Detection of Structured Fraud Supported by Shell Companies on Goods and Services Trading Operations

Mitigating tax evasion in goods and services trading operations has been one of top priorities of local governments for a long time – as they directly impact the budget available for services offered to citizens. As tax compliance mechanisms improve, however, fraud attempts also adapt and become increasingly structured to overcome them and continue to generate damage to the treasury. An increasingly recurring kind of fraud involves the use of shell companies to issue ‘cold’ invoices – which simulate an operation that never actually takes place – and, thus, simulate revenue or reduce due tax at the time of tax assessment. This work proposes the use of machine learning – more specifically, supervised and semi-supervised learning techniques – on identifying groups of companies suspected of operating exclusively in issuing of cold invoices, allowing them to be early audited by a specialist and their operation to be interrupted as soon as possible. Through two cycles of experiments, based on different labeling possibilities, we were able to achieve relevant results – mainly in the most relevant metric, Recall – even though datasets are usually widely unbalanced when dealing with tax evasion.