Skip to content

How Often to Sample a Continuous-Time Process in the Presence of Market Microstructure Noise

Bibliographic record. Follow the original-source link for the publication.

Field Value
Primary domain Volatility
Other domains —
Methods Research Methods
Facets —
Authors Yacine Aït-Sahalia, Per A. Mykland, Lan Zhang
Published 2005-02-10
Source The Review of Financial Studies
Identifiers doi:10.1093/rfs/hhi016
URL Open original source

Editorial synthesis

Why it matters

This foundational study shows why more high-frequency observations do not automatically improve realized-volatility estimation: ignoring market microstructure noise creates a finite optimal sampling interval, whereas modeling the noise allows all observations to be retained. It provides a testable statistical baseline for sampling and noise correction in high-frequency volatility features. (full_text:S10, full_text:S26, full_text:S38, full_text:S49, full_text:S51)

Main author claims

  • The authors derive a finite RMSE-optimal interval when the estimator ignores market microstructure noise; calibrations based on noise magnitudes reported in prior work range from minutes to hours for a one-day sample. (full_text:S26, full_text:S28, full_text:S29, full_text:S170, full_text:S171)
  • The authors report that an MA(1) likelihood for noise-contaminated returns restores the case for using the full highest-frequency sample instead of retaining only fixed five-minute observations. (full_text:S38, full_text:S39, full_text:S41, full_text:S49, full_text:S51)
  • The authors show that, under their second-moment and covariance structure, the volatility estimator has the same asymptotic variance even when the true noise distribution is non-Gaussian. (full_text:S43, full_text:S46, full_text:S47, full_text:S48)

Data, method, or discussion scope

The evidence is primarily analytical, built around a continuous-time diffusion, additive price noise, and likelihood derivations, with exact small-sample results, 10,000 Monte Carlo replications, and extensions to random sampling and correlated noise. This is a volatility-estimation study, not a test of alpha, option returns, or an executable strategy. (full_text:S61, full_text:S62, full_text:S67, full_text:S77, full_text:S80, full_text:S198, full_text:S574)

Main limitations

The baseline starts with constant volatility, equally spaced observations, and i.i.d. additive noise. Although several assumptions are relaxed later, the full nonparametric stochastic-volatility case is delegated to companion work. Numerical sampling intervals depend on noise calibrations borrowed from prior studies and should not be transferred mechanically across assets, market designs, or vendors. The derivations and Monte Carlo results were not independently replicated here. (full_text:S52, full_text:S69, full_text:S76, full_text:S77, full_text:S80, full_text:S170, full_text:S171, full_text:S198)

Relationships

  • None recorded.