THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
20
https://www.theamericanjournals.com/index.php/tajmei
PUBLISHED DATE: - 15-10-2024
DOI: -
https://doi.org/10.37547/tajmei/Volume06Issue10-04
EXPLORING THE ROLE OF ARTIFICIAL
INTELLIGENCE IN MANAGING EMERGING
RISKS: AN IN-DEPTH STUDY OF AI
APPLICATIONS IN FINANCIAL
INSTITUTIONS' RISK FRAMEWORKS
Md Zahidur Rahman Farazi
The University of Texas at Arlington Arlington, Texas
INTRODUCTION
The complexity of managing emerging risks in
financial institutions is therefore exacerbated by
the fast-changing nature of the markets and the
rapidly growing volume of transactional
information. Although conventional risk analysis
tools play the role of core practice, they do not
suffice in explaining and addressing the complex
and multiple risks that characterize modern
financial settings [1]. This lack of utilization points
to a need for more complex methods especially
those using improvements in artificial intelligence
(AI). Interest in the contribution of AI to financial
risk management has been significant due to its set
of capabilities such as big data processing capacity,
pattern-recognition, and high predictive accuracy
than conventional approaches [2]. Therefore, this
study seeks to establish how the advancements in
AI such as the use of machine learning and artificial
neural networks can be adopted in the risk
management sector of financial institutions
specially to manage the emerging risks.
RESEARCH ARTICLE
Open Access
Abstract
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
21
https://www.theamericanjournals.com/index.php/tajmei
Due to the constraints that permeate the use of
conventional mechanisms of risk appraisal,
financial institutions have a big problem effecting
the management of emerging risks. These methods
are some of the times unable to adapt themselves
to the compounded and constantly changing global
financial systems, resulting in poor risk control and
less accurate risk forecasting [3]. This research
aims at establishing whether AI can improve risk
assessment tasks and especially the ML and ANN
models to offer better solutions in the daily
management of financial risk situations that
present said challenges.
The key research question in this research is that
whether by building and testing the models based
on the data which has imbalanced and missing
values, the risk assessment can be improved
substantially by applying the techniques of ML and
ANN as against the conventional approaches. Some
of the typical methods of risk assessment are look
at the static models do not take into account into
consideration the dynamism and un-predictability
of the financial risks. On the other hand, ML and
ANN are capable of learning from past occurrences,
including changing patterns, and able to provide
better predictions than the rule-based systems.
This flexibility is so necessary in a context that
evaluates that new risks appear regularly and that
they are evolving.
There are three research objectives of the study.
First, it seeks to establish the performance of the
financial risk management using ML and ANN
model given the type of risk dynamic and diverse.
This includes comparing the above stated AI
integrated strategies to the conventional risk
assessment in determining risks ratings as well as
assessing potential threats. Secondly, the study
aims at establishing the specific ML and ANN
approaches that are more relevant for improving
risk management practices. This involves exploring
different algorithms of machine learning like the
decision trees, the support vector machines and the
deep networks to establish their efficacy in view of
evaluating the levels of risk in the financial world.
Last, the study will seek to proffer practical
suggestions to the financial institutions on the
effective implementation of the ML and ANN
techniques into the risk management models as
proposed by the analysis.
Embracing the potential threats of AI demands a
deep and sophisticated comprehension not only to
enhance the tactics and strategies of individual
institutions but also to advance the science and art
of financial technology [4]. As the world markets
for financial instruments develop ever more
complex, the requirement for new approaches
toward risk management is amplified. Specifically,
some of the shining areas of application of AI
involve the use of AI for improving the forecast
precision that will be useful in solving some parts
of the problems associated with financial risk
management [5]. Thus, the effective application of
ML and ANN to process enhanced information and
to control new risks serve as a key to increase the
competitive advantage of financial institutions,
enhance the efficiency of the risk assessment
operations, and thus promote the enhancement of
the financial stability.
Consequently, the present research is significant
for the development of the financial risk
management as it investigates the possibilities of
applying the AI systems and tools, with the
emphasis on the machine learning and artificial
neural networks. Specifically, this research seeks to
assess the potential of ML and ANN in enhancing
risk assessment procedures and proffer
recommendation on how best to implement them
within the framework of the financial institutions
[6]. As such, this study aims at achieving the
following objectives: To contribute to the existing
knowledge of the performance and effectiveness of
listed firms’ AI applications in financial risk
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
22
https://www.theamericanjournals.com/index.php/tajmei
management and assist with enhancing the
stability of global financial systems.
LITERATURE REVIEW
Risk management from financial risks has emerged
as one of the important areas of investigation in AI
research, as firms try to improve their capabilities
in anticipating and mitigating new risks. The
applicability of risk measures for analysing risk
exposures in today’s complex and c
onstantly
evolving economic structures remains a challenge
to conventional approaches involving the use of
risk assessment tools such as ML and ANNs [7].
This literature review is a comparison of the
findings of prior research works on use of AI in
financial risk management to determine the
progress in the field and existing research gaps.
The Classical Risk Management Models
The conventional approach used in financial
institutions with regards to risk management
usually involves the use of statistical models and
past information. Traditional models that have
played significant roles in risk measurement
include, for instance, Value at Risk (VaR) as well as
stress testing. VaR calculates the expected
maximum loss for a specified time horizon and
confidence interval while stress testing displays
how specific volatile states would affect the
stability of an organization’s financial status [8].
However, these methods have some constraints;
they are static in nature, which mean they cannot
dynamically adjust when the market conditions are
changing and they rely on historical data which
may not reflect future risks.
Machine
Learning
in
Financial
Risk
Management
Advanced techniques in the field of the financial
risk management have recently been shifted
towards utilizing the machine learning approaches
because they are capable in terms of handling large
data and at the same time, they are capable to
identify the new relationship which is not been
known with the traditional methods [9]. The
research [10] suggested that the traditional
statistical model can be improved by using the
machine learning techniques including decision
trees, support vector machines and the ensemble
methods. Like decision trees and random forests
have been applied as a tool to enhance credit
scoring models due to increased input variables
and interactions [11]. In contrast, SVMs have
proven successful in a class of classification
problems such as, fraud detection, and credit risk
assessment because of the capacity to tackle non-
linearity [12].
However, ML models also pose certain difficulties.
For instance, they need large volumes of quality
data so that they can execute their function and
their functionality may depend with the selection
of the hyperparameters [13]. In addition, it is
always possible for the experts to identify the
patterns using the ML models while at the same
time, they are capable of overfitting in cases where
the model has a high level of complexity compared
to available data [14].
Artificial Neural Networks (ANNs) in Financial
Risk Management
ML can be categorized into various types and
among them, ANNs are the most popular in
financial risk management because of their
capabilities of capturing non-linear structures in
data [15]. ANNs, including deep learning models,
have been shown to be used in a number of risk
management areas, including credit risk
assessment, market forecasting and fraud
detection [16]. Several pieces of research have
outlined the benefits of ANNs in enhancement of
prediction, as well as another consideration of
other concealed pattern in financial data. For
instance, the study [17] indicated that using ANNs,
credit risks could be predicted better than by a use
of traditional statistical techniques due to the
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
23
https://www.theamericanjournals.com/index.php/tajmei
ability of the former ones to consider non-linear
relations between financial factors.
Although, it has been seen that ANNs have their
own disadvantages also. To train deep neural
networks it takes a lot of computational power and
time as well [18]. Further, ANNs pose the problem
of interpretability in that the way it arrives at its
results is not very clear, especially in risk
management fields where it is crucial to
understand why the decision in certain areas has
been made [19].
There are a number of published works that look at
the performance of AI techniques such as ML and
ANNs compared to more traditional approaches to
risk assessment. The research [20] conducted a
study on the performances of ensemble methods
and neural networks as compared to the
conventional logistic regression techniques for
credit scoring. Based on their works, they saw that
ensemble methods especially gradient boosting
machines and deep learning models have higher
accuracy to logistic regression. In the same way, the
research [21] recognised that the use of ANNs
provided better accuracy in predicting financial
distress and bankruptcy than traditional statistical
methods as an indication of the usability of neural
networks when it comes to dealing with non-linear
and complex data.
The study [22] noted that despite achieving high
accuracy levels in credit risk, ANNs performed
poorly on unbalanced datasets whereby non
defaulted cases outnumbered default cases
considerably. Another problem is the issue of
imbalance of data, where this is also a problem in
the field of financial risk management impacting
the effectiveness of not only conventional and
complex artificial intelligence models.
Recent Developments and New Trends
The advancement of AI techniques has been
studied in the recent past with an emphasis on
addressing the shortcoming and expanding the
utility of these methods in the FSM context. For
instance, new models have been developed to
incorporate feature of artificial intelligence and at
the same time feature of the traditional human
expert systems [23]. In the paper [24] credit risk
prediction enhanced by using ANNs coupled with
econometric models proving that the incorporation
of different models lead to better models’
performance.
The last trend involves applying the explainable AI
(XAI) methods to deal with the interpretational
challenges of ANNs and other similar architectures
[25]. The main goal of Explainable AI is to extend
AI’s interpretability so that the decision
-making
process can be explained, especially in legal
contexts [26]. The study [27] proposed methods
like LIME: Local Interpretable Model-agnostic
Explanations and SHAP: SHapley Additive
explanations for better understanding of model
decisions thus enhancing their application in
financial risk management decision.
The discussion of AI in financial risk management
shows that machine learning and artificial neural
networks can complement the traditional
approaches to risk assessment, but can also pose
certain risks. The comparison in the efficiency of
traditional methods with ML/ANN models shows
that the latter offers higher predictive accuracy,
mechanisms to address complex data, yet, data
quality,
interpretability
of
models
and
computational requirements issues exist. Based on
comparative research, it is possible to identify
innovations that increase the effectiveness of risk
management practices and strengthen the
shortcomings of the approaches used. New
directions include concepts like hybrid schemes
and the concepts of explainable AI remain future
directions for research and development focusing
on enhancing the efficiency of the AI algorithms
and the methods of using these schemes in the
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
24
https://www.theamericanjournals.com/index.php/tajmei
mechanisms of financial risk management. This
means that with the continuation of changes in the
technologies of AI and their usage in the financial
industry, such as the risks management, there will
always be new chances and opportunities in the
future.
MATERIALS AND METHODS
This research uses an empirical research approach
that measures the use of ML and ANNs in FRM
utilising quantitative research methods. It employs
real data to test the efficiency of the methods used
by the AI techniques in recognising, measuring,
assessing and predicting financial risks as against
standard risk assessment techniques. The focus on
quantitive methods is important, as it allows to
establish a clear and numerical approach in
analyzing the performance of AI models when
dealing with financial data.
Traditional models such as logistic regression will
be used along with other models such as random
forests, SVMs as well as deep learning techniques.
Thus, such comparative analysis reveals the
benefits of applying AI techniques when
addressing the emerging risks within financial
institutions in terms of large datasets processing
and discovering patterns that cannot be addressed
within the framework of traditional approaches.
While evaluating the models, accuracy, precision,
recall, and F1-score will be employed as the metrics
of measurement. Therefore, the study aims at
present empirical support on the effectiveness of
using AI models particularly the ML and ANNs on
financial risk analysis and management in contrast
to the traditional models. Hence, through using this
empirical design, the research seeks to provide
information on how AI can be applied in the
improvement of risk management frameworks in
the financial institutions.
RESEARCH OBJECTIVES
•
To assess the effectiveness of the ML and
ANN models which are applied to mitigate
the financial risks.
•
To further compare the mentioned models
with the conventional risk assessment tools
and techniques.
•
To explore the best practices and possible
strategies to incorporate AI into FRM
strategies.
DATA COLLECTION
The financial data in this study are obtained from
the Kaggle dataset site, which is the largest site for
sharing and providing various datasets for machine
learning and data analysis [28]. Such datasets
include detailed descriptive data such as
demography, finance, and behaviors crucial in
evaluating the credit risk. The datasets involve real
life data and mimic the real data that is processed
in the financial organizations hence being relevant
to the objectives of the study. The specific dataset
contains attributes like income, a credit score of the
applicant, job status, and loan amount which are
important in terms of measuring the certain credit
risk. Nonetheless the utilization of public
repositories such as Kaggle keeps the data
transparent and easily accessible for empirical
research. Also, Kaggle datasets with high quality
data are required for the for building reliable
models through the training and testing phases.
SAMPLE SIZE
The dataset contains 15,000 records which is more
than the minimum number of records required for
analysis of 10,000. Exemplary in this regard is the
peace, as it affords the machine learning models
and the ANNs abundant data that enable them to
learn patterns well and generate good forecasts. A
large sample size allows the study to investigate
various many facets of financial risk taking based
on the participants demographics and financial
situation. The variations within the data set, in
form of age, income, credit score, employment
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
25
https://www.theamericanjournals.com/index.php/tajmei
history and the likes makes ability of the model to
handle issues such as imbalanced classes and
missing values robust. The large sample size also
poses external validity benefits because the
research results can be astounding across various
parties,
financial
organizations
and
risk
management systems.
DATA HANDLING
In this study, Python will be used in data handling
and analysing using Pandas, NumPy and Scikit-
learn for handling missing values and conducting
imputation. First, the empty cells in relation to the
set dataset will be checked in order to determine if
there are any missing values in the columns. This
will be achieved by using Pandas function of
isnull() and sum() to run over all the thirty numeric
predictors and check if any of the value equals to
null or NaN among the 20 variables.
When the level of missing data is determined, an
appropriate means of handling missing data must
be taken based on the type of variables. In the case
of numerical columns which i
nclude ‘Income’,
‘Credit Score’, and ‘Loan Amount’, missing values
will be replaced by mean or median depending on
the situation. For the categorical variables like
‘Gender’, ‘Marital Status’ and ‘Loan Purpose’, we
shall adopt the mode imputation on the basis of
mode of the category that has the highest
frequency of occurrence.
In situations, where there is large amount of
missing data, or if imputation might skew the
results, K Nearest Neighbors (KNN) imputation will
be considered for error-based estimation of the
missing values based on record similarity. These
processes will help to make the dataset as perfect
as possible for training and testing of the machine
learning models effectively.
PROPOSED FRAMEWORK
Integration of Artificial Intelligence Techniques
The framework that was suggested is the
enlargement of applying machine learning and
artificial neural network in addition to the
conventional methods of financial risk assessment
to achieve higher predictive accuracy and
flexibility. VaR and stress testing models are quite
elaborate but still based on the models of relatively
stable environment, hence not reflecting the
dynamic character of modern threats and risks. To
overcome these limitations, integration of ML and
ANN in the proposed framework has been planned
to use the high-level processing and pattern
recognition abilities of AI for analysis of large
amount of data, identification of intricate patterns
and dynamic risk environment.
Here in this integration, a few of the ML models are
the decision trees, Random Forest and also the
support vector machines are used to increase the
predictive accuracy as well as for managing the
non-linear associations in the financial data set.
Furthermore, more of ANNs will be used in
modeling complex relationships and improving the
chances of risk pattern detection that a normal
approach may miss. In particular, the framework is
going to employ Feedforward Neural Networks
(FNNs) and Deep Neural Networks (DNNs).
Feedforward Neural Networks (FNNs) will be
applied to capture basic non-linear mappings as
well as patterns, in the financial data. It comprises
of an input layer, one or more of the hidden layers,
and the output layer as this network enables
learning of historical financial data.
More complex and hierarchical relationships
within the data will be modeled with the help of
DNNs which are really multilayer perceptron’s. The
insights from Big Data sources will be particularly
valuable when it comes to learning from different
patterns and using them in enhancing the
possibilities for better risk prediction and
flexibility.
The work will include data preparation of the
financial data, training the AI Models and
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
26
https://www.theamericanjournals.com/index.php/tajmei
comparing the AI results against the Conventional
Models with respect to metrics such as accuracy,
precision, recall, and F1-score. This approach is
expected to improve reliability of prediction and
flexibility in the existing financial risk assessment.
Hybrid Model Approaches
To enhance the risk assessment process, the
framework will include the models that are part
machine learning and part conventional analysis,
exploring the advantages of each approach. It can
be seen that there are full hybrid models that
combine the strong analytical performance of the
numeric approach of traditional methods with the
higher pattern recognition ability of AI.
For instance, a hybrid model could be logistic
regression with additional ML techniques, where
the logistic regression model will give the basic risk
ratings while other ML techniques will improve on
the general ratings by factoring in other
parameters and details from the dataset. Another
approach that could be taken would be to integrate
Feedforward Neural Networks (FNNs) with
econometric models’ qualitative approach. That
way, FNNs could be used for dealing with complex
non-linear interactions, while the econometric
models could be used in terms of structural
specifications and constraints.
Further, while DNNs could be used in combination
with conventional methods of risk assessment for
higher detection of the patterns, they were found to
be reasonably accurate for predictions. The
possibility of various negative events is taken into
account by these approaches; however, when
integrated into the proposed framework, the risk
management system would be stronger and more
flexible. Analyticity of the hybrid models will be
reviewed for balancing of data, increasing
predictive power, and informing risk assessment
decisions for better assessment of financial risk
hence improving on the risk assessment process.
DATA ANALYSIS
Exploratory Data Analysis (EDA)
Exploratory Data Analysis (EDA) will be the first
and crucial step in gaining familiarity with the
Financial Risk Assessment Dataset. During this
phase, attention will be paid to the patterns,
distribution of the data and relationship in order to
get general insight. Income, Credit Score and Loan
Amount will be presented by histograms so as to
determine the mean, variance, and pattern of
positivity or negativity skewness of these
distributions. For the purpose of comparing two
variables, the option of scatter plots shall be
utilized with emphasis on correlations and trends
in variables such as ‘Income’ and ‘Credit Score’.
Further, correlation matrices will be employed in
determining the nature or direction while
measuring the strength of linear relationships
between features. This matrix will help enumerate
which variables are strongly associated, and this
will provide understanding of the relations of the
variables and importance for model building. By
using such techniques, EDA will highlight some of
the structural aspects and quality problems in the
datasets that should guide the subsequent steps in
data engineering and modeling strategies.
Data cleaning and processing
Methods used to impute or remove the
successfully identified missing values during EDA
include imputation techniques or removal
strategies that have an impact on the model [29].
Where numerical data is involved, missing data will
be imputed by mean or median while if categorical
data is involved missing values will be imputed
with the mode or most frequent value. Certain rows
or columns, which highly contain missing values or
imputation may introduce huge bias, can be
omitted.
Data Preparation
Feature selection will be the essential step in the
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
27
https://www.theamericanjournals.com/index.php/tajmei
ML workflow to improve the model’s performance
and its interpretability [30]. The current research
will thus be implementing a multi-dimensional
approach to the use of relevant features for
assessing financial risks.
To start with, correlation analysis will be
spearheaded in a bid to assess the correlation
matrix and in the long run aid in the identification
of features that are highly correlated. These
features will be omitted in a bid to deal with
multicollinearity problem and ease the model.
Recursive Feature Elimination (RFE) will then be
followed in order to progressively delete less
significant features which ultimately leads to a
condensation of the predictor variables based on
model significance. Further, feature importance of
each feature to the model for perusing will be
computed by the Random Forest models and the
XGBoosts. Depending on the importance level
assigned to them features with high importance
level will be given preference.
Through the application of these techniques, a rich
set of features would be developed which on its
part, enhances model efficiency, minimizes
overfitting,
and
makes
models
easily
understandable, while dealing with imbalance and
structure of the data which is typical in assessing
financial risk.
Model Selection
The Model Selection phase thus consists of
applying and selecting several of these algorithms
for evaluating their efficiency in financial risk
management. It entails assessing the classical ML
methods and the more sophisticated ANNs to find
out which methods give the best and more reliable
risk estimates.
Machine Learning (ML) Models
Financial risk analysis is a way in which Machine
Learning (ML) Models, which include Random
Forest, Logistic Regression, SVM, and XGBoost are
used to make a firm’s financial risk assessment. The
selection is made depending on the discriminant
capabilities of these models when dealing with
large data samples, increasing the precise of
prediction, and providing the capacity to manage
new types of financial risks with different
algorithms.
Random Forest Classifier
Random Forest Classifier is an enhanced decision
tree learning method where the outcome of many
decision trees is combined in order to reduce error
rate and enhance the performance of classifier [31].
Every tree in the forest is learnt on a bootstrap
sample of the data and a final prediction is made
through voting or averaging over the trees. Before
applying the Random Forest Classifier, the dataset
will be split into training and testing dataset. As
feature selection is performed in the preceding
step, in the training phase you will set the number
of trees, maximum depth and minimum samples
per leaf. Various hyperparameters will be tuned by
Grid search or Random search in order to improve
the performance [32]. Random Forest Classifier is
highly accurate with an added advantage of not
being sensitive to over fitting. It addresses both
numerical and categorical data and captures the
interaction between the features and the
complexity of risk prediction making it very
appropriate in financial risk prediction [33].
Logistic Regression
Logistic Regression is used in decision making
particularly in binary classification problems [34].
Logistic means is used to predict the likelihood of a
binary event occurrence depending on features
that are taken as inputs; It is used frequently in
credit scoring or other assessments of risk [35].
Logistic Regression will be used by training the
model by adjusting the parameters in order to fit
the training data through optimization algorithms
like gradient descent. L1 and L2 refers to
regularization parameters, used in order to avoid
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
28
https://www.theamericanjournals.com/index.php/tajmei
overfitting as well as improve on the generalization
capability of the model [36]. The interpretability
and simplicity of Logistic Regression make the
model ideal for determining the correlation of the
financial risk factors and outcomes. Its probability
output is therefore well suited for simple risk
evaluation and decision making especially in fiscal
relations [37].
The Support Vector Machines
The Support Vector Machine (SVM) is a
classification algorithm used for finding a unique
hyperplane that best separates classes in a high
dimensional space [38]. It can be used in linear and
non-linear classifying problem. For SVM, different
kernels will be used in training the model to decide
on the best hyperplane to use from the ones
available which include, linear, polynomial, RBF
among others. The main parameters for evaluating
SVM are its capacity to operate with relatively large
number of data characteristics and its suitable for
the tasks that differ by well-defined class
boundaries. This is due to the fact that its choice of
different kernels is flexible thus well suited for
disparate financial risk circumstances [39].
XGBoost Classifier
XGBoost is an extension of gradient boosting
strategy used to improve upon algorithm result
and time taken to execute the same with fewer
resources [40]. It is an ensemble learning model
that collects several weak learners and constructs
a very powerful predictor. XGBoost will be utilized
in the model with an option to train on the dataset
with variables like learning rate, number of
estimators and maximum depth of the tree. In the
process of model selection, grid search will be
employed and it was described in detail by in the
paper [41]. XGBoost is known for its better
performance, better scalability and better
performance on large datasets with intricate
correlations. It improves the predictive precision
and stability in the evaluation of financial risks by
the advanced boosting methods [42].
Artificial Neural Networks (ANN)
Feedforward Neural Networks as well as Deep
Neural Networks are used for modeling the
complex data relationships and to improve the
forecast accuracy. ANNs uses multiple neurite
layers and complex structures in order to detect
complex patterns in the financial data which
enhance the level of risk management.
Feedforward neural networks
Feedforward neural networks also abbreviated as
FNNs are a class of neural networks which function
using a forward pass neural network architecture
where signals flow forward through the layers of
the network but have no feedback loops, the FNNs
do not incorporate feedback loops through their
architecture [43].
Feedforward Neural Networks (FNNs) are
multilayer perceptron structures that contain
input layer, hidden layers and output layer with ac
concern proportional to one direction. They can
analyse intricate structures and relationships in
data [44]. This is due to the fact that the number of
hidden layers and number of neurons will be
specified during the designing of the FNN we are to
use. The cornerstone of this network will be using
activation functions such as ReLU to help in tuning
the neural network and in the process of
backpropagation together with Adam [45]
optimization technique. FNNs are useful as a first
step in learning basic non-linear relationships and
to serve as a reference for comparison of more
sophisticated methods. Because they are excellent
in learning patterns, they are suitable in
performing initial comparisons when it comes to
financial risk prediction.
DNNs
FNNs are expanded through adding more than one
hidden layer with precisely defined connectivity by
making use of Deep Neural Networks (DNNs) as
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
29
https://www.theamericanjournals.com/index.php/tajmei
they are capable of analyzing most advanced traits
and relationships in the gathered data [46]. The
architectures of DNNs will contain more than one
hidden layer; the number of neurons in each layer
will also be varied. Some of the methods such as
dropout and batch normalization will be employed
in order to enhance generalization and training
time Indeed, DNNs are ideal for processing such
data as it relates to various elements and abstract
and intricate patterns that may elude more basic
models to the financial risk assessment’s
enhancement [47].
Evaluation Metrics
The use of these evaluation metrics offers a way
through which the effectiveness of the various
predictive models in the management of financial
risks can be tested. This led to the identification of
the following key performance measures; Accuracy
which is a measure of the proportion of correct
number of instances out of all the instances made
by the model and serves as an overall measure of
effectiveness of the model but is likely to be skewed
where there is denser class distribution. Precision
and recall are crucial metrics: precision looks at the
ability of the model to correctly predict positive
cases while recall looks at how the model is able to
capture all the cases that are positive. These
metrics are especially helpful when there is a large
number of classes and most classes have few data
points in it. The F1 Score that is obtained as the
harmonic mean of both precision and recall
enables the inclusion of both false positives and
false negatives, which are essential in scenarios of
unequal risk.
RESULTS
Figure 1: Distribution of the Numerical Columns
Fig illustrates the dispersion of several numeric
columns in the dataset we are dealing with. Some
of the columns include; Age, Income, Credit Score,
Loan Amount, Years at Current Job, Debt-to-
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
30
https://www.theamericanjournals.com/index.php/tajmei
Income Ratio, Assets Value, Number of
Dependents, Previous Defaults and change in
Marital Status. Every column is depicted by a
histogram wherein the numbers of different values
in the certain column are indicated. They can see
several features of the data at first glance of the
histograms. For example, Age data seems to be
normally distributed with the most values centered
around middle age. The histograms of the Variables
Income and Assets Value also the skewed
rightwards which means that most people earn or
possess lesser income and assets value than others.
Credit Score and Debt to Income Ratio in this case,
has a more normal distribution since they have a
peak in the middle and symmetrical curves tending
towards zero on either extreme. Presenting the
results in this manner facilitates comparisons since
it reduces the variability in the distributions of the
other columns like Years at Current Job and
Number of Dependents to show more of the
distinct values at the respective peaks. These
observations make for interesting conclusions that
can be useful for data analysis that comes next and
includes modeling.
Figure 2: Distribution of the Risk Rating
The risk rating’s distribution is illustrated in Fig.
The x axis shows the risk rating categories which
are Low, medium and high while the y axis shows
the number of observations in each category. The
bar plot does show the number of observations that
belongs to Low risk, then Medium and High risks.
This implies that the dataset includes few numbers
of risky people or items according to their risk
ratings.
The correlation heatmap illustrates the connection
between numerical metrics of a data set. The color
intensity signifies the strength level and direction
to the correlation. The positive correlation is
represented by a red square while the negative
correlation is represented by a blue square. The
diagonal line of perfect correlation (1. 00) obtained
is expected. In other cases, some of the features will
have a low level of correlation with another feature
while some will have a high level of correlation. For
example, there is a weak negative relationship of
Assets Valued with Number of Dependents which
means that; when the number of household
dependents increases, the asset value is likely to be
low.
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
31
https://www.theamericanjournals.com/index.php/tajmei
Figure 4: Random Forest Classification Report
The Random Forest model has a high recall of class
1 at 1. 00, which highlights the classification of this
class as efficient. Specifically, it has zero precision
and recall for class 0 and zero recall for class 2
though it has reasonable accuracy for class 1. The
accuracy of the classification is 59% while the
macro average F1-score is low at 0. 25 indicating
that the car has a problem with balancing and it
tends to perform poorly in different classes.
Figure 5: SVM Classification Report
The SVM model successfully classify a total of 767 instances where out of 1000 samples it got a right
Figure 1: Correlation Matrix
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
32
https://www.theamericanjournals.com/index.php/tajmei
classification of 59.3% it’s not too different from
that attained using the Random Forest model. It has
1. 00 recall for class 1, however, it can neither
recognize class 0 and 2 meaning that it has 0
precision and recall for class 2 while the precision
and recall for class 0 exists but is equal to zero. The
macro average F1 score is 0. 25, indicating
performance issues.
Figure 6: XBoost Classification Report
The evaluated XGBoost model has an accuracy of
55.93%. The metric shows that it is quite good in
recall class 1=0. 88; however, it is very poor in
recognizing class 0=0. 01 and class 2=0.12. On the
macro average level, the F1-score if 0. 30, which as
a whole is below the optimal level of job
performance.
Figure 7: Logistic Regression Classification Report
Logistic Regression model achieves an accuracy of
59.3 % which is in line with both the SVM and
Random Forest models. The accuracy calculated
also reveal complete recall for class 1(response=1.
00) but miss both class 0 and class dt has zero
precision and recall on both class 0 and class 2.
respectively 0.50 for Macro-average F1 score and
this is below the average F1-score when all the
classes are considered at once. When the computed
total score is 25 it denotes that the firm has a poor
performance.
Table 1: ML Comparison Table
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
33
https://www.theamericanjournals.com/index.php/tajmei
To improve model performance, consider
addressing class imbalances using techniques like
SMOTE, which may enhance precision and recall
across
all
classes.
Experiment
with
hyperparameter tuning to optimize each model's
performance. Additionally, explore advanced
ensemble methods or hybrid models combining the
strengths of different algorithms. Regularly
evaluate and validate models with cross-validation
to ensure robust performance across various
scenarios and avoid overfitting.
Figure 8: Feedforward Neural Network Accuracy Graph
The training and validation accuracy of the
feedforward neural network gradually increase
initially and are nearly about the 60% in few
epochs and do not change considerably for the next
5 epochs. From the loss curves, the change in values
resembles those of the training and the validation
set in that the training as well as validation loss
decrease rapidly in the initial epochs successively
decreasing to a point close to one another. This is
an indication that the model does not over-train
and the loss on the validation set is similar to the
training loss.
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
34
https://www.theamericanjournals.com/index.php/tajmei
Figure 9: Deep Neural Network Accuracy Graph
As shown in Fig. 12 the deep neural network
representation of the model suffers from
overfitting. The training accuracy increases
dramatically, it reaches the level of 90% whereas,
the validation accuracy decreases after several
epochs, it fluctuates on the average of 55-60%.
Likewise, the training loss kept on declining slowly
but the validation loss first fell and then very
steeply rose up significantly. This suggests that
while the model is successfully capturing patterns
and ‘overfitting’ the training data it cannot do the
same for the validation set.
Comparison
When comparing the performance of different
models, several trends are identified. The
feedforward neural network shows good accuracy,
Training accuracy and Validation accuracy are
almost constant at around 60% after the first few
epochs and there is no sign of over fitting. On the
other hand, we have the deep neural network
where, although, the training accuracy is at 90%,
the validation accuracy is only at 55-60% which is
a clear show of overfitting where the model does a
good job in fitting the training data but does a very
bad job at the other data.
The value of accuracy is relatively low but
comparable across all the models; Random Forest,
SVM, XGBoost, and Logistic Regression all fall in the
range of 55%-59%. It can be seen that on average
these models have low macro-average F1-scores
for class 1 as they struggle with class 0 and class 2
respectively due to class imbalance. This means
that they are good in balanced classification where
they cannot handle minority classes hence are not
as useful in balanced classification problems.
Table 2:ML and ANN Model’s Accuracy Comparison
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
35
https://www.theamericanjournals.com/index.php/tajmei
DISCUSSION
The analysis results offer several insights on the
used dataset as well as evaluation of various
models. It can be readily observed from the
histograms that there exists variation in the
distribution of the data. For instance, the
distributions of Age, Income and Assets Value
indicate that Age is normally distributed
suggesting that most people falls within middle age
bracket, Income and Assets value are right skewed,
implying that more individuals have low income
and low asset values respectively. Credit Score and
Debt-to-Income Ratio are closer to be normally
distributed. It is useful for the purpose of
recognizing patterns that might impact on the
model in the future. Risk rating bar plot indicate
that the number of high-risk observations is less
while that of low risk and medium risk observation
are more suggesting that the most are of low to
medium risk.
The model evaluations reveal the positive change
to a certain extent success rate. The precision and
F1-measure of the Random Forest is very low for
both, class 0 and class 2 while the recall of class 1 is
very high, which gives an accuracy of 59% and a
macro-average F1 score of 0. 25. Similarly, the
model has observed again high recall for class 1
although the classes 0 and 2 are underrepresented
as similar to the observation made under the
confusion matrix, the accuracy of the model
including the SVM and the XGBoost models are low
and the class balancing is poor. The Logistic
Regression model has the similar performance
trends, the accuracy is 59.3 % along with problem
of misclassification of class 0 and 2.
Training and testing feedforward neural network
also strongly indicate that the accuracy increases
during the first epochs of training and validation
and then stabilize at 60 percent. The loss curves
indicate a large reduction in the early stages and
hence suggest that the models have learned well
and do not over-fit, as the validation loss curve and
the training loss curve are very close. On the other
hand, the deep neural network exhibits another
form of the model’s error called overfitting
since
the training accuracy improves to 90% while the
validation accuracy declines to 55-60%. Validation
loss rises rapidly after some fluctuations down,
which shows the fact that the model has bad ability
on data generalization for the validation set.
The feed forward neural network fluctuates
between 60 % accuracy and does not over train the
sample while the deep neural network has a 90 %
training accuracy but suffers from over training the
sample. Random Forest and SVM, XGBoost, and
Logistic Regression models have rather equal
accuracy (55-59%) and have poor results in the
minority classes. They all have low F1-scores
because of the effect of class imbalance for these
Model
Accuracy
(%)
Feedforward Neural
Network
60
Deep Neural Network
90
Random Forest
59
SVM
59
XGBoost
55.93
Logistic Regression
59.3
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
36
https://www.theamericanjournals.com/index.php/tajmei
models.
To address this the following strategies could be
employed: Over sampling or under sampling for
this case to ensure the different classes are
balanced in terms of sample size. Additionally, for
the deep neural network some hyperparameter
tuning could or applying of dropout and L2
regularization could enhance the learning model’s
performance for its generalization ability. Other
useful techniques that may assist in enhancing the
classification accuracy of the function for all the
classes include ensemble methods or boosting
techniques.
CONCLUSION
The research shows that mass and traditional
models of risk management are not very effective
because they are static models used for managing a
large amount of data, and do not focus on the
quickly changing environment of the financial
market. On the other hand, modern methods such
as the ML and ANNs possess greater benefits given
the fact that they can work with massive numbers
and look for intricate relationships. Out of the
evaluated models, feed forward neural network
and deep neural network are showing potential;
while feed forward has balanced learning
capability and can work with all three data sets,
deep learning network has high training accuracy
but overfits. Despite a stable performance of
Random Forest, SVM, and XGBoost, they scantly
work well with the minority classes. Combining of
ML and ANN into the framework of risk
management in the financial industry can help to
improve the prediction quality, control new risks,
and, therefore, improve the stability of the financial
sector.
REFERENCES
1.
R. S. Kaplan and A. Mikes, “Risk Management
-
the Revealing Hand,” Journal of Applied
Corporate Finance, vol. 28, no. 1, pp. 8
–
18, Mar.
2016,
doi:
https://doi.org/10.1111/jacf.12155.
2.
A. Deiva Ganesh and P. Kalpana, “Future of
artificial intelligence and its influence on
supply chain risk management
–
A systematic
review,” Computers & Industrial Engineering,
vol. 169, no. 169, p. 108206, Jul. 2022.
3.
A. J. Pitman et al., “Acute climate risks in the
financial system: examining the utility of
climate
model projections,” Environmental
Research: Climate, vol. 1, no. 2, p. 025002, Aug.
2022, doi: https://doi.org/10.1088/2752-
5295/ac856f.
4.
H. Benbya, T. H. Davenport, and S. Pachidi,
“Artificial Intelligence in Organizations:
Current State and Future O
pportunities,”
papers.ssrn.com, Dec. 03, 2020.
https://papers.ssrn.com/sol3/papers.cfm?abs
tract_id=3741983.
5.
T. Lynn, J. Mooney, P. Rosati, and M. Cummins,
“PALGRAVE STUDIES IN DIGITAL BUSINESS
AND ENABLING TECHNOLOGIES SERIES
EDITORS: Disrupting Finance FinTech and
Strategy in the 21st Century Edited by,” 2019.
Available:
https://library.oapen.org/bitstream/handle/
20.500.12657/23126/1007030.pdf?sequence
=1#page=54
6.
R. Y. Choi, A. S. Coyner, J. Kalpathy-Cramer, M.
F. Chiang, and J. P. Campbell, “Intr
oduction to
Machine Learning, Neural Networks, and Deep
Learning,” Translational Vision Science &
Technology, vol. 9, no. 2, pp. 14
–
14, Jan. 2020,
doi: https://doi.org/10.1167/tvst.9.2.14.
7.
S. Bhatore, L. Mohan, and Y. R. Reddy, “Machine
Learning Techniques for Credit Risk
evaluation: a Systematic Literature Review,”
Journal of Banking and Financial Technology,
vol. 4, no. 1, pp. 111
–
138, Apr. 2020, doi:
https://doi.org/10.1007/s42786-020-00020-
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
37
https://www.theamericanjournals.com/index.php/tajmei
3.
8.
T. Adrian, F. Grinberg, N. Liang, S. Malik, and J.
Yu, “The Term Structure of Growth
-at-
Risk,”
American Economic Journal: Macroeconomics,
vol. 14, no. 3, pp. 283
–
323, Jul. 2022, doi:
https://doi.org/10.1257/mac.20180428.
9.
G. Baryannis, S. Validi, S. Dani, and G. Antoniou,
“Supply Chain Risk Management
and Artificial
intelligence: State of the Art and Future
Research Directions,” International Journal of
Production Research, vol. 57, no. 7, pp. 1
–
24,
Oct.
2019,
doi:
https://doi.org/10.1080/00207543.2018.153
0476.
10.
J. Dou et al., “Improved landslide a
ssessment
using support vector machine with bagging,
boosting, and stacking ensemble machine
learning framework in a mountainous
watershed, Japan,” Landslides, vol. 17, no. 3,
pp.
641
–
658,
Oct.
2019,
doi:
https://doi.org/10.1007/s10346-019-01286-
5.
11.
Y
. Xia, X. Guo, Y. Li, L. He, and X. Chen, “Deep
learning meets decision trees: An application
of a heterogeneous deep forest approach in
credit scoring for online consumer lending,”
vol. 41, no. 8, pp. 1669
–
1690, Jul. 2022, doi:
https://doi.org/10.1002/for.2891.
12.
R. Sweet et al., “Credit Risk Assessment Using
Statistical and Machine Learning: Basic
Methodology and Risk Modeling Applications,”
International Journal on Computational
Engineering, vol. 1, no. 3, pp. 62
–
67, 2024, doi:
https://doi.org/10.62527/comien.1.3.21.
13.
L. Yang and A. Shami, “On hyperparameter
optimization of machine learning algorithms:
Theory and practice,” Neurocomputing, vol.
415, pp. 295
–
316, Nov. 2020, doi:
https://doi.org/10.1016/j.neucom.2020.07.0
61.
14.
A. Sanyal, Dokania, Puneet K, V. Kanade, and
Philip, “How benign is benign overfitting?,”
arXiv.org,
2020.
https://arxiv.org/abs/2007.04028 (accessed
Sep. 12, 2024).
15.
D. Marcek, “Forecasting of financial data: a
novel fuzzy logic neural network based on
error-correcti
on concept and statistics,”
Complex & Intelligent Systems, vol. 4, no. 2, pp.
95
–
104,
Sep.
2017,
doi:
https://doi.org/10.1007/s40747-017-0056-6.
16.
T. Sun and M. A. Vasarhelyi, “Predicting credit
card delinquencies: An application of deep
neural networks,” Intelligent Systems in
Accounting, Finance and Management, vol. 25,
no. 4, pp. 174
–
189, Aug. 2018, doi:
https://doi.org/10.1002/isaf.1437.
17.
E. Dumitrescu, S. Hué, C. Hurlin, and S. Tokpavi,
“Machine Learning for Credit Scoring:
Improving Logistic Regression with Non-
Linear Decision-
Tree Effects,” European
Journal of Operational Research, vol. 297, no. 3,
Jun.
2021,
doi:
https://doi.org/10.1016/j.ejor.2021.06.053.
18.
M. Chen, U. Challita, W. Saad, C. Yin, and M.
Debbah, “Artificial Neural Networks
-Based
Machine Learning for Wireless Networks: A
Tutorial,” IEEE Communications Surveys &
Tutorials, vol. 21, no. 4, pp. 3039
–
3071, 2019,
doi:
https://doi.org/10.1109/comst.2019.292662
5.
19.
M. Selvaggi, “A new metric for the
interpretability of artificial neural networks in
medical diagnosis applications. -
Webthesis,”
Polito.it,
Oct.
2022,
doi:
https://webthesis.biblio.polito.it/secure/245
21/1/tesi.pdf.
20.
M. Soui, I. Gasmi, S. Smiti, and K. Ghédira, “Rule
-
based credit risk assessment model using
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
38
https://www.theamericanjournals.com/index.php/tajmei
multi-
objective evolutionary algorithms,”
Expert Systems with Applications, vol. 126, pp.
144
–
157,
Jul.
2019,
doi:
https://doi.org/10.1016/j.eswa.2019.01.078.
21.
S. Marso and M. El Merouani, “Predicting
financial distress using hybrid feedforward
neural network with cuckoo search algorithm,”
Procedia Computer Science, vol. 170, pp. 1134
–
1140,
2020,
doi:
https://doi.org/10.1016/j.procs.2020.03.054.
22.
M. Satheesh and S. Nagaraj, “Applications of
Artificial Intelligence on Customer Experience
and Service Quality of the Banking Sector,”
International Management Review, vol. 17, no.
1,
p.
2021,
2021,
Available:
http://www.americanscholarspress.us/journ
als/IMR/pdf/IMR-1-2021/V17n121-art2.pdf
23.
Y. Lu, “Artificial intelligence: a survey on
evolution, models, applications and future
trends,” Journal of Management Analytics, vol.
6, no. 1, pp. 1
–
29, Jan. 2019, doi:
https://doi.org/10.1080/23270012.2019.157
0365.
24.
D. van Thiel and W. F. van Raaij, “Artificial
Intelligent Credit Risk Prediction: An Empirical
Study of Analytical Artificial Intelligence Tools
for Credit Risk Prediction in a Digital Era,”
Journal of Accounting and Finance, vol. 19, no.
8,
Dec.
2019,
Available:
https://articlearchives.co/index.php/JAF/arti
cle/view/68
25.
M. T. Keane and E. M. Kenny, “How Case
-Based
Reasoning Explains Neural Networks: A
Theoretical Analysis of XAI Using Post-Hoc
Explanation-by-Example from a Survey of
ANN-CBR
Twin-
Systems,”
Case
-Based
Reasoning Research and Development, pp.
155
–
171,
2019,
doi:
https://doi.org/10.1007/978-3-030-29249-
2_11.
26.
V. Beaudouin et al., “Flexible and Context
-
Specific AI Explainability: A Multidisciplinary
Approach,” arXiv.org,
Mar. 13, 2020.
https://arxiv.org/abs/2003.07703
27.
S. Bramhall, H. Horn, M. Tieu, and N. Lohia,
“QLIME
-A Quadratic Local Interpretable
Model-
Agnostic Explanation Approach,” SMU
Data Science Review, vol. 3, no. 1, Apr. 2020,
Available:
https://scholar.smu.edu/datasciencereview/v
ol3/iss1/4/
28.
C. S. Bojer and J. P. Meldgaard, “Kaggle
Forecasting competitions: an Overlooked
Learning Opportunity,” International Journal
of Forecasting, vol. 37, no. 2, Sep. 2020, doi:
https://doi.org/10.1016/j.ijforecast.2020.07.
007.
29.
D. Chicco, L. Oneto, and E. Tavazzi, “Eleven
quick tips for data cleaning and feature
engineering,” PLOS Computational Biology,
vol. 18, no. 12, p. e1010718, Dec. 2022, doi:
https://doi.org/10.1371/journal.pcbi.101071
8.
30.
Pronaya Prosun Das and L. Wiese,
“Explainability Based on Feature Importance
for Better Comprehension of Machine Learning
in Healthcare,” Communications in computer
and information science, pp. 324
–
335, Jan.
2023, doi: https://doi.org/10.1007/978-3-
031-42941-5_28.
31.
M. Belgiu and L. Drăguţ, “Random forest in
remote sensing: A review of applications and
future
directions,”
ISPRS
Journal
of
Photogrammetry and Remote Sensing, vol.
114, no. 114, pp. 24
–
31, Apr. 2016, doi:
https://doi.org/10.1016/j.isprsjprs.2016.01.0
11.
32.
G. A. Lujan-Moreno, P. R. Howard, O. G. Rojas,
and D. C. Montgomery, “Design of experiments
and response surface methodology to tune
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
39
https://www.theamericanjournals.com/index.php/tajmei
machine learning hyperparameters, with a
random forest case-
study,” Expert Systems
with Applications, vol. 109, pp. 195
–
205, Nov.
2018,
doi:
https://doi.org/10.1016/j.eswa.2018.05.024.
33.
H. Lingjun, R. Levine, Fan, Juanjuan, J. Beemer,
and J. Stronach, “Random Forest as a Predictive
Analytics Alternative to Regression in
Institutional Research,” Pr
actical Assessment,
Research, and Evaluation, vol. 23, 2018, doi:
https://doi.org/10.7275/1wpr-m024.
34.
Y. Geng, Q. Li, G. Yang, and W. Qiu, “Logistic
Regression,” pp. 99–
132, Jan. 2024, doi:
https://doi.org/10.1007/978-981-97-3954-
7_4.
35.
M. Mahbobi, S. Kimiagari, and M. Vasudevan,
“Credit risk classification: an integrated
predictive accuracy algorithm using artificial
and deep neural networks,” Annals of
Operations
Research, Jul.
2021,
doi:
https://doi.org/10.1007/s10479-021-04114-
z.
36.
R. Morad
i, R. Berangi, and B. Minaei, “A survey
of regularization strategies for deep models,”
Artificial Intelligence Review, vol. 53, no. 6, pp.
3947
–
3986,
Dec.
2019,
doi:
https://doi.org/10.1007/s10462-019-09784-
7.
37.
S. Kost, O. Rheinbach, and H. Schaeben,
“Using
logistic regression model selection towards
interpretable machine learning in mineral
prospectivity modeling,” Geochemistry, p.
125826,
Oct.
2021,
doi:
https://doi.org/10.1016/j.chemer.2021.1258
26.
38.
S. Ghosh, A. Dasgupta, and A. Swetapadma, “A
Study on Support Vector Machine based Linear
and Non-
Linear Pattern Classification,” IEEE
Xplore,
Feb.
01,
2019.
https://ieeexplore.ieee.org/abstract/docume
nt/8908018
39.
B. Ghaddar and J. Naoum-
Sawaya, “High
dimensional data classification and feature
s
election using support vector machines,”
European Journal of Operational Research, vol.
265, no. 3, pp. 993
–
1004, Mar. 2018, doi:
https://doi.org/10.1016/j.ejor.2017.08.040.
40.
R. Mitchell and E. Frank, “Accelerating the
XGBoost algorithm using GPU comp
uting,”
PeerJ Computer Science, vol. 3, p. e127, Jul.
2017, doi: https://doi.org/10.7717/peerj-
cs.127.
41.
N.-H. Nguyen, J. Abellán-García, S. Lee, E.
Garcia-
Castano, and T. P. Vo, “Efficient
estimating compressive strength of ultra-high
performance conc
rete using XGBoost model,”
Journal of Building Engineering, vol. 52, p.
104302,
Jul.
2022,
doi:
https://doi.org/10.1016/j.jobe.2022.104302.
42.
T. Kavzoglu and A. Teke, “Advanced
hyperparameter optimization for improved
spatial prediction of shallow landslides using
extreme gradient boosting (XGBoost),” Bulletin
of Engineering Geology and the Environment,
vol.
81,
no.
5,
Apr.
2022,
doi:
https://doi.org/10.1007/s10064-022-02708-
w.
43.
Z. Xu, C. Sun, T. Ji, J. H. Manton, and W. Shieh,
“Feedforward and Rec
urrent Neural Network-
Based Transfer Learning for Nonlinear
Equalization in Short-
Reach Optical Links,”
Journal of Lightwave Technology, vol. 39, no. 2,
pp.
475
–
480,
Jan.
2021,
Available:
https://opg.optica.org/abstract.cfm?uri=jlt-
39-2-475
44.
S. Moldovanu, C.-D. Obreja, K. C. Biswas, and L.
Moraru, “Towards Accurate Diagnosis of Skin
Lesions Using Feedforward Back Propagation
Neural Networks,” Diagnostics, vol. 11, no. 6, p.
936,
May
2021,
doi:
THE USA JOURNALS
THE AMERICAN JOURNAL OF MANAGEMENT AND ECONOMICS INNOVATIONS (ISSN- 2693-0811)
VOLUME 06 ISSUE10
40
https://www.theamericanjournals.com/index.php/tajmei
https://doi.org/10.3390/diagnostics1106093
6.
45.
F. Farhadi, Nia, Vahid Partovi, and A. Lodi,
“Activation Adaptation in Neural Networks,”
arXiv.org,
2019.
https://arxiv.org/abs/1901.09849 (accessed
Sep. 12, 2024).
46.
V. Asghari, Y. F. Leung, and S.-
C. Hsu, “Deep
neural network based framework for complex
correlations in engineering metrics,” Advanced
Engineering Informatics, vol. 44, p. 101058,
Apr.
2020,
doi:
https://doi.org/10.1016/j.aei.2020.101058.
47.
I. H. Sarker, “Deep Learning: a Comprehensive
Overview
on
Techniques,
Taxonomy,
Applicati
ons and Research Directions,” SN
Computer Science, vol. 2, no. 6, Aug. 2021, doi:
https://doi.org/10.1007/s42979-021-00815-
1.
