online transaction research paper

Buyers' trust and mistrust in e-commerce platforms: a synthesizing literature review

• Original Article
• Published: 11 November 2021
• Volume 20 , pages 57–78, ( 2022 )

Cite this article

• Marzieh Soleimani   ORCID: orcid.org/0000-0003-1299-4391 1  

22k Accesses

36 Citations

Explore all metrics

This article has been updated

Electronic markets have grown substantially, and they are considered an effective form of retail in recent years. Despite such growth, lack of physical transactions between different parties, as well as users' concerns about their privacy and security of transactions in electronic commerce (e-commerce) platforms have jeopardized users' trust. Thus, trust as a key issue for reducing consumers' perceived risk and the successful promotion of e-commerce has motivated many researchers to study it. This paper created a comprehensive and up-to-date framework that synthesized the previous studies in the literature conducted on trust in e-commerce environments. A systematic literature review method was selected to achieve this aim. The initial search in 17 top-ranked information systems journals and conferences resulted in 129 papers that met the inclusion criteria. Then these studies underwent an in-depth examination to determine how trust had been conceptualized in e-commerce environments. Further, the theoretical bases in relation to trust in e-commerce contexts used in the literature were investigated. The study concludes with implications for practice and a critical agenda for future research.

Similar content being viewed by others

Perception is Reality

An examination of trust assurances adopted by top internet retailers: unveiling some critical determinants.

An analysis of consumer’s trusting beliefs towards the use of e-commerce platforms

Explore related subjects.

• Artificial Intelligence

Avoid common mistakes on your manuscript.

1 Introduction

Trust is of high significance that has been argued “a complete lack of trust would prevent [us] from getting up in the morning” (Luhmann 2018 , p. 4). Moreover, this concept has an ever-evolving history, going from restricted trust to family and friends to strangers in peer-to-peer platforms (Mazzella et al. 2016 ). Specifically, this concept has found its way into electronic commerce (e-commerce), which is perceived as a critical factor for success in online commerce (Chang et al. 2013 ) because it is considered as an essential factor separating buyers from non-buyers (Kim and Park 2013 ).

In a survey examining 6000 customers' data, trust in e-commerce platforms was given more importance, even, than price (Ernst and Young 2000 ). However, similar surveys showed that a small number of users could trust these platforms, especially when their privacy and security came into conflict (Connolly and Bannister 2007 ). With the growth of e-commerce as a market and economic force over the past two decades (Lim et al. 2006 ), the concept of trust has inevitably attracted many researchers' attention, causing to have been studying it using some models from various disciplines.

While several reviews have examined antecedents and consequences of online trust (e-trust), there appears to be confusion regarding the attribute in which e-trust is developed (Kim et al. 2006 ). According to Forbes, global e-commerce sales surged rapidly, from 2.9 trillion U.S. dollars in 2020 to 4.2 trillion U.S. dollars in 2021 (Verdon 2021 ). In addition, the COVID-19 pandemic has created an increasing desire to switch to online modes of shopping (Barnes 2020 ). Therefore, the growing number of studies and changes in online marketing necessitates an updated and comprehensive literature review taking various dimensions of trust, including the studies' theoretical bases into account.

Few, if any, systematic literature review has yet been done on the role of trust in e-commerce platforms, investigating the antecedents leading to improving trust in e-commerce platforms; and thus to help practitioners develop a framework for improving their platforms. Accordingly, the following research questions were sought:

RQ1. What factors are mentioned in the literature do affect trust in e-commerce platforms?

Furthermore, this study investigates the impact of trust on both other tangible and intangible features of e-commerce websites. Therefore, answering this second research question can help the investigation to illustrate the effects and benefits of e-trust better:

RQ2. What are the consequences of trust in e-commerce put forth in the literature?

Finally, the study looks at the theoretical concepts and areas used in the literature and synthesizes them in a single framework that can inform future studies. A third research question is proposed to cover this aspect of the research:

RQ3. What are the possible implications of the present study for future research?

The rest of the paper is structured as follows: an overview of the theme of trust in electronic commerce research is presented in Sect.  2 . Section  3 outlines the steps of this literature review and the criteria for including and excluding research papers in the final analysis. In Sect.  4 , the literature review results are presented and explained. These results are further discussed in Sect.  5 by outlining the contributions to theory and practice with an agenda for future research, and the study concludes in Sect.  6 .

2 Background

Various scholars have provided different definitions for trust based on their outlook. McKnight et al. ( 2002 ) categorized these definitions into two groups: conceptual types and referents types. Conceptual types include attitudes, beliefs, behaviors, and dispositions, whereas referent types include trust in something, trust in someone, or trust in a specific characteristic of someone (e.g., honesty). Later, drawing on this categorization, McKnight furthered different types of trust, including disposition to trust, which means one's general disposition to trust others; institutional trust, which means one's trust in situations or structures; and interpersonal trust, such as trust in e-vendor. These led to the multidimensional definitions of trust: “to willingly become vulnerable to the trustee, whether another person, an institution or people generally having taken into consideration the characteristics of the trustee.”

Mayor, on the other hand, listed three varying perspectives for trust, including psychology (a tendency to trust others), social psychology (cognition considerations of a trustee), and sociology (characteristics of the institutional environment). With this respect, Mayor defined trust as "the willingness of a party to be vulnerable to the actions of another party based on the expectation that the other will perform a particular action important to the trustor, irrespective of the ability to monitor or control that other party" (Mayer et al. 1995 , p. 712, Chiu et al. 2019 ; Tomlinson et al. 2020 ). Although this definition seems outdated, it has been used in recent publications and is considered the primary source of trust definition in e-commerce literature. The article in which this definition was published has been getting thousands of citations every year, and in total, it has been cited over 25,000 times.

The notion of trust has been an area of interest in organization studies and information systems for decades (Mayer et al. 1995 ; Li et al. 2008 ). In particular, the concept of electronic commerce trust was initiated in the late 1990s, with studies focusing on trust antecedents (Fung and Lee 1999 ). Earlier reviews have studied some limited aspects of trust, including the impact of uncertainty, the (new) meaning and typology of trust (Grabner-Kräuter and Kaluscha 2003 ; McKnight and Chervany 2001 ), the use of existing theories regarding trust (Huang et al. 2007 ), and the early models of trust (Papadopouou et al. 2001 ). However, the recent advancements and changes in electronic markets and the massive amount of work published in the past years are not assumed mainly beneficial for providing insight into state-of-the-art research.

Previous studies have also reviewed the impact of trust on disruptive models and platforms used in electronic markets. As an illustration, in a study conducted by Hawlitschek et al. ( 2018 ) in the context of sharing economy, trust was categorized into "trust in peers" and "trust in the platform". In the end, they suggested blockchain as a technological solution to improving trust. In the realm of e-commerce, several systematic literature reviews were sought to understand the importance and the evolving nature of trust. Beldad et al. ( 2010 ), for instance, examined antecedents of trust for commercial and non-commercial online firms, realizing that there were a host of antecedents for which they designed a framework, including three clusters: customer/client-based, website-based, and company/organization-based. A meta-analysis on this topic was carried out by Kim and Peterson ( 2017 ), surveying 150 empirical studies revealed that antecedents such as perceived service quality, perceived privacy, and perceived reputation were inherently associated with the antecedents of online trust. The study also listed the consequences, often dealt with mentioned in previous empirical studies, including satisfaction, attitude, loyalty, repeating purchase intention, and intention to use the website. While, so far, the systematic literature reviews have mainly focused on the antecedents and consequences of trust, factors such as disposition to trust, security, familiarity, and risk perception received less attention.

Although extensive research has been carried out on trust in e-commerce, few writers have thus far striven to draw on a systematic literature review, focusing on the characteristics of the antecedents and consequences of trust in the context of electronic commerce. Therefore, the current study was motivated to shed light on this area by focusing not only on the components of research models but also on the theoretical concepts deeply.

3 Methodology

To investigate the factors leading to and impacted by trust in the electronic commerce environment, the Systematic Literature Review (SLR) approach was adopted, being a methodical way to identify, evaluate, and interpret the available empirical studies conducted on a particular topic, research question, or phenomenon of interest (Kitchenham 2004 ). Considering the research aims in this study, the SLR approach was built on synthesizing the available literature by summarizing and organizing published articles, as well as clarifying how prior literature has contributed to knowledge development in this area (Schryen et al. 2020 ). To do so, First, 17 high-ranked IS journals and conferences, among the considerable number of research conducted on trust in e-commerce, were selected as a representative of the whole body of knowledge and searched with a predefined set of keywords. It is important to add that the reasons behind selecting those top journals and conferences were IS journals' available rankings (Fisher et al. 2007 ), previous SLR work (Tallon et al. 2019 ; Amrollahi et al. 2013 ), and their tendency for emphasis on e-commerce and related areas. Also, it should be noted that the focus of selected conferences was AIS sponsored conferences, including ICIS, PACIS, ECIS, and AMCIS. Then, through the initial search, 601 papers were found. Then, irrelevant articles were excluded after reviewing papers' titles, abstracts, and full texts. The final set of papers were investigated against the research questions in this study.

3.1 Keywords

To select the keywords with the best results, Scopus was first searched using broad keywords “trust in electronic commerce” and “online shopping trust.” After reviewing the first ten pages of the search results, the keywords were refined. Finally, the following terms were applied to limit the search in titles, keywords, and abstracts in the search engine: Trust AND (“retail” OR e-commerce” OR “electronic commerce” OR “electronic business” OR “e-business” OR “shop” OR “sale” OR “buy” OR “purchase” OR “e-Trust” OR “market”).

Table 1 An overview of selected outlets shows the outlets used as the sources for this study, the rationale for selecting the outlet, and the number of papers found in each outlet.

3.2 Data extraction and analysis

The initial search ran based on the above keywords in October 2019. To build a meticulous review of the recent literature, with the articles published in the past twenty years, a time span from 2000 was defined. The first stage search reached 601 studies. After limiting the document types to only articles and conference papers by excluding book chapters, books, reviews, conference reviews, and short surveys, as well as filtering the subject areas to business, management, and accounting, social science, economics and finance, decision sciences, arts and humanities, and psychology, I ended up with 482 studies. In the next step, to choose the relevant studies, among which their titles and abstracts were screened while putting emphasis on the role of trust, the way of its development, and its consequences in exclusively e-commerce environments, which led to obtaining a total of 129 studies for main and in-depth investigation. For this reason, technical articles excluded those investigating trust in contexts other than electronic commerce such as tourism, sharing economy, Internet of things, and those that focused on different aspects, including supply chain.

3.3 Data analysis

To analyse the final list of articles, first, they were differentiated based on their perceptions of trust and the different theories they used. Then, content analysis was performed to extract the factors that were believed to impact various forms of trust, along with its potential consequences in electronic markets. Afterward, these factors were categorized into multiple groups. To triangulate the data, the whole set of papers and the finalized categories of factors impacting trust and its consequences were presented to two colleagues with expertise in e-commerce to check them and create their own categories. The results of the data analysis are explained in the following section (Fig.  1 ).

Stages of research methodology

4.1 Theories

By meticulously examining the literature review, 29 theories from various disciplines were found, including Marketing, Psychology, Economy, Sociology, Management and Organization Science, Computing, Information Systems, and Philosophy. It should be noted that due to the interdisciplinary characteristics of many of these theories in most cases, the origin of the discipline, where it had been published, was checked. Then, I probed into each theory to see if and how the constructs might fit into the notion of trust in e-commerce. Among these, Technology Acceptance Model (Davis et al. 1989 ; Venkatesh and Davis 2000 ) and Theory of Planned Behavior (Ajzen 1991 ) had been cited most as explained the processes leading to trust-related behavior and how technological components can be trusted in an e-commerce environment.

Concerning trust theories within the discipline of marketing, Signalling Theory (Spence 1974 ) and Expectation Confirmation Theory (Li et al. 2015 ), had been cited in the literature to explain customers' behavior in seeking relevant information and modeling factors, leading to their satisfaction. Typically, marketing theories look at social factors that influence trust; on the other hand, psychological theories mainly focus on customers' individually trusting behavior. Finally, sociological and organizational theories, such as Social Exchange Theory (Woisetschläger et al. 2011 ) and Social Capital Theory (Coleman 1988 ), are used to model social relationships between human actors and how trust can be developed as a result of these interactions.

As illustrated in Fig.  2 , four elements were identified to be considered as the antecedents in these theories, including intention, behavior, consequence, and environmental factors. Such antecedent factors can result in a specific behavior. More specifically, such antecedents usually lead to a set of factors, broadly named intention factors by which any intrinsic motivation or resistance to perform a behavior can be triggered and resulting in some consequences, such as reward or satisfaction. Furthermore, in e-commerce trust literature, we also found three other different antecedent factors based on the theory's level and origin (individual, market-related, and social factors).

Synthetization of theories in the final set of research studies

In Fig.  2 , the numbers in front of each item shows the theories used in the particular component. For example, TAM depicted with number 17, posits perceived usefulness and perceived ease of use (being a part of perception in antecedent factors) as an individual's intention to use a system (being a part of intention in intention factors); also, intention of an individual as a mediator of actual system use (being a part of transaction in behavior factors and reward in consequences).

4.2 Types of trust

As a part of the systematic literature review, I tried to carefully examine and differentiate various types of trust in the e-commerce context. To differentiate these types of trust, I found multiple stakeholders in a trusted transaction and the mechanism that they trust each other. Under such scrutiny, I recognized four types of trust and demonstrated them in Table 2 .

4.2.1 Customers' trust in sellers

The most common type of trust we found in this study is when customers as trustors expect another party (usually sellers) to do an accepted behavior. For example, Carter et al. ( 2014 ) studied the impact of trust on travelers' loyalty to online service providers. While many studies have focused on the effect of trust on behaviors, like performing a transaction (McKnight et al. 2002 ; Pennington et al. 2003 ; Kim et al. 2009 ), other studies have looked at long-term factors like loyalty (Li et al. 2015 ) and adoption of e-commerce (Pavlou and Fygenson 2006 ).

4.2.2 Community trust

Community trust serves as a generalized trust (one-to-many). It refers to trusting many trustees or trustors, especially those considered an unknown group of sellers or buyers, with the help and support of a specific online marketplace (Pavlou and Gefen 2004 ). It suggests that there would be a slight possibility for online buyers encountering the same seller twice in this research category (Pavlou and Gefen 2004 ). Sun ( 2010 , p. 6) defined trust in the community of buyers as a "seller's subjective beliefs that buyers will behave in accordance with the seller's confident expectations by showing ability, integrity, and benevolence."

4.2.3 Technology trust

In this study, we considered trust in online shopping as a shopping mode, the Internet as an online store or platform, and social commerce websites as elements of technology trust. McKnight ( 2005 ) defined technology trust as the trustor's beliefs in Information Technology (IT)'s trustworthiness to perform a task. With higher reliance on technology in recent years, trust in technology has gained more attention, too. For instance, upon conducting an online purchase, customers expect the technological infrastructures to provide appropriate conditions to help with online tracking, online supports, pictures, quality, and information specificity. However, technology trust is considered beyond transaction fulfillment as many websites offer such features, including product recommendations, product comparisons, and customer reviews (Li et al. 2009 ).

4.2.4 Sellers' trust in customers

Although there are many studies conducted on the role of trust in sellers, far too little attention has been paid to sellers' trust in other stakeholders, especially buyers. In this line, Sun ( 2010 ), in his study, mentioned that there is a substantial difference in trust behavior between sellers and buyers, stemming from various technical, political, and institutional dimensions. Therefore, sellers' trust is defined as the willingness of sellers to risk participating in a transaction, even when uncertainties occur. Sellers need to trust that buyers can make transactions with competence, benevolence, and integrity (Chong et al. 2003 ; Resnick and Zeckhauser 2002 ). For instance, Guo et al. ( 2018 ) surveyed Chinese sellers in a Business-to-Business (B2B) platform to understand the mechanism under which they could trust an online trading transaction. Similarly, in other studies, the possibility of a transaction as the outcome of B2B model was considered (Sun and Zhang 2008 ).

4.3 Antecedents to trust

As explained in the methodology part, after analyzing the final set of papers, the factors that impact trust, along with outcomes for online trust were extracted and categorized into different groups. In this section, you can see these categorized factors in detail as antecedents and consequences of trust.

Antecedents to trust are categorized based on various actors in an online transaction. Figure  3 illustrates these identified categories, which are customer-related antecedents, seller-related antecedents, technology and third-party antecedents, and environment-related antecedents. The findings of this section allow us to begin answering RQ1: What factors are mentioned in the literature do affect trust in e-commerce platforms. The remainder of this section explains each category in more detail.

The relationship among identified antecedents of trust

4.3.1 Customer-related antecedents

Understanding the antecedent of customers' trust can provide invaluable insights into the factors that can possibly urge them to create trust and improve their intentions to make an online transaction. Based on the literature, customer concerns, disposition to trust, trusting beliefs, familiarity, calculative-based trust, accessibility of information, and other similar terms were considered as the main antecedents for this category.

In e-commerce, the process of building trust for customers is affected by customers' concerns, which are considered severe obstacles in electronic transactions (Agag et al. 2020 ; Kim 2008 ). According to the literature, four primary concerns for online customers are privacy, security, perceived technology risk, and integrity concerns (Connolly and Bannister 2007 ; Shukla 2014 ).

Disposition to trust, also known as propensity to trust , is "the extent to which a person displays a tendency to be willing to depend on others across a broad spectrum of situations and people" (McKnight et al. 2002 , p. 339). McKnight and Chervany ( 2002 ) suggested that the effects of dispositional factors on trust are more than other factors such as institution-based trust. Such a general propensity to trust in others can influence the intentions and beliefs of trustors about e-vendors.

Another type of customer-related antecedent is familiarity, which relates to a customer's prior behavior. Bhattacherjee ( 2002 , p. 220) noted that "familiarity refers to one's understanding of another's behavior based on prior interactions or experiences." He also mentioned that familiarity gradually develops over time as trustees become accustomed to trustors' behavior and, in turn, improve trust in online buyers.

Based on calculative-based trust, an online customer can build trust through cost–benefit analysis of trustees whose behavior shows whether they are cheating or cooperating. Since calculative-based trust is deterrence-based, customers will not engage in opportunist behavior when they feel that the e-vendor is untrustworthy. Customers will trust e-vendors when they believe that the e-vendor either has more to lose by cheating or has nothing to gain by breaking the consumer's trust (Gefen et al. 2003a ).

4.3.2 Seller related antecedents

The second category of factors that has an impact on trust is those related to sellers. Elements such as institution-based trust, reputation, communication, and interaction are considered sellers' characteristics in the literature.

As the first characteristic of sellers studied in the literature, reputation refers to the extent to which a trustee believes that a trustor has integrity and is concerned about its consumers (Kit et al. 2013 ). Researchers seem to have adopted different terms, referring to reputation, such as perceived effectiveness of feedback mechanism, brand awareness, brand image, perceived accreditation, portal affiliation, and online aesthetic appeal. In fact, these researchers in their studies indicated that reputation leads to trust in the e-commerce context. Therefore, there is a positive relationship between reputation and online trust (Hoffmann et al. 2014 ; Kit et al. 2013 ; Shiau and Chau 2015 ). In a similar vein, other researchers suggest that reputation has a vital role in engendering trust and in repurchase intention (Qureshi et al. 2009 ).

Effective communication is also another element regarded as crucial for trust in e-commerce. Moreover, interpersonal relationships or the ability to interact intimately in social networks are considered examples of trust antecedents in this category. And finally, institution-based mechanisms such as warranty and structural assurance are significant factors leading to trust in e-commence (Huang et al. 2005 ; Wang and Benbasat 2008 ).

4.3.3 Technology and third-party related antecedents

The positive impact of website quality and third-party institutions on trust has been supported in many studies. For instance, Jones and Leonard ( 2008 ) examined trust in customer-to-customer (C2C) environments. They confirmed that when customers do not know each other in online environment, they take cues from social signals, including website quality and third-party institutions. Perceived website quality demonstrates user's perceptions of some features such as the ease of use and usefulness of information (Awad adn Ragowsky 2008 ). Web users need to feel the website is well designed, organized, timely, and accurate to believe web vendors' trustworthiness (Flavin et al. 2006 ; Qureshi et al. 2009 ).

The role of intermediaries in developing consumers' trust in e-commerce markets is remarkably important, especially for small e-vendors (Datta abd Chatterjee 2008 ). McKnight and Chervany ( 2002 ) believe that since trust is transferable, an intermediary can have such responsibility for transferring consumers' trust in a brand to e-vendors. With this regard, Pavlou and Gefen ( 2004 , p. 44) defined an online intermediary as a "third-party institution that uses internet infrastructure to facilitate transactions among buyers and sellers in its online marketplace by collecting, processing, and disseminating information." Consumers need to receive strong signals to trust other parties. A trusted third party, thus, can play this role to facilitate transactions (Clemons et al. 2016 ). Intermediaries can reduce the risk of making transactions in an e-commerce environment by producing a reliable and secure environment, instantiating fair and open rules and procedures, presenting accredit and evaluation, getting rid of problematic sellers, and encouraging benevolent transaction norms. In this way, for instance, some intermediaries such as eBay and Amazon try to reduce transactions risk by providing coverage up to a limit of $250 to guarantee their auction transactions so that buyers could reduce their actual risk in transactions (Pavlou and Gefen 2004 ).

For sellers, online intermediaries can help them obtain market signals, reduce search costs, discover better prices, deliver products at a lower price, facilitate transaction settlements, and monitor buyers (Giaglis et al. 2002 ). Sellers need to trust that the intermediary performs these functions honestly, competently while having the interests of sellers in mind.

4.3.4 Environment-related antecedents

According to the literature, environment antecedents consist of word of mouth (WOM), culture, trusting beliefs, and perceived size. E-WOM is at the center of consumer behavior and is defined as any positive, negative, or neutral comments, recommendations or statements about a product, service, brand, or company based on prior experiences and knowledge created by former, potential, or actual web users. E-WOM will usually spread via the Internet and social networks (Hennig-Thurau et al. 2004 ). An influential study with respect to the investigation of e-WOM and its effect on trust by Awad and Ragowsky ( 2008 ) approved that e-WOM quality can positively impact online trust. Kim ( 2008 ) also supported the idea that positive referrals and posts can directly affect customer's trust in an e-vendor.

Regarding investigating the impact of culture on trust and consumer behavior, previous studies have mainly employed Hofstede's cultural dimensions. He defined culture as “the collective programming of the mind that distinguishes the members or category of people from another” (Hofstede et al. 2010 , p. 6). Hofstede's original four dimensions of culture, including uncertainty avoidance, power distance, collectivism, and masculinity, have been frequently studied in the e-trust literature. Some studies focused on examining the effect of uncertainty avoidance on trust and online consumer behavior (Bui et al. 2013 ; Hwang 2005 ), while other studies have cited individualism/collectivism as a primary antecedent of trust in Internet shopping (Sia et al. 2009 ).

4.4 Consequences of trust

Different studies demonstrate various outcomes for online trust. For instance, McKnight et al. ( 2002 ) stated that online trust significantly affects online consumers' purchase intention; similarly, Sun ( 2010 ) referred to the positive impact of trust on repurchase intention. Trust can change risk perception, consumer attitude, and consumer's perception regarding website and seller (Jarvenpaa et al. 2000 ). Trust can also result in e-loyalty and satisfaction in online shopping procedures (Honglei Li et al. 2015 ; Shankar et al. 2002 ). The results of this section provides an answer to RQ2: What are the consequences of trust in e-commerce put forth in the literature? Each category related to the consequences of trust is discussed below in more detail.

4.4.1 Transaction intention

Transaction intention is the trustor's willingness to participate in an online transaction with a trustee (McKnight et al. 2002 ; Pavlou and Gefen 2004 ). According to the theory of reasoned action (TRA), trust can be regarded as a behavioral belief that makes a positive attitude toward purchase intention (Jarvenpaa et al. 2000 ; Pavlou and Gefen 2004 ). As discussed earlier, this behavioral intention is a common component in many other theories used in online trust literature. Consistent with previous studies in the literature, online consumer trust demonstrates a substantial positive influence on transaction intention (Kim et al. 2015 ; Kim and Ahn 2007 ).

4.4.2 Retention and loyalty

Customer retention can be achieved when one believes in the trustworthiness of an e-vendor and his/her ability in fulfilling promises, which in turn, increasing the possibility of repurchase intention (Hong and Cho 2011 ; Liu and Tang 2018 ). Customer loyalty in an online environment, also known as e-loyalty, refers to "an enduring psychological attachment by a customer to a particular online vendor or service provider" (Cyr et al. 2007 , p. 44).

The direct and piercing impact of trust on retention (Qureshi et al. 2009 ) and e-loyalty (Carter et al. 2014 ) have been discussed in previous literature. For example, trust has been considered a factor engendering affective commitment, which helps develop an online customer's intention to revisit the website and purchase in the upcoming future (Liu and Tang 2018 ). As for trusting beliefs, when online consumers already establish trusting beliefs, they will hardly switch to other e-vendors because of the risk and uncertainty involved with finding a new online seller and difficulties associated with establishing a new trusting relationship (Carter et al. 2014 ).

4.4.3 Perception about seller

To trust sellers, customers and buyers are expected to have an improved image of sellers, the appropriate technological conditions used for online transactions, and improved conditions of the services offered to. This feature comprises perceived usefulness of used technology, perceived enjoyment, perceived benefits, perceived value, price premium, and e-WOM intention about the seller. When buyers trust sellers as a result of an improved perception, they allow sellers to obtain premium prices and yield above-average profits (Klein and Leffler 1981 ; Shapiro 1983 ) to compensate seller transaction risks (Ba and Pavlou 2002 ).

4.4.4 Satisfaction

A desirable outcome of a trust-based relationship is satisfaction (Cannon 1999 ). When online customers trust a web vendor, they are likely to be satisfied with their transactions (Pavlou 2002 ). The relationship between trust and satisfaction has been investigated in many studies. For instance, Chang et al. ( 2016 ) investigated the effect of trust on satisfaction, indicating that customers' trust had significant and positive impacts on perceived satisfaction and transaction intention. Likewise, Pavlou ( 2002 ) stated that trust in the sellers' credibility positively influences buyers' satisfaction. Even recently, more studies have investigated a complicated relationship between trust, satisfaction, and transaction (Chang et al. 2016 ).

4.4.5 Risk perception

Finally, according to the literature, customers' and sellers' perceptions of the risks involved in an online transaction can be impacted by their trust in the other party. This relationship is different from research studies that considered risk perception an antecedent of trust (Gefen et al. 2003b ). Perceived risk is defined as “the possibility of loss” and “an inherently subjective construct.” However, the notion of risk is closely related to more general items and, consequently, emphasizes the possibility of economic loss (Dinev and Hart 2006 ). Electronic vendors try to reduce the risk perceptions of e-consumers through IT tools, e.g., third-party insurances (Gefen et al. 2008 ). In summary, trust can substantially help decrease negative risk perceptions toward an e-commerce context (Connolly and Bannister 2007 ; Guo et al. 2018 ; Shukla 2014 ; Heijden et al. 2001 ; Verhagen et al. 2006 ).

5 Discussion

This study presented a comprehensive and systematic review of antecedents and consequences of trust in online markets. Unlike the previous reviews, the present study considered both antecedents and consequences of trust from both customers' and sellers' points of view by providing a model using a synthesize of conceptually related theories in this area. Figure  4 illustrates a comprehensive model of trust based on my review of related studies in the literature.

A detailed model of antecedents and consequences of trust in e-commerce

Although the factors demonstrated in Fig.  4 have used with differing frequencies in the literature, their comprehensive and unified review can usefully provide a big picture of the research studies conducted earlier on this topic. This section will discuss these findings by listing potential implications for research and practice.

5.1 Implications for research

In answer to RQ3, we looked at the theoretical concepts and contextual dimensions in the literature beyond investigating the factors impacting on or being impacted by online trust. The present study can benefit future research in several different ways. First, reviewing the theories used in this area can help future studies identify their similarities and differences.

In addition, a synthesis of the theories used in the review can highlight the factors which have been less focused on. For example, the review of literatures indicates that little is known about interpreting trust by different parties. Therefore, in the current study, we suggest three future directions which can be conducted on online trust studies. First, future research can make perfect use of different models and theories, such as motivational model (Keller 1983 ) and accountability theory (Lerner and Tetlock 1999 ) to shed light on the less studied factors in online trust, like social presence, accountability, relevance, and confidence.

Furthermore, models and theories concerned with communication and media (Shannon and Weaver 1949 ; Toulmin 2003 ; Miller 1956 ) can be used to identify the way trust is understood and perceived in an online environment and the way it can be transferred to various situational contexts (Schultz 2006 ). Such theoretical perspectives can also be extended to explore the unobservable events that impact and create trustworthiness in an online environment and give inter-subjective meaning to trust by different stakeholders in various contexts (Dobson et al. 2007 ; Phillips and Brown 1993 ). In terms of research approaches and perspectives, the literature, so far, seems to have ignored them. Future studies can benefit from available research design and methods (Hevner 2007 ; Gregor and Jones 2007 ) and extend the existing literature on prescriptive work to improving trust in e-commerce.

Moreover, as mentioned in the previous sections, most studies in the literature (almost 96%) have focused on the mechanisms under which customers trust buyers, technology, and a community. Apart from customers, other stakeholders involved in an online transaction can receive considerable attention. In particular, sellers as trustees should be paid more attention in future research. Future studies should also go beyond available research on B2B e-commerce (Tsatsou et al. 2010 ; Pavlou 2002 ) and online auctions (Pavlou and Dimoka 2006 ) to focus on disruptive technologies that promote the position of customers in e-commerce (Hawlitschek et al. 2018 ). Such a trend brings new changes and possibilities to electronic markets in which customers are put a central position. Additionally, future research calls for more attention to new forms of transaction and technological advancements in payment such as blockchain and distributed ledger (Lacity 2018 ; Lindman et al. 2017 ) and how customers put trust in these technologies.

Surprisingly, the existing knowledge on trust in e-commerce in the literature seems to be very constrained in terms of demographic characteristics since most studies have mainly taken North American, European, and East Asian contexts into account. In the same way, there are rare studies addressing significant differences of trust in e-commerce platforms in developed and developing countries (Hajli 2019 ; Shareef et al. 2018 ). In terms of subjects' age, although a wide range of subjects differing in their age has been reported in studies in the literature (Hoffmann et al. 2014 ), many of which are usually targeted at subjects with young-age groups, like students. However, in the long term, this type of research will result in the exclusion of certain parts of society (Cushman et al. 2008 ). Therefore, future research on trust in e-commerce calls for the inclusion of a diversity of subjects, especially those who belong to the older generation of users.

5.2 Implications for practice

The present study is of high significance from different aspects. In effect, the classification proposed in this study is beneficial for practitioners, managers, and owners of e-commerce platforms and online businesses as they can be empowered to put those factors that influence customers' trust in e-commerce into practice to bring about improvements in their e-commerce platforms and online businesses. In particular, reliability, coherence, visual appearance, and website qualities that are likely to affect customers' trust in e-commerce could be enhanced. Likewise, to make buyers and customers trust more on e-commerce platforms, their owners can make use of a trusting third party or an intermediary. Also, practitioners need to consider online seals, encryption certificates, assurance, and guarantees in the process of trust.

Further, it will help practitioners to form new approaches to develop and refine the notion of trust in e-commerce flourishingly. Finally, as the current study suggested, e-commerce platform developers are required to be technologically accustomed to improving the conditions under which customers can trust their platforms.

6 Conclusion

Despite long-standing research conducted in the past two decades on trust in e-commerce environments, no comprehensive study has examined the research body of knowledge in this area, aimed to develop a comprehensive framework. Hence, this paper used a systematic literature review approach to investigate the factors that can impact trust in e-commerce platforms, as well as to find out its possible consequences. Based on my review of 129 papers in high-ranked IS journals and conferences, I identified various antecedents and consequences of trust.

Like any other review type of research, the study is not devoid of some limitations. For example, the study did not include those high-ranked journals and conferences that could not satisfy the designated metrics with respect to the topic of interest. Also, despite my attempts to form a comprehensive set of keywords, some of which may have been missed due to different terms that could be used in a specific interface. However, the study made an endeavour to be successful in developing a comprehensive framework in relation to trust among customers and sellers in e-commerce platforms using a synthesis of related theories in this area, which may be beneficial for future research and practice.

Change history

27 march 2022.

In “Results” section, the numbering of headings were incorrect and it is corrected now.

Agag GM, Khashan MA, Colmekcioglu N, Almamy A, Alharbi NS, Eid R et al (2020) Converting hotels website visitors into buyers How online hotel web assurance seals services decrease consumers ’ concerns and increase online booking intentions. Inf Technol People 33(1):129–159

Article   Google Scholar  

Ajzen I (1991) The theory of planned behavior. Organ Behav Hum Decis Process 50(2):179–211

Amrollahi A, Ghapanchi AH, Talaei-Khoei A (2013) A systematic literature review on strategic information systems planning: insights from the past decade. Pacif Asia J Assoc Inf Syst 5(2):39–66

Google Scholar  

Awad NF, Ragowsky A (2008) Establishing trust in electronic commerce through online word of mouth: an examination across genders. J Manag Inf Syst 24(4):101–121

Ba S, Pavlou PA (2002) Evidence of the effect of trust building technology in electronic markets: price premiums and buyer behavior. MIS Q 26:243–268

Barnes SJ (2020) Information management research and practice in the post-COVID-19 world. Int J Inf Manag 55:102175

Beldad A, De Jong M, Steehouder M (2010) How shall I trust the faceless and the intangible? A literature review on the antecedents of online trust. Comput Hum Behav 26(5):857–869

Bhattacherjee A (2002) Individual trust in online firms: Scale development and initial test. J Manag Inf Syst 19(1):211–241

Bui SN, Kettinger WJ, Park I (2013) Personalization to new website users: the role of trust and culture. In: Proceedings of the nineteenth americas conference on information systems

Cannon JPA (1999) Buyer–seller relationships in business markets. J Mark Res 36(4):439–460

Carter M, Wright R, Thatcher JB, Klein R (2014) Understanding online customers’ ties to merchants: the moderating influence of trust on the relationship between switching costs and e-loyalty. Eur J Inf Syst 23(2):185–204

Chang MK, Cheung W, Tang M (2013) Building trust online: Interactions among trust-building mechanisms. Inf Manag 50(7):439–445

Chang S-H, Chih W-H, Liou D-K, Yang Y-T (2016) The mediation of cognitive attitude for online shopping. Inf Technol People 29:618–646

Chiu CL, Chiu JL, Mansumitrchai S (2019) Stages in the development of consumers’ online trust as mediating variable in online banking system: a proposed model. Int J Electr Finance 9(3):170–201

Chong B, Yang Z, Wong M (2003) Asymmetrical impact of trustworthiness attributes on trust, perceived value, and purchase intention: a conceptual framework for cross-cultural study on consumer perception of online auction. In Proceedings of the 5th international conference on Electronic commerce, pp 213–219

Clemons EK, Wilson J, Matt C, Hess T, Ren F, Jin F, Koh NS (2016) Global differences in online shopping behavior: Understanding factors leading to trust. J Manag Inf Syst 33(4):1117–1148

Coleman JS (1988) Social capital in the creation of human capital. Am J Sociol 94:S95–S120

Connolly R, Bannister F (2007) Consumer trust in Internet shopping in Ireland: towards the development of a more effective trust measurement instrument. J Inf Technol 22(2):102–118

Cushman M, McLean R, Zheng Y, Walsham G (2008) Inequality of what? Social exclusion in the e‐society as capability deprivation. Inf Technol People

Cyr D, Hassanein K, Head M, Ivanov A (2007) The role of social presence in establishing loyalty in e-service environments. Interact Comput 19(1):43–56

Datta P, Chatterjee S (2008) The economics and psychology of consumer trust in intermediaries in electronic markets: the EM-trust framework. Eur J Inf Syst 17(1):12–28

Davis FD, Bagozzi RP, Warshaw PR (1989) User acceptance of computer technology: a comparison of two theoretical models. Manag Sci 35(8):982–1003

Dinev T, Hart P (2006) An extended privacy calculus model for e-commerce transactions. Inf Syst Res 17(1):61–80

Dobson P, Myles J, Jackson P (2007) Making the case for critical realism: examining the implementation of automated performance management systems. Inf Resour Manag J (IRMJ) 20(2):138–152

Ernst CG, Young (2000) Perceptions of trust. ECommerce Alerts Issued by Bennett Gold, Chartered Accountants

Fisher J, Shanks G, Lamp JW (2007) A ranking list for information systems journals. Aust J Inf Syst 14(2):5–15

Flavin C, Guinalu M, Gurrea R (2006) The role played by perceived usability, satisfaction and consumer trust on website loyalty. Inf Manag 43(1):1–14

Fung R, Lee M (1999) EC-trust (trust in electronic commerce): exploring the antecedent factors. AMCIS 1999:179

Gefen D, Benbasat I, Pavlou P (2008) A research agenda for trust in online environments. J Manag Inf Syst 24(4):275–286

Gefen D, Karahanna E, Straub DW (2003a) Trust and TAM in online shopping: an integrated model. MIS Q 27(1):51–90

Gefen D, Rao VS, Tractinsky N (2003b) The conceptualization of trust, risk and their relationship in electronic commerce: The need for clarifications. 192b

Giaglis GM, Klein S, O’Keefe RM (2002) The role of intermediaries in electronic marketplaces: developing a contingency model. Inf Syst J 12(3):231–246

Grabner-Kräuter S, Kaluscha EA (2003) Empirical research in online trust: a review and critical assessment. Int J Hum Comput Stud 58(6):783–812

Gregor S, Jones D (2007) The anatomy of a design theory. J Assoc Inf Syst 8(5):312–335

Guo Y, Bao Y, Stuart BJ, Le-Nguyen K (2018) To sell or not to sell: exploring sellers’ trust and risk of chargeback fraud in cross-border electronic commerce. Inf Syst J 28(2):359–383

Hajli N (2019) The impact of positive valence and negative valence on social commerce purchase intention. Inf Technol People

Hawlitschek F, Notheisen B, Teubner T (2018) The limits of trust-free systems: A literature review on blockchain technology and trust in the sharing economy. Electron Commer Res Appl 29:50–63

Heijden HVD, Verhagen T, Creemers M (2001) Predicting online purchase behavior: replications and tests of competing models. In: Proceedings of the 34th Annual Hawaii international conference on system sciences

Hennig-Thurau T, Gwinner KP, Walsh G, Gremler DD (2004) Electronic word-of-mouth via consumer-opinion platforms: what motivates consumers to articulate themselves on the Internet? J Interact Mark 18(1):38–52

Hevner AR (2007) A three cycle view of design science research. Scand J Inf Syst 19(2):4

Hoffmann CP, Lutz C, Meckel M (2014) Digital natives or digital immigrants? The impact of user characteristics on online trust. J Manag Inf Syst 31(3):138–171

Hofstede G, Hofstede GJ, Minkov M (2010) Cultures and organizations: software of the mind. Revised and expanded, 3rd edn. McGraw-Hill, New York

Hong IB, Cho H (2011) The impact of consumer trust on attitudinal loyalty and purchase intentions in B2C e-marketplaces: intermediary trust vs seller trust. Int J Inf Manag 31(5):469–479

Huang L-T, Farn C-K, Yin K-L (2005) On initial trust-building for e-commerce: revisiting from the perspective of signal theory and trust transference. In: ECIS 2005 Proceedings, p 94

Huang S-Y, Li C-R, Lin C-J (2007) A literature review of online trust in business to consumer e-commerce transactions 2001–2006. Inf Syst 8(2):63–69

Hwang Y (2005) An empirical study of online trust and consumer behavior: cultural orientation, social norms, and personal innovativeness in information technology. In: ICIS 2005 Proceedings, p 69

Jarvenpaa SL, Tractinsky N, Vitale M (2000) Consumer trust in an Internet store. Inf Technol Manag 1(1–2):45–71

Jones K, Leonard LN (2008) Trust in consumer-to-consumer electronic commerce. Inf Manag 45(2):88–95

Keller JM (1983) Motivational design of instruction. Instruct Des Theor Models Overv Their Curr Status 1(1983):383–434

Kim DJ (2008) Self-perception-based versus transference-based trust determinants in computer-mediated transactions: a cross-cultural comparison study. J Manag Inf Syst 24(4):13–45

Kim M-S, Ahn JH (2007) Management of trust in the e-marketplace: the role of the buyer’s experience in building trust. J Inf Technol 22(2):119–132

Kim D, Benbasat I (2006) The effects of trust-assuring arguments on consumer trust in Internet stores: application of Toulmin’s model of argumentation. Inf Syst Res 17(3):286–300

Kim S, Park H (2013) Effects of various characteristics of social commerce (s-commerce) on consumers’ trust and trust performance. Int J Inf Manage 33(2):318–332

Kim Y, Peterson RA (2017) A meta-analysis of online trust relationships in e-commerce. J Interact Mark 38:44–54

Kim DJ, Ferrin DL, Rao HR (2009) Trust and satisfaction, two stepping stones for successful e-commerce relationships: a longitudinal exploration. Inf Syst Res 20(2):237–257

Kim DJ, Sugumaran V, Rao HR (2015) Web assurance seal services, trust and consumers ' concerns: an investigation of e-commerce transaction intentions across two nations. Eur J Inf Syst 1–22

Kit M, Cheung W, Tang M (2013) Building trust online: interactions among trust building mechanisms. Inf Manag 50:439–445

Kitchenham B (2004) Procedures for performing systematic reviews. Keele UK, Keele Univ 33(2004):1–26

Klein B, Leffler KB (1981) The role of market forces in assuring contractual performance. J Polit Econ 89(4):615–641

Lacity MC (2018) Addressing key challenges to making enterprise blockchain applications a reality. MIS Q Exec 17(3):201–222

Lerner JS, Tetlock PE (1999) Accounting for the effects of accountability. Psychol Bull 125(2):255

Li X, Hess TJ, Valacich JS (2008) Why do we trust new technology? A study of initial trust formation with organizational information systems. J Strateg Inf Syst 17(1):39–71

Li X, Rong G, Thatcher JB (2009) Do we trust the technology? People? or both? Ruminations on technology trust. In: AMCIS 2009 proceedings, p 459

Li H, Aham-Anyanwu N, Tevrizci C, Luo X (2015) The interplay between value and service quality experience: e-loyalty development process through the eTailQ scale and value perception. Electron Commer Res 15(4):585–615

Lim KH, Sia C, Lee MKO, Benbasat I (2006) Do I trust you online, and if so, will I buy? An empirical study of two trust-building strategies. J Manag Inf Syst 23(2):233–266

Lindman J, Tuunainen VK, Rossi M (2017) Opportunities and risks of blockchain technologies—a research agenda. In: Proceedings of the 50th Hawaii international conference on system sciences

Liu Y, Tang X (2018) The effects of online trust-building mechanisms on trust and repurchase intentions: an empirical study on eBay. Inf Technol People 31:666–687

Luhmann N (2018) Trust and power. Wiley, New York

Mayer RC, Davis JH, Schoorman FD (1995) An integrative model of organizational trust. Acad Manag Rev 20(3):709–734

Mazzella F, Sundararajan A, d’Espous VB, Möhlmann M (2016) How digital trust powers the sharing economy. IESE Bus Rev 26(5):24–31

McKnight DH (2005) Trust in information technology. Blackwell Encycl Manag 7:329–331

McKnight DH, Chervany NL (2001) What trust means in e-commerce customer relationships: an interdisciplinary conceptual typology. Int J Electron Commer 6(2):35–59

McKnight DH, Choudhury V, Kacmar C (2002) Developing and validating trust measures for e-commerce: an integrative typology. Inf Syst Res 13(3):334–359

Miller GA (1956) The magical number seven, plus or minus two: some limits on our capacity for processing information. Psychol Rev 63(2):81

Papadopouou P, Kanellis P, Martakos D (2001) Investigating trust in e-commerce: a literature review and a model for its formation in customer relationships. In: AMCIS 2001 proceedings, p 155

Pavlou PA (2002) Institution-based trust in interorganizational exchange relationships: the role of online B2B marketplaces on trust formation. J Strateg Inf Syst 11(3–4):215–243

Pavlou PA, Dimoka A (2006) The nature and role of feedback text comments in online marketplaces: implications for trust building, price premiums, and seller differentiation. Inf Syst Res 17(4):392–414

Pavlou PA, Fygenson M (2006) Understanding and predicting electronic commerce adoption: an extension of the theory of planned behavior. MIS Q 30:115–143

Pavlou PA, Gefen D (2004) Building effective online marketplaces with institution-based trust. Inf Syst Res 15(1):37–59

Pennington R, Wilcox HD, Grover V (2003) The role of system trust in business-to-consumer transactions. J Manag Inf Syst 20(3):197–226

Phillips N, Brown JL (1993) Analyzing communication in and around organizations: a critical hermeneutic approach. Acad Manag J 36(6):1547–1576

Qureshi I, Fang Y, Ramsey E, McCole P, Ibbotson P, Compeau D (2009) Understanding online customer repurchasing intention and the mediating role of trust–an empirical investigation in two developed countries. Eur J Inf Syst 18(3):205–222

Resnick P, Zeckhauser R (2002) Trust among strangers in internet transactions: empirical analysis of ebay’s reputation system. Econ Internet E-Commerce 11(2):23–25

Schryen G, Wagner G, Benlian A, Paré G (2020) A knowledge development perspective on literature reviews: validation of a new typology in the IS field. In: Communications of the AIS, p 46

Schultz CD (2006) A trust framework model for situational contexts. In: Proceedings of the 2006 international conference on privacy, security and trust: bridge the gap between PST technologies and business services, pp 1–7

Shankar V, Urban GL, Sultan F (2002) Online trust : a stakeholder perspective, concepts, implications, and future directions. J Strat Inf Syst 11:325–344

Shannon CE, Weaver W (1949) The mathematical theory of com-munication. University of Illinois Press, Urbana, p 96

Shapiro C (1983) Premiums for high quality products as returns to reputations. Q J Econ 98(4):659–679

Shareef MA, Dwivedi YK, Kumar V, Davies G, Rana N, Baabdullah A (2018) Purchase intention in an electronic commerce environment. Inf Technol People

Shiau WL, Chau PYK (2015) Does altruism matter on online group buying? Perspectives from egotistic and altruistic motivation. Inf Technol People 28(3):677–698

Shukla P (2014) The impact of organizational efforts on consumer concerns in an online context. Inf Manag 51(1):113–119

Sia C, Lim KH, Lee MKO, Benbasat I (2009) Web strategies to promote internet shopping: is cultural-customization needed? MIS Q 33(2009):491–512

Spence AM (1974) Market signaling: informational transfer in hiring and related screening processes. Harvard University Press, Cambridge, p 143

Sun H (2010) Sellers’ trust and continued use of online marketplaces. J Assoc Inf Syst 11(4):2

Sun H, Zhang P (2008) Online sellers trust and use of online auction marketplaces. In: PACIS 2008 proceedings, p 78

Tallon PP, Queiroz M, Coltman T, Sharma R (2019) Information technology and the search for organizational agility: a systematic review with future research possibilities. J Strat Inf Syst 28(2):218–237

Tomlinson EC, Schnackenberg AK, Dawley D, Ash SR (2020) Revisiting the trustworthiness–trust relationship: exploring the differential predictors of cognition-and affect-based trust. J Organ Behav 41(6):535–550

Toulmin SE (2003) The uses of argument. Cambridge University Press, Cambridge

Book   Google Scholar  

Tsatsou P, Elaluf-Calderwood S, Liebenau J (2010) Towards a taxonomy for regulatory issues in a digital business ecosystem in the EU. J Inf Technol 25(3):288–307

Venkatesh V, Davis FD (2000) A theoretical extension of the technology acceptance model: four longitudinal field studies. Manag Sci 46(2):186–204

Verdon J (2021) Global E-Commerce Sales To Hit $4.2 Trillion As Online Surge Continues, Adobe Reports . Forbes . https://www.forbes.com/sites/joanverdon/2021/04/27/global-ecommerce-sales-to-hit-42-trillion-as-online-surge-continues-adobe-reports/?sh=77033a450fd4

Verhagen T, Meents S, Tan YH (2006) Perceived risk and trust associated with purchasing at electronic marketplaces. Eur J Inf Syst 15(6):542–555

Wang W, Benbasat I (2008) Attributions of trust in decision support technologies: a study of recommendation agents for e-commerce. J Manag Inf Syst 24(4):249–273

Woisetschläger DM, Lentz P, Evanschitzky H (2011) How habits, social ties, and economic switching barriers affect customer loyalty in contractual service settings. J Bus Res 64(8):800–808

Download references

Acknowledgements

I wish to express my sincere appreciation to Dr. Alireza Amrollahi, for his comments on the early version of this paper.

Author information

Authors and affiliations.

Faculty of Management, University of Tehran, Tehran, Iran

Marzieh Soleimani

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Marzieh Soleimani .

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Reprints and permissions

About this article

Soleimani, M. Buyers' trust and mistrust in e-commerce platforms: a synthesizing literature review. Inf Syst E-Bus Manage 20 , 57–78 (2022). https://doi.org/10.1007/s10257-021-00545-0

Download citation

Received : 19 January 2021

Revised : 01 October 2021

Accepted : 21 October 2021

Published : 11 November 2021

Issue Date : March 2022

DOI : https://doi.org/10.1007/s10257-021-00545-0

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Electronic commerce platforms
• Online shopping
• Systematic literature review
• Find a journal
• Publish with us
• Track your research

Information

• Author Services

Initiatives

You are accessing a machine-readable page. In order to be human-readable, please install an RSS reader.

All articles published by MDPI are made immediately available worldwide under an open access license. No special permission is required to reuse all or part of the article published by MDPI, including figures and tables. For articles published under an open access Creative Common CC BY license, any part of the article may be reused without permission provided that the original article is clearly cited. For more information, please refer to https://www.mdpi.com/openaccess .

Feature papers represent the most advanced research with significant potential for high impact in the field. A Feature Paper should be a substantial original Article that involves several techniques or approaches, provides an outlook for future research directions and describes possible research applications.

Feature papers are submitted upon individual invitation or recommendation by the scientific editors and must receive positive feedback from the reviewers.

Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world. Editors select a small number of articles recently published in the journal that they believe will be particularly interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the most exciting work published in the various research areas of the journal.

Original Submission Date Received: .

• Active Journals
• Find a Journal
• Proceedings Series
• For Authors
• For Reviewers
• For Editors
• For Librarians
• For Publishers
• For Societies
• For Conference Organizers
• Open Access Policy
• Institutional Open Access Program
• Special Issues Guidelines
• Editorial Process
• Research and Publication Ethics
• Article Processing Charges
• Testimonials
• Preprints.org
• SciProfiles
• Encyclopedia

Article Menu

• Subscribe SciFeed
• Recommended Articles
• Google Scholar
• on Google Scholar
• Table of Contents

Find support for a specific problem in the support section of our website.

Please let us know what you think of our products and services.

Visit our dedicated information section to learn more about MDPI.

JSmol Viewer

Financial fraud detection based on machine learning: a systematic literature review.

1. Introduction

2. research methods, 2.1. review planning, 2.2. conducting the review, 2.2.1. research questions, 2.2.2. search strategy, 2.2.3. study selection criteria, 2.3. data extraction and synthesis, 3. search results and meta-analysis, 3.1. description of studies, 3.2. synthesis results, 3.2.1. rq1: what are the different categories of fraudulent activities that are addressed using ml techniques, credit card fraud, financial statement fraud, insurance fraud, financial cyber-fraud, other financial fraudulent types, 3.2.2. rq2: what are the ml-based techniques for financial fraud detection employed in the literature, support vector machine (svm), fuzzy-logic-based method, hidden markov model (hmm), artificial neural network (ann), knn algorithm, bayesian method, decision tree, genetic algorithm, ensemble methods, clustering based methods, logistic regression, 3.2.3. rq3: what are the performance evaluation metrics used for financial fraud detection using machine learning methods, 3.2.4. rq4 what are the gaps and future research direction in machine-learning-based fraud detection, imbalanced dataset, data size feature vectors, unstructured data, machine-learning-based technique, 4. discussion, 5. limitation and threat to validity.

• This SLR is only limited to conference and journal papers that discuss machine learning (ML) in the context of detecting financial fraud. By using our search approach in the early stages of the review, several non-relevant research papers were identified and excluded from this review. This ensures that the selected research papers satisfied the criteria for the study. However, it is believed that using more sources, such as additional source books, would have further enhanced this review.
• Although major databases were taken into consideration when exploring the research articles, there may be other digital libraries with relevant studies that were overlooked. We compared search terms and keywords to a well-known list of research studies to mitigate this limitation. However, some synonyms may be overlooked when searching for the keywords. The SLR protocol has been revised to address this problem by ensuring no essential terms are left out.
• We restricted our search to only English-language articles. Thus, this results in linguistic bias because some related papers in this field of study may exist in other languages. However, fortunately, all the gathered papers in this study were written in English. As such, we have no language bias.

6. Conclusions

Author contributions, institutional review board statement, informed consent statement, data availability statement, acknowledgments, conflicts of interest.

• Hilal, W.; Gadsden, S.A.; Yawney, J. Financial Fraud: A Review of Anomaly Detection Techniques and Recent Advances. Expert Syst. Appl. 2021 , 193 , 116429. [ Google Scholar ] [ CrossRef ]
• Ashtiani, M.N.; Raahemi, B. Intelligent Fraud Detection in Financial Statements Using Machine Learning and Data Mining: A Systematic Literature Review. IEEE Access 2021 , 10 , 72504–72525. [ Google Scholar ] [ CrossRef ]
• Albashrawi, M. Detecting Financial Fraud Using Data Mining Techniques: A Decade Review from 2004 to 2015. J. Data Sci. 2016 , 14 , 553–570. [ Google Scholar ] [ CrossRef ]
• Choi, D.; Lee, K. An Artificial Intelligence Approach to Financial Fraud Detection under IoT Environment: A Survey and Implementation. Secur. Commun. Netw. 2018 , 2018 , 1–15. [ Google Scholar ] [ CrossRef ]
• Ngai, E.W.T.; Hu, Y.; Wong, Y.H.; Chen, Y.; Sun, X. The application of data mining techniques in financial fraud detection: A classification framework and an academic review of literature. Decis. Support Syst. 2011 , 50 , 559–569. [ Google Scholar ] [ CrossRef ]
• Ryman-Tubb, N.F.; Krause, P.; Garn, W. How Artificial Intelligence and machine learning research impacts payment card fraud detection: A survey and industry benchmark. Eng. Appl. Artif. Intell. 2018 , 76 , 130–157. [ Google Scholar ] [ CrossRef ]
• Al-Hashedi, K.G.; Magalingam, P. Financial fraud detection applying data mining techniques: A comprehensive review from 2009 to 2019. Comput. Sci. Rev. 2021 , 40 , 100402. [ Google Scholar ] [ CrossRef ]
• Chaquet-ulldemolins, J.; Moral-rubio, S.; Muñoz-romero, S. On the Black-Box Challenge for Fraud Detection Using Machine Learning (II): Nonlinear Analysis through Interpretable Autoencoders. Appl. Sci. 2022 , 12 , 3856. [ Google Scholar ] [ CrossRef ]
• Da’U, A.; Salim, N. Recommendation system based on deep learning methods: A systematic review and new directions. Artif. Intell. Rev. 2019 , 53 , 2709–2748. [ Google Scholar ] [ CrossRef ]
• Zeng, Y.; Tang, J. RLC-GNN: An Improved Deep Architecture for Spatial-Based Graph Neural Network with Application to Fraud Detection. Appl. Sci. 2021 , 11 , 5656. [ Google Scholar ] [ CrossRef ]
• Delamaire, L.; Hussein, A.; John, P. Credit card fraud and detection techniques: A review. Banks Bank Syst. 2009 , 4 , 57–68. [ Google Scholar ]
• Zhang, D.; Zhou, L. Discovering Golden Nuggets: Data Mining in Financial Application. IEEE Trans. Syst. Man Cybern. Part C Appl. Rev. 2004 , 34 , 513–522. [ Google Scholar ] [ CrossRef ]
• Raj, S.B.E.; Portia, A.A. Analysis on credit card fraud detection methods. In Proceedings of the 2011 International Conference on Computer, Communication and Electrical Technology (ICCCET), Tirunelveli, India, 18–19 March 2011; pp. 152–156. [ Google Scholar ] [ CrossRef ]
• Phua, C.; Lee, V.; Smith, K.; Gayler, R. A Comprehensive Survey of Data Mining-based Fraud Detection Research. arXiv 2010 , arXiv:1009.6119. [ Google Scholar ]
• West, J.; Bhattacharya, M. Intelligent financial fraud detection: A comprehensive review. Comput. Secur. 2016 , 57 , 47–66. [ Google Scholar ] [ CrossRef ]
• Abdallah, A.; Maarof, M.A.; Zainal, A. Fraud detection system: A survey. J. Netw. Comput. Appl. 2016 , 68 , 90–113. [ Google Scholar ] [ CrossRef ]
• Popat, R.R.; Chaudhary, J. A Survey on Credit Card Fraud Detection Using Machine Learning. In Proceedings of the 2018 2nd International Conference on Trends in Electronics and Informatics (ICOEI), Tirunelveli, India, 11–12 May 2018; pp. 1120–1125. [ Google Scholar ]
• Gyamfi, N.K.; Abdulai, J. Bank Fraud Detection Using Support Vector Machine. In Proceedings of the 2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), Vancouver, BC, Canada, 1–3 November 2018; pp. 37–41. [ Google Scholar ]
• Carneiro, E.M.; Dias, L.A.V.; Da Cunha, A.M.; Mialaret, L.F.S. Cluster Analysis and Artificial Neural Networks: A Case Study in Credit Card Fraud Detection. In Proceedings of the 2015 12th International Conference on Information Technology-New Generations, Mumbai, India, 11–14 December 2011; pp. 122–126. [ Google Scholar ] [ CrossRef ]
• Iyer, D.; Mohanpurkar, A.; Janardhan, S.; Rathod, D.; Sardeshmukh, A. Credit card fraud detection using Hidden Markov Model. In Proceedings of the 2011 World Congress on Information and Communication Technologies, Mumbai, India, 11–14 December 2011; pp. 1062–1066. [ Google Scholar ]
• Patil, S.; Nemade, V.; Soni, P. ScienceDirect Predictive Modelling For Credit Card Fraud Detection Using Data Analytics. Procedia Comput. Sci. 2018 , 132 , 385–395. [ Google Scholar ] [ CrossRef ]
• Mohammadian, V.; Navimipour, N.J.; Hosseinzadeh, M.; Darwesh, A. Comprehensive and systematic study on the fault tolerance architectures in cloud computing. J. Circuits Syst. Comput. 2020 , 29 , 2050240. [ Google Scholar ] [ CrossRef ]
• Kitchenham, B.; Charters, S. Guidelines for Performing Systematic Literature Reviews in Software Engineering ; Keele University: Keele, UK, 2007; p. 65. [ Google Scholar ]
• Pourhabibi, T.; Ong, K.-L.; Kam, B.H.; Boo, Y.L. Fraud detection: A systematic literature review of graph-based anomaly detection approaches. Decis. Support Syst. 2020 , 133 , 113303. [ Google Scholar ] [ CrossRef ]
• Marcotte, P.; Petrillo, F. Multiple Fault-tolerance Mechanisms in Cloud Systems: A Systematic Review. In Proceedings of the 2019 IEEE International Symposium on Software Reliability Engineering Workshops (ISSREW), Berlin, Germany, 28–31 October 2019; pp. 414–421. [ Google Scholar ]
• Isong, B.E.; Bekele, E. A systematic review of fault tolerance in mobile agents. Eng. Appl. 2013 , 2 , 111–124. [ Google Scholar ] [ CrossRef ]
• Nassif, A.B.; Abu Talib, M.; Nasir, Q.; Dakalbab, F.M. Machine Learning for Anomaly Detection: A Systematic Review. IEEE Access 2021 , 9 , 78658–78700. [ Google Scholar ] [ CrossRef ]
• Randhawa, K.; Loo, C.K.; Seera, M.; Lim, C.P.; Nandi, A.K. Credit Card Fraud Detection Using AdaBoost and Majority Voting. IEEE Access 2018 , 6 , 14277–14284. [ Google Scholar ] [ CrossRef ]
• Bhattacharyya, S.; Jha, S.; Tharakunnel, K.; Westland, J.C. Data mining for credit card fraud: A comparative study. Decis. Support Syst. 2011 , 50 , 602–613. [ Google Scholar ] [ CrossRef ]
• Srivastava, A.; Yadav, M.; Basu, S.; Salunkhe, S.; Shabad, M. Credit card fraud detection at merchant side using neural networks. In Proceedings of the 2016 3rd International Conference on Computing for Sustainable Global Development (INDIACom), New Delhi, India, 16–18 March 2016; pp. 667–670. [ Google Scholar ]
• de Sá, A.G.; Pereira, A.C.; Pappa, G.L. A customized classification algorithm for credit card fraud detection. Eng. Appl. Artif. Intell. 2018 , 72 , 21–29. [ Google Scholar ] [ CrossRef ]
• Robinson, W.N.; Aria, A. Sequential fraud detection for prepaid cards using hidden Markov model divergence. Expert Syst. Appl. 2018 , 91 , 235–251. [ Google Scholar ] [ CrossRef ]
• Hajek, P.; Henriques, R. Mining corporate annual reports for intelligent detection of financial statement fraud—A comparative study of machine learning methods. Knowl.-Based Syst. 2017 , 128 , 139–152. [ Google Scholar ] [ CrossRef ]
• Craja, P.; Kim, A.; Lessmann, S. Deep learning for detecting financial statement fraud. Decis. Support Syst. 2020 , 139 , 113421. [ Google Scholar ] [ CrossRef ]
• Ravisankar, P.; Ravi, V.; Rao, G.R.; Bose, I. Detection of fi nancial statement fraud and feature selection using data mining techniques. Decis. Support Syst. 2011 , 50 , 491–500. [ Google Scholar ] [ CrossRef ]
• Gao, Y.; Sun, C.; Li, R.; Li, Q.; Cui, L.; Gong, B. An Efficient Fraud Identification Method Combining Manifold Learning and Outliers Detection in Mobile Healthcare Services. IEEE Access 2018 , 6 , 60059–60068. [ Google Scholar ] [ CrossRef ]
• Huang, S.-Y.; Tsaih, R.-H.; Yu, F. Topological pattern discovery and feature extraction for fraudulent financial reporting. Expert Syst. Appl. 2014 , 41 , 4360–4372. [ Google Scholar ] [ CrossRef ]
• Peng, J.; Li, Q.; Li, H.; Liu, L.; Yan, Z.; Zhang, S. Fraud Detection of Medical Insurance Employing Outlier Analysis. In Proceedings of the 2018 IEEE 22nd International Conference on Computer Supported Cooperative Work in Design (CSCWD), Nanjing, China, 9–11 May 2018; pp. 341–346. [ Google Scholar ]
• van Capelleveen, G.; Poel, M.; Mueller, R.M.; Thornton, D.; van Hillegersberg, J. Outlier detection in healthcare fraud: A case study in the Medicaid dental domain. Int. J. Account. Inf. Syst. 2016 , 21 , 18–31. [ Google Scholar ] [ CrossRef ]
• Anbarasi, M.S.; Dhivya, S. Fraud detection using outlier predictor in health insurance data. In Proceedings of the 2017 International Conference on Information Communication and Embedded Systems (ICICES), Chennai, India, 23–24 February 2017; pp. 1–6. [ Google Scholar ]
• Sundarkumar, G.G.; Ravi, V.; Siddeshwar, V. One-class support vector machine based undersampling: Application to churn prediction and insurance fraud detection. In Proceedings of the 2015 IEEE International Conference on Computational Intelligence and Computing Research (ICCIC), Madurai, India, 10–12 December 2015; pp. 1–7. [ Google Scholar ] [ CrossRef ]
• Subudhi, S.; Panigrahi, S. Effect of Class Imbalanceness in Detecting Automobile Insurance Fraud. In Proceedings of the 2018 2nd International Conference on Data Science and Business Analytics (ICDSBA), ChangSha, China, 21–23 September 2018; pp. 528–531. [ Google Scholar ]
• Fayyomi, M.; Eleyan, D.; Eleyan, A. A Survey Paper On Credit Card Fraud Detection Techniques. Int. J. Adv. Res. Comput. Eng. Technol. 2021 , 3 , 827–832. [ Google Scholar ]
• Wang, Y.; Xu, W. Leveraging deep learning with LDA-based text analytics to detect automobile insurance fraud. Decis. Support Syst. 2018 , 105 , 87–95. [ Google Scholar ] [ CrossRef ]
• Gepp, A.; Kumar, K.; Bhattacharya, S. Lifting the numbers game: Identifying key input variables and a best-performing model to detect financial statement fraud. Account. Financ. 2021 , 61 , 4601–4638. [ Google Scholar ] [ CrossRef ]
• Perols, L.; Lougee, B.A. The relation between earnings management and financial statement fraud. Adv. Account. 2011 , 27 , 39–53. [ Google Scholar ] [ CrossRef ]
• Wang, Q.; Xu, W.; Huang, X.; Yang, K. Enhancing intraday stock price manipulation detection by leveraging recurrent neural networks with ensemble learning. Neurocomputing 2019 , 347 , 46–58. [ Google Scholar ] [ CrossRef ]
• Islam, S.R.; Ghafoor, S.K.; Eberle, W. Mining Illegal Insider Trading of Stocks: A Proactive Approach. In Proceedings of the 2018 IEEE International Conference on Big Data (Big Data), Seattle, WA, USA, 10–13 December 2018; pp. 1397–1406. [ Google Scholar ] [ CrossRef ]
• Kulkarni, P.M.; Domeniconi, C. Network-based anomaly detection for insider trading. arXiv 2017 , arXiv:1702.05809. [ Google Scholar ]
• Mirtaheri, M.; Abu-El-Haija, S.; Morstatter, F.; Steeg, G.V.; Galstyan, A. Identifying and Analyzing Cryptocurrency Manipulations in Social Media. IEEE Trans. Comput. Soc. Syst. 2021 , 8 , 607–617. [ Google Scholar ] [ CrossRef ]
• Monamo, P.M.; Marivate, V.; Twala, B. A Multifaceted Approach to Bitcoin Fraud Detection: Global and Local Outliers. In Proceedings of the 2016 15th IEEE International Conference on Machine Learning and Applications (ICMLA), Anaheim, CA, USA, 18–20 December 2016; pp. 188–194. [ Google Scholar ] [ CrossRef ]
• Vasek, M.; Moore, T. There’s No Free Lunch, Even Using Bitcoin: Tracking the Popularity and Profits of Virtual Currency Scams BT–Financial Cryptography and Data Security. In Proceedings of the International Conference on Financial Cryptography and Data Security, Kota Kinabalu, Malaysia, 1–5 March 2015; pp. 44–61. [ Google Scholar ]
• Monamo, P.; Marivate, V.; Twala, B. Unsupervised learning for robust Bitcoin fraud detection. In Proceedings of the 2016 Information Security for South Africa (ISSA), Johannesburg, South Africa, 17–18 August 2016; pp. 129–134. [ Google Scholar ] [ CrossRef ]
• Li, X.; Ying, S. Lib-SVMs Detection Model of Regulating-Profits Financial Statement Fraud Using Data of Chinese Listed Companies. In Proceedings of the 2010 International Conference on E-Product E-Service and E-Entertainment, Henan, China, 7–9 November 2010; pp. 1–4. [ Google Scholar ] [ CrossRef ]
• Throckmorton, C.S.; Mayew, W.J.; Venkatachalam, M.; Collins, L.M. Financial fraud detection using vocal, linguistic and fi nancial cues. Decis. Support Syst. 2015 , 74 , 78–87. [ Google Scholar ] [ CrossRef ]
• Glancy, F.H.; Yadav, S.B. A computational model for fi nancial reporting fraud detection. Decis. Support Syst. 2011 , 50 , 595–601. [ Google Scholar ] [ CrossRef ]
• Mareeswari, V.; Gunasekaran, G. Prevention of credit card fraud detection based on HSVM. In Proceedings of the 2016 International Conference on Information Communication and Embedded Systems (ICICES), Chennai, India, 25–26 February 2016; pp. 1–4. [ Google Scholar ]
• Humpherys, S.L.; Mof, K.C.; Burns, M.B.; Burgoon, J.K.; Felix, W.F. Identi fi cation of fraudulent fi nancial statements using linguistic credibility analysis. Decis. Support Syst. 2011 , 50 , 585–594. [ Google Scholar ] [ CrossRef ]
• Li, X.; Xu, W.; Tian, X. How to protect investors? A GA-based DWD approach for financial statement fraud detection. In Proceedings of the 2014 IEEE International Conference on Systems, Man, and Cybernetics (SMC), San Diego, CA, USA, 5–8 October 2014; pp. 3548–3554. [ Google Scholar ] [ CrossRef ]
• Karlos, S.; Fazakis, N.; Kotsiantis, S.; Sgarbas, K. Semi-supervised forecasting of fraudulent financial statements. In Proceedings of the 20th Pan-Hellenic Conference on Informatics, Patras, Greece, 10–12 November 2016. [ Google Scholar ] [ CrossRef ]
• Özçelik, M.H.; Duman, E.; Işik, M.; Çevik, T. Improving a credit card fraud detection system using genetic algorithm. In Proceedings of the 2010 International Conference on Networking and Information Technology, Manila, Philippines, 11–12 June 2010; pp. 436–440. [ Google Scholar ]
• Rizki, A.; Surjandari, I.; Wayasti, R.A. Data mining application to detect financial fraud in Indonesia’s public companies. In Proceedings of the 2017 3rd International Conference on Science in Information Technology (ICSITech), Bandung, Indonesia, 25–26 October 2017; pp. 206–211. [ Google Scholar ]
• Chen, S. Detection of fraudulent financial statements using the hybrid data mining approach. SpringerPlus 2016 , 5 , 1–16. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Yao, J.; Zhang, J.; Wang, L. A financial statement fraud detection model based on hybrid data mining methods. In Proceedings of the 2018 international conference on artificial intelligence and big data (ICAIBD), Chengdu, China, 26–28 May 2018; pp. 57–61. [ Google Scholar ] [ CrossRef ]
• Rajak, I.; Mathai, K.J. Intelligent fraudulent detection system based SVM and optimized by danger theory. In Proceedings of the 2015 International Conference on Computer, Communication and Control (IC4), Indore, India, 10–12 September 2015; pp. 1–4. [ Google Scholar ] [ CrossRef ]
• Jeragh, M.; Alsulaimi, M. Combining Auto Encoders and One Class Support Vectors Machine for Fraudulant Credit Card Transactions Detection. In Proceedings of the 2018 Second World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4), London, UK, 30–31 October 2018; pp. 178–184. [ Google Scholar ] [ CrossRef ]
• Kho, J.R.D.; Vea, L.A. Credit card fraud detection based on transaction behavior. In Proceedings of the TENCON 2017-2017 IEEE Region 10 Conference, Penang, Malaysia, 5–8 November 2017; pp. 1880–1884. [ Google Scholar ] [ CrossRef ]
• Behera, T.K.; Panigrahi, S. Credit Card Fraud Detection: A Hybrid Approach Using Fuzzy Clustering & Neural Network. In Proceedings of the 2015 Second International Conference on Advances in Computing and Communication Engineering, Dehradun, India, 1–2 May 2015; pp. 494–499. [ Google Scholar ]
• HaratiNik, M.R.; Akrami, M.; Khadivi, S.; Shajari, M. FUZZGY: A hybrid model for credit card fraud detection. In Proceedings of the 6th International Symposium on Telecommunications (IST), Tehran, Iran, 6–8 November 2012; pp. 1088–1093. [ Google Scholar ]
• Malini, N.; Pushpa, M. Analysis on credit card fraud identification techniques based on KNN and outlier detection. In Proceedings of the 2017 third international conference on advances in electrical, electronics, information, communication and bio-informatics (AEEICB), Chennai, India, 27–28 February 2017; pp. 255–258. [ Google Scholar ] [ CrossRef ]
• Benchaji, I.; Douzi, S.; ElOuahidi, B. Using Genetic Algorithm to Improve Classification of Imbalanced Datasets for Credit Card Fraud Detection. In Proceedings of the International Conference on Advanced Information Technology, Services and Systems, Mohammedia, Morocco, 17–18 October 2018; pp. 1–5. [ Google Scholar ] [ CrossRef ]
• Case, B. Recognizing Debit Card Fraud Transaction Using CHAID and K-Nearest Neighbor: Indonesian Bank case. In Proceedings of the 2016 11th International Conference on Knowledge, Information and Creativity Support Systems (KICSS), Yogyakarta, Indonesia, 10–12 November 2016. [ Google Scholar ]
• Bhusari, V.; Patil, S. Study of Hidden Markov Model in credit card fraudulent detection. In Proceedings of the 2016 World Conference on Futuristic Trends in Research and Innovation for Social Welfare (Startup Conclave), Coimbatore, India, 29 February–1 March 2016; pp. 1–4. [ Google Scholar ]
• Sahin, Y.; Bulkan, S.; Duman, E. A cost-sensitive decision tree approach for fraud detection. Expert Syst. Appl. 2013 , 40 , 5916–5923. [ Google Scholar ] [ CrossRef ]
• Duman, E.; Ozcelik, M.H. Detecting credit card fraud by genetic algorithm and scatter search. Expert Syst. Appl. 2011 , 38 , 13057–13063. [ Google Scholar ] [ CrossRef ]
• Sahin, Y.; Duman, E. Detecting credit card fraud by ANN and logistic regression. In Proceedings of the 2011 International Symposium on Innovations in Intelligent Systems and Applications, Istanbul, Turkey, 15–18 June 2011; pp. 315–319. [ Google Scholar ]
• Ghobadi, F.; Rohani, M. Cost sensitive modeling of credit card fraud using neural network strategy. In Proceedings of the 2016 2nd International Conference of Signal Processing and Intelligent Systems (ICSPIS), Tehran, Iran, 14–15 December 2016; pp. 1–5. [ Google Scholar ]
• Awoyemi, J.O.; Adetunmbi, A.O.; Oluwadare, S.A. Credit card fraud detection using machine learning techniques: A comparative analysis. In Proceedings of the 2017 international conference on computing networking and informatics (ICCNI), Ota, Nigeria, 29–31 October 2017; pp. 1–9. [ Google Scholar ] [ CrossRef ]
• Mishra, A.; Ghorpade, C. Credit Card Fraud Detection on the Skewed Data Using Various Classification and Ensemble Techniques. In Proceedings of the 2018 IEEE International Students’ Conference on Electrical, Electronics and Computer Science (SCEECS), Bhopal, India, 24–25 February 2018; pp. 1–5. [ Google Scholar ] [ CrossRef ]
• Kirlidog, M.; Asuk, C. A Fraud Detection Approach with Data Mining in Health Insurance. Procedia-Soc. Behav. Sci. 2012 , 62 , 989–994. [ Google Scholar ] [ CrossRef ]
• Peng, H.; You, M. The Health Care Fraud Detection Using the Pharmacopoeia Spectrum Tree and Neural Network Analytic Contribution Hierarchy Process. In Proceedings of the 2016 IEEE Trustcom/BigDataSE/ISPA, Tianjin, China, 23–26 August 2016; pp. 2006–2011. [ Google Scholar ] [ CrossRef ]
• Bauder, R.; da Rosa, R.; Khoshgoftaar, T. Identifying Medicare Provider Fraud with Unsupervised Machine Learning. In Proceedings of the 2018 IEEE International Conference on Information Reuse and Integration (IRI), Salt Lake City, UT, USA, 7–9 July 2018; pp. 285–292. [ Google Scholar ]
• Bauder, R.A.; Khoshgoftaar, T.M.; Richter, A.; Herland, M. Predicting Medical Provider Specialties to Detect Anomalous Insurance Claims. In Proceedings of the 2016 IEEE 28th International Conference on Tools with Artificial Intelligence (ICTAI), San Jose, CA, USA, 6–8 November 2016; pp. 784–790. [ Google Scholar ]
• Badriyah, T.; Rahmaniah, L.; Syarif, I. Nearest Neighbour and Statistics Method based for Detecting Fraud in Auto Insurance. In Proceedings of the 2018 International Conference on Applied Engineering (ICAE), Batam, Indonesia, 3–4 October 2018; pp. 1–5. [ Google Scholar ] [ CrossRef ]
• Zhou, Y.; Wang, X.; Zhang, J.; Zhang, P.; Liu, L.; Jin, H.; Jin, H. Analyzing and Detecting Money-Laundering Accounts in Online Social Networks. IEEE Netw. 2017 , 32 , 115–121. [ Google Scholar ] [ CrossRef ]
• Mhamane, S.S.; Lobo, L.M.R.J. Internet banking fraud detection using HMM. In Proceedings of the 2012 Third International Conference on Computing, Communication and Networking Technologies (ICCCNT’12), Karur, India, 26–28 July 2012; pp. 1–4. [ Google Scholar ]
• Faraji, Z.; States, U. A Review of Machine Learning Applications for Credit Card Fraud Detection with A Case study. J. Manag. 2022 , 5 , 49–59. [ Google Scholar ] [ CrossRef ]
• Bhavitha, B.K.; Rodrigues, A.P.; Chiplunkar, N.N. Comparative study of machine learning techniques in sentimental analysis. In Proceedings of the 2017 International Conference on Inventive Communication and Computational Technologies (ICICCT), Coimbatore, India, 10–11 March 2017; pp. 216–221. [ Google Scholar ] [ CrossRef ]
• Carta, S.; Fenu, G.; Recupero, D.R.; Saia, R. Fraud detection for E-commerce transactions by employing a prudential Multiple Consensus model. J. Inf. Secur. Appl. 2019 , 46 , 13–22. [ Google Scholar ] [ CrossRef ]
• Rb, A.; Kr, S.K. Credit card fraud detection using artificial neural network. Glob. Transit. Proc. 2021 , 2 , 35–41. [ Google Scholar ] [ CrossRef ]
• Pradeep, G.; Ravi, V.; Nandan, K.; Deekshatulu, B.L.; Bose, I.; Aditya, A. Fraud Detection in Financial Statements Using Evolutionary Computation Based Rule Miners. In Proceedings of the International Conference on Swarm, Evolutionary, and Memetic Computing, Hyderabad, India, 18–19 December 2015; pp. 239–250. [ Google Scholar ] [ CrossRef ]
• Hajek, P. Interpretable Fuzzy Rule-Based Systems for Detecting Financial Statement Fraud. In Proceedings of the IFIP International Conference on Artificial Intelligence Applications and Innovations, Crete, Greece, 25–27 June 2019; pp. 1–12. [ Google Scholar ]
• Khan, A.; Singh, T.; Sinhal, A.; Khan, A.; Singh, T. Implement credit card fraudulent detection system using observation probabilistic in hidden Markov model. In Proceedings of the 2012 Nirma University International Conference on Engineering (NUiCONE), Ahmedabad, India, 6–8 December 2012; pp. 1–6. [ Google Scholar ] [ CrossRef ]
• Wang, X.; Wu, H.; Yi, Z. Research on Bank Anti-Fraud Model Based on K-Means and Hidden Markov Model. In Proceedings of the 2018 IEEE 3rd International Conference on Image, Vision and Computing (ICIVC), Chongqing, China, 27–29 June 2018; pp. 780–784. [ Google Scholar ] [ CrossRef ]
• Song, R.; Huang, L.; Cui, W.; Vanthienen, J. Fraud Detection of Bulk Cargo Theft in Port Using Bayesian Network Models. Appl. Sci. 2020 , 10 , 1056. [ Google Scholar ] [ CrossRef ]
• Dang, T.K.; Tran, T.C.; Tuan, L.M. Machine Learning Based on Resampling Approaches and Deep Reinforcement Learning for Credit Card Fraud Detection Systems. Appl. Sci. 2021 , 11 , 10004. [ Google Scholar ] [ CrossRef ]
• Bouchti, E.; Chakroun, A.; Abbar, H.; Okar, C. Fraud detection in banking using deep reinforcement learning. In Proceedings of the 2017 Seventh International Conference on Innovative Computing Technology (INTECH), Luton, UK, 16–18 August 2017; pp. 58–63. [ Google Scholar ]
• Zouboulidis, E.; Kotsiantis, S. Forecasting fraudulent financial statements with committee of cost-sensitive decision tree classifiers. In Hellenic Conference on Artificial Intelligence ; Springer: Berlin/Heidelberg, Germany, 2012; pp. 57–64. [ Google Scholar ]
• Hassanzadeh, R. A Nomaly Detection in Online Social Networks: Using Data-Mining Techniques and Fuzzy. Ph.D. Thesis, Queensland University of Technology, Brisbane City, QLD, Australia, 2014. [ Google Scholar ]
• Shah, V.; Shah, P.; Shetty, H.; Mistry, K. Review of Credit Card Fraud Detection Techniques. In Proceedings of the 2019 IEEE International Conference on System, Computation, Automation and Networking (ICSCAN), Pondicherry, India, 29–30 March 2019; pp. 1–7. [ Google Scholar ] [ CrossRef ]
• Deng, Q. Detection of fraudulent financial statements based on Naïve Bayes classifier. In Proceedings of the 2010 5th International Conference on Computer Science & Education, Hefei, China, 24–27 August 2010; pp. 1032–1035. [ Google Scholar ] [ CrossRef ]
• Ahmed, M.; Mahmood, A.N.; Islam, R. A survey of anomaly detection techniques in financial domain. Futur. Gener. Comput. Syst. 2016 , 55 , 278–288. [ Google Scholar ] [ CrossRef ]
• Uchhana, N.; Ranjan, R.; Sharma, S.; Agrawal, D.; Punde, A. Literature Review of Different Machine Learning Algorithms for Credit Card Fraud Detection. Int. J. Innov. Technol. Explor. Eng. 2021 , 10 , 101–108. [ Google Scholar ] [ CrossRef ]
• Abbasi, A.; Albrecht, C.; Vance, A.; Hansen, J. Metafraud: A meta-learning framework for detecting financial fraud. Mis Q. 2012 , 36 , 1293–1327. [ Google Scholar ] [ CrossRef ]
• Moepya, S.O.; Nelwamondo, F.V.; Twala, B. Increasing the detection of minority class instances in financial statement fraud. In Proceedings of the Asian Conference on Intelligent Information and Database Systems, Kanazawa, Japan, 3–5 April 2017; Volume 2, p. 2017. [ Google Scholar ]
• Chen, S.; Goo, Y.-J.J.; Shen, Z.-D. A Hybrid Approach of Stepwise Regression, Logistic Regression, Support Vector Machine, and Decision Tree for Forecasting Fraudulent Financial Statements. Sci. World J. 2014 , 2014 , 1–9. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Patel, H.; Parikh, S.; Patel, A.; Parikh, A. An Application of Ensemble Random Forest Classifier for Detecting Financial Statement Manipulation of Indian Listed Companies. In Recent Developments in Machine Learning and Data Analytics ; Springer: Singapore, 2019. [ Google Scholar ]
• Hobson, L.; Mayew, W.J. Analyzing Speech to Detect Financial Misreporting Analyzing Speech to Detect Financial Misreporting. J. Account. Res. 2010 , 2 , 349–392. [ Google Scholar ]
• Li, Y.; Yan, C.; Liu, W.; Li, M. Research and application of random forest model in mining automobile insurance fraud. In Proceedings of the 2016 12th International Conference on Natural Computation, Fuzzy Systems and Knowledge Discovery (ICNC-FSKD), Changsha, China, 13–15 August 2016; pp. 1756–1761. [ Google Scholar ] [ CrossRef ]
• Kowshalya, G.; Nandhini, M. Predicting Fraudulent Claims in Automobile Insurance. In Proceedings of the 2018 Second International Conference on Inventive Communication and Computational Technologies (ICICCT), Coimbatore, India, 20–21 April 2018; pp. 1338–1343. [ Google Scholar ]
• Bauder, R.; Khoshgoftaar, T. Medicare Fraud Detection Using Random Forest with Class Imbalanced Big Data. In Proceedings of the 2018 IEEE International Conference on Information Reuse and Integration (IRI), Salt Lake City, UT, USA, 7–9 July 2018; pp. 80–87. [ Google Scholar ]
• Li, S.-H.; Yen, D.C.; Lu, W.-H.; Wang, C. Identifying the signs of fraudulent accounts using data mining techniques. Comput. Hum. Behav. 2012 , 28 , 1002–1013. [ Google Scholar ] [ CrossRef ]
• Bartoletti, M.; Pes, B.; Serusi, S. Data Mining for Detecting Bitcoin Ponzi Schemes. In Proceedings of the 2018 Crypto Valley Conference on Blockchain Technology (CVCBT), Zug, Switzerland, 20–22 June 2018; pp. 75–84. [ Google Scholar ] [ CrossRef ]
• Zhang, W.; He, X. An Anomaly Detection Method for Medicare Fraud Detection. In Proceedings of the 2017 IEEE International Conference on Big Knowledge (ICBK), Hefei, China, 9–10 August 2017; pp. 309–314. [ Google Scholar ]
• Deng, Q.; Mei, G. Combining self-organizing map and K-means clustering for detecting fraudulent financial statements. In Proceedings of the 2009 IEEE International Conference on Granular Computing, Nanchang, China, 17–19 August 2009; pp. 126–131. [ Google Scholar ] [ CrossRef ]
• Sael, N.; Benabbou, F. ScienceDirect ScienceDirect Performance of machine learning techniques in the detection of Performance of machine learning techniques in the detection of financial frauds financial frauds. Procedia Comput. Sci. 2018 , 148 , 45–54. [ Google Scholar ]
• Liang, J.; Lv, W. Research on detecting technique of financial statement fraud based on Fuzzy Genetic Algorithms BPN. In Proceedings of the 2009 International Conference on Management Science and Engineering, Nanchang, China, 17–19 August 2009; pp. 1462–1468. [ Google Scholar ] [ CrossRef ]
• Xiaoyun, W.; Danyue, L. Hybrid outlier mining algorithm based evaluation of client moral risk in insurance company. In Proceedings of the 2010 2nd IEEE International Conference on Information Management and Engineering, Chongqing, China, 17–19 September 2010; pp. 585–589. [ Google Scholar ] [ CrossRef ]
• Bauder, R.A.; Khoshgoftaar, T.M. Medicare Fraud Detection Using Machine Learning Methods. In Proceedings of the 2017 16th IEEE International Conference on Machine Learning and Applications (ICMLA), Cancun, Mexico, 18–21 December 2017; pp. 858–865. [ Google Scholar ]
• Pejic-bach, M. Invited Paper: Profiling Intelligent Systems Applications in Fraud Detection and Prevention: Survey of Research Articles Profiling intelligent systems applications in fraud detection and prevention: Survey of research articles. In Proceedings of the 2010 International Conference on Intelligent Systems, Modelling and Simulation, Liverpool, UK, 27–29 January 2010. [ Google Scholar ]
• Chawla, N.V.; Bowyer, K.W.; Hall, L.O.; Kegelmeyer, W.P. SMOTE: Synthetic Minority Over-sampling Technique. J. Artif. Intell. Res. 2002 , 16 , 321–357. [ Google Scholar ] [ CrossRef ]
• D’Addio, R.M.; Manzato, M.G. A Collaborative Filtering Approach Based on User’s Reviews. In Proceedings of the 2014 Brazilian Conference on Intelligent Systems, Washington, DC, USA, 18–22 October 2014; pp. 204–209. [ Google Scholar ] [ CrossRef ]
• Paruchuri, H. Credit Card Fraud Detection using Machine Learning: A Systematic Literature Review. ABC J. Adv. Res. 2017 , 6 , 113–120. [ Google Scholar ] [ CrossRef ]
• Silva, B.; Marques, N.; Panosso, G. Applying neural networks for concept drift detection in financial markets. CEUR Workshop Proc. 2012 , 960 , 43–47. [ Google Scholar ]

Click here to enlarge figure

Share and Cite

Ali, A.; Abd Razak, S.; Othman, S.H.; Eisa, T.A.E.; Al-Dhaqm, A.; Nasser, M.; Elhassan, T.; Elshafie, H.; Saif, A. Financial Fraud Detection Based on Machine Learning: A Systematic Literature Review. Appl. Sci. 2022 , 12 , 9637. https://doi.org/10.3390/app12199637

Ali A, Abd Razak S, Othman SH, Eisa TAE, Al-Dhaqm A, Nasser M, Elhassan T, Elshafie H, Saif A. Financial Fraud Detection Based on Machine Learning: A Systematic Literature Review. Applied Sciences . 2022; 12(19):9637. https://doi.org/10.3390/app12199637

Ali, Abdulalem, Shukor Abd Razak, Siti Hajar Othman, Taiseer Abdalla Elfadil Eisa, Arafat Al-Dhaqm, Maged Nasser, Tusneem Elhassan, Hashim Elshafie, and Abdu Saif. 2022. "Financial Fraud Detection Based on Machine Learning: A Systematic Literature Review" Applied Sciences 12, no. 19: 9637. https://doi.org/10.3390/app12199637

Article Metrics

Article access statistics, further information, mdpi initiatives, follow mdpi.

Subscribe to receive issue release notifications and newsletters from MDPI journals

• Open access
• Published: 13 March 2023

Online payment fraud: from anomaly detection to risk management

• Paolo Vanini   ORCID: orcid.org/0000-0003-0391-4847 1 ,
• Sebastiano Rossi 2 ,
• Ermin Zvizdic 3 &
• Thomas Domenig 4  

Financial Innovation volume  9 , Article number:  66 ( 2023 ) Cite this article

13k Accesses

20 Citations

Metrics details

Online banking fraud occurs whenever a criminal can seize accounts and transfer funds from an individual’s online bank account. Successfully preventing this requires the detection of as many fraudsters as possible, without producing too many false alarms. This is a challenge for machine learning owing to the extremely imbalanced data and complexity of fraud. In addition, classical machine learning methods must be extended, minimizing expected financial losses. Finally, fraud can only be combated systematically and economically if the risks and costs in payment channels are known. We define three models that overcome these challenges: machine learning-based fraud detection, economic optimization of machine learning results, and a risk model to predict the risk of fraud while considering countermeasures. The models were tested utilizing real data. Our machine learning model alone reduces the expected and unexpected losses in the three aggregated payment channels by 15% compared to a benchmark consisting of static if-then rules. Optimizing the machine-learning model further reduces the expected losses by 52%. These results hold with a low false positive rate of 0.4%. Thus, the risk framework of the three models is viable from a business and risk perspective.

Introduction

Fraud arises in the financial industry via numerous channels, such as credit cards, e-commerce, phone banking, checks, and online banking. Juniper Research ( 2020 ) reports that e-commerce, airline ticketing, money transfer, and banking services will cumulatively lose over $ 200 billion due to online payment fraud between 2020 and 2024. The increased sophistication of fraud attempts and the increasing number of attack vectors have driven these results. We focus on online and mobile payment channels and identity theft fraud (i.e., stealing an individual’s personal information to conduct fraud )(Amiri and Hekmat 2021 ). The aim is to identify external fraudsters who intend to initiate payments in their interests. As fraudsters gain access to the payment systems as if they were the owners of the accounts, they cannot be identified based on the account access process. However, the fraudster behaves differently during a payment transaction than the account owner and/or the payment has unusual characteristics, such as an unusually high payment amount or transfer to an account in a jurisdiction that does not fit the life context and payment behavior of the customer. The assumption is that algorithms can detect anomalies in behavior during payment transactions.

West and Bhattacharya ( 2016 ), Abdallah et al. ( 2016 ), Hilal et al. ( 2021 ), and Ali et al. ( 2022 ) reviewed financial fraud. They found a low number of articles regarding online payment fraud. For example, Ali et al. ( 2022 ) cited 20 articles on financial statement fraud and 32 articles on credit card fraud, see Li et al. ( 2021 ) for credit card fraud detection. Online payment fraud was not listed. The reviews also clarified that many articles utilized aggregated characteristics. However, we emphasize that fraud in online payments can only be detected based on individual data, as such fraud can only be detected through the possible different behavior of the fraudster and the account holder during payments. As fraudsters learn how best to behave undetected over time, they adapt their behavior. Therefore, self-learning defense methods are expected to outperform static-rule-based algorithms. The correctness of the expectation was shown by Abdallah et al. ( 2016 ) and Hilal et al. ( 2021 ). Various machine-learning algorithms for fraud detection have been proposed in the literature, including decision trees, support vector machines, and logistic regression to neural networks. Aggarwal and Sathe ( 2017 ) discussed various methods for outlier ensembles, and Chandola et al. ( 2009 ) provided a taxonomy and overview of anomaly detection methods.

A common feature in many studies is imbalanced data (i.e., the low proportion of fraud events in the dataset, see Wei et al. 2013 ; Carminati et al. 2015 ; Zhang et al. 2022a ; Singh et al. 2022 ). Risk detection involves detecting fraudulent transactions and stopping them before execution.

In addition to the efficiency of the algorithms, the data basis is an important reason for the differences in fraud-detection performance. While many studies have utilized either often less rich synthetic or Kaggle data, we were able to work with real data. Log files, which have substantial information content in our work, are hardly expected in Kaggle data. The difference in the data complexity is also reflected in the number of features. Singh et al. ( 2022 ) showed that the feature space consists of 31 features compared to our 147 features. Moreover, the proportion of fraudulent transactions in Singh et al. ( 2022 ) is more than a hundred times higher than in our case. Consequently, our data are much more unbalanced than any other study we know of, and the task of finding efficient fraud detection algorithms is more difficult.

However, limiting risk management to the optimal detection of anomalies does not ensure that losses caused by fraud are minimal. Optimal fraud detection can be economically suboptimal if, for example, it is efficient for small amounts of money but unsuccessful for large amounts. Thus, the machine learning outputs for risk identification must be optimized from an economic perspective. We call this optimization the triage model. Yet, neither fraud detection nor the triage model can provide an answer to the question of how large the losses in a payment channel are. Therefore, we develop a statistical risk model that considers the effects of countermeasures on loss potential. The risk model provides risk transparency and makes it possible to assess which measures in the fight against fraud in various payment channels make sense from an economic and risk perspective.

Literature on fraud risk models often refers to qualitative or assessment models for assessing fraud risk or risk assessment models (Sabu et al. 2021 ). We are not aware of any quantitative fraud risk management framework that explicitly considers the impact of the fraud detection process statistically in risk modelling. For organizational, procedural, and legal risk aspects, we refer to the literature. Montague’s ( 2010 ) book focuses on fraud prevention in online payments but does not consider machine learning and risk management in detail. The Financial Conduct Authority’s Handbook (FCA 2021 ) provides a full listing of the FCA’s legal instruments, particularly those relating to financial crime in financial institutions. Power ( 2013 ) highlights the difference between fraud and fraud risk from historical and business perspectives. Van Liebergen ( 2017 ) looks at “regtech” applications of machine learning in online banking. Fraud risk events in cryptocurrency payment systems are different from the online banking cases under consideration; see Jung et al. ( 2019 ) for fraud acting on a decentralized infrastructure and the review article of Trozze et al. ( 2022 ).

The development and validation of the three linked models are the main contributions of our work. To our knowledge, this is the first study to develop, validate, and link components of the risk management process. The output of the anomaly detection model (i.e., the ROC curves), is the input for the triage model, which provides economically optimized ROC curves. Fraud statistics data were utilized to calibrate the various components in the risk model. With these three models, the fraud risk management process can be qualitatively implemented at the same level as the risk management of market or counterparty risks (see Bessis ( 2011 ) to describe risk management in banks).

The performance of our risk management framework is the second contribution, although the performance comparison of our fraud detection method with the literature is limited and cautious, due the use of synthetic data instead of real data, a consideration of different channels in payments with different behavioral characteristics of bank customers, and the publication of incomplete statistics. Nevertheless, we compared our work with Wei et al. ( 2013 ) and Carminati et al. ( 2015 ), both of which analyze online banking fraud based, in part, on real data. The true positive rate (TPR) at a false positive rate (FPR) of \(1\%\) was \(45\%\) . In Wei et al. the TPR is between \(49\%\) and \(60\%\) , but unfortunately, the FPR is unknown. In the relevant scenario of Carminati et al. ( 2015 ), the TPR is \(70\%\) with an FPR of \(14\%\) . This FPR is not acceptable to any bank. Processing by specialists leads to high costs. We discuss all these statements in detail in the " Validation results " section. Considering all three models, the theoretical and practical importance of our approach becomes clear. The expected losses in a scenario of CHF 2.023 million, which utilizes the results of machine learning without economic optimization in the triage model, common in the literature, are reduced to CHF 0.800 million with the triage model (i.e., a reduction in the loss potential by more than 60% follows). In addition, if fraud detection is implemented without a risk model, fraud risk can be massively overestimated. Applying our models to three different payment channels, the overestimation of risk ranged from 54% to over 700%.

The remainder of this paper is organized as follows. In " Fraud risk management framework " section, the model selection for the fraud risk management framework is motivated and described. In " Online payment fraud anomaly detection " section, we consider the anomaly-detection model. " Fraud detection triage model " section links fraud detection to an economic perspective utilizing the triage model. " Risk model " presents the statistical risk model. " Conclusion " section concludes.

Fraud risk management framework

We provide an overview of the three interrelated quantitative models in the context of risk management: online payment anomaly detection, triage model, and risk model.

Online payment fraud anomaly detection

The goal of anomaly detection is to detect fraudulent activities in e-banking systems and to maintain the number of false alarms at an acceptable level. The implementation of the model consists of three steps: pre-filter, feature extraction, and machine learning.

Non-learning pre-filters ensure that both obvious fraud and normal transactions are sorted early to reduce the false positive rate. Only transactions that pass the pre-filter step are passed on to the machine-learning model. Banks utilize non-customer-based static if-then rules, such as blacklists or whitelists. Pre-filters free the algorithms from obvious cases. The adaptability and flexibility of the machine-learning model is necessary to counter the ever-improving attacks of fraudsters with effective fraud detection.

Our data face the following general challenges in payment fraud detection (per Wei et al. 2013 ): large transaction volume with the need for real-time fraud detection, a highly imbalanced dataset, dynamic fraud behavior, limited forensic information, and varying customer behavior.

Given the extremely imbalanced data, fully supervised algorithms typically struggle. Aggarwal and Sathe ( 2017 ) proposed unsupervised and weakly supervised approaches based on features that encode deviations from normal behavior. For each customer participating in an e-banking session, we assess whether the agent’s behavior is consistent with the account holder’s normal behavior. The key information for behavioral analysis lies in the sequence of the customer’s clicks during the session. We show that, unlike online e-commerce transactions (see Wei et al. 2013 ), transaction data, customer behavior data, account data, and booking data are also important for the performance of the algorithm. More precisely, these features are divided into behavioral, transactional, and customer-related features. Starting with nearly 800 features, 147 were extracted utilizing a Bagged Decision Tree Model (BDT). These numbers are many times higher than those for credit card fraud with one to two dozen features (see Table 8 in Hilal et al. 2022 ). The high dimensionality of the feature space also arises in machine learning corporate default prediction models, where several steps are needed to extract all noisy features (see Kou et al. 2021 ).

Our e-fraud model operates according to the following principles.

The model learns the “normal” behavior of each customer based on historical payment data and online banking log data.

Each new transaction is checked against the learned “normal” behavior to determine if it is an anomaly by extracting the 147 features from the data.

If an anomaly is detected, it is flagged as suspected fraud.

Detected transactions that are not found to be fraudulent after manual review are reported back to the model for learning purposes.

As there are very few known fraud cases, all base learners are trained on fraud-free data only in step one. Fraud cases are only utilized in step two of ensemble aggregation when the base learners are combined to form the final predictive function. The first step is to define base learners who are rich enough to detect a wide range of suspicious transactions or online user sessions. We consider three base learners: the density-based outlier detection model (unsupervised, Local Outlier Factor (LOF)), the isolation-based outlier detection model (unsupervised, Isolation Forest (IF)), and a model for normal customer behavior (supervised, Bagged Decision Trees (BDT)) as base learners (see Breunig et al. ( 2000 ), Chandola et al. ( 2009 ); Zhang et al. ( 2022b ) for LOF, Liu et al. ( 2012 ); Tokovarov and Karczmarek ( 2022 ) for IF). We refer to individual instances of LOF, IF, or BDT as base learners. The BDT model is not only a base model, but it is also utilized for feature selection in the other two base models: LOF and IF. The LOF method is suitable for outlier detection, where each observation is assigned an outlier level based on its distance from the nearest cluster of neighboring observations. The aim is to detect outliers in inhomogeneous data, for which classical global outlier methods typically do not provide satisfactory results. Conversely, IF explicitly isolates anomalies without capturing all normal instances. These two methods consider the heterogeneity in the data.

In the second stage, the base learner’s fraud score was aggregated. We consider two approaches to determine the weights in the ensembles: a simple averaging and a supervised approach, although our model largely consists of unsupervised procedures because of the limited availability of fraud cases for which we can extract all the required features. However, we introduce supervision where we utilize scarce labelled data to adjust the importance of certain base learners in the voting scheme, ultimately deciding whether an observation is fraudulent. The penalized logistic regression chosen for classification allows for a better interpretation of the model, as the weights can be utilized to identify base learners, subsets of features, and subsets of samples that have been particularly useful in detecting a particular type of fraud.

Triage model

The fraud detection model calculates scores and, in comparison with a threshold value, decides whether a transaction is flagged as an anomaly. This process results in the probability of detection for a given investigation effort, as indicated by the ROC curve. By making the threshold dependent on the transaction size, we can ensure that larger transaction amounts are more likely detected than smaller ones. This gives up part of the true positive rate (TPR) to reduce overall economic losses (i.e., the TPR decreases for a given FPR). This economic optimization that leads to adjusted ROC curves defines the triage model.

To minimize expected cumulative losses, the constant fraud anomaly detection threshold becomes a function of the transaction amount. Here, the transaction amounts are random variables whose distributions are estimated. In the optimization problem, the transaction function is chosen to maximize the average cumulative sum of the detected fraudulent transactions, where the expected FPR must not exceed a certain threshold. Utilizing this optimal threshold function, the fitted ROC curves were obtained.

The optimization problem has a unique solution if the ROC curve is a concave function of the false positive rate function of the threshold and if the acceptance set of the expected false positive function constraint is convex. With the chosen piecewise linear false positive constraint function, the assumptions regarding the existence of an optimum are satisfied. The ROC curves that result when fraud anomalies are detected serve as inputs for the optimization. However, because only a vanishingly small number of fraud cases exist, the TPR values for certain FPR levels are subject to considerable uncertainty. Hence, cubic spline functions were utilized for the ROC curve of the optimization.

The UK Finance (2019) report states that the recovery value for online and mobile banking in the UK is 18% of the potential loss. Therefore, we introduced an extension to the optimization program to include recovery.

Losses from transaction fraud are included in operational risk incurred by banks. As for other operational risks, one of the key questions from a risk-management perspective is whether the allocated resources and countermeasures are adequate. To answer this, one needs some way of quantifying the risk incurred, ideally a Value-at-Risk (VaR)-model that fits the general risk framework of the bank. Simply, the model calculates the loss \(L=E(\lambda )\times E(\tau )\) where \(\lambda\) is the expected event frequency (fraud), and \(\tau\) is the expected loss per event. The challenge is to determine the distributions of these variables in a tractable and plausible manner and define a model while having very scarce data on past events. We chose the path of an aggregated simulation of many scenarios per e-channel to account for the inherent uncertainty in the choice of these parameters.

Unlike market or credit risk, fraud risk is borne by comparatively few individuals or groups who utilize very specific strategies and technologies to overcome vulnerability in the payment process. Simultaneously, defenders analyze attack plans and update their countermeasures. In this constantly changing environment, neither the frequency of attacks nor the transaction amounts can be assumed to be statistically regular with great certainty. Therefore, we propose a simple, flexible stochastic model composed of basic building blocks. With such a model, risk managers can quickly adjust the model as needed, perform what-if analyses, or simulate changes in payment infrastructure.

The basic model structure for the e-fraud risk model consists of (i) independent models for the three channels, whose components and parameters can be flexibly assembled and adjusted, (ii) sub-models in each channel based on three model types, and (iii) a recovery model for each channel. The three model types for the three different online payment channels in this study are a Beta model (restricted distribution of transaction amounts), a Generalized Pareto Distribution (GPD, unrestricted distribution of transaction amounts), and a “mass attack model” (many simultaneous Beta-type attacks).

Countermeasures against fraud and recovery measures after fraud events play an essential role in determining risk potential. Therefore, they were integrated into the risk models. Countermeasures against online fraud can be divided into those that strengthen the general infrastructure of the payment process and those that focus on defense against actual attacks. The former is conceptually part of the risk model described above, as it affects the frequency and possibly the transaction size of the attacks. However, the latter is better understood in the context of recovery and is considered in the triage model.

Raw data consisted of transaction data, interaction data between customer and e-banking interface, account, booking, and customer reference data. All users with fewer than 10 logged online sessions were removed as input for ensemble learning. The removed cases were handled separately by utilizing a case-back model.

The transaction history in our dataset consists of 140 million transactions over three years. One hundred fraud cases are reported, but only 11 cases can be linked to the recorded 900’000 online session logs: a \(0.0012\%\) fraud rate. Only 900’000 of the 140 million transactions were possible as the log files were only stored in the bank for three months. This change occurred after the project.

A feature vector is created for each e-banking session based on raw data. The interaction pattern features consist of n -grams constructed from customers’ request sequences and normalized deviations from the expected duration between each pair of consecutive requests sent by a customer in an online session. Particular attention was paid to the typical time required to complete a two-step verification process during enrolment. Payment pattern characteristics were calculated for weekday, weekly, and monthly seasonality. These include normalized deviations from the expected payment amount and remaining account balance. Technical data included the IP address of the online session, the HTML agent, and the number of JavaScript scripts executed. Finally, we utilized historically observed confirmed fraudulent transaction identifiers as the ground truth for the weakly monitored part of the pipeline.

Several quality checks were performed. Consistency tests ensure that the session interaction data and transactions match, for example, that the account exists or that a recipient is listed in the transaction. We also checked for missing or non-parsable values, the latter removed.

The data are extracted from several different data sources within the bank in a two-step Python extract, transform, and load (ETL) process, and converted into features for the algorithm. First, we introduce all raw data into a standard structured format for all data sources. Then, we perform the feature engineering described in the following sections to compute the data input for the ensemble.

Our fraud rate \(0.0012\%\) is much lower than that reported in the literature. The figures in the two online banking fraud papers, Wei et al. ( 2013 ) and Carminati et al. ( 2015 ), are \(0.018\%\) and \(1\%\) , respectively. For credit card fraud, the fraud case number is larger, such as \(2\%\) in Piotr et al. ( 2008 ). Outside the banking fraud sector, anomalies account for up to 40% of the observations (see Pang et al. 2020 ). Similar numbers hold in Zhang et al. ( 2022b ), who tested their ensemble-based outlier detection methods on 35 real datasets from various sectors outside the financial sector, with an average fraud rate of 26%.

Feature extraction

For weak supervision, we utilized historically observed confirmed fraudulent transaction identifiers as the ground truth. For training and inference, we created a feature vector for each e-banking session based on the raw data. Each feature aims to encode deviation from expected (“normal”) customer behavior, as observed in historical interactions with the online banking interface and executed transactions. Three types of features are considered.

Behavioral features

The underlying motivation for utilizing features derived from customers’ online session logs is that a large fraction of online payment fraud involves hijacking, where a foreign agent (human or robot fraudster) takes control of the e-banking session. Consequently, it is expected that the timing and sequence of requests posted by the user in a fraudulent session will be significantly different from those in a non-fraudulent session. We utilize the information about user request types (e.g., “get account balance”, “create payment”) and corresponding timestamps to construct the following features:

Normalized n-gram frequency of user’s requests within the online session. We utilized single, pairs and triplets of consecutive requests (1-, 2- and 3-grams) to derive a fixed-size representation. We performed normalization by dividing the number of occurrences for each request with the total number of requests in each session.

Normalized time between consecutive user request n-grams. For each pair of recorded consecutive n-grams, we transformed absolute time between them into deviations by computing z-scores relative to respective historical observations.

Technical attributes of a session (e.g., IP address of the online session, HTML agent, screen size, and the number of executed Javascript scripts) in binary format - 0 if previously observed, otherwise 1.

Transactional features

Transactional features aim to quantify how anomalous aggregate payments scheduled in an online session are compared with previously executed payments and the remaining balance on the account. They are designed to capture attempts to empty a victim’s account through a single large or many small transactions, while being mindful of seasonal patterns (e.g., holidays, travel expenses, bills, etc.).

Normalized ratio of the payment amount relative to remaining account balance. We normalize by computing z-scores relative to historically observed ratios.

Deviation of the scheduled payment amount from the seasonally expected amount. We compute four deviations per session using z-scores relative to historical payments executed in the same hour of the day, day of the week, week of the month, and month of the year, respectively.

Scheduled time for payment execution in binary format - 0 if immediate, 1 if lagged.

A short payment history and many accounts with relatively infrequent transactions proved detrimental to seasonality modelling, hence, these features were omitted from the final model.

Customer-related features

Customer-related features provide insight into peer groups, relationships with other customers, and the service packages they utilize. These include:

Sociodemographic (e.g., age, nationality, profession, and income)

Seniority of client relationship

Product usage (i.e., savings, investment accounts, and mortgages)

Relationship to other customers (i.e., shared accounts, spouses, and family)

These features were not considered within the scope of our study because of data limitations and time constraints.

Functionality and structure of the fraud model

Base learner: bagged decision tree.

Bagged decision trees (BDT) are trained utilizing the concept of transfer learning, which assumes that distinguishing between the behaviors of different clients within their online sessions is a related problem in distinguishing between fraudulent and non-fraudulent sessions. The underlying motivation considers that a large fraction of online payment fraud involves hijacking or when a foreign agent (human or robot fraudster) takes control of the e-banking session. As fraudulent sessions are rare and non-fraudulent sessions abound, the utilization of transfer learning enables the extraction of custom patterns from a much broader dataset and the use of supervised learning. Transfer learning comprises two phases. A learning phase where one discriminates behavioral characteristics of each customer vs. non-customers and a prediction phase where discrimination between non-fraudulent and fraudulent users sessions were considered in this study. The “non-customer” class label is then attributed to fraudulent behavior.

The decision function in BDT base learners is the probability that an observation, a planned transaction x , is a “non-customer behavior.” This value is equal to the average probability observed for all decision trees in the forest.

where M is the number of trees in the bagged forest, \(c_j\) is the corresponding customer behavior class, and \(P_i( Y\ne c_j|X=x)\) is the probability that observation x is a “non-customer behavior” as predicted by the i th tree. Customer behavior class \(c_j\) consists of the collected sessions and transactions, excluding potential fraud. The model was fitted as follows:

For each customer \(c_j\) , we collect the associated sessions, excluding the potential fraud cases. This set represents the “behavior of customer \(c_j\) .”

From the pool of all customer sessions \(C_{-j}\) (excluding \(c_j\) ) we draw a uniform sample of observations to generate a set representing the class “behavior of customer \(c_j\) not,” \(\not c_j\) for short, and equal in size to the \(c_j\) set. Equal sampling is performed to ensure that none of the other customers are overrepresented in \(\not c_j\) .

Each bagged forest in the ensemble is trained on a matching feature subspace utilizing all these observations. The forests consist of 100 decision trees each.

BDTs provide variable selection as an additional benefit, owing to the large amount of data involved in supervised classification. Therefore, they can be utilized to estimate the importance of variables and adequacy of feature engineering. We achieve this by calculating the Gini impurity decreases at each split point within each tree for each feature. Gini impurity is a measure of the likelihood that a randomly selected observation would be incorrectly classified by a specific node m :

where \(p_{mi}\) is the portion of samples classified as i at node m . These impurity decreases were averaged across all decision trees and outputs to estimate the importance of each input variable. The greater the decrease in impurity, the greater the importance. Utilizing these results, we can identify the relevant subsets of input variables.

Relying on the concept of transfer learning (if the problem described in ?? is sufficiently like fraud detection), we use BDT to select a subset of \(N=147\) features. Particularly important were features measuring the deviation from the typical time required to complete a two-step verification process during login. Features that encode the relative time between n -grams and specific user request sequences are important. The following additional base model was built using the features selected by BDT.

Base learner: local outlier factor

The Local Outlier Factor (LOF) detection method assigns an outlier level to each observation based on its distance from the nearest cluster of neighboring observations (Breunig et al. 2000 ). The general intent of the LOF model is to identify outliers in the interior region of data, for which classical global outlier methods and the other considered algorithm, isolation forest, usually do not provide satisfactory results. The LOF decision function is as follows:

with K a k -neighborhood and LD ( x ) the local reachability distance density from x to its k -th nearest neighbor. We fit the model for each customer by collecting the first associated sessions or transactions, excluding potential fraud. Each LOF in the ensemble is created utilizing all these observations on a subspace of the relevant features selected by BDT and the sampled hyper-parameter. Finally, each time a new observation is available, its decision function value is computed regarding the observations from the training set.

Base learner: isolation forest

The IF algorithm recursively splits the data into two parts based on a random threshold, until each data point is isolated. The algorithm randomly selects a feature at each step, and then randomly selects a division value between its minimum and maximum values. The algorithm filters out data points that require fewer steps to be isolated from the entire dataset. In our case, IF separates one observation from the rest of a randomly selected subsample of the original dataset (Fei Tony et al. 2008 ). Anomalies are instances with short average isolation path lengths. The IF decision function is

where E ( H ( x )) is the average number of edges traversed to isolate the node and C is the average number of edges traversed in an unsuccessful search. To fit the model, we first collected each client’s associated sessions or transactions, excluding potential fraud cases. In step two, each isolation forest in the ensemble was created utilizing all these observations in a matching feature subspace. Each forest consisted of 100 isolated trees. Finally, each time a new observation was available, its decision function value was computed regarding the isolation trees created based on the training set.

Base learner scores combination

The decision functions of the base learners produced by our ensembles must be combined into a single fraud score. As these are in different ranges and scales to render the decision functions comparable, we first replace the original scores by their ranks, regarding the non-fraudulent training scores. Rank normalization is more robust and numerical stable as opposed to z-scores, for example. Therefore, we replaced the original scores with their ranks regarding the non-fraudulent training scores for each base learner:

where V is the set of all \(\delta _{\text {Base}}(p )\) over all observations \(p'\) in the learners’ training subsample, with Base being LOF, IF, or BDT.

Owing to the few fraud cases, our model largely consists of unsupervised procedures. However, we introduced supervision utilizing scarce labelled data to readjust the importance of particular base learners in the voting scheme, ultimately deciding whether an observation is fraudulent.

The following score combination procedure was established. First, a training set comprises all fraud cases in the sample, along with healthy transactions uniformly sampled over customers from the ensemble training data. Second, a logistic regression is trained to classify observations as fraudulent or not utilizing the 6 N normalized decision function features and the known fraud status on past transactions as the label.

The following binary-class penalized cost function is minimized:

where \(y_i\) is the fraud label of transaction i , \(X_i'\) is a row of 6 N decision functions describing transaction i , R is the regularization factor, and ( w ,  c ) is a set of weights defining the decision boundary between the two classes: fraud and non-fraud. To account for the imbalance between fraud and non-fraud transactions in our sample, we assign asymmetric penalties for fraud misclassification, as opposed to non-fraud classification.

Choosing logistic regression to optimize the weights of the base learners ensures that the final score combination \(x'w\) represents the log-odds of an observation to be fraudulent.

where \(\delta (x)\) is the probability that the observation x is fraudulent. Finally, to assign a fraud label to a session x , we compare the output combined score, or equivalently, the probability \(\delta (x)\) , to a threshold y , which is chosen based on ROC curve analysis such that the defined maximum allowed false positive rate is not exceeded. The decision boundary of logistic regression is linear, where each base learner is assigned a weight \(w_i\) to determine its relative importance for fraud classification. This structure simplifies the interpretation of the model because these weights can be utilized to identify the base learners, feature subsets, and sample subsets, which are particularly useful in detecting a particular type of fraud associated with a high weight \(w_i\) . Appendix A provides a detailed description of the ensemble design.

Normal customer behavior model

In summary, we created an ensemble model for each client, which is re-trained with new data at regular intervals and can be described by the following steps:

We consider the disjoint sets of behavioral features on session observations and transactional features on transaction observations.

For each of the two features/observations-pairs, we define \(N=1000\) learners for each of the three model categories as follows.

We fix N random sub-samples of features from the feature set. Each sub-sample remains fixed for all customers.

For each customer, we fix N random observation samples from the customer-specific sessions or transactions observations.

For each of the three model categories, for each customer, and for \(i=1,...,N\) , a base learner is defined by applying the model algorithm to the i -th features sub-sample and i -th observations sub-sample. Thus, this results in 6 N base learners per customer, 3 N for sessions, and another 3 N for transaction data.

The decisions for the three base learners are aggregated utilizing supervision, where the knowledge obtained from existing fraud cases is utilized to adjust the base learner weights.

Utilizing this representation, we train a model that (i) outputs an indicator of how likely a scheduled transaction is fraudulent, (ii) aggregates the overall provided decision functions to derive the unified hypothesis, while assigning more importance to learners that showcase the capability to better distinguish between fraud and non-fraud, and (iii) deals with a large imbalance in class representation.

Validation results

The training, test, and validation sets consisted of data collected from July to October 2017, as dictated by the availability of online session log files. Around 900’000 sessions formed the dataset.

Raw data were processed using ETL to derive customer-specific feature representations of each recorded online session. The data were then split into non-fraud and fraud sets. The fraud set was not utilized to train the unsupervised base learners. Non-fraudulent (“healthy”) sessions were separated into training and test sets utilizing a 3-fold cross-validation split. We then sequentially trained the models on each derived training fold and computed the scores for observations in the corresponding test folds. Following, we obtained an out-of-sample decision function value for each healthy session and each base learner. We then assigned base learner scores to each fraudulent session utilizing base learners trained on all healthy data.

The out-of-sample logistic regression decision function values were aggregated by averaging within their respective ensembles (LOFs, IFs, and BDTs). This step yields a 3-dimensional representation of each customer’s online session. Finally, we utilized leave-one-out cross-validation to report the ROC curve measures. Hence, the logistic regression model is consecutively trained on an all-but-one observation, followed by computing the probability of an observation that was left out. Thus, we again obtain an out-of-sample fraud probability for each observation in the sample. We opted for leave-one-out cross-validation to maximize the number of fraudulent observations in each training set, because these are particularly scarce. Once we have obtained the aforementioned out-of-sample probabilities for each observation, we construct an ROC curve to display the FPR and TPR relationship depending on the decision threshold.

Resultantly, when utilizing no transaction data, the detection rate of the machine learning model was in a realistic range of 18% true positives. These primary results can be easily optimized to increase the TPR and simultaneously reduce the FPR utilizing different measures. This led to an increase in true positives by up to 45%, see Table  1 .

Overall, LOF seems to perform best over the entire dataset compared to IF and BDT. However, BDT has a slightly steeper ROC curve at the beginning, thus showing better pure outlier detection capabilities. Furthermore, because BDT seems to detect frauds, as discussed below, involving larger amounts than those detected by LOF, we cannot conclude that LOF outperforms the other approaches considered. Aggregating the decision functions of the ensembles utilizing simple means outperformed supervised aggregation. Through analysis of logistic regression weights assigned to each ensemble of learners, we determined that significantly higher weights were assigned to the LOF ensemble, most likely due to its best performance over the whole dataset. This dampened the input from the other two ensembles. However, this is not the case when the mean for aggregation is utilized. The results were affected by the small number of frauds and the size of the sample analyzed.

Different ensembles detected different types of fraud, and by observing figures depicting money saved per raised alarm, we see that different ensembles (LOF and BDT) detect different types of fraud cases, displayed by a large difference in saved money per trigger. Logistic regression supervision alarms were affected mainly by LOF, thus making it miss large embezzlement detected by the BDT ensemble. This motivates the triage model described in the next section. As of the restriction of the FPR to no greater than 2%, the entire ROC curve is of less interest. The ROC AUC values for the LOF ensemble is 0.93, for the BDT ensemble 0.82, and for the mean decision function ensemble 0.91.

We compared our results with those of Wei et al. ( 2013 ), and Carminati et al. ( 2015 ). These are two of the few studies dealing with online fraud detection that use real-world data, at least in part. Wei et al. ( 2013 ) utilized an unsupervised approach, whereas Carminati et al. ( 2015 ) utilized a semi-supervised approach. Table  2 compares the performance of our model with those of Wei et al. ( 2013 ) and Carminati et al. ( 2015 ). The results of this table should be interpreted with caution. First, different payment channels were considered. Second, the data of Carminati et al. ( 2015 ) were anonymized and did not include fraud cases. These are artificially added to tune \(1\%\) of the data volume, compared to \(0.018\%\) in Wei et al. ( 2013 ) and \(0.0012\%\) in our dataset. Third, Wei et al. ( 2013 ) did not report the FPR. Finally, Carminati et al. ( 2015 ) published an almost perfect error detection for scenarios I + II, but in scenario III, the false positives are too high; they generate too much manual work for the bank. The former scenarios are simple fraud scenarios that would be blacklisted and filtered out in our data before machine learning engages.

Fraud detection triage model

Formalization.

We formalize the triage model and denote by \(\Omega\) the set of all transactions with \(\omega\) as a single transaction, \(T(\omega )\) as the transaction amount function, and \(\chi _F(\omega )\) as the fraud indicator function, where \(\chi _F(\omega )=1\) represents fraud. Space \(\Omega\) is attributed to a probability distribution P with p ( x ) as the density function of the transaction amounts. The threshold value L of the fraud score function S is a function of the transaction amount x . We define:

If we assume stochastic independence of the transaction amount T , score S and fraud indicator \(\chi _F\) , we obtain the following interpretation:

Note that the assumptions of independence are strong, as transaction sizes are utilized as the input of the machine-learning model underlying the score. Conversely, the independence of \(\chi _f\) and T implies that the transaction amounts of fraudulent transactions have the same distribution as those of non-fraudulent transactions. In the context of the value-at-risk model in the next section, we argue that there is little evidence to support this. This considered, the assumption of independence is theoretically difficult to uphold, but in practice quite necessary to obtain our results.

We formulate our optimization problem as follows:

under the constraint of the integrated FPR

The expectation in ( 11 ) is the average cumulated sum of fraudulent transaction amounts detected by the detection model. By letting \(q_0:=E(\chi _F)\) and utilizing ( 10 ), we can rewrite it as

The constant \(q_0\) is irrelevant to the optimization. Setting \(g(x)=\text {FPR}(L(x))\) , we reformulate the optimization problem in terms of the ROC curve as

under the constraint:

To account for the recovery, we introduce a recovery function \(\theta :\Omega \rightarrow [0,1]\) . This function changes the objective function in the optimization problem, as follows:

whereas this constraint does not change.

Optimization

To put our formal model into practice, we need to fix a distribution for transaction amounts. Utilizing approximately 12 million transactions from online banking and 1.2 million from mobile banking, we approximated the distribution for both channels utilizing lognormal distributions. Although this choice does not particularly focus on the distribution’s tails, it will be seen that the optimal model still places strong emphasis on the detection of anomalies with large transaction amounts. Some basic statistics of the fitted distributions are given in Table  3 .

The ROC curve is conceptually the output of the detection model described in the previous section. However, owing to the limited number of actual fraud cases available, the TPR values for the given FPR levels are tainted with considerable uncertainty. The ROC curve utilized in our optimization was obtained by fitting a cubic spline function to the base points, as presented in Table  4 . The support points were adjusted to avoid unwanted spikes in the cubic interpolation.

As the triage model aims to prevent large losses with a higher probability than smaller ones, the optimal FPR will be an increasing function of the transaction size. To avoid possible optimization problems, we choose a simple form for FPR as a function of transaction size, namely, a piecewise linear function satisfying \(g(0) = 0\) , \(g(T_1) = a\) , \(g(T_2) = 1\) for the parameters \(a > 0\) and \(0< T_1 < T_2\) (see Fig.  1 ).

The optimization problem can be simplified by assuming equality in ( 12 ) and solving a as a function of \(T_1\) and \(T_2\) . For a target integrated FPR of 0.4%, we obtained the solutions listed in Table  5 .

Figure  1 illustrates the results for an online banking channel. The concave shape of the FPR curve up to \(T_2\) shows that the optimal solution emphasizes the detection of large transaction fraud cases, accepting, in turn, the less rigorous testing of small and moderate transactions up to \(T_1\) . For transaction amounts larger than \(T_2\) , FPR and TPR are equal to 1 by construction. Hence, all such transactions are automatically flagged as anomalies.

Total Effectiveness

is the average percentage of integrated fraudulent transaction amounts detected as anomalies. In our optimized case, the rate was 39%.

Panel Left: False positive rate as a function of the transaction amount under the constraint that the total false positive rate is smaller than 0.4 percent. Right Panel: True positive rate as a function of the transaction amount. The total effectiveness is 39 percent

Compound Poisson processes were utilized as basic building blocks. We utilize beta marginal distributions for modelling bounded transaction amounts and generalized Pareto marginal distributions (GPD) for unbounded ones. The so-called mass-attack model is formulated as a nested compound Poisson process with a marginal beta distribution. All subprocesses are aggregated independently. Loss statistics, such as value-at-risk or other quantiles of the distribution, are obtained by running Monte Carlo simulations.

Utilizing the limited available fraud data and drawing on discussions with practitioners, we develop the following model for online banking fraud:

Isolated attacks with a moderate transaction size of up to CHF 70’000 are modelled by a compound Poisson process with beta marginal distribution.

Isolated attacks with transaction amounts larger than CHF 70’000 are modelled by a compound Poisson process with GPD marginal.

“Mass attacks” are modelled as a nested compound Poisson process, where the inner Poisson process simulates the individual transactions triggered by the mass attack. The inner process has a beta marginal distribution and generates transaction amounts up to CHF 20’000.

The intensities of the Poisson processes constituting the submodels vary. In our case, isolated attacks of moderate size were by far the most frequent, followed by isolated attacks of large size. Mass attacks were the least frequent.

Mobile banking fraud is modelled analogously, albeit with transaction sizes only up to CHF 20’000, because larger amounts were inadmissible on this channel during our investigation. Hence, there is no Poisson process with GPD marginal in this case. Contrastingly, in the EBICS channel, which is an internet-based payment channel between banks, only the possibility of large fraudulent transactions was of interest. Hence, this model consists of a single compound Poisson process with GPD marginals above CHF 100’000. The details of the parametrization are given in Appendix A .

Countermeasures against fraud and recovery measures after fraud events play an essential role in determining risk potential. Therefore, they were integrated into the risk models. Countermeasures against online fraud fall into two categories: those that strengthen general infrastructure of the payment process to make it harder for attackers to find a weak spot, and those that are geared towards fighting off actual attacks. The first type is conceptually part of the base model described above, as it affects the frequency and possibly the transaction size of attacks. However, the second type is better understood in the context of recovery.

A recovery variable is introduced in the triage model, which accounts for it often being possible to recover money even after it has been transferred to another bank through fraudulent transactions. Conversely, by monitoring transactions utilizing the fraud detection and triage model, a certain percentage of attacks can be identified even before the transactions are released. The ROC curve of the detection model’s ROC curve, in combination with the triage model, allows us to infer the probability of detection from the transaction size: such that this component of the recovery process is readily integrated into the stochastic framework.

Owing to the nonlinearity of the risk statistics, the aggregation of the models was performed at the level of individual scenarios. Thus, for each scenario, \(s_i\) , the loss of the overall model for one-year was calculated from the simulated loss events of the channel models:

For each sub-model, the loss is calculated by pulling the event frequency for the year according to the Poisson intensity, loss magnitude according to the marginal distribution, and stochastic recovery:

where \(\text {Rec}\) denotes recovery function. Simulated loss figures were obtained by simulating the nested overall model, from which the risk statistics could be calculated empirically. Juniper Research ( 2020 ) estimated the recovery rate as \(18\%\) .

The simulation results for online banking are presented in Table  6 . The table shows the simulation results without applying fraud detection utilizing a constant FPR level of 0.4% and the triage model for an integrated FPR of 0.4%, respectively. In this simulation, no additional recovery was applied.

The above table shows the strong mitigation of risk due to fraud detection. The triage model performs better than the constant FPR benchmark in all submodels, particularly for the GPD submodel. Recall that the triage model places strong emphasis on detecting large fraudulent transactions, even flagging all transactions larger than CHF \(192'000\) .

As a second application, we compare the results of this risk model for the three e-channels with the bank’s overall 2019 risk policy. This means that we compare the capital-at-risk (CaR) limits for market and credit risks with operational risk limits, where the e-channel part is now calculated in our model. The following allocation of CaR holds according to the annual report of the bank Footnote 1 : Credit Risk, 69%; operational risk, 11%; market risk trading, 4%; market risk treasury, 11%; market risk real estate, 2%; and investment, 4%.

Approximately 1% of operational risk capital can be attributed to these three channels. Even if we add another 4–5% of the total volume to all payment services, including corporate banking and interbank payments, less than 10% of the operational risk capital is attributed to payment systems. As payment systems account for a significant portion of operational risk, our results confirm serious doubts about the accuracy of the chosen operational risk capital in banks. Without reliable models and data, capital is determined by utilizing dubious business indicators. Our models, which represent a micro-foundation of risk, show that, at least in payment systems, trustworthy risk quantities can be derived by combining machine learning and statistics.

Defense against sophisticated online banking fraud involve several resources and methods. These include risk models, algorithms, human action, knowledge, computer tools, web technology, and online business systems in the context of risk management.

We show that anomaly detection is not only useful per se, identifying a significant proportion of fraud while controlling false alarms, but that linking anomaly detection with statistical risk management methods can significantly reduce risk. A bank equipped with an anomaly detection system will be exposed to orders of magnitude of higher risks in payments than a bank implementing our end-to-end risk management framework with the three components of fraud detection, fraud detection optimization, and risk modelling.

As fraud is part of regulated operational risk, our model allows us to analytically capture these operational risks without crude benchmarking. This also provides a microeconomic foundation for capital adequacy. In the area of operational risk, these results put internal models that are not risk sensitive or difficult to verify on a solid footing.

A complicated problem, such as online payment fraud detection, requires a comprehensive understanding. A prerequisite for this is access to a large dataset. To evaluate our method, we utilized a real dataset from a private bank. Regardless of the chosen algorithm, feature extraction is an essential part of developing an effective fraud detection method. We utilized historically observed and confirmed fraudulent transaction identifiers as the ground truth. Each feature in the feature vectors for each e-banking session aims to encode deviations from normal customer behavior. Thus, behavioral, transactional, and customer-specific features are important.

Our framework opens interesting directions for future research. Roughly speaking, the framework goes in only one direction, from machine learning methods in fraud detection to statistical risk modelling. The feedback process from the risk model to the triage model and from the triage model back to the fraud detection model is a challenging task that can be addressed utilizing reinforcement-learning methods. With such a feedback loop, the entire risk-management framework becomes a learning system. Another research direction is to extend the optimization of fraud detection (triage model) by considering transaction-dependent loss risks and other features such as customer segmentation. More emphasis is placed on segments that are known or suspected to be less alert or more vulnerable to fraudulent attacks. This resulted in a higher-dimensional triage model.

Availability of data and materials

The bank provided real transaction data and data on transactions (“raw data”) of the customers. The legal basis of the Swiss Federal Data Protection Act (2020) prevents the raw data from leaving the bank in any form or being accessible to any party other than the bank.

CaR for credit risk is VaR on the bank’s quantile level and for market risk CaR was in the past chosen on an annual basis and a risk budgeting process was defined to align present risk with the annual risk budget.

Abbreviations

Swiss Franc

Receiver operating curve

Area under the ROC curve

True positive rate

False poistive rate

Bagged decision tree

Local outlier factor

Isolation forest

Value-at-risk

Generalized pareto distribution

Payment channel for corporate banking clients

Extract, transform, load is a three-phase process where data is extracted, transformed and loaded into an output data container

Abdallah A, Maarof MA, Zainal A (2016) Fraud detection system: a survey. J Netw Comput Appl 68:90–113

Article   Google Scholar  

Ali A, Shukor AR, Siti HO, Abdu S (2022) Financial fraud detection based on machine learning: a systematic literature review. Review Appl Sci 12:9637

Amiri M, Hekmat S (2021) Banking fraud: a customer-side overview of categories and frameworks of detection and prevention. J Appl Intell Syst Inf Sci 2(2):58–68

Google Scholar  

Aggarwal CC, Sathe S (2017) Outlier ensembles: an introduction. Springer

Bessis J (2011) Risk management in banking. Wiley, New York

Bolton RJ, Hand DJ (2002) Statistical fraud detection: a review. Stat Sci 17(3):235–249

Bolton RJ, Hand DJ (2001) Unsupervised profiling methods for fraud detection, Credit Scoring and Credit Control VII, pp 235–255

Breunig MM, Kriegel H-P, Ng RT, Sander J (2000) LOF: Identifying density based local outliers. In: Proceedings of the 2000 ACM SIGMOD international conference on management of data, pp 93–104

Carminati M, Caron R, Maggi F, Epifani I, Zanero S (2015) BankSealer: a decision support system for online banking fraud analysis and investigation. Comput Secur 53:175–186

Chandola V, Banerjee A, Kumar V (2009) Anomaly detection: a survey. ACM Comput Surveys 41(3):1–58

Embrechts P, Klüppelberg C, Mikosch T (2013) Modelling extremal events: for insurance and finance (Vol 33). Springer Science & Business Media

FCA (2021) Financial conduct authority handbook. www.handbook.fca.org.uk

Fei Tony L, Kai T, Zhi-Hua Z (2008) Isolation forest. In: 2008 Eighth IEE E International Conference on Data Mining, IEEE, pp. 413-422

Haixiang G, Yijing L, Shang J, Mingyun G, Yuanyue H, Bing G (2017) Learning from class-imbalanced data: review of methods and applications. Expert Syst Appl 73:220–2

Hilal W, Gadsden SA, Yawney J (2021) A review of anomaly detection techniques and applications in financial fraud. Exp Syste Appl 116429.

Hilal W, Gadsden SA, Yawney J (2022) Financial fraud: a review of anomaly detection techniques and recent advances. Expert Syst Appl 193:11

Jung E, Le Tilly M, Gehani A, Ge Y (2019, July) Data mining-based ethereum fraud detection. In: 2019 IEEE international conference on blockchain (Blockchain) (pp 266-273). IEEE

Juniper Research (2020) Online payment fraud: Emerging threats, segment analysis and market forecasts 2020-2024. www.juniperresearch.com

KPMG (2019) Global banking fraud survey, KPMG International

Kou G, Xu Y, Peng Y, Shen F, Chen Y, Chang K, Kou S (2021) Bankruptcy prediction for SMEs using transactional data and two-stage multiobjective feature selection. Decis Support Syst 140:113429

Krawczyk B (2016) Learning from imbalanced data: open challenges and future directions. Prog Artif Intell 5(4):221–232

Li T, Kou G, Peng Y, Philip SY (2021) An integrated cluster detection, optimization, and interpretation approach for financial data. IEEE Trans Cybern 52(12):13848–13861

Liu FT, Ting KM, Zhou ZH (2008) Isolation forest. In: 2008 eighth IEEE international conference on data mining, pp 413-422. IEEE

Liu FT, Ting KM, Zhou ZH (2012) Isolation-based anomaly detection. ACM Trans Knowl Discov Data (TKDD) 6(1):1–39

McNeil AJ, Frey R, Embrechts P (2015) Quantitative risk management: concepts, techniques and tools-revised edition. Princeton University Press, Princeton

Montague DA (2010) Essentials of online payment security and fraud prevention, vol 54. Wiley, New York

Molloy I, Chari S, Finkler U, Wiggerman M, Jonker C, Habeck T, Schaik RV (2016) Graph analytics for real-time scoring of cross-channel transactional fraud. In: International conference on financial cryptography and data security, pp 22–40. Springer, Berlin, Heidelberg

Pang G, Shen C, Cao L, Hengel AVD (2020) Deep learning for anomaly detection: a review. arXiv preprint arXiv:2007.02500

Piotr J, Niall AM, Hand JD, Whitrow C, David J (2008) Off the peg and bespoke classifiers for fraud detection. Comput Stat Data Anal 52:4521–4532

Power M (2013) The apparatus of fraud risk. Account Organ Soc 38(6–7):525–543

Sabu AI, Mare C, Safta IL (2021) A statistical model of fraud risk in financial statements. Case for Romania companies. Risks 9(6):116

Shen H, Kurshan E (2020) Deep Q-network-based adaptive alert threshold selection policy for payment fraud systems in retail banking. arXiv preprint arXiv:2010.11062

Singh A, Ranjan RK, Tiwari A (2022) Credit card fraud detection under extreme imbalanced data: a comparative study of data-level algorithms. J Exper Theor Artif Intell 34(4):571–598

Tokovarov M, Karczmarek P (2022) A probabilistic generalization of isolation forest. Inform Sci 584:433–449

Trozze A, Kamps J, Akartuna EA, Hetzel FJ, Kleinberg B, Davies T, Johnson SD (2022) Cryptocurrencies and future financial crime. Crime Sci 11(1):1–35

Van Liebergen B (2017) Machine learning: a revolution in risk management and compliance? J Financ Trans 45:60–67

Vanini P (2022) Reinforcement Learning in Fraud Detection, Preprint University of Basel

Wei W, Li J, Cao L, Ou Y, Chen J (2013) Effective detection of sophisticated online banking fraud on extremely imbalanced data. World Wide Web 16(4):449–475

West J, Bhattacharya M (2016) Intelligent financial fraud detection: a comprehensive review. Comput Secur 57:47–66

Zhang W, Xie R, Wang Q, Yang Y, Li J (2022a) A novel approach for fraudulent reviewer detection based on weighted topic modelling and nearest neighbors with asymmetric Kullback-Leibler divergence. Decis Support Syst 157:113765

Zhang G, Li Z, Huang J, Wu J, Zhou C, Yang J, Gao J (2022b) efraudcom: An ecommerce fraud detection system via competitive graph neural networks. ACM Trans Inform Syst (TOIS) 40(3):1-29.

Zhou ZH (2012) Ensemble methods: foundations and algorithms. CRC Press, Boca Raton

Book   Google Scholar  

Download references

Acknowledgements

The authors are thankful to P. Senti, B. Zanella, A. Andreoli and R. Brun all from Zurich Cantonal Bank for the discussions and for providing us with the resources to perform his study. The authors are grateful to P. Embrechts (ETH Zurich) for the model discussions.

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Author information

Authors and affiliations.

University of Basel, Basel, Switzerland

Paolo Vanini

Novartis AG, Basel, Switzerland

Sebastiano Rossi

swissQuant Group, Zurich, Switzerland

Ermin Zvizdic

IT Couture, Zurich, Switzerland

Thomas Domenig

You can also search for this author in PubMed   Google Scholar

Contributions

Sebastiano Rossi and Ermin Zvizdic designed the fraud detection model and analysed the data. Thomas Domenig and Paolo Vanini designed the triage model. Thomas Domenig designed the risk model and did the calculations for the triage and risk model. Paolo Vanini wrote the manuscript with contributions from all authors. Paolo Vanini was involved in the analysis of the triage and risk model. Ermin Zvizdic led the project in the bank. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Paolo Vanini .

Ethics declarations

Competing interests.

The authors declare that they have no competing interests.

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

The structure of our ensemble results from several design decisions based on both insights from the field (online banking) and experience in building machine learning models. Our ensemble model was built according to the following guidelines:

Customer-specific models The features used in our approach encode patterns in customers’ session behaviour and transactions. These patterns vary widely from client to client, which is why we chose to use a client-specific model rather than a global model.

Global feature space Although behaviours vary, we have chosen to use the same feature representation for each session/transaction, which allows us to assign weights to specific (model, feature) pairs based on their performance across all clients. This in turn allows for consistent scoring across all clients and information sharing between clients when fraudulent activity occurs. In other words, our approach makes it easy for learners to adjust their weights.

Separation of models based on feature type We have chosen to form separate ensembles, one based on behavioural features and one based on transactional features, rather than concatenating all features into a single vector and forming a single ensemble based on concatenation. This ensures better interpretability and reduces the likelihood of constructing nonsensical feature subspaces during feature bagging.

Modified Bootstrap aggregation (Bagging) To build an ensemble of weak learners, we use a modification of bootstrap aggregation (bagging). Bagging is a meta-algorithm for ensembles that is used to reduce the variance and improve the stability of the prediction as well as to avoid overfitting.

Bagging Pipeline

Observational sampling (bagging): Bagged ensembles for classification generate additional data for training by resampling with replacement from the initial training data to produce multiple sets of the same size of initial training data, one for each base learner. This is done to reduce the prediction variance. For the two outlier detection ensembles, we used variable subsampling (without replacement) to avoid problems associated with repeated data and to mimic random selection of the neighbourhood count hyperparameter (cf. Aggarwal and Sathe 2017 ).

Feature bagging: An important task in outlier detection is to identify the appropriate features on which to base the analysis. However, these features may differ depending on the fraud mechanism. Therefore, instead of pre-selecting features, a more robust approach is to create an ensemble of models that focus on different feature sets and assign different weights to the models that use different features depending on their performance. The procedure is applied to each base learner \(b_j\) as follows:

Randomly select a number \(r_{b_j }\) in range \([ d/10,d-1]\) , where d denotes the feature dimension.

Sample a subspace of features of size \(r_{b_j }\)

Train the base learner \(b_j\) on the sampled subspace.

No Hyperparameter bagging: Due to limited fraud, tuning the hyperparameters via validation may lead to overfitting. For this reason, similar to feature selection, we could instead randomly select a set of different hyperparameters. In our case, however, IF and BDT are not expected to be sensitive to the choice of hyperparameters, and resampling the hyperparameter from LOF would be redundant to the subsampling of the data performed. We therefore set all hyperparameters to reasonable ones found in the literature.

Sharing bagged features and parameters across customers : Subspaces for sampled features and parameters for local outlier factors are shared between all client models and all types of base models (manifested in the respective weak learners). This allows, for example, the introduction of supervision in the aggregation step and increased interpretability of the model, as it is easier to identify features relevant to the detection of certain types of fraud.

Model aggregation: Each base model provides a decision function \(\delta (x)\) for a given observation x . The base model ensemble directly aggregates (majority voting or averaging) the weak learner results based on different subsamples to form a single hypothesis that determines the class membership of an observation. Usually, this aggregation is directly extended after a normalisation step to include models of different types, parameters or feature groups. In our approach, however, this final aggregation is performed based on a monitoring step that uses knowledge of available frauds to assign different weights to each pair (model, feature set). Essentially, these weights quantify how appropriate each model and feature pair is for fraud detection.

We refer to a composite Poisson process whose marginal distribution corresponds to a beta distribution as a beta model or as a GPD model if the marginal distribution corresponds to a generalised Pareto distribution. The mass attack model is a nested compound Poisson process. The outer Poisson process models the mass attack event, while an inner Poisson process models the number of affected transactions. The extent of damage of the individual affected transactions is modelled with a beta distribution.

Online banking:

Beta model: Intensity 35, \(\alpha =0.42, \beta =2.4\) and \(\text {scale}: 71.000.\) By shifting and scaling, explicitly by the transformation \(x\rightarrow \alpha + (\beta -\alpha )x\) , the beta distribution is shifted from [0, 1] to the interval \([\alpha ,\beta ]\) . The parameter \(\alpha\) is called location, and \(\beta -\alpha\) scale.

GPD model: Intensity 3, \(\text {shape}=0.25, \text {location}=60.000, \text {scale}=100.000\) .

Mass Attack model: Intensity 0.1, intensity nested model 1000, Beta model \(\alpha =0.42, \beta =2.4\) and \(\text {scale}: 20.000.\)

Mobile banking:

Beta model: Intensity 45, \(\alpha =0.42, \beta =2.4\) and \(\text {scale}: 20'000.\)

Mass Attack model: Intensity 0.1, intensity nested model 1000, Beta model \(\alpha =0.42, \beta =2.4\) and \(\text {scale}: 20'000.\)

GPD model: \(\text {shape}=0.25, \text {location}=100'000, \text { scale}=300'000\) .

Recovery model:

The recovery model models the percentage recovery in a fraud case. It has the following form:

With a probability \(p_1=65\%\) , a complete recovery is simulated, i.e. no damage remains. This resulted from the fact that in the 159 fraud cases considered, it was actually possible to reduce the loss amount to zero even in \(80\%\) of the cases.

With a probability \(p_2=18\%\) , a recovery of zero is simulated, i.e. the damage corresponds to the full amount of the offence.

With probability \(1-p_1-p_2\) , a recovery between 0 and 1 is simulated. A beta distribution is chosen as the distribution of these partial recoveries.

The beta distribution parameters for the online banking channel were fitted on the fraud cases recorded from 13/03/2013 to 13/03/2018. These are 159 fraud cases, of which both the initial fraud transaction amounts and the effective loss amount, i.e. the residual amount after recovery, were recorded. Of the 159 cases, 152 have a fraud amount between CHF 0 and 60,000, while the remaining 7 fraud amounts range between CHF 100,000 and 300,000. A beta distribution was fitted on the 152 cases with fraud amounts up to 60,000 CHF, whereby the scaling parameter, i.e. the upper limit of the distribution, was defined as a free parameter of the fitting procedure and estimated by it to be 71,000 CHF. Similar procedures apply to the marginal distribution fits of the GPD and mass attack models.

There exists significant statistical uncertainty and variability in the driving forces of the defined models. Putting the intended flexibility of the model structure into practice, we distinguish between ’easily accessible’ parameters, which should be subject to discussion at any time in the context of risk assessments, and ’deeper’ parameters, whose mode of action is less obvious and whose adjustment is subject to the process of model reviews. Roughly speaking, Poisson intensities, which determine the expected frequency of events, as well as upper and lower boundaries of the marginal distributions belong to the former category, while shape parameters for the Beta and GPD marginal distributions belong to the latter.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Vanini, P., Rossi, S., Zvizdic, E. et al. Online payment fraud: from anomaly detection to risk management. Financ Innov 9 , 66 (2023). https://doi.org/10.1186/s40854-023-00470-w

Download citation

Received : 22 March 2022

Accepted : 18 February 2023

Published : 13 March 2023

DOI : https://doi.org/10.1186/s40854-023-00470-w

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Payment fraud risk management
• Anomaly detection
• Ensemble models
• Integration of machine learning and statistical risk modelling
• Economic optimization machine learning outputs

We apologize for the inconvenience...

To ensure we keep this website safe, please can you confirm you are a human by ticking the box below.

If you are unable to complete the above request please contact us using the below link, providing a screenshot of your experience.

https://ioppublishing.org/contacts/

IEEE Account

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.