Modelos lineares e não lineares da Curva de Phillips para previsão da taxa de inflação no Brasil

Abstract

This paper presents an analysis of forecasting the Brazilian monthly inflation rate from different models linear and nonlinear time series and the Phillips Curve, with the goal of identifying the best predictive mechanism for this variable. The model used as a comparative basis of forecasts in this study was the case with autoregressive moving average. Within this class of models, the model that generated the lowest mean square forecasting error (MSE) was the AR (1). In general, the threshold models used indeed had a better performance in forecasting of the rate of inflation that the linear models. The model autoregressive indeed threshold (TAR) presented a forecast of MSE equal to 4.3%, result around 10.41% better than the forecast of the linear AR (1) process. Among the Phillips curve models which submitted the lowest estimate of MSE was the Phillips curve extended with threshold effect that had an MSE equal to 3.4%, 28.5% better result than the model AR (1) and 32.6% better than the Phillips curve extended linear. In addition to a lower estimate of MSE, the graphic analysis revealed that the Phillips curve model extended with Threshold effect also provides better forecasts for sign changes. The test proposed by Diebold and Mariano (1995) was also performed and showed a result that indicates a significant difference between AR model (best linear model) MSE and the Phillips curve model expanded with threshold (best non-linear model). That is, the non-linear model also presented better results according to this second test.