Scribd
Upload
17 views181 pages

T.C. Ankara Yildirim Beyazit University Graduate School of Social Sciences

:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:

Uploaded by

galatzphodunbavxjk
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
17 views181 pages

T.C. Ankara Yildirim Beyazit University Graduate School of Social Sciences

:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:.:

Uploaded by

galatzphodunbavxjk
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 181

T.C.

ANKARA YILDIRIM BEYAZIT UNIVERSITY


GRADUATE SCHOOL OF SOCIAL SCIENCES

THE EFFECTIVENESS OF ARTIFICIAL INTELLIGENCE


IN FINANCIAL ANALYSIS: ISE100

PHD THESIS

Fikriye KARACAMEYDAN

THE DEPARTMENT OF FINANCE AND BANKING

Ankara, 2023
T.C.
ANKARA YILDIRIM BEYAZIT UNIVERSITY
GRADUATE SCHOOL OF SOCIAL SCIENCES

THE EFFECTIVENESS OF ARTIFICIAL INTELLIGENCE


IN FINANCIAL ANALYSIS: ISE100

PHD THESIS

Fikriye KARACAMEYDAN

THE DEPARTMENT OF FINANCE AND BANKING

Supervisor
Assoc. Prof. Dr. Erhan

Ankara, 2023
PAGE OF APPROVAL

The thesis study prepared by Fikriye KARACAMEYDAN and title as


Effectiveness of Artificial Intelligence in Financial Analysis: ISE is unanimously
accepted as a PhD thesis at the Department of Finance and Banking of the Institute of Social

Juror Institution Signature


Assoc. Prof. Dr.
Accept X Rejection University
Prof. Dr. kay UNVAN
Accept X Rejection University
Asst. Prof. Dr. Haroon MUZAFFAR
Accept X Rejection University
Prof. Dr. M.
Accept Gazi University
X Rejection
Asst. Prof. Dr. Cevat RAHEBI
Accept Rejection University
X

Date of Defence: 08/02/2023

I approve that this thesis satisfies all the requirements to be accepted as a thesis for
the degree of doctor of philosophy at the Department of Finance and Banking of the Institute

Director of the Institute of Social Sciences Prof.


PLAGIARISM

I hereby declare that all information in this thesis has been obtained and presented in
accordance with academic rules and ethical conduct. I also declare that, as required by these
rules and conduct, I have fully cited and referenced all materials and results that are not
original to this work; otherwise I accept all legal responsibility. (08/02/2023)

Fikriye KARACAMEYDAN
ACKNOWLEDGEMENTS

I would like to express my deepest gratitude to my supervisor Assoc. Prof. Dr. Erhan

to The Council of Higher Education for the support and permission during my studies.

Besides, I am appreciative to the members of my thesis committee, Prof. Dr. Y


Akay UNVAN, Prof. Dr. Atilla Asst. Prof. Dr. Haroon MUZAFFAR, for their
contribution to my thesis.

Finally, I am grateful to my beloved mother Saliha and my dear daughters nd


Defne for their patience and support.
The Effectiveness of Artificial Intelligence in Financial Analysis: ISE100

In this thesis, the effects of macroeconomic indicators on the Borsa Istanbul


(BIST100) index are analyzed using a deep learning method based on artificial neural
networks, which is one of the technologies of artificial intelligence.

All the variables selected in the present study consist of the monthly closing prices
for the period January 2001-January 2022. The effects of macroeconomic indicators
consisting of the Dollar Rate (ERUSD, TL/$), Money Supply (M2), Producer Price Index
(PPI), Industrial Production Index (IPI), Active Bond Interest Rate (TYBY), Brent Oil Price
(BRT), and Gold Prices (GP, TL/Gr) on the BIST100 index are examined with both Logit
and Probit regression models with qualitative responses and LSTM (Long-Short Term
Memory) architecture and a deep learning model that is a machine learning method. The
performances of the methods are compared with each other. The research has revealed that
the effects of selected macroeconomic variables on the BIST100 index increase during crisis
periods and during the COVID-19 pandemic. It has been determined that the change in the
prices of stocks in BIST100 is largely explained by money supply, exchange rate, industrial
production index, and consumer price index. It has been determined that the changes in the
real sector have a significant impact on the capital sector. It has been determined that the
deep learning method is more successful in predicting financial crises.

The results of the analysis revealed that the LSTM deep learning model, developed
to analyze the effects of macroeconomic indicators on the Borsa Istanbul (BIST100) index,
has a very low error level and gives more effective and satisfactory results than Logit and
Probit regression models.

Keywords: Artificial Intelligence, Artificial Neural Networks, Deep Learning,


Financial Forecast, ISE100, Logit-Probit Model, LSTM, Stocks.

i
: BIST100

nitel tepkili regresyon modelleri, hem de LSTM (Long- Short Term Memory) mimarisi

-19

kuru,

etkin ve tatmin edici

Anahtar Kelime: -
.

ii
TABLE OF CONTENTS

ABSTRACT.................................................................................................................... i
............................................................................................................................ ii
TABLE OF CONTENTS ............................................................................................iii
LIST OF ABBREVIATIONS .................................................................................... vii
LIST OF FIGURES ..................................................................................................... ix
LIST OF TABLES ...................................................................................................... xi
1. INTRODUCTION .................................................................................................... 1

1.1. Stock Exchanges ............................................................................................... 7

1.2. Istanbul Stock Exchange (ISE) ......................................................................... 9

1.3. Purpose of the Research .................................................................................. 11

1.4. Contribution of the Research .......................................................................... 12

1.5. Definition and Limits of the Problem ............................................................. 13

1.6. Problem Solving ............................................................................................. 13

1.7. Organization of the Thesis .............................................................................. 14


2. QUALITATIVE RESPONSE MODELS ............................................................... 16

2.1. Linear Probability Model (LPM) .................................................................... 16

2.2. Logit Model .................................................................................................... 17

2.3. Probit Model ................................................................................................... 19


3. ARTIFICIAL INTELLIGENCE ............................................................................ 22
4. DEVELOPMENT PROCESS OF ARTIFICIAL INTELLIGENCE ..................... 25
5. ARTIFICIAL INTELLIGENCE TECHNOLOGIES ............................................. 31

5.1. Expert System ................................................................................................. 31

5.2. Machine Learning ........................................................................................... 31

5.3. Fuzzy Logic .................................................................................................... 32

5.4. Genetic Algorithms ......................................................................................... 32

5.5. Artificial Neural Networks ............................................................................. 33

5.6. Deep Learning................................................................................................. 34

iii
6. ARTIFICIAL NEURAL NETWORK .................................................................... 35

6.1. Definitions of Neural Networks in the Literature ........................................... 35

6.2. Biological Nervous System ............................................................................ 36

6.3. Artificial Neural Networks Elements ............................................................. 37

6.3.1. Inputs ................................................................................................................ 38

6.3.2. Weights ............................................................................................................. 39

6.3.3. Summation Function ........................................................................................ 39

6.3.4. Activation Function .......................................................................................... 41

6.3.5. Output ............................................................................................................... 42

6.4. General Structure of Artificial Neural Networks ............................................ 43

6.4.1. Input Layer ....................................................................................................... 43

6.4.2. Hidden (Intermediate) Layer ............................................................................ 44

6.4.3. Output Layer ..................................................................................................... 44

6.5. Architectural Structures of the Artificial Neural Networks ............................ 45

6.5.1. Classification of ANNs According to Connection Structures .......................... 45

6.5.2. Artificial Neural Network Classification Based on Learning Styles ................ 48

6.5.3. Artificial Neural Network Classification Based on the Number of Layers ...... 51

6.6. Learning Algorithms in Artificial Neural Networks....................................... 55

6.7. Learning Rules ................................................................................................ 62

6.7.1. The Hebbian Learning Rule ............................................................................. 62

6.7.2. The Hopfield Learning Rule ............................................................................. 62

6.7.3. The Delta Rule .................................................................................................. 62

6.7.4. The Gradient Descent Learning Rule ............................................................... 63

6.7.5. The Kohonen Learning Rule ............................................................................ 64

6.8. Application Areas and Properties of Artificial Neural Networks ................... 64

6.9. The Design of Artificial Neural Networks ..................................................... 68

6.10. Training and Testing of Artificial Neural Networks .................................... 68

iv
6.11. Selection of Network Structure..................................................................... 72

6.11.1. Determining the Number of Input Neurons and Output Neurons .................. 73

6.11.2. Determination of the Hidden Layer and the Number of Hidden Neurons ..... 73

6.11.3. Data Standardization ...................................................................................... 74

6.12. Determination of Neural Network Performance ........................................... 75

6.12.1. Determining the Stopping Criteria ................................................................. 77

6.12.2. Selection of Learning Algorithm .................................................................... 77

6.13. Advantages and Disadvantages of Artificial Neural Networks Applications77

6.14. Disadvantages of Artificial Neural Network Applications ........................... 78


7. DEEP LEARNING ................................................................................................. 80

7.1. Fundamentals of Deep Learning Architecture ................................................ 81

7.1.1. Input Layer ....................................................................................................... 82

7.1.2. Convolution Layer ............................................................................................ 82

7.1.3. Activation Layer ............................................................................................... 83

7.1.4. Pooling Layer ................................................................................................... 83

7.1.5. Dropout Layer .................................................................................................. 83

7.1.6. Fully Connected Layer ..................................................................................... 83

7.1.7. Classification Layer .......................................................................................... 84

7.2. Deep Learning Architectures .......................................................................... 84

7.2.1. Multi Layer Perceptron (MLP) ......................................................................... 85

7.2.2. Convolutional Neural Network (CNN or ConvNet) ........................................ 85

7.2.3. Recurrent Neural Network (RNN) ................................................................... 86

7.2.4. Long Short-Term Memory (LSTM) ................................................................. 87

7.2.5. Restricted Boltzmann Machine (RBM) ............................................................ 87

7.2.6. Deep Belief Network (DBN) ............................................................................ 88

7.2.7. Deep Auto-Encoder (DAE) .............................................................................. 88


8. LITERATURE REVIEW ....................................................................................... 89
9. METHODOLOGY AND DATA ......................................................................... 106

v
10. EMPIRICAL ANALYSIS .................................................................................. 121

10.1. Descriptive Statistics................................................................................... 121

10.2. Unit Root Test Results ................................................................................ 122

10.3. Logit ve Probit Models and Results ............................................................ 128

10.4. Deep Learning Models and Analyses ......................................................... 134

10.4.1. Preprocessing of Data Used in the Study ..................................................... 134

10.4.2. Deep Learning Network Architecture Used in the Study ............................. 135

10.4.3. Training and Test Phase of the Network ...................................................... 139

10.4.4. Deep Learning Analysis Results................................................................... 140

10.5. Comparison of Model Results .................................................................... 146


11. ARGUMENT ..................................................................................................... 149
12. CONCLUSION AND PROPOSAL ................................................................... 150
13. REFERENCES ................................................................................................... 153

vi
LIST OF ABBREVIATIONS

ADALINE : Adaptive Linear Elements


ADF : Augmented Dickey-Fuller Test
AI : Artificial Intelligence
ANN : Artificial Neural Networks
AIC : Akaike Information Criterion
ARIMA : Autoregressive Integrated Moving Average
ART : Adaptive Resonance Theory
BIC : Bayesian Information Criteria
CAE : Contractive Autoencoder
CAPM : Capital Asset Pricing Model
CML : Capital Market Line
CSD : Central Securities Depository
DAE : Denoising Autoencoder
DW : Durbin-Watson d statistic
DL : Deep Learning
EFAMA : European Fund and Asset Management Association
EMT : Efficient Market Hypothesis
GRNN : General Regressions Networks
GNP : Gross National Product
GDP : Gross Domestic Product
ICI : Investment Company Institue
IPI : Industry Price Index
ISE100 : Istanbul Stock Exchange 100
IEEE : Institute of Electrical and Electronics Engineers
LMS : Least Mean Squares
LVQ : Linear Vector Quantization
MAE : Mean Average Error
MSE : Mean Apsolute Error
MADALINE : Multiple Adaptive Linear Elements
MAPE : Mean Absolute Percentage Error
ML : Machine Learning
MLP : Multi Layer Perception

vii
MPT : heory
PPI : Wholesale Price Index/Domestic Producer Index
RBF : Radial Basis Funcitons
SML : Security Market Line
SPK : Capital Markets Borad of T rkiye
OM : Self-Organizing Feature Maps
TCCB :
TUIK : Turkish Statistical Institute
UCITS : Undertaking for Collective Investment in Transferable Securities
LSTM : Long Short Term Memory
ML : Machine Learning
MAPE : Mean Absolute Percentage Error
MLP : Multi Layer Perceptron
RCNN : Recurrent Neural Network National
SE : Stock Exchange
NYSTE : New York Stock Exchange
SAE : Sparse Autoencoder
VAE : Variational Autoencoder

viii
LIST OF FIGURES

Figure 1.1. Stock market performances (2022/03-2021/12) ................................................. 9

Figure 1.2. Number of investors and portfolio value in CSD ............................................. 10

Figure 1.3. Market capitalization of stock exchange companies (million dollars) ............. 11

Figure 2.1. Linear probability model ................................................................................... 16

Figure 2.2. Conditional probabililty model ......................................................................... 16

Figure 2.3. The expected value of the dependent variable .................................................. 17

Figure 2.4. The conditional expected value of the dependent variable ............................... 17

Figure 2.5. Logistic distribution function ............................................................................ 18

Figure 2.6. Probability of failure ......................................................................................... 18

Figure 2.7. Odds ratio .......................................................................................................... 18

Figure 2.8. Logistic regression model ................................................................................. 18

Figure 2.9. Utility index ...................................................................................................... 19

Figure 2.10. The probability that the threshold value is less than or equal to the utility
index .................................................................................................................................... 20

Figure 2.11. The standard normal cumulative distribution function ................................... 20

Figure 2.12. The graphs of the Logit and Probit models ..................................................... 21

Figure 3.1. AI, ML, ANN and DL ....................................................................................... 23

Figure 6.1. Structure of a typical neuron ............................................................................. 37

Figure 6.2. General structure of artificial neuron ................................................................ 38

Figure 6.3. Summation function .......................................................................................... 40

Figure 6.4. General structure of the ANN ........................................................................... 44

Figure 6.27. Graident descent .............................................................................................. 63

Figure 7.2. Elements of the deep neural network architecture ............................................ 82

Figure 9.1. ISE100-TYBY................................................................................................. 110

Figure 9.2. ISE100-ERUSD .............................................................................................. 111

ix
Figure 9.3. ISE100-M2 ...................................................................................................... 112

Figure 9.4. ISE100-IPI ...................................................................................................... 113

Figure 9.5. ISE100-PPI...................................................................................................... 113

Figure 9.6. ISE100-GP ...................................................................................................... 114

Figure 9.7. ISE100-BRT .................................................................................................... 115

Figure 9.8. Logit model ..................................................................................................... 117

Figure 9.9. Probit model .................................................................................................... 117

Figure 10.1. The probabilistic unit root process ................................................................ 123

Figure 10.2. ADF random walk equation .......................................................................... 124

Figure 10.3. PP model ....................................................................................................... 124

Figure 10.4. Original graphics of variables (level) ............................................................ 125

Figure 10.5. Graphs of variables with normalization ........................................................ 135

Figure 10.6. Simple architecture of deep learning network .............................................. 138

Figure 10.7. Training progress with adam ........................................................................ 142

Figure 10.8. Train and test results with adam ................................................................... 143

Figure 10.9. Training progress with sgdm ......................................................................... 144

Figure 10.10. Train and test results with sgdm .................................................................. 144

x
LIST OF TABLES

Table 4.1. Some important advantages in neural network research .................................... 30

Table 6.1. Summation functions in ANNs .......................................................................... 40

Table 6.2. Some activation functions .................................................................................. 42

Table 6.3. Network types and areas of success ................................................................... 72

Table 6.4. Performance criteria ........................................................................................... 76

Table 10.1. Some descriptive statistics for the dependent and independent variables ...... 122

Table 10.2. Unit root test results at level ........................................................................... 126

Table 10.3. Unit root test results at ln................................................................................ 127

Table 10.5. Logit model compatibility results ................................................................... 130

Table 10.6. Probit model compatibility results.................................................................. 130

Table 10.7. Results of the Logit model & the Probit model .............................................. 132

Table 10.8. Correct classification rate table for the Logit model ...................................... 133

Table 10.9. Correct classificaiton rate table for the Probit ................................................ 133

Table 10.10. Deep learning network architecture parameters ........................................... 139

Table 10.11. Simulation results for adam ......................................................................... 141

Table 10.12. Simulation results for sgdm .......................................................................... 141

Table 10.13. Deep learning results .................................................................................... 142

Table 10.14. Deep learning performance results ............................................................... 145

Table 10.15. Logit model performance ............................................................................. 147

Table 10.16. Probit model performance ............................................................................ 148

Table 10.17. DL model performance................................................................................. 148

xi
1. INTRODUCTION

The main purpose of economics and economic activities, which try to find the cause-
and-effect relationships related to the subjects it deals with, and to reveal them in the form
of scientific rules, is to increase the level of social welfare. Social welfare and economic
development are indicators of the social and economic development of societies. The easier
it is for individuals in a society to reach basic needs such as social security, education, health,
and housing opportunities, the more they feel safe and free, and the more savings turn into
investments, the higher the level of social welfare in that society. For this reason, the concept
of social welfare has become one of the issues that play an important role in scientific
platforms as well as in the political field for both developed and developing countries.

Social welfare and economic development are closely related to the growth of the
money market and capital markets, which constitute the financial markets of the countries.
However, there is no common view on the extent of the development of capital markets, a
good financial system has five components; a reliable public finance and public debt
management, reliable and stable financial regulations, the presence of various banks,
national and international, or both, a central bank that can balance domestic finances and
manage international financial relations, and a well-functioning stock market. A regular
financial system with these components plays an important role in the economic growth and
development of countries (Rousseau & Sylla, 2003, pp. 374 375). As it can be understood
from these defined components, the more developed of the financial market of a country, the
lower of the risk level and vulnerability in the society. In addition, this predictable structure
will play an important role in the development of countries, thereby increasing their social
welfare and by encouraging the private sector, and savings will be transformed into
investments, leading to an increase in economic growth.

Developed and developing countries have made various reforms in order to protect
their economies and minimize the effects of financial fluctuations due to the economic crises
experienced in the world from the 1980s to the 1990s. These reforms have been in the form

1
of practices that include financial liberalization policies such as liberalizing interest rates,
removing credit ceilings, reducing or completely removing the deposit reserve ratios that
banks have to keep at the Central Bank, opening the banking sector to both foreign and
residents and liberalizing capital movements 134). It
has been observed that the financial crises experienced in this period, especially in
developing countries, were experienced due to the financial liberalization steps taken to
sustain growth and the hot money that left the country due to economic and political risks,
or due to international capital movements, excessive borrow money, exchange rate policies
and high inflation rates. Also, some important examples of the crises and the costs caused
by these crises in the country s economies arising from speculative capital movements, are
the crises experienced in Turkey in 1994 and 2001, the Asian crisis in 1997, and the Russian
crisis in 1997-98.

In addition to these economic crises, the Mortgage crisis, which emerged in the USA
in 2007 and then spread to England and turned into a global crisis as of the second quarter
of 2008, is an important example as it shows the reflections of the crises in the world
53). It has been shown this and the other financial crises mentioned
above, regardless of the level of development, the markets of the countries greatly affect
each other, cause great financial fluctuations and ruins in the world, and that financial shocks
in large economies can lead to the bankruptcy of various financial institutions, especially
banks, and even the bankruptcy of countries.

These developments in the world financial market have shown how important it is to
forecast the possible chaos and uncertainties in financial markets and the fluctuations that
may occur in micro and macroeconomic factors affecting price movements, in order to
prevent and/or reduce the effects of financial breaks and uncertainties in the markets. This
situation has revealed the fact that economic decision-makers can gain significant
advantages and opportunities if they can accurately forecast the future value of capital
market instruments and asset prices.

The unexpected changes in stock prices in the United States, many European
countries, and Japan in the 1980s and 1990s attracted the attention of researchers and it was
concluded that these fluctuations may be due to macroeconomic factors (Kaymaz & Yilmaz,
2022). On the other hand, Fama (1990), Barro (1990), and Schwert (1990) concluded in their

2
studies that the changes in the prices of stocks representing the firm value are mainly caused
by the changes in the future cash flows and the changes in the discount rates (Barro, 1990;
Fama, 1990; Schwert, 1990). Similarly, Chakravarty et al., in their studies have
concluded that basic macroeconomic variables affect stock prices (Chakravarty, 2005; Ilahi
et al., 2015).

In traditional estimation fusion methods, various methods including averaging, least


squares, and mean absolute deviations are chosen, and a single superior estimation is tried
to be produced from the linear combination of individual estimations. In this case, the
possible disadvantage of this type of modeling is that it does the linear modeling. In other
words, it is the process of assuming a linear relationship between the variables. In reality,
there may be a linear or curvilinear relationship between the variables. Therefore, making
estimations based on linear relationships shows that traditional models do not guarantee the
accuracy of their estimations (Donaldson & Kamstra, 1999, p. 228). For this reason, model
estimation is made by combining models arising derived from different econometric and/or
computer technologies, and thus the model is tried to be strengthened by avoiding
specification errors, measurement errors, and deviations and minimizing these effects.

The developments and innovations in computer technology and capital markets in


recent years have led to an increase in the diversity of products in the financial markets, thus
the diversification of the study subjects in the field of finance and an increase in the interest
in forecasts for economic analysis. The processing of data that symbolizes real-life facts and
transforms them into a meaningful and decisionable form is called information . In the
process of transforming information into knowledge and knowledge into decisions , the
necessity of making financial analyses by means of statistical and econometric models in a
fast, economical, and detailed manner has led to the emergence of new analysis methods. In
financial forecasting, the methods used in the information processing process are: general
grouping and regression analysis methods; volatility methods; parametric statistical
techniques such as autoregression analysis, where the probability distribution depends on
certain parameters; they can be listed as non-parametric statistical techniques such as the
Friedman test, Sperman s rank correlation test, and covariance estimation, where the
probability distribution does not depend on certain parameters. In addition to these methods,
various Artificial Intelligence (AI) technologies, which are very compatible with the

3
structure of financial data based on computer technologies, have been widely used in recent
years (Henrique et al., 2019).

The fact that accurate estimations that reduce financial uncertainty and costs will
bring successful results and thus play an important role in reducing the problems and costs
that may arise for both investors and governments, and that the benefit functions can be
maximized, has increased interest in forecasting modeling and led to the emergence of new
forecasting techniques. The Artificial Neural Networks (ANN) method from AI
technologies, which is one of these new prediction techniques, has the ability to learn,
generalize, and work with incomplete, faulty, flawed, and even wrong data. It has become a
prominent method because it gives very successful results in classification, optimization,
pattern recognition processes, and especially in estimating nonlinear time series (Cheng et
al., 2022a).

It is formed by ordering the data observed in the financial markets and following each
other according to the order of occurrence; data such as stock prices, stock market index,
exchange rates, and interest rates constitute financial time series, which are the main source
of econometric studies . Time series are defined as numerical quantities in
which the values of the variables are observed consecutively from one period to the next
. Financial time series are not stationary
because they are statistically composed of trends, seasonal movements, cyclical movements,
and random movements (Kennedy, 1998, p. 288). For this reason, it is important to be able
to make forecasts based on the behavior of the series by using methods that will purify the
series from these components. In this context, AI technologies have become a very popular
method used in financial analysis.

ANNs can be defined as computer software developed to imitate the working


mechanism of the human brain to realize the brain s abilities such as learning, remembering,
and generating new information by generalizing without any assistance (Donaldson &
Kamstra, 1999, p. 228). They can accurately and perfectly model the relationship between
input and output through learning, despite being software that mimics the human brain,
without relying on any prior information or assumptions. Because of their ability to learn
from data without any prior knowledge, neural networks are very suitable for classification
and regression models. In addition, the ANNs are not linear in nature. Unlike many

4
traditional statistical and econometric methods, ANNs produce successful models using
complex, incomplete, faulty, and flawed data, both qualitative and quantitative data, and for
analyses where traditional methods cannot find a solution and produce weak or ineffective
results. For the reasons listed above, ANNs have started to be used as the best alternative for
data analysis (Kamruzzaman et al., 2006a, pp. 2 3). It is still successfully applied in many
fields such as engineering, industry, business, finance, education, military defense, and
health (Henrique et al., 2019; Quazi, 2022).

ANNs, which find a wide area of use in the field of economics and finance as well as
in many other fields, have become a very interesting method due to their features such as
measuring and analyzing without the need for prior knowledge, making easily continuous
function estimation, and making generalizations. Because of these features, it has started to
be used in many studies (Henrique et al., 2019; Patil et al., 2021). This interest in ANNs,
which have been shown to give successful results in modeling and estimating complex
problems, has also led to the diversification and enrichment of AI applications.

It is thought that applications produced using artificial neural network technology in


the field of corporate finance, such as financial simulation, investor behavior forecasts,
financial evaluation, loan approval, stock and asset portfolio management, pricing of public
offerings, and determination of optimal capital structure, can be developed with artificial
neural network technology (J. Hsieh et al., 2019, p. 12). Major fluctuations in the world
economies, major company bankruptcies, financial crises, and the increase in the importance
of determining the chaos and uncertainties experienced in financial markets have revealed
the necessity of benefiting from applications developed and/or to be developed by AI
technologies.

It is very important to make the right investment decisions for investors who want to
turn their savings into investments. Therefore, they need data or a signal to help them make
a decision. Accurate preliminary indicators to be obtained in estimating the prices of stocks
traded on the stock exchange or the direction of price movements will offer many
opportunities to investors. Political events, general economic conditions, basic economic
data, and investors expectations are important factors affecting the prices of instruments
traded on the stock exchange. However, by simplifying the analysis and assuming that each
variable behaves linearly prevents us from reaching the correct results. This situation leads

5
to the use of more accurate calculation and estimation methods offered by advanced
computer technologies in time series analysis due to their non-linear, high-frequency,
polynomial, dynamic, and complex structures. Thus, in order to solve complex real-world
problems, it is attempted to overcome the shortcomings of a single method by utilizing
various advantages of AI technologies. At the same time, the disadvantages arising from the
use of hybrid models and techniques are trying to be eliminated. Therefore, the number of
such studies is increasing (Enke et al., 2011, pp. 201 206; Enke & Mehdiyev, 2013, p. 636;
Quazi, 2022).

Capital markets contribute to economic growth and development by transferring the


funds held by investors who want to transform their savings into investments to people in
need of funds. Stocks are one of the various investment instruments offered by the capital
markets, especially the stock markets, which are regulated according to specific rules. The
stock market is significantly affected by economic, social, and political developments at both
national and international levels. Therefore, any information to be obtained about these high-
risk investment instruments will undoubtedly also affect the decision-making process.

Two types of risk that affect the capital markets are systematic risk and unsystematic
risk. Financial risk, management risk, and industry risk are also included in the list of
unsystematic risks while inflation risk, interest rate risk, political risk, market risk, and
currency risk are regarded to be systematic risks (Hull, 2007, pp. 9 12). Macroeconomic
variables are one of the most important findings that will provide an understanding of the
effects of these risk factors. Because of these characteristics, it can be observed that studies
on stock market indices in the literature frequently use fundamental macroeconomic
indicators such as inflation, interest rates, money supply, the industrial production index, the
gross national product, gold prices, oil prices, and exchange rates.

In this study, the predictive efficiency of these methods will be tested by using the
Logit and Probit models, which are traditional econometric methods, and a deep learning
neural network model from AI technologies to determine whether the changes in the value
of the Stock Exchange Istanbul 100 Index can be explained by selected macroeconomic
variables. Thus, it will be revealed from the methods used, that which modeling technique
can best predict the change in economic and financial system dynamics.

6
1.1. Stock Exchanges

Stock exchanges are markets where long-term investment instruments such as stocks
and bonds are bought and sold. The direct confrontation between fund suppliers and fund
demanders is the main feature of these markets. Fund transfers are made in a transparent
environment using standard methods in exchanges that have been established and operate
according to certain rules. Therefore, stock markets are a safe and easy investment area for
investors (Karan, 2013, p. 37). The fact that investors can take risks easily is very important
for the development of the economies of countries. So, the necessary fund transfer is
provided for the savings, which form the basis of economic activities, to be transformed into
investments and investments into production.

Although there are different opinions about the establishment of modern stock
exchanges, stock market activities started with the trading of shares of the The United East
India Company (Vereenigde Oost-Indische Compagnie (VOC) is considered the first
multinational company in the world and was the first company to issue stock) in the early
1600s. In the 1700s, London emerged as the financial and stock market center followed by
America and New York in the early 1800s. The world stock markets, which started to
develop all over the world after the 1950s, gained a global character after the 1900s (Karan,
2013, p. 37). As a result of technological developments, the increase in communication
opportunities and the developments in computer technologies have led to the development
and growth of financial institutions and the expansion of the borders within which they

markets. In addition, the globalization of financial markets has increased both capital
movements and the diversity of securities, thus the depth of the capital markets.

The world s most important stock exchanges are located in a group of countries
located in Europe, America, and Asia, which have come a long way in the industrialization
process. As of 2022, the world s four largest stock exchanges in terms of trading volume can
be listed as follows. The New York Stock Exchange (NYSE) in the United States with a
valuation of $20 trillion and by a wide margin is in first place. The National Association of
Securities Dealers Automated Quotations (NASDAQ), an over-the-counter exchange
founded in 1971, is in second place. The third place is the London Stock Exchange, founded

7
in 1801 and the fourth largest stock exchange is the Far East, Tokyo Stock Exchange (TSE),
which was established in the 1870s.

Emerging stock markets are the stock markets of the developing BRICS countries
(Brazil, Russia, India, China, and South Africa). These countries are rich in natural resources
such as oil and natural gas and have high growth potential. A significant part of the basic
capital needs of developing countries is met from stock portfolio investments, that is, from
stock market activities. In other words, businesses, especially public joint stock companies,
meet their medium and long-term funding needs from the capital markets. Thus, stock
exchanges, which enable individuals to transfer their savings to companies as capital, also
prevent companies from turning to high-interest, high-risk, and short-term loans. In this case,
resources turn into investments and contribute to the national economy. In addition, the book
values and equities of companies listed on the stock exchange are increasing.

The main function of capital markets is to optimize resource distribution in the


economy by directing savings to rational investment areas and to balance income distribution
by spreading ownership of the means of production to the base. In order for capital markets
to fulfill this function, they must have an effective market functioning. This issue, depends
on the transparency and trustworthiness of the market and the protection of investors rights.
The mechanism to ensure this is the existence of an institution that will effectively supervise
the functioning of capital market 23).

The chart below, Figure 1.1, shows the performance comparison of selected world
stock markets between 2022/03 and 2021/12. In this chart, it is seen that the stock markets
of Brazil, Indonesia, Mexico, and Luxembourg have been increased and the biggest increase
was experienced in ISE100 at a rate of %20.

On the other hand, it is observed that the performance of the stock markets such as
NASDAQ, NYSE, Hongkong, Shanghai, and Deutsche, which are among the few world
stock exchanges with high trading volumes, dropped quite high in the range of %1-%16.
This situation can also be interpreted as the reflections of the Covid-19 pandemic
experienced all over the world and the effects of the pandemic on the world economies,
which are already struggling with economic crises, to the capital markets.

8
Athens Exchange

Hong Kong Exchanges


Santiago SE

Mexican Exchange

Tel Aviv SE

Philippine SE
Lima SE

Luxembourg SE

Thailand SE

Warsaw SE

Shanghai SE
Budapest SE

Irish SE

Colombo SE
Indonesia SE

Bursa Malaysia

Egyptian Exchange

Shenzhen SE
BSE India

LSE Group Borsa Italiana


Nasdaq
Borsa Istanbul

TMX Group

NYSE
Buse Merval

Source: https://www.world-exchanges.org/our-work/statistics /WFE.


Figure 1.1. Stock market performances (2022/03-2021/12)

Although stock market indices are a basic indicator for national economies, it is
necessary to know what the factors affecting the index are before making an evaluation.
Internal factors such as the economic data of the countries, unemployment figures, interest
rates, geopolitical position, and risks, and external factors such as global data and the
economic relations of the countries with each other, are the factors that affect the economic
course of the countries and, of course, the stock market index.

1.2. Istanbul Stock Exchange (ISE)

The markets where precious metals, financial instruments, and capital market
instruments are bought and sold in a reliable and audited environment based on laws and
certain legislation and prices are formed in this transparent environment, are called the Stock
Exchange. After the financial liberalization movements that started in Turkey in the 1980s,
the Istanbul Stock Exchange was established in 1986. Since its establishment, it has
developed continuously and has become one of the important stock exchanges in the regional
area. With the Capital Markets Law dated December 30, 2012, and numbered 6362, the
capital markets, the Istanbul Stock Exchange market, the Izmir Futures and Options
Exchange market, and the Istanbul Gold Exchange market were gathered under one roof
is law, Istanbul Stock Exchange was removed
from the status of a public institution as of April 3, 2013, and continued its activities as an
autonomous and independent company ( , n.d.).

9
According to the statistics from the Central Securities Depository system show that
in July 2022, there were 28.941.853 investors who transacted a total of 20.386.350 million
TL in market value. A total of 449 companies are listed on the Istanbul Stock Exchange.
Index created to measure the sectoral and overall performances of the prices and returns of
these stocks are also alternative investment tools for the portfolio of investors. Therefore,
individual and corporate portfolio managers, who carry out their investments with an index-
linked portfolio strategy, tend to create their mutual funds according to the returns of the ISE
index.

As of June, there were 5.351.006 investors with a portfolio of 13.411.636.000.000


registered in the Central Securities Depository (CSD) system. The chart below, Figure 1.2,
illustrates how the number of investors and portfolio size has changed over the years. It is
seen that there has been a rapid increase in the number of investors and portfolio value after
2020. It can be observed that a record surge brought it to the 5 million mark in 2021. The 52
firms that went public and their shares strong performance in the initial trading days are
believed to be the leading causes of this.

Source: https://www.mkk.com.tr/en/trade-repository-services/e-veri/fund-management-fees

Figure 1.2. Number of investors and portfolio value in CSD

The market capitalization values of the companies registered in the Istanbul stock
market by year are given in Figure 1.3. Market capitalization is the value of a company as
determined by the stock market. It is observed that the market value of companies, which
had a continuous upward trend from 1997 to 2021, started to decrease significantly after
2012. It is observed that the value of all stocks in actual circulation, which experienced

10
decreases and increases between 2013 and 2019, increased again in 2020. Due to the Covid-
19 pandemic experienced all over the world in 2020, there were significant reflections on
the Turkish market as well as in the world markets, and there were decreases in the capital
markets. It has been observed that the reflections of the Covid-19 pandemic on the Turkish
stock market as a decrease with the data between 2020-2022.

315.198
307.052
286.572

233.997 237.474
219.763 227.512
197.074
195.746 184.966
161.538 157.702
149.264 140.207
162.399 188.862
112.716
98.299 118.329 147.876
61.095 69.659 68.379
47.150
33.646
34.217

Source: https://cmb.gov.tr

Figure 1.3. Market capitalization of stock exchange companies (million dollars)

1.3. Purpose of the Research

AI, which became an industry in 1980, emerged as a science in 1987 and started to
take over every aspect of our lives with the developments in computer technologies in
today s world. It is widely and effectively used in many fields, from health to the military
and defense industry, industrial applications, and even the construction industry to
archeology. In the field of finance, AI applications have become more popular in recent years
due to their increasing accuracy, speed, and data size.

The aim of this research is to predict the long-term relationship between the ISE100
index value and the selected macroeconomic variables in the January 2001-January 2022
period by using Logit and Probit analysis and the deep neural networks method and to
compare the estimation performances of the results obtained for both methods. Thus, the
effectiveness of AI technologies, which emerged from developments in traditional

11
econometric methods and computer technologies and continued to develop rapidly in recent
years, in time series analysis will also be measured. For the models used in the research, the
most commonly used variables in studies on the market index in the literature were selected
(W. Chen et al., 2017; Cheng et al., 2022a; Donaldson & Kamstra, 1999; Kara et al., 2011;
Staub et al., 2015) and the meaningful variables obtained from the created meaningful
models were given as input to the models created by ANNs and DL methods, and the
effectiveness of the modeling methods was tested.

1.4. Contribution of the Research

Especially after the global economic crisis of 2008, researchers, policymakers, and
investors have realized how important financial forecasts are in preventing financial crises
and uncertainties. Therefore, economic decision makers needed modeling in order to develop
policies that would minimize future uncertainties and create a portfolio.

Although there are many empirical studies in the literature examining the relationship
between stock market indexes and macroeconomic factors (Henrique et al., 2019, pp. 226
251), the study estimating whether macroeconomic variables have any long-term effects on
stock market index values using the Logit and Probit regression methods and the deep neural
network method is very superficial and limited in number (Aggarwal, 2020; Davis & Karim,
2008; Kantar & Akkaya, 2018).

In the empirical study, both the degree of interaction between the real sector and the
financial sector in Turkey will be measured, and a comparison of the DL method, which is
a ML method from AI technologies and a new estimation method, with qualitative response
of the Logit and Probit regression models from econometric methods will be made. The
results of econometric and AI models created for this purpose are presented comparatively.
Thus, a decision-making system that can be used to predict the short-term movement, trend,
and price of the stock market has been developed and evaluated. It also offers a survey of
the literature on the use of AI in financial analysis, with an emphasis mostly on the modeling
process.

12
1.5. Definition and Limits of the Problem

In this research, the effectiveness of the DL modeling technique has been tested. The
DL modeling technique is a ML method that has become increasingly important in our
country and around the world in recent years and has been widely used in many fields but is
used very limitedly in financial time series estimations for Turkey.

Qualitative regression models were created by using the monthly values of seven
macroeconomic variables in order to show the relationship between ISE100 index values
and macroeconomic variables in the 11-years (253-month) period between January 2001 and
January 2022 in Turkey. Then, using the obtained models, DL models were created using
the macroeconomic variables used.

1.6. Problem Solving

The first parts of the research will be formed as a result of theoretical and literature
research. The conceptual and theoretical framework of financial time series and ANNs and
DL methodology will be drawn from the information to be obtained from the Turkish and
international literature, such as articles, theses, books, papers, and reports published in these
sections.

In the application part, statistical analysis will be made using the ISE100 index and
macroeconomic variable data. Eviews12, STATA15, and MATLAB R2021b package
programs will be used in the application.

The data sets used in the analysis were compiled as monthly series from the Central
Bank of the Republic of Turkey Electronic Data Distribution System (EVDS), the Turkish
Statistical Institute (TUIK) statistics, the monthly statistical bulletins of the Capital Markets
Board, the statistics of the Ministry of Treasury and Finance, and the Eurostat databases.

The macroeconomic variables used in this study were chosen by considering the
studies by Chen, Roll, and Ross (1986) and Prantik and Vani (2004). The studies of Chen,
Roll and Ross (1986) show that macroeconomic variables such as long- and short-term

13
interest rates, expected and unexpected inflation rates, industrial production, and price
indices systematically affect securities prices and, thus, net asset values and returns. In
addition, studies conducted in recent years have shown that various macroeconomic
variables such as exchange rates, oil prices, gold prices, interest rates, production and price
indices, and money supply significantly affect stock market indices (Aggarwal, 2020;
Henrique et al., 2019; Kantar & Akkaya, 2018; Peng et al., 2021).

The ISE100 index was determined as the dependent variable in the established models,
and seven macroeconomic variables were determined as the independent variables. The
selected variables are the main variables that are most widely used in the literature and have
a direct impact on the pricing of capital market instruments and this effect has been proven
in the literature. In addition, these macroeconomic variables in the model are also used as
indicators (benchmarks or fund benchmarks) for the ISE100 index.

1.7. Organization of the Thesis

In the first part of the research, in addition to the introduction, a brief overview of the
world stock markets was made and information were given about the Istanbul Stock
Exchange. After briefly introducing the purpose, scope, methodology, and data set of the
research, in the second part, information is given about Logit and Probit models, which are
qualitative response regression models and econometric modeling techniques that constitute
the application part of the study.

From the third to the end of the seventh chapter, the concept of AI and its history, the
development of AI, AI technologies, and the latest DL method are covered. In these sections,
first of all, ANNs, which constitute the basic working system of the DL method, are
theoretically examined, including their definition, history, general structure, types, learning
algorithm, and learning rules of ANNs are given. Then, from the steps showing how to
design an ANN model, information is provided on how to determine the input, output, and
hidden layers of the network and the number of neurons in these layers, how to enter data,
how to determine the parameters of the network, or, in other words, how the structure of the
network can be created. In addition, the application areas, advantages, and disadvantages of
ANNs are also given in this section. In the seventh chapter, which discusses the DL method,

14
which is based on ANNs, the elements that make up the architecture of the network and the
methods used in learning are given wide coverage.

In the eighth chapter, detailed and current literature studies are given, and the
references to the current thesis study are examined. Thus, a wide literature review, including
financial studies using AI technologies, is included. Especially within the scope of the thesis,
financial time series and AI techniques were analyzed, and the methods to be used in the
thesis were determined.

In the last chapters, after the methodology and data set of the study were introduced,
the application of analysis techniques was started. By establishing Borsa Istanbul Index
prediction models, Logit and Probit regression analyses were made, and a DL design is made
using the variables used in these prediction models, Logit and Probit regression analyses are
made, and a DL design is made using the variables used in these models. The effectiveness
of the techniques was tested with the findings obtained from the prediction models created
using analysis techniques. A performance evaluation is carried out within the framework of
the purpose of the study, and the prediction results of the models used are presented
comparatively.

As a general evaluation, in this thesis study, the ability of the DL method to forecast
the future trend of the nonlinear, variable, and complex ISE100 index will be tested. The DL
method is one of the AI technologies that has become increasingly important in recent years
all over the world and has been widely used in many fields but is seen to be used very
limitedly in our country in financial time series estimations. It has also been observed that
DL methods are used to a limited extent or are not used at all in the field of finance, and so
this fact brings forward to new approaches in the field of finance.

15
2. QUALITATIVE RESPONSE MODELS

In regression models, it is assumed that the dependent variable, or explained variable,


Y, is quantitative and, in contrast, the explanatory variables are quantitative and qualitative
variables, or a mixture of them. On the other hand, qualitative response models are models
in which the dependent variable is a qualitative variable and takes only two values. Three
regression models developed for the two-valued response variable are briefly explained
below.

2.1. Linear Probability Model (LPM)

Linear probability models are simple models that consist of two-valued dependent
variables and can be estimated by the ordinary least squares method. It indicates the
probability of choosing one of the two options presented. Since the dependent variable takes
binary values 0 or 1, it is a restricted qualitative variable defined as a probability.

Figure 2.1. Linear probability model

The model expressed as a linear regression model is a typical linear regression model,
and it is called the Linear Probability Model since the independent (explanatory) variable
has two choices. Because when is given, the conditional expected value of can be
interpreted as E ( ) conditional probability of the occurrence of the event when is
given. Here, when =1, the first option is preferred and when =0, the second option is
preferred, so the dependent variable will be able to take the values of 0 ve 1. Assuming E( )
= 0 when the expected value is taken, the obtained model will be as follows:

Figure 2.2. Conditional probabililty model

16
In this model, the probability of choosing the first option will be =P ( =1) and if
the second option is preferred, =P ( =0). In this case, the expected value of the
dependent variable is (Gujarati & Porter, 2009, p. 543);

Figure 2.3. The expected value of the dependent variable

Figure 2.4. The conditional expected value of the dependent variable

In classical regression models, the parameters are found by the least squares method.
However, in the Linear Probability Model (LPM), like the dependent variable, the error term
also fits the Bernoulli distribution as it only takes a binary value of 0-1. Therefore, although
the model parameters are estimated by the least squares method, they do not meet the linear
deviation-free estimator specifications. Therefore, there are some criticisms regarding the
estimation and interpretation of this model. Because there is a problem of varying variance
in error terms, parameter estimates will not be efficient, and this will make traditional
significance tests questionable. For these reasons, the logit and probit models discussed in
the literature are preferred as alternative modeling options instead of linear probability
models to model binary variables (Gujarati, 2011, p. 154).

2.2. Logit Model

Although the linear probability model is simple to forecast and use, it is thought that
the limitations arising from its disadvantages, such as the fact that the appropriate
probabilities can be less than or greater than zero and the partial effect of any explanatory
variable being fixed, can be overcome by using more complex binary response models
(Wooldridge, 2010, p. 584).

When the linear probability model described in the previous section is expressed as
and (cumulative) logistic distribution function , is the i probability of
the second individual making one of the two choices;

17
Figure 2.5. Logistic distribution function

Here, e = 2.7182 while the variable takes values in the range of - to + , since
takes values between 0 and 1, the relationship between (or and will not be linear
and will not meet the two requirements mentioned earlier. For this reason, since this
relationship will need to be converted to a linear form to be estimated can be written as
;

Figure 2.6. Probability of failure

Since the probability of an event occurring is , the probability of failure will be


in this case will be;

Figure 2.7. Odds ratio

This ratio is the probability of an event occurring to the probability of not occurring
and is expressed as the Odds Ratio when the natural logarithm of this ratio is taken.

Figure 2.8. Logistic regression model

The resulting model is called the Logit regression model. So the logarithm L of the
odds ratio is linear not only with respect to X but also with respect to the population
coefficients (in terms of coefficient estimation). L is called Logit. The probability of making
a particular choice while estimating the parameters of this model will be = ln (1/0), if not
= ln (0/1). These expressions are obviously meaningless, so Logit models cannot be

18
estimated using the standard least squares method. In this case, the highest likelihood method
can be used to estimate the population parameters (Gujarati & Porter, 2009, p. 556).

In Logit models, according to the number of independent variables, each slope


coefficient is a partial slope coefficient and measures the change in the estimated logit when
there is a one-unit change in the value of a particular explanatory variable (other variables
are constant). A more meaningful interpretation of the slope coefficients is in terms of the
odds obtained by taking the inverse logarithm of the slope coefficients (Gujarati & Porter,
2009, p. 563). Again, in binary models whose dependent variable is 0 and 1, the traditional
measure of goodness of fit , does not matter, what is important is the expected signs of
the regression coefficients and their statistical significance (Gujarati, 2011, p. 158).

2.3. Probit Model

As mentioned earlier, an appropriately chosen cumulative distribution function


should be used to explain the behavior of a binomial dependent variable. The error term of
the linear probability model exhibits a nonlinear distribution, while the error term of the
Logistic model exhibits a logistic distribution. The Probit model error term shows a normal
distribution. Therefore, while a cumulative logistic distribution function is used in the Logit
model, as in Figure 2.5, the normal cumulative distribution function is used in the Probit
model, which is based on the utility theory developed by McFadden and is sometimes called
the Normit Model. In the Probit model the i individual s decision to make or not make a
particular choice depends on an unobservable utility index (also called a latent variable)
and is determined by one or more explanatory variables such as , the higher the value of
the index, the more likely the individual will make a choice. This index is expressed as
follows (Gujarati, 2011, p. 161; Gujarati & Porter, 2009, pp. 566 567):

Figure 2.9. Utility index

What is the relationship of this unobservable index to the individual s decision to


make a choice? Since Y = 1 if the individual makes a choice, and Y = 0 if he does not,
assuming that there is a threshold value called , for this index for each individual, if
19
exceeds this threshold value, the individual makes the choice, if not, he does not. The
threshold value cannot be observed like the utility index, but if we assume that they have the
same mean and variance, not only predicts population coefficients given in Figure 2.9, but
also obtains information about the unobservable index itself.

Under the assumption of normality, the probability that and are less than or
equal to the standardized normal cumulative probability distribution is expressed and
calculated as follows:

Figure 2.10. The probability that the threshold value is less than or equal to the utility
index

Here, is the standardized normal variable that is, Z N (0, ). F , being the
standard normal cumulative distribution function, it can be written as:

Figure 2.11. The standard normal cumulative distribution function

Similar to the Logit model, it is not appropriate to use the classical least squares
method for the estimation of the Probit model. With the highest likelihood method, it is
possible to obtain consistent estimates of the parameters of the Probit model (Gujarati &
Porter, 2009, p. 567).

Although the Logit and Probit models give similar results, the main difference
between the two models is that the logit distribution has slightly thicker tails. A logistically
distributed random variable has a variance of about /3 hereas a standard normally
distributed variable has a variance of 1. However, in practice, there is no reason to prefer
one over the other, and the Logit model is preferred over the Probit model because of its
comparative mathematical simplicity (Gujarati, 2011, p. 163). The response probabilities of

20
these models are complex and difficult to forecast due to their nonlinear nature, but they
have become popular recently (Wooldridge, 2010, p. 596).

Logit

Source:(Gujarati & Porter, 2009, p. 572).


Figure 2.12. The graphs of the Logit and Probit models

While natural logarithms of odds ratios are used in Logistic regression analysis, the
cumulative normal distribution is used in Probit regression analysis.

21
3. ARTIFICIAL INTELLIGENCE

The work of human beings for self-knowledge has made significant progress with
technological developments. At the latest point reached today, an AI chat robot called
ChatGPT has been put into use by a research company called OpenAI. In the modern world,
where computers and computer systems have become an indispensable part of life, major
developments have occurred in software and hardware systems and data science. These
developments have led to the widespread use of AI technologies, which have the ability to
process information, learn, reason, solve problems, and make decisions quickly, accurately,
and effectively in many areas where more information is obtained every day and the obtained
information is used in many areas where computer technologies are used.

Studies of ANNs in neurobiology were drawn to the field of engineering with the
creation of the first artificial nerve cell in 1943 by neurologist Warren McCulloch and
mathematician Walter Pitts (Haykin, 1999, p. 38). The concept of AI has also taken its place
in the literature after this study. The concept of AI, whose ability to solve problems using
the ability to think and learn, which is one of the most important features of the human brain,
has been discovered, has led to the emergence of new and more complex technologies due
to the intense interest it has received from both researchers and commercial vendors.

A neural network is a computer program or wired machine designed to learn in a


manner similar to the human brain. AI, on the other hand, is a combination of neural
networks and neural network theory that was born from AI research or from research on
designing machines with cognitive ability (Kutsurelis, 1998, p. 2). The AI research started
with the Turing Test, which is the first software example that the British mathematician AM
Turing, who managed to crack the Enigma code used by the German army for
communication in the Second World War, explained in his article Computing Machinery
and Intelligence published in the philosophy journal Mind in 1950. According to the test,
which is known by its name today, if a person talking to a computer does not understand
whether the person they are talking to is a computer or a human, that computer should be
considered intelligent (Turing, 1950, pp. 433 460). Turing s work deeply influenced AI
studies, and the modern age of AI began. After this developments, AI and ML have turned
into studies aimed at ensuring that machines have the ability to perform complex tasks. The

22
first attempt at learning neural networks was with the Hebbian learning rule, a
neurophysiological theory developed by Donald Hebb in 1949. According to the theory,
learning occurs through the exchange of messages between neurons in the human brain. In
other words, he suggested that learning ability can be assigned to network structures by
performing logical operations through networks of nerve cells (Haykin, 1999, p. 38).

Neural networks (NN), or more precisely ANN, are a technology with roots in a
variety of fields, including neurological sciences, mathematics, statistics, physics, finance,
computer science, and engineering (Haykin, 1999, p. 1). These computer systems, called
intelligent systems, have continued their development with the contribution of AI. The
developments in these two areas trigger each other, and their development processes
continue.

ANNs are neural networks, also known as simulated neural networks It is a subset
of ML and is at the heart of DL algorithms. These systems were inspired by the human brain,
and they took their names and formed their structures by mimicking the way biological nerve
cells (neurons) send signals to each other. Artificial neural networks (ANNs), Machine
Learning (ML), and Deep Learning (DL) are called artificial intelligence (AI) (Quazi, 2022).
These three concepts are shown in the literature in the Figure 3.1 below.

Figure 3.1. AI, ML, ANN and DL

AI and machine learning (ML) are two emerging and interrelated technologies in
finance. Concepts such as AI, ML, DL, and data science are confusing and intertwined.
ANNs, a sub-branch of AI, are a statistical approach created to develop forecasting models.
23
Basically, ML is a particular embodiment of AI that develops techniques and uses algorithms
that enable machines to recognize patterns in datasets. On the other hand, DL is a subset of
ML that equips machines with the techniques needed to solve complex problems. Data
science is a separate field of study that applies AI, ML, and DL to solve complex problems
15).

AI, defined as understanding the human thinking structure and attempting to develop
computer operations that will reveal the likes of it, is an attempt to give programmed
computers the ability to think. AI studies, which have been going on since the 1950s for the
development of systems that think and act like human beings, have spread to fields such as
engineering, neurology, and psychology since they were aimed at imitating humans at one
point. Despite significant advances, the point reached today in studies for developing
systems that can think and behave like humans is that AI has not yet been fully developed.
Leaving aside the discussions on the possibility of AI, studies on this subject continue along
with studies in different fields that support this field.

ANNs technology is one of the various fields that emerged within the scope of AI
studies and that at some point provided support to those studies. Therefore, ANNs, which is
a sub-branch of AI, forms the basis of systems that can learn. ANNs that imitate the neuron,
which is the basic processing element of the human brain, in a formal and functional way,
are programs created for a simple simulation of the biological nervous system in this way.
Therefore, AI technology, which is thought to be able to transfer the ability of learning by
experimenting (experience) to the computer environment, provides an incredible learning
from input data capacity to a computer system and offers many advantages. This
technology, which offers various advantages and develops day by day, is used in the fields
of economy and statistics, as it is in many other fields today. They are frequently used in
areas that require the definition of the structure contained in the data, such as forecasting and
prediction, because they are known as Universal Function Approximators

Technically, the most basic task of an ANN is to determine an output set that can
correspond to an input set shown to it. In order to do this, the network is trained with
examples of the relevant event and gains the ability to generalize. ANNs are also called
correlated networks, parallel distributed networks, and neuromorphic systems
2012, p. 30).

24
4. DEVELOPMENT PROCESS OF ARTIFICIAL
INTELLIGENCE

The history of AI begins with an interest in neurobiology and the application of the
obtained information to computer science. When AI studies are examined, it is discovered
that the majority of the research is a continuation of each other. In other words, the oldest
developed AI structures and learning algorithms are still used today, and developments in
this field are strongly related to previous developments. Due to the succession of these
studies, AI studies developed rapidly, but some studies conducted later showed that previous
studies were insufficient, which caused the studies in the field of AI and the support given
to these studies to pause until these problems were resolved mel, 2006, p. 37).

For many years, scientists have studied how the human brain works and its functions.
The first work providing information on brain functions was published in 1890. It is known
that before 1940, some scientists such as Helmholtz, Pavlov, and Poincare worked on the
concept of an ANN. However, it cannot be said that these studies have engineering value
. The first foundations of ANNs were laid in an article published in
1943 by neuropsychologist McCullogh and mathematician Pitts, who started research in the
early 1940s. McCullog and Pitts made a model in the form of a simple mathematical
representation of the biological neuron (Zurada, 1992, p. 30). Thus, inspired by the
computational ability of the human brain, the first neuron model or simple neural network
was modeled, and the first mathematical model of the artificial nervous system was
developed. They tried to determine the learning rules by revealing that cells work in parallel
with each other (McCulloch & Pitts, 1943, p. 118).

In 1949, Donald Hebb, a psychologist at McGill University, explained a learning rule


that improves the weight values of an artificial neural network consisting of artificial cells
in his book called The Organization of Behavior (Haykin, 1999, p. 38). This rule, called
the Hebb Learning Rule, still forms the basis of many learning rules today. The concept of
generating adaptive responses with random networks was introduced by Farely and Clark in
1954 (Anderson & McNeill, 1992, p. 12). This concept was developed by Rosenblatt in 1958
and by Caianiello in 1961. The perceptron model, especially developed by Rosenblatt, was

25
developed later and formed the basis of multi-layer perceptrons, which were revolutionary
in ANNs (Zurada, 1992, p. 19).

With the development of computers in the 1950s, it became possible to model the
foundations of theories about human thought. Nathanial Rochester, a researcher at IBM
research labs, led efforts to create a neural network simulation. Although the first attempt
was unsuccessful, subsequent attempts were successful. After this stage, the search for
traditional computational methods has left its place for the investigation of neural
computation methods. Until then, Rochester and his team had defended thinking machines
by citing evidence of their own work. The Artificial Intelligence Dartmouth Summer
Research Project , which was held in 1956 and later on, increased the talk and therefore the
evidence on AI and ANNs. One of the results of this research is that it encourages interest
in both AI and ANNs research (Anderson & McNeill, 1992, p. 17).

In 1957, Frank Rosenblatt, a neurobiologist at Cornell University, developed the


Simple Perceptron Model (Perceptron), which will have an important role in both the
development and pause of artificial neural network studies. Rosenblatt was interested in the
eye processes of a fly. Inspired by the fact that most of the processes that tell a fly to escape
are done inside the fly s eye instead of its brain, a hardware device called the perceptron was
established and the network structure developed, which is the oldest artificial neural network
model of today (Anderson & McNeill, 1992, p. 17). Perceptron, developed by Rosenblatt,
pioneered an important development in the history of ANNs. Because this model forms the
basis of Multilayer Perceptrons that will be developed later and will be revolutionary in
ANNs, this simple perceptron model is still used today (Zurada, 1992, p. 19).

In 1959, Bernard Widrow from Stanford University and his student Marcian Hoff
developed two Perceptron-like models, which they named ADALINE (Adaptive Linear
Elements) and MADALINE (Multiple Adaptive Linear Elements 8).

The work done by Marvin Minsky and Seymour Papert in 1969 turned into a book
called Perceptrons . The authors specifically claimed in this book that sensors based on
ANNs have no scientific value and cannot solve nonlinear problems, and they used the

26
example of not solving the XOR1 logic problem to prove their point. This situation has
caused AI studies to enter a period of stagnation. Although the studies were interrupted in
1969 and the necessary financial support was cut, some scientists continued their studies. In
particular, the studies of researchers such as Shun-Ichi Amari, Stephen Grossberg, Gail A.
Carpenter, Teuvo Kohonen, and James A. Anderson came to fruition in the 1980s, and new
studies on AI began to be put forward. Adaptive Resonance Theory (ART), developed by
Grossberg and Carpenter in 1976, has been the most complex artificial neural network
developed for teacherless learning .

Studies by Hopfield in 1982 and 1984 showed that ANNs can be generalized and can
produce solutions to problems that are difficult to solve, especially with traditional computer
programming. Hopfield used similar studies created in 1972 by electrical engineer Kohonen
and neuropsychologist Anderson, working in different disciplines and unaware of each other.
In order to solve technical problems such as optimization, he developed the nonlinear
dynamic Hopfield network and the Kohonen Self-Organizing Feature Maps
network, which is a learning network without a trainer (Akel & Karacameydan, 2018, p. 53).
Hopfield also revealed the mathematical foundations of ANNs in these studies (Anderson &
McNeill, 1992, p. 18). These studies formed the basis for unsupervised learning rules that
will be developed later.

In 1986, David Rumelhart and James McClelland, in their work called Parallel
Distributed Processing , showed that the previously claimed flaws on this subject could be
overcome by developing a new learning model, the Back Propagation Algorithm, which is a
learning model in feedforward models. Because the XOR problem, which the single-layer
perceptron could not solve, was solved by the discovery of multi-layer perceptrons. Today,
various versions of this learning method are used in a variety of settings
38). The development of this algorithm, which is still one of the most used training
algorithms, has revolutionized the field of ANNs (Zurada, 1992, p. 20). These developments
have led to an increased interest in ANN.

(Executive OR) a logic problem can be defined as a special case of the problem of classification of points in a unit
hypercube. Each point in the hypercube belongs to either class 1 or 0. For detailed information, see (Zurada, 1992).

27
As a result of the increasing interest in AI, various conferences were held on this
sub -Japan Joint Conference on

participants in Kyoto. During the conference, the Japanese presented their fifth-generation
research hereupon, there was a fear among Americans that they are falling behind Japan.
Because of this fear, funds were quickly transferred to AI research in the United States, and
AI was developed on a regular basis (Anderson & McNeill, 1992, p. 18).

By 1985, the American Institute of Physics began to organize annual meetings on


(Anderson & McNeill, 1992, p. 18). With the 1st
International Conference on ANNs (1st International Conference on Neural Networks) held
by the Institute of Electrical and Electronics Engineers (IEEE) in 1987, studies in the field
of ANNs started to become widespread all over the world.

In 1988, as an alternative to multilayer sensors, Broomhead and Lowe developed the


Radial Basis Functions (RBF) model and produced very successful results, especially for
filtering problems. Later, Specht developed Parabolistic Networks (PNN) and General
Regression Networks (GRNN), which are more advanced forms of these networks. Since
1987, ANNs have been discussed in different symposiums and conferences, and new
learning techniques and models have been put forward. Today, with the decrease in the size
of computers and the increase in their capacities, ANNs have ceased to be theoretical and
laboratory studies and have begun to create systems used in daily life and to be useful to
people practically .

The widespread use of ANNs in the field of finance is occurring in parallel with the
development of the backpropagation algorithm by Rumelhart and McClelland. Multi-layered
networks are needed in the solution of financial problems since financial problems include
a linear as well as a non-linear structure. Although there are multi-layered networks that have
been developed before, these networks have not been trained. However, with this algorithm
developed by Rumelhart and McClelland, this problem has been overcome, and ANNs have
attracted great interest in the field of finance. Today, ANNs are effectively applied in
financial forecastings, such as for financial indices, stock prices, portfolio diversification,
bond valuation, loan repayment rates, real estate prices, and the bankruptcy of enterprises.

28
In summary, AI was used in customer orders by DEC (Digital Equipment
Corporation), an American computer and technology company, and when it was seen that
40 million dollars were saved, the interest in AI increased again, and countries started to
make serious investments in this field. In 1975, Holland produced the genetic algorithm
based on the principle of natural development with living things, and in 1976 he developed
the Sejnowski Boltzman Machine and applied the backpropagation algorithm to it. In 1978,
Sutton and Barto developed a reinforcement learning model, and in 1982, hundreds of
studies in this field were presented at the international ANN conference on PC-based neuro,
with Hecht-Nielsen s TRW MARK III, the first modern electronic neuro computer (Kubat,
2019, p. 621).

Martens (2010) and Martens and Sutskever (2011) developed powerful deep and
recurrent neural network (RNN) models that employ momentum stochastic gradient descent
in datasets with well-designed long-term dependencies (Long Short Term Memory, LSTM)
Hessian-Free found that it showed higher training performance than the performance
obtained from optimization (Sutskever et al., n.d., pp. 1 9). AI is a structure that is directly
affected by these disciplines, which affect many disciplines such as computer science,
engineering, biology, psychology, mathematics, logic, philosophy, business, finance, and
linguistics. For this reason, its applications are seen in many fields, from medicine to the
military, from economics to meteorology. Of course, its use in so many different fields
appears impressive, but it has found such a wide range of applications as a result of years of
research from the 1940s to the present r, 2021, p. 3).

29
Table 4.1. Some important advantages in neural network research

Researcher Contribution

McCulloch & Pitts ANN models with adjustible weights (1943)


Rosenblatt The Perceptron Learning Algorithm (1957)
Widrow & Hoff Adaline (1960), Madaline Rule I (1961) & Madaline Rule II (1988)
Minsky & Papert The XOR Problem (1969)
Werbos (Doctoral Backpropagation (1974)
Dissertation)
Hopfield Hopfield Networks (1982)
Rumelhart, Hinton & Renewed interest in backpropagation: multilayer adaptive
Williams backpropagation (1986)
Vapnik, Cortes Support Vector Networks (1995)
Hochreiter& Long Short Term Memory Networks (1997)
Schmidhuber
LeCunn et.al. Convolutional Neural Networks (1998)
Hinton & Ruslan Hierarchical Feature Learning in Deep Neural Networks (2006)

In the Table 4.1, the main developments in AI studies are summarized by year. After
the 2000s, advances in AI research accelerated. The increasing interest in this field has led
to the emergence of new AI techniques based on ANNs. One of them is DL technology.
However, in order to understand DL, it is necessary to understand the structure and workings
of ANNs. These issues will be discussed in detail in the following sections.

30
5. ARTIFICIAL INTELLIGENCE TECHNOLOGIES

AI systems, also called intelligent systems in the literature, are technology


applications developed to perform tasks similar to human cognitive abilities, although there
is no generally accepted definition of their concept (Alaloul & Qureshi, 2020, p. 1).
Developments in software and hardware systems have led to access to more information and
data and the efficient use of the obtained data, making the work of decision units easier. The
fact that problems that cannot be solved or that can be difficult to solve with human
intelligence can be solved and decisions can be made thanks to algorithms in code sequences
for many fields has increased interest in this field and led to the emergence of many new
technologies in the field of AI. In these systems, the ways and stages of people s decision-
making are imitated. The most prominent of these technologies today are summarized below.

5.1. Expert System

Computer programs are technologies that have been developed and can solve a
problem in the same way that experts in a certain field can solve based on their expertise. Its
main fields of application are medicine and biomedicine (Staub et al., 2015, p. 1478). Experts
use their knowledge and experience to solve problems. The computer should understand and
store this knowledge and experience . It is used in applications that
need both machine and human intervention. These systems consist of three parts: The Rule
Base, The Database, and The Rule Analyzer (Kubat, 2019, p. 625).

5.2. Machine Learning

ML is the field of computer science that creates algorithms to process large amounts
of data and learn from it. It can learn the relationships between the inputs and outputs of
events using examples. By interpreting the learned information and similar events, decisions
are made or problems are solved. In order to understand ML, it is necessary to understand
learning. Learning is the process of improving behaviors through the discovery of new
information over time -21). Since this is classical programming in AI
technology, linear or non-linear relationships cannot be established between the data and the
results, and rules cannot be created. Since this is classical programming in AI technology,

31
linear or non-linear relationships cannot be established between the data and the results, and
rules cannot be created. In this case, the solution cannot be found. For this reason, ML
algorithms, which have mostly been used in military and commercial fields in recent years,
are divided into three subgroups: supervised learning, unsupervised learning, and
reinforcement learning algorithms, due to their ability to establish a relationship between
data and results .

It contains classical algorithms for various tasks such as classification, regression,


and grouping in ML technology. Since it is a technology that trains algorithms on data, the
more data that is provided, the better the algorithm learns. In other words, ML is an
optimization algorithm.

5.3. Fuzzy Logic

Fuzzy logic, which was first used by L. Zadeh in 1965, is based on the combination
of the advantages of ANNs such as learning and decision-making. In addition, fuzzy logic-
neural networks (ANNs) are one of the most suitable methods for time series calculations
(Staub et al., 2015, p. 1478). Today, many events take place in uncertain conditions. For
events that are not known with certainty, experts use definitions such as normal high,
approximate and low Fuzzy sets are developed to represent and use non-statistical
uncertainties with data and information. Every problem has its own uncertainty. This method
tries to describe this uncertainty mathematically. Therefore, these systems are suitable for
uncertain or approximate inference and especially for systems whose model cannot be
expressed mathematically (Kubat, 2019, p. 623). They are technologies that make it easier
to process uncertain information and make decisions in situations where precise numbers
cannot be expressed .

5.4. Genetic Algorithms

The genetic algorithm process, which is likened to the natural evolution process,
includes operators used in natural evolution such as reproduction, crossover, and mutation.
It is a technology used in solving complex optimization problems. In order to solve a
problem, random initial solutions are determined, and better solutions are sought by

32
matching these solutions with each other. This search continues until the best result is
produced. In the technology of genetic algorithms, it is accepted that the features that will
produce the desired result in solving problems pass to new solutions obtained from initial
solutions, and from them to later solutions, through inheritance . In
other words, it is a method dependent on natural genetics and natural selection mechanisms
by coding parameter sets and using objective function and functioning information.
Mechanical learning, economics, social sciences, and information systems are application
areas (Staub et al., 2015, p. 1478).

5.5. Artificial Neural Networks

ANNs, or in other words, neural networks , are systems that examine the structure
of the human brain, consisting of neurons and learning methods. In the 19th century, the
studies of psychologists and neuropsychologists to understand the human brain formed the
basis of ANNs. The first scientific study on artificial neural networks started with McCulloch
and W. Pitts, as mentioned earlier.These systems try to fulfill the functions of the brain,
developed by taking into account the brain structure and the interaction of neurons with each
other (Kubat, 2019, p. 623). These systems also focus on learning. They learn the
relationships between events from examples, and then make decisions using the information
they have learned about examples they have never seen .

All of the AI technologies were developed to serve people in their daily lives and are
still developing rapidly. The basic structure of AI technologies is an ANN system, which is
created by simulating the working style of the system consisting of simple biological nerve
cells (neurons), and an ANN system created according to this structure also forms the basis
of all artificial intelligence studies. ANNs, which enable computers to learn, will learn to
operate by trial and error and use this knowledge in problem-solving.

In the literature, it is seen that ANNs are used in many fields, from medicine to
defense, from psychology to finance, due to their superior success in forecasting and
classification studies. ANNs, which are briefly introduced here, will be discussed in more
detail in a separate section, since DL, which forms the basis of our study, is a sub-branch of
ANNs and works according to this system.

33
5.6. Deep Learning

Deep learning is a complex ML algorithm modeled on the workings of the human


brain (Bengio et al., 2015, p. 8). DL describes algorithms developed through supervised and
unsupervised learning that analyze data with a logical structure similar to how a human
would draw conclusions from an event. Therefore, it has to use a layered algorithm called a
deep learning ANN to achieve the main goal. For this reason, the interest in this field has
increased even more, as it has been seen that the human brain gives better results than the
standard ML models inspired by the biological neural network, and the studies reach results
that were thought to be impossible (Fathi & Maleki Shoja, 2018b, pp. 229 230).

Since DL technology and the ANNs on which it is based are the main subjects of the
study, they will be explained in detail in the following sections.

34
6. ARTIFICIAL NEURAL NETWORK

6.1. Definitions of Neural Networks in the Literature

There is no single, agreed-upon definition for ANNs. Many definitions can be found
in broad and narrow contexts. In fact, some researchers argue that instead of giving a general
definition for ANN, ANN types should be defined within themselves. However, some of the
definitions in the literature related to ANN are as follows:

ANNs are logical software developed to imitate the working mechanism of the
human brain and to perform the basic functions of the brain such as learning, remembering,
and deriving new information through generalization (Yazici et al., 2007, p. 65).

ANNs can be defined as a computer program written for a mathematical formula that
will enable the parameters to be adapted with the help of a set of examples, in the shortest
and simplest way without going into technical details (Anderson & McNeill, 1992, p. 4).

The ANN is a system formed by connecting neurons in different ways, which are a
simple processing elements that, in essence, mimic the way the human brain works. Each
neuron receives signals from other neurons or from outside, combines them, transforms
them, and produces numerical results (G. Zhang et al., 1998, p. 37).

This model, inspired by the biological working system of the human brain, is an
information-processing model that describes processes and gives examples. ANN is a
network of highly connected and organized neurons in layers (G. P. Zhang, 2003, pp. 163
164).

According to the definition of an artificial neural network made by Kohonen, who is


well-known in the artificial neural network literature, ANNs are a hierarchical organization
of many simple elements connected in parallel, interacting with real-world objects in a way
similar to the biological nervous system (Kohonen, 2001, pp. 71 72).

35
An ANN is a parallel and distributed single- or multi-layered computing system
consisting of many simple processing elements interconnected with one-way signal channels
in a weighted form (Kamruzzaman et al., 2006a, p. 3).

A more comprehensive and generally accepted definition made by Haykin is: A


neural network is a densely parallel distributed processor composed of simple processing
units that has a natural tendency to accumulate experiential information and enable it to be
used (Haykin, 1999, pp. 1 2).

This processor is similar to the brain in two ways:

Knowledge is acquired by the network from the environment through a


learning process.
Interneuron connection strengths, also known as synaptic weights, are used
to accumulate the obtained information.

6.2. Biological Nervous System

The human nervous system has a very complex structure. The brain is the central
element of this system, and it is estimated that there are about neurons (nerve cells)
connected to each other by subnets, and they have more than 6x connections. The
human nervous system has a very complex network structure. Nerve cells are specialized
cells that carry information through an electrochemical process. These nerve cells come in
different shapes and sizes. Some are only 4 microns (4/1000 millimeters) wide, while others
are 100 microns wide. Although different types of nerve cells have differences in terms of
shape and function, as shown in Figure 6.1, they all consist of 4 different regions: the
dendrite, the nucleus and soma or cell body, the axon, and the junction or synapse (Anderson
& McNeill, 1992, p. 3).

Synapses can be viewed as connections between nerve cells. These are not physical
connections, but rather gaps that allow electrical signals to pass from one cell to another.
These signals go to the soma, which processes them. The nerve cell creates an electrical
signal and sends it to the dendrites via the axon . Dendrites form the

36
input channels for a cell. Dendrites convert these signals into small electric currents and
transmit them to the cell body. The cell body processes the signals coming through the
dendrites and converts them into output. These outputs produced by the cell body are sent
via axons to be input to other neurons (Anderson & McNeill, 1992, p. 4).

Source: hhttps://en.wikipedia.org/wiki/Dendritetml

Figure 6.1. Structure of a typical neuron

In a neuron, the dendrite receives input signals, the soma processes input signals, the
axon converts input signals into output signals, and finally, the synapses provide
electrochemical contact between neurons. In the simulation of this structure adapted to
ANNs, each process element is named as follows: Synapses represent weights; dendrites
represent the summation function, the soma activation function, and the axon output.

6.3. Artificial Neural Networks Elements

An ANN consists of a large number of simple processing units called neurons, units,
cells, nodes, and process elements. Just as biological neural networks have nerve cells,
artificial neural networks also have artificial nerve cells. ANNs are a program designed to
simulate the way the simple biological nervous system works. Artificial neural networks
contain simulated nerve cells, and these cells connect to each other in different ways to form
the network. Figure 6.2 presents the structure of a general artificial neuron (Kamruzzaman
et al., 2006a, p. 3).

37
In the artificial neuron in Figure 6.2, the inputs are given by and i=0, 1, n;
and the weights are shown with mathematical symbols as ve . Each of these
input values is multiplied by a link weight. In the simplest form, these products are summed
up and sent to a transfer function, where the result is produced. This result is then converted
to output (Y). Here, the input vector is shown as {X1, X2, Xn} and the appropriate weights
vector is shown as {Wj1, Wj2 Wjn}.

Inputs
Sum Output
Activation Function

Weights

Source: Haykin, 1999, p.11.

Figure 6.2. General structure of artificial neuron

ANNs, inspired by the biological nervous system, have the capacity to learn,
memorize, and reveal the relationship between data. Just as biological nervous systems are
made up of nerve cells, ANNs are made up of artificial nerve cells. An artificial nerve is a
simple structure that mimics the basic functions of a biological nerve. Neurons in the
network take one or more inputs and give a single output. This output can be output to the
outside of the artificial neural network, or it can be used as an input to other neurons.
Artificial nerve cells are called process elements (neurons) in engineering science. Although
there are differences in the cell models developed, in general, each processing element
consists of 5 basic components: inputs (X), weights (W), summation function , activation
function and cell output ). These components can be explained as follows (Haykin,
1999, pp. 11 51):

6.3.1. Inputs

It is the information coming to an artificial nerve cell from the outside world or other
cells. These inputs are determined by the examples the network is asked to learn. The

38
artificial neuron can receive information from the outside world, as well as from other cells
or from itself. A neuron can usually receive many simultaneous inputs. Each input has its
own relative weight.

6.3.2. Weights

Inputs show the information coming to the artificial neuron, while weights show the
importance of the information coming to the artificial neuron and its effect on the cell. The
weight 1 (w1), in Figure 6.2 shows the effect ot input 1 (x1) on the cell. The fact that the
weights are large or small does not mean that they are important or unimportant. A weight
value of zero may be the most important event for that network. Weights can have positive
or negative values. A weight of value plus or minus indicates whether the effect is positive
or negative, while a zero indicates that there is no effect. Weights can have variable or fixed
values. All connections that provide the transmission of inputs between neurons in the
artificial neural network have different weight values. Thus, the weights act on each input of
each neuron. They are connected to a cell with information, inputs, and weights received
from the external environment. The value of the net input is calculated by using the activation
function by calculating with different methods and the net output value is obtained. The
important thing here is that the error between the actual values and the output values is the
least.

6.3.3. Summation Function

The summation function is a function that calculates the net input to a cell. Although
different functions are used for this, the most common one is to find the weighted sum. Here,
each incoming input value is multiplied by its own weight. Thus, the net incoming input is
found. However, in some cases, the addition function may not be such a simple operation.
Instead, it can be much more complex, such as minimum (min), maximum (max), mode,
product, majority, or a few normalization functions. The algorithm that will combine the
inputs is usually also determined by the chosen network architecture. These functions can
generate values differently, and these values are forwarded. The total input of any neuron is
equal to the weighted sum of the values from other neurons and the threshold value
(Kamruzzaman et al., 2006a, p. 3).

39
Figure 6.3. Summation function

The summation function is written as follows:

NET : Network, the net input of the neuron,


: the weight of the connection between neurons j and i,
: the output of the neuron i,
: b: represents the threshold value (bias)

Table 6.1. Summation functions in ANNs

Summation Equation Explain


Function

Weighted Total NET = Inputs and weight values are multiplied.

Multiplication Inputs and weight values are multiplied.

Minimum NET = Min ( ) The lowest calculated value is taken.

Maximum NET = Max ( ) The highest calculated value is taken.

Incremental NET = NET (old)+ Weighted total is calculated.


Total

Source

In Table 6.1, examples of different addition and combination functions used in


research in the literature are given. Usually, the best addition function is determined by trial
and error. Besides, determining the aggregation function completely depends on the
designer s own foresight.

40
6.3.4. Activation Function

It is the function that determines the output that the cell will produce in response to
this input by processing the net input obtained by the summation function. The purpose of
the activation function is to allow the output of the addition function to change when time is
involved. As in the addition function, various types of activation functions can be used in
the activation function, depending on the function the cell will perform. For example, in
multilayer perceptron models, the activation function must be a differentiable function.

With the NET value obtained by the summation function in ANNs, the next step
which is the activation function, is passed. It determines the output to be produced by the
current network structure by processing the obtained NET input value with the selected
activation function. By means of nonlinear functions that can be used instead of a threshold
function or a linear function of the inputs, it is possible to model nonlinear input-output
relations with ANNs. The activation function is the function that provides linear or nonlinear
matching between the input and the output .

The choice of the activation function largely depends on the data of the neural
network and what information it is desired to learn. Generally, the chosen activation function
is a nonlinear one rather than a linear one. Because in linear functions, the output is linearly
proportional to the input. The most suitable function for a problem is determined as a result

6.2. The sigmoid function (Logsig) or hyperbolic tangent function (Tansig) is preferred as
the activation function in the Multilayer Perceptron (MLP) model, which is the most widely
model used today. There is no strict rule about which of the join or activation functions to
use. The researcher decides which one to use in which layer by trial and error.

41
Table 6.2. Some activation functions

Activation Function Equation Graph

Linear Function F(NET) = NET

Step (Heaviside)
F(NET) =
Function

Sigmoid Function
(Logsig)

Hyperbolic Tangent
(Tansig)

Threshold Value
Function F(NET) =

Rectifier Linear Unit,


ReLu

Source: Howard Demuth and Mark Beale (2004); p.1-7; Kamruzzman Joarder, Rezaul K. Beggand
and Ruhul Sarker (2006); p.4.

6.3.5. Output

It is the output value determined by the activation function. The output produced is
sent to the outside world or to another cell. The cell can also send its own output as input to
itself. In fact, there is only one output value from a neuron. The same value goes to more
than one neuron as input. An artificial neural network can have a single or multi-layered
structure. The number of cells in the layers may differ from each other. As the number of
cells in the intermediate layers increases, the process may become more difficult, but the

42
result may be better. Therefore, the number of neurons to be used in the study, the number
of layers, the number of neurons in the layers, and the selection of the addition and activation
functions will enable the network to learn and make the closest estimation to the truth with
the least error. This is possible only by trial and error. This procedure is carried out using
computer programs that have been created. Matlab, Python, and Neurosolution are some of
these programs.

6.4. General Structure of Artificial Neural Networks

Through connections, millions of biological nerve cells (neurons) combine to form


the brain; ANNs are formed by the combination of many artificial nerve cells. In other words,
layers are formed when neurons come together in the same direction, and ANNs are formed
by the combination of these several layers. Some neurons in this structure are in contact with
the external space to receive inputs and some to transmit outputs. All other neurons are
hidden layers. In this case, connecting the layers to each other in different ways leads to the
emergence of different network architectures. In the early years of ANNs, some researchers
randomly created connections between neurons and encountered negative results
(Kamruzzaman et al., 2006b, p. 5; Kubat, 2019, pp. 668 669).

The easiest way to design a structure is to separate the elements into layers. There
are three parts to the layering here. These are grouping neurons into layers, grouping
connections between layers, and finally grouping summation and transfer functions. This
situation is illustrated in detail in Figure 6.4. In other words, the coming together of nerve
cells cannot be random. In general, they come together in 3 layers: the input layer, hidden
layer or intermediate layer, and output layer, and in each layer, in parallel, they form the
network (Kubat, 2019, pp. 668 53). These layers are:

6.4.1. Input Layer

Neurons in this layer receive information from the outside world and transfer it to the
hidden layers. In other words, inputs are information coming to neurons. In some networks,
there is no information processing at the input layer. They just pass the input values to the
next layer. For this reason, some researchers do not add this layer to the number of layers of
networks.
43
6.4.2. Hidden (Intermediate) Layer

Information from the input layer is processed and sent to the output layer. The
processing of this information is carried out in hidden layers. In a network, the input and
output layers consist of a single layer, while the hidden layer can consist of more than one
layer. The hidden layers contain a large number of neurons, and these neurons are fully
connected with other neurons. The selection of the number of neurons in the hidden layer is
very important in terms of defining the size of the network and knowing its performance. In
addition, increasing or decreasing the number of neurons and layers in the hidden layer
affects whether the network is simple or complex.

6.4.3. Output Layer

The neurons in this layer process the information from the hidden layer and produce
the output that the network should produce for the input set (sample) presented from the
input layer. The output produced is sent to the outside world.

OUTPUTS

INPUTS

INPUT LAYER HIDDEN LAYER OUTPUT LAYER

Source: p.53.

Figure 6.4. General structure of the ANN

The neurons in each of these three layers and the relationships between the layers are
shown in Figure 6.4. The round shapes in the figure show neurons. There are parallel
elements in each layer, and the lines connecting these elements show the connections of the
network. Neurons and their connections form an artificial neural network. These
connections weights are determined during learning.

44
6.5. Architectural Structures of the Artificial Neural Networks

In general, ANNs are classified according to three main criteria. One of these criteria
is the connection structure of ANNs, which is also called the architecture of the network.
The artificial neural network's connection structure differs from one another based on the
directions of the connections between the neurons or the flow directions of the signals in the
network. While some networks are configured as feedforward, some networks contain a
feedback network structure.

The second classification criterion is the classification of the network according to


its learning patterns. According to this criterion, there are three basic learning styles. These
are supervised (teacher) learning, unsupervised (teacher) learning, and supportive learning.

The third classification criterion is the classification made according to the number
of layers. According to this classification, networks are divided into single-layer and multi-
layer networks.

6.5.1. Classification of ANNs According to Connection Structures

In ANNs, some networks are configured as feedforward, while others contain


feedback structures. In feedforward neural networks, the connections between neurons do
not form a loop, and these networks usually respond quickly to input data. In feedback
networks, connections contain loops and can use new data each time the loop is closed. These
networks generate the input response slowly due to the loop. Therefore, the training process
for such networks is longer. There are also network structures that can be defined as both
feedforward and backward propagation (Paul D. McNelis, 2005, p. 21). These two types of
networks are discussed in detail below.

Feedforward Networks

Neurons in a feedforward network are usually divided into layers. Neurons in each
layer are associated with neurons in the next layer by connection weights. However, the
layers do not have any connection among themselves. In feedforward networks, information
flow is carried out in one direction from the input layer to the output layer without feedback

45
(Fathi & Maleki Shoja, 2018a, p. 251). This is also called the activation direction. An
example of this type of artificial neural network is a single or multi layer perceptron. Such
networks are trained using the supervised learning technique.

Output

Y Output
Layer
Bias w jk
+1

H1 H 2 H p Hidden
Layer

Bias
+1 v ij Weights

x1 x2 x3 xn Input
Layer

Source: Kaastra and Boyd (1996), p. 218.

Figure 6.5: Feedforward network structure

The feedforward network structure is given in Figure 6.5. There is no feedback in


such networks. Here, input neurons are shown as{X1, X2, Xn}, and hidden neurons are

shown as {H1, H2, Hp}, Y represents output neuron, and vij shows, the weight of the
connection from the input neuron i to the hidden neuron j and Wjk, shows the weight of the
connection from the hidden neuron j to the output neuron k. The units shown as +1 are the
threshold values.

The input layer transmits the information it receives from the external environment
to the neurons in the hidden layer without making any changes. This information is processed
in the input layer and output layer to determine the network output. With this structure,
networks perform a nonlinear static function. The most well-known backpropagation
learning algorithm is used effectively in the training of this type of artificial neural network,
and sometimes these networks are also called backward propagation networks. The
feedforward backpropagation architecture was developed in the 1970s. Generally, a
backpropagation network has an input layer, an output layer, and at least one hidden layer.
There is no restriction on the number of hidden layers, but usually, one or two hidden layers
are used (Haykin, 1999, p. 22).
46
In backward propagation networks, the number of layers and the selection of the
number of neurons in each layer are important decisions in terms of affecting the
performance of the network. There are no clear selection criteria for what these numbers will
be. However, there are general rules that must be followed. These rules can be expressed as:

1. As the relationship between input data and output becomes more complex,
the number of neurons in the hidden layers also increases.
2. If the subject matter can be divided into several stages, it may be necessary
to increase the number of layers.
3. The amount of training data used in the network, constitutes the upper limit
criterion for the number of processes in the hidden layers.

Recurrent Networks

A recurrent neural network is a network structure in which the outputs in the output
layer and hidden layers are recurrent to the input units or previous intermediate layers. In
this network, there are turns or recurrent connections between neurons and therefore they
are said to have dynamic memory. In recurrent networks, the output of any cell can be sent
directly to the input layer and used as input again. Recurrent can be between cells in a layer
as well as between neurons between layers. ANNs with recurrent exhibit nonlinear dynamic
behavior. As examples of these networks, Hopfield, Elman, and Jordan nets can be given.
Figure 6.6 shows the structure of a recurrent artificial neural network structure (Haykin,
1999, p. 24).

47
1
Z 1
Outputs

1
Z 2

1
Z k

Z
1 N

Unit-Delay
operators
Inputs

Source: Haykin (1999),

Figure 6.6: Recurrent network structure

Recurrent ANNs can be developed with models with richer dynamics than
feedforward ANNs. However, feedforward networks are more applied than recurrent
networks in academic and practical fields. This is because recurrent networks are difficult to
implement in practice. In particular, the fact that recurrent networks can be created with
many different structures may prevent specialization in a particular model structure and
make the training phase difficult due to the inconsistency of the training algorithms (Kaastra
& Boyd, 1996, p. 217). In addition, the training of recurrent networks takes a long time. In
particular, as the number of data points in the training set increases, the training time gets
longer. For this reason, a feedforward network structure is preferred for solving problems
related to multivariate and long time series.

6.5.2. Artificial Neural Network Classification Based on Learning Styles

The main feature of ANNs is their ability to learn. The basic philosophy of learning
is to learn the relationships between the inputs and outputs of an event by using actual
examples of that event, and then determine the outputs of new examples that will occur later
according to these relationships. Here, it is accepted that the relationship between the inputs
and outputs of the examples related to an event contains information that will represent the
overall event 24).

48
The process of determining the weight values of the connections between the neurons
of an artificial neural network is called training the network. The learning process is started
by randomly assigning these weight values at the beginning. As the determined samples are
introduced to the network, the weight values of the network change, and the network
continues this process until it produces the desired outputs. Training of the network is
complete when the network is able to produce the desired outputs and make generalizations
about the events it represents. This is called network learning. Learning styles that enable
finding the best weight set during the solution of the problem are divided into three
categories: supervised (with the teacher), unsupervised (without a teacher), and supportive
learning 26). In the literature, these learning methods are also
referred to as machine learning algorithms , which are frequently used in time series data
analysis and are also a sub branch of ANNs.

Supervised Learning

In the form of supervised learning, a teacher or superviser helps the learning system
to learn the event. The supervisor illustrates the event that will be taught to the system as an
input/output set. In other words, for each sample, both the inputs and the outputs that need
to be generated in return for those inputs are sent to the system. The task of the system is to
map the inputs to the outputs determined by the consultant. In this way, the relationships
between the inputs and outputs of the event are learned.

In this learning, the difference between the outputs produced by the network and the
target outputs is considered as an error, and this error is tried to be minimized. This process
is tried continuously until acceptable accuracy is reached. In these trials, the weights of the
connections are changed to give the most appropriate output. If the intended output cannot
be produced against the given input, the connection weights should be changed to minimize
the error in the output value of the network. For this reason, supervised learning requires a
mentor. The consultant evaluates the performance of the control and guides the learning
process to gradually improve performance. Mean Absolute Error (MAE) and Root Mean
Squared Error (RMSE) performance criteria are used in supervised learning algorithms. As
examples of supervised learning algorithms, Widrow-Hoff s Delta Rule (Delta Rule) and
Rumelhart and McClelland s Generalized Delta Rule (backpropagation algorithm) can be
given. The operation of a supervised learning algorithm is shown below.

49
Inputs

Expected Output

Output Target Function

Weights

Training Algorithm

Figure 6.7. Supervised learning algorithm

Unsupervised Learning

In unsupervised learning, there is no teacher helping the system learn. Only input
values are shown to the system. The system is expected to learn the relationships between
the parameters in the examples by itself. This is an algorithm mostly used for classification
problems. Only after the learning of the system is finished, the labeling that shows what the
outputs mean be done by the user. Examples of this type of learning are Hebb, Hopfield, and
Kohonen learning algorithms and ART networks. The forecasting of such networks is not
evaluated by how well the network predicts a particular readily observed target variable
(Anderson & McNeill, 1992, p. 41).

Reinforcement Learning

In this type of learning, a teacher assists the network. However, instead of showing
the system the output set that should be produced for each input set, the supervisor expects
the system to produce the output corresponding to the inputs shown to him and generates a
signal indicating whether the produced output is true or false. The system continues the
learning process by taking this signal from the supervisor into account. As an example of
networks where this learning method is used; Linear Vector Quantization Model (LVQ) can
be given. It is used in many fields such as game theory, statistics, information theory,
simulation-based optimization, and genetic algorithms.

50
6.5.3. Artificial Neural Network Classification Based on the Number of Layers

According to the number of layers, ANNs are divided into two types: single-layer
neural networks and multilayer neural networks.

Single Layer Neural Networks

Single-layer neural networks consist of only input and output layers. Every network
has one or more inputs and outputs. Output units are connected to all input units. Every link
has a weight. In these networks, there is also a threshold value ( ) that prevents the values
of the neurons and thus the output of the network from being zero. The input of the threshold
value is always +1. Because there is no hidden layer to provide curvature, this type of mesh
is mostly used for linear problems.

In single-layer networks, each network has one or more inputs (Xi ; and
only one output (Zk; k=1), and each link has a weight shown with (Wik; i=1, 2 n; k=1).
The artificial neural network model in Figure 6.8 has a single-layer structure.

Bias
1

x1

w jk

x2 z k

xn

Input Layer Output Layer

Source

Figure 6.8: Single layer networks

51
The output of the network is found by summing the weighted input values with the
threshold value (Bias) ( ). This input value is passed through a transfer function to calculate
the output of the network. This process is formulated as below 60):
m
f( wi xi + )
i 1

Figure 6.9. Output function

In single-layer perceptrons, the output function is linear. In other words, the examples
shown to the network are shared between two classes, and the line that separates the two
classes is tring to be found. For this reason, the threshold value function is used. The most
important single-layer ANNs are the Simple Perceptron Model (SPM), the Adaptive Linear
Element Model (ADALINE), and the Multiple Adaptive Linear Element Model
(MADALINE).

The Simple Perceptron Model was first developed by Rosenblatt in 1958 to classify
successive processes in a certain order. It is a single-layered, trainable, and single-output
model that emerged as a result of studies carried out to model brain functions. It is based on
the principle that a nerve cell takes more than one input and produces an output. The output
of the network is a logical value consisting of one or zero. The threshold function is used to
calculate the value of the output . 61).

It is the first artificial neural network to be trainable. In the simple perceptron model,
the output value is derived when the net value, which is the sum of the products of the inputs
and their corresponding weight values, exceeds the threshold value. Learning about the
artificial neural network is possible by changing these weights. The most important feature
of the simple perceptron model is that it has the property of converging the correct weights
from the input variables entered into the network in the event of a solution to the problem.
The simple perceptron model is the basis for multilayer networks that will be developed later
and will be revolutionary for artificial neural network models .

Multilayer Neural Networks

A multilayer perceptron or neural network is a typical network consisting of several


layers formed by many neurons that are interconnected. In multilayer perceptrons, there is

52
an input layer, an output layer, and one or more hidden layers between these two layers. This
feature is the main feature that distinguishes multi-layer perceptrons from single-layer
perceptrons. Multilayer perceptrons are used in solving complex problems, especially in
predictions. Because this network has the ability to automatically transform into a non-linear
structure with a series of operations performed in the hidden layer of its structure (G. Zhang
et al., 1998, pp. 37 38).

The inadequacy of single-layer perceptrons in solving linear problems has led to the
idea that ANNs do not work. However, multilayer perceptrons developed as a result of
studies to solve the XOR problem have been an important step in the historical development
of ANNs. Research has resulted in the emergence of MLP and its learning rule, the back
propagation algorithm. The Multilayer Perceptron Model (MLP) is also called the Error
Propagation Model or Back Propagation Model

Obtaining successful results in solving nonlinear problems with multilayer


perceptrons and the back propagation algorithm has led to an increase in its use in many
fields, including business science. Multilayer perceptrons are a very important solution tool,
especially for problems that require classification, recognition, and generalization. Its main
purpose is to minimize the error between the expected output of the network and the output
it produces . Multilayer ANNs can be illustrated by the Linear Vector
Quantization Model (LVQ) and the Multilayer Perceptron Model.

The multi-layer artificial neural network structure is given in Figure 6.10. ANNs have
an MLP structure. The first layer of MLP is the input layer. The last layer is the output layer.
Between the input layer and the output layer, there is at least one hidden layer with hidden
neurons. Each neuron in the hidden layer has a nonlinear transfer function, and the results
obtained through these functions provide input to the next neurons (Smith & Gupta, 2000,
p. 1025). The two most commonly used transfer functions in MLP are the sigmoid function
and the hyperbolic tangent function.

53
Inputs

Input Layer

Bias

Hidden Layer

Output Layer

Outputs

Source

Figure 6.10: The multi-layer artificial neural network structure

The Figure 6.10 illustrates a simple backpropagation network structure that includes
an input layer, a hidden layer, and an output layer. Circles arranged in layers represent
processing elements, or neurons. There are three neurons in the input layer, and three
variables are introduced as input to the network. There are three neurons in the hidden layer
and three neurons in the output layer. Therefore, three variables are outputs from the
network. The values transmitted from the input layer to the hidden layer and the values
transmitted from the hidden layer to the output layer are weighted with a weight set. In the
network structure, the thick arrows represent the information flow during recall. Recall is
the process of presenting and outputting new input data to a trained network. Therefore, back
propagation is not used during the recall process. Back propagation is used only in the
training process, so the information flow in the training process is indicated by all the arrows
in the figure. There are no connections between neurons in the same layer. The links only go
forward .

They work according to a multi-layered, perceptive, supervised (teacher) learning


strategy. Both inputs and outputs to be produced in response to those inputs are shown to
these networks during training. The learning rule is a generalization of the Delta Learning
Rule based on the least squares method .

54
6.6. Learning Algorithms in Artificial Neural Networks

Before an MLP can be used for foresight, it must first be trained. In MLPs, connection
weights are determined that allow the expected output to be calculated during the training
phase . In a linear regression, the computation is easily done by finding
the sum of the squares of the error. However, ANNs include non-linear optimization by
nature, which causes the training process to be more difficult and complex than linear
regression. Although there are many training algorithms developed in the literature, the most
widely used among these algorithms is the back propagation algorithm developed by Werbos
(1974) and Rumelhart (1986) et al (G. P. Zhang, 2004a, p. 5)

The backpropagation algorithm has been one of the most important developments in
the history of ANNs. This algorithm can be applied to multilayer perceptrons consisting of
neurons with a continuous, differentiable activation function. The backpropagation
algorithm minimizes total error by adjusting weights with a gradient descent approach or
gradual decrease (G. P. Zhang, 2004b, p. 5).

The learning rule of the back propagation network is a generalization of the Delta
learning rule based on the least squares method. For this reason, the learning rule is also
called the Generalized Delta Rule 12, p. 77).

The back propagation algorithm allows one to find the weights that will produce the
most suitable solution for the given training set. The backward propagation algorithm
consists of two steps, forward computation and backward computation 2012, pp.
78 80; Smith & Gupta, 2000, p. 6):

The stage of forward computation is the stage of computing the output of the network.
In the backward calculation phase, it is the rearrangement of the weights in the model based
on the errors in the outputs. This algorithm allows for finding the most suitable weight values
for the training set presented to the network. Weights are regulated based on gradient
descent.

55
The implementation process of the back propagation algorithm of a single hidden
layer and feed-forward MLP is as follows:

Step 1: After samples are collected and the topological structure of the network and
learning parameters are determined, the initial values of the weights are randomly assigned.

Step 2: Steps 3 through 9 are repeated until the necessary condition for the training
to stop is met.
Step 3: For each set of training data, steps 4 through 8 are repeated.

Forward Propagation Calculation

Step 4: It starts the input layer of the network by displaying the inputs (G1, G2
from a sample selected from the training set. Inputs are transmitted to the middleware
without in the input layer is Ç ik

; then neuron's output is formed as Ç ik Gk .

Step 5: Each input to the neurons in the hidden layer is multiplied by the weights
{W1, W2 Wn} and the net input is calculated as follows:

n
a i
NET j w kj Ç k
k 1

Figure 6.11. Net input function

Here Wkj: represents the weight value that connects the k


the hidden layer element.

The sigmoid function is the activation function, and the important thing here is to use
a differentiable function and when used, the output is as given below. j in this equation is
the weight of the threshold value of the element in the hidden layer.

56
a 1
Çj a a

1 e NET j j

Figure 6.12. Output function

When these calculations are made by all neurons and finally the output values of the
output layer are found, the forward calculation phase ends. The activation function in the
hidden layer and the activation function in the output layer do not have to be the same. One
of them can be a sigmoid function and the other a tangent function or some other function
.

Backward Propagation Calculation

Step 6: For the input presented to the network, the output produced by the network
and the expected outputs (E1, E2 ,....) are compared, and the difference, or error value, is
distributed to the weight values of the network, and the error is tried to be reduced in the
next iteration. (em);

em = E m - O m
Figure 6.13.

This value is the error obtained for a processing element (the neuron). To find the
Total Error (TE) for the output layer, all errors must be summed. The main purpose of
training the MLP network is to minimize this error. TE is formulated as:

1
TE = 2
em
2 m

Figure 6.14. Total error

The error is distributed to the neurons to minimize the total error. This is done by
changing the neuron weights. This is achieved by changing the weights between the hidden
layer and the output layer, and between the hidden layer and the hidden layer input layer.

57
Step 7: If we call wa the amount of change in the connection weights connecting

Figure 6.15. Amount of changes ( wa)

In this equetion ( ) is the learning coefficient, and, ( ) is the momentum coefficient.


The learning coefficient is effective in improving the performance of the network. A small
learning coefficient extends the training time, while a large learning coefficient shortens it.
However, too large a learning coefficient prevents convergence. The momentum coefficient,
on the other hand, is effective in preventing the mesh from getting stuck at a local minimum
point during knitting. It ensures that the weight change value is added to the next change at
a certain rate, and the total error tends more towards zero. m in the equation is the error of
.

m = ). em
Figure 6.16. output unit error

Here, is the derivative of the activation function. When the sigmoid


activation function is used, this equation is;

m = Om (1 - Om ). em
Figure 6.17. Output unit error

Step 8: After calculating the amount of change, the new values of the weights in the
calculated as follows:

a a a
w jm t w jm t 1 w jm t
Figure 6.18. Value of weights

58
In addition, the weights of the threshold value are changed. If the threshold value of
ç
the neurons in the output layer is represented by their weights the amount of change is
given as the threshold output is constant and 1;

ç ç
m t m m t 1

Figure 6.19. The amount of changes

The new weight value of the threshold value in the t iteration is as follows:

ç ç ç
m t m t 1 m t

Figure 6.20. The new weight value

In the second case, the errors of all elements in the output layer must be taken into
account when changing the weights between the input layer and the middle layer or between
the two middle layers. If this weight change, for example, is represented by wi between the
input layer and the middle layer, the amount of change is;

Figure 6.21. The amount of weight change

a
The factor is in the form of;

a a
j f ( NET ) m w jm
m

a
Figure 6.22. factor value

a
and when the sigmoid function is used factor value is;

a
Figure 6.23. factor value

59
The new values of the weights are;

i i i
wkj t wkj t 1 wkj t
Figure 6.24. The new values of the weights

If the interlayer threshold weights are denoted by the amount of change;

a a a
j t j j t 1

Figure 6.25. The amount of change

The new values of the weights for the t iterastion are provided by the equation
below,

a a a
j t j t 1 j t

Figure 6.26. The new values of the weights

As a result, the forward and backward calculation steps are performed for one
iteration. All of the weights are rearranged, and the process of editing the weights is
continued until the termination criterion is met.

Step 9: If the error has reached the predetermined fault tolerance or the termination
criterion has been met, the training is stopped. Otherwise, the process is repeated starting
from Step 4.

The above-mentioned steps continue until the learning of the multilayer perceptron
is complete, that is, until the errors between the realized outputs and the expected outputs
are reduced to an acceptable level. For the network to learn, there must be a stopping
criterion. This is generally taken as the error falling below a certain level.

Input data is presented to the network and propagates throughout the network until it
reaches the output layer. This forward process produces the calculated output.
An error value is calculated for the network by comparing the calculated output with
the target output. To train the network, back propagation algorithms from supervised

60
learning algorithms are used. This algorithm propagates backwards through the network,
starting with the weights between the output layer neurons and the last hidden layer neurons.

When backpropagation is finished, the forward process starts again, and this cycle
continues until the error between the calculated output and the target output is minimized.

When the back propagation algorithm is chosen as the learning algorithm, two
parameters become important. The first of these is the learning rate ( ), and the other is the
momentum coefficient ( ). These two parameters will be briefly described below.

Learning Rate ( )

The learning rate has a significant impact on the performance of the network and
determines the amount of change in the weight values of the links. This coefficient usually
has a value between 0 and 1. In the case where the learning rate is small, the training process
takes a long time, while it takes a shorter time with the growth of this value. While very
small values of the learning rate cause the learning process to slow down unacceptably, very
high values of the learning rate cause instability. In a case where the learning rate is greater
than 1, the network oscillates between local minima, and no convergence event occurs. In
addition, the fact that this ratio is too low may not provide the opportunity to find the
minimum 12, p. 99; G. Zhang et al., 1998, p. 48).

Momentum Coefficient ( )

This coefficient, which affects the learning performance of the network, is defined as
the addition of a certain rate of the change in the previous one iteration to the new change
amount 12, p. 99). This coefficient has been specifically proposed to enable
networks stuck in local solutions to find better results with a jump and is a factor that helps
the network recover faster. Taking into account the momentum coefficient, which has values
between 0 and 1, produces a reduction in the number of steps and the total network error (G.
Zhang et al., 1998, p. 48). The momentum coefficient helps to speed up the educational
process with a large learning rate, while minimizing the tendency to oscillate.

61
Determining the optimal learning rate and momentum coefficient is largely
experimental and heuristic. In addition, these parameters vary greatly depending on the
problem area of interest.

6.7. Learning Rules

In the literature on neural networks, there are many learning rules used in learning
systems. The majority of these learning rules are based on the oldest and best-known Hebb
Learning Rule. Some important learning rules in use are given below.

6.7.1. The Hebbian Learning Rule

It is the oldest and most famous learning rule, developed on a biological basis by
Canadian psychologist Donald Hebb in 1949. According to this rule, if a neuron receives
input from another neuron and both neurons are highly active (i.e., they have the same
mathematical sign), the weight of the connection between neurons must be increased. In
other words, if a cell is active, it tries to make the cell it is connected to active, and if it is
passive, it tries to make it passive (Hebb, n.d., pp. 60 78).

6.7.2. The Hopfield Learning Rule

This rule, which is similar to the Hebb rule, determines how much the connections
between the artificial neural network elements should be strengthened or weakened. The
main difference is that it also determines the size of the change to be made in the connection
weight. Accordingly, if both the input and the desired output are active or both are inactive,
the connection weight is increased by the learning coefficient; otherwise, it is decreased by
the learning coefficient. Strengthening or weakening of the weights is carried out with the
help of the learning rate. The learning coefficient is a fixed and positive value assigned by
the user, which generally takes a value between 0 and 1 2012, p. 26).

6.7.3. The Delta Rule

One of the most commonly used learning rules is the Delta Rule. This rule is an
improved version of the Hebb rule. It was developed by Widrow and Hoff (1960). It is based

62
on switching input connections in order to reduce the difference (delta) between the desired
output and the actual output of the processing unit. For this reason, this algorithm is also
known as the least-squares rule. The error is reduced by simultaneously propagating back
from one layer to the previous layers. The process of debugging the network continues from
the output layer to the input layer (Anderson & McNeill, 1992, p. 30).

6.7.4. The Gradient Descent Learning Rule

This rule is similar to the Delta rule. The adjustment of the weights can be performed
in a manner proportional to the first derivative of the error (gradient) between the desired
output and the actual output for a processing unit. The aim is to get rid of the local minimum
and catch the general minimum by reducing the error function (Kamruzzaman et al., 2006b,
p. 131). It is one of the most used optimization algorithms in ML. The size or smallness of
the learning rate is an important factor in achieving the optimal result. A low learning rate
increases the processing time required to achieve the optimal result, whereas a high one does
not reduce error because the steps are too large and deviate from the optimal. Therefore, the
learning rate is chosen to be large enough to reach the global optimum by trial and error. The

Loss

very high learning rate

high learning rate

low learning rate

good learning rate

Epoch

Source: https://medium.com/deep-learning-turkiye/gradient-descent-nedir-3ec6afcb9900

Figure 6.27. Graident descent

63
6.7.5. The Kohonen Learning Rule

This rule was developed by Teuvo Kohonen (1982), inspired by learning in biological
systems. Neurons are thought to compete to learn to adjust their weight. According to this
rule, the elements of the mesh compete with each other to change their weight. The cell
producing the largest output becomes the winning cell (neuron), and the connection weights
are changed. This means that, this cell becomes stronger than the cells next to it. The cell
has the capacity to warn and ban its neighbors. The Kohonen rule does not require a target
output. Therefore, it is an unsupervised learning method. This rule, also known as self-
organizing, is used especially in studies on distributions. However, due to the fact that its
theoretical infrastructure is not fully developed, it has not become widespread in practice yet
(Anderson & McNeill, 1992, p. 30; Kohonen, 2001, pp. 71 72).

6.8. Application Areas and Properties of Artificial Neural Networks

ANNs, which have gained a wide application area in real-life problems, have been
used in the solution of problems in many different areas, especially those that are difficult
and complex to solve, and generally very successful results have been achieved. It is used in
many industries today. There is no limit to the application area. However, it is mainly used
in some areas, such as forecasting, modeling, and classification. ANNs, which emerged in
the 1950s, became sufficiently powerful for general-purpose use only in the 1980s. Since it
is the method that best describes the trend or structure in the data, it is very suitable for
forecasting and forecasting operations. The following examples can be given for the
common uses of ANNs in real life 2012, p. 203-206).

Industrial applications (quality control, inspection, system modeling, etc.),


Financial applications (economic and financial forecasting, credit rating,
macroeconomic forecasts, risk analysis, etc.),
Military and defense applications (target recognition and tracking systems, etc.),
Health applications (medical image processing, urology applications, etc.),
Applications in other fields (oil and gas exploration, handwriting recognition,
etc.).

64
ANNs need special environments in which they can work due to their densely
connected and complicated processing structures. Therefore, ANNs run on computers with
special software prepared for this purpose. Today, special hardware is being developed to
run increasingly dense and complex neural networks and to process information faster.

Properties of the Artificial Neural Networks

ANNs can produce results according to the information they have without any
intervention by performing learning, association, classification, generalization, prediction,
feature determination, and optimization processes. ANNs have the ability to make correct
decisions for subsequent inputs by arranging themselves with the information given during
the learning process. ANNs, which are used in many fields today, are used effectively in
many fields due to the following general features. ANNs have the ability to make correct
decisions for subsequent inputs by arranging themselves with the information given during
the learning process. ANNs, are used effectively in many fields today due to the following
general features 2012, p. 31).

ANNs perform ML: The main function of networks is to enable computers to learn.
By learning from the events, it tries to make similar decisions in the face of similar events.

The working style of the programs is not similar to the known programming methods:
There are different information processing methods than the ones in which traditional
programming and AI methods are applied.

Storage of information: In ANNs, information is measured by the values of the


connections in the network and stored in the connections.

Nonlinearity: The neuron, which is the basic processing element of ANNs, is not
linear. Therefore, ANNs, which are formed by the combination of neurons, are also non-
linear. However, the fact that the activation function used is not a linear function turns ANNs
into a non-linear structure. Therefore, ANNs can be used to solve nonlinear problems in real
life.

65
Learning: Neural networks learn and solve problems using examples from real
events. In order to learn about events, examples related to those events should be determined.
Artificial neural network gains the ability to make generalizations about the event by using
examples. In order for ANNs to show the desired behavior, they must be programmed in
accordance with their purpose. In other words, correct connections must be made between
neurons, and connections must have appropriate weights. Due to the complex nature of
neural networks, connections and weights cannot be preset or designed. As a result, ANN
should learn the problem by using training examples from the problem it is interested in, so
that it exhibits the desired behavior.

In order to operate ANNs safely, they must first be trained and their performance
tested: Training the network means showing the existing samples to the network one by one
and determining the relationships between the events in the example by running its own
mechanisms. The samples are divided into two sets, the training set and the test set. Each
network is first trained with the training set, and when the network starts to give correct
answers to all examples, the training work is complete.

Ability to generate information about unseen examples: The network can generate
information about examples it has not seen by making generalizations from the examples
shown to it.

Usability in detection events: Networks are mostly used to process information for
detection.

Pattern association and classification: The purpose of most networks is to associate


patterns that are given as examples to themselves or others. Another purpose is classification.

Ability to self-organize and learn: It is possible for the network to adapt to new
situations shown by examples and to learn about new events continuously.

Ability to work with incomplete information: Networks can work with incomplete
information after they are trained and can produce results even though there is missing
information in new samples. This does not degrade their performance. The network itself
can learn what information is important during training.

66
Having fault tolerance: Networks ability to work with incomplete information
ensures that they are fault-tolerant. Since the ANN consists of many neurons connected in
various ways, it has a parallel distributed structure, and the information of the network is
distributed over all connections in the network. Therefore, the inactivation of some
connections or even some neurons in a trained ANN does not significantly affect the
network's ability to produce accurate information. Therefore, their ability to fix the error is
higher than other methods.

Ability to process uncertain and incomplete information: After learning the events,
networks can make decisions by establishing relationships with the events they learned under
uncertainty.

Gradual degradation: The fact that networks are fault-tolerant causes their
degradation to be gradual. A network breaks down slowly and gently over time. This is due
to incomplete information or disruption of neurons. Networks do not degrade as soon as a
problem arises, they degrade gradually.
Distributed memory: In neural networks, information is spread over the network. The
values of the connections of neurons with each other indicate the information of the network.
A single link makes no sense. The network as a whole characterizes the entire event it learns.
Therefore, the information is distributed over the network. This results in a messy memory.

Ability to work with only numerical information: Information represented by


symbolic expressions must be converted to numerical representation in order to be
interpreted and to produce solutions.

Hardware and Speed: ANNs can be implemented with large-scale integrated circuit
(VLSI) technology due to their parallel structure. This feature is expected in real-time
applications as it increases the fast information processing capability of the neural network.

67
6.9. The Design of Artificial Neural Networks

In this part of the study, information will be given about how to design ANNs and
which criteria to consider during the creation of network architecture. The number of layers
in ANNs, how many neurons these layers will consist of, how the neurons should be
positioned relative to each other, and the flow directions of signals between neurons
determine the structure of the networks.

There are differences in performance and characteristics between the structures of


ANNs. The structures of ANNs are very important, especially since they determine the
modeling ability of the network. At the design stage of the artificial neural network, the most
suitable one is selected among these network structures. Although there are some suggested
techniques for determining the optimal network architecture, these methods are quite
complex and difficult to implement. Also, none of these methods guarantees an optimal
solution for real estimation problems. In short, there are no precise and clearly applied
methods for determining these parameters. Therefore, neural network design is more of an
art than a science (G. Zhang et al., 1998, p. 42).

6.10. Training and Testing of Artificial Neural Networks

The process of determining the weight values of the connections of neurons in ANNs
is called training the network Initially, these values are randomly assigned. Networks
change these weight values as examples are shown to them 2012, p. 55).

In the setup phase of the neural network estimator, the sample dataset is divided into
two datasets for training and testing of the network. There is no general rule for separating
data. However, the data type, amount of data, and the characteristics of the problem are
important factors in separating the data set. Inaccuracies in the selection of the training and
test dataset will affect the performance of the network (G. Zhang et al., 1998, p. 50). The
selected datasets should be at a level that can describe the sample space. There are few
suggestions in the literature for the determination of training and test sets. Many researchers
use 90% of the data as a training dataset, while the remaining 10% is used as a test dataset.
Likewise, the rates of 80%, 20%, or 70%, 30% are frequently used in the literature to divide

68
the data into periods (G. Zhang et al., 1998, p. 50). Training samples from the data separated
according to this rule are used to develop the artificial neural network model, while test
samples are used to evaluate the predictive ability of the developed model.

In the learning process, weights are randomly assigned at the beginning, and the
weight values change as the samples are shown to the network according to the chosen
learning approach. The goal here is to find the weight values that will produce the correct
outputs for the examples shown to the network. The network, which has reached the correct
weight values, has become able to make generalizations about the event represented by the
examples and has completed the network learning.

The philosophy of training multi-layer networks, which is the model to be considered


in this study, is not different from other ANNs. For this reason, the problems encountered in
the training and education of multi-layer networks will be discussed here. The important
point in the training process of the network is to find the weight values that will give the
least error in the problem space. Considering a simple problem to explain error minimization,
suppose that the event (problem space) that the network is asked to learn has an error space
as shown in Figure 6.28.

Error (E)

W* Weight (W)

Source: p..82.
Figure 6.28. Representation of learning in error space

69
W in the figure shows the weight vector with the least error. Multilayer networks are
desired to reach the W* value. This weight value represents the point at which the error for
the problem is the smallest. For this reason, by making a change by W in each iteration, it
is ensured that an error of up to E is reduced at the error level.

It should be noted that the error level of the problem will not always be simple and
two-dimensional, as seen in Figure 6.28. Complex problems in real life have many minimum
and maximum points at the error level. The minimum points will only give the W* minimum
weight vector, or global minimum point, at one level. As can be seen in Figure 6.28 although
the weight vector W*gives the least error level, it may not always be possible to catch this
point. This solution is the best the network can have. It tries to reach this solution during
network training. However, instead of stopping the training process of multilayer networks
at the global minimum, which is the best solution point, there is also the possibility of
stopping at the minimum point (W1, W3) with weight values that will give a higher error rate
(Zurada, 1992, pp. 206 207). In this case, it is said that multilayer networks fall into the
local minimum trap during training (Kamruzzaman et al., 2006b, p. 131). The local minimum
point may be at a level that is considerably higher than the global minimum level, or it may
be close to this level 2012, p. 83).

Error (E)

Weight (W)

Source: Zurada (1992), p.207.

Figure 6.29. Multidimensional error space

As seen in Figure 6.29, although W* is the weight vector that gives the least error for
solving the problem, it is often not possible to catch this error value in practice. It tries to
capture the W* solution of multilayer networks during training. Sometimes it can be stuck

70
on a different solution (the local minimum), and it is not possible to improve performance.
For this reason, users accept a margin of error up to a certain level by specifying a tolerance
value ( ) in the performances of the networks. The problem is considered learned when it
falls below the tolerance value. Since the errors of the W0 and W2 solutions in Figure 6.28
are above the acceptable error level, these solutions are unacceptable. These are called local
solutions. Although solutions W1, and W3 are not the best, their error level is lower than the
acceptable level. Although these are local solutions, they are acceptable solutions. As can be
seen, more than one solution can be produced for a problem. Therefore, it cannot be said that
multi-layer networks always produce the best solution. It would be more accurate to say that
they produced an acceptable solution. Even if the solution produced is the best solution, it is
difficult to know. In most cases, it is not possible to know 83).

Figure 6.28 also shows the following possible reasons why the best results could not
be found :

The training set presented to the network while the problem is being trained may not
represent 100% of the problem space.
The correct parameters may not have been selected when the MLP was created.
The network s weights may not have been determined exactly as desired at the start.
Then, the next day, the next.

For these and similar reasons, MLP may not reduce the error below a certain value
during training. For example, W1 finds the weights and cannot lower the error any further.
But W1 is actually a native solution, not the best solution. It can be seen as the local best
solution since the error has decreased to an acceptable level for the W1 vector. It may also
be possible that the global solution does not exist. It all depends on the design of the network,
the nature of the samples, and the training process.

Another problem encountered in the training of multi-layer networks is that the


network memorizes rather than learns. In this case, although the trained multilayer networks
produce correct answers for all samples in the training set, they cannot produce correct
answers for the samples in the test set. In other words, the network gives the impression that
it has achieved the global minimum by memorizing the sample input. The network didn't
actually learn; it just memorized the sample input set. Such a network will produce results

71
that are far from the performance of the training during the testing process. Users
encountering this type of problem should investigate the cause of this situation (Kaastra &
Boyd, 1996, pp. 229 231).

After the training of the network is completed, the attempts to measure whether it
learns (performance) are called network testing. For testing, the samples that the network
does not see during learning, which is the test set, are used. The weight values of the network
are not changed during testing. The inputs in the test set are given to the model of the artificial
neural network, and the output value of the artificial neural network is compared with the
desired output value. The purpose of this process is to see whether the ANN model can make
adequate generalizations. If the desired success is achieved in the training and testing stages,
the artificial neural network model can be used.

6.11. Selection of Network Structure

While creating an ANN model, the stages of determining the structure and structure
properties of the network, determining the properties of the functions in the neuron, and
determining the parameters by selecting the learning algorithm should be done carefully and
depending on the characteristics of the problem to be applied. Because these stages are very
important for the success of the results to be achieved. When the learning algorithm to be
used in the network is selected, the structure required by this algorithm will also be
automatically selected. Table 6.3 shows the network types that are successful based on their
intended use.

Table 6.3. Network types and areas of success

Purpose of Network Type Use of the network


Useage
Prediction Multilayer perceptron (MLP) Estimating an output value
from the inputs of the network
Classification Vector Quantization Models (LVQ) Determining which class the
Adaptive Resonance Theory Models (ART) inputs belong to
Probabilistic Neural Networks (PNN)
Counterpropagation
Data Shaping Hopfield Networks Finding incorrect information
Boltzman Machine in the entries and completing
Bidirectional Associative Memory (BAM) the missing information
Source: 2012), p.207.

72
6.11.1. Determining the Number of Input Neurons and Output Neurons

In ANNs, it is easy to determine the number of input neurons in estimation problems


based on cause-effect relationships. In such studies, the number of input neurons is
determined by the number of variables in the input vector. In this case, the number of input
neurons is equal to the number of variables. In time series estimation problems, the number
of input neurons is related to the number of delays. However, there is no definite way to
determine how this situation will be resolved.

It is easier to determine the number of output neurons than to determine the number
of neurons in other layers. In problems involving the future prediction of a time series, the
number of output neurons is equal to the length of the prediction period. While the number
of output neurons is equal to 1 in the single-period forecasting, in multi-period forecasting,
forecasting can be done in two ways. A one-period forecast is an iterative forecast. The
estimated period value is used as input for the next period. In this case, a single output neuron
is sufficient. The direct method is used in multi-period forecasting when more than one
period is estimated at the same time. In this case, the number of output neurons is equal to
the number of periods to be estimated (G. Zhang et al., 1998, pp. 44 46).

6.11.2. Determination of the Hidden Layer and the Number of Hidden


Neurons

The hidden layer and the hidden neurons in this layer are of great importance in the
success of ANNs (G. Zhang et al., 1998, p. 42).

The number of hidden layers varies depending on the problem, the amount of data,
and the design. Usually one or two hidden layers (intermediate layers) are sufficient. Using
more hidden layers significantly reduces the speed of the network. It can also cause the
network to memorize rather than learn. Preferably, a three-layer structure consisting of input-
hidden-output layers can be used. If the result is not satisfactory, 2 or 3 intermediate layers
can be tried later. Some studies have shown that there is no need for structures with more
than two hidden layers to solve most prediction problems (G. Zhang et al., 1998, p. 44). In
addition, applications have shown that more than four layers in total adversely affect the
success of the network (Kaastra & Boyd, 1996, p. 225).

73
Increasing the number of hidden layers will increase the number of connections
between all processing elements in the network, thus increasing the risk of memorization,
leading to increased computation time for the network and poor prediction results. For this
reason, when determining the number of hidden layers, the most appropriate number should
be found by trial and error, taking the number of data into account.

Similarly, there is no magic formula for the number of hidden neurons. In this regard,
the task falls to the designer s ability. It is generally preferred to work with a small number
of hidden neurons. Because they have higher generalization abilities. However, some rules
have been developed. These rules are not absolute, they are just a generally applied method.
For a 3-layer network with n input cells and m output cells, the number of neurons in the
hidden layer can be . The number of neurons in the hidden layer can vary between 1.5
and 2 times the above formula, depending on the type of problem and the number of data.
This is called the geometric pyramid rule. Baily and Thompson suggest that the number of
hidden cells for a 3-layer network is 75% of the number of cells in the input layer (Kaastra
& Boyd, 1996, p. 225).

Although there is no absolute rule for determining the number of neurons in the
hidden layer, the exact number depends on the structure of the network, the amount of data,
the type of problem, and the experience of the designer. It is determined by trial and error.

6.11.3. Data Standardization

The most important factor that provides nonlinearity in ANNs is the standardization
of the data. The data can be standardizated to recover from extreme values so that the
available data can be better modeled. In addition, the method to be chosen for normalizing
the data also affects the performance of the network. Data standardization ensures that the
negative effects of the cumulative totals of the data used during the processing process are
prevented. The activation functions used in the hidden and output layers, which determine
the data normalization range, play the role of compressing the output of a neuron to a range
of [0,1] or [-1,1]. Data standardization is done before the training process begins. The
formulas that are frequently used in data standardization approaches are as follows (G. Zhang
et al., 1998, p. 49):

74
Linear transformation in the range [0, +1]: xn = ( x 0 xmin)/( xmax xmin)
Linear transformation in the range [a,b]: xn=(b-a)(x0 xmin)/(xmax xmin)+a

Statistical normalization: xn = (x0 x )/s


Simple normalization: xn = x0 /xmax

Xn and X0 in the formulas represent the normalized and original data, and Xmin, Xmax,

x and s represent the minimum, maximum, mean, and standard deviation values along the
row or column, respectively.

Studies have been conducted on how important data normalization is in learning


ANNs, and they concluded that normalization is beneficial, but this benefit decreases as the
sample size increases.

It is usually independent of normalizing the inputs and normalizing the targeted


output values. In time series and forecasting problems, the normalization of inputs and
targeted outputs is done together. The choice of the normalization range depends on the
transfer function of the neurons in the output layer. If the sigmoid function is used in the
output layer, the normalization range is chosen as [0,1], and if the hyperbolic tangent
function is used, the normalization range is chosen as [-1,1]. However, if the identity
function, which is a linear function, is used as the transfer function of the neurons in the
output layer, then the selection of the normalization range is made according to the transfer
function of the processing elements in the hidden layer. The targeted output values should
be consistent with the normalization range of the observed network outputs as a result of
normalization. Interpretation of the results from the network can be done after conversion of
the outputs to the original range. The accuracy of the values produced by the network should
be calculated based on the original data set. The performance measure should also be
calculated after converting the outputs to the original range (G. Zhang et al., 1998, p. 50).

6.12. Determination of Neural Network Performance

Although there are many performance measures for a neural network estimator, such
as modeling time or training time, the best and most important performance metric is the

75
accuracy of the estimation. The accuracy criterion is defined as the difference between the
true value and the predicted value. This difference is called the prediction error (G. Zhang et
al., 1998, p. 51).

Determining the performance of a model means measuring the accuracy of predictive


modeling. Good forecasting leads to good decisions. When the performance of an artificial
neural network is mentioned, it is understood that the learning ability is measured. In other
words, it is measured whether the artificial neural network model learns the data well or not.
The most commonly used predictive accuracy metrics are as follows (Gujarati, 2011, p. 77;
G. Zhang et al., 1998, p. 51):

Table 6.4. Performance criteria

t|/N
MAE
SSE t)
2

MSE t)
2
/N
RMSE
MAPE
t/yt |(100)

Here;
et, is the estimation error,

yt, is the observation value of the period t,

N represents the number of error terms.

Among these predictive accuracy measures, MSE is the most widely used. An
important feature of this measure is that the prediction error can be decomposed into variance
sums. This feature shows that the MSE criterion depends only on the second moment of the
combined distribution of realization and prediction. Thus, it is a measure that provides useful
information. It should be noted, however, that it does not provide complete information on
the true distribution (G. Zhang et al., 1998, p. 52).

76
6.12.1. Determining the Stopping Criteria

The stopping criterion can be determined in two ways: if the error falls below a
certain value or below an acceptable error determined by the researcher, and when the
network completes the specified number of iterations, the training of the network is stopped.

6.12.2. Selection of Learning Algorithm

The two most important factors that determine the success of an artificial neural
network application are the structure of the network and the learning algorithm. The network
structure plays a decisive role in the selection of the learning algorithm. There are many
learning algorithms to be used in the development of ANNs. It is known that some algorithms
are more suitable for some applications. The most used algorithm is the back propagation
algorithm (G. Zhang et al., 1998, pp. 47 48).

6.13. Advantages and Disadvantages of Artificial Neural Networks


Applications

The advantages of ANNs are due to their non-linear structure and unique training
process. Generally, strengths of ANNs compared to other models include the ability to model
nonlinear structure, the ability to make generalizations, adaptability and flexibility,
information storage, error tolerance, and the absence of prerequisites and assumptions in
statistical or other modeling techniques (G. P. Zhang, 2003, p. 160).

The relationships between real-life events and the factors behind these events are
non-linear, and it is very difficult to model these relationships. However, by modeling
models more easily, ANNs can solve models that are difficult and complex to model
mathematically l, 2012, p. 207). Thanks to the transfer function they use, ANNs can
produce models for nonlinear problems and perform effective predictive modeling. For this
reason, ANNs are preferred as an effective predictive tool over traditional prediction
methods.

Through their learning ability, ANNs can generalize about situations that they have
not encountered before using known examples. By using the data in the learning phase, the

77
error between the input and the output is minimized, and an ANN model that gives the least
error between the input and output variables can be established. Due to their adaptability and
flexibility, artificial neural netw ANN can be retrained by changing the connection weights
in the case of new information or changes in the environment. This is one of the most basic
features that distinguish ANNs from traditional statistical methods (G. Zhang et al., 1998, p.
36).

In neural networks, information is measured by the value of the network s


connections and stored in the connections. As with other programs, the data is not embedded
in a database or program. In this way, ANNs have the power to store information compared
to other models 2012, p. 31). The fact that artificial neural networks, consisting
of densely interconnected neurons, can work with incomplete information, ensures that they
are tolerant to errors. There are limitations when modeling with traditional econometric
modeling techniques such as Box-Jenkins, ARIMA, moving average, and the linear model.
In addition, there is a loss of information since stagnation is required. However, ANN
modeling does not have these drawbacks (Abraham & Malik, 1973, pp. 1 17).

In order to successfully solve the problem using artificial neural networks, the
problem must be well modeled. ANNs do not need any prior knowledge other than examples
for modeling. ANNs applications are both practical and cheaper in terms of cost. Just
specifying examples and a simple program may be sufficient to solve the problem. The fact
that ANNs can work in parallel facilitates their real-time use.

6.14. Disadvantages of Artificial Neural Network Applications

Apart from the advantages provided by the use of artificial neural networks, there are
also some disadvantages that should be considered 2012, p. 34). Hardware
dependent operation of ANNs is an important problem. Most of today s machines have serial
processors. The reason networks exist is to be able to work with parallel processors. ANNs
need very fast parallel processors, especially because they do parallel processing. Performing
parallel operations on serial machines causes a waste of time. A different ANN structure
should be developed for each problem. This is done by trial and error. Neural networks also
do not guarantee that the solution found is the best solution.

78
There is no certain set of rules in the creation of ANNs, in model selection, in
determining the topology of the network, in determining the number of layers, or in
determining parameters such as the learning rate and the number of neurons that should be
included in each layer. It is determined entirely based on the experience of the researcher.
This is also an important disadvantage. There is also no method to determine when the
training of the network will end. Reducing the error of the network on the samples below a
certain value is considered sufficient to complete the training. However, it cannot be said
that an optimal trainer is provided. In this case, it may take a long time for the training to
take place.

The ANN does not give any information on how it transforms any input vector into
an output vector. From an engineering point of view, neural networks can be seen as a black
box . The black box receives the information from the outside and gives the outputs it
produces to the outside. What happens inside is unknown (Cheng et al., 2022a, p. 108-218).
In other words, the general rule or condition of the connection between input and output is
unknown to the network, and the network has no ability to explain it. While this situation
decreases trust in the obtained network, successful applications increase interest in ANNs.
In cases where it is difficult to find examples or examples that accurately represent the
problem cannot be found, it is not possible to produce healthy solutions to problems.

Despite all these disadvantages, it is possible to produce solutions for many problems
and create successful applications with ANN. In order for ANNs to be able to produce
solutions to problems by getting rid of these disadvantages, it is necessary to meticulously
create networks. Having sufficient knowledge about both the problems to be solved and
ANNs can provide successful results. It is necessary to know that it is possible to create a
network without ignoring this fact in searching for solutions to problems with the ANN
method, but it is not an easy process 2012, p. 35).

79
7. DEEP LEARNING

DL is a branch of ML that uses ANNs and algorithms to process data in a way that
is inspired by the structure of human nerve cells. DL generates results by learning from the
processed data. Just as people learn from their life experiences, DL algorithms improve their
learning abilities by learning a little more each time and making changes thanks to their
many layers. Thus, they are able to produce results that were not possible before. The concept
of deep here also refers to the number of layers in the artificial neural network. Unlike
classical neural networks, deep neural networks consist of more layers. Since it is a sub-
branch of ML, the terms machine learning and deep learning are often used as if they
were the same term. But the two systems are different and have different capabilities.

The foundations of ANNs and DL, which are created by influencing the human nerve
cell structure, date back to the 1940s. Bengio et al. have examined the general historical
development of DL in three waves. The term cybernetics was used to describe the period
between 1940 and 1960, connectionism to describe the period between 1980 and 1990,
and deep learning to describe the period after 2006 (Bengio et al., 2015, pp. 11 12).

The concept of ANNs first started with the mathematical model developed by
McCulloch-Pitts in 1943 based on the study of the human brain. Later, with the concept of
ML by Turing in 1949, the concept of the perceptron by Rosenblatt in 1958, and the
discovery of the Adaptive Linear Element-ADALINE by Widrow and Hoff in 1960, the
process called cybernetics was experienced. However, since studies in this field declined
until the 1980s, the period known as the first winter of artificial intelligence was
experienced. In the second wave, called connectionism, gained momentum in studies after
the 1980s gained momentum, and the concepts of DL networks, backpropagation algorithms
and convolutional neural networks were introduced by LeCun (1998) in 1998. In the 1990s,
AI studies continued increasingly, and with the Hebbian learning rule concept in 1949 and
Schmidhuber s very DL concept in 1993, great steps were taken in the depth and complexity
of neural networks. This process paved the way for the discovery of the Long Short Term
Memory-LSTM network (Schmidhuber, 2015, pp. 93 94). During this period, deep
networks were thought to be very difficult to train because the existing hardware was not
structured to allow too many experiments and also the cost was too high. In 2006, the third

80
wave, which started with the concept of Goffrey deep belief networks, gradually increased
its speed with the concept of Goodfellow Contested Producer Networks in 2014 and
continued its development (Bengio et al., 2015, pp. 11 17; Schmidhuber, 2015, pp. 90 98).

deep learning

connectionism
cybernetics

1940 1950 1960 1970 1980 1990 2000 2010 2020

Figure 7.1. The historical process of the deep learning

In general, this historical process of DL mentioned above is summarized by the graph


in Figure 7.1. According to this figure, there have been three waves of development of deep
learning: These are, the cybernetic process from the 1940s to the 1960s, the
connectionism process from the 1980s and the 1990s, and finally, the process that continued
rapidly after 2006 under the name deep learning (Bengio et al., 2015, p. 12).

Today, DL is used by the world s top technology companies, such as Google,


Microsoft, Facebook, IBM, Apple, Baidu, Adobe, Netflix, NVIDIA, and NEC.

7.1. Fundamentals of Deep Learning Architecture

The ultimate point of ML is DL algorithms. Contrary to traditional learning methods,


DL has a hierarchical structure. The general logic of the algorithm, in which supervised and
unsupervised learning methods are used, is based on the convolution-based artificial neural
network model 7). Convolutional neural networks come
together in different combinations and form the basis of modern DL architectures. Layers
contain many non-linear processing units, and each layer uses the output of the previous
layer as input. In DL, low-level features are extracted from layers without removing certain
features beforehand, and a hierarchical representation is formed by deriving high-level
features from low-level features. Deep neural networks are capable of processing voice,

81
image, speech, text, and video data. While doing so, the layers in the classical DL
architecture created are listed below 18).

Source: https://towardsdatascience.com/how-to-easily-draw-neural-network-architecture
-diagrams-a6b6138ed875

Figure 7.2. Elements of the deep neural network architecture

An example of deep neural network structure is shown in Figure 7.2 above. A deep
neural network structure consists of four main layers: the convolutional layer, max pooling
layer, fully connected layer, and softmax layer.

7.1.1. Input Layer

The first layer in convolutional neural networks is the input layer. It is the layer where
data is standardized and presented to the network. Data presented to the network is first
served from this layer. If the data has been subjected to any sizing process before, this layer
can be given to the network. If there is no standard for the size of the data or if there is a size
that must be followed at the entrance of the network, the sizing process is also performed in
this layer.

7.1.2. Convolution Layer

A series of filters (kernels) are used to circulate the input data. The content, size, and
number of the filters change depending on the application to be made. The size of the filters
to be used in the application should be smaller than the size of the data. It is the layer where

82
new attribute matrices are generated by circulating filters with dimensions and properties
determined by developers on data from the previous layer.

7.1.3. Activation Layer

It is the layer that brings the matrix values obtained from the convolution layer into
the range determined depending on the algorithm used. It is often used to convert negative
values to positives. The most commonly used activation functions in the literature are; the
logistic sigmoid function , the hyperbolic tangent function

= , and the Recitified Linear Unit (ReLU) function


(Fathi & Maleki Shoja, 2018a, pp. 258 259).

7.1.4. Pooling Layer

In many studies in the literature, the pooling layer is used after the activation layer.
Basically, the purpose of this layer is to reduce the preceding data to matrices of smaller
sizes. So, the network will work faster. However, downgrading may also cause data loss.
This is a predictable process that will normally be welcomed. Many matrices, such as
maximum value, minimum value, and mean value matrices, can be used in this layer for data
reduction.

7.1.5. Dropout Layer

In general, while DL works with large data sets, when it needs to work with a smaller
data set, the artificial neural network algorithm can memorize the data used during training.
As a result, the data in the network architecture may need to be forgotten. In other words, it
is the layer used to prevent the network from memorizing the training set in order to reduce
this risk, since there is a risk of over-learning in cases where the network has few data.

7.1.6. Fully Connected Layer

This layer is the one that takes all the data from the previous layers and transforms it
into a one-dimensional array matrix, or vector. Data from this layer goes directly to the

83
classification layer. It can be used after the pooling or memorization layer as well as after
the classification layer.

7.1.7. Classification Layer

In DL networks, as in artificial neural network algorithms, it is the final layer and is


the layer that creates the output value of the network by evaluating the data from the previous
full link layer. The output of the generated network can be classification information as well
as polynomial coefficients. Class information obtained from the classification layer is used
by many different classification functions. One of them is the SoftMax classifier, which uses
a probabilistic calculation method and can generate values in the range of 0-1 for each class.

7.2. Deep Learning Architectures

Since data science and the ease of accessing data increase the volume of data
obtained, it is very difficult and requires a long process to analyze big data with classical
programming methods. For this reason, many new AI algorithms have been developed that
can easily work with big data. Basically, the common feature of these algorithms is that they
can generate rules by creating a link between the data and the result. This rule concept
expresses the weight coefficients of the network and the structure of the network. DL has
found application in many areas where computer technologies are used. It is seen that it gives
very successful results. This result has led to an increase in the diversity of studies in this
field due to the interest in the deep artificial neural network method
p. 20).

In DL networks, which generally use convolutional neural network architecture, the


term deep refers to the number of layers in the network. The more layers, the deeper the
network. While classical neural networks or convolutional neural networks have only two
or three layers, deep neural networks can have hundreds of sublayers
2021, p. 21). However, deep neural networks are designed like typical feed-forward
networks. Due to the diversity of the problems and areas studied, different structures or
models of deep neural networks have been developed for the solution. Some of the main
ones are described below.

84
7.2.1. Multi Layer Perceptron (MLP)

The first single-layer perceptron model was developed by Rosenblatt in 1958. This
perceptron model, called a perceptron is based on the principle of a neuron taking multiple
inputs and producing an output using a threshold function. A perceptron is the simplest form
of neural network and consists of only one artificial neuron. The output of the network is a
logical value of 1 or 0 2012, p. 61). Later, in 1959, Widrow and Heff developed a
single-layer network called Adaptive Linear Element (ADALINE), which learns using the
least squares method. In 1990, Widrow and Lehr (1990) took this model one step further and
developed Multiple Adaptive Linear Elements (MADALINE) networks, which are formed
by combining more than one ADALINE unit 2012, pp. 68 74). These early
models of ANNs are used to solve problems with a linear structure but it was seen that they
could not learn nonlinear relationships. For this reason, multilayer perceptrons have been
developed.

Multi layer perceptrons (MLP), also called feedforward neural networks or deep
feedforward neural networks, are models consisting of three or more layers or multiple
perceptrons (McNelis, 2005, p. 251). MPL is also known as the core architecture of deep
neural networks or DL. The learning phase of multi-layered DL networks is done with the
backpropagation algorithm, which is a supervised learning technique, as in classical ANNs.
Various optimization approaches, such as Stochastic Gradient Descent (SGD), Limited
Memory BFGS (L-BFGS), and Adaptive Moment Estimation (Adam), are applied during
the training process. The output of an MLP network is determined using transfer functions
such as ReLU, Tanh, Sigmoid, and Softmax. The mean square error or logloss functions are
used to calculate the loss rate (Sarker, 2021, pp. 6-7).

Due to the large number of layers, it is called the deep network and is used for solving
more complex problems. As seen in the examples given in the literature, deep neural
networks have been used successfully in time series analysis (Jing et al., 2021).

7.2.2. Convolutional Neural Network (CNN or ConvNet)

Convolutional neural networks are a kind of feedforward-supervised deep neural


network that learns directly from input. Convolutional neural networks develop a classical

85
neural network design like multilayer networks. It creates the optimal parameters at each
layer to achieve a meaningful output, thus reducing the model s complexity. It uses the
dropout layer to reduce the memorization problems that can occur in a classical network. In
addition, it reduces the matrix of the outputs by using the pooling layer and transforms the
matrix into a one-dimensional vector by using the full link layer. Because of these features,
they are generally used in the fields of visual recognition, medical imaging, and natural
language processing. Convolutional network architecture is used to build applications such
as AlexNet, Xception, ResNet, and GoogleLeNet (Sarker, 2021, p. 7).

A general convolutional neural network structure is shown in Figure 7.3. The


convolutional neural network in the figure consists of six layers.

Source: https://teknoloji.org/derin-ogrenme-nedir-yapay-sinir-aglari-ne-ise-
yarar/#Multilayer_Perceptron_Cok_Katmanli_Algilayici

Figure 7.3. Convolutional network architecture

7.2.3. Recurrent Neural Network (RNN)

Recurrent neural networks are another popular network structure that uses sequential
or time-series data and feeds the output from the previous step as input to the next step.
(Sarker, 2021, p. 7). Like feedforward and convolutional networks, iterative networks also
learn from training inputs, but the difference in this network structure is that their internal
memory is used to process information from previous inputs. The purpose of the model is to
create a training model using sequential information. These networks are widely used for

86
complex tasks such as time series, handwriting learning, and language recognition due to
their learning capabilities.

Unlike a typical deep neural network, which assumes that the inputs and outputs are
independent of each other, the outputs in recurrent neural networks depend on previous
elements in the sequence. However, the backward dependence of a standard recurrent
network structure makes it difficult to train this network and causes the problem of
disappearing gradients. Due to the problem of lost information in this network, whose main
purpose is to measure long-term dependencies, it is difficult to store and learn information
with this network structure. For this purpose, methods such as recurrent neural network
architecture based on Long Short-Term Memory (LSTM), Bidirectional RNN/LSTM, Gated
Recurrent Units (GRUs) have been developed (Sarker, 2021, p. 7).

7.2.4. Long Short-Term Memory (LSTM)

It is an iterative neural network architecture based on feedback connections, unlike


feedforward neural networks. This architecture is a popular RNN architecture developed by
Hochreiter et al. (1997) that uses special units to deal with the vanishing gradient problem.
A memory cell in LSTM architecture can store data for a long time. It has a unit called a
memory cell. RNN and LSTM network structures are the same but the only difference comes
from the difference in the content of a node. In an LSTM network, information flow in and
out of the cell can be done through three ports. These are the input gate, output gate, and
forget gate. The data coming to these gates passes through a certain activation function
(sigmoid or tanh), and the incoming data leaves the node as output after certain processes
with the input content. Thus, LSTM is considered one of the most successful networks of
RNNs, as it works with a memory that remembers the inputs for a long time and solves the
training problem (Sarker, 2021, p. 7).

7.2.5. Restricted Boltzmann Machine (RBM)

A Constrained Boltzmann Machine (RBM) is also a generative stochastic neural


network that can learn a probability distribution among its inputs. An RBM neural network
typically consists of two layers, visible and hidden (Sarker, 2021, p. 10). The layers are not
interconnected. Each neural node is a unit of computation that processes the input, and this

87
node makes stochastic decisions about whether to transmit the input. Certain threshold
values are added to the inputs, multiplied by certain weights, and passed through the
activation function to form the output. In the restructuring phase, the output is compared
model consisting of several sequentially connected stacks of independent unsupervisewith
the previous output by reprocessing the process in this way, and it is tried to reach the output
value that gives the least difference by making more than one repetition (Chong et al., 2017a,
p. 190).

7.2.6. Deep Belief Network (DBN)

A Deep Belief Network (DBN) is a multi-layered generative graphicd networks that


use the hidden layer of each network as input for the next layer, such as RBMs. Clustering
with unsupervised learning is the last layer, followed by classification with the SOTFMAX
layer. The main idea in the DBN structure is to train unsupervised feedforward neural
networks with unlabeled data before fine-tuning the network with labeled input. One of the
key advantages of DBN, unlike typical surface learning networks, is that it allows for
reasoning capabilities and the detection of deep patterns, allowing the deep difference
between normal and erroneous data to be captured. Overall, the DBN model can play a key
role in a wide variety of high-dimensional data applications due to its powerful feature
extraction and classification capabilities and has become one of the hottest topics in the field
of neural networks (Sarker, 2021, p. 10).

7.2.7. Deep Auto-Encoder (DAE)

A deep automatic encoder network, consisting of two parts (encoding and decoding),
is a feed-forward but unsupervised learning model that sends data from the input layer to the
output layer using its own tags during training. Autoencoder is widely used in many
unsupervised learning tasks such as size reduction, feature extraction, efficient coding,
generative modelling, noise removal, anomaly or outlier detection, etc. There are methods
developed based on deep autoencoder architecture, such as Sparse Auto-Encoder (SAE),
Noise Reduction Auto-Encoder (DAE), Contracted Auto-Encoder (CAE), and Variational
Auto-Encoder (VAE) (Sarker, 2021, p. 9).

88
8. LITERATURE REVIEW

Stock exchanges are organized markets where long-term investment instruments


such as stocks and bonds are bought and sold. In these markets, those who supply funds and
those who demand funds come face to face. Thus, investors can take risks in a standard and
transparent environment, and funds are transferred from those who supply funds to those
who request funds. Since most of the capital needs of developing countries are provided by
stock portfolio investments, the development and volume of the securities markets are very
important for these countries. After the 1980s, rapid developments were experienced in the
capital markets due to the globalization trend that started with the liberalization of capital
markets in the USA and other developed countries. Mobility between capital markets
accelerated, and financial markets grew. Although the financial markets of America and
Europe fell into trouble with the global financial crisis in 2008, the growth in China and
India had a positive impact on the world markets (Karan, 2013, pp. 37 38).

With the great economic crises experienced in 1929 and 2008, the world stock market
activities all over the world have developed rapidly in an attempt to heal their wounds. Stock
exchanges have a significant impact on the global economy because they contribute
significantly to the economic development of countries by allowing a significant amount of
savings to be transferred to companies as capital. In addition, companies can reach long-
term, less risky, and less costly loans through securities exchanges instead of short-term,
high-interest, high-risk loans to be taken from the banking system. Undoubtedly, the
transformation of these funds arising from these savings into investments is a serious growth
driver for countries. The stock exchanges that investors invest most in the world are the New
York Stock Exchange (NYSE), an unorganized and over-the-counter stock market;
NASDAQ; the Toronto Stock Exchange; the Amsterdam Stock Exchange; the Tokyo Stock
Exchange; the Hong Kong Stock Exchange; and the Bombay Stock Exchange, which are
thought to be the starting place of the stock market in some sources.

An index is a criterion that is used to measure the change that consists of the
movements of one or more variables and thus provides general information about complex
events that are difficult to understand based on many more variables by linking them to a
single variable (Karan, 2013, p. 60). In other words, an index is a statistical composite

89
measure of movement in the market or industry. The price indices by which the stock prices
of companies are determined in a market or market segment are called stock market indexes
(Vui et al., 2013, p. 477). Since stock market indices include price movements in the sector,
they are not linear in nature, and their volatility is high. Stock markets are highly volatile
due to the effects of both micro and macroeconomic variables of countries, as well as
psychological variables such as consumer behavior, expectations, and political risk. Due to
these features, they are also important market indicators. Owing to their dynamic, complex,
and non-parametric structure, stock market index prediction is difficult. Considering that the
estimated size of the world stock market was 36.6 trillion dollars during the global crisis
period in 2008, it can be understood why stock markets are so important for national
economies and investors. For this reason, stock market indexes and stock price predictions
have attracted a lot of attention from investors and researchers recently.

In the finance literature, the number of studies examining the effect of


macroeconomic variables on asset prices is increasing. However, most of the studies are
aimed at estimating the effects of macroeconomic variables on stock returns. Because stocks
are the most risky investment instruments in the capital market, they can respond very
quickly to economic developments. For this reason, examples of studies on stock market
indexes and stock forecasts are given below.

The basis of the performance appraisal method, which is an indicator of effective


forecasts, is based on the study of the market timing of money managers in the USA by
Alfred Cowles in 1933. According to this study, Cowles concluded that 45 professional fund
management organizations cannot provide a higher return from the stock market (Cowles,
1933, pp. 309 324). In this study, 16 financial advisory companies giving investment advice
made approximately 7500 recommendations for common stocks from January 1, 1928, to
July 1, 1932, but an annual average of worse than 1.43 percent was observed, and even this
result was found to be a coincidence.

The modern portfolio theory, or Capital Asset Pricing Model (CPM), introduced by
Markowitz in 1952, which assumes that investors choose the securities they will buy into
their portfolios from a universe full of risky assets, is the oldest and most popular method
used in determining the returns on securities (Markowitz, 1952). This theory was later
developed by Sharpe (1964), Lintner (1965), and Mossin (1966) and used in many studies

90
(Lintner, 1965; Mossin, 1966; Sharpe, 1964). The fact that it considers the market factor as
the only factor affecting asset returns ensures that the model can be easily predicted. In
addition, Treynor (Treynor, 1965, pp. 63 75), Sharpe (1966), and Jensen (1968) have
developed three different performance valuation models, named after themselves and
consisting of a single parameter, which evaluate the performance of mutual funds using
SVFM and Modern Portfolio Theory. However, the necessity of using multi-factor models
in the analysis due to the ignoring of many factors has led to the emergence of the Arbitrage
Pricing Model (AFM) (Roll & Ross, 1980).

Chen, Roll, and Ross (1986) concluded in their study that the difference between
long-term interest rates, expected and unexpected inflation rates, industrial production,
differences between high and low quality bonds, and the risk premium systematically affect
stock returns. They concluded that oil prices do not have any effect on stock returns. In their
linear regression analysis using monthly time series, they concluded that stock returns are
affected by systematic economic news and changes in macroeconomic variables and are
priced in line with these effects. In addition, in their study, they found that the effects of
stock market indices such as the NYSE on asset prices are insignificant when compared to
macroeconomic variables. Innovations in macroeconomic variables are risks and are priced
meaningfully in the stock market (N.-F. Chen et al., 1986, pp. 383 403).

In contrast to Markowitz s equilibrium model and Sharpe s single index models,


Fama and French (1996) proved that changes in macroeconomic factors such as the market
index, firm size, and book value/market value ratio have an effect on security returns with
the model they established using the time series approach of monthly securities returns. In
addition, the value obtained by subtracting the risk-free interest rate from the market index,
which is the first variable used in the model working with monthly data, will be the residual
return of the index. Thus, the assumption that stock returns are based on index returns has
also been proven (Fama & French, 1996, pp. 55 84).

Fama (1990) examined the interaction between stock prices and real activities,
inflation, and money and found a strong positive correlation between stock returns and real
variables such as industrial production, GNP, money supply, interest rate, and lagging values
of inflation (Fama, 1990, pp. 1089 1108). However, most of the recent studies have drawn
attention to the short-term relationships between macroeconomic variables and stocks.

91
Schwert (1990) analyzed the relationship between real stock returns and real activity
between 1889 and 1988. Fama (1990), in his study with 65 years of data between 1953 and
1987, concluded that there is a high correlation between monthly, quarterly and annual stock
returns and future production growth rates. He also compared the two measures of industrial
production, the Miron-Romer index and the Babson index, and concluded that the new
Miron-Romer index of industrial production is less related to stock price movements than
the old Babson measure (Schwert, 1990, pp. 1237 1257).

Mukherjee and Naka (1995) used the Vector Error Correction Model in their study
with stock prices and six macroeconomic indicators. They found that there is a long-term
equilibrium relationship between exchange rate, money supply, inflation, industrial
production, long-term government bond interest rates, demand loan interest rates, and the
Tokyo Stock Exchange Index, and that the variables are co-integrated (Mukherjee & Naka,
1995, pp. 223 237).

Wongbangpo and Sharma (2002) investigated stock prices traded in five ASEAN
countries (Indonesia, Malaysia, Singapore, the Philippines, and Thailand) and concluded that
variables such as gross national product, consumer price index, money supply, interest rate,
and exchange rate are affected and that there is a causal relationship between them. In the
long run, a positive relationship was observed between the growth in production and stock
prices and an inverse relationship between inflation and stock prices. It was also observed
that the effects of exchange rates and interest rates had different negative and positive effects
according to countries (Wongbangpo & Sharma, 2002, pp. 27 51).

The three-factor model of Fama and French (1998) was applied to ISE firms
operating in the 1993 1998 period, and it was concluded that firms with low book value or
market value outperformed firms with high book value or market value. It has also been
observed that the performance of large firms is higher than that of small firms. These results
also showed that the findings of similar studies in developed and emerging markets differ
due to differences in national capital dynamics (Gonenc & Karan, 2003, pp. 1 25).

Chen, Leung, and Daouk (2003) conducted a study based on the idea that trading
strategies guided by their predictions about the direction of the Taiwan stock market, one of
the fastest-growing stock markets in Asian countries, can lead to greater effectiveness and

92
higher profits. In this study, they concluded that the performance of the Probabilistic Neural
Network model they deveped is better than that of the Generalized Moments Method,
Kalman Filter and Random Walk model (A.-S. Chen et al., 2003).

Chakravarty (2005) used monthly time series from 1991 to 2005 to investigate the
relationship between stock price and some basic macroeconomic variables in India. It has
been concluded that there is no causal relationship between stock price, gold price, and
exchange rate, but there is a one-way relationship with money supply, and industrial
production affects stock prices (Chakravarty, 2005, pp. 1 15).

Lui and Shrestha (2008) investigated the long-run relationship between the Chinese
stock market index and macroeconomic factors such as the exchange rate, inflation, money
supply, industrial production, and interest rate. They showed that there is a positive
relationship between money supply and industrial production and the stock market index and
a negative relationship between the exchange rate, inflation, and interest rates and the stock
market index (Liu & Shrestha, 2008, pp. 744 755).

Boyer and Zheng (2009) selected seven investor groups in the US capital market
during the 53-year period between 1952 and 2004 and found that there was a positive and
significant relationship between cash flows and stock returns of these groups, especially
mutual funds and foreign mutual fund groups (Boyer & Zheng, 2009, pp. 87 100).

on selected
variables between 1999 and 2006, 11 developing countries such as Turkey, Hungary, Poland,
Russia, Argentina, Brazil, Chile, Mexico, Indonesia, Malaysia, and Jordan. In their analysis,
the effects of selected macroeconomic variables on stock returns were analyzed using
balanced panel data analysis. As a result, stock returns were affected by the exchange rate,
inflation rate, and Standard and Poor s 500 index. It was determined that there is no
significant relationship between the interest rate, gross domestic product, money supply, and
oil prices 96).

In their study, Ilahi, Ali, and Jamil (2015) proved by linear regression analysis that
the Karachi Stock Market Index is weakly affected by changes in macroeconomic variables
such as the inflation rate, exchange rate, and interest rate (Ilahi et al., 2015, pp. 1 11).

93
In their studies using monthly data from the 2005
Muhammed (2016) examined the relationship between the Borsa Istanbul Index and the
variables of interest rate, exchange rate, export amount, import amount, industrial production
index, and gold price by means of a causality test and an impulse-response function. While
a one-way causality relationship was observed from BIST to the industrial production index,
exports, and imports, there was also a one-way causal relationship from the exchange rate to
BIST 74).

it was
seen that the stock returns of the banks traded in Borsa Istanbul 100 were positively affected
by the changes in the Standard & Poors 500 Index, the exchange rate, and the US interest
rate 19, pp. 1 25).

affecting the Borsa Istanbul 100 index and the New York Stock Exchange S&P 500 index
for the period 2010 2019 with the Toda-Yamamoto model. They found a bidirectional
causal relationship between the money and exchange rate and the BIST100 index, and a
causal relationship was seen only in the money supply towards the S&P500 index. In
addition, no causal relationship was found between the consumer price index, industrial
production index, export-import coverage ratio, interest rate, oil price, and gold price
(Kocabiyik & Fattah, 2020, pp. 116 151).

Vector Autoregressive Model


to see the effect of macroeconomic variables such as credit risk premium, exchange rate, oil
prices, and gold prices on Borsa Istanbul (BIST100) during the COVID-19 pandemic period.
In the pre-Covid-19 period, changes in BIST100 were mainly affected positively by gold
prices, while credit risk premium, exchange rate, and oil prices negatively affected BIST100.
In the Covid-19 period, while the effect of credit risk premium and oil prices on BIST100
increased, gold prices had a negative effect (Kaymaz & Yilmaz, 2022, pp. 206 217).

After the great economic depression in 1929, the science of econometrics, which is
based on testing economic theories and modeling them mathematically and statistically,
gained importance. Because the experience gained from the crises in the capital markets
around the world has shown that the modeling and forecasting of economic and financial

94
variables is extremely important for the economies of countries. Therefore, the science of
econometrics has begun to be used more by policymakers to guide economic policies. These
experiences have shown that modeling and predicting economic and financial variables are
extremely important for the economy and finance. Since economic and financial variables
are time series that are formed by ordering the observation values according to time, the use
of econometric methods in economic and financial modeling has gradually increased due to
the developments in the field of time series analysis in recent years. In addition, AI
technologies, which have found a wide application area as a powerful statistical modeling
technique, have become an alternative to econometric methods and a predictive method with
comparable performance. This method, which can produce very successful results in the
classification and prediction of time series, is also used extensively in the fields of statistics,
economics, and finance (Kaastra & Boyd, 1996, p. 216).

There are many studies in the literature on the modeling and estimation of financial
variables using AI methods. These studies deal with the financial performance and financial
forecasting of the markets, the forecasting of financial crises, the forecasting of exchange
rates, and the forecasting of stock prices. The following studies and their findings are
examples of studies conducted with AI methods in the literature.

Wu and Lu (1993), in their study using ANNs and Box Jenkins ARIMA model to
estimate the Standard & Poor s 500 index for the period 1971 1990, they concluded that the
ANNs outperformed the time series (Wu & Lu, 1993, pp. 257 264).

Hsieh (1993) showed that it gives very successful results in financial management by
using artificial neural networks, which try to simulate the physical process on which intuition
is based, or the adaptive biological learning process, and have the ability to produce solutions
even for uncertain and incomplete data for the given problem (C. Hsieh, 1993, p. 12).

Yao, Poh and Jasic (1996) estimated the exchange rates between the US dollar and
the five main currencies of the period, using the exchange rates for 2,910 days for the period
from 18 May 1984 to 7 February 1995. They used the Japanese Yen, German Mark, British
Pound, Swiss Franc and Australian Dollar as the five basic currencies of the period. They
estimated the exchange rates using ARIMA and ANN models and found that the ANN model
gave more effective results in estimating exchange rates (Yao et al., 1996, pp. 754 759).

95
In recent studies, it has been seen that AI technologies are used in analysis together
with traditional statistical methods, so that better results can be obtained by reflecting the
strengths and advantageous aspects of both methods. One of the studies that gave these
successful results is the study of Hu and Tsoukalas in 1999, which estimated the volatility
of the European Monetary System exchange rates by combining the GARCH, EGARCH,
and IGARCH models with the moving average variance (MAV) model and an artificial
neural network model. In this study, it was concluded that ANNs outperformed the least
squares and simple averaging methods (Hu & Tsoukalas, 1999).

Donaldson and Kamstra (1999) estimated the price volatility of the S&P 500 stock
index using ANNs with the GARCH and moving average variance (MAV) models. They
showed that, compared to the traditional weighted least squares method, they could take into
account the interaction effects in time series estimations and that ANNs were more effective
in non-linear time series (Donaldson & Kamstra, 1999, pp. 227 236).

between 1993 and 1997 with the alpha coefficient obtained by linear regression of fund
returns and various benchmarks and concluded that the funds generally did not perform well
(Dahlquist, Engstr m ve Soderlind, 2000, 409-423).

Zhang (2001) tested the applicability of ANNs to linear time series problems. Zhang
compared eight different ARIMA models with nonlinear multilayer ANN models for
predicting IBM stock closing prices. As a result, the study revealed that ANNs are successful
in solving linear problems as well as nonlinear problems (G. P. Zhang, 2001, pp. 1183
1202).

Bollen and Buse (2001), in their study on 230 funds during the 1985 1995 period,
investigated whether there is a difference in timing ability when the daily returns and
monthly returns of the funds are used. Regression analyses developed by Treynor and Mazuy
(1966) and Henriksson and Merton (1981) were used as a method. In the method developed
by Treynor and Mazuy, it was determined that 11.9% of the funds were used in the case of
monthly returns and 34.2% of the funds showed timing ability in the case of daily returns.
Similar results were obtained with the method developed by Henriksson and Merton (Bollen
& Busse, 2001, pp. 1075 1094).

96
Chen, Leung, and Daouk (2003) have tested The direction of the Taiwan Stock
Exchange Index, one of the fastest growing financial stock markets from developing Asian
countries, with probabilistic neural networks and the Kalman Filter, which is one of the
parametric statistical methods, using the generalized moment method and the random walk
model, and the probabilistic neural. It has been observed that the network model makes
predictions with superior performance compared to other traditional methods, and higher
returns are obtained from investment strategies made according to this method (A.-S. Chen
et al., 2003, pp. 901 923).

Kim, Oh, Sohn, and Hwang (2004) examined the effects of the financial crisis in
South Korea in 1997 on the economic structure of Korea and developed an early warning
system with multi-layer ANN models. They divided the year 1997 into three main periods:
the stable period between January 3 and September 18, the unstable period between
September 19 and October 21, and the crisis period between October 22 and December 27.
They used the KOSPI stock index as an input variable in the model, and, considering that
index volatility gives information about the direction of the market, they calculated the
index's end-of-day closing value, daily return, 10-day moving average, variance, and
variance ratio. In the study covering the period 1994-2001, they found that the volatility of
the KOSPI index was a harbinger of the crisis and realized that ANN models were
impressively successful in early detection of the 1997 economic crisis, in classifying the
market movements, and in following the fundamental trend of the economy (T. Y. Kim et
al., 2004, pp. 583 590).

Yu, Tresp, and Schwaighofer (2005) on the other hand, combined parametric linear
models with non-parametric Gaussian processes and concluded that nonlinear estimations
yield better results than traditional estimations (Yu et al., 2005, pp. 1012 1019).

Dutta, Jha, Kumar, and Mohan (2006) used the multi-layer ANN method in modeling
the Mumbai Stock Exchange index. They concluded that the ANN model successfully
predicted the index values (Dutta et al., 2006, pp. 283 295).

Panda and Narasimhan (2007) estimated the future value of exchange rates with a
linear autoregressive model, a random walk model, and an ANN model using weekly data
of Indian Rupee/US Dollar (INR/USD) exchange rates for the period January 1994 June

97
2003. They found that the ANN model gave more effective results for the firm and investors
in estimating the exchange rate (Panda & Narasimhan, 2007, pp. 227 236).

Tseng, Cheng, Wang, and Peng (2008) estimated the volatility of Taiwan Stock Index
(TXO) prices with a new hybrid asymmetric volatility approach, the multi-layer ANN option
pricing model. They used EGARCH and Grey-GARCH models as comparison criteria. They
concluded that the ANN model predicts market volatility more effectively (Tseng et al.,
2008, pp. 3192 3200).

Liang, Zhang, Xiao, and Chen (2009) estimated option prices using Hong Kong
options stock market data for the period 2006 2007 using ANN, finite difference, and Monte
Carlo methods. Forecasting was done first using traditional option pricing methods, and then
ANN and support vector regression (SVR) were used to reduce forecast errors. Thus, future
option prices were estimated using parametric and non-parametric methods, and they
concluded that the ANN method showed superior forecasting performance (Liang et al.,
2009, pp. 3055 3065).

According to Tsai and Wang (2009), there are two methods used in estimating stock
prices in the literature. These are fundamental analysis, which uses information from the
company's financial statements, and technical analysis, which uses figures and graphs based
on historical data. However, since fundamental and technical analysis alone are not sufficient
to make the right decision, they used two analysis methods based on computer technologies
that have better forecasting performance. With a hybrid model combining ANNs and
decision tree management (DT), they achieved forecasts with an accuracy of 77%, which is
higher than the individual models (Tsai & Wang, 2009, p. 60).

At the beginning of October 2008, the size of the world stock market was estimated
at approximately $36.6 trillion in the United States. Considering that the total world
derivatives market is 11 times the size of the entire world economy and has a nominal value
of approximately 791 trillion dollars, it also shows why stock market forecasting is so
important and why it is the subject of so many academic studies. Dase and Pawar (2010)
compiled the studies in the literature that make stock market index prediction with the
artificial neural network method, which gives successful results in large data sets. In these
studies, it has been shown that the artificial neural network method has proven to have quite

98
high predictive performance in predicting the stock index and predicting whether it is best
to buy, hold, or sell stocks (Dase & Pawar, 2010, pp. 14 10).

In Vadlamudi s (2017) study, in which studies in the literature on stock market


estimation were examined, it was shown that AI methods made more effective predictions
than traditional methods when methods such as linear regression, AI networks, and genetic
algorithms were compared (Vadlamudi, 2017, pp. 123 128).

Chen, Zhang, Yeo, Lau, and Lee (2017) used the parametric statistical methods of
linear regression (LR) and Support Vector Regression (SVR) to predict the volatility of
stocks on the Chinese Stock Exchange, which is very popular in both the business world and
the academic community. They estimated and analyzed with Repetitive Artificial
Intelligence (RNN) and Gated Repetitive Units (GRU), which are AI methods, and it was
seen that the estimation results of AI methods were more successful and showed higher
estimation performance (W. Chen et al., 2017, pp. 1 6).

Chong, Han, and Park (2017) used three data representation methods with a data set
consisting of high-frequency daily stock returns. Principal component analysis, an
autoencoder and constrained Boltzmann machine, and a three-layer deep neural network
(DNN) model to predict future stock returns were applied to the Korean stock market. They
observed that DNNs performed better than the linear autoregressive model in the training
set, but the advantage was mostly lost in the test set (Chong et al., 2017b, pp. 187 205).

According to Lin, Chien, and Cheng (2018), according to behavioral finance,


investors behaviors such as greed and fear during transactions in financial markets also
affect market trends, so traditional financial analysis or technical analysis can be used to
forecast short-term market trends from these highly complex data. argued that their method
was inadequate. Using neural networks in their work, they have proven that financial trading
markets follow a certain trading logic by integrating physical momentum behavior into
financial engineering technology analysis and market profile theory. They showed that
buying and selling behavior in financial markets can be explained by the physical trends of
the quantitative and technical analysis of the market profile theory (C.-C. Lin et al., 2018,
pp. 756 764).

99
Nunes, Gerding, McGroarty, and Niranjan (2019) showed that the multilayer AI
model gave better results, being the first comprehensive study to use a multivariate linear
regression model and a multilayer artificial neural network model to predict the European
yield curve. This result also paved the way for the development of better forecasting systems
for fixed income markets (Nunes et al., 2019, pp. 362 375).

Cao and Wang (2020) used principal component analysis and ANNs for an accurate
and effective stock prediction model and proved that the artificial neural network model they
created offers an effective stock selection strategy (Cao & Wang, 2020, pp. 7851 7860).

Jing, Wu, and Hefei (2021) created a hybrid model by combining the stock prices
traded on the Shanghai Stock Exchange with investor sentiment analysis and DL analysis,
which is an AI method. They concluded that the hybrid model they created by using the
Long-Short-Term Memory Neural Network approach to classify the hidden emotions of the
investors and to analyze the technical indicators in the stock market using the Convolutional
Neural Network model outperformed the basic classifiers in classifying investor sentiment
and outperformed the prediction of stock prices (Jing et al., 2021, p. 115019).

Cheng, Yang, Xiamg, and Lui (2022), noting that the financial market capitalization
of listed companies in the US reached $30 trillion in 2019, which is more than 1.5 times the
US Gross Domestic Product, emphasized how important it is for both investors and financial
institutions to predict the price movements of stocks in this huge but volatile market. In their
analysis using Multimodal Graph Neural Networks (MAGNN), they explained the
construction process of the leading lag effect and heterogeneous graphs, which is a new
approach for financial time series analysis that plays a major role in hedging market risks
and optimizing investment decisions, and concluded that this method shows superior
performance in market forecasting. They have shown that, this method provides investors
with a profitable as well as interpretable option and enables them to make informed
investment decisions (Cheng et al., 2022b, pp. 108 2018).

Many studies have been conducted using various econometric analysis methods to
predict financial crises, detect chaos and uncertainties, and determine the mobility of capital
markets, and the effectiveness of the methods has been compared. In this sense, besides the

100
econometric methods mentioned above, studies using Logit and Probit models, although few
in number, have also taken their place in the literature. These studies are listed below.

Frankel and Rose (1996) used a probit model to identify leading indicators of
currency crises for more than 100 developing countries using annual data for the period from
1971 to 1992. They concluded that high domestic credit growth, high foreign interest rates,
and persistently low foreign direct investment to debt ratios indicate a high probability of a
collapse (Frankel & Rose, 1996, pp. 351 366).

Kaminsky and Reinhart (1999) examined the sources and extent of 76 currency crises
and 26 banking crises using monthly data on 16 macroeconomic variables, whose findings
they considered likely to be indicative. They concluded that financial liberalization leads to
a banking crisis, banking sector problems trigger the currency crisis, and the currency crisis
caused by the banking crisis weaklened the already weak banking sector, and that this
process, which triggered each other by deepening the banking crisis, became a vicious circle.
They aslo concluded that when currency and banking crises occur together, the impact is
deeper than when they occur separately (Kaminsky & Reinhart, 1999, pp. 473 500).

Kim (2003) forecasted the stock price index, a financial time series, with a logistic
model and ANNs and concluded that ANNs are a promising method in stock market
forecasting (K. Kim, 2003, pp. 307 319).

the indicators of corporate financial crises in emerging markets during periods of economic
depression and emphasized that the separation model, which was the primary method for
predicting financial failure until the 1980s, has now been replaced by logistic regression and
that ANNs have yielded better results in predicting financial failure in recent years
& Aksoy, 2006, pp. 277 295).

the period 1990 2002, using the logit and signal methods, which are widely used in
estimating currency crises, for the monthly data of the 2003 2005 period in Turkey. Except
for two of these variables, the findings of the signal analysis made with 29 variables were
combined to form a common cluster, and they used five macroeconomic variables, which

101
they defined as "traditional indicators" for this period, containing the findings of this
common cluster, to form a financial pressure index. They tested the performances of the
models they established and concluded that both methods gave successful results, but the
leading indicators differed in terms of periods and economies (Kaya & Yilmaz, 2007).

Davis and Karim (2008) used the polynomial logit model and an early warning
system to predict leading indicators to enable the early detection of crises in the banking
sector. They found that the polynomial logit model performed better in predicting a global
banking crisis, while the early warning system for a country-specific banking crisis did better
(Davis & Karim, 2008, pp. 89 120).

Lin, Khan, Chang, and Wang (2008); Kaminsky and Reinhart, in their 1999 study,
created four models to predict the exchange rate crisis using data from 20 countries between
1970 and 1998. They measured the model performance of the early warning systems they
created by using four modeling techniques, namely the logistic regression model which is
one of the traditional methods, the KLR model, ANNs, and fuzzy logic. They concluded that
the most successful models are fuzzy logic, ANNs, and the logistics models, respectively
(C.-S. Lin et al., 2008, pp. 1098 1121).

ANNs, Support Vector


Machine, and multivariate statistical techniques (discriminant analysis and logistic
regression methods) to detect manipulations in the prices of stocks on the Istanbul Stock
Exchange. They concluded that multivariate techniques performed better
pp. 11944 11949).
Kovacova and Klistik (2017) concluded that the probit model performed better in
their study using logit and probit regression models, in which Slovak companies create their
bankruptcy forecasts (Kovacova & Kliestik, 2017, pp. 775 791).

Kantar and Akkaya (2018) estimated the effects of financial liberalization in Turkey
with the financial pressure index and logit and probit models they created using 19
macroeconomic variables for the period of January 2005-January 2017 and tried to reveal
the leading indicators of the financial crisis. They concluded that the increase in deposit
rates, the increase in the domestic debt stock, and the decrease in gross reserves cause an
increase the probability of a crisis (Kantar & Akkaya, 2018, pp. 575 590).

102
According to Akkaya and Kantar (2019), the importance of predicting banking crises,
which became a more important issue after the 2008 global crisis, has increased due to the
effects they have on the economy. For this reason, in their study, they examined the fragility
structure of the Turkish banking sector using annual data and using logit and probit models
(limited dependent variable models) for the period 1996-2017. While the exchange rate and
deposit interest variables, which have high explanatory power in all three models, are
statistically significant in the logit model, they have reached the conclusion that the loan
amount, deposit amount, and deposit interest variables are statistically significant in the
probit model (Akkaya & Kantar, 2019, pp. 131 145).

Some studies, taking the Turkish capital markets into account, also reveal that
artificial neural network modeling produces more successful results than other models.

the years 1983 and 1997, which are subject to the Capital Markets Board and/or traded on
the Istanbul Stock Exchange (ISE). He concluded that the ANN model is more successful
than the separation analysis in predicting financial failure (Yildiz, 2001, pp. 47 62).

Diler (2003) estimated the direction of the ISE National-100 Index the next day by
using the error back-propagation method with ANN modeling and concluded that the method
can predict the next-day value of the ISE National-100 Index by 60.81% (Diler, 2003, pp.
65 82).

Benli (2005) developed financial failure prediction models based on logistic


regression and an artificial neural network model by using data from 17 privately owned
commercial banks transferred to the Savings Deposit Insurance Fund and 21 privately owned
commercial banks in the 1997 2001 period. It has been determined that the power of the
artificial neural network model to predict financial failure (82.4%) is superior to the logistic
regression model (76.5%). Therefore, it has been determined that the artificial neural
network model can be used as a important tool to predict financial failure for all information
users (Benli, 2005, pp. 31 46).

Altay and Satman (2005) tried to estimate the returns of the ISE 30 and ISE all
indexes using multilayer ANN and linear regression methods. It has been concluded that
while ANN models do not give better results than linear regression for monthly and daily

103
returns, they are quite successful in estimating the direction of index returns (Altay &
Satman, 2005, pp. 18 33).

), sought answers to the following questions in their


study. Could the November and February crises have been prevented with the help of leading
indicators? Have the leading indicators worked for the Turkish economy? If it did, could we
predict possible crises? Can we avoid the crisis? Can we manage the crisis? By using
macroeconomic indicators such as exports, international reserves, the real exchange rate, the
real deposit interest rate, and the production index, which are accepted as leading indicators
of the crisis, and concluded that the leading indicators predict the crisis process with little
error n, 2006, pp. 237 256).

-100 index with the


multi-layered ANN model and determined that the forecast performance is strong and can
be used as a financial performance measure. With this study, it has been shown that effective
and strong predictions are achieved with the ANN modeling technique (Avci, 2007, pp. 128
142).

Altan (2008) estimated the exchange rate with an ANN and a vector autoregressive
(VAR) model using monthly data on the exchange rate (TL/USD) for the period January
1987 September 2007. It has been concluded that the predictions made with an ANN
architecture that learns by multi-layer feed-forward and back-propagation yield very
effective results (Altan, 2008, pp. 141 160).

modeled their study with ANNs, which are


increasingly important in time series analysis and are used in many fields. They analyzed
Canadian Lynx data with the hybrid model they created by combining ARIMA and Elman s
reversible artificial neural network model and concluded that this proposed model has the
best prediction accuracy performance (Aladag et al., 2009, pp. 1467 1470).

concluded that time series with


linear and non-linear structures are better predicted by ANNs and hybrid artificial neural
network models, which are more effective methods than traditional methods (Yolcu et al.,
2013, pp. 1340 1347).

104
ANNs
models and observed that the buy-and-hold strategy had a great advantage in most of the
examined periods (Avci, 2015, pp. 443 461).

105
9. METHODOLOGY AND DATA

Especially since the global economic crisis in 2008, researchers, policymakers, and
investors have understood how important financial forecasts are in preventing financial
crises and uncertainties. Although there is no clear consensus on when the global economic
crisis began, some researchers believe it began with the failure to repay subprime mortgage
(low value hold and sell) loans in the United States in 2006, while others believe it began
with the collapse of the large banking company Lehman Brothers in 2008. Everything began
with the failure of the large banking firm, Brothers. However, the impact area of the crisis
expanded and globalized in 2008, and the crisis spread to Europe and the rest of the world.

The aim of this study is to predict the long-term relationship between the ISE100
index value and the selected macroeconomic variables in the January 2001-January 2022
period by using Logit and Probit analysis and the deep neural networks method and to
compare the estimation performances of the results obtained for both methods. Thus, the
effectiveness of AI technologies, which emerged from developments in traditional
econometric methods and computer technologies and continued to develop rapidly in recent
years, in time series analysis will also be measured. For the models used in the research, the
most commonly used variables in studies on the market index in the literature were selected
(W. Chen et al., 2017; Cheng et al., 2022a; Donaldson & Kamstra, 1999; Kara et al., 2011;
Staub et al., 2015) and the meaningful variables obtained from the created meaningful
models were given as input to the models created by ANNs and DL methods, and the
effectiveness of the modeling methods was tested.

In the study, after giving information about AI technologies, it is intended to develop


and evaluate an alternative decision-making system that can be used to predict the short- and
long-term movement, trend, and price of the stock market. It also provides a review of the
literature on AI technologies applied to analyze financial problems, focusing mainly on the
modeling process. For this purpose, macroeconomic variables, which are frequently used in
studies on stock market indices in other domestic and foreign literature and proven to have
effects on stock prices, were chosen based on the studies made by Chen, Roll, and Ross
(1986) and Prantik and Vani (2004) on the basis of the literature review (W. Chen et al.,
2017; Cheng et al., 2022a; Kara et al., 2011; Staub et al., 2015)

106
While the price movements of financial assets are observed in financial markets, the
price of assets is generally used in empirical studies (Campbell et al., 1997, p. 9). They
presented two main reasons for using returns in the analysis. First, the return on a financial
asset provides a complete and independent summary of investment opportunities for an
investor. In other words, the investor who invests 1000 TL in a financial asset with an annual
return rate of 5% will have 1050 TL at the end of the year. The second reason is that asset
returns are more useful than prices for both empirical and theoretical reasons. In other words,
price series display a non-stationary appearance due to the presence of a stochastic and/or
deterministic trend and constant deviations from the unconditional mean. Asset returns, on
the other hand, are close to the average and only deviate from it in the short run. Therefore,
return series are generally stationary .

The study consists of 253 monthly observation values covering the time interval
between 2001:01 and 2022:01. In order to estimate the value of the Borsa Istanbul 100 index
selected as the input (dependent) variable, or, in other words, to investigate the effects on
the stock market in Turkey, seven macroeconomic variables were determined. These are the
Dollar Rate (TL/$), Money Supply, Producer Price Index, Industrial Production Index, Gold
Price (TL/Gr), Active Bond Interest Rate and Brent Oil Price. In addition, since the effects
of the global economic crisis experienced in the world in 2008 were more limited in Turkey
than in developed countries, these periods were not excluded from the data set in order not
to spoil the integrity of the data set and to see the effects of the crisis in Turkey in the 2001
period.

Seven macroeconomic variables were used as independent (input) variables in the


models. ISE100 index values were used as the dependent (output) variable to represent the
stock markets. This index is the basic index for the BIST stock market. In addition, since
monthly series were used in the study, the Gross Domestic Product series could not be
obtained, so the industrial production index was used to represent this variable. All variables
were subjected to statistical analysis, and the Eviews 12, STATA 15, and MATLAB R2021b
programs were used in the analysis.

The data sets used in the analysis were compiled as monthly series from the Central
Bank of the Republic of Turkey Electronic Data Distribution System (EVDS), the Turkish

107
Statistical Institute (TUIK) statistics, the monthly statistical bulletins of the Capital Markets
Board, the statistics of the Ministry of Treasury and Finance, and the Eurostat databases.

While the monthly values of the ISE100 index were used as the dependent variable
in the research, the consumer price index, industrial production index, gold bullion prices,
dollar rate, crude oil barrel prices, and monthly values of active bond interest rates were used
as independent variables. The arguments used are as described below:

Index: It is an indicator that measures the proportional change of investment


instruments traded on the stock exchange within a certain time period by collecting data such
as price, cost, sales, and production. Stock indices, on the other hand, measure the price and
return performance of stocks traded in the stock market. These measurements can be made
on both a holistic and sectoral basis. In this way, it is possible to monitor the economic
performance of stocks and sectors.

ISE100 Index: It is a fundamental index and fundamental indicator used to assess


the performance of the top 100 stocks traded on the Borsa Istanbul in terms of market and
trading volume. It consists of 100 stocks selected to be used in the futures market among the
publicly offered stocks of companies traded in the national market, excluding investment
trusts, with high market value and liquidity, taking into account their sectoral representation
ability. The transaction code of the ISE100 index, which is also accepted as an indicator of
the Turkish stock market, is XU100. Indices are calculated from prices. As the country s
financial system improves, the stock market index also rises. The ISE100 index is used as a
benchmark for other investment instruments. The ISE100 index allows one to measure the
price movements of the stocks in the index under certain conditions and thus determine the
general trend of the stock market. Price indices such as the wholesale price index and the
consumer price index are indicators that measure the change in the general level of prices.

Although stock market indices are a basic indicator for national economies, it is
necessary to know what the factors affecting the index are before making an evaluation.
Internal factors such as a country s economic data, unemployment figures, interest rates,
geopolitical position, and risks, as well as external factors such as global data and economic
relations between countries, are factors that influence a country s economic course and, of
course, stock market indices.

108
Two Years Bond Yield/ Active Bond Interest Rate: This rate is the 2-year bond
rate. Bonds are debt instruments used by governments, companies, or individual investors in
need of funds for various reasons, to provide financing for a certain period of time with a
variable or fixed interest rate. Investors prefer this investment tool to protect themselves
from economic risks and obtain a fixed return above inflation. While increasing interest rates
and increasing economic uncertainties due to the implementation of tight monetary policies
direct the investor to bonds in order to be protected from the said risks, as a result of stable
economic growth, interest rates tend to fall, leading them to stocks due to the reduction of
risks.

Interest is called the rent or price of money. If the interest rates that direct the entire
country s economy decrease, there is an economic recovery, while if the interest rates
increase, an economic slowdown occurs. Because when interest rates increase, fund owners
prefer alternative investment instruments that they anticipate will bring higher returns and
therefore do not turn into production. This will undoubtedly affect the banking sector as well.
For these reasons, stock returns significantly affect investors investment decisions in the
markets. When the market interest rate increases, the bond price decreases. A stock is a
security, like a bond. Therefore, when the Interest Rate of Active Bond increases, the stock
market index is expected to decrease. In other words, there is a negative relationship between
the two variables.

Since the investments to be made in stocks, bonds, and deposits are the types of
placements fed from the same fund source, the investor has two investment preferences. The
first is to buy bonds. The second is to earn interest income in return for deposits in the bank
or to earn stock market income by purchasing stocks. Therefore, in a situation where the
interest rates rise, the investor will turn to bonds and bills because he believes that he will
protect himself from risks, and in a situation where the interest rates decrease, he will turn
to stock investments . The graph in Figure 9.1. below shows the direction of
the relationship between the two variables.

109
Figure 9.1. ISE100-TYBY

The graphic above shows the reflections of the financial crisis in Turkey in February
2001. The financial crisis was reflected in the money market, and there were serious
increases in interest rates and exchange rates due to the increase in demand for TL and
foreign currencies and the inability to meet this increasing liquidity demand. On February
21, the overnight interest rate in the interbank markets increased to 6200%. When the foreign
exchange reserves of the central bank decreased significantly, it was unable to meet these
sharp speculative movements and left the exchange rate to fluctuate.

Exchange Rate: The exchange rate is a price and relates to commodity markets. The
stable course of the changes in the exchange rate affects economic stability positively. After
the liquidity crisis on November 22, 2000, and the currency crisis on February 19, 2001,
Turkey abandoned the disinflation program based on adjustable fixed rates and switched to
floating rates. The equilibrium exchange rate has a linear relationship with the ratio of price
levels. The exchange rate is also a macroeconomic variable that has been identified as a
benchmark for some securities. Since the crises in the foreign exchange and stock markets
cause uncertainties in both national economies and international markets, they have been the
focus of studies. It is very important both for governments to determine policy and for
investors to direct their investments. When the exchange rate increases, the returns on stocks
decrease as investors funds shift to other financial instruments. Therefore, empirical studies
have shown that while there is a negative relationship between the two variables in the short
term, there is also a positive relationship in the long term. The constant rise in foreign
exchange prices will cause investors who bind their savings to stocks to sell their stocks and
turn to foreign exchange. In this case, as the demand for stocks decreases, their returns will

110
also decrease. The graph in Figure 9.2 below shows the direction of the relationship between
the two variables.

Figure 9.2. ISE100-ERUSD

The financial crisis experienced in Turkey in February 2001 was reflected in the
money market led to an increase in the demand for TL and foreign currency in interest rates
due to the due to the inability to meet this increasing liquidity demand. This situation is
shown in Figure 9.2. Likewise, there were serious increases in the exchange rates, and while
the dollar rate was 686.500 liras on February 19, 2001, it rose to 960.000 liras with an
increase of around 40% per day. The Turkish lira depreciated by 11% in real terms. 19 banks
were closed, economic contraction was experienced, and 1.5 million people were
unemployed.

The exchange rate, which started to increase after 2015, reached 14.00 TL in the 11th
month of 2021. Since 2022, when similar problems were experienced, exchange rates and
inflation have increased rapidly, and the Central Bank s reserves have decreased. In order to
prevent investors from turning to foreign exchange deposit accounts due to the increasing
exchange rates, a new application called a Currency Protected Account has been started
by the TCCB in order to support the Turkish Lira Deposit or Participation Account opened
with banks as of May 23, 2022. This method is a type of deposit that aims to give a higher
return by adding the difference of the changes that will occur in addition to the interest in
TL deposit accounts. However, it can be seen from the definition that this method does not
differ from paying interest on interest. It is observed that the rise in exchange rates also led
to an upward movement in the stock markets.

111
Money Supply/M2: The M2/GNP ratio is an indicator of the financial depth of the
country s economy. It is also an indicator of the use of the banking system by the public.
When the money supply increases, interest rates will fall and the prices of bonds and stocks
will rise. Therefore, investments will be directed toward stocks. Then there is a positive
relationship between the two variables. The money supply is a size that includes money in
circulation and time and demand deposits.

M1= Currency in circulation + Demand deposit


M2= M1 + Time deposit

The graph in Figure 9.3 below shows that the relationship between the two variables
is in the same direction and they move together. Following the failure of 19 banks during the
2001 financial crisis, the public s use of the banking sector has increased dramatically in
Turkey. It can be said that, as of January 2022, banking transactions have replaced those on
the stock market.

Figure 9.3. ISE100-M2

IPI (Industrial Production Index): The Industrial Production Index is one of the
indices among the statistics announced by TURKSTAT. It is an index that gives the
industrial production/GNP ratio. It consists of weighting all branches of the industry
according to production classes (85% of manufacturing, 20% of electricity, 20% of mining).
Increasing the industrial production index means more production and the growth of the
Turkish economy. When this index, which measures economic activity, rises, it means that
the Gross National Product (GNP) rises as well, implying that prices have risen. In this case,

112
the value of the stocks also increases. Therefore, there is a positive relationship between
them. The relationship between the two variables is also seen in the graph below.

Figure 9.4. ISE100-IPI

PPI (Wholesale Price Index- Domestic Producer Index): In Turkey, the index
calculated as the Producer Price Index (PPI) since November 10, 2005 and the Domestic
Producer Price Index (D-PPI) since February 26, 2014, replacing the index known as the
Wholesale Price Index (WPI) by the Turkish Statistical Institute, started to be used. This
index, which measures the price level of the country, covers the goods and services that are
subject to sale in the country. Therefore, it is a measure of inflation that shows the general
level of prices. When the Domestic Producer Price Index increases, stock returns are
expected to increase as prices increase. In other words, there is a positive relationship
between the two variables. The graph in Figure 9.5 below shows the direction of the
relationship between the two variables.

Figure 9.5. ISE100-PPI

113
Gold Prices: The monthly selling price in Turkish Lira of gram gold bullion, which
is an alternative investment instrument, has been taken. Although gold has been around for
hundreds of years, it is one of the most preferred alternative investment instruments,
especially in recent years, since it has low risk. Since the variable is included in the analysis
as Turkish Lira, it will be possible to observe how the changes in Turkish Lira affect the
stock market index. The increase in gold prices will lead the investor to turn to gold as an
alternative investment tool. In an inflationary environment, however, investors will prefer a
low-risk, high-return investment instrument. Gold, as a precious metal, has always remained
a safe haven for investors. The graph in Figure 9.6 below shows the direction of the
relationship between the two variables.

Figure 9.6. ISE100-GP

Brent Oil Price: Oil, which is one of the primary energy sources, is the basic and
indispensable input of many sectors. This feature is therefore an important cost factor. The
increase in oil prices will cause cost inflation and create inflationary pressure. Inflationary
pressure will cause prices to rise and the central banks to raise interest rates to suppress price
increases, leading to investors turning to Treasury bills and bonds and thus lowering stock
prices. As can be seen, the components in question are like a loop that has an intricate
interaction structure with each other.

Oil is not only an input but also an investment tool. Because fluctuations in oil prices
cause changes in economic activity and stock prices. After the oil shock experienced in 1973

114
and because of the effects of this shock on the economies of the countries, oil started to be
used as an important factor in economic analysis. It has been observed that changes in oil
prices have a positive effect on the stock market in the long run.

Movements in Brent oil prices affect the country s economies through channels such
as the foreign trade balance, monetary policy, and inflation. Rising oil prices negatively
affect the foreign trade balance of oil-importing countries. It causes a wealth transfer to oil-
exporting countries. Since Turkey is an oil-importing country, when oil prices increase,
money demand increases, and if money supply does not increase in parallel with this
demand, interest rates increase and investment costs increase. The lack of this resource,
which is a very important input for many sectors, especially energy-intensive sectors, causes
a contraction in these sectors. Therefore, increased costs lead to shrinkage and inflation. This
continuity in price increases leads to a decrease in employment, which leads to a decrease in
demand, and a decrease in demand leads to a decrease in production and the continuation of
the spiral cycle. Again, this will lead to changes in the profitability ratios of the companies
and thus the values of the stocks. Therefore, oil-importing countries such as Turkey are very
sensitive to changes in oil prices. The graph below shows the direction of the relationship
between the two variables. In Figure 9.7, the high increases in Brent oil prices experienced
due to the global crisis in 2008 can be seen. It is seen that the price increase in 2011 almost
approached the increase in 2008. Brent oil prices, which had a decreasing trend in 2020,
increased rapidly in 2021 and continued to increase.

Figure 9.7. ISE100-BRT

115
Economic models are established by taking into account the equilibrium relations
envisaged in economic theory. The existence of significant econometric and economic
relations between the variables in the established model also depends on the stationarity of
the series. Financial time series, on the other hand, are not stationary because they generally
show a volatile structure. Therefore, first of all, the variables to be used in the analysis should
be made stationary.

It is an important step to determine the stationarity of the series in the analysis made
with time series, since it may seem like there are relationships that do not exist in the analysis
made with non-stationary series and therefore may cause spurious regression problems.
Whether the statistical tests are meaningful or not in the modeling depends on these analyses.
The Efficient Market Hypothesis for the stock market is tested with unit root tests. If the
ISE100 index is unit-rooted, the efficient market hypothesis is valid; if it is stationary, the
efficient market hypothesis is not valid. Therefore, with unit root tests in financial time
series, the trend of asset prices returning to the mean is investigated
289).

Since the time series of the macroeconomic variables utilized in the analysis have
unit root and seasonality effects by nature, the Generalized Dickey-Fuller Test (ADF) and
Phillips-Perron (PP) unit root tests were used to ascertain if the data sets were stationary or
not.

In classical regression models, the dependent variable is random and the independent
variable or variables are considered constant. When the dependent variable is defined
according to whether a certain feature is present or not, it is converted into a binary structure
as a qualitative variable. In this case, the dependent variable takes the value 0 or 1. Since
these model parameters cannot be estimated by the classical least squares method, the
estimations are made with the linear probability, Logit, and Probit models.

In the study, both Logit and Probit model estimations were made using the ISE100
index value as a dependent (output) variable and seven macroeconomic variables as
independent (input) variables. Since the dependent variable must have a binary structure in
these modeling techniques, a threshold value was calculated by taking the geometric mean
of the series for the dependent variable BIST100 index. The variable was converted into a

116
qualitative or binary structure by assigning a value of 0 to values less than the threshold
value and a value of 1 to values greater than the threshold value. Here, the geometric mean
is preferred because it is a measure of central tendency that takes into account the geometric
differences, rather than the arithmetic differences between the data. Thus, a structure was
created that covers all the observations in the data set, conforms to algebraic operations, and
allows the processing of relative numbers.

The variables used in the study are shown below:

ISE100 BIST 100 Index


ERUSD Exchange Rate (TL/$)
WPI-PPI Domestic Producer Price Index
IPI Industrial Production Index
M2 Money Supply M2
GP Gold TL/gr
TYBY Active Bond Interest Rate or 2-year Government Bond Interest
BRT Brent Oil Price (Brent/Dollar)

Based on this information and the logit and porbit regression model, which was
created with the probability of the index value , the i period being greater or less than
the geometric mean, will be as in Figure 9.8:

= + ERUSD + PPI + IPI + M2 + GP + TYBY + BRT +


Figure 9.8. Logit model

+ ERUSD+ PPI+ IPI+ M2+ GP+ TYBY+ BRT)


Figure 9.9. Probit model

The time series of the variables used in the analysis were compiled as monthly series
from the Central Bank of the Republic of Turkey Electronic Data Distribution System
database (EVDS), the statistics of the Turkish Statistical Institute (TUIK), the monthly
statistical bulletins of the Capital Markets Board, the bulletins of the Ministry of Treasury
and Finance, and the Eurostat databases. Eviews 12 for the statistical and econometric tests
of the time series data of the dependent and independent variables for the period January
117
2001 to January 2022, Stata 15 for Logit and Probit analysis, and MATLAB R2021b for DL
analysis were used.

In the DL model, ISE100 was used as the dependent (output) variable, and seven
macroeconomic variables were used as independent (input) variables. In the DL model,
which works based on the artificial neural network architecture, the data set is divided into
two parts: 70% training and 30% testing. In financial analyses in the literature, AI
applications usually make a distinction between data sets in different ratios, but this
distinction is determined by trial and error as with other parameters of the model (Chong et
al., 2017a; Jing et al., 2021; C.-C. Lin et al., 2018). In this case, the 2001:01 2015:08 period
consists of the first 177 observations as the training set and the last 76 observations as the
2015:09 2021:02 test set.

In order to solve the problems in their studies, researchers use modeling and analysis
by trying to create the best and most explanatory parameter sets to represent the problem. In
the studies carried out as a result of the development of the DL approach, the issues of how
to best design a multi-layer artificial neural network structure, how many layers it will
consist of, how many neurons it will contain, what the dropout value will be, and which
optimization algorithm or activation function to choose have been seen as how important
issues are for the solution. However, since there is no definite proposition in the selection of
these parameters, the parameters were obtained by trial and error according to the problem
and the data set (Jing et al., 2021; Kaastra & Boyd, 1996, pp. 220 224).

In the DL method, which is a ML method that learns from data, the most used
parameters while creating the artificial neural network architecture are explained in detail in
Chapter 7 of the study. Here, we will quickly touch on the key factors that must be taken
into account while determining these characteristics (Deep Learning Toolbox, n.d.):

The size and diversity of the data set: The larger the dataset, the better the model
learns, although this is not sufficient for a good model.

Mini-Batch Size: In each iteration of DL, gradient descent is calculated


retrospectively on the network with the backpropagation process, and the weight values are
updated in this way. The higher the number of variables in this calculation process, the longer

118
the calculation takes. For this reason, the data is divided into small groups, and learning is
done through these small groups. It indicates how many data points the model will process
at the same time by segmenting data or determining the learning mini-batch parameter.

Learning rate and momentum: In each iteration of learning in DL, the difference is
found by backward derivatives with the backpropagation method, and the result obtained by
multiplying the value with the learning parameter is subtracted from the weight values and
new weight values are calculated. Here, the learning rate can be determined as a fixed value,
or it can be learned as a step-by-step value, a momentum-dependent value, or during learning
by adaptive algorithms. However, the learning rate usually defaults to 0.01 and is reduced to
0.001 after a certain epoch or cycle. It is seen that the momentum beta coefficient, which is
generally used, is also 0.9 and varies in the range of 0.8 0.99.

Selection of the optimization algorithm: In DL, learning is an optimization process.


In solving nonlinear problems, algorithms such as stochastic gradient descent (sgdm),
adagard, adadelta, and adam are commonly used. The important feature of adaptive
algorithms is that they learn the learning rate themselves and are dynamic. DL generally uses
the sgdm algorithm by default.

Epoch number: While the model is being trained, the data is trained and tested in
parts, not completely, and the weights are updated by back propagation according to the
result. This is the name given to each of the training steps.

Weight value and activation function: There are various weight determination
methods that affect the learning speed of the model. It is explained in detail in Section 6.3.2.
In multilayer artificial neural network models, activation functions such as sigmoid, tanh,
and ReLu are mostly used in hidden layers for nonlinear transformation operations. It is used
for backing derivatives in hidden layers.

Dropout value: It has been observed that forgetting weak data below a certain
threshold in fully connected networks increases the success of learning. In general, the value
taken as 0.5 is defined as a value in the range [0,1] when taken as the threshold value.

119
Layer, number of neurons in each layer, and pooling: The feature that distinguishes
DL from ANNs is the high number of layers. That s why it s called deep learning As the
number of layers increases, other features of the data are learned during the learning stages,
so learning is more successful. The number of neurons indicates the amount of information
stored in memory. The better the computer technology and processor, the larger the size of
the kernels that operate on the matrix in each layer, so the pooling process made on the
output of the kernels, in other words, the filtering technique for the data, is also successful.

120
10. EMPIRICAL ANALYSIS

In this section, the models used during the study are introduced. In order to measure
the change in the value of Istanbul Stock Exchange Index by the change in the selected
macroeconomic variables, Logit and Probit models from qualitative response regression
models and a DL model from AI technologies will be estimated. Thus, the effectiveness of
these modeling techniques will be tested.

Istanbul Stock Exchange (BIST100 Index) index values and selected 7


macroeconomic variables for the period 2001:01-2022:01 in Turkey were subjected to
statistical analysis and Eviews12, STATA15 and MATLAB R2021b programs were used in
the analysis.

10.1. Descriptive Statistics

Descriptive statistics are statistics that describe the general characteristics of a data
set statistically, summarize the series, show typical values at which units are stacked, and
provide information about distributions.

Standard deviation is the most widely used measure of dispersion. Because the
standard deviation expresses how close the observations in the data set are to the mean.
Usually, the small standard deviation value indicates that the deviations from the mean are
small. In the opposite case, when there is a large standard deviation, it is understood that the
data diverge from the mean. The standard deviation is obtained by dividing the sum of the
squares of the deviations from the arithmetic mean, which is the variance, to the total number
of observations, and taking the square root of this value. In summary, the standard deviation
is equal to the square root of the variance.

Summary statistics for the original series dependent and independent variables are
given in Table 10.1 below. Analyzing the summary statistics, it is seen that macroeconomic
variables are exposed to fluctuations in the 2001:01 2022:01 period, so the variability is
high. It is seen that the series are skewed to the left or inclined to values higher than the
average, also due to the effects of the economic crisis. This shows that the financial series

121
are quadratic series. While the averages of macroeconomic variables are different, their
standard deviations also show significant differences.

Table 10.1. Some descriptive statistics for the dependent and independent variables

ISE100 BRT GP ERUSD TYBY IPI M2 PPI


Mean 629.4487 65.74174 116.4266 2.780042 21.96500 81.86682 9.86E+08 237.5752
Median 618.5821 63.52000 80.32000 1.712764 14.32000 77.58710 6.49E+08 189.6186
Maximum 2003.200 138.4000 817.6100 13.55286 276.1500 165.5610 5.14E+09 1129.030
Minimum 76.25870 14.85000 5.882500 0.673858 5.200000 36.35654 32202029 37.04517
Std. Dev. 383.8759 28.93278 139.6022 2.227352 25.78224 28.03729 1.06E+09 169.4207
Skewness 0.605083 0.353347 2.275737 2.054375 4.995143 0.369554 1.666464 2.016016
Kurtosis 3.232937 2.206649 8.639739 7.463964 41.35857 2.300925 5.591807 8.236685

Jarque-Bera 16.01029 11.89966 553.6755 388.0260 16562.91 10.91049 187.9144 460.4615


Probability 0.000334 0.002606 0.000000 0.000000 0.000000 0.004274 0.000000 0.000000

Sum 159250.5 16632.66 29455.92 703.3505 5557.144 20712.30 2.50E+11 60106.54


Sum Sq. Dev. 37134889 210950.7 4911174. 1250.197 167510.5 198094.6 2.81E+20 7233247.

Observations 253 253 253 253 253 253 253 253

10.2. Unit Root Test Results

Time series are numerical quantities in which the values of the variables are observed
consecutively from one period to the next .A
time series is considered stationary if it does not continuously increase or decrease over a
given time period and the data scatter around a horizontal axis over time. In other words, a
time series is stationary if its mean, variance, and covariance do not change systematically
.

The stationarity of the series is a necessary assumption for efficient and consistent
estimations. However, economic time series tend to increase over time, so they are not
stationary. Regression model assumptions require both dependent and independent series to
be stationary and errors to have a zero mean and finite variance. Otherwise, spurious
regressions may occur with regression analysis on a non-stationary series. While there are
-statistics in the spurious regression, the parameter estimation
results are meaningless in terms of economic interpretation. If there is cointegration between
the variables, even if the variables are not stationary when examined individually, taking the
linear functions of these variables causes the series to become stationary by removing the

122
trend of each other and making the spurious regression disappear. Co-integration or
cointegretation means that the error term obtained from linear combinations of non-
stationary series is stationary. For cointegration to occur, all series must be equally
stationary, and the error term obtained from the linear functions of these variables must also
be stationary. This correction mechanism also includes time series AR (Autoregressive)
corrections and MA (Moving averages) corrections for DF and ADF models. Therefore, we
can say that the PP test is an ARMA (Autoregressive Moving Average) process.

While there are various methods for examining the stationarity of time series, the unit
root test is one of them. A unit root in a series means that the series is not stationary. The
most commonly used unit root tests for determining stationarity are the Dickey-Fuller Test
and the Generalized Dickey-Fuller Test (ADF, Augmented Dickey-Fuller, 1979), developed
as an extension of it. In addition, the Phillips Perron (1989) Test, which is created by adding
a non-parametric addition, is used to correct the error terms, considering that these tests are
insufficient when the economic time series are subjected to a break (Wooldridge, 2010, pp.
640 642). This correction mechanism also includes time series AR (Autoregressive)
corrections and MA (Moving Averages) corrections for DF and ADF models. Therefore, it
can be said that the PP test is an ARMA (Autoregressive Moving Average) process (Phillips
& Perron, 1988, pp. 335 346).

The unit root tests and hypotheses used in the study are as follows:

Augmented Dickey-Fuller (ADF) Unit Root Test

The following process is used to explain the use of the test. The starting point is the
probabilistic unit root process, which is an autoregressive process.

Figure 10.1. The probabilistic unit root process

Here, is the error term of the white noise. If =1, which means the series contains
a unit root, the above model will be a non-stationary probabilistic process. Therefore, since
we cannot predict the coefficient in question with Ordinary Least Squares (OLS) estimators,

123
when we subtract from both sides of the model, Figure 10.2 shows -
processor;

Figure 10.2. ADF random walk equation

This model is predicted and tested with the following hypothesis.

: =0 There is a unit root in the series.


: There is no unit root in the series.

For a =0.05 significance level, if, P , is rejected and is said to be the series is
stationary. is calculated and the absolute value of the
calculated value exceeds the McKinnon DF critical values, the null hypothesis cannot be
rejected .

Phillips Perron (PP) Unit Root Test

It is stated that the Dickey-Fuller Test error terms are statistically independent and
have constant variance. In the extended Dickey-Fuller unit root test, the lagged values of the
dependent variable are added to the independent variable and the autocorrelation problem in
the error term is eliminated. The Phillips-Perron Unit Root Test uses non-coefficient or non-
parametric statistical methods without adding lagged difference values to account for the
sequential relationship, i.e. autocorrelation, in the error term (Gujarati & Porter, 2009, p.
758).

The critical values of the Phillips-Perron (PP) test and the critical values of the ADF
test are the same. The unit root test developed by Phillips-Perron (PP) adds new assumptions
to Dickey-Fuller s error term. The regression equations of Phillips-Perron are expressed
below (Phillips & Perron, 1988, pp. 335 346).

+ + ve

Figure 10.3. PP model

124
Here shows the white noise sequence process with an expected mean of zero and
the number of observations. No serial correlation or assumption of
homogeneity between error terms is required here. Phillips-Perron (PP) produced test
statistics to test the hypotheses about the and coefficients. In Dickey-Fuller unit root
tests, is used to test a model and in Phillips-Perron test it is expressed as
366).

Economic models are established by taking into account the equilibrium relations
envisaged in economic theory. The existence of significant econometric relations between
the variables in the established model also depends on the stationarity of the series. In a
stationary process, the series fluctuates around a fixed long-term average and the effect of
any shock is not permanent. On the other hand, a non-stationary series will not enter the
long-term deterministic path and will permanently reflect the effects of shocks in the short-
term to the long-term values. Time series are not stationary because they usually show a
volatile structure. Therefore, the variables to be used in the analysis must first be made
stationary.

Figure 10.4. Original graphics of variables (level)

In order to determine the stationarity of a time series, inferences are made with the
help of graphics. As can be seen in Figure 10.4, where the original graphics of the series of

125
the variables are given, it is seen that all the dependent and independent variables used in the
regression models are not stationary, in other words, the series contain unit roots. Trend or
seasonal mobility was observed in each variable.

Although there are many different unit root test techniques, the most popular ones
are the ADF (1981) and PP (1988) unit root tests for determining whether financial time
series are stationary. The stationarity test of the series was carried out by using three unit
root test models: the fixed model, the fixed and trended model, and finally the none model
without the constant term and trend. The results obtained from the stationarity analysis are
shown in Table 10.2.

Table 10.2. Unit root test results at level

Variables T-Statistics ADF T-Statistics PP Critical Value


Trend & Intercept Trend & Intercept 1% 5% 10%
ISE100 -0.5926[0] -0.6834 -3.9967 -3.4286 -3.1377
TYBY -5.4044[15] *, **, *** -6.4646 *, **, *** -3.9967 -3.4286 -3.1377
ERUSD 4.7760[12] 4.3645 -3.9967 -3.4286 -3.1377
PPI 3.7943[6] 9.58833 -3.9967 -3.4286 -3.1377
IPI -1.80939[12] -10.6153*, **, *** -3.9967 -3.4286 -3.1377
M2 7.86595[12] 10.3237 -3.9967 -3.4286 -3.1377
GP 7.2818 [10] 7.9621 -3.9967 -3.4286 -3.1377
BRT -2.64021[0] -2.02866 -3.9967 -3.4286 -3.1377
Note: All three models were tried in the level values and first order differences of the series and constant term and trend
models were used in model selection as long as they were meaningful. The values in square brackets represent the
appropriate lag length of the variables determined according to the SIC. The minimum lag length at which autocorrelation
is removed was chosen. *, ** and *** indicate stationarity at 1%, 5% and 10% significance levels, respectively.

When the test statistics in Table 10.2 are examined, it is seen that all of the ADF and
PP coefficients of the original series of the dependent and independent variables are not
It has been seen that the series where all variables
(except the interest rate of the active bond) are not stationary at level contain a unit root at
level. It is seen in Figure 10.4 above that the graphs of the series also support this result. In
the analyses made, it was seen that when the first-order differences of the original series
were taken, the series did not become stationary and the integration degrees were different.
For this reason, stationarity analyses were performed again by taking the natural logarithms
of the series.

126
If a time series becomes stationary when it is differentiated d times, the series is said
to be d-integrated and is denoted as I(d). When the first differences of the series, except the
interest of the active bond, are taken in the analysis, it is concluded that not all of them are
integrated at the same level, the PPI series is integrated at I(2), and the other series are also
integrated at the I(1) level. If the series is not stationary, the future prediction is not made by
making generalizations. While the series are not singularly stationary, their linear
combinations can be stationary. For this reason, after the ADF test, the Engle-Granger test
was performed, and the ADF test was performed on the series consisting of the residues
obtained from the regression and the null hypothesis was rejected. Therefore, according to
Engle-Granger, the series are cointegrated. Therefore, the regression will be a regression
showing a real relationship.

Table 10.3. Unit root test results at ln

Variables T-Statistics ADF T-Statistics PP Critical Value


Trend & Intercept Trend & Intercept 1% 5% 10%
lISE100 -2.7080[0] -2.5823 -3.9951 -3.4279 -3.1373
TYBY -5.4044[15] *, **, *** -6.4646 *, **, *** -3.9951 -3.4279 -3.1373
lERUSD -0.0178[2] -0.3388 -3.9951 -3.4279 -3.1373
lPPI 0.1964[2] -1.5855 -3.9951 -3.4279 -3.1373
lIPI -2.7569[12] -9.5753 *, **, *** -3.9951 -3.4279 -3.1373
lM2 -2.1832[0] -2.1813 -3.9951 -3.4279 -3.1373
lGP -3.0737[1] -2.4055 -3.9951 -3.4279 -3.1373
lBRT -2.8523[0] -2.3922 -3.9951 -3.4279 -3.1373

dlISE100 -17.4022[0] *, **, *** -17.5399*, **, *** -3.9951 -3.4279 -3.1373
dlERUSD -12.1062[1] *, **, *** -9.8556*, **, *** -3.9951 -3.4279 -3.1373
dlPPI -7.2550[1] *, **, *** -5.6983 *, **, *** -3.9951 -3.4279 -3.1373
dlIPI -6.1426[11] *, **, *** -87.3719 *, **, *** -3.9951 -3.4279 -3.1373
dlM2 -16.4110[0] *, **, *** -16.4004 *, **, *** -3.9951 -3.4279 -3.1373
dlGP -11.4887[1] *, **, *** -11.3133 *, **, *** -3.9951 -3.4279 -3.1373
dlBRT -13.7588[0] *, **, *** -13.8688 *, **, *** -3.9951 -3.4279 -3.1373
Note: All three models were tried in the level values and first order differences of the series and constant term and
trend models were used in model selection as long as they were meaningful. The values in square brackets represent
the appropriate lag length of the variables determined according to the SIC. d stands for first-order difference. The
minimum lag length at which autocorrelation is removed was chosen. *, ** and *** indicate stationarity at 1%, 5%
and 10% significance levels, respectively.

When the test statistics in Table 10.3 are examined, it is seen that all of the ADF and
PP coefficients of the first-order differences of the dependent and independent variables are
It has been determined that all variables are
127
stationary at the I(1) level and their series do not contain unit roots. All series are I(1) co-
integrated, and the Phillips-Perron unit root test results also confirm the ADF test results.
For this reason, the first-order differences of the variables are taken into account for the
estimation of the models.

Table 10.4. Correlation coefficient between the independent variables


TYBY BRT ERUSD GP IPI M2 PPI
TYBY 1 -0.03312 0.24945 0.1678 -0.00835 0.13426 0.36378
BRT -0.03312 1 -0.09913 -0.0337 0.00648 -0.04088 0.13353
ERUSD 0.24945 -0.09913 1 0.733 0.03381 0.21283 0.60858
GP 0.16779 -0.03367 0.73304 1 -0.06753 0.26173 0.48518
IPI -0.00835 0.00648 0.03380 -0.06753 1 0.09101 0.00243
M2 0.13426 -0.04088 0.21283 0.2617 0.09102 1 0.09195
PPI 0.36378 0.13353 0.60858 0.4852 0.00244 0.09195 1

As seen in Table 10.4, the correlation coefficients between the independent variables
remain below 0.73 and mostly demonstrate weak linear relationships.

10.3. Logit ve Probit Models and Results

In the linear probability model, the conditional expected value of the dependent
variable is the conditional probability of the event occurring given the independent variable
values. The linear probability model can be estimated by the classical least squares method.
However, the error term is not normally distributed, the error term has heteroscedasticity,
the estimation values can fall outside the range of 0 1, and the error term generally has a
low value. For this reason, Logit and Probit models have been developed as alternative
models to get rid of these problems (Bengio et al., 2015, p. 584; Gujarati & Porter, 2009, pp.
552 553). The difference between the two models is that the Logit curve has a thicker tail
or approaches the axis (0 or 1) more slowly than the Probit curve. The difference in the
coefficients obtained from these models is due to the different functions used for
probabilities (Gujarati & Porter, 2009, p. 571). The issue of which model will be selected in
the application is presented according to the preference of the person who will use the model.

The parameters of the Logit and Probit models are estimated using the maximum
likelihood method. In order to use this method, it is assumed that the probability distribution
of the error term is in accordance with the normal distribution. In this case, when the series
128
are stationary, the normal distribution assumption regarding the error term is fulfilled. Before
estimating the parameters of Logit and Probit regression models, which are qualitative
dependent response models, a stationarity analysis for the variables used in the model is
required.

Because the parameters of the Logit and Probit models cannot be estimated using the
classical least squares method, the maximum likelihood method is used. In this method, the
standard normal distribution which is the Z test statistic, is used for the statistical significance
of the coefficients. In order to test the co-significance of the coefficients in the Logit and
Probit models, Wald, Score (Lagrange Multiplier, LM), and Likelihood Ratio (LR) tests are
used. In the analysis, the standard normal distribution Z statistic was used to test for the
statistical significance of the coefficients, and the LR statistic was used to test the general
significance of the models.

, the coefficient of determination, which is used as a measure of goodness of fit or


how explanatory the model is in classical linear regression models, is not a good measure
for Logit and Probit models. Because these measures are not based on variance, they are not
an accurate measure of whether the model is well explained or not, as they are calculated
based on the change in the values of similarity ratios. The correct classification rate obtained
from the classification table was also used to evaluate the goodness of fit of the models
created. The obtained probability values are classified according to the determined cutoff
point, and it is predicted which of the 0 or 1 values each unit will take. A value of 0.50 is
usually taken for the cutoff point. If the probability value is greater than this value, the unit
is assigned to group 0, and if it is less than this value, the unit is assigned to group 1.

In binary logit-probit models, the STATA program calculates the Pseudo value
as a measure of goodness of fit, and the Eviews12 program calculates the McFadden .
Pseudo- was used in this study.

The coefficients obtained in the models made with Logit and Probit techniques
cannot be interpreted directly as in the linear regression model. However, the sign of the
examined coefficients indicates the direction of the relationship between the independent
variables and the probability of the event occurring. If the sign of the coefficient is negative,
there is an inverse relationship; if it is positive, there is a direct relationship.
129
In this study, both Logit and Probit model estimations were made using the
macroeconomic variables in question, and for this, a threshold value was calculated for our
independent variable, the BIST100 index. Thus, the transformation of the dependent variable
into a binary variable with a qualitative response was achieved. The geometric mean of the
BIST100 variable was taken and converted to a binomial to give a value of 0 for values
below it and 1 for values above it. The geometric mean is a measure of central tendency that
takes into account geometric differences between data, rather than arithmetic differences.
The geometric mean, which covers all the observations in the data set and conforms to
algebraic operations, allows the processing of relative numbers.

While investigating whether the variables used in this study are suitable for the Logit
model and the Probit model, first of all, the dependent variable for each independent variable
was added to the Logit and Probit models separately, and the relationship between the
dependent variable and the independent variables was observed. The test results obtained in
this context are given in Tables 10.5 and 10.6 below.

Table 10.5. Logit model compatibility results

BinaryISE100
Log- LR Prob
Variables Coefficient Prob Likelihood Statistic (LR St.)
TYBY 0.763828 0.0000 -94.087715 147.70 0.0000
LnERUSD 24294.1 0.0000 -82.945841 169.99 0.0000
LnPPI 553424.8 0.0000 -40.502838 254.87 0.0000
LnIPI 4.48e+07 0.0000 -44.480443 246.92 0.0000
LnM2 1.33e+50 0.0000 -38.382249 259.11 0.0000
LnGP 136.1367 0.0000 -41.043973 253.79 0.0000
LnBRT 12.36735 0.0000 -134.60854 66.66 0.0000

Table 10.6. Probit model compatibility results

BinaryISE100
Log- LR Prob
Variable Coefficient Prob Likelihood Statistic (LR St.)
TYBY -0.160051 0.0000 -93.265235 149.35 0.0000
LnERUSD 5.747567 0.0000 -83.387448 169.10 0.0000
LnPPI 7.528884 0.0000 -39.850861 256.18 0.0000
LnIPI 9.398173 0.0000 -44.761852 246.36 0.0000
LnM2 66.07681 0.0000 -37.872111 260.13 0.0000
LnGP 2.776939 0.0000 -40.080625 255.72 0.0000
LnBRT 1.452791 0.0000 135.30463 65.27 0.0000

130
As can be seen in the Table 10.5 and Table 10.6 above showing the compatibility
results, each variable is compatible with the dependent binary variable ISE100 for the Logit
and Probit models. A 1% change in each variable causes the ISE100 to take a higher or lower
value than its geometric mean. In other words, each variable in the model is an important
variable that explains the dependent variable and is statistically significant. It was also in
line with our expectations economically. Therefore, our variables are suitable for both the
Logit and Probit models. However, since the changes in the interest rate and gold prices in
the Probit model are in line with expectations, this model is considered more appropriate.

In this study, the Logit and Probit models were estimated with the highest likelihood
method, taking into account the obtained data and evaluations, and the results are
summarized in Table 10.7. It was seen that the variables used in the models were statistically
significant according to the 1%, 5%, and 10% significance levels and were important
variables affecting the dependent variable.

To test the overall significance of the Logit and Probit models, Hosmer-Lemeshow
and Pearson goodness-of-fit tests, which are generally right-tailed tests, were performed. The
Akaike Information Criterion (1973), which is an indicator of the goodness of fit of an
estimated statistical model, and the Schwartz or Bayesian Information Criteria (1978), which
enables model selection among a group of parametric models with different numbers of
parameters, were also calculated. According to the these test, the model with the smallest
AIC or BIC value is the best model. In addition, the correct classification rate was found to
be 95.6 percent and 94.8 percent, respectively.

In the Logit and Probit models, since each of the coefficients obtained (log-odds) is
a log-bet value, they cannot be interpreted directly as in linear regression models. However,
the signs in the coefficients obtained indicate the direction of the relationship between the
variables and the probability of the event occurring. If the coefficient obtained is negative,
the relationship between the variables will be in the opposite direction; if it is positive, the
relationship will be in the same direction.

131
Table 10.7. Results of the Logit model & the Probit model

Logit Model Probit Model


Variables Odss Ratio Prob Coefficient Prob
c 8.1e-277 0.133*** -321.2098 0.066***
TYBY 0.540201 0.004* -0.362669 0.002*
LnERUSD 13.79592 0.765*** 2.213481 0.611***
LnPPI 9.424175 0.904*** 1.408437 0.872***
LnIPI 2.43e+08 0.005* 10.29269 0.004*
LnM2 4.50e+85 0.231*** 98.83031 0.145***
LnGP 0.000015 0.066*** -6.232234 0.056***
LnBRT 2.109222 0.872*** 0.7311821 0.757***
Log-Likelihood -22.420721 -22.222259
LR Statistic 291.04 291.43
Prob (LR St.) 0.0000 0.0000
Pseudo R2 0.8665 0.8677
Hosmer-Lemeshow 3.52 0.50
chi2 (Prob>chi2) (0.8974) (0.9999)
Pearson chi2 64.49 53.70
(Prob>chi2) (1.0000) (1.0000)
AIC 60.84144 60.44452
BIC 89.10856 88.71163
Correct Classification
Rate 95.65% 94.86%

Note: *, **, ***; represent 1%, 5%, and 10% significance levels respectively.

As seen in Table 10.7, the coefficients obtained in the Logit and Probit models are
statistically significant at 5% and 10% significance levels, respectively. When the
coefficients of the variables were evaluated one by one, the results were obtained in a way
that met our expectations. While an inverse relationship was expected with the interest rate
of the active bond and gold prices, it was observed that this expectation was met and that
there was a positive relationship between other variables, again meeting the expectations. As
the interest rate of the active bond increases, the return will also increase, so the demand of
the investor shifts from the stock market to this instrument. The investor will show the same
reaction when gold prices increase and will choose gold as a safe haven. In other words,
there is a negative relationship between the stock market, the two-year bond rate, and gold
prices. It is seen that a 1% increase or decrease in other macroeconomic variables included
in the model causes the dependent variable to take higher and lower values than the average,
depending on the coefficients of the variables, by increasing or decreasing the dependent
variable at different rates.

132
It is seen that these results are confirmed by four separate tests calculated for models,
showing whether the model is completely significant or not. Since the probability values of
Hosmer-Lemeshow and Pearson tests are higher than 0.05, our models are statistically
significant. Again, AIC and BIC criteria are also undervalued. Accurate classification rates
were found to be 95.65% for the logit model and 94.86% for the probit model. In addition,
considering the LR test statistic and the probability values for this test statistic, it is seen that
all coefficients in both the Logit and Probit models are significant together. Also, the Pseudo
R2 values, which are the output of the STATA program and show the significance of the
whole model, were found to be 0.865 for the Logit model and 0.866 for the Probit model.
The correct classification rate results for the two models are shown in Table 10.8 and Table
10.9.

Table 10.8. Correct classification rate table for the Logit model

Clasisified D F Total
Positive 151 5 156
Negative 6 91 97
Total 157 96 253
Correctly
Classified(%) 96.79 93.81 95.65

Table 10.9. Correct classificaiton rate table for the Probit

Clasisified D F Total
Positive 150 6 156
Negative 7 90 97
Total 157 96 253
Correctly
Classified(%) 96.15 92.78 94.86

In the Probit model, unlike the Logit model, to meet our economic expectations, gold
prices appear as a variable that adversely affects the stock market index as well as the two-
year bond interest. The variable is also statistically significant. Gold is an important
investment tool all over the world and has been seen as a safe haven for investors. It is
evident that many investors are attracted to this area due to the continuous increase in gold
prices in recent years. There is no doubt that speculative attacks in the stock market,
interventions in the capital market, the lack of transparency and reliability of the capital
market, as well as speculative news on this subject, have a negative impact on all investors.
Since all of these factors make small investors nervous, they cause investors to turn to safer

133
investment instruments, and the failure to transform savings into investments leads to a
contraction in the economy. When this mechanism, which has a spiral structure, operates in
the opposite direction, it affects social welfare negatively.

Comparing the performance of the Logit and Probit models, the compatibility test
values of both models are very close to each other, but the main difference between the two
models is that the logistic distribution has a slightly thicker tail. However, there is practically
no reason to prefer one over the other. One is preferred over the other because of its
comparative mathematical simplicity. The correct classification rate varies from 95.65% for
the Logit model to 94.86% for the Probit model, which is quite low at 8 per thousand.
However, the variables in the Probit model were estimated to meet economic expectations.
The statistical tests of the variables included in the model are also an indication of this. In
addition, the Pseudo R2 values of both models were over 86%.

10.4. Deep Learning Models and Analyses

In this study, it was investigated how the DL model, which is a method based on the
working mechanism of AI technologies and is one of the ANNs, should be used to measure
the power of explaining the Istanbul Stock Exchange index values using the selected
macroeconomic variables. It has been shown in many empicial studies that AI applications,
especially in estimating non-linear time series, give as good or better results than well-known
traditional econometric methods, and because of this success, they are used as good
estimation tools (Akel & Karacameydan, 2018; Jing et al., 2021; G. Zhang et al., 1998, p.
40).

10.4.1. Preprocessing of Data Used in the Study

The preprocessing stage of the data means the normalization of the data groups
consisting of training and test data in accordance with the established DL networks.
Normalization is one of the most important factors that determine the success of the DL
model. In addition, the entry of the data into the system in raw or processed form should be
determined according to the researcher s purpose for the problem (Kaastra & Boyd, 1996,
p. 222). Data normalization is the process of compressing the output of the processing

134
elements into the range of [0, 1] or [-1, 1] with the activation functions used in the hidden
layer and the output layer (Altan, 2008; Gomes et al., 2016; G. Zhang et al., 1998, p. 49).

The data used in this study consists of non-linear, seasonally unadjusted, and trend-
containing time series in which seven macroeconomic variables and Borsa Istanbul Index
values are taken as independent variables. For this reason, the series are normalized in the
range of [-1, 1] using the equation ) . The graphs of the normalized
independent variable series obtained are given in Figue 10.5 below.

Figure 10.5. Graphs of variables with normalization

10.4.2. Deep Learning Network Architecture Used in the Study

Significant variables obtained from the meaningful models established as a result of


the analyses made with Logit and Probit regression modeling methods; Dollar Rate,
Industrial Production, Domestic Consumer Price Index, Active Bond Interest Rate, Gold
Prices, Money Supply and Brent Oil Price are used as input variables in the established deep
learning neural network models.

With the DL method based on working of ANNs, many different models can be
produced using different criteria. The performances of the models created are depends on
many parameters, such as the number of inputs, the number of input layers, the number of

135
hidden layers, activation functions, the learning method used, the learning rate, the number
of mini-batches, the number of epochs, weight values, memorization value, and even the
speed of the processor used in the analysis. It is closely related to critical components like
capacity. The most widely used ANN model in the literature is the multilayer perceptron and
the most widely used DL model is LSTM (W. Chen et al., 2017; Cheng et al., 2022a; Jing et
al., 2021; Sarker, 2021).

DL, one of the applications of AI that provides the opportunity to design deep neural
networks with pre-trained models, is a method used on image, time series, and text data.
Convolutional neural networks (ConvNets, CNN) and Long Short-Term Memory (LSTM)
networks can be used to classify and regress data. An LSTM network is a recurrent neural
network (RNN) that processes input data by looping the time steps and updating the network
state. The network contains information remembered at all previous time steps, and using
data from these previous time steps, it predicts the value a time series might take in the next
period. This network learns to predict the value of the next period. In this method, there are
two estimation methods: open and closed loop estimation.

A standard reversible neural network architecture (RNN) contains a single layer,


while LSTM network consists of four communicating layers. The basic core parts in a LSTM
network structure are the sequence input layer and the LSTM layer. In the sequence input
layer, the inputs are serial or time series data in the network. The LSTM layer learns the
long-term dependencies of serial data between time steps. Starting with the serial input layer,
the network architecture consists of the LSTM layer, the fully dependent layer and the output
layer.

In this study, an AI architecture will be created by using a DL algorithm, which is


the last point reached by ML algorithms from AI technologies. Convolutional neural
networks are used in various combinations to form the foundation of modern DL
architectures. The DL architecture has layers of input, convolution, activation, pooling,
memory, full coupling, and classification. In the convolution layer, filters (kernels) are used
to circulate the data in the input layer horizontally and vertically.

There are activation functions such as ReLu, Sigmoid, and Step (Step), which are the
most commonly used in the literature, in the activation layer. The ReLu activation function

136
is used in this study. In the pooling layer, the maximum pooling function, which is the most
used in the literature, is the maximum pooling, minimum pooling, and average pooling value
functions. In the dropout layer, AI algorithms function to make the network forget the data
in order to prevent the data from being memorized during training. For this reason, the
dropout layer is used in network architecture.

Since the model we created is an LSTM model, the full connection (FullConnected,
FC) layer and classification layers in the structure of this model are also used. Thus, the data
is transformed into a one-dimensional vector and passed to the classification layer. As in
ANNs, in the classification layer, which is the last layer in DL algorithms, it evaluates the
data from the previous full link layer and creates the outputs of the network. SoftMax
classifier, which is a probabilistic calculation method, is generally used in this layer. Thus,
values in the range 0 1 are generated for each possible class
4 18).

For an LSTM model, there must be a number of input neurons that match the numbers
of the input data. In the created LSTM network, seven input neurons, 200 hidden units, 50
fully connected layers units, 50% dropout, and an output layer with the same layer as input
data were created. In the training phase, several tries later, using adam and sgdm optimisation
functions, 100 cycles per iteration and minibatch size 20 with an initial learning rate of 0.01
and gradient threshold 1 were used. Each cycle, the data is shuffled. The goal here is to
prevent the network from memorizing. The general structure of the DL prediction models
created according to the parameters used is given in Figure 10.6.

137
LTSM Layer
Input Layer (200hidden
(7 varb.) units) FullyCon1(50) Droupout (50) FullyCon2(50) Output Layer

1 1 1

1
2 2 2
2
TYBY
3 3
3
ERUSD 3
4 4 4

PPI 4 ISE100
5
5 5
IPI
5
6
6
M2
6 7
GP 7
7
7 8
BRT 8 8

Figure 10.6. Simple architecture of deep learning network

Windows 10 Pro 64-bit is installed on hardware that includes an Intel(R) Core (TM)
i7-10510U CPU @ 1.80 GHz (8 CPUs), 2.30 GHz processor, 16 GB of RAM, and an Intel(R)
UHD Graphics graphics card. It was developed in the MATLAB R2021b environment with
the operating system.

138
The parameters that make up the AI architecture of DL models are given in Table
10.10.

Table 10.10. Deep learning network architecture parameters

ANN type Deep Learning LSTM


Number of input layer neurons 7
Number of output layer neurons 1
Number of fullyconnected layers 2
Number of hidden layer neurons 200
Optimization function adam, sgdm
Activation function used in hidden layer ReLU
Activation function used in the output layer Softmax
Optimization algorithm Gradient Descent
Scaling method used Normalisation
0.01
Fullyconnected layer 50
Droupout 50
Gradient threshold 1
Epoch 100
Minibatchsize 20

10.4.3. Training and Test Phase of the Network

While the data used for the solution of the established DL creates the training set, the
test set consists of data that the network did not see during the training process. The test set
tests whether the network learns well. It has been seen in the literature that the training test
set data is determined at various rates, such as 90%-10%, 80%-20%, or 70%-30%. However,
the most preferred rate is 70% 30%, although it varies according to the problem (Chong et
al., 2017a, pp. 187 205; Jing et al., 2021, pp. 5 8; G. Zhang et al., 1998, p. 50). In some
studies, the data set is divided into three groups as training, validation and test sets (Kaastra
& Boyd, 1996, p. 220). In this study, the data set was divided into training and test sets at
different rates, and models were created for each ratio by using the adam and sgdm
optimisation functions. Analyses were made with models created according to these
parameters. The findings are presented in the tables below for comparison.

139
The data set is divided into 70% training and 30% test sets, which is the most widely
used in the literature. The first 177 observation values (2001:01 2015:08 period) were
determined as the training set, and the last 76 observation values (2015:09 2021:02 period)
were determined as the test set. Another reason for determining the test set in this range is
to show whether the global crisis and its seasonal effects, which started in 2008 and showed
their effects until the beginning of 2009, could be comprehended by the model.

As an indicator of whether the established DL model is a good predictor or not, the

section 6.12 and which are frequently used in performance measurements of MLPs, are used.
These metrics are performance metrics such as R2, HKO-Mean Squared Errors, HKT-Sum
Squared Errors, and HKO-Root Mean Squared Errors. The most appropriate network
structure was obtained by comparing the findings in the training and testing phases, taking
into account these performance criteria.

10.4.4. Deep Learning Analysis Results

In a multi-layered AI and DL model, determining parameters such as the number of


hidden layers, the number of neurons in the hidden layer, the activation functions in the
hidden layer and the output layer, and learning algorithms such as the learning rate and
momentum coefficient are very important for the performance of the network. However,
since there is no precise method for determining these parameters, they were determined by
trial and error (Jing et al., 2021, pp. 5 8; Kaastra & Boyd, 1996, pp. 220 224; C.-C. Lin et
al., 2018, pp. 756 764; Wang et al., 2018, pp. 1 17).

The MATLAB R2021b computer language s algorithms were used to build


multilayer DL models using the settings listed in Table 10.7. The best performing models
were selected from the models created by changing the parameters. These models were
created using the adam and sgdm optimization algorithms, respectively, and by changing the
training and test ratios of the data without normalizing the data set used in the study. In the
next step, the data were normalized, and the same process was repeated. The analysis results
of the fourteen DL models obtained are summarized in the tables below. The aim here is to
show the effectiveness of normalization on model performance.

140
Table 10.11. Simulation results for adam

adam Train Test


Train Test
(%) (%) MSE RMSE MSE RMSE
20 80 0.1267 0.3559 0.0851 0.2917
30 70 0.0466 0.2159 0.3496 0.5913
40 60 0.0317 0.1782 0.1555 0.3943
50 50 0.2027 0.4502 0.4081 0.6388
60 40 0.0588 0.2425 0.0749 0.2738
70 30 0.1245 0.3528 0.3409 0.5839
80 20 0.1379 0.3714 0.3697 0.6081

The effectiveness of each of the seven DL models developed was evaluated using the
MSE and RMSE performance metrics. According to the results in Table 10.11, DL networks
were created with an unnormalized data set using the man optimization algorithm, and as
expected, the performance values differed according to the data set partitioning. The training
set reached the smallest MSE value when it was 40%. However, the test set did not reach the
global minimum, but the performance indicator values were high.

Table 10.12. Simulation results for sgdm

sgdm Train Test


Train Test
(%) (%) MSE RMSE MSE RMSE
20 80 0.0481 0.2195 0.0361 0.1901
30 70 0.0279 0.1672 0.0992 0.3150
40 60 0.3653 0.6044 0.5380 0.7334
50 50 0.0287 0.1695 0.0830 0.2881
60 40 0.0334 0.1829 0.1357 0.3684
70 30 0.1222 0.3496 0.1975 0.4441
80 20 0.0547 0.2339 0.1981 0.4451

As can be seen in Table 10.12, seven DL network models were created using the
sgdm optimization algorithm with the unnormalized data set. The training set reached the
smallest MSE value when it was 40%. However, the test set did not reach the global
minimum, but the performance indicator values were high. According to these two tables,
the performance values of DL network models are high even if the optimization algorithm
is changed, if the analysis is performed without normalizing the series.

141
Table 10.13 below provides a summary of the outcomes of the examination of DL
models, which were developed by normalizing the research's data set in accordance with the
normal distribution and employing the adam and sgdm optimization algorithms.

Table 10.13. Deep learning results

Train Test
Train Test
MSE RMSE MSE RMSE
(70%) (30%)
adam 0.0440 0.2099 0.1396 0.3736
sgdm 0.0448 0.2117 0.1444 0.3800

As can be seen from the tables above, the system that gives the best results is a linear
system, and the best results are achieved by normalization.

Figure 10.7. Training progress with adam

Figure 10.6 shows how effective learning is when the optimization algorithm
determines the adam algorithm and the number of epoch as 100 during the training phase of
the obtained six layers LSTM DL model (while taking into account the other determined
parameters). As the number of epoch increases, the error between the original series
gradually decreases and progresses towards the global minimum.

142
The graphs in Figure 10.8 below show the training and test curve graphs obtained
from the analysis using the adam optimization algorithm. It is seen that the complexity of
the system increases as the degree of mathematical expressions of the regression models
chosen to model the training and test curves in the graphs increases, and thus this turns into
an advantage. Thus, an obscure system becomes identifiable with a simpler expression.

Figure 10.8. Train and test results with adam

It can be seen that the two fully connected layered LSTM models aid in the
simplification of complex problems. It is also seen that the mathematical equation of the
LSTM models has a fourth-order function, and the R2 value is quite high at 0.95. This shows
that the model is completely significant and the explanatory power of the variables is quite
high. In other words, the network showed a very successful performance during the training
and testing phase. The model is a good forecasting model, and the margin of error in the
estimates was 0.0440.

143
Figure 10.9. Training progress with sgdm

Figure 10.9 shows how effective learning is when the optimization algorithm
determines the sgdm algorithm and the number of epoch as 100 during the training phase of
the obtained six layers LSTM DL model (while taking into account the other determined
parameters). As the number of epoch increases, the error between the original series
gradually decreases and progresses towards the global minimum.

The graphs in Figure 10.10 below show the training and test curve graphs obtained
from the analysis using the sgdm optimization algorithm. It is seen that the complexity of
the system increases as the degree of mathematical expressions of the regression models
chosen to model the training and test curves in the graphs increases, and thus this turns into
an advantage. Thus, an obscure system becomes identifiable with a simpler expression.

Figure 10.10. Train and test results with sgdm


144
As can be seen in the tables and figures given above, a deep neural network
architecture with the best performance was created as a result of the trials using various
parameters. It is an optimization process based on the learning process in DL applications.
In DL models, the sgdm (stochastic gradient descent with momentum) optimization
algorithm is used hypothetically. However, in this study, the financial time series was
modeled both unnormalized and normalized, using both optimization algorithms. Model
performances are compared. It has been seen that the optimization algorithm, which is slower
than the global minimum and has a lower margin of error, is sgdm and predicts the models
with the minimum error. According to the results obtained, it was seen that the selected
macroeconomic variables were important variables explaining the change in the ISE100
index values and were statistically significant.

The correctness and sufficiency of the learning of an artificial neural network is tested
with the test set. According to Figure 10.8 and Figure 10.10, which give the training and test
phase results, the trainig and test phase has been successfully completed. As can be seen
from these graphs, which show the estimation values alongside the real values, the estimation
errors are quite low and gradually approach zero. In order to measure the prediction success
of the deep neural network models created after the training and testing phases, the findings
in Table 10.14 or, and root mean square
error.

Table 10.14. Deep learning performance results

Train Test
MSE RMSE MSE RMSE
adam 0.9501 0.0440 0,2099 0,9156 0,1396 0,3736
sgmd 0,9759 0,0448 0,2117 0,9919 0,1444 0,3800

When the findings in Table 10.14 and the graphs showing the results of the training
and testing phases are evaluated together, it is seen that the error margins of the networks in
the training and testing phases are quite low and approach zero. In other words, it is seen
that multilayer LSTM deep neural network models perform quite successfully in the in-
sample and out-of-sample periods. During the in-
multilayer deep neural network model developed to estimate the ISE100 index value was
97.59%, while the MSE value was at the lowest level at 0.0448.

145
Out-of-sample predictions are of greater importance in the comparison of prediction
accuracies, as they consist of predictions from new data that are not used in model prediction.
Considering this situation, it is seen that the created multilayer deep neural network model

at a very high level of 99.19%. This model s MSE value was at its lowest with 0.1444. In
general, it is seen that there is a strong and positive relationship between the independent
variables and the ISE100 value. It is seen that this relationship between the variables is over
50% in all variables, and especially the MSE is at a very low level.

Although the variables were not seasonally adjusted and dummy variables were not
used during the creation of the models, the deep neural network models performed quite
successfully. The deep neural network models developed in general have a good
understanding of the effects of the economic crisis in Turkey in 2001 and the global
economic crisis that emerged in the USA in mid-2007 and adversely affected all world
economies in 2008 on macroeconomic variables and the ISE100 index.

The findings in Table 10.14 and the training and testing phase graphics given above
show that the deep neural network models created gave very good results and that the
networks successfully completed the training and testing phases. The positive performance
indicators show that deep neural network models can be used for prediction. However, the
prediction results will be presented comparatively in the last section.

10.5. Comparison of Model Results

Time series are numerical quantities in which the values of the variables are observed
consecutively from one period to the next .
Statistically, the economic time series has a structure consisting of trend, seasonality,
cyclical, and random movements (Kennedy, 1998, p. 288). Therefore, they are not stationary.
It is very important for decision makers to be able to predict the future performance and
behavior of economic time series due to uncertainties in the economy. In this case, the model
established for a time series based on the observed values is expected to have a performance
that can predict the possible values that the series may take in the future (one day, one month,

146
or one year later). In other words, the forecasting performance of the model is expected to
be high.

Forecasting modeling is of great importance and widely used in the field of


economics, as well as in many other fields. Forecasting modeling plays an important role in
determining the future behavior of both governments and producers, consumers, and finance
sectors, which are called micro-decision makers. It is of great importance that good forecasts
can be made, especially when making the right decisions in the field of finance. Simply put,
good foresight leads to good decisions.

In many empirical studies in the literature, econometric models have been used as
predictive models, and AI and deep neural network models have been used in comparison
with econometric models in recent years due to their high prediction performance. This is
because AI models can outperform and even outperform classical models in predicting
nonlinear, seasonal, and trending time series.

In this part of the study, the comparison of Logit, Probit, and DL methods has been
made. The comparison was made based on the estimation performances of the methods used.
The performances of the Logit and Probit models with the best performance were examined,
and the prediction results obtained from the regression analysis are presented comparatively
in Tables 10.15 and Table 10.16.

Table 10.15. Logit model performance

Clasisified D F Total
Positive 151 5 156
Negative 6 91 97
Total 157 96 253
Correctly Classified (%) 96.79 93.81 95.65
Sensitivity (%) 96.18
Specifivity (%) 94.79
Error Rate (%) 4.35

147
Table 10.16. Probit model performance

Clasisified D F Total
Positive 150 6 156
Negative 7 90 97
Total 157 96 253
Correctly Classified (%) 96.15 92.78 94.86
Sensitivity (%) 95.54
Specifivity (%) 93.75
Error Rate (%) 5.14

In this part of the study, the comparison of Logit, Probit and DL methods has been
made. The comparison was made based on the estimation performances of the methods used.
The performances of the Logit and Probit models with the best performance were examined,
and the prediction results obtained from the regression analysis are presented comparatively
in Tables 10.15 and Table 10.16.

Table 10.17. DL model performance

Accuracy Loss Total(%)


Accuracy (%) 99.61 0 100
Loss (%) 0 0.39 100
Total (%) 99.61 0.39 100

The DL model learned the in-sample period series correctly at a rate of 98.81% with
a loss of 1.19% during the training phase. In other words, the rate at which the model learned
about the crises experienced in Turkey in 2000 and 2001 and the reflections of the 2008
global crisis is 98.81%.

As can be seen from Table 10.17, the DL model correctly learned the in-sample
period series and predicted 100% of the periods in the test data set for the out-of-sample
prediction. After 2015, the effects of the developments in the markets due to the political
events in July 2016 in Turkey and the price increases due to the COVID-19 pandemic in
2019, and their reflections on the market indicators, have been estimated at 100%. In other
words, the validation rate of the model is 100%.

Although the validation rates of the created models are high, the validation
percentage of the DL model is 100%. The margin of error is zero. In other models, the margin
of error is between 4.35% and 5.14%.

148
11. ARGUMENT

After the rapid developments in computer technologies, the use of AI technologies


has become widespread in the field of finance, as well as in many other fields. The reason
for the increase in interest in AI is that artificial intelligence technologies ability to
comprehend the problem, read the data correctly, and solve it quickly is quite advanced.
Because of these characteristics and following its success in many fields such as health,
defense, and engineering, it has begun to be used in the field of finance since the great
financial crisis of 2008. Especially for policymakers and investors, it has started to turn
towards methods that will ensure the elimination of financial uncertainties.

The effectiveness of AI has been proven once again in this study, in which the DL
method, one of the new generation of AI technologies, was tested alongside traditional
statistical and econometric methods such as the Logit and Probit regression modeling
techniques. In this study, it was discovered that AI technologies outperform classical
econometric methods in the analysis of financial time series with a dynamic and volatile
structure. Changes in the seven selected macroeconomic variables will affect the future
values of the ISE100 index in which direction and strength. In fact, studies have shown that
even new modulation techniques developed by AI technologies outperform each other in
estimating each other. The new formations that have occurred all over the world in recent
years, the ongoing economic crises since 2008, the COVID-19 pandemic that started in 2020,
the military crises between Russia and Ukraine in 2022, have actually revealed the necessity
of evaluating the world as a whole, not as countries and borders.

These developments have demonstrated that a positive or negative situation in any


part of the world has reached global proportions, affecting all countries. This situation has
made all the components that countries face in maintaining or increasing the level of social
welfare uncertain. In these days when there are many discussions about whether AI
applications will bring benefit or harm in the future, it is obvious that change is inevitable
and it is clear that countries should be open to new technological structures or methods in
order to ensure their continuity and to predict and structure the future in every field.

149
12. CONCLUSION AND PROPOSAL

ANNs are computer software developed to imitate the working mechanism of the
human brain and to realize the brain's abilities, such as learning, remembering, and
generating new information through generalization, without any assistance. Due to this
feature, AI applicaitons are successfully applied in many fields, such as industry, business,
finance, education, the military, defense, and health. One of the application areas of AI is
solving problems of future prediction.

ANNs have been able to obtain very successful results in non-linear time series
analysis due to their non-linear structure. In this regard, AI has become a frequently used
method in many areas of finance. In the field of finance, although AI models are applied in
many studies examining issues such as predicting financial crises, determining the direction
of exchange rates, general price levels, and especially measuring and selecting the
performance of stocks, there are limited studies on the estimation of values of stocks market
index.

Linear stochastic regression models can have an advantage over other models if they
can understand and explain important details between variables. However, linear models are
insufficient when the relationship between the variables in the studied problem is not linear.
At this point, AI models can make successful predictions when the appropriate network
structure for nonlinear relationships is determined.

In this study, it is aimed to estimate the value of the ISE100 index for the period
January 2001 January 2022 by using the DL modeling technique, which is one of the non-
linear estimation models, after the variables to be used as input variables are determined by
using the Logit and Probit methods. General information about the size and development of
stock markets in Turkey and around the world, or in other words, stock markets, is given.
Then, AI technologies, such as ANNs, and finally DL modeling, which is ML, were
examined in detail. The DL model was designed, and the ISE100 value was modeled using
the appropriate network architectures and seven macroeconomic variables. Estimated
models were evaluated within themselves and performance comparisons were made by
performing Logit and Probit analyses. In Logit and Probit analysis, the ISE100 index was

150
used as the dependent variable, with seven independent variables: the Dollar Rate (TL/$),
the Money Supply, the Producer Price Index, the Industrial Production Index, Gold Prices
(TL/Gr), the Active Bond Interest Rate, and the Brent Oil Price.

The estimated DL models showed a consistent structure and good predictive


performance. In the prediction comparison made with Logit and Probit models, it was
concluded that the DL technique performed better than this method. In addition, it is the
most striking point that the method of DL is more successful than the Logit and Probit
methods in the prediction of financial crises.

When the Logit and Probit model results are examined, it is seen that the models
created are the best models. Since the probability values of Hosmer and Lemeshow and
Pearson goodness of fit tests were higher than 0.05, the models were statistically completely
significant, and the lowest AIC and BIC values and the highest Pseudo R2 values were found
to be 0.865 for the Logit model and 0.866 for the Probit model. The accurate classification
rates were found to be 95.65% and 94.86% for the Logit and Probit models, respectively. It
has been concluded that the estimation results of the Logit and Probit models are very close
to each other. According to the Logit and Probit analysis results, the coefficients obtained
for each independent variable are significant, and the coefficient sizes and test statistics
values are close to each other. It has been observed that the Probit model gives better results
than the Logit model.

The fact that the predictions made with Logit and Probit analyses show great
deviations from the real values and the better prediction performance of the network even
though the seasonal effects are not reflected in the deep neural networks cause a
generalization to be made that the nonlinear modeling for financial variables, that is, the
deep neural network method, is more effective.

There are very few studies that use deep neural networks or DL models to predict
financial markets and stock markets around the world. The method of ANNs has been used
in several studies to predict the direction of the Turkish stock market, but the DL model has
not been used. In this study, it is thought that the use of Logit and Probit analysis and the DL
model to predict the direction of the ISE100 index and the large number and variety of
macroeconomic variables included in the analysis add originality to the study.

151
Today, developments in the fields of communication technology and finance, the
global economic crises that started in the US loan market and affected the national markets
of all the countries of the world, and the socio-economic processes experienced after
epidemics such as the COVID-19 pandemic have shown that international capital markets
have gained a national capital market identity. It is seen that the models created with AI
techniques, which have the ability to make successful predictions and model non-linear
relationships and thus explain the changes in the capital market, are an effective estimation
tool for investors who want to hedge risk and gain returns in such a wide market. Because
of these features, artificial neural network techniques are becoming an increasingly
important problem-solving tool in the field of finance.

Based on this study, the following can be suggested as future research topics for the
ISE100 index: While expanding the basket of macroeconomic variables that affect the
ISE100 value, the stock being studied can be limited. In other words, a similar analysis can
be made for certain stocks in Turkey. The analysis of the value of the Turkish stock market
over a longer period by comparing other econometric models with AI methods and/or using
hybrid models can be another subject of study.

It has been determined that artificial intelligence technologies provide very


successful optimizations compared to traditional methods in terms of how investors or
decision makers will take a position in the capital markets, which investment instruments
they will prefer, or which instruments will be included in the portfolio basket. Therefore,
artificial intelligence technologies emerge as a very successful method in the field of finance.
These models, which are created by learning from the data with the lowest error, will take
their place in the literature as the best optimization and prediction models.

152
13. REFERENCES

Abraham, B., & Malik, H. J. (1973). Multivariate Logistic Distributions. CRC Press., 1(3),

588 590. https://doi.org/10.1214/aos/1176342430

Aggarwal, M. (2020). Probit and Nested Logit Models Based on Fuzzy Measure. I

Journal of Fuzzy Systems, 17(2), 169 181.

Anadolu University Journal of Social

Sciences, 12(2), 87 106.

deksinin Logit-Probit Model

Finans Politik& Ekonomik Yorumlar, 56(650),

131 145.

Aladag, C. H., Egrioglu, E., & Kadilar, C. (2009). Forecasting Nonlinear Time Series with

a Hybrid Methodology. Applied Mathematics Letters, 22(9), 1467 1470.

Alaloul, W. S., & Qureshi, A. H. (2020). Data Processing Using Artificial Neural

Networks. In Data Processing Using Artificial Neural Networks. IntechOpen.

pay

, 10(2),

Article 2.

Altay, E., & Satman, M. H. (2005). Stock market forecasting: Artificial neural network and

linear regression comparison in an emerging market. Journal of Financial

Management & Analysis, 18(2), 18.

Anderson, D., & McNeill, G. (1992). Artificial Neural Networks Technology. Kaman

Science Corporation,.

Avci, E. (2007). Forecasting Daily and Sessional Returns of the ISE-100 Index with Neural

Network Models. niversitesi Dergisi, 8(2), Article 2.

153
Avci, E. (2015). Stock Return Forecasts with Artificial Neural Network Models. Marmara

, 26(1), Article 1.

Barro, R. J. (1990). The Stock Market and Investment. Review of Financial Studies, 3(1),

115 131.

Bengio, Y., Goodfellow, I., & Courville, A. (2015). Deep Learning: Methods and

Applications. MIT Press.

, 16, 31 46.

Bollen, N. P. B., & Busse, J. A. (2001). On the Timing Ability of Mutual Fund Managers.

The Journal of Finance, 56(3), 1075 1094.

(n.d.). Retrieved 30 July 2022, from https://www.borsaistanbul.com

/tr/sayfa/471/borsa-istanbul-hakkinda.

Boyer, B., & Zheng, L. (2009). Investor flows and stock market returns. Journal of

Empirical Finance, 16(1), 87 100. https://doi.org/10.1016/j.jempfin.2008.06.003

Campbell, J. Y., Lo, A. W., & MacKinlay, A. C. (1997). The Econometrics of Financial

Markets. Princeton University Press. https://www.amazon.com/Econometrics-

Financial-Markets-John-Campbell/dp/0691043019

Cao, J., & Wang, J. (2020). Exploration of Stock index Change Prediction Model Based on

The Combination of Principal Component Analysis and Artificial Neural Network.

Soft Computing, 24(11), 7851 7860.

Chakravarty, S. (2005). Stock Market and Macro Economic Behavior in India. Institute of

Economic Growth [Cited 20 May 2011]., 15.

154
Chen, A.-S., Leung, M. T., & Daouk, H. (2003). Application of Neural Networks to an

Emerging Financial Market: Forecasting and Trading The Taiwan Stock Index.

Computers & Operations Research, 30(6), 901 923.

Chen, N.-F., Roll, R., & Ross, S. A. (1986). Economic Forces and the Stock Market. The

Journal of Business, 59(3), 383 403.

Chen, W., Zhang, Y., Yeo, C. K., Lau, C. T., & Lee, B. S. (2017). Stock market prediction

using neural network through news on online social networks. 2017 International

Smart Cities Conference (ISC2), 1 6.

Cheng, D., Yang, F., Xiang, S., & Liu, J. (2022a). Financial Time Series Forecasting with

Multi-Modality Graph Neural Network. Pattern Recognition, 121, 108218.

Cheng, D., Yang, F., Xiang, S., & Liu, J. (2022b). Financial time series forecasting with

multi-modality graph neural network. Pattern Recognition, 121, 108218.

Chong, E., Han, C., & Park, F. C. (2017a). Deep Learning Networks for Stock Market

Analysis and Prediction: Methodology, Data Representations, and Case Studies.

Expert Systems with Applications, 83, 187 205.

Chong, E., Han, C., & Park, F. C. (2017b). Deep learning networks for stock market

analysis and prediction: Methodology, data representations, and case studies.

Expert Systems with Applications, 83, 187 205.

Finansal Ekonometri

Finans Politik ve Ekonomik Yorumlar, 616, Article 616.

Cowles, A. (1933). Can Stock Market Forecasters Forecast? Econometrica, 1(3), 309 324.

155
Dase, R. K., & Pawar, D. D. (2010). Application of Artificial Neural Network for Stock

Market Predictions: A Review of Literature. International Journal of Machine

Intelligence, 2(2), 14 17.

Davis, E. P., & Karim, D. (2008). Comparing Early Warning Systems for Banking Crises.

Journal of Financial Stability, 4(2), 89 120.

Deep Learning Toolbox. (n.d.). Retrieved 28 August 2022, from

https://www.mathworks.com/products/deep-learning.html

Diler, A. I. (2003). Forecasting the Direction of the ISE National-100 Index by Neural

Networks Backpropagation Algorithm. Istanbul Stock Exchange Review, 7(25 26),

65 82.

Donaldson, R. G., & Kamstra, M. (1999). Neural Network Forecast Combining with

Interaction Effects. Journal of the Franklin Institute, 336(2), 227 236.

Dutta, G., Jha, P., Laha, A. K., & Mohan, N. (2006). Artificial Neural Network Models for

Forecasting Stock Price Index in the Bombay Stock Exchange. Journal of

Emerging Market Finance, 5(3), 283 295.

. Remzi Kitabevi.

https://www.nadirkitap.com/kuresel-finans-krizi-piyasa-sisteminin-elestirisi-mahfi-

egilmez-kitap14581620.html

Enke, D., Grauer, M., & Mehdiyev, N. (2011). Stock Market Prediction with Multiple

Regression, Fuzzy Type-2 Clustering and Neural Networks. Procedia Computer

Science, 6, 201 206. https://doi.org/10.1016/j.procs.2011.08.038

Enke, D., & Mehdiyev, N. (2013). Stock Market Prediction Using a Combination of

Stepwise Regression Analysis, Differential Evolution-based Fuzzy Clustering, and

156
a Fuzzy Inference Neural Network. Intelligent Automation & Soft Computing,

19(4), 636 648.

Fama, E. F. (1990). Stock Returns, Expected Returns, and Real Activity. The Journal of

Finance, 45(4), 1089 1108.

Fama, E. F., & French, K. R. (1996). Multifactor Explanations of Asset Pricing Anomalies.

The Journal of Finance, 51(1), 55 84.

Fathi, E., & Maleki Shoja, B. (2018a). Deep Neural Networks for Natural Language

Processing. In Handbook of Statistics (Vol. 38, pp. 229 316). Elsevier.

Fathi, E., & Maleki Shoja, B. (2018b). Chapter 9 Deep Neural Networks for Natural

Language Processing. In V. N. Gudivada & C. R. Rao (Eds.), Handbook of

Statistics (Vol. 38, pp. 229 316). Elsevier.

Frankel, J. A., & Rose, A. K. (1996). Currency Crashes in Emerging Markets: An

empirical Treatment. Journal of International Economics, 41(3), 351 366.

Gomes, L. F. A. M., Machado, M. A. S., Caldeira, A. M., Santos, D. J., & Nascimento, W.

J. D. do. (2016). Time Series Forecasting with Neural Networks and Choquet

Integral. Procedia Computer Science, 91, 1119 1129.

Gonenc, H., & Karan, M. B. (2003). Do Value Stocks Earn Higher Returns than Growth

Stocks in an Emerging Market? Evidence from the Istanbul Stock Exchange.

Journal of International Financial Management and Accounting, 14(1), 1 25.

Goodell, J. W., Kumar, S., Lim, W. M., & Pattnaik, D. (2021). Artificial intelligence and

machine learning in finance: Identifying foundations, themes, and research clusters

from bibliometric analysis. Journal of Behavioral and Experimental Finance, 32,

100577.

Gujarati, D. N. (2011). Econometrics by Example. Palgrave Macmillan.

Gujarati, D. N., & Porter, D. C. (2009). Basic eEconometrics (5th ed). McGraw-Hill Irwin.

157
, 27, Article 27.

irilerine Etkisi:

, 1(1 2),

Article 1 2.

Haykin, S. S. (1999). Neural Networks: A comprehensive Foundation (2. ed., [Nachdr.],

internat. ed). Prentice Hall.

Hebb, D. O. (n.d.). The Organization of Behavior. In

Henrique, B. M., Sobreiro, V. A., & Kimura, H. (2019). Literature Review: Machine

Learning Techniques Applied to Financial Market Prediction. Expert Systems with

Applications, 124, 226 251.

Hsieh, C. (1993). Some Potential Applications of Artificial Neural Systems in Financial

Management. Journal of Systems Management, 44(4), 12.

Hsieh, J., Chen, T.-C., & Lin, S.-C. (2019). Credit constraints and growth gains from

governance. Applied Economics, 51(11), 1199 1211.

Hu, M. Y., & Tsoukalas, C. (1999). Combining Conditional Volatility Forecasts Using

Neural Networks: An Application to The Ems Exchange Rates. Journal of

International Financial Markets, Institutions and Money, 9(4), 407 422.

Hull, J. (2007). Risk Management and Financial Institutions. Pearson Prentice Hall.

https://books.google.com.tr/books?id=G4nCQgAACAAJ

Ilahi, I., Ali, M., & Jamil, R. (2015). Impact of Macroeconomic Variables on Stock Market

Returns: A Case of Karachi Stock Exchange. SSRN Electronic Journal.

158
Jing, N., Wu, Z., & Wang, H. (2021). A Hybrid Model Integrating Deep Learning with

Investor Sentiment Analysis for Stock Price Prediction. Expert Systems with

Applications, 178, 115019.

Kaastra, I., & Boyd, M. (1996). Designing a Neural Network for Forecasting Financial and

Economic Time Series. Neurocomputing, 10(3), 215 236.

Kaminsky, G. L., & Reinhart, C. M. (1999). The Twin Crises: The Causes of Banking and

Balance-of-Payments Problems. American Economic Review, 89(3), 473 500.

Kamruzzaman, J., Begg, R. K., & Sarker, R. A. (2006a). Artificial Neural Networks in

Finance and Manufacturing. Idea Group Publishing.

Kamruzzaman, J., Begg, R. K., & Sarker, R. A. (2006b). Artificial Neural Networks in

Finance and Manufacturing. Idea Group Publishing.

Logit- International Journal of

Management Economics and Business, 14(3), 575 590.

Price Index Movement Using Artificial Neural Networks and Support Vector

Machines: The Sample of The Istanbul Stock Exchange. Expert Systems with

Applications, 38(5), 5311 5319.

, 13(2), Article 2.

Karan, M. B. (2013). (4th ed.). Gazi Kitapevi.

Kaya, V., & Yilmaz, O. (2007).

(Working Paper No. 2007/1). Discussion

Paper.

159
change

During Covid-19 Pandemic. , 37(1), Article 1.

Kennedy, P. (1998). A Guide to Econometrics (Fourth Edition). MIT Press.

Kim, K. (2003). Financial Time Series Forecasting Using Support Vector Machines.

Neurocomputing, 55(1), 307 319.

Kim, T. Y., Oh, K. J., Sohn, I., & Hwang, C. (2004). Usefulness of Artificial Neural

Networks for Early Warning System of Economic Crisisrisis. Expert Systems with

Applications, 26(4), 583 590.

kenlerin Borsa Endeksleri

, 12(22), 116 151.

Kohonen, T. (2001). Self-Organizing Maps. Springer Science & Business Media.

Kovacova, M., & Kliestik, T. (2017). Logit and Probit Application for The Prediction of

Bankruptcy in Slovak Companies. Equilibrium. Quarterly Journal of Economics

and Economic Policy, 12(4), 775 791.

Kubat, C. (2019).

Kutsurelis, J. E. (1998). Forecasting Financial Markets Using Neural Networks: An

Analysis of Methods and Accuracy

Liang, X., Zhang, H., Xiao, J., & Chen, Y. (2009). Improving Option Price Forecasts with

Neural Networks and Support Vector Regressions. Neurocomputing, 72(13 15),

3055 3065.

Lin, C.-C., Chen, C.-S., & Chen, A.-P. (2018). Using intelligent computing and data

stream mining for behavioral finance associated with market profile and financial

physics. Applied Soft Computing, 68, 756 764.

160
Lin, C.-S., Khan, H. A., Chang, R.-Y., & Wang, Y.-C. (2008). A New Approach to

Modeling Early Warning Systems for Currency Crises: Can a Machine-Learning

Fuzzy Expert System Predict the Currency Crises Effectively? Journal of

International Money and Finance, 27(7), 1098 1121.

Lintner, J. (1965). Security Prices, Risk, and Maximal Gains From Diversification. The

Journal of Finance, 20(4), 587 615.

Liu, M., & Shrestha, K. M. (2008). Analysis of t

economic variables and the Chinese stock market using heteroscedastic

cointegration. Managerial Finance, 34(11), 744 755.

Markowitz, H. (1952). Portfolio Selection. The Journal of Finance, 7(1), 77 91.

McCulloch, W. S., & Pitts, W. (1943). A Logical Calculus of the Ideas Immanent in

Nervous Activity. The Bulletin of Mathematical Biophysics, 5(4), 115 133.

McNelis, P. D. (2005). Neural Networks in Finance Gaining: Predictive Edge in the

Market. Elsevier Academic Press.

(1st

Mossin, J. (1966). Equilibrium in a Capital Asset Market. Econometrica, 34(4), 768 783.

Mukherjee, T. K., & Naka, A. (1995). Dynamic Relations Between Macroeconomic

Variables and The Japanese Stock Market: An Application of a Vector Error

Correction Model. Journal of Financial Research, 18(2), 223 237.

Nunes, M., Gerding, E., McGroarty, F., & Niranjan, M. (2019). A Comparison of

Multitask and Single Task Learning with Artificial Neural Networks for Yield

Curve Forecasting. Expert Systems with Applications, 119, 362 375.

161
-price manipulation in an

emerging market: The case of Turkey. Expert Systems with Applications, 36(9),

11944 11949.

Panda, C., & Narasimhan, V. (2007). Forecasting Exchange Rate Better with Artificial

Neural Network. Journal of Policy Modeling, 29(2), 227 236.

Patil, P. R., Parasar, D., & Charhate, S. (2021). A Literature Review on Machine Learning

Techniques and Strategies Applied to Stock Market Price Prediction. Springer.

Paul D. McNelis. (2005). Neural Networks in Finance: Gaining Predictive Edge in the

Market. Academic Press; eBook Subscription Super Collection - Turkey

(EBSCOhost).

Peng, Y., Albuquerque, P. H. M., Kimura, H., & Saavedra, C. A. P. B. (2021). Feature

Selection and Deep Neural Networks for Stock Price Direction Forecasting Using

Technical Analysis Indicators. Machine Learning with Applications, 5, 100060.

Phillips, P. C. B., & Perron, P. (1988). Testing for a unit root in time series regression.

Biometrika, 75(2), 335 346.

Quazi, S. (2022). Artificial Intelligence and Machine Learning in Precision and Genomic

Medicine. Medical Oncology, 39(8), 120.

Roll, R., & Ross, S. A. (1980). An Empirical Investigation of the Arbitrage Pricing Theory.

The Journal of Finance, 35(5), 1073 1103.

Rousseau, P. L., & Sylla, R. (2003). Financial Systems, Economic Growth, and

Globalization. In Globalization in Historical Perspective (pp. 373 416). University

of Chicago Press.

162
Sarker, I. H. (2021). Deep Learning: A Comprehensive Overview on Techniques,

Taxonomy, Applications and Research Directions. Sn Computer Science, 2(6), 420.

BDDK

, 5(1), Article 1.

(5.). Turhan Kitapevi.

Schmidhuber, J. (2015). Deep Learning in Neural Networks: An Overview. Neural

Networks, 61, 85 117.

Schwert, G. W. (1990). Stock Returns and Real Activity: A Century of Evidence. The

Journal of Finance, 45(4), 1237 1257.

, M. (2010). Ekonometik Zaman Serileri Analizi (3.).

Sharpe, W. F. (1964). Capital Asset Prices: A Theory of Market Equilibrium Under

Conditions of Risk. The Journal of Finance, 19(3), 425 442.

Smith, K. A., & Gupta, J. N. D. (2000). Neural Networks in Business: Techniques and

Applications for The Operations Researcher. Computers & Operations Research,

27(11 12), 1023 1044.

Network and Agility. Procedia - Social and Behavioral Sciences, 195, 1477 1485.

Sutskever, I., Martens, J., & Dahl, G. (n.d.). On The Importance of Initialization and

Momentum in Deep Learning. 9.

Treynor, J. L. (1965). How to Rate Management of Investment Funds. Harvard Business

Re , 43(1), 63 75.

163
Tsai, C. F., & Wang, S. P. (2009). Stock Price Forecasting by Hybrid Machine Learning

Techniques. Proceedings of the International Multiconference of Engineers and

Computer Scientists, 1(755), 60.

Tseng, C.-H., Cheng, S.-T., Wang, Y.-H., & Peng, J.-T. (2008). Artificial neural network

model of the hybrid EGARCH volatility of the Taiwan stock index option prices.

Physica A: Statistical Mechanics and Its Applications, 387(13), 3192 3200.

Turing, A. M. (1950). Computing Machinery and Intelligence. Mind, 49, 433 460.

market: The case of Turkey. Cross Cultural Management: An International

Journal, 13(4), 277 295.

Vadlamudi, S. (2017). Stock Market Prediction using Machine Learning: A Systematic

Literature Review. American Journal of Trade and Policy, 4(3), 123 128.

Vui, C. S., Soon, G. K., On, C. K., Alfred, R., & Anthony, P. (2013). A review of stock

market prediction with Artificial neural network (ANN). 2013 IEEE International

Conference on Control System, Computing and Engineering, 477 482.

Wang, L., Wang, Z., Qu, H., & Liu, S. (2018). Optimal Forecast Combination Based on

Neural Networks for Time Series Forecasting. Applied Soft Computing, 66, 1 17.

Wongbangpo, P., & Sharma, S. C. (2002). Stock Market and Macroeconomic Fundamental

Dynamic Interactions: ASEAN-5 Countries. Journal of Asian Economics, 13(1),

27 51.

Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data (The

second edition). The MIT Press.

Wu, S.-I., & Lu, R.-P. (1993). Combining Artificial Neural Networks and Statistics for

Stock-Market Forecasting. Proceedings of the 1993 ACM Conference on Computer

Science - , 257 264.

164
Yao, J., Poh, H., & Jasic, T. (1996). Foreign Exchange Rates Forecasting with Neural

Networks. 754 759.

Artificial Nneural Networks: Review. Turkiye Klinikleri Journal of Medical

Sciences, 27, 65 71.

Yildiz, B. (2001). Prediction of Financial Failure With Artificial Neural Network

Technology and an Empirical Application on Publicly Held Companies. Istanbul

Stock Exchange Review, 5(17), 47 62.

Yolcu, U., Egrioglu, E., & Aladag, C. H. (2013). A New Linear & Nonlinear Artificial

Neural Network Model for Time Series Forecasting. Decision Support Systems,

54(3), 1340 1347.

Yu, K., Tresp, V., & Schwaighofer, A. (2005). Learning Gaussian Processes from Multiple

Tasks. Proceedings of the 22nd International Conference on Machine Learning -

, 1012 1019.

Zhang, G., Eddy Patuwo, B., & Y. Hu, M. (1998). Forecasting with Artificial Neural

Networks: The state of the Art. International Journal of Forecasting, 14(1), 35 62.

Zhang, G. P. (2001). An Investigation of Neural Networks for Linear Time-Series

Forecasting. Computers & Operations Research, 28(12), 1183 1202.

Zhang, G. P. (2003). Time Series Forecasting Using a Hybrid ARIMA and Neural

Network Model. Neurocomputing, 50, 159 175.

Zhang, G. P. (2004a). Business Forecasting with Artificial Neural Networks: An

Overview. Neural Networks in Business Forecasting.

Zhang, G. P. (2004b). Neural Networks in Business Forecasting. Idea Group Inc (IGI).

Zurada, J. M. (1992). Introduction to Artificial Neural Systems. West.

165

You might also like