Most long memory forecasting studies assume that the memory is generated by the fractional difference operator. We argue that the most cited theoretical arguments for the presence of long memory do not imply the fractional difference operator, and assess the performance of the autoregressive fractionally integrated moving average (ARF IM A) model when forecasting series with long memory generated by nonfractional processes. We find that high-order autoregressive (AR) models produce similar or superior forecast performance than ARF IM A models at short horizons. Nonetheless, as the forecast horizon increases, the ARF IM A models tend to dominate in forecast performance. Hence, ARF IM A models are well suited for forecasts of long memory processes regardless of the long memory generating mechanism, particularly for medium and long forecast horizons. Additionally, we analyse the forecasting performance of the heterogeneous autoregressive (HAR) model which imposes restrictions on high-order AR models. We find that the structure imposed 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, among others, Climate Econometrics and Financial Econometrics models dealing with long memory series at different forecast horizons. We show in an example that while a short memory autoregressive moving average (ARM A) model gives the best performance when forecasting the Realized Variance of the S&P 500 up to a month ahead, the ARF IM A model gives the best performance for longer forecast horizons.JEL classification: C53, C22.