Volatility is central to financial theory. In practice volatility is not directly observed and must be estimated. Merton's (1980) seminal work suggests that as the sampling interval approaches zero arbitrarily precise volatility estimates can be obtained. Realistically, however, the limiting case is not attainable since the sampling frequency cannot be any higher than transaction by transaction. We examine the precision of volatility estimates that use the high-frequency data. We report several prominent high frequency data characteristics, including autocorrelation, deterministic intra-day volatility pattern, leptokurtosis and volatility clustering in the intra-day returns, which are consistent with other findings in literature, and we analyze their impacts on the estimation of volatility. Once these features are accounted for, we find that large amounts of high frequency data do not necessarily translate into very precise volatility estimates. Our results provide a measure of the usefulness of high frequency data in estimating volatility.; The relation of leptokurtic property with volatility clustering and non-normality is also studied. We find that, for LARCH model or linear stochastic volatility model, the volatility clustering and non-normality contribute interactively and symmetrically to the overall kurtosis of the series. |