Tempo de atravessamento de ações cíveis na Justiça Federal de 2o grau: ajuste de modelos baseados em redes neurais e máquina de vetor suporte

Abstract

The Courts as an example of public institutions, are fundamental for the economic and social development. However, the main problems of the Brazilian judiciary, pointed out by the Ministry of Justice, are the high number of cases in stock, lack of access to justice and slowness, considered as the main aspect of the crisis of the judiciary. In this sense, this thesis proposes to structure and compare models of lead time of civil lawsuits in the federal court of second instance, to serve as information for the parties of lawsuit and the administration. The study was conducted at the Tribunal Regional Federal da 4a Região, with data from civil lawsuits termined in 2017. Four models for lead time lawsuits were compared. The first was fitted by neural network for regression with the backpropagation algorithm; the second model was fitted by using support vector machine for regression with the Libsvm library. The performance of these two models, calculated by the RMSE measurement, was compared to the performance of the survival analysis model, considered as the usual technique for the analysis of quantitative time studies. The dependent variable used was the time in days between the arraignment date and the case disposition date, chosen among indicators used in academic studies and by judicial Courts of Brazil and Europe. The fourth model was fitted using vector support machine for classification, using the Libsvm algorithm. The dependent variable was transformed into ordinal by means of the stratification in time bands, which allowed the calculation of the measurement as accuracy and precision. The independent variables were categorical and were available in the TRF database. After, rules of association were applied to the time bands in order to find the characteristics of the most frequent band and the more time consuming lawsuits. The feasibility of establishing parameters of reasonable times was also analyzed. The time-band classification model is suggest to use in forecasts and it can use the model adjusted by neural networks or by the support vector machine in use to establish standards of time. Among the suggestions for future work are the construction of a life tables of lawsuits, analogous to the actuarial tables, and the establishment of standards to consider reasonable lead time.