Formação inteligente do preço de referência em licitações brasileiras

Abstract

Context: The absence of detailed regulatory guidelines for the formation of the reference price generates inaccurate estimates that are vulnerable to market distortions. Objective: To demonstrate that the integration of classical statistical metrics and a large-scale trained logistic classifier reduces the deviation between calculated reference prices, in addition to identifying overpricing and underpricing with greater precision. Method: Legislation and regulatory instruments were analyzed, 4,875 purchase records were collected via open data, and a hexagonal architecture software was developed that applies a decision tree based on the coefficient of variation and Benford’s Law, complemented by logistic regression via machine learning, with analysis of 900 thousand bidding records. Results: The solution achieved a Mean Average Percentage Error (MAPE) of 2.53% against 15.58% of official estimates and obtained an accuracy of 70% in detecting overpricing or underpricing, in addition to representing a smaller deviation in relation to the median of the Price Panel in the three versions of the solution, compared to the reference prices estimated by the agencies themselves. Conclusions: The platform demonstrates the feasibility of increasing accuracy and transparency in the bidding process through the integration of advanced statistics and machine learning, offering a solid basis for adoption in public agencies and future extensions with machine learning techniques.