Long memory components in macroeconomic series: Are we missing something?

摘要

In this paper it is shown that there are strong statistical evidences towards the presence of long memory components in three American macroeconomic time series (Output Gap, M1 Quantity of Money and Real Interest Rates). Moreover, in this paper is presented two Fractionally Integrated Vector Autoregression (FIVAR) models with the fractional difference coefficients estimated using two different procedures aiming to forecast key macroeconomic variables of the American economy. They are compared against a standard Vector Autoregression (VAR) model, and it is shown that the FIVAR models outperform significantly the traditional VAR in forecasting capabilities, suggesting that the usage of fractional difference parameter and the consideration of long memory provides a more robust estimation (as a consequence of the reduction of parameters to be estimated).