Using Machine Learning Algorithms to Signal Potential Anticompetitive Behavior By Firms in Federal Procurement Processes
*Names in bold indicate Presenter
Public procurement represents an important share of the Brazilian economy (about 8.2% of GDP (IMF Government Finance Statistics database, 2014)). Free and fair competition on public procurement can be impacted by the modality of bid and by the behavior of the stakeholders. The Brazilian public procurement legal framework provides six types of bidding, in addition to the possibility to waive the bidding process. According to data of the Purchasing Panel of the Federal Government, between 2012 and 2017, the government carried out 775,688 procurement processes, totalizing more than R$ 320 billion (US$ 99 billion). Those expenses resulted in a total contracted value of R$ 169,923 billion and 35,063 contracted companies. Furthermore, most of the purchasing processes, 49.67% in terms of monetary values, were carried out in the trading floor modality. However, considering the number of procurement processes performed, 79.67% of them were made due to the nonexistence of a bid.
2. Research goal
The objective of this work is to answer this question: How public managers could identify anticompetitive behavior by firms in federal procurement? In this way, it is expected to contribute to the improvement of the quality of public procurement procedures through the use of machine learning algorithms that can signal potential anti-competitive conducts by companies in bidding processes and to provide a unique perspective on this research topic. This research will be executed in partnership with the Department of Research and Strategic Information at the Brazilian Office of the General Comptroller (CGU), which uses machine learning techniques to audit public expenditures. The research will also present efforts by other Brazilian Institutions in order to enhance public spending quality.
This research combine administrative data from the Integrated General Services Administration System (SIASG), which is an electronic platform that contains all information about federal government procurement procedures. To complement this data, it will be used metadata provided by the Brazilian General Comptroller Officer and it is also intended to work with the Ministry of Planning, Development and Management in order to clarify some underlying questions that would remain.
This project is divided in three main stages. First, it is intended to apply unsupervised learning algorithms in order to try to identify key structures, or patterns, within prices practiced by the participating companies in bidding processes. In the second stage, it would be used supervised learning algorithms intending to make predictions or assessments about the behavior of companies participating in public procurement bidding based on the type and price of the product. Finally, it is intended to investigate the possibility and the applicability to use reinforcement learning algorithms to enhance the capacity of the system to provide ex-ante guidance for the public manager during the bidding process.