Alexandra Barysheva^a, Jevgenij Egle^a, Alexander Markov^a, Sergejs Solovjovs^a

^aEconophysica Ltd., Annecy Court, Ferry Works, Summer Road, Thames Ditton, Surrey KT7 0QJ, United Kingdom

Abstract

Motivating by the growing threat of fraud facing banks and payments firms, this short review considers presently most popular approaches to fraud detection in financial institutions including, e.g., credit card fraud, financial statement fraud, insurance fraud, money laundering, and terrorist financing.

Keywords: credit card fraud, data mining, fraud detection system, machine learning, money laundering, mortgage fraud, operational risk, securities and commodities fraud, suptech

1.     Introduction

Modern financial institutions face the growing threat of fraud in its different and ever-changing ways. For example, fraudulent transactions with cards could amount up to $44 billion by 2025 as shown in Figure 1.

Financial fraud is a broad term, which can be defined as «the intentional use of illegal methods or practices for the purpose of obtaining financial gain» [19], and also as «the use of one’s occupation for personal enrichment through the deliberate misuse or misapplication of the employing organization’s resources or assets» [1, p. 91], and finally as «any intentional or deliberate act of depriving another of property or money by cunning, deception or other unfair acts» [16, p. 46]. The presently most common fraud types are depicted in Figure 2 and include bank fraud, corporate fraud, and insurance fraud.

Fraudulent activities are part of operational risk, which is defined as «the risk of loss resulting from inadequate or failed internal processes, people and systems or from external events» [4, p. 762]. This definition includes legal risk but excludes strategic and reputational risk, where legal risk includes, but is not limited to, exposure to fines, penalties, or punitive damages resulting from supervisory actions, as well as private settlements. In particular, in the Basel classification of operational risk types [4, pp. 783–786], financial fraud is mostly subsumed under «Internal fraud», «External fraud», and «Clients, products, and business practices» (the latter includes, e.g., money laundering). Moreover, there exist general principles of managing operational risk, which state in particular that «risk management generally encompasses the process of identifying risks to the bank, measuring exposures to those risks (where possible), ensuring that an effective capital planning and monitoring programme is in place, monitoring risk exposures and corresponding capital needs on an ongoing basis, taking steps to control or mitigate risk exposures and reporting to senior management and the board on the bank’s risk exposures and capital positions» [3, p. 3]. All these recommendations though the amount to either calculation of regulatory capital to account for operational losses (using Basic Indicator, Standardised or Advanced Measurement Approach) or following fundamental principles of operational risk management (like, e.g., establishing three lines of defense: (1) business line management; (2) functionally independent corporate operational risk function; and (3) independent review and the challenge of the bank’s operational risk management controls, processes, and systems). To be more proactive, one needs methods for direct fraud detection and prevention. The purpose of this paper is to provide a brief review of the currently available methods and techniques. For the sake of shortness and clarity, we mostly concentrate on the fraud types mentioned in Figure 2.

2.     Intelligent fraud detection

This section provides a brief overview of the possible fraud detection methods and techniques used now and in recent history. We rely on several review papers, namely, [1, 16, 19] (mostly, however, on [19]), which conveniently list the available methods, making additionally their comprehensive comparison based on the results obtained in a number of research papers on the topic of fraud detection and prevention. 

2.1.     Types of financial fraud

We begin with a brief explanation of the considered fraud types following mostly the convenient exposition of [19] (which in turn is based, in particular, in the material of [12]). Moreover, [13] gives another good and general insight into the nature of some fraud types like, e.g., false financial disclosures and financial scams.

2.1.1.     Credit card fraud

Credit card fraud refers to the unauthorized use of a personal credit card to perform fraudulent transactions without this personal knowledge. The transactions can be performed using the physical card, where the card was either lost or stolen, as well as remotely. The cardholder information may be acquired through several methods. Phishing involves a fraudster impersonating a finance official to convince the user to disclose their details, swipers or skimmers provide an interface to an ATM or POS device, which can read the card directly, or entire databases of user information can be obtained provided that the fraudster is able to breach the respective financial institution's network security or get the help of an accomplice within the company. Obtaining a person's card could even be as simple as intercepting mail containing a new or replacement card. The anonymity and availability of these remote methods gave rise to the prevalence of organized crime in credit card fraud. A typical method for identifying credit card fraud is to analyze customer regular spending habits and flag transactions that are noticeably off the model.

2.1.2.     Mortgage fraud

Mortgage fraud is a specific form of financial fraud that refers to the manipulation of a property or mortgage documents. It is rather often committed to misrepresenting the value of a property for the purpose of influencing a lender to fund a loan for it.

2.1.3.     Money laundering

Money laundering is a method used by criminals to insert proceeds obtained from illicit ventures into valid businesses. This conceals the origin of the money, giving them the appearance of legitimate income and making it difficult to track their crimes. Money laundering is extremely undesirable since it enables criminals to have vast economic influence.

2.1.4.     Financial statement fraud

Financial statements are the documents released by a company explaining its details such as expenses, loans, income, and profits. They can also include comments from management on the business performance and expected issues that may arise in the future. The various financial statements that the company releases give an overall picture of the company status, and can be used to indicate how successful the company is, influence stock prices, and determine if the stocks are eligible for loans. Financial statement fraud, also known as corporate fraud, involves doctoring these statements to make the company appear more profitable.

Reasons for committing financial statement fraud include improving stock performance, reducing tax obligations, or exaggerating business performance due to managerial pressure. Financial statement fraud can be difficult to disclose due to a general lack of understanding in the field, infrequency of its occurrence, and the fact that it is often committed by experienced people in the industry who can disguise their deceit.

2.1.5.     Securities and commodities fraud

Securities fraud, also known as commodities fraud, refers to a variety of methods through which a person is deceived into investing in a company based on false information. It includes pyramid schemes, Ponzi schemes, hedge fund fraud, foreign exchange fraud, and embezzlement.

2.1.6.     Securities and commodities fraud

Insurance fraud is a fraud that can be committed at any point during the insurance process, and by any person(s) in the chain. Insurance claim fraud occurs when a customer submits a fraudulent insurance claim as a result of an exaggerated injury or loss of assets, or a completely fraudulent event. A common form of claim fraud is automobile insurance fraud, which is often committed by faking or intentionally committing accidents that result in excessive repair and injury costs. Larger-scale claim fraud also occurs as, for example, crop insurance fraud, where an insurance purchaser overstates the losses due to declining agricultural prices or the effects of natural disasters. Insurance fraud can additionally include excessive billing, duplicate claims, kickbacks to brokers, as well as «upcoding» of items.

2.1.7.     Synthetic identity fraud

According to [14, p. 2], synthetic identity fraud is the fastest-growing type of financial crime in the United States, accounting for 10−15% of charge-offs in a typical unsecured lending portfolio. In particular, synthetic identity fraud cost the US lenders $6 billion in 2016 [9, p. 3]. Following [10, pp. 2–3], synthetic identity is a scenario, where fraud perpetrators combine fragments of stolen or fake information to create a new identity and apply for financial products. All synthetic identity fraud forms, namely, traditional (a fusion of valid information from multiple real people), manipulated (all real information about a single person with a fake national ID), and manufactured (wholly fake information, including national ID), can exist only because of inadequate onboarding and customer due diligence. As stated in [14, p. 2], the largest (up to this moment) detected synthetic identity ring incurred losses for banks of $200 million from 7,000 synthetic identities and 25,000 credit cards.

2.1.8.     Corruption in public procurement

Certain types of corruption make a part of financial fraud, which is a long-standing problem. For example, the main aim of institutionalized grand corruption lies in earning corruption rents. Corruption rents in public procurement can be earned when the winning contractor is a pre-selected company that earns extra profit due to higher than market price for the delivered quantity and/or quality [8, p. 372].

The winning company has to be pre-selected in order to control rent extraction in an institutionalized manner. There has to be an extra profit to create the pot of money from which rents can be paid [8, p. 372].

2.2.     Fraud detection techniques

There exist various mathematical (e.g., statistical) techniques used to detect the above-mentioned as well as other types of fraud, some of which are considered in this section. It is important to emphasize that these techniques were not developed specifically to address numerous fraud issues but are just used as convenient existing tools. The techniques in question include Bayesian belief networks, logistic regressions or logistic models, neural networks, support vector machines, genetic algorithms and programming, decision trees, forests, and classification and regression trees (CART), group method of data handling, text mining, self-organizing maps, process mining, artificial immune systems, fuzzy logic, as well as hybrid methods, which collectively use several of the just mentioned approaches. A brief evaluation of these techniques is provided in Table 1. Moreover, Figure 3 shows fraud detection methods applicable to particular fraud types.

A number of metrics have been used to determine the performance of a particular method, where the three most commonly used ones are accuracy, sensitivity, and specificity. Accuracy measures the ratio of all successfully classified samples to unsuccessful ones. Sensitivity compares the number of items correctly identified as fraud to the amount incorrectly listed as fraud, also known as the ratio of true positives to false positives. Specificity refers to the same concept with legitimate transactions or the comparison of true negatives to false negatives. It is usually the case that computational intelligence (CI) methods perform better than classical statistical methods. Sensitivity is slightly better for random forests and support vector machines than logistic regression with comparable specificity and accuracy. Genetic programming, support vector machines, probabilistic neural networks, and group method of data handling outperform regression in all three areas. However, Bayesian belief networks can sometimes be more accurate than neural networks and decision trees. The comparative performance of various detection methods is provided in Figure 4. Finally, Table 2 shows the best method for each fraud type based on the accuracy measure.

To conclude, we notice that [8] used binary logistic regression and also linear regression to construct the corruption risk index of a public procurement contract, which claimed to give a lower bound estimate of «true» corruption. However, as stated in the paper itself, validating this index could take up to several years of work in multiple countries using many different indicators.

2.3.     Fraud detection systems

This section provides a brief outlook of fraud detection systems (FDS) and fraud prevention systems (FPS) as described in [1]. These systems rely on fraud detection techniques of the previous section and deal mostly with cyber-related fraud, that is, credit card fraud and insurance fraud. The basic fraud combating mechanism is shown in Figure 5. This mechanism consists of the above-mentioned two parts: FPS and FDS.

FPS is the first layer of protection to secure technological systems against fraud. The purpose of this phase is to stop fraud from occurring in the first place. Mechanisms in this phase restrict, suppress, destruct, destroy, control, remove, or prevent the occurrence of cyber-attacks in computer systems (hardware and software systems), networks, or data. Examples of such mechanisms include using an encryption algorithm that is applied to scramble data. Another mechanism is a firewall, which forms a blockade between the internal privately-owned network and external networks. It does not only help to secure systems from unauthorized access but also allows an organization to enforce a network security policy on traffic flowing between its network and the internet. However, this layer is not always efficient and strong. There are sufficiently many occasions when the prevention layer could be breached by fraudsters.

FDS is the next layer of protection. Fraud detection tries to discover and identify fraudulent activities as they enter the system and report them to a system administrator. Previously, manual fraud audit techniques such as discovery sampling have been used to detect fraud. These complicated and time-consuming techniques transact with various areas of knowledge like economics, finance, law, and business practices. Therefore, to raise the effectiveness of detection, computerized and automated FDS was invented. However, FDS capabilities were limited since the detection fundamentally depended on the predefined rules that were stated by experts. More complex FDSs integrating a wide range of data mining methods are required and are being developed for effective fraud detection. Data mining involves statistical, mathematical, artificial intelligence, and machine learning techniques to extract and identify useful information and subsequent knowledge from large databases (decision support systems and intelligent systems). These systems have a number of benefits: (1) obtaining fraud patterns automatically from data; (2) specification of «fraud likelihood» for each case so that efforts in investigating suspicious cases can be prioritized; and (3) revelation of new fraud types that were not defined before. Data mining methods consist of six main categories which are classification, clustering, regression, outlier detection, visualization, and prediction. Fraud detection now integrates anomaly-based and misuse-based approaches through using data mining techniques.

Anomaly or outlier detection approach relies on behavioral profiling methods, modelling each individual behavioral pattern, and monitoring it for any deviation from the norm. Anomaly-based FDSs have the potential to detect novel fraud. This method can be further categorized into the following three types: unsupervised, semi-supervised, and supervised anomaly detection.

Supervised learning techniques require a data set that has been labeled as «fraud» and «non-fraud» and involves training a classifier. Unsupervised learning techniques detect fraudulent activities in an unlabelled test data set under the assumption that the majority of the instances in the data set are non-fraud. Unlike supervised technique, unsupervised means that there is no class label for model construction. Semi-supervised learning lies between supervised and unsupervised learning since it involves a small number of labelled samples and a large number of unlabelled samples. The main goal of the semi-supervised approach is to train a classifier from both labelled and unlabelled data to improve the obtained machinery.

In the misused detection approach, fraudulent behaviors are first defined by using fraudsters signatures, and then other behaviours are defined as normal behaviours. The misused approach adopted by FDS utilizes rule-based, statistics, or corresponding heuristic methods to reveal an occurrence of a specific suspicious transaction.

There are several challenges underlying a successful fraud detection system, which are described below.

Concept drift: There exist several definitions for the concept drift issue in the literature. In data mining, concept drift refers to the phenomenon that the underlying model (or concept) is changing over time. FDSs work in a dynamic environment, where the behavior of a legitimate user or fraudster is continuously changing, which is called the drift phenomenon concept. For example, in the credit card area, the cardholder behaviour may be subject to change due to a variety of external causes, e.g., the transaction amount and frequency are closely related to the spending habits of a person, which is actually influenced by income, resource availability, and lifestyle of a person that may change with time. In addition, fraudster tricks are continuously evolving and detection has to adapt to these new fraud types.

Skewed class distribution: Skewed distribution (imbalanced class) is considered as one of the most critical issues faced by FDS. Generally, the imbalanced class issue is the situation where there are much fewer samples of fraudulent instances than normal instances. In a supervised learning approach, the class imbalance problem happens when the minority class (fraudulent transactions) is very small, leading to numerous problems such as disability of learners to discover patterns in the minority class data. Furthermore, the imbalanced class has a serious impact on the performance of classifiers that tend to be overwhelmed by the majority class (normal transactions) and ignore the minority class.

Reduction of a large amount of data: Large-scale and high dimensions of fraud data set and presence of a number of features/attributes/inputs/variables make the process of data mining and detection extremely difficult and complicated. Besides, this situation also slows down the detection process. Therefore, the existing FDSs use data reduction approaches to reduce the size of the data set, producing small model size which may be useful with respect to real-time processing. In addition, small data will reduce the size of the model, consequently reducing the computation time. Data reduction approaches include dimensionality reduction and reduction of the number of data points.

Possibility of real-time detection: Fraud detection systems work in two different modes, which are offline detection or online detection that is based on different fraud types. There are areas with real-time applications that require online fraud detection. For example, fraud in the online payment application in the credit card area needs immediate detection and response. In contrast, there are applications that require offline detection. Online fraud detection should be able to deal with limited resources (time and memory) in ensuring that the detection process works efficiently. Therefore, the efficiency of any proposed online fraud detection solution does not only benefit from the reduced amount of data but also from the reduced computational complexity of methods used for detection.

To conclude the section, we notice that, as of recent, machine learning has become one of the most prominent techniques in detecting financial fraud. A convenient account of the presently used methods can be found in, e.g., [16], which emphasizes that hybrid fraud detection techniques are the most used ones since they combine the strength of several traditional detection methods. In addition, it rightly appears that there is no single strategy valid for all fraud types since every fraud has its specific properties to address. However, as stated in [19], given the diversity of common categories of fraud, it would be useful to have some form of generic framework that could apply to more than one fraud category. Such a framework could be used to study the differences between various types of fraud, or even specific details of a particular fraud type such as differentiating between stolen and counterfeited credit cards.

As stated in [14, p. 2] the sophisticated technology that helps detect other types of fraud is not of much assistance for synthetic identity fraud. Machine learning techniques such as deep neural networks that find patterns associated with fraud are of little use since the small number of cases of synthetic identity fraud have been found to train models. Unsupervised machine learning techniques that look for anomalies in data also do not cope, since there are few, if any, differences between real and synthetic identities at the time of application. However, an approach to identifying synthetic identities that entails leveraging third-party data and which is grounded in the fact that real people have real histories, evidence of which they scatter behind them in dozens of different data systems, physical and digital is described in [14]. By evaluating the depth and consistency of information available about applicants in third-party data systems, institutions can determine whether the applicants are real or not [14, p. 3].

3.     Money laundering and terrorism financing combating techniques

In this section, we briefly discuss the problems of money laundering (ML) and the financing of terrorism (FT). As stated in, e.g., [6, p. 3], one estimate of ML/FT compliance cost by financial firms put it at $25.3 billion per year for US financial services firms alone. ML/FT authorities are also devoting substantial resources to fight ML/TF to the extent possible. However, many of them, particularly those in emerging markets and developing economies, have severe capacity constraints. Despite these efforts, estimates of money laundered worldwide are still staggering. In addition, there seems to be slow progress globally in reducing ML/TF risks. This has wide-ranging economic and social repercussions. The Financial Stability Board (FSB), for example, noted that jurisdictions most frequently exited by global correspondent banks seem to be those with weak ML/FT supervisory and regulatory frameworks. This has significant implications for international trade and remittance flows to these jurisdictions.

ML/FT authorities typically receive substantial amounts of transactional data from reporting institutions. These are then supplemented by non-transactional data (e.g., data from tax, customs, and property registration authorities) to provide context. Some ML/FT authorities are now actively collaborating with other government agencies and private entities to expand the scope of data available to them. Some authorities are also exploring the use of non-traditional sources of information (e.g., newspaper articles, social media) and integrating them with traditional information to come up with richer analyses. Big data analytics enables the integration of all this information from different sources for a coherent story [6, p. 3].

ML/FT authorities in general pursuе similar advanced data analytics tools, such as network analysis, natural language processing, text mining, and machine learning. These tools increase their ability to detect networks of related transactions, to identify unusual behaviours, and in general, to transform significant amounts of structured and unstructured data into useful information that contributes to their respective processes [6, p. 1]. A number of methods used currently by different institutions can be found in, e.g., [5].

Financial institutions traditionally relied on rule-based ML/FT measures, which are shown to generate false positive alerts of around 90−95% leading to substantial resource implications. So they are now increasingly using machine learning applications to reduce the number of false positives. The same improvements could be expected from the use of these advanced data analytics tools by ML/FT authorities [6, pp. 3–4]. The use of advanced data collection by financial authorities and analytics tools enabled by new technologies is collectively called suptech. These tools give opportunities to enhance financial authorities’ capacity [6, p. 1].

Figures 6, 7, 8 show several examples of institutional techniques in combating fraudulent activities. Notice that «R2A» stands for «Regtech for Regulators Accelerator», «FIU» stands for «financial intelligence unit», and «STR» stands for «suspicious transaction report».

One should also notice that the process of proving that a money laundering activity has occurred takes time. Once an FIU detects that a suspicious activity warrants more investigation, it passes this information on to the law enforcement agency. The law enforcement agency, in turn, refers the case to the court if it deems that it warrants prosecution. Court proceedings can be protracted so that it could take years to determine whether a particular tool has correctly identified ML [6, p. 14].

Finally, we notice that [18] provides an overview of some methods of machine learning used to detect fraud in general and ML/TF in particular (saying, for example, that «in recent years, machine learning and artificial intelligence have seen increasing interest and popularity in the financial services community»). In particular, [18, p. 66] states that «Machine-learning systems have the potential to improve detection of money laundering activity significantly, due to their ability to identify complex patterns in the data and combine transactions information at network speed, with data from many other sources to obtain a holistic picture of a client’s activity. Indeed, these systems have already been shown to bring false positives down significantly.» Despite this high and promising potential though, when it comes to ML/TF detection, «many banks are still relying on conventional rule-based systems, which focus on individual transactions or simple transaction patterns. These systems are often unable to detect complex patterns of transactions or obtain a holistic view of transaction behavior on payment infrastructures. Due to their coarse selection methods, the number of false positives created by these systems is substantial. As a result, significant human capacity is required for the assessment of alerts and filtering false positives from actual suspicious observations.» [18, p. 65] Lastly, some methods of data analytics related to ML/TF can also be found in a review paper [11].

 

3.1.     Convergence of anti-ML (AML) and fraud functions (FRAML)

A typical financial organisation of any size has both fraud detection and compliance units. These units require much of the same information and both struggle with a legacy of multiple-point solutions and data inconsistency embedded into core business systems. Such a stance allows them neither to share information nor to have a full picture of the customer.

Data is captured in multiple applications and data repositories which are often imperfectly synchronised. Numerous systems are usually managed and owned by different vendors or separate parts of the same organisation resulting in multiple teams with their own skill sets that are not transferrable [7]. Such an approach is costly to maintain and it has limited value.

Taking a more holistic approach to handling feeds from all associated systems will allow developing data-sharing platforms, with the goal to combine financial fraud detection efforts with AML and other financial crime detection efforts among collaborating stakeholders in order to bring together the teams, processes, and solutions they use and make better decisions with an appropriate contextual view and value insights.

The following are the main advantages of taking a holistic approach to FRAML [15, 17]: 

• a single customer view improves discovery; 
• advanced analytics for fraud detection can be applied to AML, and vice versa;
• cost savings due to the elimination of duplicate alert reviews and case investigations of common subjects and elimination of re-keying data. 
  

4.     Conclusion

This short review paper tried to provide an account of fraud detection and prevention techniques presently available in the research literature. Following [19], we conclude our review with the most important challenges arising from combating financial fraud.

• Typical classification problems. Computational intelligence and data mining-based financial fraud detection are subject to the same issues as other classification problems, such as feature selection, parameter tuning, and analysis of the problem domain.
• Fraud types and detection methods. Financial fraud is a diverse field and there has been a large imbalance in both fraud types and detection methods studied: some have been studied extensively while others, such as hybrid methods, have only been looked at superficially.
• Privacy considerations. Financial fraud is a sensitive topic and stakeholders are reluctant to share information on the subject. This has led to experimental issues such as undersampling.
• Computational performance. As a high-cost problem, it is desirable for financial fraud to be detected immediately. Very little research has been conducted on the computational performance of fraud detection methods to be used in real-time situations.
• Evolving problem. Fraudsters are continually modifying their techniques to remain undetected, which means such detection methods are required to be able to constantly adapt to new fraud techniques.

• Disproportionate misclassification costs. Fraud detection is primarily a classification problem with a vast difference in misclassification costs. Research on the performance of detection methods with respect to this factor is an area that needs further attention.
• Generic framework. Given that there are many varieties of fraud, a generic framework that can be applied to multiple fraud categories would be valuable.
  

References

[1] A. Abdallah, M. A. Maarof, and A. Zainal, Fraud detection system: a survey, Journal of Network and Computer Applications 68 (2016), 90–113.

[2] L. Anan, R. Hayden, K. Joshi, M.-C. Nadeau, and J. Steitz, Fraud management: recovering value through next-generation solutions, McKinsey & Company, McKinsey on Payments, Volume 11, Number 27, pp. 30–36, June 2018.

[3] BCBS, Principles for the sound management of operational risk, Bank for International Settlements, June 2011.

[4] BCBS, The Basel framework, Bank for International Settlements, 2020.

[5] D. Broeders and J. Prenio, Innovative technology in financial supervision (suptech) { the experience of early users, Bank for International Settlements, Financial Stability Institute, FSI Insights on Policy Implementation No. 9, July 2018.

[6] R. Coelho, M. De Simoni, and J. Prenio, Suptech applications for anti-money laundering, Bank for International Settlements, Financial Stability Institute, FSI Insights on Policy Implementation No. 18, August 2019.

[7] M. Cox, Convergence of fraud and compliance in fight against financial crime, Raconteur, Special Report, September 2018, available online at https://www.raconteur.net/sponsored/convergence-of-fraud-and-compliance-in-fightagainst-financial-cr....

[8] M. Fazekas, I. J. T´oth, and L. P. King, An objective corruption risk index using public procurement data, European Journal on Criminal Policy and Research 22 (2016), no. 3, 369–397.

[9] FED, Synthetic identity fraud in the U.S. payment system: a review of causes and contributing factors, The Federal Reserve, Payments Fraud Insights, July 2019.

[10] S. Hasham, R. Hayden, and R. Wavra, Combating payments fraud and enhancing customer experience, McKinsey & Company, McKinsey on Payments, Volume 11, Number 28, pp. 1–8, August 2018.

[11] O. Kraft, Sharpening the Money-Laundering Risk Picture: How Data Analytics Can Support Financial Intelligence, Supervision and Enforcement, Royal United Services Institute for Defence and Security Studies, Occasional Paper, November 2018.

[12] E. W. T. Ngai, Y. Hu, Y. H. Wong, Y. Chen, and X. Sun, The application of data mining techniques in financial fraud detection: a classification framework and an academic review of literature, Decision Support Systems 50 (2011), no. 3, 559–569.

[13] A. Reurink, Financial fraud: a literature review, Journal of Economic Surveys 32 (2018), no. 5, 1292–1325.

[14] B. Richardson and D. Waldron, Fighting back against synthetic identity fraud, McKinsey & Company, Risk Practice, January 2019.

[15] S. Rutherford, What is FRAML? And why is it important?, Fico Blog, Fraud Protection & Compliance, September 2018, available online at https://www.fico.com/blogs/what-framl-and-why-it-important.

[16] I. Sadgali, N. Sael, and F. Benabbou, Performance of machine learning techniques in the detection of financial frauds, Procedia Computer Science 148 (2019), 45–54.

[17] D. Stewart and D. Barta, Financial crimes convergence - the case for integrating fraud and anti-money laundering processes, ACAMS Today, September 2011, available online at https://www.acamstoday.org/case-for-integratingfraud-and-anti-money-laundering-processes.

[18] B. van Liebergen, Machine learning: a revolution in risk management and compliance?, The Capco Iinstitute Journal of Financial Transformation, available online at https://www.iif.com/portals/0/Files/private/32370132_van_liebergen_- _machine_learning_in_compliance_risk_management.pdf.

[19] J. West and M. Bhattacharya, Intelligent financial fraud detection: a comprehensive review, Computers & Security 57 (2016), 47–66.