Ciência de dados aplicada à gestão fiscal: um estudo sobre a inadimplência do Imposto Predial e Territorial Urbano em São Luís-MA

Abstract

The Urban Property Tax (IPTU) is one of the main sources of own-source revenue for Brazilian municipalities; however, its collection is often affected by high levels of tax delinquency and weaknesses in cadastral records. In the municipality of São Luís, there is a significant gap between the revenue potential and the amounts effectively collected. Within this context, this study investigates the factors associated with IPTU delinquency through the application of data science techniques and machine learning algorithms. Cadastral and financial IPTU data for the period from 2019 to 2023 were used, combining descriptive analyses, supervised modeling, and unsupervised methods. The initial diagnosis revealed structural delinquency, with approximately 70% of properties without tax exemptions remaining delinquent five years after the tax assessment, and about 45% of the assessed value not collected by the end of the statutory limitation period. In the supervised modeling stage, a multiclass classifier based on the XGBoost algorithm was applied, enabling the distinction between compliant, partially compliant, and delinquent properties, even in a context of strong class imbalance. The results indicate that tax delinquency is strongly associated with poor cadastral quality, particularly the absence of a valid CPF/CNPJ, as well as variables related to urban infrastructure, property construction characteristics, and spatial location. The clustering analysis using the K-prototypes algorithm revealed a high concentration of revenue potential within a small number of groups of delinquent properties, characterized by higher average IPTU values and greater operational feasibility for collection efforts. In addition, the identification of atypical patterns and cadastral inconsistencies demonstrated the potential of these approaches to support the improvement of property records and the prioritization of administrative actions. It is concluded that IPTU delinquency in São Luís is predominantly driven by structural, cadastral, and territorial factors. The findings show that the integrated use of machine learning techniques can significantly contribute to the improvement of municipal tax administration, promoting greater revenue efficiency and tax fairness without the need to increase tax rates.