Realized Volatility and the HAR Model
Also known as: realized variance, HAR model, heterogeneous autoregressive model of realized volatility, HAR-RV, Gerçekleşmiş Volatilite ve HAR Modeli
Realized volatility estimates an asset's variance directly from high-frequency intraday returns rather than from a parametric latent process. The Heterogeneous Autoregressive (HAR) model of Corsi (2009), building on the realized-volatility framework of Andersen, Bollerslev, Diebold and Labys (2003), forecasts this measure by combining daily, weekly, and monthly volatility components, and is a strong alternative to GARCH for volatility prediction.
Key highlights
- Uses observable high-frequency data to estimate variance directly, avoiding the latent-process assumptions of parametric volatility models.
- The HAR model reproduces the long-memory-like persistence of volatility while remaining a simple OLS regression that is fast to estimate.
- A strong, well-documented forecasting alternative to GARCH, capturing daily, weekly, and monthly dynamics in separate components.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use realized volatility and the HAR model when you have high-frequency (intraday, typically 5-minute or finer) return data and want to measure or forecast variance directly. It suits continuous time-series data with a reasonably long history; at least about 250 days of data is recommended for reliable estimates, and shorter histories favour GARCH-type models instead. The series should be stationary — non-stationary volatility series can give misleading measurements and may need differencing. Because high-frequency data carries market-microstructure noise, sampling at five minutes or coarser, or applying subsampling or kernel estimators, is advisable.
Strengths & limitations
- Uses observable high-frequency data to estimate variance directly, avoiding the latent-process assumptions of parametric volatility models.
- The HAR model reproduces the long-memory-like persistence of volatility while remaining a simple OLS regression that is fast to estimate.
- A strong, well-documented forecasting alternative to GARCH, capturing daily, weekly, and monthly dynamics in separate components.
- High-frequency data contains market-microstructure noise, so measurements need careful sampling, subsampling, or kernel correction.
- With insufficient high-frequency data (fewer than about 250 days) realized-volatility estimates become unreliable and GARCH is preferable.
- On non-stationary series the volatility measurement is misleading; the data may need differencing first.
- HAR mimics long memory but is actually a short-memory structure, so it is an approximation rather than a true long-memory model.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does realized volatility differ from GARCH?
GARCH infers a latent conditional variance from a parametric model of returns, while realized volatility measures variance directly by summing squared high-frequency intraday returns. The HAR model then forecasts this observable measure with a simple regression and is a strong alternative to GARCH when intraday data is available.
What is microstructure noise and why does it matter?
At very fine sampling frequencies, observed prices are contaminated by bid-ask bounce and other trading frictions, biasing the summed squared returns. Sampling at five minutes or coarser, or applying subsampling or kernel-based estimators, mitigates this noise.
Why does HAR use daily, weekly, and monthly components?
Different market participants operate over different horizons. By combining the daily measure, the past week's average, and the past month's average, the HAR model reproduces the slow-decaying, long-memory-like persistence of volatility while staying a parsimonious linear regression.
Should I model realized variance in logs?
Realized-volatility series are usually right-skewed, so a log-RV specification often fits better and stabilises the variance of the regression errors.
Sources
- 1.Corsi, F. (2009). A Simple Approximate Long-Memory Model of Realized Volatility. Journal of Financial Econometrics, 7(2), 174-196.
- 2.Andersen, T. G., Bollerslev, T., Diebold, F. X., & Labys, P. (2003). Modeling and Forecasting Realized Volatility. Econometrica, 71(2), 579-625.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 1). Realized Volatility. ScholarGate. https://scholargate.app/finance/realized-volatility