International Journal of Information Management Data Insights 5 (2025) 100324

Contents lists available at ScienceDirect

International Journal of Information

Management Data Insights
journal homepage: [Link]/locate/jjimei

Machine learning in banking risk management: Mapping a decade

of evolution
Valentin Lennart Heß , Bruno Damásio *
Nova Information Management School (NOVA IMS), Universidade Nova de Lisboa, Campus de Campolide, 1070-312 Lisboa, Portugal

A R T I C L E I N F O A B S T R A C T

Keywords: The techniques used in banks’ risk management are evolving as opposed to the process of risk management. It is
Machine learning necessary to respond to these market- and technology-driven changes appropriately. Innovative approaches are
Bank needed to overcome the limitations of traditional methods. Machine learning (ML) algorithms are suitable for
Risk management
dealing with the various risk types banks face. Academic literature focuses on applying ML in credit risk man­
Algorithm
Artificial intelligence
agement. This article addresses market, operational, liquidity, and other risk types, with the objective to examine
how ML algorithms predict, assess, and mitigate these risks and identify both their advantages and challenges.
This article systematically reviews 46 recent studies and highlights the expanding role of ML in enhancing risk
management strategies. The article has revealed that ML is adequately covered in the context of market and
operational risk. The learning ability and predictive capabilities of artificial neural networks and other algo­
rithms are promising for risk management. Our findings offer a concise overview of current ML applications for
multiple risk types in banking, identifying research gaps, highlighting opportunities and challenges and
providing actionable directions for further studies. By providing a focused overview of the expanding role of ML
in banking risk management, we underscore the potential to strengthen the robustness of banks’ strategies and
practices.

1. Introduction sustainability in a dynamic environment. Risk management is subject

to constant change, characterized by varying market conditions, regu­
Machine learning (ML) has become a transformative force in banking lations, and advances in practices and technology. This risk management
risk management, evolving significantly over the last decade. This evolution synergizes strongly with ML properties and developments.
article addresses the broad implications of ML applications across The implementation of ML techniques can have significant benefits on
various risk types in banking, highlighting the gaps left by previous the quality and execution of risk management in comparison to classical
research that primarily focused on credit risk. By exploring market, approaches. By leveraging advanced pattern recognition, data analysis,
operational, and liquidity risks, this study aims to provide a wide- and predictive modeling, ML allows banks to enhance their risk miti­
ranging overview of ML’s potential to enhance risk management stra­ gation by predicting and treating potential threats (Leo et al., 2019). To
tegies, thus contributing to the stability and resilience of financial maintain control over risks, banks employ various methods, tools, and
institutions. techniques that increasingly leverage ML. These techniques include
In fact, banks face various factors and events, such as market vola­ stress testing, risk assessment, portfolio analysis, setting risk limits,
tility, credit defaults, operational failures, and regulatory changes that scoring systems, estimating Value at Risk (VaR) and Expected Shortfall
could harm their operational stability, profitability, or reputation. To (ES), and many more (Milojević & Redzepagic, 2021).
prevent these threats, banks must identify, assess, and mitigate them by Several studies have intensively researched ML in risk management,
employing risk management practices to anticipate and respond to these and further developments are expected. The literature categorizes
challenges. Stable risk management is essential for a bank to deal with banks’ risks into four major categories: credit, market, operational,
these challenges, especially in times of crisis. Effective risk management liquidity, and other risk types (Milojević & Redzepagic, 2021). Most
safeguards investors’ interests, enhances trust, and ensures long-term studies and literature reviews deal with the application of machine

* Corresponding author.
E-mail address: bdamasio@[Link] (B. Damásio).

[Link]

Available online 30 January 2025

2667-0968/© 2025 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license ([Link]
nc-nd/4.0/).
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

learning in credit risk management, which is one of the most significant housing market and the global financial system (Gorton & Metrick,
risks for banks (Bhatore et al., 2020; Chen et al., 2017; Shi et al., 2022; 2012). With the introduction of sophisticated risk management, banks
Tyagi, 2022). However, research is lacking in other risk areas (Leo et al., nowadays seek to prevent such events from recurrence or better manage
2019). Since this finding, more research on other risk types has been the consequences of these incidents. Risk management aims to protect
done, and machine learning models have been trained and tested but not assets, maintain liquidity, and ensure profitability and compliance by
uniformly summarized. This research gap serves as a guideline for this identifying, assessing, and managing different risks (Alazzabi et al.,
study, leading to the following research question: 2023). Therefore, banks’ risk management is paramount for economic
activity and existence in the market; thus, they must consider several
RQ: How is machine learning utilized in scientific literature in influencing factors. Loan borrowers could fail to make their payments,
dealing with market, operational, liquidity, and other risks within market interest rates could change and affect bank loan values, banks’
banking risk management? investments could lose value or human input errors, or fraud could affect
the business (Apostolik et al., 2012). These examples show that banks
The research objective is to identify application areas in risk man­ must be prepared for several types of risks to be able to act according to
agement that can benefit from machine learning. Additionally, an the situation.
objective is to develop a guideline for banks, showing which machine Due to the great importance of banks for society, they are subject to
learning algorithms can be used in the respective areas and risk types. state regulations, which in Europe are defined by the frameworks Cap­
The methodology consists of a qualitative and quantitative approach. A ital Requirements Directive (CRD IV) and Capital Requirements Regu­
systematic literature review (SLR) is conducted on scientific databases lations (CRR) of the European Parliament. This regulatory requirement
using the Preferred Reporting Items for Systematic Reviews and Meta- package implements the Basel III framework (DEPC, 2023; DEPC 2023).
Analysis framework (PRISMA), and the data is analyzed based on The Basel III Framework is an international set of reforms published by
qualitative and quantitative aspects (Page et al., 2021). By following the the Basel Committee on Banking Supervision in response to the
proposed guidelines, a collection of scientific work is developed and 2007/2008 financial crisis to enhance regulation, supervision, and risk
evaluated on aspects such as used algorithms, risk categories, perfor­ management in the banking sector (Chockalingam et al., 2018). Basel III
mance metrics, practical results, and more. This research will unveil the is a set of regulations that describes measures to strengthen the banking
spectrum of machine learning capabilities in banking risk management. sector and make it more resilient by developing new capital re­
These findings will contribute to the ongoing academic discourse on quirements that provide banks with sufficient capital for future crises
cutting-edge technologies in the financial sector. The empirical results and different risk types (Varotto, 2011, BCBS 2023).
can help banks develop comprehensive guidelines for integrating ma­ Since risks of various types affect the banking industry, the regula­
chine learning into their risk management strategies. It may also un­ tions stipulate that banks must have capital set aside for each type of risk
cover challenges and opportunities that inspire further research. to absorb them (Leo et al., 2019). The size and risk of a bank´s assets are
This study is structured as follows: The theoretical background in­ the most critical factors in determining how much capital the bank must
formation is covered in the second section, together with related work hold for specific risk types. In the literature, a wide range of risk types
that highlights the study’s relevance. This section provides a general are discussed. The main types of risks are credit, market, and operational
overview of risk management in banks and explains the different types risks. The different risk types vary in intensity, reflected in the risk
of risks. Additionally, machine learning is introduced, and the most distribution functions triggered by various factors (Apostolik et al.,
important algorithms used in the banking sector are outlined. Section 2012, Rosenberg and Schuermann, 2006). More recent studies have
three deals with the research methodologies. A more detailed explana­ shown that liquidity risk is another vital aspect banks need to care about
tion of the systematic literature review approach is provided. Next, the as a systematic risk factor (Dang and Nguyen, 2020). Additionally,
answers to the research question are presented in the results section. The regulatory and strategic risks are listed in the literature, as well as many
results are analyzed in more detail in the discussion section, and prac­ more types like reputational risk (Leo et al., 2019, Chockalingam et al.,
tical implications, limitations, and further research directions are 2018, Aziz and Dowling, 2019).
pointed out and finally summarized.
2.1.1. Risk types
2. Theoretical background It is essential to look closely at the individual risk types mentioned to
identify the potential for using machine learning in a bank´s risk man­
2.1. Risk management foundations agement. This exercise is necessary to reveal the exact focus areas and
application domains.
Financial institutions such as banks play a crucial role in society. Credit risk, also called credit default risk, can be defined as the
They act as intermediaries for the circulation of money and credit in the possibility that a borrower or the counterparty will not meet its liabil­
economy and private households. Thus, they offer a place to store money ities according to the agreed terms. This risk is the most significant risk
and provide various other services, such as giving credit loans, mort­ of most banks and arises when either credits or bonds disbursed by the
gages, and many more amenities. Banks primarily contribute to a soci­ bank are either partially or fully not repaid (Apostolik et al., 2012).
ety’s economic growth and stability by enabling lending, promoting Banks review loan applications, require securities, diversify risk, and
investment, and supporting financial stability (Chockalingam et al., analyze credit risk to reduce the inherent risk. This analysis is used to
2018). Banks are exposed to various internal and external risks, like evaluate the risk posed by the borrower to determine the interest rate
those faced by ordinary companies. Risk can be defined as the proba­ that will be charged. This risk depends on the borrower´s behavior,
bility of losses resulting from different events (Horcher, 2005). Risk ability, and willingness to repay the loan. Thus, when granting a loan,
management practices seek to protect the financial system from risks the borrower’s creditworthiness is checked, which depends on their
like credit defaults, market volatility, or operational and liquidity risks statistical and personal background (Chen, 2023). In the literature,
(Leo et al., 2019). However, a significant difference to regular com­ credit risk is traditionally considered the most significant risk, requiring
panies is that risk-triggered failures in the financial system can lead to the most capital to be assigned (Leo et al., 2019).
systemic crises, as was observed in the global financial crisis of The risk of incurring losses to a bank resulting from movements in
2007/2008. A failure in bank risk management led to one of the most market prices due to changes in foreign exchange rates, interest rates,
significant financial crises of the last century. It happened due to banks and equity and commodity prices is called market risk (Apostolik et al.,
accepting high risks in the form of risky mortgage loans, performing 2012). In the Basel III framework, it is described as the risk of losses in
financial engineering, and poor regulation, which led to a collapse of the the on-balance sheet and off-balance sheet risk positions due to market

2
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

price fluctuations (BCBS 2023). A bank needs to monitor market risk as algorithms because the used model might increase the model risk. As­
it affects sensitive business segments. For example, interest rate changes pects like interpretability, data sensitivity, or overfitting of ML models
are crucial as they can lead to a divergence between revenues and assets can impact the model risk (Aziz and Dowling, 2019).
and the upcoming cost of liabilities. Interest rate changes can also
impact financial securities held by banks or influence customer 2.2. Machine learning principles
behavior. A similar impact on banks’ profitability can be seen in the
exchange rates of foreign currencies. When banks trade equities or Machine learning can be understood as a subset of artificial intelli­
commodities, these can also be subject to fluctuations that directly affect gence that makes it possible for computer systems to learn from data,
profitability (Chen, 2023). The most widely used metric to measure identify patterns, and predict outcomes without explicitly programming
market risk is Value at Risk. It uses statistical probability estimates to them so that decisions can be made on this basis (Pallathadka et al.,
assess the potential of financial losses, usually of a portfolio of assets or 2023). Machine learning can detect meaningful patterns in data and thus
exposures (Horcher, 2005). extract useful information by learning and adapting. The quality of the
Operational risk is defined in the Basel III framework as the risk of predictions strongly depends on the quality and volume of the data (Leo
loss due to insufficient or incorrect internal processes, people, and sys­ et al., 2019). One possible classification of machine learning models is
tems or due to external events (BCBS 2023). Therefore, this risk covers the division into four types, namely supervised, unsupervised, rein­
many events that impact the business. Errors in internal procedures, forcement, and ensemble learning. Supervised models use labeled data
such as in transactions, regulatory compliance, and human error, to train a model and make predictions like classification and regression
contribute to operational risk. In addition, technical system and soft­ algorithms. Classification is the process of mapping an input collection
ware errors or data protection breaches are assigned to operational risk. of several instances into a unique collection of characteristics, also called
Furthermore, external risks such as regulatory changes, cyber-attacks, or targets or labels (Pallathadka et al., 2023). Accordingly, the classifier
natural disasters are added (Aziz and Dowling, 2019). Operational risk uses the information of labeled data inputs to classify a new observation
covers legal risks but does not include strategic or reputational risks by learning the relationships between these inputs. Examples of classi­
(Apostolik et al., 2012). Due to the technology-driven age characterized fication models are Artificial Neural Networks (ANN), Decision Trees
by constant change, operational risks have become increasingly (DT), Support Vector Machines (SVM), and many more. Unsupervised
important in recent decades (Milojević and Redzepagic, 2021). learning models identify patterns in unlabeled datasets to extract
Liquidity risks pose a significant challenge for banks and arise when exploratory information and perform actions like grouping similar ob­
they cannot execute transactions at reasonable market prices due to servations in clusters or storing data patterns. Exemplary unsupervised
asset liquidity (Dang and Nguyen, 2020, Johri et al., 2022). A bank must learning algorithms are k-means clustering, hierarchical clustering, or
ensure enough liquidity to meet credit demands and withdrawals from self-organizing maps (SOM), among others. Reinforcement Learning
depositors, which is crucial to meeting short-term financial obligations (RL) algorithms learn by interacting with an environment to maximize a
(Swankie and Broby, 2019). The Basel III framework includes two reward or achieve a goal. Therefore, it is a kind of dynamic program­
quantitative ratios that measure liquidity risk. The Liquidity Coverage ming used, for example, in portfolio allocation or trade execution over a
Ratio, which can be seen as a stress test, aims to ensure that a bank holds specific time horizon (Dixon et al., 2020). Ensemble learning is a tech­
sufficient assets to meet all ongoing short-term requirements within 30 nique that combines multiple models to improve predictive performance
days. Furthermore, the Net Stable Funding Ratio, designed to promote and enhance accuracy (Kubat, 2017).
medium and long-term liquidity funding for banks, must demonstrate The SLR conducted in this study analyzes research papers that have
sufficiently stable funding sources. It is calculated as the ratio of the implemented the following steps and actions. To guarantee compara­
available amount of stable funding to the required amount of stable bility and, thus, unity, it is essential to outline the implementation
funding. Both ratios should be at least 100 % (BCBS 2023, Swankie and process, as critical differences in the research papers will be revealed.
Broby, 2019). Before applying machine learning models to a dataset, it is crucial to
In addition to the main types of risk, there are numerous other prepare the data so the algorithms can process it without problems. This
essential risks that banks should consider. Reputational risk is the po­ preparation guarantees that qualitative data is made available to the
tential loss resulting from a reduction in a bank´s reputation in the public model. This data process is called data preprocessing. Before this begins,
domain. This risk is also of great importance, as damage to reputation it is vital to divide the data set into training, validation, and test data.
causes customers to lose confidence in the bank, and possible customer This is necessary so that the model also works well with new data. Only
losses may come as a result (Apostolik et al., 2012). Strategic risk in­ data that has not been used during training should be tested to evaluate
cludes incorrect or improper implementation of a strategy and a lack of how well the algorithm performs on new data. Adjustments can be made
response to changes in the business environment. Regarding strategy with the validation data set by fine-tuning different parameters before
risk, the bank´s risk profile is generally based on whether the strategy is the algorithm is thoroughly tested. The next step is to identify and
moving in the right direction, being implemented correctly, and whether handle missing values and outliers. In addition, the so-called feature
the respective objectives are reasonable (Chockalingam et al., 2018). In engineering takes place, adding value by creating new variables based
a related but different category lies the business risk, characterized by on existing ones.
the potential financial loss due to the weakening of the competitive Furthermore, categorical data is encoded into numerical data. Some
position and market changes (Apostolik et al., 2012). Both risk types machine learning algorithms can only process data within a specific
relate to the profitability and economic direction of the bank, but they range. Therefore, in some cases, the variables must be scaled to the same
have a crucial difference. The distinction is that strategic risk looks at a range. Finally, only variables that add value to the machine learning
long-term overarching period with the strategic orientation of the bank. algorithm, disregarding unimportant variables, are selected in the
Business risk considers immediate operational business changes and feature selection step. All these steps are done to ensure the quality of
their impact (Chockalingam et al., 2018). Regulatory and legal risks the machine learning algorithm to avoid overfitting and create robust­
relate to costs arising from changes in the regulatory environment in the ness. A model is overfitting when the algorithm performs well on
form of new laws and banking regulations (Kelliher et al., 2013). This training data but badly on test data since it learns the irregularities of the
risk, in particular, affects all risk types and is therefore considered training data set, which do not occur in new data sets (Kotsiantis et al.,
individually based on its impact and significance (Aziz and Dowling, 2006, García et al., 2016).
2019). An additional significant risk is the model risk, which measures Due to the reliance on mathematics, probability theory, and statis­
the incompleteness or incorrectness of used risk management and tics, machine learning might produce incorrect classifications (Milana
assessment methods. This risk is crucial for the application area of ML and Ashta, 2021). The performance of machine learning algorithms

3
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

must be measured to keep track of these errors. This is done with Decision Tree-based classifiers that average the results of many trees,
so-called performance metrics or measures. These performance metrics thus improving the predictive accuracy and decreasing the outcome’s
determine the quality of a model based on the correctly and incorrectly interpretability, resulting in a black box (Breiman, 2001). RFs offer high
classified records. The most common performance metric is the F1 accuracy and robustness, making them useful for complex classification
Score. It is the harmonic mean of precision and recall and is suitable for tasks in the financial domain González-Carrasco et al., (2019). One of
unequal class distributions, as in the case of most machine learning the common machine learning algorithms in the financial sector is the
projects (Guerra et al., 2022). In addition, accuracy is another important Logistic Regression model used for binary classification tasks. The model
performance metric for machine learning models. The following for­ determines the probability that an input belongs to a specific category. A
mulas from Dixon et al. (2022) show how each metric is calculated. The logistic transformation function is applied to a feature set, which maps
associated meanings are described in Table 2.1. them into a range of probabilities between 0 and 1, whereby 1 equals
100 percent probability of belonging to the corresponding class. The
TP TP 2*Precission* Recall
Precision = , Recall = , F1 Score = , maximum likelihood function, which maximizes the prediction simi­
TP + FP TP + FN Precision + Recall
larity to the actual outcome, is applied to find the best parameters to
train and fit the model to the input data (Hosmer et al., 2013).
TP + TN
Accuracy = Artificial Neural Networks are models inspired by the functioning of
All Samples a human brain. A neural network typically consists of several layers,
The integration of ML into banking risk management represents a where information from the external environment is inserted into the
paradigm shift from traditional methods to more adaptive and predic­ input layer. This information is processed in hidden layers and returned
tive frameworks. This transformation enables banks to better anticipate to an output layer. The structure of an ANN is displayed in Fig. 2.1. Each
and mitigate potential threats, leveraging advanced algorithms for data such layer consists of multiple neurons that receive, process, and pass on
analysis and pattern recognition. The broad implications of this shift data. When a neuron receives an input, a weighted sum is applied to that
include improved accuracy in risk prediction, enhanced decision- input and adds a bias term. This sum is then passed through an activa­
making capabilities, and a more robust approach to managing diverse tion function that produces the output. Learning algorithms such as
risk types. backpropagation and optimization techniques like gradient descent are
applied to learn from the data and minimize loss functions by updating
2.2.1. Models weights (Milana and Ashta, 2021). This architecture makes it possible to
Several machine learning models are already being applied in today´s recognize complex structures and relationships of variables, including
risk management, and new models and approaches are being investi­ nonlinear relationships. A common example of a typical feedforward
gated in scientific research. In the following, the core principles of the network is the Multi-Layer Perceptron (MLP), where information flows
most essential algorithms in finance are further analyzed. from input to output in one direction through at least one hidden layer.
Supervised Learning Algorithms In Recurrent Neural Networks (RNN), data flows in both directions and
Support Vector Machines are widely used in finance for classification sequential information about a hidden state is stored, preserving past
tasks like analyzing stock markets (Rundo et al., 2019). The SVM algo­ information. Popular types of RNNs are long short-term memory (LSTM)
rithm aims to find a hyperplane that correctly classifies the input data and gated recurrent unit (GRU) neural networks. LSTM models have
while maximizing the distance between the different classes. Since an memory cells that retain information for long durations, enabling the
SVM is also a supervised classification algorithm, mathematical opti­ evaluation of hidden non-linear correlations and capturing long-range
mizations must first be performed on a labeled training data set. Support dependencies (Rundo et al., 2019). A GRU is more simplified than the
vectors are the training examples that determine the maximum margin LSTM, as it combines the memory and forget cells, resulting in a leaner
of the SVM. The margin is the distance between these support vectors. architecture that is computationally less expensive (Dutta et al., 2020).
SVMs are especially effective in handling high-dimensional data and Another ANN is the Convolutional Neural Network (CNN). This type can
nonlinear relationships (Cortes and Vapnik, 1995). By adjusting various automatically adapt and learn spatial hierarchies of features from the
parameters, the dimension of the input training data can be changed, input data. A CNN is particularly effective for tasks that have a grid-like
and thus, the optimal hyperplane can be found. Therefore, overfitting structure because it can extract local translation-invariant features from
can be avoided since the margin of the support vectors is maximized. images, time-series data, or spatial data.
Since an SVM has a finite number of controlling parameters, conver­ Additionally, they can learn spatial hierarchies for data with com­
gence can be reached faster, unlike other machine learning algorithms. plex hierarchical patterns like images (Mashrur et al., 2020). A neural
As the financial industry deals with many highly complex data and network with many hidden layers is defined as a deep learning model
nonlinear trends, SVMs are suitable for overcoming inconsistent and and is even better suited for complex tasks. However, neural networks
unpredictable performance on noisy data and reliably learning and are considered black-box models, as the increasing complexity with
recognizing patterns (Rundo et al., 2019). multiple layers makes it difficult to interpret the decision-making and
A Decision Tree is a highly interpretable machine learning algorithm functioning (Giudici et al., 2023).
that splits the dataset based on the values of features in a recursive Bayesian methods are statistical techniques for clustering, classifi­
manner as an objective to maximize the reparation of the target variable. cation tasks, and general probabilistic modeling. By updating proba­
At each tree node, a decision is made, which forms leaves that produce bilities based on prior knowledge and new findings, Bayesian methods
the final prediction (Kubat, 2017). Since Decision Trees sometimes enable the estimation of unknown quantities and the creation of pre­
derive complex and deep tree structures from the input data, they tend dictions. A popular algorithm in the financial sector is the Naïve Bayes
to overfit. That is why Random Forests (RFs) were developed. RFs are Classifier. It calculates the likelihood that a data point belongs to a
particular class using Bayes’ theorem based on the features. The algo­
rithm assumes that the individual features are independent, simplifying
Table 2.1 the calculations but making the algorithm ‘naïve’ (Berrar, 2018).
Confusion matrix. Another important Bayesian method is the Bayesian Network. It repre­
Predicted sents relationships of variables using a directed graph and conditional
probability distributions. These distributions show how one variable
Actual 1 0
might influence another one. So, it tells the likelihood of each variable to
1 True Positive (TP) False Positive (FP) be in a particular state based on the states of the other variables. This
0 False Negative (FN) True Negative (TN)
makes it possible for complex reasoning under uncertainty (Lowd and

4
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

Fig. 2.1. Artificial Neural Network structure from Dixon et al., (2020).

Domingos, 2005). aim to optimize the policy that specifies the agent´s behavior without a
Unsupervised Learning Algorithms value function. These methods work with stochastic policies that define
The K-Means clustering algorithm is a model of unsupervised probability distributions over a set of possible actions. REINFORCE and
learning that attempts to group data into K distinct clusters so that the Actor-Critic Methods are algorithms that belong to this category.
sum of squares within a cluster is minimal. It tries to find local optima, Model-based methods are based on an explicit internal model of the
which are states where moving a data point to another cluster does not environment. This is used to simulate future states and rewards to make
cause a decrease in the sum of squares of the cluster. It starts with a better decisions. Model-based methods may be more efficient on a
certain number of K clusters and a random selection of K starting points, sampling basis but require more computing power. Monte Carlo Tree
called centroids. Each data point is added to the nearest starting point, Search or dynamic programming can be categorized as model-based
forming the first clusters. Cluster centers are then determined and set as methods (Dixon et al., 2020, Gašperov et al., 2021). Model-free
the next centroids. This is repeated until the centroids do not change or a methods, on the other hand, do not rely on a model of the environ­
specified iteration count is reached (Hartigan and Wong, 1979). Prin­ ment. They learn to make decisions directly from interactions with the
cipal Component Analysis (PCA) is a technique from multivariate sta­ environment. This is done to simplify the implementation and to
tistics. It is strictly not a machine learning algorithm but is often used in improve scalability. An example of model-free methods is Deep Q-Net­
conjunction with other machine learning models in Finance and is thus works, which extend the Q-Learning with deep neural networks to
considered essential for this work (Bhatore et al., 2020). The PCA aims handle high-dimensional state spaces (Dixon et al., 2020).
to extract the most crucial information from the dataset by reducing the Ensemble Methods
dimensionality. The algorithm identifies those variables with the most Ensemble methods combine multiple individual models to create
tremendous variance as principal components as a linear combination of more robust and accurate predictions than individual models. Bootstrap
the original data. These principal components can explain most of the aggregation, or bagging, is a method in which n subsets of the training
variance in the original dataset. The following principal components are data are formed, and a base model is applied to each subset. The subsets
calculated with the condition that they are orthogonal to the first ones are formed by bootstrapping. If a training record has already occurred in
and have the largest possible inertia. By projecting the original data onto a subset, it has the same chance to be included in the subsequent subsets.
this coordinate system, unimportant dimensions can be eliminated, and The individual results of the n subsets are aggregated and combined.
the dataset can be simplified (Abdi and Williams, 2010). Bagging achieves good results with low error rates of single classifiers
Reinforcement Learning because these errors can be corrected by the other classifiers (Kubat,
Reinforcement learning is a category of ML in which an agent in­ 2017). Boosting is a set of methods in which the models are trained
teracts with an environment to make a sequence of decisions given the sequentially, with each model attempting to correct the errors of the
environmental states. The critical difference from classical machine previous model. The outcomes of the models are weighted and then
learning methods lies in the type of feedback provided. While in su­ aggregated to make a final prediction. Misclassified instances are
pervised learning, an exact output is given, in an RL model, the agent weighted higher and are therefore rated as more important for the next
receives partial feedback in the form of rewards. These rewards instance (Kubat, 2017).
encourage the desired behavior without explicitly telling the agent what
action to perform. A Markov decision process is commonly used to
2.3. Study relevance
formalize the problem space. The agent makes decisions based on the
current state. These decisions are guided by a policy that updates the
The integration of structured and unstructured data, coupled with
probabilities for choosing each action depending on the current state.
the exponential growth in data volume, has made machine learning an
This can be important in the financial sector, where the market is not
essential component of decision-making processes in the financial
fully observable or is high dimensional (Dixon et al., 2020, Singh et al.,
sector. Traditional methods often cannot handle the variety of data types
2022). RL includes different types of algorithms. Value-based methods
that characterize modern finance (Bhatore et al., 2020). The algorithms
indirectly search for the optimal policy, representing the expected return
described can meet the growing demands by predicting market fluctu­
for certain actions in specific states. This category includes algorithms
ations, making data-driven decisions, or identifying unusual activity.
such as Q-Learning, SARSA, or DEEP Q-Network. Policy-based methods
This is particularly advantageous for risk management in banks, as the

5
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

algorithms may help determine and treat the identified risk types. starts with the banking-related keywords since machine learning algo­
However, machine learning is not universally superior to traditional rithms should be analyzed for the banking risk management sector. This
methods in every aspect of risk management. For example, some algo­ is followed by keywords related to the identified risk types, excluding
rithms’ black-box problems, in other words, the inability to interpret the credit risk. Finally, keywords are added to the search string for machine
decision-making process, can be a significant disadvantage. Especially learning and specific algorithms. A detailed presentation of the searched
for risk management, which is subject to many regulations, this aspect topics with the corresponding keywords and connecting operators is
can become a problem, as decisions must be made comprehensibly shown in Table 3.1. The keywords and the resulting structure of the
(Giudici et al., 2023). Accordingly, an analysis of the usage of ML for the search string were identified via the initial review of the literature. The
different risk types is of significant use for banks to identify potentials, fundamentals chapter and related systematic literature reviews on
problems, and limitations. Comparable analyses have already been similar topics were the basis for the search string (Bhatore et al., 2020,
carried out in academic work. However, these studies do not include the Kitchenham et al., 2009). Some acronyms are omitted from the search
most recent research results and focus mainly on credit risk (Leo et al., string as they were used for other keywords. For example, in this
2019, Shi et al., 2022). The other types of risk are also crucial to banks as context, the acronym RF of Random Forest stands for “Regulatory
they affect the financial stability and the bank’s operating business. Framework” or “Risk-Free Rate,” and the DT of Decision Tree stands for
Understanding and managing these risks provides an approach to ensure “Digital Transformation.” The search is performed on all fields and is not
a bank’s resilience under different circumstances. limited to abstract, title, and keywords, as a comparison showed that
important literature was excluded with a corresponding filter.
3. Methodology
3.2. Eligibility criteria
The application of machine learning in banking risk management is
extensively documented in the literature. Many research papers deal The literature determined by the search string is filtered by criteria
with different algorithms and analyze their suitability for different defined here to be used for the analysis. The literature is picked for
scenarios that can be subordinated to the risk types. Considering the analysis if it is in the subject area of computer science, artificial intel­
large amount of work and various methodological approaches in the ligence, finance, economics, or other business and additional
domain, a systematic literature review is the most appropriate method to mathematical-related fields. This excludes areas such as environmental
summarize the results in terms of the research objectives. A SLR is and engineering sciences or other non-relevant categories. The respec­
particularly useful for capturing the complexity and heterogeneity of tive areas can be automatically filtered on the Web of Science or
financial risk management, identifying research gaps, and providing a manually analyzed by the title, abstract, and keywords on IEEE Xplore.
structured foundation for future research efforts. Likewise, the princi­ A study is included if it discusses a relevant topic regarding this study’s
ples and benefits of SLRs that Kitchenham (2004) points out, such as objective. This involves a manual quality assessment by analyzing the
providing a framework for new research activities and analyzing the title, abstract, and conclusion.
actual impacts of the research, are more valid reasons for choosing this Furthermore, an article or paper is excluded if it is not accessible or
research method (Kitchenham, 2004). A suitable framework for con­ not published in English. The review does not include studies that have
ducting the SLR is provided by the PRISMA 2020 statement, which offers merely referred to machine learning or artificial intelligence in banking
an updated guideline for reporting systematic reviews (Page et al., risk management. The studies must either have tested and validated
2021). Following the proposed guideline, the process is recorded in individual algorithms for corresponding use cases or deal with the exact
detail to ensure a thorough and transparent documentation of the review application possibilities and potentials for the different risk types. Only
process, which is crucial for the reproducibility and understanding of the studies with defined research objectives and a degree of experimental
steps and decisions performed. rigor and thoroughness are included in the review to ensure further
quality. The workflow of selecting studies is displayed in Fig. 3.1. The
3.1. Search process records tracking is illustrated in Fig. 3.2, which shows that many studies
were excluded during the review process. This is initially because the
A search strategy is developed for this systematic literature review to large search string identifies a wide range of literature, which is neces­
identify the relevant literature. This strategy was adapted to two sci­ sary to ensure no information is omitted. As only a few filters, such as the
entific databases: Web of Science and IEEE Xplore. Numerous databases data and subject area filter, are used, many results are included at the
were analyzed for the quality of their entries, and these two databases beginning. A breakdown of the search results using more filters would
provide the best and most differentiated results. They cover the most
important journals and conferences according to which the literature Table 3.1
was filtered. The search process was performed for journal articles, Search string.
conference proceedings, and review articles from 2013 onwards for the Topic Operator Keywords
last ten years. This time limitation was undertaken to analyze only the
Banking AND (“Banking” OR “Bank”)
latest research findings. A first analysis of existing literature showed that
Risk AND (“Risk Management” OR “Market Risk” OR
relevant research in the context of this study increased significantly from Management “Operational Risk” OR “Liquidity Risk” OR “Market
2013 onwards. From this year onwards significant advancements in Volatility” OR “Model Risk” OR “Business Risk” OR
computational power, the expansion of large data sets and the devel­ “Strategic Risk” OR “Reputational Risk” OR
opment of new ML algorithms led to a rapid increase in the adoption of “Security Risk” OR “Cyber Security”)
Credit Risk NOT (“Credit Risk”)
ML techniques in different areas, including risk management in banking. Machine AND (“Machine Learning” OR “ML” OR “Artificial
As a result, numerous studies have been carried out in the last few years. Learning Intelligence” OR “AI” OR “Algorithm” OR “Deep
Accordingly, the latest research findings are analyzed, and the most Learning” OR “DL” OR “Reinforcement Learning”
recent advancements and trends are captured. Scientific work published OR “RL” OR “Supervised” OR “Unsupervised” OR
“Ensemble Learning” OR “Support Vector Machine”
until November 2023 is included in the analysis.
OR “SVM” OR “Decision Tree” OR “Random Forest”
The search starts with identifying relevant keywords and search OR “Logistic Regression” OR “Neural Network” OR
strings aligning with the research objectives. It should be mentioned that “ANN” OR “Naive Bayes” OR “Bayesian Network”
the syntax of the search strings and the input options vary in the sci­ OR “K-Means” OR “Principal Component Analysis”
entific databases. A general search string is given for simplification, OR “PCA” OR “Boosting” OR “Bagging” OR
“Classification” OR “Clustering” OR “Regression”)
which must be adapted for the respective databases. The search string

6
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

ultimately leads to a significant reduction of the records considered in

the review.

4. Results

4.1. Quantitative analysis

Looking at Table 4.1, the focus of the studies is on the analysis of

market and operational risks. Based on the identified study topics that
deal with market risk, a trend can be derived for predicting various
situations, such as market fluctuations or rates. The prevalence of
studies focusing on market risk underlines the importance and
complexity of this aspect of risk banks face. As dealing with market risk
often requires accurate forecasting of different scenarios, the properties
of machine learning algorithms seem to fit well, and therefore, extensive
research is crucial. In addition, many studies deal with operational risk,
which examines weaknesses in internal processes and systems (Aziz and
Dowling, 2019). It can be observed that sometimes, this risk is associated
with regulatory and reputational risk. This connection to other types of
banks comes as they are interconnected. When internal processes like
fraud detection are disrupted or fail, it can damage the reputation or
affect compliance with regulations (Qasaimeh et al., 2022). The topics of
machine learning in operational risk are very diverse, as they bring
value to many areas like cybersecurity, fraud detection, and
customer-related issues. In comparison, liquidity risk, as well as other
types, are analyzed less intensively.
As seen in Fig. 4.1 more studies have been published in recent years.
The figure shows the publication fluctuations, with a notable increase
observed in the last several years from 2019 onwards. The initial small
Fig. 3.1. Workflow of selecting studies (Shi et al., 2022).
number of studies published indicates that the application of machine
learning to banking risk management was not a research focus, and the
have been possible. However, additional filters and further breakdowns
recent increase in publications included in this study reflects the
during test runs revealed that relevant studies were excluded. Hence, a
growing interest in leveraging ML in the banking domain.
large proportion of the results from the first phase are retained. The
Furthermore, Fig. 4.2 shows the number of algorithms used, whereby
following manual assessment of titles, abstracts, and keywords leads to a
it can be seen that methods based on ANN predominate. In addition,
further reduction of records. Numerous papers with topics irrelevant to
SVM and RF are also often used for risk management-related issues.
the SLR’s objective are excluded. The specialization of this topic

Fig. 3.2. Tracking of records based on Prisma Flow diagram (Page et al., 2021).

7
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

Table 4.1 Table 4.1 (continued )

Systematic literature review results. Risk Type Topic Algorithm Reference
Risk Type Topic Algorithm Reference
Reputational
Market Risk Cryptocurrency RNN (Dutta et al., 2020) Risk
Market Risk Portfolio ANN, RL (Gu et al., 2021) Operational Risk Operational Risk ANN (Taweerojkulsri &
Management Measurement Limpiyakorn, 2014)
Market Risk Stock Market ANN (Alamsyah & Zahir, Operational Risk Cyber Security Naïve Bayes, RF (Ajeetha & Madhu,
Prediction 2018) 2019)
Market Risk Stock Market Regression (Sarangi et al., 2023) Operational Risk Cyber Security DT, Boosting, (Sahingoz et al.,
Prediction Algorithms KNN, RF, Naïve 2019)
Market Risk Stock Market ANN, CART, (Hua et al., 2022) Bayes
Prediction Bagging Operational Risk Operational Risk Semi-supervised (Zhou et al., 2021)
ensemble Measurement few-shot
learning learning
Market Risk Portfolio DL (Lin et al., 2021) Operational Risk Operational Risk Bayesian (Sanford & Moosa,
Management Measurement Networks 2015)
Market Risk Portfolio RL, DL (Ngo et al., 2023) Operational Risk Customer Logistic (Shetu et al., 2021)
Management Relationship Regression, RF,
Market Risk Financial Risk ANN, PCA (Liang, 2017) Naïve Bayes,
Prediction SVM, ANN, DT,
Market Risk Market Indicator SVM, ANN, DL (Cheevirot et al., KNN
Prediction 2023) Operational Risk Cybersecurity SVM (Masduki et al.,
Market Risk Stock Market RNN, ANN; (W. Chen et al., 2015)
Prediction XGBoost 2023) Operational Risk, Fraud Detection K-Means, RF (Liu et al., 2019)
Market Risk Market Indicator Genetic (Tan, 2019) Regulatory
Prediction Algorithm, RNN Risk,
Market Risk Cryptocurrency RL (Shahbazi & Byun, Reputational
2022) Risk
Market Risk Market Indicator ANN (Stege et al., 2017) Liquidity Risk Asset-Liability- DL, PCA (Krabichler &
Prediction Management Techmann, 2023)
Market Risk Performance ANN (Balci & Ogul, 2021) Liquidity Risk Liquidity Risk ANN, Bayesian (Tavana et al., 2018)
Prediction Assessment Network
Market Risk Performance ANN (Wanke et al., 2016) Liquidity Risk Cashflow Graph-Based ML, (Kawahara &
Prediction Prediction Gradient Takeuchi, 2021)
Market Risk Financial Risk CNN (Taylor & Keselj, Boosting
Prediction 2021) Liquidity Risk Financial Risk DL (Xu & Yang, 2022)
Market Risk Market Indicator DL (Daniali et al., 2021) Control
Prediction Liquidity Risk Liquidity Risk Logistic (Guerra et al., 2022)
Market Risk Cryptocurrency RNN (Freeda et al., 2021) Assessment Regression,
Market Risk Stock Market RNN, CNN, (Tengxi, 2023) SVM, Naïve
Prediction XGBoost Bayes, RF,
Market Risk Stock Market RNN (Chatterjee et al., XGBoost
Prediction 2022) Business Risk Bankruptcy ANN (Zaychenko &
Market Risk Stock Market ANN (Wang et al., 2019) Forecasting Zgurovsky, 2019)
Prediction Business Risk Financial Crisis ANN (Zhailybayevich &
Market Risk Performance RF, Linear (González-Rossano Prediction Hamada, 2023)
Prediction Regression, DT, et al., 2023)
PCA
4.2. Qualitative analysis
Operational Risk Cybersecurity ANN (Qasaimeh et al.,
2022)
Operational Risk Fraud Detection KNN, SVM RF, (Tadesse, 2022) 4.2.1. Market risk
Stochastic The SLR has identified various application areas and topics for
Gradient Descent market risk. It can be observed that the most significant subject area
Operational Risk Customer Outlier (Ullah et al., 2019)
deals with the prediction of the stock market. In addition, other topics
Relationship Detection, K-
Means deal with market-specific problems, such as portfolio management or
Operational Risk Customer Logistic (Seid & interest rate and performance predictions. Studies dealing with crypto­
Relationship Regression, RF, Woldeyohannis, currencies were also included in the analysis due to their growing
SVM, KNN, DL 2022) importance for the financial market and banking sector in recent years.
Operational Risk Fraud Detection Deep RL (El Bouchti et al.,
2017)
Several studies deal with predicting the stock market (Alamsyah and
Operational Risk, Money ANN (Yu et al., 2022) Zahir, 2018, Sarangi et al., 2023, Hua et al., 2022, Chen et al., 2023).
Regulatory Laundering The researchers pursued different approaches. Alamsyah and Zahir
Risk, Detection (2018) improved the accuracy and the mean squared error for predicting
Reputational
the Indonesian stock exchange composite based on macroeconomic
Risk
Operational Risk Operational Risk Fuzzy CNN (Pena et al., 2021) variables using ANN (Alamsyah and Zahir, 2018). In addition, the
Measurement suitability of various regression algorithms and the application of
Operational Risk, Fraud Detection SVM, K-Means (Viji et al., 2021) Classification and Regression Trees in combination with bagging
Regulatory ensemble learning for short-term stock price prediction were also
Risk,
Reputational
investigated to achieve improved results compared to traditional
Risk methods (Sarangi et al., 2023, Hua et al., 2022). Moreover, different
Operational Risk, Fraud Detection Logistic (Gangula et al., types of neural networks were used to predict internet money funds and
Regulatory Regression, 2023) stock volatility (Tengxi, 2023, Chatterjee et al., 2022). ML algorithms
Risk, SVM, RF,
can predict share and market prices and their corresponding risks. In a
XGBoost, ANN
study on forecasting China’s stock market risk, various risks were first
identified using VaR. The risks and various fundamental, technical, and

8
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

Fig. 4.1. Amount of studies published per year.

macroeconomic indicators were used for prediction. Subsequently, With limited data, these ANNs can effectively regularize past data to
several different types of neural networks and boosting algorithms were learn from non-linear patterns (Dutta et al., 2020). Another study sup­
used for prediction, significantly improving the financial early warning ports these findings by showing that RNNs are more effective for Bitcoin
system through enhanced performance metrics (Chen et al., 2023). Since price predictions than other ML algorithms. RNNs have better accuracy
Value at Risk is one of the most used traditional risk measures, its pre­ than Naïve Bayes, RF, SVM, or K-Nearest-Neighbor (KNN), a simple
diction may be helpful for enhanced risk management and informed classification and regression algorithm (Freeda et al., 2021). Shahbazi
decision-making. Wang et al. (2019) present a novel ANN-based VaR and Byun (2022) applied RL techniques with an asset allocation method
forecasting method by incorporating external market information to to manage risks within the cryptocurrency framework. RL showed high
improve the accuracy of the S&P 500 index. The researchers chose this performance compared to other ML approaches due to its learning-based
approach since traditional methods are usually based on model as­ nature, which increases the accuracy of information delivery (Shahbazi
sumptions. ML models, on the other hand, can better capture non-linear and Byun, 2022).
patterns and complex relationships, which can lead to improved pre­ Machine learning algorithms’ ability to analyze historical data,
dictions as market dynamics are considered and changes can be adapted identify patterns, and optimize strategies is particularly beneficial for
(Wang et al., 2019). portfolio management. In the review, three papers that deal with ma­
Another widely covered area is cryptocurrency. Two studies propose chine learning in portfolio management were identified. In the first
an approach for bitcoin price prediction using Recurrent Neural Net­ study, the researchers analyze the portfolio optimization problem in the
works. Dutta et al. (2020) compared different types of Artificial Neural digital currency market. The advantages of DL for Feature Expression
Networks using the root means squared error. The feature engineering in Learning and Reinforcement Learning in decision-making are combined
this study is conducted by considering Bitcoin as an alternative invest­ (Gu et al., 2021). Second, DL is used to manage portfolio risk and
ment that offers diversification benefits and another investment option maintain investment performance. Lin et al. (2021) propose a DL solu­
than traditional methods. Thus, a holistic approach was considered to tion to facilitate the design of risk factors. By employing a neural
select the predictor variables. Based on this, an analysis has shown that network, the researchers demonstrate superior performance in
Recurrent Neural Networks like LSTM and GRU outperform traditional explaining stock return variance, improving portfolio volatility, and
machine learning models with smaller root mean squared error (RMSE). achieving stability (Lin et al., 2021). Another approach is comparing the

Fig. 4.2. Counts of algorithms used in studies.

9
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

performance of DL with RL and traditional portfolio optimization relationships.

models like PCA or Hierarchical Risk Parity in frontier and developed Operational risk measurement is essential to risk management,
markets. By analyzing sharp ratios and other metrics with different test serving as the cornerstone for comprehensive assessment, quantifica­
settings, the researchers show that RL consistently performs better than tion, and mitigation of potential threats. The first study dealing with this
traditional methods in a typical market environment and an environ­ topic describes the development of a tool that provides predictions of
ment with high market volatility. In contrast, DL performs well in a operational risk events, aggregate operational loss distributions, and
normal condition. However, in a high degree of market fluctuation, operational VaR within a central Australian bank. The authors propose a
traditional and other ML methods perform better due to the sensitivity to Bayesian Network approach to model smaller, more frequent, attritional
data distribution indicated by the market condition (Ngo et al., 2023). operational loss events. As small failures can cause significant opera­
Assessing the resilience to market fluctuations by analyzing banks’ tional failures, tools that address these local-level events may have
performance and profitability is crucial to banking market risk. significant impacts. The researchers developed a large model in
González-Rossano et al. (2023) analyze the profit income drivers in collaboration with professionals from the Australian banking sector. The
Mexican banks. Several prediction models were tested using machine model is partitioned into seven categories: (I) skills, experience, and
learning algorithms like Linear Regression, DT, RF, and PCA. The re­ working environment; (II) transaction characteristics; (III) human er­
searchers found that RF can be considered reliable in profit prediction rors; (IV) error types; (V) payment failure events; (VI) exposure man­
with a lower mean absolute error than the other models agement events; (VII) regulatory/ legal/ tax events. The model aims to
(González-Rossano et al., 2023). Another study focused on the perfor­ generate probabilities of identified operational loss events within
mance prediction of Asian banks by utilizing ANN and contextual vari­ structured financial operations. The Bayesian Network model encapsu­
ables. The variables contain criteria like capital adequacy, asset quality, lates the probability-based relationships between the model’s risk fac­
management quality, earnings, liquidity, and sensitivity to market risk. tors (Sanford and Moosa, 2015). A different method addresses the
Different banks’ profitability was compared based on RMSE (Wanke challenge of operational risk classification by employing a
et al., 2016). Balci & Ogul (2021) also employ an ANN for performance semi-supervised few-shot learning approach called MetaRisk. It aims to
measurement by predicting the return on equity of Turkish state deposit enhance classification accuracy in identifying new risk types in the
money banks (Balci and Ogul, 2021). system. As the model learns from labeled and unlabeled data, it is
Another area in which ML algorithms can be applied is the prediction helpful for scenarios with a small amount of data, improving the ability
of market indicators. Tan describes the prediction of the London Inter to generalize to new tasks. Because of the limited data availability, the
Banking Offered Rate, one of the most crucial interest rate markets researchers used the few-shot approach, enabling the model to learn
worldwide, using a combination of genetic algorithms and RNN. Genetic from a few labeled examples. This hybrid approach is beneficial for risk
algorithms are based on the principles of natural selection and genetics management in banks, where obtaining labeled data for different risk
(Tan, 2019). In another study, the projections of mapping interest rates scenarios is limited. Compared to other methods like Logistic Regres­
using neural networks in combination with cointegration analysis are sion, Deep Learning, Support Vector Machines, and more, the re­
investigated. The findings show that combining traditional time series searcher’s MetaRisk model showed better performance metrics like
analysis with ANN improves the ability to project interest rates, offering different variations of the F1 Score and accuracy (Zhou et al., 2021).
an addition to relevant stress testing in the financial sector (Stege et al., Besides the classification of operational risk, the assessment of losses
2017). also plays a role in risk management. Taweerojkulsri and Limpiyakorn
Additionally, Daniali et al.’s (2021) research on predicting the (2014) researched a learning model for assessing loss severity of oper­
volatility index is based on DL and aims to improve accuracy in socially ational risk using a backpropagation neural network as a predictive
responsible portfolio management, including sustainable investment model. The focus lies on developing an alternative to expert-based risk
decisions. The results highlight the improved accuracy of the model in assessments, especially in larger banks where operational loss data is
estimating the volatility index, underlining its practicality and potential available. The model is trained on expert judgments to understand
for promoting sustainable investment strategies (Daniali et al., 2021). causal chains and effects. The neural network’s input layer comprises
Cheevirot et al. (2023) evaluate traditional financial models and ML four categories: actor, threat type, event type, and resource, with further
approaches in predicting foreign exchange market volatility. SVM, ANN, subcategories. With an accuracy of 94.72 %, the model is a solid method
and DL exhibit superior predictive capabilities than the traditional for predicting the loss severity of operational risk scenarios
models. The ANN and DL models performed well in providing precise (Taweerojkulsri and Limpiyakorn, 2014). A different approach to esti­
volatility prediction due to their ability to capture intricate relationships mating operational risk capital is proposed by Pena et al. (2021). Their
(Cheevirot et al., 2023). study aims to estimate the operational loss component within a bank by
A further area of market risk identified in the SLR is the prediction of using a fuzzy convolutional deep learning model. In this context, fuzzy
financial risks in the market environment. Two studies have dealt with logic deals with reasoning that approximates the degree of truth and is
different aspects of this topic. An ANN-based approach deals with a not absolute in binary values. The model can integrate internal data­
financial risk early warning system for China. It is based on PCA, which bases within an organization (observed loss events) and external data­
was used to determine the weights of indicators to integrate the bases made available by other organizations (available databases of loss
comprehensive financial security index (Liang, 2017). Taylor and Keselj events). This ensures compliance with the guidelines of the Basel
(2021) examine the prediction of financial distress in financial in­ agreements. The model successfully identified credibility features
termediaries by combining traditional financial ratios, sentiment anal­ within the databases to estimate the loss component from multiple
ysis, and additional factors like interest rates in their study. Using a CNN sources of risk scenarios, showing its adaptability, stability, and credi­
on a dataset of 20 intermediaries over four years, the model achieved an bility. The model’s flexibility makes evaluating risk for new financial
accuracy of 88.24 % in classifying distressed and non-distressed in­ products or technological platforms valuable (Pena et al., 2021).
termediaries (Taylor and Keselj, 2021). As the banking industry has been characterized by the technological
advances of recent decades, computer systems are used in all areas and
4.2.2. Operational risk processes of a bank. These information technology systems must be
During the systematic literature review, numerous areas that deal protected against cyber-attacks to ensure continued operation. Machine
with operational risk were identified. Researchers address many learning can also be helpful for cyber security. Qasaimeh et al. (2022)
different subject areas assigned to the operational domain of risk man­ propose advanced security testing using a network-based cyber-attack
agement. These categories are operational risk measurement, cyberse­ forecasting model for financial institutions based on deep neural net­
curity, fraud detection, money laundering detection, and customer works. The model is designed to better protect banks from unknown

10
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

suspicious activities by anticipating the occurrence of cyber-attacks. The The tested ML models are LR, SVM, RF, Extreme Gradient Boosting
model was able to forecast suspicious behavior with 90.36 % accuracy in (XGBoost), and ANN. While all algorithms performed well, the ANN
the validation phase in a real-life banking test environment and thus consistently outperformed the other algorithms with better performance
offers a solid foundation for identifying network-based threats metrics, highlighting its effectiveness in accurately classifying bank­
(Qasaimeh et al., 2022). The following studies were not conducted in a notes (Gangula et al., 2023).
direct banking context, but the benefits for banks are illustrated. Money laundering is a critical threat to banks as it is an unlawful
Therefore, the studies are included in this review. Another possible process that hides the origin of illegal money, exposing financial in­
threat area is distributed denial of service attacks. These attacks involve stitutions to potentially serious regulatory sanctions or reputational
abnormal network traffic to overwhelm a target server or network, damage (Apostolik et al., 2012). Effectively detecting and thus pre­
making it inaccessible to legitimate users. Ajeetha and Madhu (2019) venting money laundering is a crucial process for banks, which can be
used the Naïve Bayes and Random Forest algorithms to detect that type enhanced as well by machine learning. Yu et al. (2022) propose an
of attack. In the test phase of the two algorithms, the researchers found anti-money laundering risk identification model using a graph-based
that Naïve Bayes delivered better results than RF with 90.9 % accuracy Convolutional Neural Network using transaction information to iden­
(Ajeetha and Madhu, 2019). Another intrusion detection system was tify risky customers. Due to the nature of the data being
proposed by Masduki et al. (2015), who implemented an SVM-based graph-structured, the neural network learns from its ability to model and
approach with close to 96 % accuracy (Masduki et al., 2015). Due to learn effectively from the relationships and dependencies of trans­
the rapid growth in Internet use in the banking sector in recent years, an actions. This approach offers compelling predictions with better accu­
increasing number of phishing attacks are occurring. These attacks are racy and recall scores than an RF model (Yu et al., 2022).
characterized by fraudulent attempts to obtain sensitive information by Analyzing customer relationships is essential for the operational
pretending to be trustworthy sources via emails, websites, or other business of a bank as it can be used to identify potential churn or
channels. Sahingoz et al. (2019) propose a real-time anti-phishing sys­ dissatisfaction influencing operational continuity and profitability.
tem using seven classification algorithms and natural language pro­ Thus, it directly impacts the day-to-day running of banking operations.
cessing features. Analyzing a dataset of 73,575 Uniform Resource Transaction volumes or usages of services like customer service will
Locators (URL), the Random Forest algorithm, in contrast to the other change with varying customer numbers. Churn prediction can help
algorithms, performed best in all performance metrics tested (Sahingoz banks to improve internal processes and services or to ensure quality. It
et al., 2019). can also be of a strategic or business nature, but as there is an impact on
Since fraud is an unsolicited part of today’s financial system, several daily operations, it is classified as an operational risk. Ullah et al. (2019)
researchers approach this problem with different concepts. First, used K-means and outlier detection algorithms to predict customer
Tadesse (2022) suggests a fusion of supervised machine learning models churn and highlight this process’s importance for banks (Ullah et al.,
and control rules. The study shows that a targeted application of pre­ 2019). Another study analyzing bank customer churn prediction com­
dictive models combined with control rules aids in detecting actual pares ML algorithms like Logistic Regression, RF, SVM, KNN, and DL.
fraud occurrence. Additionally, the researcher emphasizes the crucial For an Ethiopian commercial bank dataset, the DL approach out­
role of graph analysis in identifying fraud networks. An RF model per­ performed the other methods with adequate performance metrics (Seid
formed best with accuracy, precision, and recall scores above 90 %. and Woldeyohannis, 2022). A bank’s operational business is heavily
False alarms have significantly reduced (Tadesse, 2022). Another dependent on its customers. The extent of customer churn is directly
approach to counter fraud uses a deep RL method. Combining DL linked to customer satisfaction, and thus, it is also characterized as an
techniques with RL principles allows an agent to make sequential de­ operational risk. Shetu et al. (2021) predict the satisfaction of an online
cisions in a complex environment. The author only deals with the theory banking system in Bangladesh using ML. They compare various ML al­
of this approach but presents the benefits of risk management, including gorithms used on survey-based data, and with accuracy scores of 96 %,
identifying patterns in customer data that signal potential fraud and KNN, Logistic Regression, and RF achieved the best scores (Shetu et al.,
enhancing security in general (El Bouchti et al., 2017). The research of 2021).
Liu et al. (2019) on detecting suspicious transactions within bank ac­
counts includes a K-means clustering algorithm combined with RF to 4.2.3. Liquidity risk
identify and analyze suspicious transactions effectively. They focus on The number of studies dealing with liquidity risk is smaller than that
highly imbalanced data where fraudulent transactions comprise only a of market and operational risk. In addition, very different topics are
tiny part of the data set. By combining these two algorithms, the re­ dealt with, which cannot be summarized in a few categories. One article
searchers aim to address the imbalance in data and enhance the accu­ deals with the potential of Deep Learning in asset-liability management.
racy. Combining the algorithms results in improved performance The researchers state that their approach is computationally less inten­
metrics, such as an F1 Score of 92 % (Liu et al., 2019). Viji et al. (2021) sive than traditional methods. Additionally, risk appetite can be
present another study dealing with fraud by introducing an intelligent controlled by choosing appropriate objectives associated with rewards
anomaly detection model for automatic teller machine booth surveil­ in the learning algorithm. Furthermore, regulatory constraints can be
lance using machine learning. With automatic teller machines, banks enforced by choosing penalties adequately. Therefore, machine learning
face a point of attack for criminal activities, which must be considered provides a robust framework that supports balanced risk-taking and risk-
by operational risk management. Accordingly, the researchers present adjusted pricing (Krabichler and Teichmann, 2023).
an intelligent surveillance system that identifies abnormal activities The assessment of liquidity risk is another area considered by the
with the help of SVM. First, video clips from the machines were encoded identified studies. Guerra et al. (2022) investigate if ML algorithms can
and clustered using the K-Means algorithm. The results of the SVM were model liquidity risk and provide insights for stress-testing scenarios.
then compared with the RF classifier and showed better predictive They classify credit institutions from Portugal regarding their liquidity
performance (Viji et al., 2021). Financial losses can arise from inade­ risk using actual data from 2014 until 2021. Then, a comparison of
quate internal processes or systems. various ML algorithms to a traditional statistical model as a benchmark
Banknote authentication is an essential internal process for banks to is conducted. The comparison included Logistic Regression, SVM, Naïve
maintain the integrity of transactions and to prevent counterfeit bank­ Bayes, RF, and XGBoost classifiers. As a result, the other ML techniques
notes leading to financial losses. Gangula et al. (2023) investigate outperform the traditional Logistic Regression approach, with good
various ML models for banknote authentication. They use a dataset precision and F1 Scores. The XGBoost classifier mainly results in good
containing variance, skewness, kurtosis, entropy, and class values to performance metrics, indicating a suitable solution as a decision support
distinguish between genuine and forged banknotes from image data. or early warning system for liquidity risk modeling (Guerra et al., 2022).

11
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

Tavana et al. (2018) analyze the usage of ANN and Bayesian Networks Accordingly, the studies’ findings can help reduce regulatory risk. In
for liquidity risk assessment. As the banking environment is constantly addition, this has further-reaching consequences on reputational risk. If
exposed to changes, a neural network is particularly capable of pre­ banks do not counteract violations of money laundering and fraud reg­
dicting liquidity risk metrics like the LCR because of its learning fea­ ulations, this negatively affects customer trust and the bank’s
tures. The two models were applied to a dataset from a U.S. bank over reputation.
eight years. The ANN was structured with one hidden layer to approx­
imate the liquidity risk function. In the second phase, a Bayesian 5. Discussion
Network identified influential indicators affecting the liquidity risk, like
liquidity and loan/deposit ratios. Then, the fitted Bayesian Network was 5.1. Literature synthesis
compared with the risk function developed by the ANN, showing that
the BN approximates the risk function with high precision. Thus, the As stated in the introduction, this review aims to understand how
Bayesian Network confirmed the ANN trend, demonstrating the appro­ and in which areas of risk management banks can use ML algorithms to
priateness of both models for measuring liquidity risk (Tavana et al., deal with market, operational, liquidity, and other risks. As a general­
2018). ization, it can be said that ML techniques can benefit all areas of risk
Cash flow prediction is a crucial aspect of liquidity risk as it enables management. The advantages of ML have already been extensively
anticipating liquidity shortfalls and managing the working capital demonstrated, especially in the management of credit risks (Bhatore
effectively. Kawahara and Takeuchi (2021) used a graph-based ML et al., 2020, Chen et al., 2017, Shi et al., 2022). However, other risk
approach for this task to improve the predictive capabilities. The pre­ types can benefit from ML as well. The utilization and effects of various
diction model is DT-based combined with a gradient boosting method, algorithms were analyzed in detail for the market risk in the literature.
enabling time series prediction and incorporating non-linear relation­ The predictive capabilities of ML algorithms are used for the prognosis
ships (Kawahara and Takeuchi, 2021). of stock markets, cryptocurrencies, banking performance, market in­
Financial risk control is crucial for liquidity risk management as it dicators, and financial risk, as well as for portfolio management. The
identifies and treats potential threats to safe banks’ liquidity. Xu and recent increase in publications compared to previous years shown in
Yang (2022) focus on assessing banks’ liquidity risks using deep neural Fig. 4.1 is due to technological advances, increasing industry adoption,
networks. The researchers measure and predict liquidity risks from static and changing requirements in banking. These aspects drive increased
and dynamic perspectives and propose suggestions for the financial risk interest in applying ML methods to address various risk-related issues.
control of commercial banks. In this case, static measurements include This shift underlines the increasing recognition of the potential of uti­
assessing liquidity at a specific point in time, whereas dynamic mea­ lizing ML in banking risk management.
surement analyzes over a period, considering changes and trends in a The utilization of ANN and their different variations from DL to RNN
bank’s financial condition. The predictions are performed using an ANN or CNN is particularly significant. In most studies, researchers use ele­
of the analytic hierarchy process, which connects subjective and ments of these models to deal with the many use cases. ANN’s suitability
objective assessment methods. This approach combines the strengths of for treating market risks can be attributed to several causes. ANN can
a decision-making framework that helps to prioritize and analyze mul­ recognize complex data patterns and relationships. This is particularly
tiple criteria or goals with the predictive capabilities of an ANN. The useful for the market risk domain, as the markets of numerous industries
authors compared it with other ML models to verify the effectiveness of and sectors are linked to the banking sector. Therefore, market risk in­
their model, resulting in better performance measures (Xu and Yang, volves many variables like interest rate, currency, and price fluctuations,
2022). which must be considered to generate reliable models. ANNs enable
diverse data inputs by incorporating different network configurations by
4.2.4. Other risk types parameter tuning to adapt to different market environments. Since
The SLR identified numerous studies that deal with market, opera­ ANNs can handle non-linear high-dimensional data, they are suitable for
tional, and liquidity risks. The analyzed studies also deal with other risks mapping the complex interrelationships of the financial markets (Dixon
but to a smaller extent. Two studies have dealt with topics that can be et al., 2020). As a result, ANN-based models can provide better perfor­
subordinated to business risk (Zaychenko and Zgurovsky, 2019, Zhai­ mance metrics than traditional approaches to address market-specific
lybayevich and Hamada, 2023). Zaychenko and Zgurovsky (2019) problems (Freeda et al., 2021, Cheevirot et al., 2023). RL is another
propose a method for Ukrainian banks’ bankruptcy forecasting. It ad­ algorithm that is often used for market risk. It can be used well for this
dresses the importance of predicting bankruptcy risk, which is related to risk type as it can overcome market-related issues by continuously
the instability or even potential failure of banks, characterizing business adapting strategies to the changing market environment. As RL can
risk. The researchers tested different fuzzy neural networks and interact with an environment, it can learn from market feedback and
compared them to traditional financial risk assessment methods. The iteratively tune decision-making processes to handle dynamic market
study highlights the superior predictive accuracy of fuzzy neural net­ scenarios and manage volatility (Dixon et al., 2020). This adaptability
works over traditional methods (Zaychenko and Zgurovsky, 2019). A makes RL valuable for optimizing risk strategies customized to the dy­
further investigation into the use of ANN in this domain describes namic character of market risks.
developing a predictive model for financial crises in banks. The re­ Further algorithms, such as ensemble methods or regression algo­
searchers emphasize the importance of tailored predictive models for rithms, are widely used, highlighting their usefulness in dealing with
banks depending on the circumstances and suggest using various pre­ specific market risk challenges. Depending on the use case, corre­
diction algorithms to enhance early warning systems (Zhailybayevich sponding ML algorithms can make predictions, optimize portfolios, or
and Hamada, 2023). Both studies aim to predict bankruptcy in banks. handle diverse risks in the market landscape. These strengths are re­
Thus, they are considered a business risk due to their direct impact on flected in the number of studies identified to support risk management in
financial stability, investors’ trust, and other effects on the business an evolving market environment.
environment. ML is also proving to be a promising approach for banks in opera­
The studies do not directly cover other types of risk. Studies only tional risk. The SLR shows the various applications of ML in areas
cover regulatory risks indirectly that deal with money laundering and ranging from operational risk measurement to cybersecurity, fraud
some aspects of fraud detection (Liu et al., 2019, Viji et al., 2021, detection, money laundering prevention, and customer relationship
Gangula et al., 2023, Yu et al., 2022). In the realm of banks, adapting to analysis. The effectiveness of various algorithms is demonstrated by
and complying with changes in the regulatory environment is crucial to robust performance metrics, underlining the value of operational risk
mitigate possible negative consequences (Kelliher et al., 2013). management. These models show potential for improving risk

12
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

identification, classification, and mitigation strategies. The broad range operational risk, ML algorithms are crucial for fraud detection and
of algorithms shows different performances in overcoming specific cybersecurity, showcasing their versatility and robustness in identifying
problems. In contrast to market risk, no algorithm can be identified as and mitigating diverse threats
universally applicable. This is due to the complex nature of the opera­
tional risk domain, which addresses a range of tasks in internal processes 5.2. Practical implications
and systems, fraud detection, or the analysis of customer relationships.
Each area is diverse and requires specific methods, so a single algorithm This research contributes to improving banks’ risk management by
cannot universally solve all operational risk problems. However, RF, exploring and analyzing the application of ML techniques across various
SVM, and ANN were analyzed quite extensively. The properties of RF risk types. By performing the SLR, the existing studies that leverage ML
allow to deal with outliers and noise in the data and show good accuracy addressing different aspects of risk within banks have been synthesized.
(Breiman, 2001). RFs help deal with different data types, feature sets, This synthetization of insights provides an overview of current practices,
and high-dimensional data (Liu et al., 2019, Shetu et al., 2021). strengths, and limitations in employing ML algorithms in risk manage­
Additionally, RFs perform well in settings with imbalanced labels ment. The SLR contributes to practitioners’ understanding of the current
and provide insights into the importance of features, thus being valuable maturity of utilizing ML. It is shown that incorporating ML techniques
for fraud detection or cyber-attack prediction (Breiman, 2001). As SVM into risk management strategies allows banks to analyze large datasets
can handle linear and non-linear relationships, they can be used for with varying data types. This leads to enhanced risk assessment, iden­
various scenarios, especially showing good performance in classification tification of threats, and decision-making.
tasks in cybersecurity, fraud detection, or customer churn prediction Furthermore, ML enables improved market trends, operational de­
(Masduki et al., 2015, Viji et al., 2021, Seid and Woldeyohannis, 2022). ficiencies, and liquidity shortage prediction. This helps to analyze risk
As ANNs are suitable for pattern recognition tasks, they can detect po­ proactively and develop mitigation strategies. Moreover, the SLR high­
tential threats like irregularities in data and continuously learn and lights the utility of different algorithms in dealing with specific chal­
adapt, enhancing the ability to detect anomalies and irregularities. In lenges. This is useful for selecting and implementing the most
operational risk management, these features are perfect for dealing with appropriate algorithms for particular needs. The algorithms can be
unusual behavior in network traffic, fraud, money laundering, or adapted to the risk types by adjusting various parameters to fine-tune
assessing loss severity (Qasaimeh et al., 2022, Taweerojkulsri and risk treatment.
Limpiyakorn, 2014, Gangula et al., 2023, Yu et al., 2022). Despite their Additionally, resources can be better allocated as more data is
predictive solid performance, some algorithms like ANN or RF have a collected in banks because of technological change. Due to the large
black-box nature, resulting in a lack of interpretability, which is a amounts of data, numerous areas can be analyzed, and thus, banks can
problem for the highly regulated banking industry (Giudici et al., 2023). focus more quickly on areas in need. Strategical decision support also
Compared to the other risk types, research on liquidity risk in the benefits from ML. The derived insights help to make strategic decisions
context of ML has been less extensive. The main reason for this is the lack by better understanding risk patterns, market behavior, and customer
of qualitative and available data. Banks rarely provide internal liquidity relationships. This is additionally beneficial to develop new dynamic
data, which leads to less research on this risk type. Nevertheless, strategies to adapt to evolving risks. Collected customer data can be used
essential areas of liquidity risk are highlighted. For example, the po­ to build predictive customer analyses that offer insights to build
tential of DL for asset liability management and the advantages of risk- customer-centric products and services. Fundamentally, integrating ML
adjusted pricing are emphasized (Krabichler and Teichmann, 2023). algorithms into banking risk management enhances the robustness and
Another focus is comparing the liquidity risk for stress test analyses of the ability to adjust strategic positioning to different risk types to
traditional methods and boosting algorithms or Bayesian methods. The respond to threats in a dynamic environment.
ML methods are more appropriate, indicating their suitability as a Indeed, the advancements in ML contribute to the overall stability
support system (Guerra et al., 2022, Tavana et al., 2018). ML is also used and reliability of the banking sector. By enabling more accurate risk
to forecast cash flow and control financial risks, making it an effective assessments and proactive threat mitigation, ML helps banks maintain
instrument for managing liquidity risks. Despite the limited number of operational continuity and safeguard against financial losses. These
published studies, the SLR demonstrates the benefits of ML for liquidity improvements not only enhance the financial health of individual in­
risk. The ability to make predictions and risk assessments is crucial for stitutions but also contribute to the stability of the broader financial
liquidity risk. No algorithm emerges as dominant for individual tasks. system.
The algorithms stand out in different areas due to their inherent Also, some practical implications are in order. On the one hand,
strengths. Banks can strategically integrate ML into their risk management
The other less significant risk types are even less studied. Some frameworks to enhance their predictive capabilities and operational
studies indirectly cover only business, regulatory, and reputational risks. efficiency. This integration requires addressing challenges such as data
ANNs are used for the business risk domain to forecast bankruptcy and quality, model interpretability, and regulatory compliance. By adopting
financial crises (Zaychenko and Zgurovsky, 2019, Zhailybayevich and a strategic approach, banks can leverage ML to gain a competitive edge
Hamada, 2023). As business risk focuses on risks threatening the in risk management. On the other hand, ML’s ability to provide more
competitive position and operations, these studies can be subordinated transparent and accurate risk assessments can help banks meet stringent
to business and not to market risk (Apostolik et al., 2012). Both studies regulatory requirements. Regulators can also benefit from ML’s insights
utilize ANNs due to their predictive and learning abilities. Some studies to better understand the risk landscape and develop more effective
that deal with operational risk also focus on regulatory and reputational oversight mechanisms.
risks. This is because a failure in fraud detection and money laundering
might have significant consequences, as regulations stipulate. Breaking 5.3. Limitations
these regulations can seriously affect a bank’s reputation (Aziz and
Dowling, 2019). Different algorithms are selected in the studies One of the main limitations of this review is the focus on the Web of
depending on the use case. Science and IEEE Xplore databases. This focus could have excluded
In a nutshell, the systematic literature review reveals significant relevant studies from journals or conferences published on other data­
trends in the application of ML across different risk types. For instance, bases. A manual comparison of various databases before the SLR showed
artificial neural networks (ANNs) and support vector machines (SVMs) that the two databases publish the best-fitting studies and include the
are prominently used for market risk prediction, demonstrating their most important journals and conferences regarding the nature of the
ability to capture complex, non-linear relationships in financial data. In research objectives. Thus, further studies may be relevant, but no

13
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

groundbreaking new findings are expected to be found that have not also with interdisciplinary cooperation between scientists and industry
been identified by the analyzed databases. The included studies show a stakeholders. These efforts can foster innovative approaches by sharing
mainly positive outcome of employing ML in banking risk management data concepts and findings to integrate the benefits of ML into risk
domains. This might indicate a publication bias, skewing the findings management in a sustainable fashion.
towards favorable ML outcomes. Some crucial negative aspects, like the In the same breath, it is important to mention four key topics:
lack of interpretability, are sometimes mentioned but not further emerging risk areas; ethical and responsible AI; scalability and real-time
investigated (Krabichler and Teichmann, 2023, Jensen and Iosifidis, processing; and cross-disciplinary approaches. Significant progress has
2023). Another significant limitation is the sole focus on scientific been made in applying ML to traditional risk areas like credit and market
research. Since this review only analyzed scientific papers from journals risk. However, emerging risks such as cyber risk, climate risk, and
and conferences, grey literature is left out. Potentially valuable findings operational resilience require further exploration. Research should
from the banking sector, such as reports or industry papers, have not investigate how ML can predict and mitigate these emerging risks,
been considered. Furthermore, this study focuses exclusively on the considering their unique characteristics and the dynamic nature of the
potential of ML, whereby other technological methods and approaches risk landscape.
in risk management were not examined. In addition, the possibility of From another angle, the ethical implications of using ML in banking
integrating ML based on readiness and maturity in the banking sector must be considered. Future research should address issues related to bias
has not been analyzed. in ML models, ensuring these models do not perpetuate existing in­
equalities or unfairly disadvantage certain groups of customers. Devel­
5.4. Further research oping frameworks for ethical AI in banking is essential for maintaining
public trust and regulatory compliance.
Further research directions can be derived based on the findings of As the volume of financial data continues to grow, scalable ML so­
this review. Initially, it is essential to mention that market risk has been lutions capable of real-time processing will become increasingly
dealt with extensively. However, no studies have been conducted that important. Research should focus on developing ML models and archi­
specifically analyze market risk in the risk management context of tectures that can efficiently handle large-scale data and provide real-
banks. The studies deal with topics relevant to market risk and can be time insights. This capability is crucial for applications like fraud
transferred but are not in an explicit risk management context. Further detection, where timely intervention is necessary.
research efforts in this area would be advisable. An actual analysis of a Lastly, the complexity of banking risk management requires a cross-
bank’s risk management procedures and processes with subsequent disciplinary approach that combines insights from finance, computer
identification of the application potential through case studies and stress science, statistics, and regulatory studies. Future research should pro­
testing could offer a possible solution. This proposal is also supported by mote collaboration across these disciplines to develop more compre­
Leo et al., who call for more studies directly related to risk management hensive and effective ML applications. Such collaboration can lead to
(Leo et al., 2019). Since then, research efforts specifically related to innovative solutions that address the multidimensional challenges of
operational and liquidity risk in the context of ML have been conducted modern banking risk management.
(Guerra et al., 2022, Sanford and Moosa, 2015, Zhou et al., 2021,
Taweerojkulsri and Limpiyakorn, 2014). Additionally, the integration of 6. Conclusion
nowcasting techniques could provide valuable real-time insights into
risk management, allowing for more responsive and dynamic The constantly increasing amount of data accumulated in the
decision-making processes (Sadok et al., 2023). financial sector requires continuous analysis to leverage it effectively.
Nevertheless, the scope of the studies analyzing liquidity risk can Banks’ risk management requires the integration of advanced analyses
expand to gain more detailed insights. An analysis of the relationships of various data. ML is becoming increasingly common as a suitable op­
between individual risk types could be helpful to understand how tion for overcoming the changes and challenges in risk management.
different categories influence each other. This would allow more in- This SLR provides an overview of ML learning algorithms used in banks’
depth risk management strategies to be drawn up. For instance, inte­ risk management about market, operational, liquidity, and other risk
grating sentiment analysis into risk management by analyzing social types. Research into the application of ML has revealed innovative
media data could be beneficial for a comprehensive understanding of methods and their benefits in the banking sector. The review results
market movements and customer-related issues, reputation, or credit highlight the central role that ML algorithms play in the assessment,
risk. An initial analysis of the feasibility of this would provide further prediction, and mitigation of the different risk types. The analysis
insights. Another area that can be further explored is the integration of revealed that scientific research focuses on market and operational risks.
existing risk management frameworks with machine learning algo­ In particular, the potential of Artificial Neural Networks and other al­
rithms. Traditional methods have often been compared with ML gorithms, such as Support Vector Machines or Random Forests, was
(Sarangi et al., 2023, Hua et al., 2022, Zaychenko and Zgurovsky, 2019). researched. The studies have successfully predicted market dynamics,
Integration into existing frameworks would involve examining the portfolio management, fraud detection, and cyber security. Research in
challenges and potentials, focusing on adoption challenges, imple­ the direct context of a bank’s risk management is desirable regarding
mentation scalability, and institutional barriers. For this purpose, the market risk. In addition, further analysis of ML for liquidity risk is
actual processes and maturity levels must be examined to generate needed to gain more precise insights. The results underline the oppor­
added value with ML. tunities for using ML in risk management. It is highlighted that further
Additionally, qualitative studies like expert interviews, case studies, research is required to unlock the full potential of ML and enable banks
or surveys can further analyze institutional barriers like governance or to develop their risk mitigation strategies further. Utilizing the potential
stakeholder engagement. This is useful for measuring the acceptance, requires further research so that ML can be integrated into existing
maturity level, and deployment potential of ML in a bank’s risk man­ frameworks by overcoming institutional barriers. This SLR underlines
agement. Explaining decision-making processes or portfolio allocation the central role of ML in redesigning risk management practices. The
with ML to different stakeholders is very important (Fritz-Morgenthal results provide insights into current practices and highlight areas for
et al., 2022). Researching the black-box problem is an elementary further research to exploit the full potential of ML. Leveraging the
component of further research efforts. The integration of ML methods inherent capabilities of ML for risk management in the banking sector
with regulatory requirements and the effect on ethical aspects should be not only meets current challenges but also leads institutions into a data-
examined. Scientific research has made progress in employing ML in driven future of resilience and strategic decision-making.
banking risk management. The findings should be linked to the practice,

14
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

CRediT authorship contribution statement Daniali, S. M., et al. (2021). Predicting volatility index according to technical index and
economic indicators on the basis of deep learning algorithm. Sustainability, 13(24).
[Link] Dec.
Valentin Lennart Heß: Writing – review & editing, Writing – orig­ DEPC. (2023b). Capital Requirements Directive IV. Official Journal of the European Union
inal draft, Software, Methodology, Conceptualization. Bruno Damásio: Accessed: Oct. 16[Online]. Available Https://[Link]/legal-conten
Writing – review & editing, Supervision, Methodology, Investigation, t/EN/TXT/?uri=CELEX%3A32013L0036.
DEPC. (2023a). Capital Recuirements Regulations. Official Journal of the European Union
Conceptualization. Accessed: Oct. 16[Online]. Available Https://[Link]/legal-content/en
/TXT/?uri=celex%3A32013R0575.
Declaration of competing interest Dixon, M. F., Halperin, I., & Bilokon, P. (2020). Machine learning in finance. Cham:
Springer International Publishing. [Link]
Dutta, A., Kumar, S., & Basu, M. (2020). A gated recurrent unit approach to bitcoin price
The authors declare that they have no known competing financial prediction. Journal of Risk and Financial Management, 13(2). [Link]
interests or personal relationships that could have appeared to influence 10.3390/jrfm13020023. Feb.
El Bouchti, A., Chakroun, A., Abbar, H., & Okar, C. (2017). Fraud detection in banking
the work reported in this paper. using deep reinforcement learning. In The Seventh International Conference on
Innovative Computing Technology (pp. 58–63).
Acknowledgements Freeda, S. E., Selvan, T. C. E., & Hemanandhini, I. G. (2021). Prediction of bitcoin price
using deep learning model. In Proceedings of the 5th International Conference on
Electronics, Communication and Aerospace Technology, ICECA 2021 (pp. 1702–1706).
This work was supported by national funds through FCT (Fundação Institute of Electrical and Electronics Engineers Inc.. [Link]
para a Ciência e a Tecnologia), under the project - UIDB/04152/2020 - ICECA52323.2021.9676048
Fritz-Morgenthal, S., Hein, B., & Papenbrock, J. (2022). Financial risk management and
Centro de Investigação em Gestão de Informação (MagIC)/NOVA IMS)
explainable, trustworthy, responsible AI. Frontiers in Artificial Intelligence, 5. https://
([Link] [Link]/10.3389/frai.2022.779799. Feb.
Gangula, Y., Srivastava, S., Mohanty, A., & Prakasam, P. (2023). Investigation of various
References machine learning techniques for banknote authentication. In 2023 3rd Asian
Conference on Innovation in Technology. Institute of Electrical and Electronics
Engineers Inc. [Link]
Abdi, H., & Williams, L. J. (2010). Principal component analysis. WIREs Computational García, S., Ramírez-Gallego, S., Luengo, J., Benítez, J. M., & Herrera, F. (2016). Big data
Statistics, 2(4), 433–459. [Link] Jul. preprocessing: Methods and prospects. Big Data Analysis, 1(1), 9. [Link]
Ajeetha, G., & Madhu, P. G. (2019). Machine learning based DDos attack detection. 10.1186/s41044-016-0014-0. Dec.
Innovations in power and advanced computing technology (i-PACT). Gašperov, B., Begušić, S., Posedel Šimović, P., & Kostanjčar, Z. (2021). Reinforcement
Alamsyah, A., & Zahir, A. N. (2018). Artificial neural network for predicting Indonesia learning approaches to optimal market making. Mathematics, 9(21), 2689. https://
Stock Exchange composite using macroeconomic variables. In 2018 6th International [Link]/10.3390/math9212689. Oct.
Conference on Information and Communication Technology (ICoICT) (pp. 44–48). IEEE. Giudici, P., Centurelli, M., & Turchetta, S. (2023). Artificial intelligence risk
[Link] May. measurement. Expert systems with Applications, 235, Article 121220. [Link]
Alazzabi, W. Y. E., Mustafa, H., & Karage, A. I. (2023). Risk management, top 10.1016/[Link].2023.121220. Jan.
management support, internal audit activities and fraud mitigation. Journal of González-Carrasco, I., Jiménez-Márquez, J. L., López-Cuadrado, J. L., & Ruiz-Mezcua, B.
Financial Crime, 30(2), 569–582. [Link] Feb. (2019). Automatic detection of relationships between banking operations using
Apostolik, R., Donohue, C., & Went, P. (2012). Foundations of banking risk. Wiley. https:// machine learning. Information Science (N Y), 485, 319–346. [Link]
[Link]/10.1002/9780470555996 10.1016/[Link].2019.02.030. Jun.
S. Aziz and M. Dowling, “Machine learning and AI for risk management,” 2019, pp. González-Rossano, C., Terán-Bustamante, A., Velázquez-Salazar, M., & Martínez-
33–50. doi: 10.1007/978-3-030-02330-0_3. Velasco, A. (2023). What drives profit income in Mexico’s main banks? Evidence
Balci, T., & Ogul, H. (2021). Predicting bank return on equity (ROE) using neural using machine learning. Sustainability, 15(7). [Link]
networks. In SAMI 2021 - IEEE 19th World Symposium on Applied Machine Intelligence Apr.
and Informatics, Proceedings (pp. 279–285). Institute of Electrical and Electronics Gorton, G., & Metrick, A. (2012). Getting up to speed on the financial crisis: A one-
Engineers Inc.. [Link] Jan. weekend-reader’s guide. Journal of Economic Literature, 50(1), 128–150. [Link]
BCBS, “Basel Committee on Banking Supervision the Basel Framework,” 2023. Accessed: org/10.1257/jel.50.1.128. Mar.
Sep. 29, 2023. [Online]. Available: Https://[Link]/basel_framework/. Gu, F., Jiang, Z., & Su, J. (2021). Application of features and neural network to enhance
Berrar, D. (2018). Bayes’ Theorem and Naive Bayes classifier. Encyclopedia of the performance of deep reinforcement learning in portfolio management. In 2021
bioinformatics and computational biology (pp. 403–412). Elsevier. [Link] IEEE 6th International Conference on Big Data Analytics, ICBDA 2021 (pp. 92–97).
10.1016/B978-0-12-809633-8.20473-1 Institute of Electrical and Electronics Engineers Inc.. [Link]
Bhatore, S., Mohan, L., & Reddy, Y. R. (2020). Machine learning techniques for credit risk ICBDA51983.2021.9403044. Mar.
evaluation: A systematic literature review. Journal of Banking and Financial Guerra, P., Castelli, M., & Côrte-Real, N. (2022). Machine learning for liquidity risk
Technology, 4(1), 111–138. [Link] Apr. modelling: A supervisory perspective. Economic Analysis and Policy, 74, 175–187.
Breiman, L. (2001). Random forests. Machine learning, 45(1), 5–32. [Link] [Link] Jun.
10.1023/A:1010933404324 Hartigan, J. A., & Wong, M. A. (1979). Algorithm AS 136: A K-means clustering
Chatterjee, A., Bhowmick, H., & Sen, J. (2022). Stock volatility prediction using time Algorithm. Applied Statistics, 28(1), 100. [Link]
series and deep learning approach. In 2022 IEEE 2nd Mysore Sub Section International Horcher, K. A. (2005). Essentials of financial risk management. Wiley. [Link]
Conference. Institute of Electrical and Electronics Engineers Inc.. [Link] 10.1002/9781118386392
10.1109/MysuruCon55714.2022.9972559 Hosmer, D. W., Lemeshow, S., & Sturdivant, R. X. (2013). Applied logistic regression.
Cheevirot, S., Smanchat, S., & Nuchitprasitchai, S. (2023). Evaluation of traditional Wiley. [Link]
finance models and machine learning approaches in FOREX volatility prediction. In Hua, Y., Zhu, R., & Duan, Y. (2022). Construction of short-term stock price prediction
2023 27th International Computer Science and Engineering Conference (ICSEC) (pp. algorithm based on MLP and CART Bagging ensemble learning. In 2022 IEEE
315–318). IEEE. [Link] Sep. Conference on Telecommunications, Optics and Computer Science (TOCS) (pp.
Chen, M., Dautais, Y., Huang, L. G., & Ge, J. (2017). Data driven credit risk management 371–376). Institute of Electrical and Electronics Engineers (IEEE). [Link]
process: A machine learning approach. In ACM International Conference Proceeding 10.1109/tocs56154.2022.10015919. Jan.
Series (pp. 109–113). Association for Computing Machinery. [Link] Jensen, R. I. T., & Iosifidis, A. (2023). Fighting money laundering with statistics and
10.1145/3084100.3084113. Jul. machine learning. IEEE Access : Practical Innovations, Open Solutions, 11, 8889–8903.
Chen, S. (2023). Research on risk management in banking system. Highlights in Business, [Link]
Economics and Management, 3, 267–275. [Link] Johri, S., Qureshi, N. I., Mehta, K., Othman, B., Waghmare, S., & Pant, B. (2022).
Jan. A critical significance of using machine learning in strengthening financial risk
Chen, W., Chen, B., & Cai, X. (2023). Forecasting China’s stock market risk under the management in banking firms. In 2022 2nd International Conference on Advance
background of the Stock Connect programs. Application of Soft Computing. https:// Computing and Innovative Technologies in Engineering (ICACITE) (pp. 1933–1937).
[Link]/10.1007/s00500-023-08496-z IEEE. [Link] Apr.
Chockalingam, A., Dabadghao, S., & Soetekouw, R. (2018). Strategic risk, banks, and Kawahara, R., & Takeuchi, M. (2021). Cash flow prediction of a bank deposit using
Basel III: Estimating economic capital requirements. The Journal of Risk Finance, 19 scalable graph analysis and machine learning. In 2021 IEEE International Conference
(3), 225–246. [Link] May. on Big Data (pp. 1647–1656). Institute of Electrical and Electronics Engineers Inc..
Cortes, C., & Vapnik, V. (1995). Support-Vector Networks. Machine Learning, 20(3), [Link]
273–297. [Link] Kelliher, P. O. J., Wilmot, D., Vij, J., & Klumpes, P. J. M. (2013). A common risk
Dang, T. L., & Nguyen, T. M. H. (2020). Liquidity risk and stock performance during the classification system for the actuarial profession. British Actuarial Journal, 18(1),
financial crisis. Research of International Business and Finance, 52, Article 101165. 91–121. [Link] Mar.
[Link] Apr. B. Kitchenham, “Procedures for performing systematic reviews,” 2004.

15
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

B. Kitchenham, O. Pearl Brereton, D. Budgen, M. Turner, J. Bailey, and S. Linkman, Information Systems and Computer Networks. Institute of Electrical and Electronics
“Systematic literature reviews in software engineering - A systematic literature Engineers Inc. [Link]
review,” Jan. 2009. doi: 10.1016/[Link].2008.09.009. Seid, M. H., & Woldeyohannis, M. M. (2022). Customer churn prediction using machine
Kotsiantis, S., Kanellopoulos, D., & Pintelas, P. E. (2006). Data preprocessing for learning: Commercial Bank of Ethiopia. In 2022 International Conference on
Supervised learning. International Journal of Computer Science, 1(1), 111–117 Information and Communication Technology for Development for Africa (ICT4DA) (pp.
[Online]Available Https://[Link]/publication/228084519. 1–6). IEEE. [Link] Nov.
Krabichler, T., & Teichmann, J. (2023). A case study for unlocking the potential of deep Shahbazi, Z., & Byun, Y. C. (2022). Machine learning-based analysis of cryptocurrency
learning in asset-liability-management. Frontiers in Artificial Intelligence, 6. https:// market financial risk management. IEEE Access : Practical Innovations, Open Solutions,
[Link]/10.3389/frai.2023.1177702 10, 37848–37856. [Link]
Kubat, M. (2017). An introduction to machine learning. Cham: Springer International Shetu, S. F., Jahan, I., Islam, M. M., Ara Hossain, R., Moon, N. N., & Nur, F. Narin (2021).
Publishing. [Link] Predicting satisfaction of online banking system in Bangladesh by machine learning.
Leo, M., Sharma, S., & Maddulety, K. (2019). Machine learning in banking risk In 2021 International Conference on Artificial Intelligence and Computer Science
management: A literature review. Risks, 7(1). [Link] Technology (ICAICST) (pp. 223–228). IEEE. [Link]
risks7010029. Mar. ICAICST53116.2021.9497796. Jun.
Liang, Y. (2017). Empirical study on China’s financial risk early warning under the new Shi, S., Tse, R., Luo, W., D’Addona, S., & Pau, G. (2022). Machine learning-driven credit
normal. In 4th International Conference on Industrial Economics System and Industrial risk: A systemic review. Springer Science and Business Media Deutschland GmbH.
Security Engineering (IEIS) [Link] [Link] Sep. 01.
Lin, H., Zhou, D., Liu, W., & Bian, J. (2021). Deep risk model: A Deep learning solution Singh, V., Chen, S.-S., Singhania, M., Nanavati, B., kumar kar, A., & Gupta, A. (2022).
for mining latent risk factors to improve covariance matrix estimation. In ICAIF 2021 How are reinforcement learning and deep learning algorithms used for big data
- 2nd ACM International Conference on AI in Finance. Association for Computing based decision making in financial industries–A review and research agenda.
Machinery, Inc [Link] Nov. International Journal of Information Management Data Insights, 2(2). [Link]
Liu, Y., Tang, Z., & Zheng, W. (2019). Suspicious bank card transaction recognition based 10.1016/[Link].2022.100094. Nov.
on K-means clustering and random forest algorithm. In 2019 International Conference Stege, N., Basse, T., Wegener, C., & Kunze, F. (2017). Mapping interest rate projections
on Intelligent Informatics and Biomedical Sciences (ICIIBMS). IEEE. using neural networks under cointegration. In ACM International Conference
Lowd, D., & Domingos, P. (2005). Naive Bayes models for probability estimation. In Proceeding Series. Association for Computing Machinery [Link]
Proceedings of the 22nd international conference on Machine learning - ICML ’05 (pp. 3109761.3109774. Oct.
529–536). New York, New York, USA: ACM Press. [Link] G. Swankie and D. Broby, “Examining the impact of artificial intelligence on the
1102351.1102418. evaluation of banking risk,” 2019. [Online]. Available: [Link]
Masduki, B. W., Ramli, K., Saputra, F. Astika, & Sugiarto, D. (2015). Study on net/publication/337908452.
implementation of machine learning methods combination for improving attacks Tadesse, T. (2022). Combining control rules, machine learning models, and community
detection accuracy on intrusion detection system (IDS). In International Conference on detection algorithms for effective fraud detection. In 2022 International Conference on
Quality in Research (pp. 56–64). Information and Communication Technology for Development for Africa, ICT4DA 2022
A. Mashrur, W. Luo, N.A. Zaidi, and A. Robles-Kelly, “Machine learning for financial risk (pp. 42–46). Institute of Electrical and Electronics Engineers Inc.. [Link]
management: A survey,” 2020, Institute of Electrical and Electronics Engineers Inc. 10.1109/ICT4DA56482.2022.9971235
doi: 10.1109/ACCESS.2020.3036322. Tan, X. (2019). LIBOR prediction using genetic algorithm and genetic algorithm
Milana, C., & Ashta, A. (2021). Artificial intelligence techniques in finance and financial integrated with recurrent neural network. In 2019 Global Conference for Advancement
markets: A survey of the literature. Strategic Change, 30(3), 189–209. [Link] in Technology (GCAT) (pp. 1–8). IEEE. [Link]
org/10.1002/jsc.2403. May. GCAT47503.2019.8978299. Oct.
Milojević, N., & Redzepagic, S. (2021). Prospects of artificial intelligence and machine Tavana, M., Abtahi, A. R., Di Caprio, D., & Poortarigh, M. (2018). An Artificial Neural
learning application in banking risk management. Journal of Central Banking Theory Network and Bayesian Network model for liquidity risk assessment in banking.
and Practice, 10(3), 41–57. [Link] Sep. Neurocomputing, 275, 2525–2554. [Link]
Ngo, V. M., Nguyen, H. H., & Van Nguyen, P. (2023). Does reinforcement learning Jan.
outperform deep learning and traditional portfolio optimization models in frontier Taweerojkulsri, C., & Limpiyakorn, Y. (2014). Learning model for assessing loss of
and developed financial markets? Research in International Business and Finance, 65, severity of operational risk. In International Conference on Information Science &
Article 101936. [Link] Apr. Applications (ICISA) [Link]
Page, M. J., et al. (2021). The PRISMA 2020 statement: An updated guideline for Taylor, S., & Keselj, V. (2021). Predicting the distress of financial intermediaries using
reporting systematic reviews. BMJ (Clinical Research Ed.), n71. [Link] convolutional neural networks. In 2021 IEEE 23rd Conference on Business Informatics
10.1136/bmj.n71. Mar. (pp. 71–77). Institute of Electrical and Electronics Engineers Inc.. [Link]
Pallathadka, H., Ramirez-Asis, E. H., Loli-Poma, T. P., Kaliyaperumal, K., 10.1109/CBI52690.2021.10057
Ventayen, R. J. M., & Naved, M. (2023). Applications of artificial intelligence in Tengxi, W. (2023). Research on forecasting and risk measurement of internet money
business management, e-commerce and finance. Material Today Proceedings, 80, fund returns based on error-corrected 1DCNN-LSTM-SAM and VaR: Evidence from
2610–2613. [Link] China. IEEE Access : Practical Innovations, Open Solutions. [Link]
Pena, A., et al. (2021). Fuzzy convolutional deep-learning model to estimate the ACCESS.2023.3309007
operational risk capital using multi-source risk events. Applied Soft Computing, 107. S. Tyagi, “Analyzing machine learning models for credit scoring with explainable AI and
[Link] Aug. optimizing investment decisions,” 2022. Accessed: Jan. 05, 2025. [Online].
Qasaimeh, M., Hammour, R. A., Yassein, M. B., Al-Qassas, R. S., Torralbo, J. A. L., & Available: [Link]
Lizcano, D. (2022). Advanced security testing using a cyber-attack forecasting Ullah, I., Hussain, H., Ali, I., & Liaquat, A. (2019). Churn prediction in banking system
model: A case study of financial institutions. Journal of Software: Evolution and using K-means, LOF, and CBLOF. In 1st International Conference on Electrical.
Process, 34(11). [Link] Nov. Communication and Computer Engineering.
Rosenberg, J. V., & Schuermann, T. (2006). A general approach to integrated risk Varotto, S. (2011). Liquidity risk, credit risk, market risk and bank capital. International
management with skewed, fat-tailed risks. Journal of Financial Economics, 79(3), Journal of Managerial Finance, 7(2), 134–152. [Link]
569–614. [Link] Mar. 17439131111122139. Apr.
Rundo, Trenta, di Stallo, & Battiato. (2019). Machine learning for quantitative finance Viji, S., Kannan, R., & Jayalashmi, N. Y. (2021). Intelligent anomaly detection model for
applications: A survey. Applied Sciences, 9(24), 5574. [Link] atm booth surveillance using machine learning algorithm : Iintelligent ATM
app9245574. Dec. survillance model. In 2021 International Conference on Computing, Communication,
Sadok, H., Mahboub, H., Chaibi, H., Saadane, R., & Wahbi, M. (2023). Applications of and Intelligent Systems, ICCCIS 2021 (pp. 1007–1012). Institute of Electrical and
artificial intelligence in finance: Prospects, limits and risks. In Proceedings - 2023 Electronics Engineers Inc.. [Link]
International Conference on Digital Age and Technological Advances for Sustainable Feb.
Development, ICDATA 2023 (pp. 145–149). Institute of Electrical and Electronics Wang, Y., Jiang, Y., & He, G. (2019). Using artificial neural network to predict value-at-
Engineers Inc.. [Link] risk of S&P 500 market index with external information. In 2019 IEEE International
Sahingoz, O. K., Buber, E., Demir, O., & Diri, B. (2019). Machine learning based phishing Conference on Big Data, Cloud Computing, Data Science & Engineering (BCD) (pp.
detection from URLs. Expert Systems with Applications, 117, 345–357. [Link] 90–93). IEEE. [Link] May.
org/10.1016/[Link].2018.09.029. Mar. Wanke, P., Kalam Azad, M. A., Barros, C. P., & Hadi-Vencheh, A. (2016). Predicting
Sanford, A., & Moosa, I. (2015). Operational risk modelling and organizational learning performance in ASEAN banks: An integrated fuzzy MCDM–neural network approach.
in structured finance operations: A bayesian network approach. Journal of the Expert Systems, 33(3), 213–229. [Link] Jun.
Operational Research Society, 66(1), 86–115. [Link] Xu, J., & Yang, D. (2022). Financial risk control model based on deep neural networks.
Jan. Scientific Programming. [Link]
Sarangi, P. K., Mohanty, J., Mohapatra, S. K., & Sahu, P. (2023). Computing stock market Yu, Y., Xu, Y., Wang, J., Li, Z., & Cao, B. (2022). Anti-money laundering risk
price behavior using machine learning approach. In 6th International Conference on identification of financial institutions based on aspect-level graph neural networks.
In 2022 IEEE 22nd International Conference on Software Quality, Reliability and Security

16
V.L. Heß and B. Damásio International Journal of Information Management Data Insights 5 (2025) 100324

Companion (pp. 542–546). Institute of Electrical and Electronics Engineers Inc.. International Conference on Applied Artificial Intelligence and Computing, ICAAIC 2023
[Link] (pp. 661–665). Institute of Electrical and Electronics Engineers Inc.. [Link]
Zaychenko, Y., & Zgurovsky, M. (2019). Banks bankruptcy risk forecasting with 10.1109/ICAAIC56838.2023.10140628
application of computational intelligence methods. In IEEE 14th International Zhou, F., Qi, X., Xiao, C., & Wang, J. (2021). MetaRisk: Semi-supervised few-shot
Conference on Computer Sciences and Information Technologies (CSIT) (pp. 38–41). operational risk classification in banking industry. Information Science (N Y), 552,
[Link] 1–16. [Link] Apr.
Zhailybayevich, Z. K., & Hamada, M. A. (2023). Development of a predictive intellectual
model for predicting the financial crisis in banks. In Proceedings of the 2nd

17