Detecção de Casos Suspeitos de Conluio em Licitações Públicas: uma Aplicação do Algoritmo a Priori de Aprendizado de Máquina para o Estado da Paraíba

Abstract

Purpose: This paper aims to identify potentially collusive bidding schemes in municipal government biddings of Paraíba state during 2005 to 2016. Methodology: For this purpose, data from the Court of Auditors of the State of Paraíba and the Apriori machine learning algorithm were used to build association rules in a sample of 104 thousand biddings, with 60 thousand bidders. The association rules were evaluated considering patterns of cooperative strategies in repeated games, taking account the bid winner likelihood and the bidders competition on average. How companies mapped in the association rules were ranked by the firm's suspicion indicator (ISE), directly related to the probability of victory, association with false tests and spatially concentrated performance, in which ISE is a contribution proposed by this research. Findings: The results revealed indications of suspected collusion for various companies in the meals suppliers, cleaning services, rental of vehicles, copiers, stages, chemical toilets, fuel suppliers, automotive parts, sports equipments, medicines and hospital material, consulting in accounting, civil engineering and advertising, considering each of the three municipal management periods (2005–2008, 2009–2012 and 2013–2016). We also find several patterns of association between bidders with evidence of competition simulation and concentrated activity in a few municipalities. Contribution: the results, particularly by the firm's suspicion indicator, provide important initial guidelines to optimize inspection and auditing processes in justice control bodies, whose undisputed identification of a potential bidding fraud, necessarily, requires complementary and specific procedural evaluations for each case.