Are the statistical improvements yield by long memory models in the computation of Value at Risk substantial? The performance of the GARCH and the ARFIMA models, the latter estimated using daily variance obtained from high frequency data, are compared on various criteria. The results show that the long memory model provides a superior performance in terms of multi-step point forecasting. Allowing for time-varying variance of the realized variance process in the context of an ARFIMA-FIGARCH model also substantially improves VaR forecasting.