Previsão de fraude em licitações no Brasil

• Ano de defesa: 2024
• Autor(a) principal: Morais, Vinícius Souza
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Gestão Organizacional (Mestrado Profissional)
• Palavras-chave em Português: Licitação; Bidding; Licitação Pública; Public Procurement; e-Procurement; Fraudes; Fraud; Serviço Público; Public Service; Inteligência Artificial; Artificial Intelligence; Random Forest; CNPQ::CIENCIAS SOCIAIS APLICADAS; Administração; ODS::ODS 9. Indústria, Inovação e infraestrutura - Construir infraestrutura resiliente, promover a industrialização inclusiva e sustentável, e fomentar a inovação.; ODS::ODS 16. Paz, justiça e instituições eficazes - Promover sociedades pacíficas e inclusivas par ao desenvolvimento sustentável, proporcionar o acesso à justiça para todos e construir instituições eficazes, responsáveis e inclusivas em todos os níveis.
• Link de acesso: https://repositorio.ufu.br/handle/123456789/43695; http://doi.org/10.14393/ufu.di.2024.5149

Electronic procurement portals have transformed the process of acquiring goods and services, making it more efficient, competitive, and transparent. Simultaneously, the audit process must also keep up with these changes and adapt to this new system. In the academic context, there has been progress with the use of Artificial Intelligence techniques for fraud prediction, having this tool a wide range of applications in public management, particularly in monitoring procurement processes. Thus, this study proposes the use of the Random Forest technique for predicting signs of fraud in public administration procurements in Brazil. The Random Forest model was applied to the dataset related to 'Bidding and Contracts' of the year 2022, obtained from the Transparency Portal of the General Comptroller of the Union. Eleven independent variables (including budget, purpose of the bid, type of process, day the contract was signed, execution period, waiting period, interval between the signing date and the grant date, interval between the start date of execution and the signing date, distance to the next election, product category, state) were used to predict whether the company received fines (dependent variable). The performance of the model was evaluated monthly and annually. The classifier was effective in detecting fraud in procurements with an average monthly F1 Score of 78.5% and annual of 80%. It stood out for its ability to detect 90% (recall) of all actual cases of fraud in both analyses. In the monthly analysis, a significant reduction in the quality parameters of the irregularity cases was observed, with maximum values in January (recall=1.00 and F1 Score=0.96) and minimum in November (recall=0.47 and F1 Score=0.54). The most important variables fluctuated significantly over the months, but overall were: Product Category, Purpose of the Bid, and State. The potential impact of this work is highlighted not only in academia but also for all citizens and public managers, offering an effective tool in detecting potential misconduct in the public sector.