents of Datasets 1 and 2 were the already encoded two-dimensional
matrix data. The experimental results are shown in Tables 7 and 8; the contents are
the average of the results after conducting 10-fold cross-validation experiments.

Table 7

Experimental results with 10-fold cross-validation in Experiment 1

Batch size Experimental result Model 1 Model 2 Model 3 Model 4

Accuracy 90.24% 89.35% 89.80% 89.80%

32
Loss 0.206 0.213 0.192 0.224

Accuracy 90.62% 89.49% 90.02% 90.01%

64
Loss 0.206 0.215 0.212 0.216

Accuracy 90.90% 89.17% 90.36% 90.34%

128
Loss 0.209 0.225 0.198 0.222

Accuracy 91.10% 89.41% 90.51% 90.57%

256
Loss 0.201 0.226 0.202 0.194

Accuracy 90.98% 89.17% 90.37% 90.65%

512
Loss 0.198 0.237 0.196 0.186

Table 8

Experimental results with 10-fold cross-validation in Experiment 2

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 14/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

Batch size Experimental result Model 1 Model 2 Model 3 Model 4

Accuracy 90.51% 89.21% 90.08% 89.92%

32
Loss 0.225 0.227 0.233 0.228

Accuracy 90.72% 89.38% 89.94% 89.96%

64
Loss 0.239 0.214 0.217 0.232

Accuracy 90.78% 89.26% 90.20% 90.20%

128
Loss 0.213 0.246 0.229 0.228

Accuracy 91.02% 89.51% 90.40% 90.52%

256
Loss 0.208 0.227 0.232 0.208

Accuracy 91.22% 89.31% 90.52% 90.95%

512
Loss 0.205 0.233 0.229 0.214

The accuracy of the results of Experiments 1 and 2 were very close. Their accuracy
rates were approximately 90%, but Model 2 was slightly lower at about 89%. The
difference between Models 1 and 2 was the size of the pooling layer. In the pooling
layer of Model 2, a 5 × 5 convolution filter was used for sampling. The reason for the
poor result might be that the two-dimensional matrix used in this research was
16 × 16. Compared with the pixel matrix of a photo, the matrix of this research is
indeed too small. Direct oversized sampling could have destroyed the structure of the
small matrix, resulting in feature losses that eventually affected the accuracy. Although
the size of the convolution filter in Model 4 was the same as in Model 2, the sampling
method and the sequence were different in the convolutional layer; hence the matrix
structure was not likely to be destroyed, ensuring a more accurate prediction. Model 3
was the deepest model. Normally, performance is better when there are more
convolutional and pooling layers. There are two reasons for the setbacks. The first
reason is insufficient data volume. Deep learning requires a large amount of data in
order to raise the performance level of the model. The approximately 4,000 entries in
our dataset were insufficient in terms of data volume. The second reason is insufficient
matrix size. The convolutional neural network model was originally applied to image
sampling, and digital images can often measure more than 16 pixels on a side. Taking
the commonly used handwriting recognition MNIST dataset as an example, the image
size is 28 × 28 pixels, not to mention the much larger image sizes used in other studies
or in videos.

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 15/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

Increasing the batch size is likely to reduce the time required to train the model. The
loss value and accuracy would also be improved because the deviations could be more
easily identified in a larger batch size, and the loss value could hence be lowered. As
shown in the results, the batch size must be adjusted according to the structure and
data characteristics of the model.

4.3 Results of Experiment 3

This section shows the experimental results of the four models after using Dataset 3
under the conditions of different batch sizes. Dataset 3 was the summation of Datasets
1 and 2. The training data volume is also extremely important to deep learning. Hence
the datasets of Experiments 1 and 2 were summed up and trained in Experiment 3.
The experimental results are depicted in Table 9.

Table 9

Experimental results with 10-fold cross-validation in Experiment 3

Batch size Experimental result Model 1 Model 2 Model 3 Model 4

Accuracy 72.80% 71.43% 72.32% 71.84%

32
Loss 0.451 0.461 0.477 0.478

Accuracy 72.97% 71.37% 72.25% 71.97%

64
Loss 0.464 0.468 0.456 0.473

Accuracy 73.11% 71.18% 72.13% 71.81%

128
Loss 0.459 0.464 0.471 0.471

Accuracy 73.03% 71.29% 72.24% 71.57%

256
Loss 0.457 0.463 0.459 0.485

Accuracy 73.19% 71.07% 72.01% 71.53%

512
Loss 0.455 0.475 0.472 0.479

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 16/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

The accuracy of Experiment 3 was about 70%; the loss value rose to 0.4 or even 0.5.
Compared with Experiments 1 and 2, the accuracy and loss value were rather different.
Since Experiment 3 used twice the data volume for training, the results were expected
to be much better, and it took much more time to train.

After comparing the results of Experiment 3 with those of Experiments 1 and 2, it was
determined that the integrated dataset had failed to accurately predict the results
because of the different features of the datasets. In other words, when Dataset 1 was
flipped as Dataset 2, the results was two datasets that were completely different from
each other. As shown in the result of Experiment 3, the flipped matrix could not retain
the characteristics of the original images. In object recognition, the target for
identification is usually symmetrical in terms of left and right; i.e., after flipping, the
relative positional relationships of the features can still be found. However, in
Experiment 3, such a relationship could not be found. This made accurate predictions
difficult in Experiment 3.

4.4 Results of Experiment 4

This section shows the experimental results of the four models after using Dataset 4
under the conditions of different batch sizes. The purpose of Experiment 4 was to
determine whether after randomly swapping the matrix with a precondition, the
dataset could retain certain characteristics in order for the models to extract the
feature values. The results of Experiment 4 are shown in Table 10.

Table 10

Experimental results with 10-fold cross-validation in Experiment 4

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 17/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

Batch size Experimental result Model 1 Model 2 Model 3 Model 4

Accuracy 86.07% 85.59% 85.65% 85.09%

32
Loss 0.235 0.280 0.256 0.259

Accuracy 86.43% 85.72% 85.93% 85.63%

64
Loss 0.239 0.285 0.261 0.249

Accuracy 86.61% 85.74% 85.77% 85.71%

128
Loss 0.247 0.284 0.260 0.244

Accuracy 86.81% 85.69% 85.84% 85.89%

256
Loss 0.240 0.282 0.266 0.259

Accuracy 87.12% 85.78% 86.12% 86.04%

512
Loss 0.238 0.279 0.258 0.254

The results of Experiment 4 were similar to those of Experiments 1 and 2. Although

the results were not as good as Experiments 1 and 2, the accuracy of each model was
about 86%, and the loss value was about 0.25. The results of Experiment 4 were very
close. The differences between models were nearly identical to those of Experiments 1
and 2. The reasons were also the same.

With respect to the significance of randomly swapping the two-dimensional matrix in

Experiment 4, it should be noted that no matter how the matrix contents were
randomly swapped, only the rows and columns of the two-dimensional matrix were
used as units to perform random swapping, rather than completely randomizing the
entire two-dimensional matrix. In other words, regardless of the arrangement
sequence of each player in the home and visiting teams, the players were compared
one by one eventually. Though not as good as in Experiment 1, the prediction results in
Experiment 4 were pretty good, and the prediction capability was rather stable.

4.5 Discussion

A total of four different datasets, four different convolutional neural network

architectures, and five different batch sizes were used to verify the feasibility of the

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 18/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

encoding methods proposed in this research. Only Experiment 3 failed to achieve

better performance. Experiments 1, 2, and 4 achieved a good level of performance.
Among the four experiments, Model 1 stood out most prominently. In Table 11 shows
the best results of Model 1 in Experiments 1 to 4.

Table 11

Best results of Model 1 in each experiment

Batch Experimental Experiment Experiment Experiment Experiment

size result 1 2 3 4

Accuracy 90.74% 90.86% 73.21% 86.65%

32
Loss 0.182 0.246 0.421 0.274

Accuracy 91.31% 91.12% 73.31% 86.71%

64
Loss 0.177 0.211 0.451 0.193

Accuracy 91.43% 91.57% 73.59% 87.01%

128
Loss 0.211 0.183 0.455 0.253

Accuracy 91.78% 91.62% 73.41% 87.44%

256
Loss 0.207 0.205 0.457 0.255

Accuracy 91.24% 91.67% 73.60% 87.48%

512
Loss 0.172 0.212 0.450 0.230

We compared the foregoing with the artificial neural network, which was established
using Weka version 3.8. The setup parameters are shown in Table 12. The artificial
neural network used Dataset 1, for which the accuracy rate was about 91.64%. On the
other hand, the best accuracy in this research was approximately 91.78%.

Table 12

The artificial neural network parameter settings and experimental results

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 19/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

  ANN CNN

Hidden layer 1 1

Iteration 200 200

Batch size 256 256

Learning rate 0.0015 0.0015

Accuracy 91.64% 91.78%

The encoding method proposed in this research for the convolutional neural network
worked better than the artificial neural network. We found that, after encoding the
data, the regional feature extraction method of the convolutional neural network could
effectively extract features. Through the coding method proposed in this research, the
calculated values were used to compare and analyze the relative capabilities of the
home and visiting team players. These values were then encoded into two-dimensional
matrices, which were used as input data in the convolutional neural network for
training and modeling. From the coding in the two-dimensional matrices, the
differences in strengths between players could also be observed.

5 Conclusions

This study used a total of four experiments to verify the feasibility of the proposed
encoding method. The final results were compared with those of the traditional
artificial neural network. We found the accuracy of the convolutional neural network
to be slightly better than that of the artificial neural network. By taking advantage of
the regional feature extraction capability, the convolutional neural network was able to
achieve results that could be superior to that of the traditional artificial neural
network. Moreover, the encoding method proposed in this research allowed the
strengths of the players to be perceived via the matrices. This eliminated the need to
analyze and compare the players separately in the original data.

This research has several limitations. First, in comparison with other kinds of ball
sports, basketball has more controllable factors. Besides, its statistics are available to
the public. Thus, this study selected basketball as the research topic. Second, the
Python suite and programming language we selected for our deep learning
development requires the Linux operating system in order to be fully supported. Third,
the size of the matrix encoded in this research was limited to 16 × 16 because the NBA
allows no more than 16 players to be registered at a time for a given game.

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 20/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

As mentioned in Sect. 3.2, the variables of the NBA data are under analysis in this
research. The method in this research encodes the characteristics of basketball only. In
the future, more sports events will be encoded for the modeling of the convolutional
neural network. Adjustments of the model architecture can also raise the performance
and increase the accuracy of the prediction results.

References
1. Google: Google Trend. https://trends.google.com.tw/trends/?geo=TW
(https://trends.google.com.tw/trends/?geo=TW) (2018). Accessed 4 June
2018
2. Zhao, W.X., Wu, H.H.: Japan’s first–full of actuarial science. Business Weekly
1500 (2016)
Google Scholar (https://scholar.google.com/scholar?
q=Zhao%2C%20W.X.%2C%20Wu%2C%20H.H.%3A%20Japan%E2%80%99s
%20first%E2%80%93full%20of%20actuarial%20science.%20Business%20Wee
kly%201500%20%282016%29)
3. Fong, R.S.: Studies on predicting the outcome of professional baseball games
with data mining techniques: MLB as a case. Department of Information
Management of Chinese Culture University. Unpublished Thesis (2013)
Google Scholar (https://scholar.google.com/scholar?
q=Fong%2C%20R.S.%3A%20Studies%20on%20predicting%20the%20outcom
e%20of%20professional%20baseball%20games%20with%20data%20mining%
20techniques%3A%20MLB%20as%20a%20case.%20Department%20of%20Inf
ormation%20Management%20of%20Chinese%20Culture%20University.%20U
npublished%20Thesis%20%282013%29)
4. Hubel, D.H., Wiesel, T.N.: Receptive fields, binocular interaction and functional
architecture in the cat’s visual cortex. J. Physiol. 160(1), 106–154 (1962)
CrossRef (https://doi.org/10.1113/jphysiol.1962.sp006837)
Google Scholar (http://scholar.google.com/scholar_lookup?
title=Receptive%20fields%2C%20binocular%20interaction%20and%20functio
nal%20architecture%20in%20the%20cat%E2%80%99s%20visual%20cortex&a
uthor=DH.%20Hubel&author=TN.%20Wiesel&journal=J.%20Physiol.&volum
e=160&issue=1&pages=106-154&publication_year=1962)
5. Fukushima, K., Miyake, S.: Neocognitron: a self-organizing neural network
model for a mechanism of visual pattern recognition. In: Competition and
Cooperation in Neural Nets, pp. 267–285. Springer, Berlin (1982)
CrossRef (https://doi.org/10.1007/978-3-642-46466-9_18)
Google Scholar (http://scholar.google.com/scholar_lookup?
title=Neocognitron%3A%20A%20Self-
Organizing%20Neural%20Network%20Model%20for%20a%20Mechanism%2
0of%20Visual%20Pattern%20Recognition&author=Kunihiko.%20Fukushima&
author=Sei.%20Miyake&pages=267-285&publication_year=1982)
6. Indolia, S., Goswami, A.K., Mishra, S.P., Asopa, P.: Conceptual understanding
of convolutional neural network-a deep learning approach. Procedia Comput.
Sci. 132, 679–688 (2018)
CrossRef (https://doi.org/10.1016/j.procs.2018.05.069)

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 21/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

Google Scholar (http://scholar.google.com/scholar_lookup?

title=Conceptual%20understanding%20of%20convolutional%20neural%20net
work-
a%20deep%20learning%20approach&author=S.%20Indolia&author=AK.%20
Goswami&author=SP.%20Mishra&author=P.%20Asopa&journal=Procedia%2
0Comput.%20Sci.&volume=132&pages=679-688&publication_year=2018)
7. Craig, C., Overbeek, R.W., Condon, M.V., Rinaldo, S.B.: A relationship between
temperature and aggression in NFL football penalties. J. Sport. Health. Sci.
5(2), 205–210 (2016)
CrossRef (https://doi.org/10.1016/j.jshs.2015.01.001)
Google Scholar (http://scholar.google.com/scholar_lookup?
title=A%20relationship%20between%20temperature%20and%20aggression%2
0in%20NFL%20football%20penalties&author=C.%20Craig&author=RW.%20
Overbeek&author=MV.%20Condon&author=SB.%20Rinaldo&journal=J.%20S
port.%20Health.%20Sci.&volume=5&issue=2&pages=205-
210&publication_year=2016)
8. Maszczyk, A., Gołaś, A., Pietraszewski, P., Roczniok, R., Zając, A., Stanula, A.:
Application of neural and regression models in sports results prediction.
Procedia Soc. Behav. Sci. 117, 482–487 (2014)
CrossRef (https://doi.org/10.1016/j.sbspro.2014.02.249)
Google Scholar (http://scholar.google.com/scholar_lookup?
title=Application%20of%20neural%20and%20regression%20models%20in%2
0sports%20results%20prediction&author=A.%20Maszczyk&author=A.%20Go
%C5%82a%C5%9B&author=P.%20Pietraszewski&author=R.%20Roczniok&aut
hor=A.%20Zaj%C4%85c&author=A.%20Stanula&journal=Procedia%20Soc.%2
0Behav.%20Sci.&volume=117&pages=482-487&publication_year=2014)
9. Bačić, B.: Towards the next generation of exergames: flexible and personalised
assessment-based identification of tennis swings. In: 2018 International Joint
Conference on Neural Networks (2018).
https://doi.org/10.1109/ijcnn.2018.8489602
(https://doi.org/10.1109/ijcnn.2018.8489602)
10. Bunker, R.P., Thabtah, F.: A machine learning framework for sport result
prediction. Appl. Comput. Inform. 15(1), 27–33 (2019)
CrossRef (https://doi.org/10.1016/j.aci.2017.09.005)
Google Scholar (http://scholar.google.com/scholar_lookup?
title=A%20machine%20learning%20framework%20for%20sport%20result%2
0prediction&author=RP.%20Bunker&author=F.%20Thabtah&journal=Appl.%
20Comput.%20Inform.&volume=15&issue=1&pages=27-
33&publication_year=2019)
11. Kipp, K., Giordanelli, M., Geiser, C.: Predicting net joint moments during a
weightlifting exercise with a neural network model. J. Biomech. 74, 225–229
(2018)
CrossRef (https://doi.org/10.1016/j.jbiomech.2018.04.021)
Google Scholar (http://scholar.google.com/scholar_lookup?
title=Predicting%20net%20joint%20moments%20during%20a%20weightliftin
g%20exercise%20with%20a%20neural%20network%20model&author=K.%20
Kipp&author=M.%20Giordanelli&author=C.%20Geiser&journal=J.%20Biomec
h.&volume=74&pages=225-229&publication_year=2018)

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 22/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

12. NBA Media Ventures: NBA official website. https://nba.udn.com/nba/index

(https://nba.udn.com/nba/index) (2018). Accessed 4 June 2018
13. Manley, M.: Martin Manleys Basketball Heaven. Doubleday Books (1989)
Google Scholar (https://scholar.google.com/scholar?
q=Manley%2C%20M.%3A%20Martin%20Manleys%20Basketball%20Heaven.
%20Doubleday%20Books%20%281989%29)
14. Kline, D.M., Berardi, V.L.: Revisiting squared-error and cross-entropy functions
for training neural network classifiers. Neural Comput. Appl. 14(4), 310–318
(2005)
CrossRef (https://doi.org/10.1007/s00521-005-0467-y)
Google Scholar (http://scholar.google.com/scholar_lookup?
title=Revisiting%20squared-error%20and%20cross-
entropy%20functions%20for%20training%20neural%20network%20classifiers
&author=DM.%20Kline&author=VL.%20Berardi&journal=Neural%20Comput.
%20Appl.&volume=14&issue=4&pages=310-318&publication_year=2005)

Copyright information
© Springer Nature Switzerland AG 2020

About this chapter

Cite this chapter as:
Chen MY., Chen TH., Lin SH. (2020) Using Convolutional Neural Networks to Forecast Sporting Event
Results. In: Pedrycz W., Chen SM. (eds) Deep Learning: Concepts and Architectures. Studies in
Computational Intelligence, vol 866. Springer, Cham. https://doi.org/10.1007/978-3-030-31756-0_9

First Online
30 October 2019
DOI
https://doi.org/10.1007/978-3-030-31756-0_9
Publisher Name
Springer, Cham
Print ISBN
978-3-030-31755-3
Online ISBN
978-3-030-31756-0
eBook Packages
Intelligent Technologies and Robotics
Intelligent Technologies and Robotics (R0)

Reprints and Permissions

Personalised recommendations

© 2020 Springer Nature Switzerland AG. Part of Springer Nature.

Not logged in
National Taiwan University (1000835717) - National Taiwan University Medical Library
(1600136001) - 5166 SpringerLink Taiwan TAEBC eBook Consortium (3000182799) - 4803 SpringerLink Taiwan
eJournal Consortium 2010 (3000185072) - 7452 SpringerLink Taiwan TAEBC eBook Consortium - 2010
(3000222544) - 8429 SpringerLink Taiwan eJournal Consortium (3000251287) - 9243 SpringerLink Taiwan
Maths eJournal Consortium (3000651002) - Taiwan Trial Consortium (3001059271) - TAEBDC consortium

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 23/24
2021/12/25 上午6:33 Using Convolutional Neural Networks to Forecast Sporting Event Results | SpringerLink

(3001336515) - Taiwan TAEBC eBook Consortium 2015 (3001997788) - Taiwan eBook Consortium - various
(3002062831) - Taiwan Trial Consortium (3002083039) - Taiwan TAEBC eBook Consortium 2016 CRY
(3002339428) - Taiwan Consortium - Academic (3002622062) - Taiwan TAEBDC eBook Consortium 2021 Section
1 (3902483890) - 17007 Taiwan Consortium (3991454581) - Taiwan TAEBDC eBook Consortium 2014
(3991456190)
140.112.217.46

https://link.springer.com/chapter/10.1007%2F978-3-030-31756-0_9 24/24