Title: Forecasting volatility using long memory dynamics: How effective is the use of a realised measure
Authors: Hidde Jelsma - VU Amsterdam (Netherlands)
Katarzyna Lasak - VU Amsterdam (Netherlands) [presenting]
Abstract: A closer look is taken at multiple models to estimate the unobserved volatility component of financial returns. This is done for three stocks, Nike Inc., AEGON n.v., and Idex Corporation over the period of August 3rd 2009 till August 1st 2014. We use models from the GARCH and GAS frameworks and extend these models to incorporate long memory dynamics. We look at the forecasting performance and the predictive capabilities of risk. Forecasting performance is tested by comparing the forecasts with the Realised kernel described previously in the literature, which is used as a proxy for the true volatility, and applying the Diebold-Mariano test based on squared and absolute errors. The predictive capabilities of risk are compared by looking at the Value at Risk and applying the backtesting methods also previously described. We find that models imposing long memory dynamics do not provide better one-day-ahead forecasts but become more powerful if we extend the forecasting horizon to ten days, and that the models using the Realised kernel have provided more accurate forecasts and predict risk better than models using only the returns as input, but that the Realised kernel is not always an accurate measure of the true volatility.