Aplicação de aprendizagem de máquina para monitoramento de atividade econômica utilizando dados de comércio exterior

Abstract

The use of data science and machine learning techniques has obtained a prominent position both in the technological field and in other fields of knowledge that realize the scientific potential brought by the study of massive amounts of data. Considering this interdisciplinarity, this work of course conclusion proposes to explore an application of data science in time series of economic variables. More specifically, aims the development of machine learning models to perform estimates concerning the behavior of the Brazilian economic activity. To this end, historical foreign trade data and the Economic Activity Index from Brazilian Central Bank, IBC-Br, are used. An exploratory analysis is conducted on the data and four classes of models are studied: Multiple Linear Regression, Vector Autoregression (VAR), ARIMA and Random Forest. The performances are evaluated using classical statistical metrics and comparison with a base model. The best model was able to achieve an error reduction of about 48forecasts when compared to the base model, showing potential applicability in real economic activity estimation situations.