Machine Learning Based Risk Scoring for Money Mule Detection

Abstract

Money laundering is a serious problem for the financial systems, as it helps criminals hide illegal funds and damages trust in banks. Money mules are individuals who transfer money on behalf of others, often unknowingly, and play a central role in these activities. Detecting money mules is a major challenge for financial institutions. Traditional detection methods rely on fixed rules that are not flexible enough to keep up with changing criminal tactics. This thesis explores the application of machine learning models for detecting money mules by assigning client-level risk scores based on transaction habits, product usage, and personal information. Multiple models were evaluated, with XGBoost demonstrating the highest area under the precision–recall curve (AUPRC) of 0.1314, indicating strong performance in this highly imbalanced setting. The model shows potential to detect over half of known money mules while maintaining a manageable false positive rate. These findings suggest that integrating machine learning into anti-money laundering systems can improve anti-money laundering efforts, helping banks prevent criminal activity, protect their customers and maintain their reputation. Rahapesu on tõsine probleem finantsüsteemides, kuna see aitab kurjategijatel varjata ebaseaduslikke summasid ja kahjustab pankade usaldusväärsust. Rahamuulad on isikud, kes tihti eneselegi teadmata kannavad raha teiste eest ja mängivad rahapesus keskset rolli. Rahamuulade tuvastamine on finantsasutustele suur väljakutse. Traditsioonilised tuvastamismeetodid tuginevad kindlatele reeglitele, mis ei ole piisavalt paindlikud, et kohaneda muutuvate kuritegelike taktikatega. Käesolev magistritöö uurib masinõppe mudelite rakendamist rahamuulade tuvastamiseks, määrates klientidele riskiskoori nende tehinguharjumuste, toodete kasutamise ja isikuandmete põhjal. Hinnati mitmeid mudeleid, millest parima tulemuse andis XGBoost, saavutades täpsuse-saagis kõvera aluse pindala (AUPRC) väärtuseks 0,1314. See on hea tulemus arvestades tasakaalustamata andmestikus. Mudel tuvastas üle poole teadaolevatest rahamuuladest, hoides samal ajal valepositiivsete määrad mõistlikul tasemel. Töö tulemused viitavad, et masinõppe integreerimine rahapesuvastastesse süsteemidesse võib aitada kaasa kuritegude ennetamisele, aidata pankadel kaitsta oma kliente ning säilitada usaldusväärsust.