Skip to content

A Simple Approximate Long-Memory Model of Realized Volatility (HAR): summary and key takeaways

The paper proposes the HAR-RV model, a regression of next-day realized volatility on daily, weekly and monthly average realized volatility. Although it is not a true long-memory model, it reproduces long memory and fat tails in simulation. Out of sample it beats simple short-memory autoregressions and performs about as well as an ARFIMA model.

Corsi, F. · 2008 · Journal of Financial EconometricsSummary updated 20 November 2026
Published2008Journal article
Asset focusVolatility/VIXVolatility forecasting
TypeResearchVolatility & VIX
Reading time4 minthis summary

A Simple Approximate Long-Memory Model of Realized Volatility (HAR) at a glance

The question
Can a simple three-component volatility regression reproduce long-memory behaviour and forecast realized volatility as well as fractionally integrated models?
Dataset & period
Tick-by-tick data for USD/CHF spot (14 years, December 1989 to December 2003, 3599 daily observations), S&P500 futures (January 1990 to July 2007, 4344 observations) and 30-year T-Bond futures (January 1990 to October 2003, 3391 observations). Daily realized volatility.
Method
Daily realized volatility is computed from tick data with a two-scales estimator. It is averaged into weekly and monthly measures. Next-day realized volatility is regressed by OLS on the daily, weekly and monthly values, with Newey-West standard errors. The model is justified by traders with different horizons creating a volatility cascade from long to short horizons. It is compared with AR(1), AR(3) and ARFIMA(5, d, 0) using RMSE, MAE and Mincer-Zarnowitz R2.

Key findings of A Simple Approximate Long-Memory Model of Realized Volatility (HAR)

  • The HAR model is a restricted version of an AR(22) regression, with step-function weights on daily, weekly and monthly volatility.
  • In a simulation calibrated with weights of 0.36, 0.28 and 0.28, the model matches USD/CHF return kurtosis: daily kurtosis is 4.89 simulated versus 4.74 actual.
  • The simulated volatility autocorrelation decays as slowly as the empirical one, even though the model is formally short-memory.
  • In the estimates, all three horizon components are highly significant for USD/CHF and S&P500. For T-Bond the daily term is not significant, which the paper links to noisier realized volatility.
  • Out of sample, HAR(3) outperforms AR(1) and AR(3) at one-day, one-week and two-week horizons, with the gap largest at longer horizons.
  • HAR(3) and ARFIMA are comparable out of sample. HAR is slightly better at daily and weekly horizons and ARFIMA slightly better at two weeks.

Robustness, as reported in the paper

  • Forecasts come from models re-estimated daily on a moving window of 1000 observations.
  • An F-test rejects the 19 restrictions of HAR relative to AR(22) for all three series, though BIC prefers HAR(3) and AIC is similar for both.
  • The ARFIMA benchmark is not fully out of sample: its fractional differencing coefficient is fitted once on all the data.

Caveats and limits

  • Only three series are tested, each ending by 2007, so the paper offers no evidence on later periods (inferred).
  • The model is not a true long-memory process; the long memory it shows is approximate and depends on the aggregation range tested.
  • Residuals show clear heteroskedasticity, which the paper says more sophisticated estimation could address.
  • Forecast comparisons rely on tick data and a noise-robust estimator, which may be hard to replicate with lower-quality data (inferred).

Items marked (inferred) are our reading of the design, not statements by the authors.

Why it matters

HAR-RV became a common baseline for realized volatility forecasting because it needs only OLS on three averages. The paper shows a parsimonious model can match a more complex long-memory benchmark. Practitioners can use it as a transparent starting point and extend it with extra regressors.

Questions about this paper

What does the HAR-RV model regress on?

Next-day realized volatility is regressed on the current daily realized volatility and on its weekly and monthly averages.

Is the HAR model a true long-memory model?

No. The paper states it is formally short-memory, but it reproduces long-memory-like persistence over the ranges tested.

How does HAR compare with ARFIMA out of sample?

The paper finds the two comparable, with HAR slightly ahead at daily and weekly horizons and ARFIMA slightly ahead at two weeks.

Where can I read the original paper?

The original is at its publisher: https://doi.org/10.1093/jjfinec/nbp001. We link to the paper and do not host the PDF.

Follow new quant research

Download MacroFinance market data as CSV or Excel, or pull it by API, and test the ideas in these papers yourself. New summaries are added as we review them.

Create a free account
Summaries are written by MacroFinance from the published papers, in our own words, for information only. They are not investment advice and may omit detail. Papers are not hosted here: each page links to the original at its publisher.