Predictive Models in the Assessment of Tax Fraud Evidences

Abstract

The aim of the work is to verify the possibility of improving the selection of taxpayers to be inspected through projections of the results of future audits, based on the results of the inspections already carried out. The analysis of information about the process, obtained from the auditors involved in the selection of taxpayers and in the inspection of companies, allowed the selection of the variables used in the models, and the literature review allowed to define the techniques and tools necessary for their creation and training. The research generated predictive models of logistic regression and neural networks, whose forecasts identified sets of companies that correspond to approximately half of the audited companies and account for more than 80% of the credit constituted (89% in the case of the model neural network), of so that these models have the potential to optimize the application of available resources and maximize results, assisting in the selection of indications of irregularities and fraud with greater potential for the constitution of the due credit.