Regime de metas no Brasil e previsão de inflação: acurácia e encompassing

Abstract

This study seeks to determine if the adoption of inflation targeting regime in Brazil in 1999, has been able to reduce inflation forecasts errors generated by Naïve models, ARIMA, GARCH, UC-SV, VAR and Phillips Curve, and how the accuracy of the forecasts generated by these models has behaved throughout the post-adoption period of the regime. In addition, we seek to verify the occurrence of predictions encompassing generated by the models mentioned. The reporting period comprises January 1996 to December 2013 and the horizon of the forecast is twelve months. Using statistical Root Mean Square Error (RMSE) and Diebold-Mariano modified test (mDM), the results show that the prediction errors have been reduced over time, so that inflation appears to be easier to be predicted. On the other hand, the Naive model is more accurate than other models examined, so that inflation is more difficult to predict. The HLN modified test (mHLN) shows that inflation forecasts generated by the Naive model contains in most of the time, all the necessary information to make accurate predictions, not requiring to incorporate information available on predictions generated by other models. However, this situation appears to be changing over time.