Machine learning methods for anti-money laundering monitoring
Kuupäev
2023-11-02
Autorid
Ajakirja pealkiri
Ajakirja ISSN
Köite pealkiri
Kirjastaja
Abstrakt
Rahapesu (RP) kujutab endast märkimisväärset ohtu ülemaailmsetele finantssüsteemidele, võimaldades kurjategijatel varjata raha ebaseaduslikku päritolu ja integreerida need seaduslikku majandusse. Sellel ei ole mitte ainult rahalised tagajärjed, vaid see õõnestab ka finantssüsteemide stabiilsust, ohustab riigi julgeolekut ja kahandab üldsuse usaldust finantsasutuste vastu. Valitsused ja õiguskaitseasutused kogu maailmas on mures salakuritegevuse tuvastamise pärast. Finantsasutused kasutavad omakorda mitmesuguseid seiremehhanisme, et tuvastada ja teavitada võimalikust rahapesust. Need süsteemid järgivad tavaliselt lihtsaid reegleid, kuid neil on keeruliste ja uute RP-skeemide tuvastamisel piirangud. Masinõppe algoritmide – algoritmid, mis kasutavad otsuse tegemiseks märgistatud ajaloolisi andmeid – kasutamine RP tuvastamise kontekstis võib oluliselt parandada olemasolevate süsteemide tõhusust.
Selle lõputöö eesmärk on luua lahendus, mis kombineerib erinevaid raamistikke, et tuvastada RP automaatselt masinõppe abil. Sellise lahenduse väljatöötamisel on aga palju väljakutseid. Digitaalsete maksete ja ülemaailmsete tehingute kasv on toonud analüüsimiseks tohutul hulgal andmeid. Erinevad finantstooted muudavad rahapesu tuvastamise veelgi raskemaks, kuna kurjategijad võivad kasutada nende toodete erinevaid kombinatsioone. Lisaks on RP väga haruldane sündmus, mis raskendab tõhusate masinõppemudelite väljatöötamist. Lõpuks muutuvad RP-skeemid pidevalt, nõudes lähenemisviisi regulaarset värskendamist ja kohandamist.
Lõputöö annab sellesse uurimisvaldkonda neli peamist panust: (i) raamistik, mis tuvastab RP individuaalse kliendi tasandil; (ii) raamistik, mis tuvastab RP grupi kliendi tasandil; (iii) süsteem, mis määrab kindlaks, millal teavitada tuvastatud rahapesust tingitud käitumist, mida hiljem töötlevad spetsialiseerunud eksperdid; (iv) süsteem, mis annab tekstilisi selgitusi nende hoiatusteadete esitamise põhjuste kohta. Need panused koos moodustavad tervikliku lahenduse automaatseks RP tuvastamiseks, mis vastab olulistele ärinõuetele. Lahendust on testitud tegelike andmete põhjal koos kliendiprofiilide, tehingute ajaloo ja rahapesuvastaste ekspertide sisendiga. Tulemusi hinnati arvutuslike katsete ja domeeniekspertide tagasiside kaudu.
Money laundering (ML) poses a significant threat to global financial systems, enabling criminals to disguise the illicit origins of funds and integrate them into the legitimate economy. It not only has financial consequences but also undermines the stability of financial systems, threatens national security, and erodes public trust in financial institutions. Governments and law enforcement agencies worldwide are concerned about identifying ML activities. In turn, financial institutions deploy a variety of monitoring mechanisms to detect and report potential ML activities. These systems usually follow simple rules, but they have limitations in detecting complex and new ML schemes. Usage of machine learning algorithms – algorithms which use labelled historical data to form a decision – in the context of ML detection can significantly improve the effectiveness of the existing systems. The goal of this thesis is to create a solution that combines different frameworks to automatically detect ML using machine learning. However, there are many challenges in developing such a solution. The increase in digital payments and global transactions has generated a huge amount of data to analyse. Different financial products make it even harder to detect ML because criminals can use various combinations of those products. Moreover, ML is a very rare event, which makes it difficult to develop effective machine learning models. Finally, ML schemes are constantly changing, requiring regular updates and adjustments of the approach. The thesis makes four main contributions to this research area: (i) a framework that detects ML on individual customer level; (ii) a framework that detects ML on group customer level; (iii) a system that defines when to raise alerts for detected ML behaviour, which later are processed by specialized experts; (iv) a system that provides textual explanations on why those alerts were raised. These contributions together form a comprehensive solution for automated ML detection that meets important business requirements. The solution has been tested on real-life data with customer profiles, transaction histories, and input from anti-money laundering experts. The results were evaluated through computational experiments and feedback from domain experts.
Money laundering (ML) poses a significant threat to global financial systems, enabling criminals to disguise the illicit origins of funds and integrate them into the legitimate economy. It not only has financial consequences but also undermines the stability of financial systems, threatens national security, and erodes public trust in financial institutions. Governments and law enforcement agencies worldwide are concerned about identifying ML activities. In turn, financial institutions deploy a variety of monitoring mechanisms to detect and report potential ML activities. These systems usually follow simple rules, but they have limitations in detecting complex and new ML schemes. Usage of machine learning algorithms – algorithms which use labelled historical data to form a decision – in the context of ML detection can significantly improve the effectiveness of the existing systems. The goal of this thesis is to create a solution that combines different frameworks to automatically detect ML using machine learning. However, there are many challenges in developing such a solution. The increase in digital payments and global transactions has generated a huge amount of data to analyse. Different financial products make it even harder to detect ML because criminals can use various combinations of those products. Moreover, ML is a very rare event, which makes it difficult to develop effective machine learning models. Finally, ML schemes are constantly changing, requiring regular updates and adjustments of the approach. The thesis makes four main contributions to this research area: (i) a framework that detects ML on individual customer level; (ii) a framework that detects ML on group customer level; (iii) a system that defines when to raise alerts for detected ML behaviour, which later are processed by specialized experts; (iv) a system that provides textual explanations on why those alerts were raised. These contributions together form a comprehensive solution for automated ML detection that meets important business requirements. The solution has been tested on real-life data with customer profiles, transaction histories, and input from anti-money laundering experts. The results were evaluated through computational experiments and feedback from domain experts.
Kirjeldus
Märksõnad
machine learning, methods, money laundering, monitoring