| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 237.16 KB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
A fraude em seguros tem levado a perdas significativas, na ordem dos milhões de euros, pelo que se tem verificado um interesse crescente no estudo deste fenómeno.
A maioria das seguradoras continua a recorrer à inspeção manual, que é bastante complexa e demorada, por exigir especialistas para rever cada caso individualmente. O enorme número de pedidos de reembolso para analisar aliado à limitação da inspeção manual em detetar padrões de fraude emergentes, torna crucial o desenvolvimento de um Sistema de Deteção de Fraude robusto e eficaz.
Os modelos preditivos de deteção de fraude têm potencial para detetar casos suspeitos em tempo real e, consequentemente, reduzir significativamente as perdas económicas, tanto para as seguradoras como para os restantes segurados.
Contudo, a implementação destes modelos apresenta vários desafios, nomeadamente a distribuição enviesada de classes, que se caracteriza por existir uma proporção muito elevada de pedidos de reembolso legítimos face a pedidos de reembolso suspeitos. Este tipo de desafios exige a adoção de abordagens específicas para a sua resolução.
Neste estudo apresenta-se o desenvolvimento de modelos preditivos de deteção de fraude, na área de Multirriscos Habitação e Multirriscos Condomínio, com dados de uma seguradora portuguesa. Para esse efeito, usou-se um dataset constituído pelos pedidos de reembolso referentes aos anos de 2018 a 2022, que haviam sido classificados como legítimos ou suspeitos. Foram consideradas quatro famílias de classificadores – Árvores de Decisão, Regressão Logística, K-Nearest Neighbors e Random Forest, sendo os seus desempenhos medidos e comparados através de métricas adequadas.
Os resultados obtidos evidenciaram a utilidade destes classificadores, tendo as Árvores de Decisão apresentado melhores resultados.
O objetivo principal deste trabalho é contribuir para a mitigação dos riscos associados à fraude em seguros e, assim, fortalecer a integridade do setor de seguros como um todo.
Insurance fraud has led to significant losses amounting to millions of euros, which has sparked a growing interest in studying this phenomenon. Most insurance companies continue to rely on manual inspection, which is quite complex and time-consuming because it requires specialists to review each case individually. The sheer volume of refund requests to analyze, coupled with the limitations of manual inspection in detecting emerging fraud patterns, makes it crucial to develop a robust and effective Fraud Detection System. Predictive models for fraud detection have the potential to identify suspicious cases in real-time, thereby significantly reducing economic losses for both insurers and policyholders. However, implementing these models poses several challenges, including the biased class distribution, characterized by a very high proportion of legitimate refund requests compared to suspicious ones. Overcoming such challenges requires the adoption of specific approaches. This study presents the development of predictive fraud detection models using data from a Portuguese insurer. For this purpose, a dataset comprising refund requests from the years 2018 to 2022, classified as either legitimate or suspicious, was used. Four classifier families were considered: Decision Trees, Logistic Regression, K-Nearest Neighbors, and Random Forest, with their performances measured and compared using appropriate metrics. The results obtained underscored the usefulness of these classifiers, with Decision Trees yielding the best results. The primary objective of this work is to contribute to the mitigation of risks associated with insurance fraud, thus strengthening the integrity of the insurance sector.
Insurance fraud has led to significant losses amounting to millions of euros, which has sparked a growing interest in studying this phenomenon. Most insurance companies continue to rely on manual inspection, which is quite complex and time-consuming because it requires specialists to review each case individually. The sheer volume of refund requests to analyze, coupled with the limitations of manual inspection in detecting emerging fraud patterns, makes it crucial to develop a robust and effective Fraud Detection System. Predictive models for fraud detection have the potential to identify suspicious cases in real-time, thereby significantly reducing economic losses for both insurers and policyholders. However, implementing these models poses several challenges, including the biased class distribution, characterized by a very high proportion of legitimate refund requests compared to suspicious ones. Overcoming such challenges requires the adoption of specific approaches. This study presents the development of predictive fraud detection models using data from a Portuguese insurer. For this purpose, a dataset comprising refund requests from the years 2018 to 2022, classified as either legitimate or suspicious, was used. Four classifier families were considered: Decision Trees, Logistic Regression, K-Nearest Neighbors, and Random Forest, with their performances measured and compared using appropriate metrics. The results obtained underscored the usefulness of these classifiers, with Decision Trees yielding the best results. The primary objective of this work is to contribute to the mitigation of risks associated with insurance fraud, thus strengthening the integrity of the insurance sector.
Descrição
Trabalho de projeto de mestrado, Estatística e Investigação Operacional (Estatística), 2023, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
Fraude Multirriscos Habitação Multirriscos Condomínio Modelos Preditivos Machine Learning Trabalhos de projeto de mestrado - 2023
