Avaliação de risco de irregularidades em processos de aposentadoria no âmbito do Tribunal de Contas do Estado da Paraíba: uma aplicação baseada em aprendizagem de máquina supervisionada

Abstract

With the accelerated increase in the share of public budgets allocated to retirees each year, the efficient management of resources spent by retirement funds on this type of social benefit has become increasingly strategic. This study addresses the growing demand for audits in retirement benefit approval processes and the limitations of available human resources to meet this demand. Using historical and structured data related to retirement processes submitted to the State Court of Auditors of Paraíba (TCE/PB) by pension institutes, the research seeks to evaluate whether a model can automatically and preliminarily identify the intrinsic risk of irregularities in retirement processes, thereby enabling the prioritization of cases for analysis by the Court. To achieve this, eleven different binary classification models are trained, with the test dataset segmented according to the degree of consensus among the models. For each process, in addition to the individual classification derived from each model, the collective choice derived from the majority classification of the 11 models was also considered. The performance evaluation of the models and the collective choice was guided by a benefit-cost metric, which seeks to simultaneously weigh the efficiency gains from correctly classifying low-risk processes (true negatives) and the losses from incorrectly classifying processes that, in reality, carry risk (false negatives). The evaluation based on this metric indicates that, for processes where 9 or more models reached consensus, automation demonstrates superior performance compared to the purely random selection performed by the Audit. Conversely, for processes where 8 or fewer models agreed on the classification, the random choice shows better performance. In this context, the optimal solution was found to be a mixed strategy, applying the collective choice for processes where 9 or more models agree and random selection for processes where 8 or fewer models agree. This combination yields an average Benefit-Cost ratio of 2.197, which is 64.0% higher than the purely random selection currently employed by the TCE/PB.