Recent studies have found indications of
long-range dependence in financial time series and used conventional,
non-robust estimates of the memory parameter, which measures the degree of
long-range dependence, for the calculation of buy and sell signals. In this
paper, new robust estimators are proposed which are possibly more appropriate
for financial data. The new estimators are compared with various robust and
non-robust competitors by means of extensive simulations. In addition to
additive outliers and heavy-tailed distributions, also conditional heteroscedasticity is considered. The results show that the robust estimators
do not generally deliver better results than the conventional estimators but
only in special cases, the existing robust estimators with respect to the
root-mean-square error and the new robust estimators with respect to the bias.
Finally, the different estimators are used to investigate possible long-range
dependence both in developed and developing stock markets. The results of this empirical study suggest
that long-range dependence is present only in the volatility and is therefore
of no use for directional forecasting and trading.
JEL classification numbers: C13, C14, C22, C58, G15
Keywords: Long-range dependence, Frequency-domain estimation,
Periodogram, Truncated F-distribution, Volatility, Stock markets.