Planning for Money Laundering Investigations

Abstract

According to the United Nations, the amount of money laundered worldwide each year is an estimated 2 - 5% of global GDP (equivalent to $800 billion to $2 trillion in US dollars). This is money that criminal enterprises rely on to oper- ate. For that reason, the European Union demands that gatekeepers (banks and other obliged entities) apply measures to counteract money laundering. Current industry state of the art anti-money laundering (AML) techniques ultimately revolve around investigations by specialized financial investigators of suspicious behaviour. Due to the human nature of this work, this process is relatively slow and has limited capacity. Deciding in the most optimal way what financial en- tities to investigate and when is not a trivial problem. However, optimizing this sequential decision making problem could significantly decrease the time-scale in which fraudulent actors are caught. This thesis will formulate the AML problem as a Partially Observable Markov Decision Problem. It will design and imple- ment an AML model and investigate the challenges associated with optimizing it. In particular, several Partially Observable Monte-Carlo Planning based methods are proposed that exploit the combinatorial structure of the actions to overcome the challenges associated with a large action space. The methods are empirically evaluated on the AML problem and compared to a baseline solution. The results indicate that exploiting the combinatorial structure increases the performance in this problem scenario. However, it seems that exploiting the structure to the highest degree does not always lead to the best performance. Additionally, we show that the proposed methods can match or even outperform the upper bound set by the baseline solution.