| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 3.98 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
As últimas décadas foram marcadas por dois fatos importantes: organizações de combate ao crime de Branqueamento de Capitais (BC) foram criadas; e os Bancos Centrais, normalmente responsáveis pelo controle e definição de normas, ampliaram as leis de Anti-Branqueamento de Capitais (ABC). A agilidade na adaptação do modus operandi dos fraudadores e a falta de informação sistematizada que associe os clientes suspeitos comunicados com a comprovação do crime (classificado neste trabalho como um risco sistémico) são fatores que dificultam a automatização do processo de ABC e pode explicar as raras publicações com soluções inovadoras. A abordagem apresentada nesta tese modifica o tratamento genérico utilizado pela maioria dos trabalhos publicados, indo além da sinalização de transações suspeitas, auxiliando o Analista de ABC na tomada de decisão. Esta tese apresenta uma forma inovadora de integração de processos de aprendizagem, numa perspetiva de grupos de risco, com agentes auxiliando a análise e a tomada de decisão, resultando na implementação do sistema multi-agente denominado Jano, que levou a uma melhoria clara nos resultados relativos à identificação e sinalização de clientes suspeitos de BC. Para cada cliente foi gerado um perfil representando seu padrão de comportamento transacional, permitindo a adoção de uma abordagem que identifica e classifica o nível de risco de BC de cada perfil. Os dados utilizados referem-se a dois anos de movimentações, um com 30 e outro com 32 milhões de transações relevantes.
De modo a avaliar a relevância das propostas apresentadas, seis meses de transações foram utilizadas para o teste final, um conjunto dos perfis sinalizados foi submetido aos Analisas de ABC da instituição financeira (IF) que financiou a pesquisa e o resultado foi comparado com outros sistemas em uso naquela IF. Os números obtidos nas métricas F1-score e Matthews Correlation Coefficient permitem concluir que com a metodologia proposta nesta tese foi possível obter resultados melhores, destacando que 76% dos perfis confirmados como suspeitos não foram reportados ao órgão regulador na época da sua ocorrência, porque nenhum sistema em utilização na IF sinalizou-os como suspeitos.
The last decades have been marked by two important facts: organizations to combat Money Laundering (ML) have been created, and Central Banks, usually responsible for control and definition of norms, have expanded Anti-Money Laundering (AML) laws. The agility in the adaptation of the fraudsters’ modus operandi and the lack of systematic information that associates the suspicious clients communicated with the proof of the crime (classified in this work as a systemic risk), are factors that hinder the automation of the AML process and can explain the rare publications with innovative solutions. The approach presented in this thesis modifies the generic treatment used by most published works, going beyond the signaling of suspicious transactions, assisting the AML Analyst in decision making. This thesis presents an innovative way of integrating learning processes, from a risk groups perspective, with agents helping the analysis and decision making, resulting in the implementation of the multi-agent system called Jano, which has led to a clear improvement in the results related to the identification and signaling of ML suspicious clients. For each client, a profile representing their transactional behavior pattern has been generated, allowing the adoption of an approach that identifies and classifies the ML risk level of each profile. The data used refer to two years of transactions, with an average of 32 million relevant transactions. To evaluate the relevance of the proposals presented, six months of transactions were used for the final test, a set of signaled profiles was submitted to the AML Analyzes of the financial institution (FI) that funded the research and the result was compared with other systems in use in that FI. The numbers obtained in the F1-score and Matthews Correlation Coefficient metrics allow us to conclude that with the methodology proposed in this thesis it was possible to obtain better results, highlighting that 76% of the profiles confirmed as suspicious were not reported to the regulator at the time of their occurrence, because no system in use at the FI flagged them as suspicious.
The last decades have been marked by two important facts: organizations to combat Money Laundering (ML) have been created, and Central Banks, usually responsible for control and definition of norms, have expanded Anti-Money Laundering (AML) laws. The agility in the adaptation of the fraudsters’ modus operandi and the lack of systematic information that associates the suspicious clients communicated with the proof of the crime (classified in this work as a systemic risk), are factors that hinder the automation of the AML process and can explain the rare publications with innovative solutions. The approach presented in this thesis modifies the generic treatment used by most published works, going beyond the signaling of suspicious transactions, assisting the AML Analyst in decision making. This thesis presents an innovative way of integrating learning processes, from a risk groups perspective, with agents helping the analysis and decision making, resulting in the implementation of the multi-agent system called Jano, which has led to a clear improvement in the results related to the identification and signaling of ML suspicious clients. For each client, a profile representing their transactional behavior pattern has been generated, allowing the adoption of an approach that identifies and classifies the ML risk level of each profile. The data used refer to two years of transactions, with an average of 32 million relevant transactions. To evaluate the relevance of the proposals presented, six months of transactions were used for the final test, a set of signaled profiles was submitted to the AML Analyzes of the financial institution (FI) that funded the research and the result was compared with other systems in use in that FI. The numbers obtained in the F1-score and Matthews Correlation Coefficient metrics allow us to conclude that with the methodology proposed in this thesis it was possible to obtain better results, highlighting that 76% of the profiles confirmed as suspicious were not reported to the regulator at the time of their occurrence, because no system in use at the FI flagged them as suspicious.
Description
Keywords
Anti-Branqueamento de Capitais Sistema Multi-agente Agentes Inteligentes Aprendizado de Máquina Avaliação de Risco Anti-Money Laundering Multi-agent System Intelligent Agents Machine Learning Risk Evaluation