Microstructure noise, realized volatility, and optimal sampling

Recorded prices are known to diverge from their "efficient" values due to the presence of market microstructure contaminations. The microstructure noise creates a dichotomy in the model-free estimation of integrated volatility. While it is theoretically necessary to sum squared returns that are computed over very small intervals to better identify the underlying quadratic variation over a period, the summing of numerous contaminated return data entails substantial accumulation of noise. Using asymptotic arguments as in the extant theoretical literature on the subject, we argue that the realized volatility estimator diverges to infinity almost surely when noise plays a role. While realized volatility cannot be a consistent estimate of the quadratic variation of the log price process, we show that a standardized version of the realized volatility estimator can be employed to uncover the second moment of the (unobserved) noise process. More generally, we show that straightforward sample moments of the noisy return data provide consistent estimates of the moments of the noise process. Finally, we quantify the finite sample bias/variance trade-off that is induced by the accumulation of noisy observations and provide clear and easily implementable directions for optimally sampling contaminated high frequency return data for the purpose of volatility estimation