Métodos de classificação supervisionados aplicados à identificação de fraudes de fornecedores do Estado do Rio de Janeiro.

Abstract

Bidding fraud occurs by the attempt to frustrate its competitive character, causing overpricing and favoring a specific group. This research aims to verify which set of indicators and algorithms have best properties to identify the propensity of a supplier to have committed fraud against the public administration in the contracting process of the Government of the State of Rio de Janeiro. The delimitation of the research was the suppliers of legal entities of the Government of the State of Rio de Janeiro that had contracts or participated in bids between 2017 and 2021. Data from bids and suppliers published on the Transparency Portal of the Government of the State of Rio de Janeiro, and data from CNPJs published by the Brazilian Internal Revenue Service were used. The hypothesis test based on Benford’s Law demonstrated that the contracts database was not in compliance with the law, being prone to have had manipulation of the numbers. The proposed research showed satisfactory results for the classification of suppliers sanctioned by this state. The classification algorithm with the best result in the three metrics of evaluation was the Support Vector Machine (SVM), with an accuracy of 99.30%, and only one false positive in the verification of the test base. The attributes collected through the databases also proved to be relevant in the model result.