Automated triage of financial intelligence reports

Abstract

The Federal Police of Brazil initiated many money-laundering investigations through Financial Intelligence Reports generated by the Financial Operations Control Board of Brazil. When they arrive at the Federal Police, these reports are first analyzed and then distributed among the country's police stations, which will have to decide whether to initiate a formal investigation or whether the reports will be filed as intelligence information. This triage is usually tricky and laborious, as it involves the analysis of diverse and complex information related to the operations, people and companies mentioned. As a result, the present research showed the possibility of applying artificial intelligence through machine learning to help this task. The implementation of the Random Forest classification algorithm and its application achieved a performance of 70% accuracy. The dataset consisted of 793 Financial Intelligence Reports, and the algorithm rated 522 reports as "archive as intelligence information" and 271 reports as "initiate a formal investigation". A feature selection process was carried out to achieve this accuracy, revealing that the sum of values, the amount of communications, and the number of people involved form the optimal subset of characteristics that best represents the classification model.