What's the best way to estimate realized volatility from high-frequency data?

Question

AcadiFi · Accepted Answer

Each estimator trades bias against variance and microstructure noise. ```mermaid flowchart TD A[Price
Data] --> B{Frequency} B -->|Daily OHLC| C[Close-to-Close
Stdev] B -->|Daily OHLC| D[Parkinson
High-Low] B -->|Daily OHLC| E[Garman-Klass
OHLC] B -->|Intraday| F[Realized
Variance] F --> G[Noise
Bias] G --> H[Subsample
or TSRV] ``` Close-to-close uses only the end-of-day price. It's unbiased but noisy — you need many days to pin down vol. Parkinson uses daily high and low, yielding an estimator roughly 5x more efficient under geometric Brownian motion with no drift: sigma_P^2 = (1/(4 ln 2)) * (ln(H/L))^2. Garman-Klass adds open and close and gets another efficiency bump. Rogers-Satchell allows drift. Realized variance from intraday returns (say 5-minute) is the sum of squared log returns over the day. In continuous time this converges to integrated variance. The catch is market microstructure noise: bid-ask bounce inflates RV as sampling frequency rises. Andersen-Bollerslev suggested 5-minute sampling as a rule of thumb; modern practice uses two-scale realized volatility (TSRV) or pre-averaging to debias. A concrete example: Northfield Research estimates 1-month vol on KalmarTech. Close-to-close gives 24%, Parkinson 23.5%, and 5-minute RV 23.8%. Moving to 30-second sampling without correction gives 31% — mostly noise. For FRM exam purposes: know that intraday RV is the theoretically best estimator but requires noise correction, and range-based estimators like Parkinson are dramatically more efficient than close-to-close.

What's the best way to estimate realized volatility from high-frequency data?

Master Part I with our FRM Course

Related Questions

Practice Questions