Inflation in Brazil

an application of Time Series and Recurrent Neural Networks

Abstract

The realization of inflation forecasts is essential when one wants to improve strategic planning in order to reduce uncertainties and increase the planning capacity of families, companies and government, and thus ensure the proper functioning of the country's economy. Because of this, the objective of this study was to make forecasts of the Extended National Consumer Price Index (IPCA), which is the official metric for measuring inflation by the federal government. The Box and Jenkins and recurrent neural network techniques were used to build models that capture the information contained in the series, in order to evaluate the behavior of the variable analyzed over time, as well as to make predictions of future values. By applying the Box and Jenkins model, a SARIMA-type forecasting model was adopted, which considers both the trend and the seasonality in the series, being selected the one with the smallest mean square error (RMSE) and the smallest number of parameters. For the models of recurrent neural networks, the LSTM algorithm was used, and the one with the lowest RMSE was selected. In order to compare the two forecasting techniques, the criteria adopted were root mean square error (RMSE), and the SARIMA model proved to be the best model for predicting 12 direct months of inflation, while for short-term forecasts with feedback , the LSTM technique proved to be very effective.