Title: Long horizon forecasts
Authors: Jose Eduardo Vera Valdes - Aalborg University (Denmark) [presenting]
Abstract: Most forecasting studies for long memory assume that the series are generated by fractional processes. We assess the performance of the ARFIMA model when forecasting long memory series where the long memory generating mechanism may be different from fractional differencing. We consider cross sectional aggregation and the error duration model as long memory generating mechanisms. We find that ARFIMA models produce similar forecast performance compared to high-order AR models at shorter horizons. As the forecast horizon increases, the ARFIMA models tend to dominate in forecast performance. Hence, ARFIMA models are well suited for long horizon forecasts of long memory processes regardless of how the long memory is generated. Additionally, we analyze the forecasting performance of the heterogeneous autoregressive model (HAR), we find that the structure enforced by the HAR model produces better long horizon forecasts than AR models of the same order at the price of inferior short horizon forecasts in some cases. Our results have implications for Climate Econometrics and Financial Econometrics models dealing with larger forecast horizons. In an example, we show that a short memory ARMA model gives the best performance when forecasting the Realized Variance for the S\&P 500 up to a month ahead, while an ARFIMA model gives the best performance for longer forecast horizons.