Modelos de classificação em fraudes financeiras: comparação de desempenho em casos de crime de smurfing

Abstract

The difficulty in identifying financial fraud is directly related to technological advances, as the new possibilities of forms of financial transactions, in turn, generate new forms of fraudulent agents to act. In this context, the aim of this study is to explore the theoretical construction of six machine learning (ML) models, in addition to comparing them through specific performance evaluation metrics. Furthermore, this work develops an algorithm to detect a type of financial crime known as smurfing. This algorithm does not use ML techniques, but aims to classify financial transactions as possible fraud through the analysis of pooled data. Given the impossibility of using real financial data, due to its confidentiality, this work is using simulated data. Two different scenarios were generated, both highly unbalanced, in which the behavior of financial fraud varies according to specific parameters. The chosen classification models were logistic model, Fuzzy Rule Based Systems, Artificial Neural Networks, Random Forest, Extreme Gradient Reinforcement and Support Vector Machine. The comparison of the models in the different scenarios was done through a combination of the metrics Area Under de Curve, Recall and Fb , once data are imbalanced. The results showed that the Random Forest and Extreme Gradient Boosting models had the best performances, therefore, it is believed that the use of such models in real data, even with different parameters, can help in tracking illegal financial transactions and identifying fraudsters