Previsão de ICMS por rede neural backpropagation

Abstract

The ICMS is a major source of revenue for the 26 Brazilian states and the Federal District, in 2018 was responsible for 7% of the Brazilian GDP. However, it is a tribute to petition the complex form, and its own indicators of external influence, such as political leaders, profits, financing, and climate, among many others. All the variables make the ICMS nothing biased. But his prediction is of utmost importance to the government at all levels. This study was developed that decided to develop an application to make a prediction of ICMS using artificial neural networks of backpropagation, not the latter as a whole of this work. As a specific objective, the different network structures and data utilization strategies (ICMS collections in the monthly range from January 1997 to October 2018) in order to find a smaller margin of error in the forecast of future ICMS collections from past collections. Among the strategies, the 6 ways of using the samples, 3 different access functions, 3 neuron quantity spheres in the hidden layer, 2 ways to start as samples, 2 ways to start weights, 5 strategies for using the learning rate and 4 different strategies for using collections. With this, we executed 4,320 times the project, where each execution had a purpose of testing a strategy and using the data information. When finalizing the tests, observe the results of the strategies, comparing the assumptions with their sources and, also comparing, in general, through the different data samples.