Range-Based Volatility

Estimators: Overview and

Examples of Usage
! 18 minute read

Volatility estimation and forecasting plays a crucial role in many areas of finance.

For example, standard risk-based portfolio allocation methods (minimum variance,

equal risk contributions, hierarchical risk parity…) critically depend on the ability to
build accurate volatility forecasts1.

Multiple methods for estimating volatility have been proposed over the past several
decades, and in this blog post I will focus on range-based volatility estimators.

These estimators, the first of which introduced by Parkinson2 as a way to compute

the true variance of the rate of return of a common stock2, rely on the highest and
lowest prices of an asset over a given time period to estimate its volatility, hence their
name3.

After describing the four most well known range-based volatility estimators, I will
reproduce the analysis of Arthur Sepp (https://artursepp.com/) in his presentation
Volatility Modelling and Trading4 made at Global Derivatives Conference 2016 and
test the predictive power of the naive volatility forecasts produced by these
estimators for various ETFs.

Notes:
A very accessible series of papers about range-based volatility estimators has
recently5 been released by people at Lombard Odier
(https://www.lombardodier.com/home.html), c.f. here
:
 (https://www.lombardodier.com/home.html), c.f. here
(https://am.lombardodier.com/fr/en/contents/news/investment-
viewpoints/2023/february/1148-NA-NA-NA-volatility.html), here
(https://am.lombardodier.com/fr/en/contents/news/investment-viewpoints/2023/may/1148-
NA-NA-NA-volatility.html) and here
(https://am.lombardodier.com/fr/en/contents/news/investment-viewpoints/2023/august/1148-
MARS-PROD-risk-based.html).

Mathematical preliminaries
Volatility modelling
One of the main6 assumptions made when working with range-based volatility
estimators7 is that the price movements St of the asset under consideration follow a
geometric Brownian motion (https://en.wikipedia.org/wiki/Geometric_Brownian_motion) with
unknown volatility coefficient8 σ and unknown drift coefficient µ, that is

dSt = µSt dt + σSt dWt

, where Wt is a standard Brownian motion.

Under this working assumption, σ represents the volatility of the asset.

Volatility and variance estimators

Although anyone can empirically observe the impact of “volatility” on the prices of a
given asset, the volatility coefficient σ of this asset is not directly observable9 and
must be estimated using stock market information.

A statistical estimator (https://en.wikipedia.org/wiki/Estimator) of σ is then called a volatility

estimator, and a statistical estimator of σ 2 is called a variance estimator.

Efficiency of a volatility estimator

In order to determine the quality of a volatility estimator, two measures are commonly
used:

Bias (https://en.wikipedia.org/wiki/Bias_(statistics))
:
 Bias (https://en.wikipedia.org/wiki/Bias_(statistics))

The bias of a volatility estimator measures whether this estimator produces, on

average, too high or too low volatility estimates.

More formally, a volatility estimator σA is said to be unbiased when E[σA ] = σ

and biased otherwise.

Efficiency (https://en.wikipedia.org/wiki/Efficiency_(statistics))

The efficiency of a volatility estimator measures the uncertainty of the volatility

estimates produced by this estimator, with the greater the efficiency of the
estimator, the more accurate the volatility estimates.

More formally, the relative efficiency Eff (σA ) of a volatility estimator σA

compared to a reference volatility estimator σB is defined as the ratio of the
variance of the estimator σB2 over the variance of the estimator σA2 , that is,

Var (σB2 )
Eff (σA ) =
Var (σA2 )

To be noted that bias and efficiency are sometimes conflicting, which is more
generally known in statistics as the bias-variance tradeoff
(https://en.wikipedia.org/wiki/Bias%E2%80%93variance_tradeoff).

Close-to-close volatility estimators

Let C1 , … , CT be the closing prices of an asset for T time periods t = 1..T 10.

Then,

Ci 2
T
1
σcc,0 (T ) = ∑ ln
T − 1 i=2 Ci−1

is a biased11 estimator of the asset volatility σ over the T time periods, assuming zero
drift (i.e., µ = 0), c.f. Parkinson2.

In addition,

T 2
1
∑ (ln − µcc )
Ci
σcc (T ) =
T −2 Ci−1
:
 ∑ (ln −µ )
T −2 i=2
Ci−1
1
ln CCi−1i , is a biased11 estimator of the asset volatility σ over the T
T
, with µcc = T −1 ∑i=2
= 0), c.f. Yang and Zhang12.
time periods, assuming non-zero drift (i.e., µ 

These two estimators are known as close-to-close volatility estimators.

Range-based volatility estimators

Let be:

t = 1..T , T time periods10

(O1 , H1 , L1 , C1 ) , … , (OT , HT , LT , CT ), the opening, highest, lowest and closing

prices of an asset for time periods t = 1..T

As mentioned in the introduction, a volatility estimator fully or partially relying on the

highest prices Ht , t = 1..T and on the lowest prices Lt , t = 1..T is called a range-
based volatility estimator.

The underlying idea behind such estimators is that information contained in the asset
high-low price ranges Ht − Lt , t = 1..T should allow to build volatility estimators that
are more efficient than the close-to-close volatility estimators, which use only one
price inside this range13.

This quest for efficiency is important because, contrary to one of the working
assumptions6, the volatility of an asset is known to be time-varying14, so that the less
the number of time periods required to estimate its volatility, the more chances that
its volatility is constant(ish) over the time periods under consideration.

As Rogers et al.15 put it:

[…] volatility may change over long periods of time; a highly efficient
procedure will allow researchers to estimate volatility with a small number of
observations.

Parkinson volatility estimator

Parkinson2 introduces an estimator for the diffusion coefficient of a Brownian motion
without drift that relies on the highest and lowest observed values of this Brownian
:
 without drift that relies on the highest and lowest observed values of this Brownian
motion over a given time period.

When applied to the estimation of an asset volatility, this gives the Parkinson volatility
estimator σP (T ) defined over T time periods by

T 2
1 1
∑ (ln )
Hi
σP (T ) =
T 4 ln 2 Li
i=1

Intuitively, the Parkinson estimator should be “better” than the close-to-close

estimators because large price movements impacting the high-low price range Ht −
Lt but leaving the closing price Ct unchanged might occur within any time period t.

This is confirmed by the efficiency of this estimator, up to 5.2 times higher than the
efficiency of the close-to-close estimators16.

Garman-Klass volatility estimator

Garman and Klass17 propose to improve the Parkinson estimator by taking into
account the opening prices Ot , t = 1..T and the closing prices Ct , t = 1..T .

This leads to the Garman-Klass volatility estimator σGK (T ), defined over T time
periods by

T 2 2
1 1
∑ (ln ) − (2 ln 2 − 1) (ln )
Hi Ci
σGK (T ) =
T
i=1
2 Li Oi

For the historical comment, Garman and Klass17 establish in their paper that σGK is
the “best reasonable”18 volatility estimator that depends only on the high-open price
range Ht − Ot , the low-open price range Lt − Ot and the close-open price range Ct −
Ot , t = 1..T .

The Garman-Klass estimator is up to 7.4 times more efficient than the close-to-close
estimators16.

Rogers-Satchell volatility estimator

The Parkinson and the Garman-Klass estimators have both been derived under a zero
drift assumption.

When this assumption is not verified for an asset, for example because of a strong
:
 When this assumption is not verified for an asset, for example because of a strong
upward or downward trend in the asset prices or because of the usage of large time
periods (monthly, yearly…), these estimators should in theory not be used because
the quality of their volatility estimates is negatively impacted by the presence of a
non-zero drift1915.

In order to solve this problem, Rogers and Satchell19 devise the Rogers-Satchell
volatility estimator σRS (T ), defined over T time periods by

T
1 Hi Hi Li Li
σRS (T ) = ∑ ln ln − ln ln
T Ci Oi Ci Oi
i=1

The Rogers-Satchell estimator is up to 6 times more efficient than the close-to-close

estimators19, which is less than the Garman-Klass estimator20.

Yang-Zhang volatility estimator

The range-based volatility estimators discussed so far do not take into account
opening jumps in an asset prices21, that is, the potential difference between an asset
opening price Ot and its closing price Ct−1 for a time period t22.

This limitation causes a systematic underestimation of the true volatility12.

When trying to integrate opening jumps into the Parkinson, the Garman-Klass and the
Rogers-Satchell estimators, Yang and Zhang12 discover that it is unfortunately not
possible for any “reasonable” single-period23 volatility estimator to properly handle
both a non-zero drift and opening jumps.

This leads them to introduce the multi-period23 Yang-Zhang volatility estimator

σY Z (T ), defined over T time periods by

σY Z (T ) = 2 + kσ 2 + (1 − k)σ 2 )
σov oc RS

, where:

σco (T ) is the close-to-open volatility, defined as

T 2
1
∑ (ln − µco )
Oi
σco =
T −2 Ci−1
:
 ∑ (ln −µ )
T −2 i=2
Ci−1
1 T
, with µco = T −1 ∑i=2 ln COi−1i

σoc is the open-to-close volatility, defined as

T 2
1
∑ (ln − µoc )
Oi
σoc (T ) =
T −2 Ci
i=2
1 T
, with µoc = T −1 ∑i=2 ln O
Ci
i

σRS is the Rogers-Satchell volatility estimator over the time periods t = 2..T

0.34
k= 1.34+ T T−2

In addition to the new estimator σY Z , Yang and Zhang12 also provide multi-period
versions of the Parkinson, the Garman-Klass and the Rogers-Satchell estimators that
support opening jumps24.

The Yang-Zhang estimator is up to 14 times more efficient than the close-to-close

estimators12, a result that Yang and Zhang12 comment as follows

The improvement of accuracy over the classical close-to-close estimator is

dramatic for real-life time series

Other estimators
The family of range-based volatility estimators has many other members:

The Kunitomo25 volatility estimator

The Alizadeh-Brandt-Diebold26 volatility estimator
The Meilijson27 volatility estimator
…

Still, the Parkinson, the Garman-Klass, the Rogers-Satchell and the Yang-Zhang
volatility estimators are representative of this family, so that I will not detail any other
range-based volatility estimator in this blog post.
:
 From volatility estimation to volatility
forecasting
Range-based volatility estimators are based on the assumption of independent
sample and observations within the sample4, so that the corresponding volatility
forecasts are simply naive forecasts under a random walk model.

In other words, with such volatility estimators, the “natural” forecast of an asset
volatility over the next T time periods is the (past) estimate of the asset volatility over
the last T time periods.

That being said, it is perfectly possible to use range-based volatility estimates

together with any volatility forecasting model such as:

A time series forecasting model (simple moving average, exponentially weighted

moving average…), as detailed for example in Jacob and Vipul28
An econometric forecasting model (GARCH
(https://en.wikipedia.org/wiki/Autoregressive_conditional_heteroskedasticity#GARCH)
model…), c.f. Mapa29
A specific range-based forecasting model (Chou’s30 Conditional AutoRegressive
Range model, Harris and Yilmaz’s31 hybrid multivariate exponentially weighted
moving average model…)

Performance of range-based volatility

estimators
Theoretical and practical performances of range-based volatility estimators are
studied in several papers, for example Shu and Zhang32, Jacob and Vipul28 and
Brandt and Kinlay33, among others.

Most of these studies agree that range-based volatility estimators are biased11, but
other conclusions differ depending on the exact methodology used.

In particular, as highlighted by Brandt and Kinlay33, the results from empirical

:
 In particular, as highlighted by Brandt and Kinlay , the results from empirical
research differ significantly from those seen in simulation studies in a number of
respects33.

One perfect example of these differences is Shu and Zhang32 concluding, using a
Monte Carlo simulation, that

If the drift term is large, the Parkinson estimator and the [Garman-Klass]
estimator will significantly overestimate the true variance […]

, while Jacob and Vipul28 concluding, using real stock market data, that

Overall, the [Garman-Klass] estimator, which indirectly adjusts for the drift,
performs better for the high-drift stocks.

Motivated by such inconsistencies, Lyocsa et al.34, building on Patton and

Sheppard35, introduced what I will call the Lyocsa-Plihal-Vyrost volatility estimator
σLP V , defined as the arithmetic average of the Parkinson, the Garman-Klass and the
Rogers-Satchell volatility estimators36

σP + σGK + σRS
σLP V =
3
As Lyocsa et al.34 explain, the motivation behind using the (naive) equally weighted
average is based on the assumption that we have no prior information on which
estimator might be more accurate34.

I personally like the idea of an averaged estimator, but at this point, I think it is safe to
highlight that there is no “best” range-based volatility estimator…

Implementation in Portfolio Optimizer

Portfolio Optimizer implements all the volatility estimators discussed in this blog
post:

The close-to-close volatility estimators, through the endpoint

/assets/volatility/estimation/close-to-close (https://docs.portfoliooptimizer.io/)

The Parkinson volatility estimator, through the endpoint

/assets/volatility/estimation/parkinson (https://docs.portfoliooptimizer.io/)
:
 /assets/volatility/estimation/parkinson (https://docs.portfoliooptimizer.io/)

The Garman-Klass volatility estimator, through the endpoint

/assets/volatility/estimation/garman-klass (https://docs.portfoliooptimizer.io/)

The original Garman-Klass volatility estimator18, through the endpoint

/assets/volatility/estimation/garman-klass/original
(https://docs.portfoliooptimizer.io/)

The Rogers-Satchell volatility estimator, through the endpoint

/assets/volatility/estimation/rogers-satchell (https://docs.portfoliooptimizer.io/)

The Yang-Zhang volatility estimator, through the endpoint

/assets/volatility/estimation/yang-zhang (https://docs.portfoliooptimizer.io/)

, as well as their jump-adjusted variations, whenever applicable.

Examples of usage
To illustrate possible uses of range-based volatility estimators, I propose to reproduce
a couple of results from Sepp4:

The estimation and the forecast of the SPY ETF monthly volatility
The forecast of the monthly volatility of misc. ETFs representative of different
asset classes (U.S. treasuries, international stock market, gold…)

Such examples will allow to compare the empirical behavior of the different volatility
estimators and maybe reach a conclusion as to their relative performance in this
specific setting.

Estimating SPY ETF volatility

I will estimate the SPY ETF monthly volatility using all the daily open/high/low/close
prices37 observed during that month38.

Figure 1, limited to 5 volatility estimators for readability purposes, illustrates the

results obtained over the period 31 January 2005 - 29 February 201639.
:
 Figure 1. SPY ETF monthly volatility estimates, using daily returns over the period 31 January 2005 - 29 February
2016.

Figure 1 is mostly identical to the figure on slide 22 from Sepp4, on which it seems in
particular that the close-to-close and the Yang-Zhang volatility estimators provide
higher estimates of volatility when the overall level of volatility is high4.

Overall, though, the behavior of the different volatility estimators is essentially the
same on this specific example, which is confirmed by their correlations displayed in
Figure 2.

Figure 2. Correlations of SPY ETF monthly volatility estimates, using daily returns over the period 31 January 2005 -
29 February 2016.

Forecasting misc. ETFs volatility

Using the same methodology as in Sepp4, I will now evaluate the quality of the naive
forecasts produced by all the range-based volatility estimators implemented in
Portfolio Optimizer against the next month’s close-to-close observed volatility40, for
10 ETFs representative of misc. asset classes:
:
 10 ETFs representative of misc. asset classes:

U.S. stocks (SPY ETF)

European stocks (EZU ETF)

Japanese stocks (EWJ ETF)

Emerging markets stocks (EEM ETF)
U.S. REITs (VNQ ETF)
International REITs (RWX ETF)
U.S. 7-10 year Treasuries (IEF ETF)
U.S. 20+ year Treasuries (TLT ETF)
Commodities (DBC ETF)

Gold (GLD ETF)

These ETFs are used in the Adaptative Asset Allocation strategy from ReSolve Asset
Management (https://investresolve.com/), described in the paper Adaptive Asset
Allocation: A Primer41.

For each ETF, Sepps’s methodology is as follows:

At each month’s end, compute the volatility estimates σcc,t , σP ,t , … using all the
ETF daily open/high/low/close prices37 observed during that month38

Under a random walk volatility model, each of these estimates represents the
:
 Under a random walk volatility model, each of these estimates represents the
next month’s volatility forecast σ
^t+1

At each month’s end, also compute the next month’s close-to-close volatility
estimate σcc,t+1 using all the ETF daily close prices37 observed during that
month38

This estimate is the volatility benchmark, which represents how the ETF
“volatility” is perceived by an investor monitoring her portfolio daily.

Once all months have been processed that way, regress the volatility forecasts
on the volatility benchmarks by applying the Mincer-Zarnowitz42 regression
model:

^t+1 = α + βσcc,t+1 + ϵt+1

σ
, where ϵt+1 is an error term.

Then, the estimator producing [the best] volatility forecast is indicated by [a] high
explanatory power R^2, [a] small intercept α and [a] β coefficient close to one4.

Forecasting SPY ETF volatility

In the case of the SPY ETF, Figure 3 illustrates Sepps’s methodology for the Lyocsa-
Plihal-Vyrost volatility estimator σLP V over the period 31 January 2005 - 29 February
2016.
:
 Figure 3. SPY ETF Lyocsa-Plihal-Vyrost naive monthly volatility forecasts v.s. next month's close-to-close volatility
estimates, using daily returns over the period 31 January 2005 - 29 February 2016.

Detailed results for all regression models over the period 31 January 2005 - 29
February 2016:

Volatility estimator α β R2

Close-to-close 4.1% 0.75 57%

Close-to-close (zero drift) 3.9% 0.77 57%

:
 Close-to-close (zero drift) 3.9% 0.77 57%

Parkinson 3.5% 0.95 58%

Parkinson (jump-adjusted) 3.4% 0.79 58%

Garman-Klass 3.7% 0.92 57%

Garman-Klass (jump-adjusted) 3.6% 0.77 58%

Garman-Klass (original) 3.7% 0.92 57%

Garman-Klass (original, jump-adjusted) 3.6% 0.77 58%

Rogers-Satchell 4.0% 0.88 56%

Rogers-Satchell (jump-adjusted) 3.9% 0.74 57%

Yang-Zhang 3.8% 0.75 58%

Lyocsa-Plihal-Vyrost 3.7% 0.92 57%

While these figures are far43 from those on slide 42 from Sepp4, with for example
nearly no variation in terms of R2 among the different volatility estimators, two
observations are similar:

All volatility estimators have comparable α

The Parkinson, the Garman-Klass and the Rogers-Satchell volatility estimators
have a β much closer to 1 than the close-to-close volatility estimator

Forecasting the other ETFs volatility

Going beyond the SPY ETF, averaged results for all ETFs/regression models over each
ETF price history44 are the following:

Volatility estimator ˉ
α βˉ Rˉ2

Close-to-close 5.8% 0.66 44%

Close-to-close (zero drift) 5.6% 0.67 45%

:
 Close-to-close (zero drift) 5.6% 0.67 45%

Parkinson 5.6% 0.94 44%

Parkinson (jump-adjusted) 4.9% 0.70 45%

Garman-Klass 5.7% 0.93 43%

Garman-Klass (jump-adjusted) 5.0% 0.70 44%

Garman-Klass (original) 5.7% 0.93 43%

Garman-Klass (original, jump-adjusted) 5.0% 0.70 44%

Rogers-Satchell 6.1% 0.88 42%

Rogers-Satchell (jump-adjusted) 5.2% 0.68 43%

Yang-Zhang 5.1% 0.69 44%

Lyocsa-Plihal-Vyrost 5.7% 0.92 43%

A couple of remarks:

Forecasts produced by all the volatility estimators explain on average only ~45%
of the variability of the ETFs monthly volatility
Forecasts produced by the jump-adjusted volatility estimators seem to offer no
improvement on average over the forecasts produced by the close-to-close
volatility estimator
Forecasts produced by the Parkinson, the Garman-Klass and the Rogers-
Satchell volatility estimators seem to be much less biased on average than the
forecasts produced by the close-to-close volatility estimator, a property
inherited by the Lyocsa-Plihal-Vyrost volatility estimator

As an empirical conclusion, it is disappointing that the naive monthly volatility

forecasts produced by range-based volatility estimators have about the same
predictive power as the forecasts produced by the close-to-close volatility estimator.
Nevertheless, because these forecasts are much less biased than their close-to-close
counterparts, they still represent an improvement for the many investors who
currently rely on close prices only45.
To also be noted, similar to one of the conclusions of Lyocsa et al.34, that the Lyocsa-
Plihal-Vyrost volatility estimator should probably be preferred to the Parkinson, the
Garman-Klass or the Rogers-Satchell volatility estimators because using only one
range-based estimators has occasionally led to very inaccurate forecasts, which
:
 range-based estimators has occasionally led to very inaccurate forecasts, which
could successfully be avoided by using the average of the three range-based
estimators34.

Conclusion
One aspect of range-based volatility estimators not discussed in this blog post is their
capability to capture important stylized facts about asset returns46.

This, together with possible ways to incorporate them in more predictive volatility
models than the random walk model, will be the subject of future blog posts.

Meanwhile, for more volatile discussions, feel free to connect with me on LinkedIn
(https://www.linkedin.com/in/roman-rubsamen/) or to follow me on Twitter
(https://twitter.com/portfoliooptim).

1. As well as correlation forecasts.

2. See Parkinson, Michael H., The Extreme Value Method for Estimating the Variance of the Rate of
Return, The Journal of Business 53 (1980), 61-65 (https://www.jstor.org/stable/2352357), which is the final
version of the working paper The random walk problem: extreme value method for estimating the
2 3 4
variance of the displacement (diffusion constant) started 4 years before.

3. Because the range of prices of an asset over a given time period is contained, by definition, within its
highest and its lowest price.

4. See Sepp, Artur, Volatility Modelling and Trading. Global Derivatives Workshop Global Derivatives
Trading & Risk Management, Budapest, 2016 (https://papers.ssrn.com/sol3/papers.cfm?
2 3 4 5 6 7 8
abstract_id=2810768).

5. At the date of publication of this post.

6. Other working assumptions are also commonly made, like assuming that the asset does not pay
dividends, assuming that the volatility coefficient σ remains constant, assuming that the geometric
Brownian motion model also applies during time periods with no trading activity (e.g., stock market
closure), etc. 2

7. In details, the geometric Brownian motion assumption slightly differs between authors; for example,
Garman and Klass17 assume that asset prices follow a more generic diffusion process, which includes
the geometric Brownian motion as a specific case.

8. σ is also called the diffusion coefficient of the geometric Brownian motion, but in the context of this
:
 blog post, I think it is clearer to explicitly call it the volatility coefficient.

9. See Andersen, T., Bollerslev, T., Diebold, F., & Labys, P. (2003). Modeling and forecasting realized
volatility. Econometrica, 71, 579–625 (https://onlinelibrary.wiley.com/doi/10.1111/1468-0262.00418).

10. In practice, a time period t usually corresponds to a trading day, a week or a month, so that the closing
2
prices Ct , t = 1..T are simply the daily, weekly or monthly closing prices of the asset.

11. These estimators are biased, due to Jensen’s inequality

(https://en.wikipedia.org/wiki/Unbiased_estimation_of_standard_deviation#Motivation); c.f. also
Molnar46. 2 3

12. See Yang, D., and Q. Zhang, 2000, Drift-Independent Volatility Estimation Based on High, Low, Open,
and Close Prices, Journal of Business 73:477–491
(https://www.jstor.org/stable/10.1086/209650). 2 3 4 5 6

13. The asset closing price Ct , t = 1..T .

14. See French, K. R., Schwert, G. W., & Stambaugh, R. F. (1987). Expected stock returns and volatility.
Journal of Financial Economics, 19, 3–29
(https://www.sciencedirect.com/science/article/abs/pii/0304405X87900262).

15. See L. C. G. Rogers, S. E. Satchell & Y. Yoon (1994) Estimating the volatility of stock prices: a
comparison of methods that use high and low prices, Applied Financial Economics, 4:3, 241-247
(https://www.tandfonline.com/doi/abs/10.1080/758526905?journalCode=rafe20). 2

16. See Colin Bennett, Trading Volatility, Correlation, Term Structure and Skew (https://www.trading-
volatility.com/). 2

17. See Garman, M. B., and M. J. Klass, 1980, On the Estimation of Security Price Volatilities from
Historical Data, Journal of Business 53:67–78 (https://www.jstor.org/stable/2352358). 2 3

18. More precisely, Garman and Klass17 establish that a variation of σGK is the “best” reasonable estimator
but note that σGK is 1) more practical and 2) as efficient as this variation, which I will call the original
Garman-Klass volatility estimator σGKo . 2

19. See L. C. G. Rogers and S. E. Satchell, Estimating Variance From High, Low and Closing Prices, The
Annals of Applied Probability, Vol. 1, No. 4 (Nov., 1991), pp. 504-512
2 3
(https://www.jstor.org/stable/2959703).

20. Such a decrease in efficiency cannot be avoided because the Rogers-Satchell estimator belongs to
class of estimators studied in Garman and Klass17, so that its efficiency is necessarily smaller than the
efficiency of the Garman-Klass estimator (maximal by definition).

21. Garman and Klass17 provide a volatility estimator that takes into account opening jumps, but this
estimator has a dependency on an unknown f parameter which makes it unusable in practice; Yang
and Zhang12 show that this dependency is actually spurious and provide a usable form of this
estimator.
:
 22. When the time periods t are measured in trading days, opening jumps are called overnight jumps.

23. A single-period volatility estimator is a volatility estimator that can be used to estimate the volatility of
an asset over a single time period t using price data for this time period only; for example, the
Parkinson, the Garman-Klass and the Rogers-Satchell estimators are single-period estimators while
2
the close-to-close estimators are multi-period estimators.

24. C.f. also Molnar46 on this subject.

25. See Kunitomo, N. (1992). Improving the Parkinson method of estimating security price volatilities.
Journal of Business, 65, 295–302 (https://www.jstor.org/stable/2353167).

26. See Alizadeh ,S., Brandt, W. M., and Diebold, X.F., 2002. Range-based estimation of stochastic
volatility models. Journal of Finance 57: 1047-1091 (https://onlinelibrary.wiley.com/doi/10.1111/1540-
6261.00454).

27. See Meilijson , I. (2011). The Garman–Klass Volatility Estimator Revisited. REVSTAT-Statistical Journal,
9(3), 199–212 (https://revstat.ine.pt/index.php/REVSTAT/article/view/104).

28. See Jacob, J. and Vipul, (2008), Estimation and forecasting of stock volatility with range-based
estimators. J. Fut. Mark., 28: 561-581 (https://onlinelibrary.wiley.com/doi/abs/10.1002/fut.20321). 2 3

29. See Mapa, Dennis S., 2003. A Range-Based GARCH Model for Forecasting Volatility, MPRA Paper
21323, University Library of Munich, Germany (https://ideas.repec.org/p/pra/mprapa/21323.html).

30. See Chou, R.Y. (2005). Forecasting Financial Volatilities with Extreme Values: The Conditional
Autoregressive Range (CARR) Model. Journal of Money Credit and Banking, 37(3): 561-582
(https://www.jstor.org/stable/3839168).

31. See Harris, R. D. F., & Yilmaz, F. (2010). Estimation of the conditional variance–covariance matrix of
returns using the intraday range. International Journal of Forecasting, 26, 180–194
(https://www.sciencedirect.com/science/article/abs/pii/S0169207009000466).

32. See Shu, J. and Zhang, J.E. (2006), Testing range estimators of historical volatility. J. Fut. Mark., 26:
2
297-313 (https://onlinelibrary.wiley.com/doi/10.1002/fut.20197).

33. See Brandt, Michael W. and Kinlay, J, Estimating Historical Volatility (March 10, 2005)
2 3
(https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4384038).

34. See Lyocsa S, Plihal T, Vyrost T. FX market volatility modelling: Can we use low-frequency data? Financ
Res Lett. 2021 May;40:101776. doi: 10.1016/j.frl.2020.101776. Epub 2020 Sep 30
(https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7526631/#bib0024). 2 3 4 5

35. See Patton A.J., Sheppard K. Optimal combinations of realised volatility estimators. Int. J. Forecast.
2009;25(2):218–238 (https://www.sciencedirect.com/science/article/abs/pii/S0169207009000107).

36. The Yang-Zhang volatility estimator is excluded to avoid mixing jump-adjusted volatility estimators
with non-jump-adjusted ones.

37. (Adjusted) prices have have been retrieved using Tiingo (https://api.tiingo.com/). 2 3

38. The jump-adjusted Yang-Zhang volatility estimator, as well as the close-to-close volatility estimators,
:
 38. The jump-adjusted Yang-Zhang volatility estimator, as well as the close-to-close volatility estimators,
require the closing price of the last day of the previous month as an additional price. 2 3

39. This period more or less matches with the period used in Sepp4.

40. The next month’s close-to-close volatility is then taken as a proxy for the next month’s realized
volatility; this choice is important, because different proxies might result in different conclusions as to
the out-of-sample forecast performances.

41. See Butler, Adam and Philbrick, Mike and Gordillo, Rodrigo and Varadi, David, Adaptive Asset
Allocation: A Primer (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2328254).

42. See Mincer, J. and V. Zarnowitz (1969). The evaluation of economic forecasts. In J. Mincer (Ed.),
Economic Forecasts and Expectations (https://econpapers.repec.org/bookchap/nbrnberch/1214.htm).

43. This is due to slight differences in methodology, with mainly 1) the definition of “monthly volatility” in
Sepp4 taken to be the volatility from the 3rd Friday of a month to the 3rd Friday of the next month and
2) the usage in Sepp4 of a linear regression model robust to outliers.

44. The common ending price history of all the ETFs is 31 August 2023, but there is no common starting
price history, as all ETFs started trading on different dates.

45. For example, for all investors running some kind of monthly tactical asset allocation strategy.

46. See Peter Molnar, Properties of range-based volatility estimators, International Review of Financial
Analysis, Volume 23, 2012, Pages 20-29,
(https://www.sciencedirect.com/science/article/abs/pii/S1057521911000731). 2 3

! Tags: volatility

" Updated: September 20, 2023

: