See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228431910

Volatility Cones and Their Sampling Properties

Article in The Journal of Derivatives · October 2002

DOI: 10.3905/jod.2002.319188

CITATIONS READS

15 2,052

2 authors, including:

Robert Tompkins
Rotterdam School of Management
24 PUBLICATIONS 496 CITATIONS

SEE PROFILE

All content following this page was uploaded by Robert Tompkins on 15 September 2014.

The user has requested enhancement of the downloaded file.

 The Sampling Properties of Volatility Cones
By Stewart Hodges* and Robert Tompkins#

* Financial Options Research Centre, University of Warwick, Coventry, CV4 7AL, Telephone:
44-1203-523-606, Facsimile: 44-1203-524-167, email: [email protected].
#
Department of Finance, University of Technology, Vienna, Favoritenstrasse 11, A-1040, Vienna,
Austria, Telephone: 43-1-726-0919, Facsimile: 43-1-729-6753, email: [email protected].

This piece of was research was partially supported by the Austrian Finance Foundation (FWF) under
grant SFB #10 (Adaptive Information Systems and Modelling in Economics and Management Science).
Further financial support was provided by Corporate Members of the Financial Options Research
Centre.
 The Sampling Properties of Volatility Cones
ABSTRACT

In this research, we extend the original work on volatility cones by Burghardt and Lane
(1990) to consider of the sampling properties of the variance of variance (and the
standard deviation of volatility) under a rich class of models that includes stochastic
volatility and conditionally fat-tailed distributions.

Because the volatility cone examines volatility at quite long horizons, the estimation
requires the use of overlapping data. This theory confirms the casual observation that
the estimation of the variance of variance is downward biased when estimation is
done on an overlapping basis. Our principal contribution is to identify what this bias is
and derive an adjustment factor that approximates an unbiased estimate of the true
variance of variance when overlapping data is used. Another contribution is the
derivation of a formula that describes the variance of the quadratic variation over
different time horizons.

Using the theory presented, we tested the bias adjustments to the standard deviation of
volatility using simulations. Two cases were examined: a GBM i.i.d. process and a
non-i.i.d. process associated with the stochastic volatility model suggested by Heston
(1993). In both cases, the bias introduced by estimation of volatility with overlapping
data becomes insignificant after making the theoretical adjustments..

These results are relevant to those who must sell options and must understand the
nature of quadratic variation in asset prices. This will provide clearer insights into the
nature of hedging errors when dynamically hedging options.

This research also suggests a new method for the estimation of stochastic volatility
models, where estimation over a long horizon is likely to provide robustness not
associated with current methods.

JEL classifications: C13, G13

Keywords: Overlapping Data Observations, Volatility Estimation, Quadratic

variation, Volatility Cones, Stochastic Volatility

1
1. INTRODUCTION
The seminal work on option pricing by Black and Scholes (1973) launched the field of

contingent claims analysis. Ever since, the sampling properties of realized volatility

over different time horizons has been of some concern to academics and practitioners

alike. However, the first published empirical examination of the average levels of

volatilities at different time horizons seems to lie in the 1990 work of Burghardt and

Lane. Since then the One critical assumption of this paper was that the volatility of

the underlying asset was known and constant. Volatility is defined as the square root

of the prospective (annualised) quadratic variation. Quadratic variation is the integral

[or approximated as the sum] of squared returns over some time period. Despite the

success of the Black Scholes approach, it is generally agreed that volatility is neither

known nor constant and that forecasting it is paramount in successful derivatives

trading.

Merton (1973) relaxed the assumption of constant volatility. He showed that the

Black Scholes formula still yields a unique option price and the appropriate riskless

hedge by incorporating the (time varying) quadratic variation of the underlying price

process integrated over the life of the option. In subsequent research, Hull & White

(1987) showed that for a certain class of stochastic volatility situations, option values

can be obtained as integrals over the probability distribution of future (realised)

volatility. These papers establish the importance of the prospective quadratic variation

in both deterministic and stochastic volatility settings.

Neuberger (1992) demonstrated the strength of the link between realised quadratic

variation and the profit or loss on a hedged options position. He shows that by

dynamic hedging against a static options position, it is possible to engineer the future

profit [or loss] as an exact linear function of the realised quadratic variation.

Therefore, option traders must understand the sampling properties of the volatility for

the time horizon of the option, not only to determine the expected average level but

also to estimate the possible gains or losses.

2
Burghardt and Lane (1990) were the first to examine empirically the average levels of

volatilities at different time horizons and consider the variability of the realisations.

Using sample these samples of realised volatilities (measured on an overlapping basis),

they determined the average levels and the maximum/minimum ranges. Neither

Burghardt and Lane (1990) nor subsequent researchers examined the sample properties

of these estimates. It is apparent that to apply this approach for forecasting, we must be

sensitive to the biases introduced by the use of overlapping data. The purpose of this

paper is to examine these biases. We will restrict our analysis initially to the variance

of the quadratic variation and use these results to explore the effects of overlapping

biases on the standard deviation of the volatility.

We develop a theory of the sampling properties of volatility estimated on an

overlapping basis. The resulting estimate of the bias can be used to determine an

adjustment to the standard deviation of the volatility estimated using overlapping data.

This yields a truer picture of the future variability of asset prices. The major

contribution of this research is to develop a theory of the variance (or standard

deviations) of quadratic variation (volatility) in asset price processes over different

time horizons.

The sampling properties of cones will be analyzed under a variety of different models.

These models will include jump processes and stochastic volatility. We argue that

cones are relevant as they reflect the sampling properties of volatility at a range of

time horizons. In contrast, most other methods used in understanding volatility

dynamics are estimated directly from daily returns (e.g. ARCH, GARCH models

estimated using maximum likelihood methods) which makes them less robust to

specification error. Our analysis is a prerequisite for the development of estimation

procedures that fit simultaneously across different time horizons [see Tompkins

(2000)].

The paper is organised as follows. The second section provides a brief literature review

and develops a theory of sampling properties of variance estimation using overlapping

data. Initially, we will focus on the variance of quadratic variation in asset price
3
processes over different time horizons. This will entail a detailed description of the

implicit weighting scheme for observations when data is overlapping. This is

developed for a general model related to the Heston (1993) volatility process [which is

similar to the Cox, Ingersoll and Ross (1985) paper on interest rates]. Nested in our

model are all i.i.d. processes, which are considered ahead of the stochastic volatility

case. In this section, we precisely quantify the bias in estimating the variance of

quadratic variation and approximately quantify the bias for the standard deviation of

volatility. The third section tests this model by simulating the sampling properties of

the standard deviation of volatility for an i.i.d. model and a stochastic volatility model.

Using our sampling theory derived in the second section, we assess how effectively we

can form unbiased estimates of the standard deviation of volatility at different time

horizons. A summary and implications of these results follow.

2. A THEORY OF THE SAMPLING PROPERTIES OF VOLATILITY CONES

One crucial assumption of Black and Scholes (1973) analysis of option pricing is that

the volatility of the underlying asset is known and constant. Volatility is defined as the

square root of the prospective (annualised) quadratic variation. Quadratic variation is

the integral [or approximated as the sum] of squared returns over some time period.

Despite the success of the Black Scholes approach, it is generally agreed that volatility

is neither known nor constant and that forecasting it is paramount in successful

derivatives trading.

Merton (1973) relaxed the assumption of constant volatility. He showed that the

Black Scholes formula still yields a unique option price and the appropriate riskless

hedge by incorporating the (time varying) quadratic variation of the underlying price

process integrated over the life of the option. In subsequent research, Hull & White

(1987) showed that for a certain class of stochastic volatility situations, option values

can be obtained as integrals over the probability distribution of future (realised)

volatility. These papers establish the importance of the prospective quadratic variation

in both deterministic and stochastic volatility settings.

4
Neuberger (1992) demonstrated the strength of the link between realised quadratic

variation and the profit or loss on a hedged options position. He shows that by

dynamic hedging against a static options position, it is possible to engineer the future

profit [or loss] as an exact linear function of the realised quadratic variation.

Therefore, option traders must understand the sampling properties of the volatility for

the time horizon of the option, not only to determine the expected average level but

also to estimate the possible gains or losses.

Burghardt and Lane (1990) were the first to examine empirically the average levels of

volatilities at different time horizons and consider the variability of the realisations.

They achieved this by use of the volatility cone technique. This method measures the

unconditional volatility for different time horizons using a given set of underlying

returns. With these samples of realised volatilities (measured on an overlapping

basis), they determined the average levels and the maximum/minimum ranges. The

reason why their technique is referred to as a volatility cone is that the difference

between the maximum and minimum levels narrows as the time horizon of volatility

estimation is extended.

The choice of the range as a measure of variability is somewhat arbitrary (although it

suited the purpose of their research). Neither Burghardt and Lane (1990) nor

subsequent researchers examined the sample properties of volatility cones, yet it is

readily apparent that the use of overlapping data must introduce biases into these

estimates.

Consider a typical procedure in which, following from Burghardt and Lane (1990),

volatility cones are estimated using daily price relatives (natural logarithm of closing

prices) for the entire period of analysis. Assuming the time period of ten years of

trading days (250 trading days per year), the number of daily price relatives will be

approximately T=2500. These price relatives will be grouped into periods of analysis

from a minimum of one day (h=1) to a maximum period of two years (h=500) in one

day increments. With these selected horizon periods, standard deviations (or the

volatility) can be estimated for the 500 horizon periods.

5
For the estimates of the standard deviation to be unbiased, observations must be

independent. The usual manner to achieve this is to restrict the analysis solely to non-

overlapping data. When this is done, the sampling properties of such an approach are

well known. An equally well known problem is that as the time horizon of estimation

is extended, the number of non-overlapping samples is reduced (n = T/h), sampling

properties are unable to be determined as the degrees of freedom are also reduced.

Said simply, the estimation method runs out of samples. The penultimate problem

exists for the time horizon h = T, where only one estimate is possible. An alternative

is to estimate the standard deviation on an overlapping basis. Now, the number of

observed samples, n, is increased (n = T−h+1) but they are no longer independent.

Unfortunately, the sampling properties of the standard deviation will be biased due to

interdependence in the samples. This is what is examined in this paper.

In this research, we focus initially on the variance of the variance. The high degree of

correlation between such overlapping samples will dampen the true variability of the

variance and will provide little insight into its true nature. A number of authors have

considered this problem. One method of addressing the problem of overlapping data

in variance estimation is the use of panel regression techniques. Hansen and Hodrick

(1980) first suggested this approach. Dunis and Keller (1995) modified this for the

examination of currency option volatilities. Other proposed solutions to the

overlapping data problem include the bootstrap method proposed by Efron (1983) and

the jack-knife method suggested by Yang and Robinson (1986).

The approach adopted here is to develop a theory of the biases of the estimated

variances for overlapping and non-overlapping periods. We examine these variances

for a range of time horizons under a rich case of processes, which nest stochastic

volatility and all i.i.d. processes (including potentially non-normal ones).

2.1 Assumptions
We will consider the sampling properties of volatility cone analysis conducted on a

time series containing t = 1,.., T observations of single period security returns rt. The
6
analysis focuses on the variance properties of these returns. With daily (or even

weekly) returns data, variance estimates are unaffected by the level of the mean return.

Without any real loss of generality we can therefore make the convenient assumption

that each return rt is drawn from a distribution with zero mean. We assume the

distribution of rt, conditioned on the information at the beginning of the return period,

is constant in shape, but that its variance vt changes through time. In particular we

assume that its kurtosis exists as a finite constant K. The fact that vt may have a

complex autocorrelation structure precludes the possibility of devising any simple

bootstrapping method.

Formally, we will assume that vt is a stationary stochastic process with long run mean
of v . We can write the squared returns as:
( )
st = rt 2 = vt 1 + K − 1 ε t , where εt has zero mean, unit variance, is i.i.d. and

vt = v + u t , where ut is independent of εt. (1)

Note that if our returns are measured over time intervals of length ∆t then
v = σ 2 δt , where σ 2 is the average variance rate.

2.2 Weighting Of Observations

We will consider the sampling properties of variance estimates based on overlapping
(contiguous) sub-series of data of length h. For a given horizon of length h, the

number of distinct sub-series available is n = T – h +1. Our analysis is restricted to

cases where h < T/2, which is also a sensible practical restriction. If we restricted

ourselves to non-overlapping sub-series we would have at most T/h observations

available on the quadratic variation (measured as the average variance rate) to horizon

h.

Each variance estimate for the cone is obtained as a weighted average of the single
period squared returns, st = rt2 . The variance estimate, φi, from the i sub-series is
th

simply
1 i + h−1
φi = ∑ st . where i = 1,.. , n. (2.1)
h t =i

7
We can more conveniently regard it as the vector product
φi = w′i s (2.2)

where wi and s are T×1 vectors, and wi has values of 1/h in elements i through
i + h − 1 and zeros elsewhere. The mean of these variance estimates, φ , weights the

squared returns with a mean weighting vector, m, as

′
1 n 1 n 
φ = ∑ φi =  ∑ w i  s = m′ s . (3)
n i =1  n i =1 

The deviations of these estimates from their mean, zi, are given by
z i = φ i − φ = (w i − m)′ s = x ′i s .

(4.1)
Finally, by stacking the n (1 × T) row-vectors x ′i = ( w i − m ) ′ vertically on top of each

other to form an n × T matrix X′, we can represent the n × 1 vector z of deviations, zi,

from the mean as

z = X′ s . (4.2)

2.3 The Cross-Sample Variance of Estimated Variances

By construction, the zi sum to zero. The cross-sample variance of h-period variances
obtained from the overlapping volatility cone technique is therefore given by
1
θˆ = z ′ z . (5.1)
n
We are interested in the behaviour of this, and in particular in its bias.
1 1 1
θˆ = z ′ z = s ′ X X′ s = 1′ (s s ′) ! ( X X′) 1 , (5.2)
n n n
where ! denotes the Hadamard product cij = a ij × bij of corresponding elements of a

matrix and 1 is a T × 1 vector of ones.

The expected variance of variance is obtained as:

1
E[θˆ] = 1′ V ! W 1 , (6)
n
where V = E[ss′] and W = XX′. In other words, we form the two T × T matrices V

and W, sum the products of their corresponding elements and divide by n.

8
We will consider the V and W matrices separately. We will first describe the
structure of X′ and W in a little more detail. We will examine V and the E[ θˆ ] which

results from the product with W. We do this first for the constant volatility (i.i.d.)

case as this is the most tractable, and then provide an extension to the stochastic

volatility one.

Table 1 Summary of Principal Notation and Assumptions

Single period returns rt (where t = 1,…, T) are drawn from a distribution with mean
zero, variance vt, and kurtosis K. These are all conditioned on the information at t−1.
vt is a stationary stochastic process with long run mean of v .
( )
st = rt2 = vt 1 + K − 1 ε t where εt has zero mean, unit variance and is i.i.d.

vt = v + ut , where ut is independent of εt.

z = X′ s is the n × 1 vector of deviations of the variances from the sample mean

where X′ is a n × T matrix of fixed weights which are applied to the T × 1 vector s of squared
returns. The cross-sample variance of the variance is:
1 1 1
θˆ = z ′ z = s ′ X X′ s = 1′ (s s ′) ! ( X X′) 1 ,
n n n
where ! denotes the Hadamard product cij = a ij × bij of corresponding elements of a matrix.
The expected variance of variance is obtained as:
1
E[θˆ] = 1′ V ! W 1 ,
n
where V = E[ss′] and W = XX′.

2.4 The Structure of The X′ Matrix

The basic structure of the X′ matrix has already been outlined above, but because the

form of XX′ is central to our analysis, it is worth providing rather more detail. Figure 1
th
shows the nature of the weights involved: it shows the construction of the 4 row of the

matrix X′ for the case T = 20, h = 5.

9
 W e ig h tin g o f O b s e rv a tio n s

0 .2 0 G ro s s : w 4
M e a n : m
0 .1 5 N e t: x 4

0 .1 0
Weight

0 .0 5

0 .0 0

-0 .0 5

-0 .1 0
1 2 3 4 5 6 7 8 9 1 0 1 1 1 2 1 3 1 4 1 5 1 6 1 7 1 8 1 9 2 0

O b s e rv a tio n

Figure 1, Weighting of Observations

For any row i of the matrix, the “raw” weights, w′i, contains a group of h contiguous
1/h’s, and the remaining T−h entries are zeros. The average of these, m′, is subtracted

from each w′i to obtain the corresponding x′i row of X′. Table 2 shows the form of

the coefficients in this matrix. A property we shall later make use of is that both the

column sums and the row sums of this matrix are zero.

Table 2 Form of the X′ Matrix (multiplied by nh)

n−1 n−2 n−h −h −h −h −h −2 −1
−1 n−2 n−h n−h −h −h −h −2 −1
−1 −2 n−h n−h n−h −h −h −2 −1
−1 −2 −h n−h n−h n−h −h −2 −1
−1 −2 −h −h n−h n−h n−h −2 −1
−1 −2 −h −h −h n−h n−h n−2 −1
−1 −2 −h −h −h −h n−h n−2 n−1

(h −1) columns (n−h+1) columns (h − 1) columns

[Dimensions correspond to the case T = 9, h = 3, n = 7, with n rows and T = n + h − 1 columns.]

2.5 The Structure Of The W Matrix

We next consider the form of the W matrix ( = XX′). Table 3 continues the previous
2
example to show the general form of the W matrix ( = XX′) multiplied by nh . The

10
matrix is symmetric, and the upper triangle provides a guide to the general formulae

which are given below the table.

2
Table 3 Form of the W (=XX′) Matrix (multiplied by nh )

[This table also with T = 9, h = 3, n = 7, and W is T × T.]

6 e11 e12 b1 b1 b1 b1 a12 a11
5 10 e21 f22 b2 b2 b2 a22 a21
4 8 12 g1 g2 c c b2 b1
-3 1 5 12 g1 g2 c b2 b1
-3 -6 -2 5 12 g1 g2 b2 b1
-3 -6 -9 -2 5 12 g1 f22 b1
-3 -6 -9 -9 -2 5 12 e21 e12
-2 -4 -6 -6 -6 1 8 10 e11
-1 -2 -3 -3 -3 -3 4 5 6

The formulae for these cell values are:

aij = −ij, bi = −ih, c = −h , eid = i (n − i − d), fid = h (n − i ) − d n , and

2 2

gd = h (n − h) − d n ,
2

where i, j < h denote a row or column distance from the edge of the matrix, and d < h

denotes a distance from the diagonal of the matrix.

The values of the diagonal itself are i (n − i) for i < h, and h (n − h) elsewhere.

As before, this example is for T = 9, h = 3, n = 7.

2.6 The I.I.D. Case

In the case where the underlying volatility is constant, we have vt = v for all t, and st

simplifies to:
(
st = rt2 = v 1 + K − 1 ε t ) (7)

where εt has zero mean, unit variance and is i.i.d. This makes V a diagonal matrix
with entries on the diagonal equal to ( K − 1) v 2 . (8)

We will consider first the true variance of the rate of quadratic variation in samples of

length h, and then the variance obtained from overlapping estimates.

The annualized variance rate σ is v divided by the length of the time interval ∆t. For
2

a sequence of h returns, the realized (annual) rate of quadratic variation is

11
 1 h
∑ st / ∆t . (9)
h t =1

Its mean is σ and its variance is (K − 1) σ / h .

2 4

We next derive the comparable expression for estimates derived from volatility cones

employing n overlapping sub-series each containing h returns.

In this i.i.d. case, since V is diagonal with constant elements, we require the sum of

the diagonal elements of W (or its trace).

From the expressions derived previously, it is clear that this sum amounts to:
h −1
2∑ i (n − i ) + h(n − h)(n − h + 1)
trace( W) = i =1
(10.1)
nh 2
which simplifies to
3n 2 − 3nh + h 2 − 1
trace( W ) = . (10.2)
3nh
The variance of the estimated annualized quadratic variation is this multiplied by

(K − 1) σ / n, which gives
4

( K − 1) σ 4 h h2 −1
E[θ ] =
ˆ (1 − + ), (11)
h n 3n 2

so the variance obtained from the cone is too small by a factor of the expression on the
 h 
right hand side, which is approximately (1 − h/n), or roughly 1 −  in terms of
 T −h

the total number of observations, T . From this we can derive the following rule of

thumb: to maintain a (proportional) bias (before adjustment) of less than 1/k, we

1
should restrict k not to exceed .
k +1
Under the hypothesis of an i.i.d. process, we can adjust for this bias by multiplying

quadratic variation estimated from overlapping data by the adjustment factor:

1
. (12)
h h2 −1
1− +
n 3n 2

12
2.7 The Stochastic Volatility Case
We have in mind the class of stochastic volatility models which provide the a negative

exponential covariance structure for the covariance of the asset’s returns. A

particularly suitable example was first suggested by Cox, Ingersoll and Ross (1976)

for interest rates and subsequently applied to variances by Heston (1993). The

continuous-time variance process can be expressed as:

dv(t ) = α [σ 2 − v(t )]dt + ξ v(t ) dz 2 (t ) (13)

and the spot asset at time t follows the diffusion

dS (t ) = µSdt + v(t ) S dz1 (t ) (14)

where α, ξ and µ are constants and dz1 (t ), dz 2 (t ) are increments of Brownian

Motions with correlation ρ. Although for these diffusion processes, the disturbance
term for the spot asset price is Gaussian, under our discrete time formulation we

permit a more general distribution with kurtosis K. Under the process we have just

described, the unconditional covariance between vt at any two of our (discretely

labelled) dates s and t is given by:

σ 2 ξ 2 −α t − s δt
Cov[v s , vt ] = E[u s u t ] = e . (15)
2α
In this stochastic volatility case, we now have to work with the full structure of both

the V and W matrices. The above equation provides the form of the off-diagonal
σ 2ξ 2
entries of V. The term is the steady state variance of the variance and α is the
2α
rate of mean reversion.

It should be noted that this is ignoring transient effects from having a single

starting point. For known v0 we would instead have1:

E [u s u t ] =
σ 2ξ 2 −α t − s ∆t
2α
⋅e ( )
× 1 − e − 2α s ∆t for s < t. (16)

However, this effect dies out rapidly from the mean reversion of the model and

thus, we concentrate on the analysis of equation (15) from this point forward.

The form of the V matrix is therefore:

1
For O-U processes only: C-I-R seems more complicated

13
 σ 2ξ 2 −α j − i δt
vij = ( K e + ( K − 1) σ 4 ) ∆t 2 , for i = j , and
2α
(17)
σ ξ −α j − i δ t 2
2 2
= e ∆t , for i ≠ j.
2α

This has a banded diagonal structure.

2
As before, the variance of the annualized quadratic variation comes from 1′V 1/h ,

where for this purpose V is taken to be h × h.

This gives:
2c h −1
E[θˆ] = [( K − 1) σ 4 + Kc + ∑ (h − i) a i ] / h
k i =1
(18)
2ac a h + h(1 − a) − 1
= [( K − 1) σ + Kc +
4
] / h, where
h (1 − a ) 2
σ 2ξ 2
c= , and a = e −α ∆t .
2α
This equation if of importance in its own right, and provides a temporal pattern for the

time profile of variance quite different from the i.i.d case.

In the case of overlapping observations we have to combine the V and W matrices.

The particular negative exponential form of V means that in order to obtain the

variance of the quadratic variation, besides the terms from the diagonal, we need to

evaluate sums of the form:

∑ a d ∑ wi, i +d .
d i
(19)

The sums we require can all be expressed analytically. The expressions they give rise

to are rather complicated, so we have given them in Appendix B rather than in the

main body of the paper. They provide us with analytic expressions for the variance of

the annualized quadratic variation from overlapping samples in the presence of both

stochastic volatility and conditional excess kurtosis.

2.8 The Standard Deviation of Volatility

So far we have provided formulae for the variance (and hence the standard deviation)

of the quadratic variation, expressed as an annual rate. In this section we show how

14
the preceding results may be (approximately) re-expressed in terms of the standard

deviation of volatility by using a Taylor series expansion.

We have obtained estimates of the mean, M, and standard deviation S of a quadratic

variation measure, q. We now need the mean and standard deviation of the volatility

σ = √ q. Write:
S
q = M (1 + ε) (20)
M
where ε has zero mean and unit standard deviation. Expanding σ as a Taylor series

we obtain:
S S S2 2 S3
σ = M 1+ ε = M (1 + ε− ε + ε 3 + ..) (21)
M 2M 2 3
8M 16 M

We therefore have the approximate result that:

S
SD(σ ) ≈ . (22)
2 M

The closeness of this approximation depends both on the ratio of S/M and on the

higher moments of ε. We can examine it directly for particular cases. The

approximation is often rather poor. However, because of the way we apply it to two

variances to obtain the ratio of a true to a biased standard deviation, most of the error

cancels out and it gives rise to only a second order error in our estimates.

3. TESTING THE MODELS

To assess the effectiveness of the adjustment to the standard deviation of volatility in

correcting the biases introduced by overlapping estimation, models were tested by

simulation.

In this simulation, two cases were considered: an i.i.d. (GBM) case2 and a

stochastic volatility process consistent with equations (13) and (14). Upon fixing

model parameters, a 100 year long time series of daily prices was simulated (with 252

2
We also considered alternative i.i.d. processes, which included a Student-t process. Results were
similar to those found for the i.i.d. GBM process.

15
trading days in each year).3 For both simulations, the variance (σ 2 ) used for the

generation of prices was set equal to 0.04 (or 20% volatility). For the simulations,

prices were generated using the standard Euler approach (for discrete increments in

time of one day, i.e. 1/252) consistent with equation (14):

∆S = Sµ∆t + Sσ∆z1 (23)

In this equation, we assumed the interest rate was zero, so the only adjustment to the

drift came from the risk neutral adjustment associated with a GBM process.

For the stochastic volatility case, a time series of daily variances for the same

period of 100 years was estimated. Again, this assumes discrete daily increments and

also relied upon an Euler approach. The discrete time version of equation (13) can be

expressed as:
∆vt = α σ 2 − vt −1  ∆t + ξ vt −1 ∆z2 (24)

The two disturbance terms ( ∆z1 , ∆z 2 ) are independent draws from a Gauss-Wiener

process. These were determined using a standard Box-Müller method and used the

antithetic approach suggested by Boyle (1977). For the stochastic volatility

model, the rate of mean reversion α was set to 4.00, the volatility of the variance ξ
was set to 0.6 and the long-term variance σ 2 was set equal to 0.04. As indicated

previously, the effects of the stochastic volatility will be to increase the kurtosis of the

unconditional returns. Within our sample, the unconditional kurtosis of daily returns

was 6.374.

Using this data set of 100 years, the true sampling properties of the volatility

cone were estimated with good precision4. The returns of the daily time series were

estimated using differences in the natural logarithm of prices. This result was then

annualised. This can be expressed as:

ln (St/St-1) x √252 (25)

3
Actually, 102 years of data were determined. However, The first two years of simulated prices were
eliminated so that the starting values were randomised with the correct distribution.
4
For the IID case, the theoretical standard deviation of volatility is approximately σ/√(2h). In Table 1A
in Appendix A, these have been presented along with the standard deviation of volatility from the
simulated 100 years. The errors in the simulation are within 3% of the theoretical values.

16
 With this series of daily returns, the standard deviation was determined at time

horizons from 20 days to 500 days in 20-day increments. Thereafter, we determined

the standard deviation of the standard deviation (volatility) at each time horizon.

Then the 100 years of data was split into 20 samples of five years each. In each

of these sub-periods, the volatility cones were re-estimated using overlapping data.

From these 20 sub-periods, the standard deviations of the volatility were determined

for the same time horizons. The average of the standard deviations of volatility (across

the 20 sample periods) was compared to the true standard deviation of volatility

determined using the entire 100-year sample. Finally, the square root of the

adjustment factors derived previously for the variance of variance were multiplied by

the average standard deviations of volatility to assess if the bias had been corrected.

The results of these simulations appear in Figure 2 for the IID case and in Figure 3 for

the stochastic volatility case. In Appendix A, Tables A1 and A2 appear and provide

the numerical summaries of these cases, including the adjustment factor and

significance tests of the biases prior and post correction.

For the 100-year period, the standard deviation of volatility was estimated for

all time horizons (using overlapping data). For each of the 20 sub-periods, the

standard deviations of volatility were also estimated using overlapping data. These

were averaged and this was compared to the estimate of the standard deviation for the

100-year period. Both absolute and percentage differences were computed. In Figure

2, results are presented for the IID case. In this Figure, the left two panels consider the

unadjusted case. The top panel plots the relationship between the true standard

deviation of volatility (as a dashed line) and the unadjusted average of the standard

deviations of volatility (as a solid line) relative to the time horizon of estimation. It is

clear that the unadjusted average is systematically biased downwards and is directly

related to the time horizon of estimation. The bottom panel considers the absolute

difference in these two standard deviations and further indicates the relationship

between the degree of bias and the time horizon.

17
 The two panels on the right-hand side show the adjusted average standard

deviation of volatility. This is determined by multiplying the averaged standard

deviation by the adjustment factor formula [equation (12)]. In the upper panel, the two

series are plotted as absolute levels of volatility, and they coincide almost exactly. The

lower panel plots the percentage differences between the two series. The maximum

deviation is +3.42%. In Table A1, t-tests indicate that a statistically significant bias

exists when overlapping the data (|t| > 1 for h > 60, and |t| > 2 for h > 260). After

correction, there are no cases where the bias is significant.

Figure 3 provides the same information for the stochastic volatility case. We

can see clearly that the presence of stochastic volatility has two effects. First, the

standard deviation of the volatility now decreases at a slower rate as the time horizon

is increased: the 1/√n rule no longer applies. Second, the bias from overlapping
th
observations has increased. At the 500 time horizon, the bias is –34.83%, whereas

for the i.i.d. case it was –28.91%. Retaining the same scales, the right-hand side

panels of Figure 3, display the adjusted average standard deviations of volatility. This

difference varies around zero and is no longer monotonically decreasing in the time

horizon. The most extreme percentage difference is now only –2.69%. In Table A2, t-

tests again indicate significant biases in the unadjusted estimates (|t| > 1 for all h, and |

t| > 2 for h > 200), but not in the adjusted estimates.

18
4. SUMMARY AND CONCLUSIONS
The sampling properties of realized volatility over different time horizons

concern both academics and practitioners. For academic research, this will provide

information about volatility dynamics beyond what is currently modelled using daily

data. For traders of option contracts, a better understanding of the sampling properties

of quadratic variation will provide a better forecast of possible gains or losses when

dynamically hedging these products.

In this research, we have described how volatility cones can provide

information about the sampling properties of volatility measured at different horizon

dates. We extend the original work on volatility cones by Burghardt and Lane (1990)

and present the sampling properties of the variance of variance (and the standard

deviation of volatility). The original approach gave biased information about the

variability of volatility because of its use of overlapping data. By developing a theory

of the sampling properties of volatility cone estimates, we are able to address and

correct for this bias.

We derive expressions for the variance of the variance for a general model,

which nests stochastic volatility models and alternative price processes to Gauss-

Wiener diffusions. This theory confirms the casual observation that the estimation of

the variance of variance is downward biased when estimation is done on an

overlapping basis. Our main contribution is to identify what this bias is and derive an

adjustment factor that approximates an unbiased estimate of the true variance of

variance when overlapping data is used. Equation (18) is also potentially important: it

describes the variance of the quadratic variation over different time horizons and

under a rich class of models that includes stochastic volatility and conditionally fat-

tailed distributions.

To put our work into the same metric as Burghardt and Lane (and most

derivatives practitioners), we have also extended the theoretical analysis to the

standard deviation of volatility estimated at the various horizons. This part of our work

relies on an approximation. The bias adjustments to the standard deviation of

19
volatility were testing using simulations. Two cases were examined: a GBM i.i.d.

process and a non-i.i.d. process associated with the stochastic volatility model

suggested by Heston (1993). For both cases, the bias in the estimates was significant

before the adjustment was made and insignificant afterwards.

This research has a number of implications. Clearly our results are relevant to

those who must sell options and must understand the nature of quadratic variation in

asset prices. This should lead to clearer insights into the nature of hedging errors when

dynamically hedging options.

Another application is the use of unbiased sampling properties of long-term

volatilities to estimate stochastic volatility models. This research allows an unbiased

estimate of the volatility of volatility, which is so critical to these models. Most other

methods for estimation rely directly on daily returns, which makes them less robust to

specification error. Our analysis is a prerequisite for the development of estimation

procedures that fit simultaneously across longer time horizons.

20
References:

Black, Fischer and Myron Scholes (1973), “The Pricing of Options and Corporate
Liabilities,” Journal of Political Economy, Volume 81, Number 2, pp 637-654.

Boyle, Phelim P. (1977), “Options: A Monte Carlo Approach,” Journal of Financial

Economics, Volume 4, Number 3, pp. 323-338.

Burghardt, Galen and Morton Lane (1990), “How to Tell if Options are Cheap,”
Journal of Portfolio Management, Volume 16, Number 2 (Winter), pp. 72-78.

Cox, John C., Jonathan E. Ingersoll and Stephen A. Ross (1985), “A Theory of the
Term Structure of Interest Rates,” Econometrica, Volume 53, Number 2, pp. 385-407.

Dunis, Christian and Andre Keller (1995), “Efficiency tests with overlapping data: an
application to the currency options market,” The European Journal of Finance,
Volume 1, pp. 345-366.

Efron, B (1983), “Estimating the error rate of a prediction rule: improvements on

cross-validation,” Journal of the American Statistical Association, Volume 78,
pp.316-331.

Hansen, Lars Peter and Robert J. Hodrick (1980), “Forward Exchange Rates as
Optimal Predictors of Future Spot Rates: An Econometric Analysis,” Journal of
Political Economy, Volume 88, pp 820-853.

Heston, Steven L. (1993), “A Closed-Form Solution for Options with Stochastic

Volatility with Applications to Bond and Currency Options,” Review of Financial
Studies, Volume 6, pp. 327-343.

Hull, John and Alan White (1987), “The Pricing of Options on Assets with Stochastic
Volatilities”, The Journal of Finance, Volume 42, Number 2, pp. 281-300.

Merton, Robert (1973), “The Theory of Rational Option Pricing,” Bell Journal of
Economics, Volume 4, pp 141-183.

Neuberger, Anthony (1994), “The Log Contract,” Journal of Portfolio Management,

Volume 20, Number 2 (Winter), pp. 74-80.

Tompkins, Robert G. (2000), “Stock Index Futures: Stochastic Volatility Models and
Smiles, ” The Journal of Futures Markets, Forthcoming, 2000.

Yang M.C.K. and D.H. Robinson (1986), Understanding and Learning Statistics by
Computer, Volume 4 of Series in Computer Science, World Scientific Publishing Co.
Pte. Ltd, Singapore 1986

21
 Unadjusted Adjusted
Standard Deviation of Volatility Standard Deviation of Volatility
3.5
3.5

3.0 3.0

2.5 2.5

2.0 2.0

1.5 1.5

1.0 1.0

0.5 0.5

0.0 0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Difference in the Unadjusted Standard Deviation Difference in the Adjusted Standard

of Volatility Deviation of Volatility
0.05 0.05

0.00 0.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

-0.05 -0.05

-0.10 -0.10

-0.15 -0.15

-0.20 -0.20

Figure 2, Differences in Unadjusted & Adjusted Standard Deviation of Volatility (I.I.D. Case)
 Unadjusted Adjusted
Standard Deviation of Volatility Standard Deviation of Volatility
10.0 10.0

9.0 9.0

8.0 8.0

7.0 7.0

6.0 6.0

5.0 5.0

4.0 4.0

3.0 3.0

2.0 2.0

1.0 1.0

0.0 0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Difference in the Unadjusted Standard Difference in the Adjusted Standard

Deviation of Volatility Deviation of Volatility

0.20 0.20
0.00 0.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
-0.20 -0.20
-0.40 -0.40
-0.60 -0.60

-0.80 -0.80

-1.00 -1.00

-1.20 -1.20

-1.40 -1.40

-1.60 -1.60

Figure 3, Differences in Unadjusted & Adjusted Standard Deviation of Volatility (Stochastic Volatility Case)
 IID case
True Standard Deviation of Volatility

testall 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 360 380 400 420 440 460 480 500
Theory Stdev 3.1623 2.2361 1.8257 1.5811 1.4142 1.2910 1.1952 1.1180 1.0541 1.0000 0.9535 0.9129 0.8771 0.8452 0.8165 0.7906 0.7670 0.7454 0.7255 0.7071 0.6901 0.6742 0.6594 0.6455 0.6325
testall Stdev 3.2027 2.2464 1.8185 1.5641 1.3939 1.2674 1.1764 1.1071 1.0495 1.0024 0.9629 0.9293 0.8973 0.8684 0.8428 0.8173 0.7923 0.7687 0.7466 0.7249 0.7062 0.6891 0.6720 0.6553 0.6398

test1 Stdev 3.1481 2.2285 1.8419 1.5787 1.3643 1.2113 1.0361 0.8903 0.7827 0.7201 0.6770 0.6617 0.6484 0.6391 0.6160 0.5782 0.5455 0.5160 0.5020 0.4911 0.4828 0.4786 0.4746 0.4625 0.4414
test2 Stdev 3.0869 2.1890 1.8111 1.6244 1.5033 1.3997 1.3308 1.2717 1.2367 1.2133 1.1998 1.1910 1.1857 1.1772 1.1655 1.1544 1.1331 1.1091 1.0866 1.0579 1.0257 0.9977 0.9710 0.9432 0.9110
test3 Stdev 3.3613 2.4175 1.9619 1.7240 1.5593 1.4056 1.2890 1.2147 1.1386 1.0657 1.0102 0.9700 0.9288 0.8977 0.8721 0.8497 0.8260 0.7969 0.7647 0.7272 0.6911 0.6592 0.6191 0.5760 0.5313
test4 Stdev 3.5916 2.6912 2.2589 2.0057 1.8266 1.6582 1.5518 1.4607 1.3903 1.3189 1.2398 1.1529 1.0813 1.0138 0.9562 0.9029 0.8442 0.7765 0.7113 0.6447 0.5939 0.5596 0.5201 0.4860 0.4610
test5 Stdev 3.1781 2.1938 1.7603 1.4090 1.1562 1.0018 0.9146 0.8569 0.8147 0.7901 0.7736 0.7526 0.7232 0.6853 0.6561 0.6424 0.6293 0.6169 0.6021 0.5771 0.5472 0.5139 0.4797 0.4527 0.4357
test6 Stdev 3.1085 2.0739 1.6684 1.3941 1.1893 1.0396 0.9415 0.9016 0.8718 0.8738 0.8613 0.8412 0.8199 0.7943 0.7777 0.7578 0.7509 0.7367 0.7224 0.7013 0.6771 0.6460 0.6242 0.6027 0.5791
test7 Stdev 2.9495 1.9844 1.5393 1.2879 1.1398 0.9962 0.9041 0.8501 0.8078 0.7944 0.7665 0.7513 0.7271 0.6973 0.6691 0.6474 0.6178 0.5937 0.5764 0.5475 0.5201 0.4907 0.4627 0.4394 0.4165
test8 Stdev 2.7341 1.6861 1.2470 1.0293 0.9111 0.8397 0.7793 0.7278 0.6838 0.6506 0.6241 0.5861 0.5535 0.5332 0.5161 0.4952 0.4758 0.4554 0.4264 0.4023 0.3728 0.3489 0.3157 0.2848 0.2571
test9 Stdev 3.1113 2.1487 1.8356 1.6248 1.4346 1.3256 1.2570 1.2119 1.1726 1.1314 1.0974 1.0686 1.0424 1.0194 0.9939 0.9718 0.9473 0.9270 0.9028 0.8762 0.8410 0.8091 0.7772 0.7438 0.7136
test10 Stdev 3.4088 2.4658 1.9792 1.6217 1.3271 1.0775 0.9344 0.8528 0.7871 0.7411 0.7127 0.6823 0.6617 0.6391 0.6218 0.6000 0.5690 0.5458 0.5279 0.5094 0.4882 0.4689 0.4481 0.4218 0.3891
test11 Stdev 3.3058 2.3561 1.9894 1.7766 1.6414 1.5456 1.4653 1.3970 1.3375 1.2725 1.2202 1.1658 1.1011 1.0266 0.9531 0.8875 0.8334 0.7770 0.7190 0.6670 0.6223 0.5725 0.5232 0.4767 0.4323
test12 Stdev 3.0515 2.1249 1.6423 1.3244 1.0907 0.9173 0.8187 0.7800 0.7249 0.6529 0.5765 0.5014 0.4348 0.3945 0.3855 0.3543 0.3210 0.2931 0.2665 0.2293 0.2166 0.2046 0.1844 0.1750 0.1758
test13 Stdev 2.8701 1.8023 1.3350 1.1296 1.0150 0.9397 0.8728 0.7843 0.7075 0.6466 0.5779 0.5120 0.4627 0.4108 0.3564 0.3124 0.2810 0.2553 0.2476 0.2561 0.2553 0.2561 0.2624 0.2756 0.2770
test14 Stdev 3.5141 2.5781 2.0556 1.7253 1.5063 1.3549 1.2365 1.1419 1.0808 1.0460 1.0010 0.9472 0.8879 0.8408 0.8033 0.7659 0.7245 0.6735 0.6192 0.5657 0.5029 0.4390 0.3847 0.3368 0.3020
test15 Stdev 3.0981 2.0792 1.7212 1.5439 1.3745 1.2329 1.1276 1.0494 0.9831 0.9170 0.8499 0.8002 0.7466 0.7087 0.6811 0.6488 0.6281 0.6111 0.5863 0.5584 0.5361 0.5122 0.4861 0.4621 0.4324
test16 Stdev 3.1377 2.0556 1.4870 1.0407 0.8673 0.8301 0.8131 0.7351 0.6517 0.6016 0.5870 0.5805 0.5515 0.5053 0.4728 0.4624 0.4593 0.4396 0.4117 0.3840 0.3760 0.3754 0.3609 0.3339 0.3018
test17 Stdev 3.5279 2.5515 2.0200 1.7432 1.5711 1.3816 1.2331 1.1422 1.0780 1.0262 0.9672 0.9037 0.8355 0.7792 0.7192 0.6593 0.6149 0.5809 0.5545 0.5342 0.5215 0.5075 0.4891 0.4625 0.4356
test18 Stdev 2.9241 1.9118 1.4756 1.1864 1.0755 1.0017 0.9661 0.9373 0.8993 0.8646 0.8260 0.7905 0.7565 0.7303 0.7030 0.6788 0.6522 0.6285 0.6103 0.5903 0.5659 0.5414 0.5136 0.4834 0.4540
test19 Stdev 3.1688 2.1941 1.7667 1.5466 1.4095 1.2948 1.2244 1.1809 1.1414 1.1027 1.0629 1.0241 0.9854 0.9503 0.9266 0.9120 0.8898 0.8690 0.8429 0.8217 0.8010 0.7775 0.7564 0.7341 0.7052
test20 Stdev 2.9715 2.0187 1.5745 1.3725 1.2275 1.1702 1.0812 1.0081 0.9471 0.8749 0.8353 0.7969 0.7525 0.7229 0.6796 0.6502 0.6235 0.5914 0.5671 0.5363 0.5204 0.5060 0.4857 0.4668 0.4447

Unadjusted Average 3.1624 2.1876 1.7485 1.4844 1.3095 1.1812 1.0889 1.0197 0.9619 0.9152 0.8733 0.8340 0.7943 0.7583 0.7262 0.6966 0.6683 0.6397 0.6124 0.5839 0.5579 0.5332 0.5069 0.4810 0.4548
Difference -0.0403 -0.0588 -0.0700 -0.0796 -0.0844 -0.0862 -0.0875 -0.0874 -0.0876 -0.0871 -0.0896 -0.0953 -0.1029 -0.1101 -0.1165 -0.1208 -0.1240 -0.1290 -0.1342 -0.1410 -0.1483 -0.1559 -0.1650 -0.1743 -0.1849
% Difference -1.26% -2.62% -3.85% -5.09% -6.05% -6.80% -7.44% -7.89% -8.35% -8.69% -9.30% -10.25% -11.47% -12.68% -13.82% -14.78% -15.65% -16.79% -17.97% -19.46% -21.00% -22.62% -24.56% -26.60% -28.91%
T-Test -0.767152 -0.983125 -1.182422 -1.315176 -1.442828 -1.588649 -1.692886 -1.719747 -1.70662 -1.719467 -1.793241 -1.933233 -2.100756 -2.262046 -2.420731 -2.506962 -2.596896 -2.73469 -2.898752 -3.117193 -3.393946 -3.684788 -3.97345 -4.294989 -4.686244

Adjusted Average 3.1880 2.2239 1.7931 1.5362 1.3680 1.2461 1.1605 1.0985 1.0478 1.0086 0.9742 0.9423 0.9095 0.8805 0.8557 0.8335 0.8127 0.7912 0.7711 0.7492 0.7302 0.7127 0.6927 0.6727 0.6518
Difference -0.0147 -0.0225 -0.0254 -0.0279 -0.0259 -0.0213 -0.0159 -0.0086 -0.0017 0.0063 0.0113 0.0130 0.0122 0.0121 0.0129 0.0161 0.0204 0.0225 0.0246 0.0243 0.0240 0.0236 0.0207 0.0173 0.0121
% Difference -0.46% -1.00% -1.39% -1.78% -1.86% -1.68% -1.35% -0.78% -0.16% 0.63% 1.18% 1.40% 1.36% 1.39% 1.54% 1.97% 2.57% 2.93% 3.29% 3.35% 3.40% 3.42% 3.08% 2.65% 1.89%
T-Test -0.279366 -0.375981 -0.428662 -0.460837 -0.443127 -0.392484 -0.306685 -0.169268 -0.032615 0.1237708 0.2270882 0.2636276 0.2496557 0.2476498 0.2690242 0.3346451 0.4267645 0.4766195 0.5308873 0.5366933 0.5497178 0.5571431 0.4984817 0.4274387 0.3059042

adjustment 1.0081125 1.0166059 1.0255071 1.0348452 1.0446521 1.0549628 1.0658157 1.077253 1.0893212 1.1020714 1.1155602 1.1298502 1.1450107 1.1611183 1.1782583 1.1965253 1.2160241 1.2368713 1.259196 1.2831409 1.3088632 1.3365345 1.3663412 1.3984818 1.4331628

Table A1, Comparions of True Standard Deviation of Volatility to Sample Standard Deviations of Volatility Using Overlapping Data (For I.I.D. Case)
 Stochastic Volatility case
True Standard Deviation of Volatility

testall 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 360 380 400 420 440 460 480 500
svall Stdev 9.0617 8.3317 7.8305 7.4218 7.0656 6.7546 6.4829 6.2433 6.0294 5.8372 5.6625 5.5036 5.3559 5.2168 5.0863 4.9625 4.8445 4.7325 4.6276 4.5303 4.4397 4.3542 4.2716 4.1910 4.1115

sv1 Stdev 8.9932 8.4017 7.8649 7.2808 6.6696 6.0961 5.4913 4.9042 4.3561 3.8605 3.4424 3.0856 2.7686 2.4872 2.2383 2.0186 1.8405 1.7082 1.6543 1.6738 1.7393 1.8079 1.8233 1.7916 1.7179
sv2 Stdev 7.3091 6.4180 5.6864 5.1415 4.7134 4.3626 4.0287 3.7395 3.5359 3.4019 3.3067 3.2424 3.1878 3.1167 3.0359 2.9489 2.8462 2.7152 2.5477 2.3625 2.1981 2.0813 1.9863 1.9125 1.8448
sv3 Stdev 11.0655 10.5739 10.2676 9.9939 9.7139 9.4325 9.1759 8.9566 8.7582 8.5836 8.4365 8.2882 8.1273 7.9562 7.7695 7.5733 7.3744 7.1720 6.9703 6.7657 6.5602 6.3281 6.0505 5.7273 5.3648
sv4 Stdev 11.1675 10.3929 9.7743 9.2126 8.6980 8.2382 7.8764 7.5718 7.3103 7.0654 6.8013 6.5040 6.1735 5.8401 5.5165 5.1695 4.7958 4.4154 4.0728 3.7954 3.5713 3.4249 3.3243 3.2657 3.2250
sv5 Stdev 9.4406 8.4825 7.8934 7.4369 7.0655 6.7002 6.2958 5.8960 5.5362 5.1843 4.8192 4.4869 4.1980 3.9499 3.7467 3.5724 3.4021 3.2187 3.0290 2.8265 2.6182 2.4131 2.2559 2.1382 2.0306
sv6 Stdev 10.2413 9.3608 8.7712 8.2336 7.6896 7.1823 6.7657 6.4321 6.1205 5.8043 5.4353 5.0503 4.6898 4.3719 4.1023 3.8661 3.6404 3.4224 3.1972 2.9965 2.8327 2.7085 2.6027 2.5163 2.4549
sv7 Stdev 5.2758 4.5186 4.0570 3.7607 3.4882 3.2422 2.9954 2.7195 2.4572 2.2383 2.0336 1.8674 1.7407 1.6285 1.5299 1.4343 1.3347 1.2351 1.1543 1.0837 1.0081 0.9357 0.8696 0.8236 0.8098
sv8 Stdev 8.3046 7.5677 7.0241 6.5633 6.0291 5.4827 4.9544 4.4995 4.1427 3.8839 3.6960 3.5802 3.5157 3.4609 3.4019 3.3049 3.1889 3.0590 2.9220 2.7826 2.6509 2.5304 2.4175 2.3130 2.2063
sv9 Stdev 6.7098 5.7433 5.1518 4.6119 4.0772 3.5985 3.1567 2.8009 2.5719 2.4555 2.3962 2.3791 2.3299 2.2708 2.1898 2.1103 2.0428 1.9765 1.9043 1.8409 1.7925 1.7489 1.6958 1.6177 1.5431
sv10 Stdev 7.9687 7.2147 6.7624 6.4224 6.1274 5.8617 5.6346 5.4156 5.1953 4.9720 4.7319 4.4720 4.2065 3.9422 3.6779 3.4196 3.1682 2.9489 2.7581 2.5875 2.4404 2.3192 2.2113 2.1226 2.0471
sv11 Stdev 10.9757 10.2919 9.7342 9.3067 8.9759 8.6434 8.2802 7.9409 7.6658 7.4089 7.1360 6.8638 6.6048 6.3476 6.0845 5.8273 5.5831 5.3482 5.1228 4.9082 4.7053 4.5097 4.3152 4.1123 3.9068
sv12 Stdev 9.3753 8.0893 7.2029 6.5474 5.9585 5.4238 4.9848 4.6678 4.4356 4.2691 4.1539 4.0495 3.9377 3.8194 3.7047 3.5976 3.4936 3.3884 3.2844 3.1891 3.1073 3.0232 2.9240 2.8117 2.6852
sv13 Stdev 9.0310 8.3803 7.9987 7.6950 7.3381 6.9328 6.4739 6.0249 5.5807 5.1718 4.8171 4.4743 4.1220 3.7812 3.4719 3.1958 2.9481 2.7330 2.5580 2.4295 2.3456 2.2976 2.2550 2.1914 2.1094
sv14 Stdev 6.7707 5.9187 5.3172 4.8861 4.5456 4.2822 4.0887 3.9585 3.8504 3.7320 3.5990 3.4561 3.3243 3.2226 3.1283 3.0296 2.9100 2.7664 2.6077 2.4503 2.3179 2.2292 2.1534 2.0887 2.0303
sv15 Stdev 9.2160 8.3723 7.8253 7.3939 6.9804 6.6446 6.3700 6.1249 5.9090 5.7033 5.5080 5.3071 5.1146 4.9579 4.8230 4.7044 4.5929 4.4738 4.3435 4.2104 4.0783 3.9470 3.8125 3.6713 3.5136
sv16 Stdev 8.8273 7.8777 7.4118 6.9747 6.5403 6.1134 5.7345 5.3752 5.0157 4.6845 4.3575 4.0588 3.7734 3.5083 3.2671 3.0377 2.7915 2.5349 2.3131 2.1341 2.0289 1.9367 1.8591 1.7938 1.7364
sv17 Stdev 8.8093 8.1550 7.6133 7.1546 6.7969 6.4585 6.1234 5.8113 5.5189 5.2598 5.0326 4.8207 4.6065 4.3952 4.2072 4.0512 3.9253 3.8042 3.6802 3.5475 3.3959 3.2271 3.0484 2.8653 2.6778
sv18 Stdev 8.3111 7.8561 7.5647 7.3243 7.1269 6.9430 6.7869 6.6513 6.5339 6.4288 6.3096 6.1890 6.0725 5.9608 5.8371 5.7243 5.6241 5.5127 5.3993 5.2810 5.1573 5.0400 4.9011 4.7527 4.5969
sv19 Stdev 6.2774 5.3623 4.8138 4.4090 4.0908 3.8742 3.7290 3.6257 3.5120 3.3833 3.2569 3.1355 3.0305 2.9565 2.9053 2.8478 2.7764 2.6948 2.6072 2.5127 2.4249 2.3454 2.2729 2.2032 2.1389
sv20 Stdev 9.1272 8.6483 8.3365 8.1115 7.9173 7.7782 7.6593 7.5410 7.4208 7.3110 7.2174 7.1342 7.0540 6.9680 6.8639 6.7309 6.5734 6.4064 6.2297 6.0328 5.8172 5.5929 5.3708 5.1560 4.9515

Unadjusted Average 8.6599 7.8813 7.3536 6.9230 6.5271 6.1646 5.8303 5.5329 5.2714 5.0401 4.8244 4.6223 4.4289 4.2471 4.0751 3.9082 3.7426 3.5767 3.4178 3.2705 3.1395 3.0223 2.9075 2.7937 2.6796
Difference -0.4018 -0.4504 -0.4770 -0.4988 -0.5385 -0.5900 -0.6526 -0.7104 -0.7581 -0.7971 -0.8382 -0.8813 -0.9270 -0.9697 -1.0112 -1.0543 -1.1018 -1.1558 -1.2099 -1.2598 -1.3002 -1.3319 -1.3641 -1.3973 -1.4319
% Difference -4.43% -5.41% -6.09% -6.72% -7.62% -8.74% -10.07% -11.38% -12.57% -13.66% -14.80% -16.01% -17.31% -18.59% -19.88% -21.24% -22.74% -24.42% -26.14% -27.81% -29.29% -30.59% -31.93% -33.34% -34.83%
T-Test -1.09141 -1.18508 -1.23469 -1.28473 -1.38121 -1.51388 -1.67118 -1.81124 -1.92768 -2.03133 -2.1489 -2.28483 -2.43266 -2.57523 -2.72121 -2.87793 -3.05071 -3.24711 -3.45886 -3.68308 -3.9121 -4.1484 -4.42558 -4.75263 -5.14337
Adjusted
Average 9.0050 8.3050 7.8284 7.4412 7.0837 6.7572 6.4571 6.1946 5.9694 5.7764 5.5992 5.4364 5.2827 5.1411 5.0104 4.8848 4.7596 4.6328 4.5134 4.4083 4.3242 4.2608 4.2026 4.1501 4.1058
Difference -0.0567 -0.0268 -0.0022 0.0194 0.0180 0.0026 -0.0257 -0.0487 -0.0600 -0.0609 -0.0634 -0.0672 -0.0733 -0.0757 -0.0759 -0.0777 -0.0849 -0.0997 -0.1142 -0.1220 -0.1155 -0.0935 -0.0690 -0.0409 -0.0057
% Difference -0.63% -0.32% -0.03% 0.26% 0.26% 0.04% -0.40% -0.78% -1.00% -1.04% -1.12% -1.22% -1.37% -1.45% -1.49% -1.57% -1.75% -2.11% -2.47% -2.69% -2.60% -2.15% -1.62% -0.98% -0.14%
T-Test -0.154 -0.07039 -0.00566 0.049923 0.046247 0.006584 -0.06587 -0.12409 -0.15257 -0.15508 -0.16245 -0.17412 -0.19227 -0.20111 -0.20428 -0.21208 -0.23496 -0.2802 -0.32654 -0.35671 -0.34752 -0.29113 -0.2239 -0.13912 -0.02064

adjustment 1.039856 1.053755 1.064566 1.074849 1.085265 1.09613 1.10752 1.119598 1.132431 1.14608 1.160603 1.176138 1.19277 1.210496 1.229512 1.24988 1.271731 1.295261 1.320568 1.347887 1.377353 1.409752 1.445441 1.485497 1.532253

Table A2, Comparions of True Standard Deviation of Volatility to Sample Standard Deviations of Volatility Using Overlapping Data (For Stochastic Volatility Case)
 Appendix B

The T × T matrix W is summed along each of its d = j − i = 1,..., T − 1 diagonal

bands, and these sums are then summed as geometric progressions in increasing
powers of a.

The summations were checked with the aid of a computer algebra program, and then
output as code for numerical work. The key results are summarized here.

The sums below have all been divided by a factor of n.

d=0
This is the sum of the diagonal (i = j) entries:

 h h 2 − 1
hn 1 − +
2
.
 n 3n 2 

d=1
This and the following ones occur above and below the diagonal and so must be
multiplied by two. We require a times the value of the following sum:

h
n 2 ( h − 1) − h 2 n + ( h 2 + 2) .
3

d = 2,.., h − 1, gives a times the following sums over i = 0,.., I = h − 3:

2

1 I 3 i I I
− ∑
3 i =0
i a + ( h − 2 ) ∑
i =0
i 2 i
a + ( 4 h − n 2
− 11 / 3) ∑
i =0
i ai
I
+ ( n 2 ( h − 2) − h 2 n + ( h 3 + 11h − 6) / 3) ∑ a i .
i =0

d = h,.., n − h + 1, gives a times the following sums over i = 0,.., I = n− 2h+1:

h

I I
h 2 ∑ i a i − h 2 ( n − h) ∑ a i .
i =0 i =0

d = n − h + 2,.., n, gives a − times the following sums over i = 0,.., I = h−

n h+2

2:

1 I 3 i I I I
− ∑
6 i =0
i a − ∑
i =0
i 2 i
a + ( h 2
− 11 / 6 ) ∑
i =0
i a i
+ ( − h 3
+ 2 h 2
− 1) ∑
i =0
ai .

d = n + 1,.., n+h−2, gives a times the following sums over i = 0,.., I = h− 3:

n+1

1 I 3 i h − 1 I 2 i 3h(h − 2) + 2 I h( h − 1)(h − 2) I i
∑
6 i =0
i a − ∑
2 i =0
i a +
6
∑i =0
i ai −
6
∑i =0
a .

22

View publication stats