Integral representation method based efficient rule optimizing framework for anti-money laundering
| DOI | https://doi.org/10.1108/JMLC-12-2021-0137 |
| Published date | 18 April 2022 |
| Date | 18 April 2022 |
| Pages | 290-308 |
| Subject Matter | Accounting & finance,Financial risk/company failure,Financial compliance/regulation,Financial crime |
| Author | Tamás Badics,Dániel Hajtó,Kálmán Tornai,Levente Kiss,István Zoltán Reguly,István Pesti,Péter Sváb,György Cserey |
Integral representation method
based efficient rule optimizing
framework for anti-money
laundering
Tam
as Badics
Jedlik Innovation Ltd., Budapest, Hungary
D
aniel Hajt
o
Jedlik Innovation Ltd., Budapest, Hungary and
Faculty of Information Technology and Bionics,
P
azm
any Péter Catholic University, Budapest, Hungary
K
alm
an Tornai
Faculty of Information Technology and Bionics,
P
azm
any Péter Catholic University, Budapest, Hungary
Levente Kiss
Consortix Ltd., Budapest, Hungary
Istv
an Zolt
an Reguly
Faculty of Information Technology and Bionics,
P
azm
any Péter Catholic University, Budapest, Hungary
Istv
an Pesti and Péter Sv
ab
Consortix Ltd., Budapest, Hungary, and
György Cserey
Faculty of Information Technology and Bionics,
P
azm
any Péter Catholic University, Budapest, Hungary
Abstract
Purpose –This paper aims to introduce a framework for optimizing rule-based anti-money laundering
systemswith a clear economicinterpretation, and the authors introduce the integralrepresentation method.
Design/methodology/approach –By using a microeconomic model, the authors reformulate the
threshold optimization problemas a decision problem to gain insights from economics regarding the main
properties of the optimum. The authors used algorithmicconsiderations to find an efficient implementation
by using a kind of weak mode estimate of the distributionand the authors extend this approach to classes of
alerts or cases.
Findings –The method provides a new and efficient alternative for the sampling method or the
multidimensional optimization technique described in the literature to decrease the bias emanating from
multiple alertsby smoothing the number of alerts across classes in the optimumand decrease the overlapping
The authors gratefully acknowledge the support of grant 2018–1.1.1-MKI-2018–00145 of the
Hungarian National Research, Development and Innovation Office (NRDI Office).
JMLC
26,2
290
Journalof Money Laundering
Control
Vol.26 No. 2, 2023
pp. 290-308
© Emerald Publishing Limited
1368-5201
DOI 10.1108/JMLC-12-2021-0137
The current issue and full text archive of this journal is available on Emerald Insight at:
https://www.emerald.com/insight/1368-5201.htm
between scenariosat the case level. Using the method for real bank data, the authors were able to decrease the
number of falsepositives cases by about 18% while retaining almost 98% of the true-positive cases.
Research limitations/implications –The model assumes that alerts from different scenarios are
indifferentto the bank. To include scenario-specific preferences or constraintsdemands further research.
Originality/value –The new framework presentedin the paper is a flexible extension of the usual above-
the-line method,which makes it possible to include bank preferencesand use the parallelization capabilitiesof
modern processors.
Keywords Optimization, Anti-money laundering, SIMD parallelization, Threshold tuning,
Vectorization
Paper type Research paper
1. Introduction
1.1 Fundamental concepts and relationships to the literature
This article will introducea method that can be applied to optimize the parameters of a rule-
based anti-money laundering system.The fundamental business problem is to decrease the
volume of false-positive alerts without significantly decreasing the number of true-positive
alerts. Although there is an abundance of recent articles related to anti-money laundering,
most of them deal with machine learning as an alternative to the rule-based approach, and
only a very few papers deal with the problemof the threshold tuning of a rule-based system
(Chau and Nemcsik, 2020;Gupta et al.,2021;Herhold et al., 2017;Lucchetti, 2018). We
mention that there are some articles abouthybrid approaches that use the output of a rule-
based system in a machine learning framework or in an expert system (Khan et al., 2013;
Khanuja and Adane,2013, 2014;Rajput et al.,2014), but in this article, we confine ourselves
to the tuning process itself.
The method introduced in this article can be described as a statistical tuning of a rule-
based transaction monitoring system and is an improved version of above-the-line method
(Chau and Nemcsik, 2020;Lucchetti,2018). In rule-based systems, we assume that each rule
(also called as scenario) is a logical statementmade up of different elementary statements of
the form feature
i
>threshold
i
combined by any kind of logical operations. Here the feature
can be thought of as an aggregateamount, for example, the cumulative cash deposits over a
given period of time. Our method assumes that for every single alert, the values of these
aggregate amounts are saved into a database table, and this database table (which we call
the alert table) is available for optimization. We also assume that for each entry in thealert
table, the result of the investigation of the bank expert is also available in the form of a
positive or negative label. During the expert investigation, the individual alerts which are in
close relationship to eachother are grouped into so-called cases. A case can contain different
alerts over a more extended period of time, but if two alerts are in the same case, they get the
same label. A simple but widespread practice is that we do not bother taking into account
the cases and treat all of the alerts in the alert table as equivalent (Chau and Nemcsik,2020).
Another approach, described in Jullum et al. (2020), is also possible whenthe process of the
creation of the cases is part of a triage process; then, we can use the case status as a kind of
class label, and we can use multiclass classification methods in a machine learning context.
The third, more frequent approach uses only one element from each case for the
optimization. The problem withthis approach is that taking random sample results in such
representations of cases where they are representedby the typical elements instead of those
elements that distinguish the positivecases from the negative ones. In our approach, we use
all the alerts in the alert table for the optimization;at the same time, we take into account the
diminishing information value for each additional alert in the given case and the overlaps
Anti-money
laundering
291
To continue reading
Request your trial