Previsão de variáveis macroeconômicas brasileiras usando modelos de séries temporais de alta dimensão

Abstract

This paper analyzes the performance of high-dimensional factor models to forecast four Brazilian macroeconomic variables: two real variables, unemployment rate and industrial production index, and two nominal variables, IPCA and IPC. The factors are estimated from a data set containing 117 macroeconomic variables. We applied techniques to improve factor models forecasts. Methods of statistical learning are applied aims to increase the performance of factors models. Three types of statistical learning techniques are used: shrinkage methods, forecast combinations, and selection of preditors. The factors are extracted using supervised and unsupervised version. The results indicate that statistical learning improves forecasts performance. The combination of statistical learning and supervised factor models is more accurate than all other models, with exception to the industrial production index which is best forecasted by unsupervised factor model without statistical learning.