Previsão de séries de tempo na presença de mudança estrutural: redes neurais artificials e modelos estruturais

Abstract

The Brazilian price stabilisation policies and trade liberalisation measures of this decade have considerably increased the difficulty in generating accurate time series forecasts due to structural changes in the data generation processes. In this paper we provide an empirical evaluation of the forecasting performance of Artificial Neural Networks (ANN) and Structural Time Series models (STS) in the presence of structural change. We are basically interested in evaluating the capability of ANN and STS models in terms of both identifying that a structural change has happened and the speed of adjustment of the one step ahead forecasts after the change. We use both real and simulated time series in these exercises. The simulated series are generated from ARIMA processes with imposed structural changes in the mean and trend. On the other hand, we also use real time series data for the Brazilian inflation rate and total imports. The results for the one step ahead forecasts show that the ANN models present a marginally better perfomance than the STS in the periods just after the structural change.