This article was downloaded by: [Princeton University]

On: 19 September 2012, At: 05:41

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer
House, 37-41 Mortimer Street, London W1T 3JH, UK
Journal of the American Statistical Association
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/uasa20
Vast Volatility Matrix Estimation Using High-
Frequency Data for Portfolio Selection
Jianqing Fan
a
, Yingying Li
b
& Ke Yu
a
a
Department of Operations Research and Financial Engineering, Princeton University,
Princeton, NJ, 08544, USA
b
Department of Information Systems, Business Statistics and Operations Management,
Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong,
China
Accepted author version posted online: 31 Jan 2012.Version of record first published: 11
Jun 2012.
To cite this article: Jianqing Fan, Yingying Li & Ke Yu (2012): Vast Volatility Matrix Estimation Using High-Frequency Data
for Portfolio Selection, Journal of the American Statistical Association, 107:497, 412-428
To link to this article: http://dx.doi.org/10.1080/01621459.2012.656041
PLEASE SCROLL DOWN FOR ARTICLE
Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to
anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses
should be independently verified with primary sources. The publisher shall not be liable for any loss, actions,
claims, proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or
indirectly in connection with or arising out of the use of this material.
Vast Volatility Matrix Estimation Using
High-Frequency Data for Portfolio Selection
Jianqing FAN, Yingying LI, and Ke YU
Portfolio allocation with gross-exposure constraint is an effective method to increase the efciency and stability of portfolios selection among
a vast pool of assets, as demonstrated by Fan, Zhang, and Yu. The required high-dimensional volatility matrix can be estimated by using
high-frequency nancial data. This enables us to better adapt to the local volatilities and local correlations among a vast number of assets
and to increase signicantly the sample size for estimating the volatility matrix. This article studies the volatility matrix estimation using
high-dimensional, high-frequency data from the perspective of portfolio selection. Specically, we propose the use of pairwise-refresh
time and all-refresh time methods based on the concept of refresh time proposed by Barndorff-Nielsen, Hansen, Lunde, and Shephard
for the estimation of vast covariance matrix and compare their merits in the portfolio selection. We establish the concentration inequalities of
the estimates, which guarantee desirable properties of the estimated volatility matrix in vast asset allocation with gross-exposure constraints.
Extensive numerical studies are made via carefully designed simulations. Comparing with the methods based on low-frequency daily data,
our methods can capture the most recent trend of the time varying volatility and correlation, hence provide more accurate guidance for
the portfolio allocation in the next time period. The advantage of using high-frequency data is signicant in our simulation and empirical
studies, which consist of 50 simulated assets and 30 constituent stocks of Dow Jones Industrial Average index.
KEY WORDS: Concentration inequalities; High-frequency data; Portfolio allocation; Refresh time; Risk assessment; Volatility matrix
estimation.
1. INTRODUCTION
The mean-variance efcient portfolio theory by Markowitz
(1952, 1959) has had a profound impact on modern nance.
Yet, its applications to practical portfolio selection face a num-
ber of challenges. It is well known that the selected portfolios
depend too sensitively on the expected future returns and volatil-
ity matrix. This leads to the puzzle postulated by Jagannathan
and Ma (2003) why no short-sale portfolio outperforms the ef-
cient Markowicz portfolio. The sensitivity on the dependence
can be effectively addressed by the introduction of the constraint
on the gross exposure of portfolios (Fan, Zhang, and Yu 2011).
Their results not only give a theoretical answer to the puzzle
postulated by Jagannathan and Ma (2003) but also pave a way
for optimal portfolio selection in practice.
The second challenge of the implementation of Markowitzs
portfolio selection theory is the intrinsic difculty of the esti-
mation of the large volatility matrix. This is well documented
in the statistics and econometrics literature even for the static
large covariance matrix (Bickel and Levina 2008; Fan, Fan, and
Lv 2008; Lam and Fan 2009; Rothman, Levina, and Zhu 2009).
The additional challenge comes from the time-varying nature
of a large volatility matrix. For a short and medium holding
period (1 day or 1 week, say), the expected volatility matrix in
Jianqing Fan (E-mail: [email protected]) is Frederick L. Moore 18 Pro-
fessor of Finance, Department of Operations Research and Financial Engineer-
ing, Princeton University, Princeton, NJ 08544, USA. Yingying Li (E-mail:
[email protected]) is Assistant Professor, Department of Information Systems, Busi-
ness Statistics and Operations Management, Hong Kong University of Science
and Technology, Clear Water Bay, Kowloon, Hong Kong, China. Ke Yu (E-mail:
[email protected]) is Graduate Student, Department of Operations Research
and Financial Engineering, Princeton University, Princeton, NJ 08544, USA.
Fans research was supported by NSF (National Science Foundation) grant
DMS-0704337, NIH (National Institutes of Health) grant R01-GM072611 and
NIH grant R01GM100474. The main part of the work was carried while Ying-
ying Li was a postdoctoral fellow at the Department of Operations Research
and Financial Engineering, Princeton University. Lis research was further par-
tially supported by GRF (General Research Fund) 606811 of Hong Kong. The
authors thank the editor, the associate editor, and two referees for their helpful
comments.
the near future can be very different from the average of the ex-
pected volatility matrix over a long time horizon (the past 1 year,
say). As a result, even if we know exactly the realized volatility
matrix in the past, the bias can still be large. This calls for a
stable and robust portfolio selection. The portfolio allocation
under the gross-exposure constraint provides a needed solution.
To reduce the bias of the forecasted expected volatility matrix,
we need to shorten the learning period to better capture the dy-
namics of the time-varying volatility matrix, adapting better to
the local volatility and correlation. But this is at the expense of
a reduced sample size. The wide availability of high-frequency
data provides sufcient amount of data for reliable estimation
of the volatility matrix.
Recent years have seen dramatic developments in the study of
high-frequency data in estimating integrated volatility. Statisti-
cians and econometricians have been focusing on the interesting
and challenging problem of volatility estimation in the pres-
ence of market microstructure noise and asynchronous tradings,
which are the stylized features of high-frequency nancial data.
The progress is very impressive with a large literature. In partic-
ular, in the one-dimensional case when the focus is on estima-
tion of integrated volatility, At-Sahalia, Mykland, and Zhang
(2005) discussed a subsampling scheme; Zhang, Mykland, and
At-Sahalia (2005) proposed a two-scale estimate which was
extended and improved by Zhang (2006) to multiple scales; Fan
and Wang (2007) separated jumps from diffusions in presence
of market microstructural noise using a wavelet method; the ro-
bustness issues were addressed by Li and Mykland (2007); the
realized kernel methods were proposed and thoroughly studied
by Barndorff-Nielsen et al. (2009, 2011); Jacod et al. (2009)
proposed a preaveraging approach to reduce the market mi-
crostructral noise; Xiu (2010) demonstrated that a simple quasi-
likelihood method achieves the optimal rate of convergence for
2012 American Statistical Association
Journal of the American Statistical Association
March 2012, Vol. 107, No. 497, Theory and Methods
DOI: 10.1080/01621459.2012.656041
412
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

Fan, Li, and Yu: Vast Volatility Matrix Estimation Using High-Frequency Data 413
estimating integrated volatility. For estimation of integrated co-
variation, the nonsynchronized trading issue was rst addressed
by Hayashi and Yoshida (2005) in the absence of the microstruc-
tural noise; the kernel method with refresh time idea was rst
proposed by Barndorff-Nielsen, Hansen, Lunde, and Shephard
(2008); Zhang (2011) extended the two-scale method to study
the integrated covariation using a previous tick method; At-
Sahalia, Fan, and Xiu (2010) extended the quasi-maximumlike-
lihood method; Kinnebrock, Podolskij, and Christensen (2010)
extended the preaveraging technique. For the high-dimensional
case, Wang and Zou (2010) estimated volatility matrix when
sparsity condition is satised; Hautsch, Kyj, and Oomen (2009)
studied a blocking and regularization approach; Tao et al. (2011)
aggregated daily integrated volatility matrix via a factor model;
Zheng and Li (2011) studied the empirical spectral distribution
of the volatility matrix.
The aim of this article is to study the volatility matrix esti-
mation using high-dimensional, high-frequency data from the
perspective of nancial engineering. Specically, our main top-
ics are how to extract the covariation information from high-
frequency data for asset allocation and how effective they
are. Two particular strategies are proposed for handling the
nonsynchronized trading: pairwise-refresh and all-refresh
schemes. The former retains many more data points and esti-
mates covariance matrix element-by-element, which is often not
positive semidenite, whereas the latter retains far fewer data
points and the resulting covariance matrix is more often positive
semidenite. As a result, the former has a better elementwise
estimation error and is better in controlling risk approxima-
tion mentioned in the rst paragraph of the introduction. How-
ever, the merits between the two methods are not that simple.
In implementation, we need to project the estimate of covari-
ance matrices onto the space of the semipositive denite ma-
trices. The projections distort the accuracy of the elementwise
estimation. As a result, the pairwise-refresh scheme does not
have much more advantage than the all-refresh method, though
the former is very easy to implement. However, both methods
signicantly outperform the methods based on low-frequency
data, since they adapt better to the time-varying volatilities
and correlations. The comparative advantage is more dramatic
when there are rapid changes of the volatility matrix over time.
This will be demonstrated in both simulation and empirical
studies.
As mentioned before and demonstrated in Section 2, the accu-
racy of portfolio risk relative to the theoretically optimal portfo-
lio is governed by the maximum elementwise estimation error.
How does this error grow with the number of assets? Thanks
to the concentration inequalities derived in this article, it grows
only at the logarithmic order of the number of assets. This gives
a theoretical endorsement why the portfolio selection problem
is feasible for vast portfolios.
The article is organized as follows. Section 2 gives an
overview of portfolio allocation using high-frequency data.
Section 3 studies the volatility matrix estimation using high-
frequency data from the perspective of asset allocation, where
the analytical results are presented. How well our idea works
in simulation and empirical studies can be found in Sec-
tions 4 and 5, respectively. Conclusions are given in Sec-
tion 6. Technical conditions and proofs are relegated to the
Appendix.
2. CONSTRAINED PORTFOLIO OPTIMIZATION WITH
HIGH-FREQUENCY DATA
2.1 Problem Setup
Consider a pool of p assets, with log-price processes X
(1)
t
,
X
(2)
t
, . . . , X
(p)
t
. Denote by X
s
= (X
(1)
s
, . . . , X
(p)
s
)
T
the vector of
the log-price processes at time s. Suppose they followa diffusion
process, namely,
dX
t
=
t
dt +S
1/2
t
dW
t
(1)
where W
t
is the vector of p-dimensional standard Brownian
motions. The drift vector
t
and the instantaneous variance
S
t
can be stochastic processes and are assumed to be continuous.
A given portfolio with the allocation vector w at time t and a
holding period has the log-return w
T
_
t +
t
dX
s
with variance
(risk)
R
t,
(w) = w
T

t,
w, (2)
where w
T
1 = 1 and

t,
=
_
t +
t
E
t
S
u
du, (3)
with E
t
denoting the conditional expectation given the history up
to time t. Let w
+
be the proportion of long positions and w

be
the proportion of the short positions. Then, w
1
= w
+
+w

is
the gross exposure of the portfolio. To simplify the problem,
following Jagannathan and Ma (2003) and Fan et al. (2011),
we consider only the risk optimization problem. In practice, the
expected return constraint can be replaced by the constraints
of sectors or industries, to avoid unreliable estimates of the
expected return vector. For a short-time horizon, the expected
return is usually negligible. Following Fan et al. (2011), we
consider the following risk optimization under gross-exposure
constraints:
min w
T

t,
w, s.t.w
1
c and w
T
1 = 1, (4)
where c is the total exposure allowed. Note that using w
+

= 1, problem (4) puts equivalently the constraint on the

proportion of the short positions: w

(c 1)/2. As noted by
Jagannathan and Ma (2003), the constrained optimization prob-
lem (4) is equivalent to unconstrained risk optimization with
a regularized covariance matrix. Other methods of regulariza-
tion are also possible to handle the noise-accumulation problem
(e.g., the shrinkage method of Ledoit and Wolf (2004)).
The problem (4) involves the conditional expected volatility
matrix (3) in the future. Unless we know exactly the dynamic
of the volatility process, which is usually unknown, even if we
observed the entire continuous paths up to the current time t.
As a result, we rely on the approximation even with ideal data
that we were able to observe the processes continuously without
error. The typical approximation is

t,
h
1
_
t
t h
S
u
du (5)
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

414 Journal of the American Statistical Association, March 2012
for an appropriate window width h and we estimate
_
t
t h
S
u
du based on the historical data at the time interval
[t h, t ].
Approximation (5) holds reasonably well when and h are
both small. This relies on the continuity assumptions: local time-
varying volatility matrices are continuous in time. The approx-
imation is also reasonable when both and h are large. This
relies on the ergodicity assumption so that both quantities will
be approximately ES
u
, when the stochastic volatility matrix
S
u
is ergodic. The approximation is not good when is small,
whereas h is large as long as S
u
is time varying, whether the
stochastic volatility S
u
is stationary or not. In other words, when
the holding time horizon is short, as long as S
u
is time varying,
we can only use a short-time window [t h, t ] to estimate
t,
.
The recent arrivals of high-frequency data make this problem
feasible.
The approximation error in (5) cannot usually be evaluated
unless we have a specic parametric model on the stochastic
volatility matrix S
u
. However, this is at the risk of model mis-
specications and nonparametric approach is usually preferred
for high-frequency data. With p
2
elements approximated, which
can be in the order of hundreds of thousands or millions, a natural
question to ask is whether these errors accumulate and whether
the result (risk) is stable. The gross-exposure constraint gives a
stable solution to the problem as shown by Fan et al. (2011).
We would like to close this section by noting that formulation
(4) is a one-period, not a multi-period, portfolio optimization
problem.
2.2 Risk Approximations With Gross-Exposure
Constraints
The utility of gross-exposure constraint can easily be seen
through the following inequality. Let

t,
be an estimated co-
variance matrix and

R
t,
(w) = w
T

t,
w (6)
be estimated risk of the portfolio. Then, for any portfolio with
gross exposure w
1
c, we have
|

R
t,
(w) R
t,
(w)|
p

i=1
p

j=1
|
i,j

i,j
||w
i
||w
j
|
|
t,

t,
|

w
2
1
|
t,

t,
|

c
2
, (7)
where
i,j
and
i,j
are respectively the (i, j) elements of

t,
and

t,
, and
|
t,

t,
|

= max
i,j
|
i,j

i,j
|
is the maximum elementwise estimation error. Risk approxima-
tion (7) reveals that there is no large error accumulation effect
when gross exposure c is moderate.
From now on, we drop the dependence of t and whenever
there is no confusion. This facilitates the notation.
Fan et al. (2011) showed further that the risks of optimal
portfolios are indeed close. Let
w
opt
= argmin
w
T
1=1, ||w||
1
c
R(w),
w
opt
= argmin
w
T
1=1, ||w||
1
c

R(w) (8)
be, respectively, the theoretical (oracle) optimal allocation vec-
tor we want and the estimated optimal allocation vector we get.
Then, R(w
opt
) is the theoretical minimum risk and R( w
opt
) is
the actual risk of our selected portfolio, whereas

R( w
opt
) is our
perceived risk, which is the quantity known to nancial econo-
metricians. They showed that
|R( w
opt
) R(w
opt
)| 2a
p
c
2
, (9)
|R( w
opt
)

R( w
opt
)| a
p
c
2
, (10)
|R(w
opt
)

R( w
opt
)| a
p
c
2
, (11)
with a
p
= |

, which usually grows slowly with the

number of assets p. These reveal that the three relevant risks
are in fact close as long as the gross-exposure parameter c is
moderate and the maximum elementwise estimation error a
p
is
small.
The above risk approximations hold for any estimate of
covariance matrix. It does not even require

a semipositive def-
inite matrix. This facilitates signicantly the method of covari-
ance matrix estimation. In particular, the elementwise estimation
methods are allowed. In fact, since the approximation errors in
(9), (10), and (11) are all controlled by the maximum element-
wise estimation error, it can be advantageous to use elementwise
estimation methods. This is particularly the case for the high-
frequency data where trading is nonsynchronized. The synchro-
nization can be done pairwisely or for all assets. The former
retains much more data than the latter, as shown in the next
section.
3. COVARIANCE MATRIX ESTIMATION USING
HIGH-FREQUENCY DATA
3.1 All-Refresh Method and Pairwise-Refresh Method
Estimating high-dimensional volatility matrix using high-
frequency data is a challenging task. One of the challenges is the
nonsynchronicity of trading. Several synchronization schemes
have been studied. The refresh time method was proposed by
Barndorff-Nielsen et al. (2008) and the previous tick method
was used by Zhang (2011). The former uses more efciently the
available data and will be used in this article.
The idea of refresh time is to wait until all assets are traded
at least once at time v
1
(say) and then use the last price traded
before or at v
1
of each asset as its price at time v
1
. This yields
one synchronized price vector at time v
1
. The clock now starts
again. Wait until all assets are traded at least once at time v
2
(say)
and again use the previous tick price of each asset as its price
at time v
2
. This yields the second synchronized price vector at
time v
2
. Repeat the process until all available trading data are
synchronized. Clearly, the process discards a large portion of
the available trades. We will refer this synchorization scheme as
the all-refresh time. (The method is called all-refresh method
for short.) Barndorff-Nielsen et al. (2008) advocated the kernel
method to estimate integrated volatility matrix after synchro-
nization; this can also be done by using other methods.
A more efcient method to use the available sample is the
pairwise-refresh time scheme, which synchronizes the trading
for each pair of assets separately. (The method is called pairwise-
refresh method for short.) The pairwise-refresh scheme makes
far more efcient use of the rich information in high-frequency
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

Fan, Li, and Yu: Vast Volatility Matrix Estimation Using High-Frequency Data 415
data, and enables us to estimate each element in the volatility
matrix more precisely, which helps improve the efciency of
the selected portfolio. We will study the merits of these two
methods. A third synchronization scheme is the blocking
scheme of Hautsch et al. (2009) by grouping stocks with similar
liquidities. The pairwise-refresh approach corresponds to the
case with one-stock per group.
The pairwise estimation method allows us to use a wealth
of univariate integrated volatility estimators such as those men-
tioned in the Introduction. For any given two assets with log-
price processes X
(i)
t
and X
(j)
t
, with pairwise-refresh times, the
synchronized prices of X
(i)
t
+X
(j)
t
and X
(i)
t
X
(j)
t
can be com-
puted. With the univariate estimate of the integrated volatilities

< X
(i)
+X
(j)
> and

< X
(i)
X
(j)
>, the integrated covaria-
tion can be estimated as

i,j
=

_
X
(i)
, X
(j)
_
=
_

_
X
(i)
+X
(j)
_

_
X
(i)
X
(j)
__
/4. (12)
In particular, the diagonal elements are estimated by the method
itself. When the two-scale realized volatility (TSRV) (Zhang et
al. 2005) is used, this results in the two-scale realized covariance
(TSCV) estimate (Zhang 2011).
3.2 Pairwise-Refresh Method and TSCV
We nowfocus on the pairwise estimation method. To facilitate
the notation, we reintroduce it.
We consider two log-price processes X and Y that satisfy
dX
t
=
(X)
t
dt +
(X)
t
dB
(X)
t
and dY
t
=
(Y)
t
dt +
(Y)
t
dB
(Y)
t
,
(13)
where cor(B
(X)
t
, B
(Y)
t
) =
(X,Y)
t
. For the two processes X and Y,
consider the problemof estimating X, Y
T
with T = 1. Denote
by T
n
the observation times of X and S
m
the observation times of
Y. Denote the elements in T
n
and S
m
by {
n,i
}
n
i=0
and {
m,i
}
m
i=0
,
respectively, in an ascending order (
n,0
and
m,0
are set to be
0). The actual log prices are not directly observable, but are
observed with microstructure noises:
X
o

n,i
= X

n,i
+
X
i
, and Y
o

m,i
= Y

m,i
+
Y
i
, (14)
where X
o
and Y
o
are the observed transaction prices in the
logarithmic scale, and X and Y are the latent log prices govern by
the stochastic dynamics (13). We assume that the microstructure
noise
X
i
and
Y
i
processes are independent of the X and Y
processes and that

X
i
i.i.d.
N
_
0,
2
X
_
and
Y
i
i.i.d.
N
_
0,
2
Y
_
. (15)
Note that this assumption is mainly for the simplicity of pre-
sentation; as we can see from the proof, one can, for example,
easily replace the identical Gaussian assumption with the not
necessarily identically distributed (but are of the same variance)
sub-Gaussian assumption without affecting our results.
The pairwise-refresh time V = {v
0
, v
1
, . . . , v
n
} can be ob-
tained by setting v
0
= 0, and
v
i
= max
_
min{ T
n
: > v
i1
}, min{ S
m
: > v
i1
}
_
,
where n is the total number of refresh times in the interval (0, 1].
The actual sample times for the two individual processes X and
Y that correspond to the refresh times are
t
i
= max{ T
n
: v
i
} and s
i
= max{ S
m
: v
i
},
which are indeed the previous-tick measurement.
We study the property of the TSCVbased on the asynchronous
data:

X, Y
1
= [X
o
, Y
o
]
(K)
1

n
K
n
J
_
X
o
, Y
o
_
(J)
1
, (16)
where
[X
o
, Y
o
]
(K)
1
=
1
K
n

i=K
_
X
o
t
i
X
o
t
iK
__
Y
o
s
i
Y
o
s
iK
_
and n
K
= ( n K +1)/K. As discussed by Zhang (2011), the
optimal choice of K has order K = O( n
2/3
), and J can be taken
to be a constant such as 1.
When either the microstructure error or the asynchronicity
exists, the realized covariance is seriously biased. An asymp-
totic normality result in Zhang (2011) reveals that TSCV can
simultaneously remove the bias due to the microstructure error
and the bias due to the asynchronicity. However, this result is
not adequate for our application to the vast volatility matrix esti-
mation. To understand its impact on a
p
, we need to establish the
concentration inequality. In particular, for a sufciently large
|x| = O((log p)

), if
max
i,j
P{

n|
ij

ij
| > x} < C
1
exp
_
C
2
x
1/
_
, (17)
for some positive constants C
1
, C
2
, and , then
a
p
= |

|

= O
P
_
(log p)

n
_
. (18)
We will show in the next section that the result indeed holds for
some which depends on the tail of the volatility process and
n replaced by the minimum of the subsample size ( n
K
( n)
1
3
).
Hence the impact of the number of assets is limited, only of the
logarithmic order.
3.3 Concentration Inequalities
Inequality (17) requires the conditions on both diagonal ele-
ments and off-diagonal elements. Technically, they are treated
differently. For the diagonal cases, the problem corresponds to
the estimation of integrated volatility and there is no issue of
asynchronicity. TSCV(16) reduces to TSRV(Zhang et al. 2005),
which is explicitly given by

X, X
1
=
_
X
o
, X
o
_
(K)
1

n
K
n
J
_
X
o
, X
o
_
(J)
1
, (19)
where [X
o
, X
o
]
(K)
1
=
1
K

n
i=K
(X
o
t
i
X
o
t
iK
)
2
and n
K
= (n
K +1)/K. As shown by Zhang et al. (2005), the optimal choice
of K has order K = O(n
2/3
) and J can be taken to be a constant
such as 1.
To facilitate the reading, we relegate the technical conditions
and proofs to the Appendix. The following two results establish
the concentration inequalities for the integrated volatility and
integrated covariation.
Theorem 1. Let X process be as in (13), and n be the total
number of observations for the X process during the time interval
(0,1]. Under Conditions 14 in the Appendix
(A) if
(X)
t
C

< for all t [0, 1], then for all large n,

for x [0, cn
1/6
],
P
_
n
1/6
|

X, X
1

_
1
0

(X)
t
2
dt | > x
_
4 exp{Cx
2
}
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

416 Journal of the American Statistical Association, March 2012
for positive constants c and C. A set of candidate values
for c and C are given in (A.25).
(B) if the tail behavior of
(X)
t
can be described as
P
_
sup
0t 1

(X)
t
C

_
k

exp
_
aC
b

_
, for any C

> 0
with positive constants k

, a and b, then for all large n,

for x [0, cn
4+b
6b
],
P
_
n
1/6
|

X, X
1

_
1
0

(X)
t
2
dt | > x
_
(4 +k

) exp{C x
2b
4+b
}.
Aset of candidate values for c and C are given in (A.26).
Theorem 2. Let X and Y be as in (13), and n be the total
number of refresh times for the processes X and Y during the
time interval (0,1]. Under Conditions 15 in the Appendix,
(A) if
(i)
t
C

< for all t [0, 1] and i = X and Y,

then for all large n, for x [0, c n
1/6
],
P
_
n
1/6
|

X, Y
1

_
1
0

(X)
t

(Y)
t

(X,Y)
t
dt | > x
_
8 exp{Cx
2
}
for positive constants c and C. A set of candidate values
for c and C are given in (A.29).
(B) if the tail behavior of
(i)
t
for i = X or Y satisfy
P
_
sup
0t 1

(i)
t
C

_
k

exp
_
aC
b

_
, for any C

> 0
with positive constants k

, a, and b, then for all large n,

for x [0, c n
4+b
6b
],
P
_
n
1/6
|

X, Y
1

_
1
0

(X)
t

(Y)
t

(X,Y)
t
dt | > x
_
(8 +2k

) exp
_
Cx
2b
4+b
_
.
Aset of candidate values for c and C are given in (A.30).
3.4 Error Rates on Risk Approximations
Having had the above concentration inequalities, we can now
readily give an upper bound of the risk approximations. Con-
sider the p log-price processes as in Section 2.1. Suppose the
processes are observed with the market microstructure noises.
Let n
(i,j)
be the observation frequency obtained by the pairwise-
refresh method for two processes X
(i)
and X
(j)
and n

be the ob-
servation frequency obtained by the all-refresh method. Clearly,
n
(i,j)
is typically much larger than n

. Hence, most elements are

estimated more accurately using the pairwise-refresh method
than using the all-refresh method. On the other hand, for less liq-
uidly traded pairs, its observation frequency of pairwise-refresh
time cannot be an order of magnitude larger than n

.
Using (18), an application to Theorems 1 and 2 to each ele-
ment in the estimated integrated covariance matrix yields
a
pairwiserefresh
p
= |

pairwise
|

= O
P
_
(log p)

n
1/6
min
_
, (20)
where is
1
2
when the volatility processes are bounded and
is a constant depending on the tail behavior of the volatility
processes when they are unbounded, and n
min
= min
i,j
n
(i,j)
is
the minimum number of observations of the pairwise-refresh
time.
Note that based on our proofs which do not rely on any
particular properties of pairwise-refresh times, our results of
Theorems 1 and 2 are applicable to all-refresh method as well,
with the observation frequency of the pairwise-refresh times
replaced by that of the all-refresh times. Hence, using the all-
refresh time scheme, we have
a
allrefresh
p
= |

allrefresh
|

= O
P
_
(log p)

n
1/6

_
, (21)
with the same as above. Clearly, n
min
is larger than n

. Hence,
the pairwise-refresh method gives a somewhat more accurate
estimate in terms of the maximumelementwise estimation error.
3.5 Projections of Estimated Volatility Matrices
The risk approximations (9)(11) hold for any solutions to (8)
whether the matrix

is positive semidenite or not. However,
convex optimization algorithms typically require the positive
semideniteness of the matrix

. Yet, the estimates based on
the elementwise estimation sometimes cannot all satisfy this
and even the ones from all-refresh method can have the same
problemwhen TSCVis applied. This leads to the issue of howto
project a symmetric matrix onto the space of positive semide-
nite matrices.
There are two intuitive methods for projecting a p p sym-
metric matrix A onto the space of positive semidenite
matrices. Consider the singular value decomposition: A =

T
diag(
1
, . . . ,
p
), where is an orthogonal matrix, con-
sisting of p eigenvectors. The two intuitive appealing projection
methods are
A
+
1
=
T
diag(
+
1
, . . . ,
+
n
), (22)
where
+
j
is the positive part of
j
, and
A
+
2
= (A +

min
I
p
)/(1 +

min
), (23)
where

min
is the negative part of the minimum eigenvalue. For
both projection methods, the eigenvectors remain the same as
those of A. When A is positive semidenite matrix, we have
obviously that A
1
= A
2
= A.
In applications, we apply the above transformations to the es-
timated correlation matrix A rather than directly to the volatility
matrix estimate

. The correlation matrix A has diagonal ele-
ments of 1. The resulting matrix under projection method (23)
apparently still satises this property, whereas the one under
projection method (22) does not. As a result, projection method
(23) keeps the integrated volatility of each asset intact.
In our initial simulation and empirical studies, we ap-
plied both projections. It turns out that there is no signicant
difference between the two projection methods in terms of
results. We decided to apply only projection (23) in all nu-
merical studies.
3.6 Comparisons Between Pairwise- and All-Refresh
Methods
The pairwise-refresh method keeps far richer information in
the high-frequency data than the all-refresh method. Thus, it is
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

Fan, Li, and Yu: Vast Volatility Matrix Estimation Using High-Frequency Data 417
expected to estimate each element more precisely. Yet, the esti-
mated correlation matrix is typically not positive semidenite.
As a result, projection (23) can distort the accuracy of element-
wise estimation. On the other hand, the all-refresh method more
often gives positive semidenite estimates. Therefore, projec-
tion (23) has less impact on the all-refresh method than on the
pairwise-refresh method.
Risk approximations (9)(11) are only the upper bounds. The
upper bounds are controlled by a
p
, which has rates of con-
vergence govern by (20) and (21). While the average number
of observations of pairwise-refresh time is far larger than the
number of observations n

of the all-refresh time, the minimum

number of observations of the pairwise-refresh time n
min
is not
much larger than n

. Therefore, the upper bounds (20) and (21)

are approximately of the same level. This together with the dis-
tortion due to projection do not leave much advantage for the
pairwise-refresh method.
4. SIMULATION STUDIES
In this section, we simulate the market trading data using a
reasonable stochastic model. As the latent prices and dynamics
of simulations are known, our study on the risk prole is fa-
cilitated. In particular, we would like to verify our theoretical
results and to quantify the nite sample behaviors.
In this section and the next, the risk refers to the stan-
dard deviation of portfolios returns. To avoid ambiguity, we
call
_
R(w
opt
) the theoretical optimal risk or oracle risk,
_

R( w
opt
) the perceived optimal risk, and
_
R( w
opt
) the actual
risk of the perceived optimal allocation.
4.1 Design of Simulations
A slightly modied version of the simulation model in
Barndorff-Nielsen et al. (2008) is used to generate the latent
price processes of p traded assets. It is a multivariate factor
model with stochastic volatilities. Specically, the latent log
prices X
(i)
t
follow
dX
(i)
t
=
(i)
dt +
(i)

(i)
t
dB
(i)
t
+
_
1 (
(i)
)
2

(i)
t
dW
t
+
(i)
dZ
t
,
i = 1, . . . , p, (24)
where the elements of B, W, and Z are independent standard
Brownian motions. The spot volatility obeys the independent
OrnsteinUhlenbeck processes:
d
(i)
t
=
(i)
_

(i)
0

(i)
t
_
dt +
(i)
1
dU
(i)
t
, (25)
where
(i)
t
= log
(i)
t
and U
(i)
t
is an independent Brownian
motion.
The number of assets p is taken to be 50. Slightly
modied from Barndorff-Nielsen et al. (2008), the
parameters are set to be (
(i)
,
(i)
0
,
(i)
1
,
(i)
,
(i)
) =
(0.03x
(i)
1
, x
(i)
2
, 0.75x
(i)
3
, 1/40x
(i)
4
, 0.7) and
(i)
=
exp(
(i)
0
), where x
(i)
j
is an independent realization from the
uniform distribution on [0.7, 1.3]. The parameters are kept
xed in the simulations.
Model (24) is used to generate the latent log-price values with
initial values X
(i)
0
= 1 (log price) and
(i)
0
from its stationary
distribution. The Euler scheme is used to generate latent price
at the frequency of once per second. To account for the market
microstructure noise, the Gaussian noises
(i)
t
i.i.d.
N(0,
2
) with
= 0.0005 are added. Therefore, like (14), the observed log
prices are X
o(i)
t
= X
(i)
t
+
(i)
t
.
To model the nonsynchronicity, p independent Poisson pro-
cesses with intensity parameters
1
,
2
, . . . ,
p
are used to sim-
ulate the trading times of the assets. Motivated by the U.S. equity
trading dataset (the total number of seconds in a common trad-
ing day of the U.S. equity is 23, 400), we set the trading inten-
sity parameters
i
s to be 0.02i 23400 for i = 1, 2, . . . , 50,
meaning that the average numbers of trading times for each asset
are spread out in the arithmetic sequence of the interval [468,
23,400].
4.2 An Oracle Investment Strategy and Risk
Assessment
An oracle investment strategy is usually a decent benchmark
for other portfolio strategies to compare with. There are several
oracle strategies. The one we choose is to make portfolio al-
location based on the covariance matrix estimated using latent
prices at the nest grid (one per second). Latent prices are the
noise-free prices of each asset at every time points (one per
second), which are unobservable in practice and are available
to us only in the simulation. Therefore, for each asset, there are
23, 400 latent prices in a normal trading day. We will refer to
the investment strategy based on the latent prices as the oracle
or latent strategy. This strategy is not available for the empirical
studies.
The assessment of risk is based on the high-frequency data.
For a given portfolio strategy, its risk is computed based on
the latent prices at every 15 minutes for the simulation studies,
whereas for the empirical studies, the observed prices at every 15
minutes are used to assess its risk. This mitigates the inuence
of the microstructure noises. For the empirical study, we do not
hold positions overnight therefore are immune to the overnight
price jumps (we will discuss the details in Section 5).
4.3 Out-of-Sample Optimal Allocation
One of the main purposes of this article is to investigate
the comparative advantage of the high-frequency-based meth-
ods against the low-frequency-based methods, especially in the
context of portfolio investment. Hence, it is essential for us to
run the following out-of-sample investment strategy test which
includes both the high-frequency- and low-frequency-based ap-
proaches. Moreover, since in the empirical studies, we do not
know the latent asset prices, the out-of-sample test should be
designed so that it can also be conducted in the empirical studies.
We simulate the prices of 50 traded assets as described in
Section 4.1 for the duration of 200 trading days (numbered as
day 1, day 2, . . . , day 200) and record all the tick-by-tick trading
times and trading prices of the assets.
We start investing 1 unit of capital into the pool of assets with
low-frequency- and high-frequency-based strategies from day
101 (the portfolios are bought at the opening of day 101). For the
low-frequency strategy, we use the previous 100 trading days
daily closing prices to compute the sample covariance matrix
and to make the portfolio allocation accordingly with the gross-
exposure constraints. For the high-frequency strategies, we use
the previous h = 10 trading days tick-by-tick trading data. For
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

418 Journal of the American Statistical Association, March 2012
the all-refresh strategy, we use all-refresh time to synchronize
the trades of the assets before applying TSCV to estimate the
integrated volatility matrix and make the portfolio allocation,
while for the pairwise-refresh high-frequency strategy, we use
pairwise-refresh times to synchronize each pair of assets and
apply TSCV to estimate the integrated covariance for the cor-
responding pair. With projection technique (23), the resulting
TSCV integrated volatility matrix can always be transformed to
a positive semidenite matrix which facilitates the optimization.
We run two investment plans. In the rst plan, the portfolio
is held for = 1 trading day before we reestimate the covaria-
tion structure and adjust the portfolio weights accordingly. The
second plan is the same as the rst one except for the fact that
the portfolio is held for = 5 trading days before rebalancing.
In the investment horizon (which is fromday 101 to day 200 in
this case), we record the 15-minute portfolio returns based on the
latent prices of the assets, the variation of the portfolio weights
across 50 assets, and other relevant characteristics. While it
appears that 100 trading days is short, calculating 15-minute
returns increases the size of the relevant data for computing the
risk by a factor of 26.
We study those portfolio features for a whole range of gross-
exposure constraint c fromc = 1, which stands for the no-short-
sale portfolio strategy, to c = 3. This is usually the relevant range
of gross exposure for investment purpose.
The standard deviations and other characteristics of the strat-
egy for = 1 are presented in Table 1 (the case = 5 gives
similar comparisons and hence is omitted). The standard de-
viations, which are calculated based on the 15-minute returns
as mentioned before, represent the actual risks of the strategy.
As we only optimize the risk prole, we should not look sig-
nicantly on the returns of the optimal portfolios. They cannot
even be estimated with good accuracy over such a short invest-
ment horizon. Figures 1 and 2 provide graphical details to these
characteristics for both = 1 and = 5.
Table 1. The out-of-sample performance of daily-rebalanced optimal
portfolios with gross-exposure constraint
Std Dev Max Min No. of No. of
Methods % Weight Weight Long Short
Low-frequency sample covariance matrix estimator
c = 1 (No short) 16.69 0.19 0.00 13 0
c = 2 16.44 0.14 0.05 28.5 20
c = 3 16.45 0.14 0.05 28.5 20
High-frequency all-refresh TSCV covariance matrix estimator
c = 1 (No short) 16.08 0.20 0.00 15 0
c = 2 14.44 0.14 0.05 30 19
c = 3 14.44 0.14 0.05 30 19
High-frequency pairwise-refresh TSCV covariance matrix estimator
c = 1 (No short) 15.34 0.18 0.00 15 0
c = 2 12.72 0.13 0.03 31 18
c = 3 12.72 0.13 0.03 31 18
NOTE: We simulate one trial of intra-day trading data for 50 assets, make portfolio
allocations for 100 trading days, and rebalance daily. The standard deviations and other
characteristics of these portfolios are recorded. All the characteristics are annualized (Max
Weight, median of maximum weights; Min Weight, median of minimum weights; No. of
Long, median of numbers of long positions whose weights exceed 0.001; No. of Short,
median of numbers of short positions whose absolute weights exceed 0.001).
From Table 1 and Figures 1 and 2, we see that for both
holding lengths = 1 and = 5, the all-refresh TSCV and
pairwise-refresh TSCVapproaches outperformsignicantly the
low-frequency one in terms of risk prole for the whole range
of the gross-exposure constraints. This supports our theoreti-
cal results and intuitions. First, the shorter estimation window
allows these two high-frequency approaches to deliver consis-
tently better results than the low-frequency one. Second, the
pairwise-refresh method outperforms the all-refresh method, as
expected. Finally, both the low-frequency strategy and the high-
frequency strategies outperform signicantly the equal-weight
portfolio (see Figures 1 and 2).
All the risk curves attain their minimum around c = 1.2 (see
Figures 1 and 2), which meets our expectation again, since
that must be the point where the marginal increase in estimation
error outpaces the marginal decrease in specication error. This,
coupled with the result we get in Empirical Studies section, will
give us some guidelines about what gross-exposure constraint
to use in investment practice.
In terms of portfolio weights, neither the low-frequency nor
the high-frequency optimal no-short-sale portfolios are well di-
versied with all approaches assigning a concentrated weight of
around 20% to one individual asset. Their portfolio risks can be
improved by relaxing the gross-exposure constraint (see Figures
1 and 2).
5. EMPIRICAL STUDIES
Risk minimization problem (6) has important applications
in asset allocation. We demonstrate its application in the stock
portfolio investment in the 30 DowJones Industrial Average (30
DJIA) constituent stocks.
To make asset allocation, we use the high-frequency data
of the 30 DJIA stocks from January 1, 2008, to September 30,
2008. These stocks are highly liquid. The period covers the birth
of nancial crisis in 2008.
At the end of each holding period of = 1 or = 5 trading
days in the investment period (fromMay 27, 2008, to September
30, 2008), the covariance of the 30 stocks is estimated according
to the different estimators. They are the sample covariance of the
previous 100 trading days daily return data (low-frequency), the
all-refresh TSCV estimator of the previous 10 trading days, and
the pairwise-refresh TSCV estimator of the previous 10 trading
days. These estimated covariance matrices are used to construct
optimal portfolios with a range of exposure constraints. For
= 5, we do not count the overnight risks of the portifolio. The
reason is that the overnight price jumps are often due to the
arrival of news and are irrelevant of the topics of our study. The
standard deviations and other characteristics of these portfolio
returns for = 1 are presented in Table 2 together with the stan-
dard deviation of an equally weighted portfolio of the 30 DJIA
stocks rebalanced daily. The standard deviations are for the 15
minutes returns, which represent the actual risks. Figures 3 and
4 provide the graphical details to these characteristics for both
= 1 and = 5.
Table 2 and Figures 3 and 4 reveal that in terms of the
portfolios actual risk, the all-refresh TSCV and pairwise-
refresh TSCV strategies perform at least as well as the
low-frequency-based strategy when the gross exposure is small
and outperform the latter signicantly when the gross exposure
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

Fan, Li, and Yu: Vast Volatility Matrix Estimation Using High-Frequency Data 419
1 1.5 2 2.5 3
10
15
20
25
30
35
40
Exposure Constraint
A
n
n
u
a
l
i
z
e
d

R
i
s
k

(
%
)
(a) Risk
1 1.5 2 2.5 3
0.1
0.11
0.12
0.13
0.14
0.15
0.16
0.17
0.18
0.19
0.2
Exposure Constraint
M
a
x
i
m
u
m

W
e
i
g
h
t
(b) Maximum Weight
Low frequency
Allrefresh
Pairwiserefresh
Oracle (Latent price)
Equalweight
Low frequency
Allrefresh
Pairwiserefresh
Oracle (Latent price)
Figure 1. Out-of-sample performance of daily-rebalanced optimal portfolios based on high-frequency and low-frequency estimation of the
integrated covariance matrix. (a) Annualized risk of portfolios. (b) Maximum weight of allocations. (The online version of this gure is in color.)
is large. Both facts support our theoretical results and intuitions.
Given 10 times the length of covariance estimation window, the
low-frequency approach still cannot perform better than the
high-frequency TSCV approaches, which afrms our belief that
the high-frequency TSCV approaches can signicantly shorten
the necessary covariance estimation window and capture better
the short-term time-varying covariation structure (or the local
covariance). These results, together with the ones presented in
the Simulation Studies section, lend strong support to the above
statement.
1 1.5 2 2.5 3
10
15
20
25
30
35
40
Exposure Constraint
A
n
n
u
a
l
i
z
e
d

R
i
s
k

(
%
)
(a) Risk
1 1.5 2 2.5 3
0.1
0.11
0.12
0.13
0.14
0.15
0.16
0.17
0.18
0.19
0.2
Exposure Constraint
M
a
x
i
m
u
m

W
e
i
g
h
t
(b) Maximum Weight
Low frequency
Allrefresh
Pairwiserefresh
Oracle (Latent price)
Equalweight
Low frequency
Allrefresh
Pairwiserefresh
Oracle (Latent price)
Figure 2. Out-of-sample performance of optimal portfolios based on high-frequency and low-frequency estimation of the integrated covariance
matrix with holding period = 5. (a) Annualized risk of portfolios, (b) maximum weight of allocations. (The online version of this gure is in
color.)
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

420 Journal of the American Statistical Association, March 2012
Table 2. The out-of-sample performance of daily-rebalanced optimal
portfolios of the 30 DJIA stocks
Std Dev Max Min No. of No. of
Methods % Weight Weight Long Short
Low-frequency sample covariance matrix estimator
c = 1 (No short) 12.73 0.50 0.00 8 0
c = 2 14.27 0.44 0.12 16 10
c = 3 15.12 0.45 0.18 18 12
High-frequency all-refresh TSCV covariance matrix estimator
c = 1 (No short) 12.55 0.40 0.00 8 0
c = 2 12.36 0.36 0.10 17 12
c = 3 12.50 0.36 0.10 17 12
High-frequency pairwise-refresh TSCV covariance matrix estimator
c = 1 (No short) 12.54 0.39 0.00 9 0
c = 2 12.23 0.35 0.08 17 12
c = 3 12.34 0.35 0.08 17 12
Unmanaged index
30 DJIA equally weighted 22.12
As the gross-exposure constraint increases, the portfolio risk
of the low-frequency approach increases drastically relative to
the ones of the high-frequency TSCV approaches. The reason
could be a combination of the fact that the low-frequency ap-
proach does not produce a well-conditioned estimated covari-
ance due to the lack of data and the fact that the low-frequency
approach can only attain the long-run covariation but cannot
capture well the local covariance dynamics. The portfolio risk
of the high-frequency TSCV approaches increased only moder-
ately as the gross-exposure constraint increases. From nancial
practitioners standpoint, that is also one of the comparative
advantages of high-frequency TSCV approaches, which means
that investors do not need to be much concerned about the choice
of the gross-exposure constraint while using the high-frequency
TSCV approaches.
It can be seen that both the low-frequency and high-frequency
optimal no-short-sale portfolios are not diversied enough.
Their risk proles can be improved by relaxing the gross-
exposure constraint to around c = 1.2, that is, 10% short po-
sitions and 110% long positions are allowed. The no-short-sale
portfolios under all approaches have the maximum portfolio
weight of 22%50%. As the gross-exposure constraint relaxes,
the pairwise-refresh TSCV approach has its maximum weight
reaching the smallest value around 30%34%, while the low-
frequency approach goes down to only around 40%. This is
another comparative advantage of the high-frequency approach
in practice, as a portfolio with less weight concentration is typ-
ically considered more preferable.
Another interesting fact is that the equally weighted daily-
rebalanced portfolio of the 30 DJIAstocks carries an annualized
return of only 10% while DJIA went down 13.5% during
the same period (May 27, 2008, to Sep 30, 2008), giving an
annualized return of 38.3%. The cause of the difference is that
we intentionally avoided holding portfolios overnight, hence
the portfolios are not affected by the overnight price jumps. In
the turbulent nancial market of May to September 2008, our
portfolio strategies are not affected by the numerous sizeable
downward jumps. Those jumps are mainly caused by the news of
distressed economy and corporations. The moves could deviate
far from what the previously held covariation structure dictates.
1 1.5 2 2.5 3
12
14
16
18
20
22
24
Exposure Constraint
A
n
n
u
a
l
i
z
e
d

R
i
s
k

(
%
)
(a) Risk
1 1.5 2 2.5 3
0.32
0.34
0.36
0.38
0.4
0.42
0.44
0.46
0.48
0.5
Exposure Constraint
M
a
x
i
m
u
m

W
e
i
g
h
t
(b) Maximum Weight
Low frequency
Allrefresh
Pairwiserefresh
Equalweight
Low frequency
Allrefresh
Pairwiserefresh
Figure 3. Out-of-sample performance of daily-rebalanced optimal portfolios for 30 DJIA constituent stocks with investment period from May
27, 2008, to September 30, 2008 (89 trading days). (a) Annualized risk of portfolios. (b) Maximum weight of allocations. (The online version of
this gure is in color.)
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

Fan, Li, and Yu: Vast Volatility Matrix Estimation Using High-Frequency Data 421
1 1.5 2 2.5 3
12
14
16
18
20
22
24
Exposure Constraint
A
n
n
u
a
l
i
z
e
d

R
i
s
k

(
%
)
(a) Risk
1 1.5 2 2.5 3
0.32
0.34
0.36
0.38
0.4
0.42
0.44
0.46
0.48
0.5
Exposure Constraint
M
a
x
i
m
u
m

W
e
i
g
h
t
(b) Maximum Weight
Low frequency
Allrefresh
Pairwiserefresh
Equalweight
Low frequency
Allrefresh
Pairwiserefresh
Figure 4. Out-of-sample performance of 5-day-rebalanced optimal portfolios for 30 DJIA constituent stocks with investment period from
May 27, 2008, to September 30, 2008 (89 trading days). (a) Annualized risk of portfolios. (b) Maximum weight of allocations. (The online
version of this gure is in color.)
6. CONCLUSION
We advocate the portfolio selection with gross-exposure con-
straint (Fan et al. 2011). It is less sensitive to the error of co-
variance estimation and mitigates the noise accumulation. The
out-of-sample portfolio performance depends on the expected
volatility in the holding period. It is at best approximated and the
gross-exposure constraints help reducing the error accumulation
in the approximations.
Two approaches are proposed for the use of high-frequency
data to estimate the integrated covariance: all-refresh and
pairwise-refresh methods. The latter retains far more data
on average and hence estimates more precisely element-by-
element. Yet, the pairwise-refresh estimates are often not posi-
tive semidenite and projections are needed for the convex opti-
mization algorithms. The projection distorts somewhat the per-
formance of the pairwise-refresh strategies. New optimization
algorithms need to be developed in order to take full advantage
of pairwise-refresh. Further investigations on the relative mer-
its of pairwise-refresh, blocking approach, and all-refresh
are needed.
The use of high-frequency nancial data increases signi-
cantly the available sample size for volatility estimation, and
hence shortens the time window for estimation, adapts better to
local covariations. Our theoretical observations are supported by
the empirical studies and simulations, in which we demonstrate
convincingly that the high-frequency-based strategies outper-
form the low-frequency-based one in general.
With the gross-exposure constraint, the impact of the size of
the candidate pool for portfolio allocation is limited. We derive
the concentration inequalities to demonstrate this theoretically.
Simulation and empirical studies also lend further support to it.
APPENDIX: CONDITIONS AND PROOFS
Conditions
We derive our theoretical results under the following conditions. For
simplicity, we state the conditions for integrated covariation (Theorem
2). The conditions for integrated volatility (Theorem 1) are simply the
ones with Y = X.
Condition 1. The drift processes are such that
(X)
t
=
(Y)
t
= 0 for
all t [0, 1].
Condition 2.
(i)
t
, i = X, Y are continuous stochastic processes
which are either bounded by 0 < C

< , or such that the tail behavior

can be described by
P
_
sup
0t 1

(i)
t
C

_
k

exp
_
aC
b

_
, for any C

> 0,
with positive constants k

, a, and b.
Condition 3. The observation times are independent with the X
and Y processes. The synchronized observation times for the X
and Y processes satisfy sup
1j n
n (v
j
v
j1
) C

, where
C

is a nonrandom constant, n is the observation frequency, and

V = {v
0
, v
1
, . . . , v
n
} is the set of refresh times of the processes X and
Y.
Condition 4. For the TSCV parameters, we consider the case when
J = 1 ( n
J
= n) and n
K
= O( n
1/3
) such that
1
2
n
1/3
n
K
2 n
1/3
.
Condition 5. The processes
X
and
Y
are independent.
Remark 1. Conditions 1 and 4 are imposed for simplicity. They can
be removed at the expenses of lengthier proofs. For a short horizon
and high-frequency, whether Condition 1 holds or not has little impact
on the investment. For estimating integrated volatility, the synchro-
nized time becomes observation time {
n,j
} and Conditions 3 and 4
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

422 Journal of the American Statistical Association, March 2012
become
sup
1jn
n (
n,j

n,j1
) C

< (A.1)
and
1
2
n
1/3
n
K
2 n
1/3
.
Remark 2. Note that indeed, in the derivations of the Theorems 1
and 2,
(X)
t
and
(Y)
t
are only required to be c` adl` ag. In other words, the
continuity assumption is not needed for the concentration inequalities
to hold. The continuity of the volatility processes is only needed in
approximation (5) in our study, it can be removed or relaxed for other
applications of the concentration inequalities.
Lemmas
We need the following three lemmas for the proofs of Theorems
1 and 2. In particular, Lemma 2 is exponential type of inequality for
any dependent random variables that have a nite moment generating
function. It is useful for many statistical learning problems. Lemma 3 is
a concentration inequality for the realized volatility based on discretely
observed latent process.
Lemma 1. When Z N(0, 1), for any ||
1
4
, E exp{(Z
2

1)} exp(2
2
).
Proof. Using the moment generating function of Z
2

2
1
, we have
E exp{(Z
2
1)} = exp
_

1
2
log(1 2)
_
.
Let g(x) = log(1 x) +x +x
2
with |x| 1/2. Then, g

(x) =
x(1 2x)/(1 x) is nonnegative when x [0, 1/2] and negative when
x [1/2, 0). In other words, g(x) has a minimum at point 0, namely
g(x) 0 for |x| 1/2. Consequently, for || 1/4, log(1 2)
2 (2)
2
. Hence, E exp{(Z
2
1)} exp(2
2
).
Lemma 2. For a set of random variables X
i
, i = 1, . . . , K, and an
event A, if there exists two positive constants C
1
and C
2
such that for
all || C
1
,
E (exp(X
i
)I
A
) exp(C
2

2
), (A.2)
then for w
i
s being weights satisfying

K
i=1
|w
i
| w [1, ), we
have
P

i=1
w
i
X
i

> x

2 exp
_

x
2
4C
2
w
2
_
, when 0 x 2C
1
C
2
.
Proof. By the Markov inequality, for 0 C
1
/w and w

K
i=1
|w
i
|, we have
P

i=1
w
i
X
i

> x

exp(x)E

exp

i=1
w
i
X
i

I
A

exp(x)w
1
K

i=1
|w
i
|E (exp(w|X
i
|)I
A
)
2 exp
_
C
2

2
w
2
x
_
. (A.3)
Taking = x/(2C
2
w
2
), we have
P

i=1
w
i
X
i

> x

2 exp
_

x
2
4C
2
w
2
_
,
when 0 x 2C
1
C
2
. (A.4)

Lemma 3. (A Concentration Inequality for Realized Volatility)

Let

[X, X]
1
=

n
i=1
(X
v
i
X
v
i1
)
2
be the realized volatility based on
the discretely observed X process frommodel (1) of the univariate case:
dX
t
=
t
dt +
t
dW
t
. Under Conditions 13,
(A) if
(X)
t
C

< for all t [0, 1], then for all large n, for
x [0, c

n],
P

n
1/2

[X, X]
1

_
1
0

2
t
dt

> x

2 exp{Cx
2
},
where the constants c and C can be taken as in (A.7).
(B) If the tail behavior of
(X)
t
satises,
P
_
sup
0t 1

(X)
t
C

_
k

exp
_
aC
b

_
, for any C

> 0
with positive constants k

, a, and b, then for all large n, for

x [0, cn
4+b
2b
],
P

n
1/2

[X, X]
1

_
1
0

2
t
dt

> x

(2 +k

) exp
_
Cx
2b
4+b
_
.
A set of candidate values for c and C are given in (A.8).
Proof. For any constant C

> 0, dene a stopping time

C
:=
inf{t : sup
0st

s
> C

} 1. Let

s
=
_

s
, when s
C
C

, when s >
C
and

X
t
=
_
t
0

s
dW
s
. By time change for martingales, [see, for example,
Theorem 4.6 of chapter 3 of Karatzas and Shreve (2000)], if
t
=
inf{s : [

X]
s
t } where [

X]
s
is the quadratic variation process, then
B
t
:=

X
t
is a Brownian-motion with regard to {F
t
}
0t
. We then
have that
E exp

X
2
t

_
t
0

2
s
ds

= E exp
_

_
B
2
[

X]t
[

X]
t
__
.
Note further that for any t, [

X]
t
is a stopping time with regard to
{F
s
}
0s
, and the process exp((B
2
s
s)) is a submartingale for any
. By the optional sampling theorem, using [

X]
u
C
2

u (bounded
stopping time), we have
E exp
_

_
B
2
[

X]u
[

X]
u
__
E exp
_

_
B
2
C
2

u
C
2

u
__
.
Therefore, note that X
i
=

X
i
on the set of {
C
= 1}, we have that,
under Condition 3,
E

exp

(X
i
)
2

_
v
i
v
i1

2
t
dt

I
{
C
=1}
|F
v
i1

E exp
_

n
_
B
2
C
2

C
2

n
__
= E exp
_

C
2

n
(Z
2
1)
_
, (A.5)
where Z N(0, 1) and X
i
= X
v
i
X
v
i1
.
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

Fan, Li, and Yu: Vast Volatility Matrix Estimation Using High-Frequency Data 423
It follows from the law of iterated expectations and (A.5) that
E

exp

[X, X]
1

_
1
0

2
t
dt

I
{
C
=1}

= E

exp

n1

i=1
(X
i
)
2

_
v
n1
0

2
t
dt

I
{
C
=1}

exp

X
2
n

_
vn
v
n1

2
t
dt

I
{
C
=1}
|F
v
n1

exp

n1

i=1
(X
i
)
2

_
v
n1
0

2
t
dt

I
{
C
=1}

E exp
_

C
2

n
(Z
2
1)
_
,
where Z N(0, 1). Repeating the process above, we obtain
E

exp

[X, X]
1

_
1
0

2
t
dt

I
{
C
=1}

_
E exp
_

C
2

n
(Z
2
1)
__
n
.
By Lemma 1, we have for ||

n
4C
2

,
E

exp

[X, X]
1

_
1
0

2
t
dt

I
{
C
=1}

exp
_
2
2
C
4

C
2

_
.
(A.6)
By Lemma 2, we have,
P

n
1/2

[X, X]
1

_
1
0

2
t
dt

> x

{
C
= 1}

2 exp
_

x
2
8C
4

C
2

_
, (A.7)
when 0 x C
2

n. This proves the rst half of the theorem.

For the second half,
P

n
1/2

[X, X]
1

_
1
0

2
t
dt

> x

n
1/2

[X, X]
1

_
1
0

2
t
dt

> x,
C
= 1

+P{
C
< 1}
exp{x}E

exp

[X, X]
1

_
1
0

2
t
dt

I
{
C
=1}

+P{
C
< 1} for nonnegative
2 exp
_

x
2
8C
4

C
2

_
+k

exp
_
aC
b

_
,
when 0 x C
2

n.
Now, let C

= (
x
2
8aC
2

)
1
4+b
, we have that when 0 x (8a)
2
b
C

n
4+b
2b
,
P

n
1/2

[X, X]
1

_
1
0

2
t
dt

> x

(2 +k

) exp
_
8
b
4+b
a
4
4+b
C
2b
4+b

x
2b
4+b
_
. (A.8)

Proof of Theorem 1
We rst prove the results conditioning on the set of observation
times V. Recall notation introduced in Sections 3.2 and 3.3. Let
n be the observation frequency. For simplicity of notation, with-
out ambiguity, we will write
n,i
as
i
and
(X)
t
as
t
. Again let

C
:= inf{t : sup
0st

s
> C

} 1. Denote the TSRV based on the

latent process by

X, X
(K)
1
=

[X, X]
(K)
1

n
K
n
J

[X, X]
1
(J)
, (A.9)
where

[X, X]
(K)
1
= K
1

n
i=K
(X

i
X

iK
)
2
. Then, from the deni-
tion, we have,

X, X
1
=

[X, X]
(K)
1
+

[
X
,
X
]
(K)
1
+2

[X,
X
]
(K)
1

n
K
n
J
_

[X, X]
(J)
1
+

[
X
,
X
]
(J)
1
+2

[X,
X
]
(J)
1
_
=
1
K
K1

l=0
V
(l)
K

n
K
n
J

[X, X]
(J)
1
+R
1
+R
2
, (A.10)
where R
1
=

[
X
,
X
]
(K)
1

n
K
n
J

[
X
,
X
]
(1)
1
, R
2
= 2

[X,
X
]
(K)
1

2
n
K
n
J

[X,
X
]
(1)
1
, and
V
(l)
K
=
n
K

j=1
(X

jK+l
X

(j1)K+l
)
2
, for l = 0, 1, . . . , K 1.
Note that we have assumed that n
K
=
nK+1
K
is an integer above, to
simplify the presentation.
Recall that we consider the case when J = 1, or n
J
= n. Let
I
1
=
1
K
K1

l=0

n
K

V
(l)
K

_
1
0

t
2
dt

_
n
K
n
_3
2

[X, X]
(1)
1

_
1
0

t
2
dt

n
K
R
1
+

n
K
R
2
, (A.11)
and I
2
=
n
3/2
K
n
_
1
0

t
2
dt . We are interested in

n
K

X, X
1

_
1
0

t
2
dt

= I
1
+I
2
.
The key idea is to compute the moment generating functions for each
term in I
1
and then to use Lemma 2 to conclude.
For the rst term in I
1
, since V
(l)
k
is a realized volatility based on
discretely observed X process, with observation frequency satisfying
sup
1i n
K
n
K
(
iK+l

(i1)K+l
) C

, we have, by (A.6) in Lemma

3, for ||

n
K
4C
2

,
E

exp

n
K
(V
(l)
K

_
1
0

2
t
dt

I
{
C
=1}

exp
_
2
2
C
4

C
2

_
.
(A.12)
For the second term in I
1
, we have obtained in (A6) that
E

exp

[X, X]
(1)
1

_
1
0

t
2
dt

I
{
C
=1}

exp
_
2
2
C
4

C
2

_
, when ||

n
4C
2

. (A.13)
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

424 Journal of the American Statistical Association, March 2012
We introduce an auxiliary sequence a
n
that grows with n at a mod-
erate rate to facilitate our presentation in the following. In particular,
we can set a
n
= n
1/12
. Let us now deal with R
1
, the third term in I
1
.
Note that from the denition

n
K
R
1
=

n
K
K

i=K
(
i

iK
)
2

n K +1
n
n

i=1
(
i

i1
)
2

n
K
(n K +1)
K

n
n

i=1

i1

n
K

n K +1
K

n K +1
n

i=K

iK

n
K

K 1a
n
K

1
a
n

K 1
K1

i=1
_

2
i

2
X
_

n
K

K 1a
n
K

1
a
n

K 1
n1

nK+1
_

2
i

2
X
_
+

n
K
(K 1)a
n
K

1
a
n

n
n

i=1
_

2
i

2
X
_
+

n
K
(K 1)a
n
K

1
a
n

n
n1

i=0
_

2
i

2
X
_
.
(A.14)
The rst two terms in (A.14) are not sums of independent variables.
But they can be decomposed into sums of independent randomvariables
and the moment generating functions can be computed. To simplify the
argument without losing the essential ingredient, let us focus on the
rst term of (A.14). We have the following decomposition
n

i=1

i1
=

oddi

i1
+

eveni

i1
,
and the summands in each terms of the right-hand side are now inde-
pendent. Therefore, we need only to calculate the moment generating
function of
i

i1
.
For two independent normally distributed random variables X
N(0,
2
X
) and Y N(0,
2
Y
), it can easily be computed that
E(exp{n
1/2
XY}) =
_
1
1
2
X

2
Y

2
/n
_
1/2
exp
_

2
X

2
Y

2
/n
_
when ||

2
X

Y
,
where we have used the fact that log(1 x) 2x when 0 x
1
2
.
Hence, by independence, it follows that (we assume n is even to
simplify the presentation)
E exp

2n
1/2

oddi

i1

=
_
1
1 4
4
X

2
/n
_
n/4
exp
_
2
4
X

2
_
, when ||

n
2

2
2
X
. (A.15)
The second term in R
1
works similarly and has the same bound. For
example, when n
K
is even, one can have the following decomposition
n

i=K

iK
=
n
K
/2

j=1
2jK1

i=2jKK

iK
+
n
K
/2

j=1
2jK+K1

i=2jK

iK
.
The last four terms are sums of independent
2
-distributed random
variables and their moment generating functions can easily be bounded
by using Lemma 1. Taking the term
1
an

K1

K1
i=1
(
2
i

2
X
) for exam-
ple, we have
E

exp

a
n

K 1
K1

i=1
_

2
i

2
X
_

exp
_
2
4
X

2
/a
2
n
_
when ||
a
n

K 1
4
2
X
.
For the term R
2
, we have,

n
K
R
2
=
2a
n
n
K
n
1
a
n

i=1
X
i

i1

i=1
X
i

+
2
a
n
a
n

n
K
K

i=K

(K)
X
i

i

n

i=K

(K)
X
i

iK

, (A.16)
where X
i
= X

i
X

i1
, and
(K)
X
i
= X

i
X

iK
. The rst term
above satises
E

exp

a
n
n

i=1
X
i

I
{
C
=1}

= E

exp

i=1
_

a
n
X
i
_
2

2
X
/2

I
{
C
=1}

_
E
_
exp
_

2
X
C
2

Z
2
/2na
2
n
___
n
=
_
1
1
2
X
C
2

2
/na
2
n
_
n/2
exp
_

2
X
C
2

2
/a
2
n
_
, when ||

na
n

2C

X
, (A.17)
where in the second line we have again used the optional sampling
theoremand lawof iterated expectations as in the derivations of Lemma
3; Z denotes a standard normal random variable. The second term in
R
2
works similarly and has the same bound. For the third term, by
conditioning on the X-process rst, we have
E
_
exp
_
a
n

n
K
K
n

i=K

(K)
X
i

i
_
I
{
C
=1}
_
= E

exp

a
2
n

2
n
K
2K
2
K1

l=0
n
K

j=1
_

(K)
X
jK+l
_
2

2
X

I
{
C
=1}

K1
l=0

exp

a
2
n

2
n
K

2
X
2K
n
K

j=1
_

(K)
X
jK+l
_
2

I
{
C
=1}

1
K

K1
l=0

1
a
2
n

2
X
K
C
2

n
K
/2

1
K
exp
_
a
2
n

2
n
K

2
X
K
C
2

_
when ||

2C

a
n
C

,
(A.18)
where we have used the H olders inequality above. The fourth term
works similarly and has the same bound.
Combining the results for all the terms (A.12)(A.18) together, ap-
plying Lemma 2 to I
1
, we conclude that the conditions for Lemma 2
are satised with A = {
C
= 1}, C
1
= C
1,x

n
K
,
C
1,x
= min

1
4C
2

n/ n
K
2

2
2
X
,
a
n

(K 1)/ n
K
4
2
X
,
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

Fan, Li, and Yu: Vast Volatility Matrix Estimation Using High-Frequency Data 425
a
n

n/ n
K

2C

X
C

K/ n
K

2C

a
n
C

(A.19)
=
1
4C
2

for big enough n

and
C
2
= max
_
2C
4

C
2

, 2
4
X
, 2
4
X
/a
2
n
,
2
X
C
2

/a
2
n
,
a
2
n
n
K

2
X
K
C
2

_
(A.20)
= max
_
2C
4

C
2

, 2
4
X
_
for big enough n
= 2C
4

C
2

for the typical case when C

1 and C

X
.
Let w = 8, which is larger, when n is sufciently large, than
1
K
K1

l=0
1 +
_
n
K
n
_
3/2
. ,, .
coef. in the rst two terms of I
1
+
2

n
K

n
K
+
2

n
K
a
n

K
+
2

n
K
a
n

n
. ,, .
controls coefcients in (A.14)
+
4a
n
n
K
n
+
4
a
n
. ,, .
coefcients in (A.16)
.
Set C
I
1
= (4C
2
w
2
)
1
. By Lemma 2, when 0 x 2C
1,x
C
2

n
K
,
P
_
{|I
1
| > x}
_

C
= 1
__
2 exp(C
I
1
x
2
).
Hence when 0 x 4C
1,x
C
2

n
K
,
P
__
|I
1
| >
x
2
_

C
= 1
_
_
2 exp
_

C
I
1
4
x
2
_
. (A.21)
For the term I
2
, let C
I
2
= 2
3/2
C
2

, we have, by Condition 4,
I
2
=
n
3/2
K
n
_
1
0

t
2
dt C
I
2
/

n, on the set {
C
= 1}. (A.22)
Hence
P
__
|I
2
| >
x
2
_

C
= 1
_
_

_
0, if x > 2C
I
2
/

n
1, if x 2C
I
2
/

n.
Since for all large n, 1 2 exp(
C
I
1
4C
2
I
2
4n
), which is smaller than
2 exp(
C
I
1
x
2
4
) when x
2C
I
2

n
, we have that for all large n,
P
__
|I
2
| >
x
2
_

C
= 1
_
_
2 exp
_

C
I
1
4
x
2
_
. (A.23)
Combining (A.21) and (A.23), we have, when 0 x
4C
1,x
C
2

n
K
,
P
__

n
K

X, X
1

_
1
0

t
2
dt

> x
_

C
= 1
_
_
= P
_
{|I
1
+I
2
| > x}
_

C
= 1
__
P
_
{|I
1
| > x/2}
_

C
= 1
__
+P
_
{|I
2
| > x/2}
_

C
= 1
__
4 exp
_

C
I
1
4
x
2
_
.
By Condition 4 again, we have, when 0 x cn
1/6
,
P
__
n
1/6

X, X
1

_
1
0

t
2
dt

> x
_

C
= 1
_
_
P
__

n
K

X, X
1

_
1
0

t
2
dt

> x/

2
_

C
= 1
_
_
4 exp(Cx
2
), (A.24)
where c = 4

2C
1,x
C
2

n
K
, and C =
C
I
1
8
= (32C
2
w
2
)
1
. For big
enough n and the typical case when C

1 and C

X
, we have
c = 2

2C
2

and C =
1
64w
2
C
4

C
2

. (A.25)
Notice that this conditional result depends only on the observation
frequency n and not on the locations of the observation times as long
as Condition 3 is satised, (A.24) holds unconditionally on the set of
the observation times. This proves the rst half of the Theorem 1 when

C
1.
For the second half of the theorem, we have
P
_
n
1/6

X, X
1

_
1
0

t
2
dt

> x
_
= P
__
n
1/6

X, X
1

_
1
0

t
2
dt

> x
_

C
= 1
_
_
+P
__

C
< 1
__
< 4 exp
_

x
2
64w
2
C
4

C
2

_
+k

exp
_
aC
b

_
,
when 0 x 2

2C
2

n
1/6
.
Letting C

= (
x
2
64w
2
aC
2

)
1
4+b
, we have, when 0 x
2
3b12
2b
C

(w
2
a)
2
b
n
4+b
6b
,
P
_
n
1/6

[X, X]
1

_
1
0

2
t
dt

> x
_
(4 +k

) exp
_
(64w
2
)
b
4+b
a
4
4+b
C
2b
4+b

x
2b
4+b
_
. (A.26)
Remark 3. In the above proof, we have demonstrated by using a
sequence a
n
that goes to at a moderate rate that one can eliminate
the impact of the small order terms on the choices of the constants,
as long as the terms have their moment generating functions satisfy
inequalities of form (A.2). We will use this technique again in the next
subsection.
Proof of Theorem 2
We again conduct all the analysis assuming that the observation
times are given. Our nal result holds because the conditional result
does not depend on the locations of the observation times as long as
Condition 3 is satised.
Recall notation for the observation times as introduced in Section
3.2. Dene
Z
+
= X +Y and Z

= X Y.
Z
+
and Z

are diffusion processes with volatilities satisfying Condi-

tion 2. To see this, let W
+
and W

be processes such that

dW
+
t
=

(X)
t
dB
(X)
t
+
(Y)
t
dB
(Y)
t
_
_

(X)
t
_
2
+
_

(Y)
t
_
2
+2
t

(X)
t

(Y)
t
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

426 Journal of the American Statistical Association, March 2012
and
dW

t
=

(X)
t
dB
(X)
t

(Y)
t
dB
(Y)
t
_
_

(X)
t
_
2
+
_

(Y)
t
_
2
2
t

(X)
t

(Y)
t
.
W
+
and W

are standard Brownian motions by Levys characterization

of Brownian motion [see, for example, Theorem 3.16, chapter 3 of
Karatzas and Shreve (2000)]. Write

Z
+
t
=
_
_

(X)
t
_
2
+
_

(Y)
t
_
2
+2
t

(X)
t

(Y)
t
and

t
=
_
_

(X)
t
_
2
+
_

(Y)
t
_
2
2
t

(X)
t

(Y)
t
,
we have
dZ
+
=
Z
+
t
dW
+
t
and dZ

=
Z

t
dW

t
with
0
Z
+
t
,
Z

t
2C

when
(X)
t
,
(Y)
t
are bounded by C

, or
P
_
sup
0t 1

Z
+
t
2C

_
P
_
sup
0t 1

X
t
C

_
+P
_
sup
0t 1

Y
t
C

_
2k

exp
_
aC
b

_
,
when
X
t
,
Y
t
are such that their tail probabilities are bounded as in
Condition 2.
In fact, let
C
:= inf{t : sup
0st

(X)
s
> C

or sup
0st

(Y)
s
>
C

} 1. We have P{
C
< 1} P{sup
0t 1

X
t
C

} +
P{sup
0t 1

Y
t
C

} 2k

exp{aC
b

}.
For the observed Z
+
and Z

processes, we have
Z
+,o
v
i
= X
o
t
i
+Y
o
s
i
= Z
+
v
i
+
i,+
and
Z
,o
v
i
= X
o
t
i
Y
o
s
i
= Z

v
i
+
i,
,
where t
i
and s
i
are the last ticks at or before v
i
and

i,+
= X
t
i
X
v
i
+Y
s
i
Y
v
i
+
X
i
+
Y
i
,

i,
= X
t
i
X
v
i
Y
s
i
+Y
v
i
+
X
i

Y
i
.
Note that

X, Y
1
=
1
4
(

Z
+
, Z
+

, Z

1
). We can rst prove
analogous results as in Theorem 1 for

Z
+
, Z
+

1
and

Z

, Z

1
, then
use the results to obtain the nal conclusion for TSCV.
For

Z
+
, Z
+

1
, the derivation is different from that of Theorem 1
only for the terms that involve the noise, namely

n
K
R
1
and

n
K
R
2
.
Write

X
i
= X
t
i
X
v
i
and

Y
i
= Y
s
i
Y
v
i
. Then, the rst term in

n
K
R
1
becomes

n
K

n
K

n
n

i=1

i,+

i1,+
=

n
K

n
K

i=1
_

X
i

X
i1
+

X
i

Y
i1
+

X
i
_

X
i1
+
Y
i1
_
+

Y
i

X
i1
+

Y
i

Y
i1
+

Y
i
_

X
i1
+
Y
i1
_
+
_

X
i
+
Y
i
_

X
i1
+
_

X
i
+
Y
i
_

Y
i1
+
_

X
i
+
Y
i
__

X
i1
+
Y
i1
_
_
.
The only O
P
(1) term is the last term, which involves
only independent normals, and can be dealt with the same
way as before (again assume n is even for simplicity of
presentation):
E exp
_
2 n
1/2

odd i
_

X
i
+
Y
i
__

X
i1
+
Y
i1
_
_
= E exp
_
2 n
1/2

even i
_

X
i
+
Y
i
__

X
i1
+
Y
i1
_
_
=

1
1 4
_

2
X
+
2
Y
_
2

2
/ n

n/4
exp
_
2
_

2
X
+
2
Y
_
2

2
_
, when ||

n
2

2
_

2
X
+
2
Y
_.
The other terms are of a smaller order of magnitude. By applying an
a
n
sequence which grows moderately with n as in Proof of Theorem
1 (we can set a
n
= n
1/12
), we can see easily that as long as we can
show that the moment generating functions of these terms can indeed
be suitably bounded as in (A.2), their exact bounds do not have effect
on our choice of C
1
, C
2
, or . To show the bounds for the moment
generating functions, rst note that, for any positive number a and real
valued b, by the optional sampling theorem (applied to submartingales
exp(aB
2
s
) and exp(b

yB
s
) with stopping time [X]
u
C
2

u for real
number

y), we have
E
_
exp{a(

X
i
)
2
}I
{
C
=1}
|F
i1
_

_
E
_
exp
_
aC
2

Z
2
/ n
__
_
for Z N(0, 1)
=
_
1
1 2aC
2

/ n
_
1/2
, (A.27)
where F
i
is the information collected up to time v
i
. Inequality (A.27)
holds when

X
i
is replaced by

Y
i
. Similarly,
E
_
exp{b

X
i

Y
i1
}I
{
C
=1}
|F
i2
_
E
_
E(exp{b

X
i

Y
i1
}I
{
C
=1}
|F
i1
)|F
i2
_
E
_
exp{b
2
C

C
2

(

Y
i1
)
2
/2 n}I
{
C
=1}
|F
i2
_

_
1
1 b
2
C
4

C
2

/ n
2
_
1/2
. (A.28)
Inequalities (A.27) and (A.28) can be used to obtain the bounds we
need. For example, by (A.28) and the law of iterated expectations,
E
_
exp
_

oddi

X
i

Y
i1
_
I
{
C
=1}
_

_
1
1
2
C
4

C
2

/ n
2
_
n/4
exp
_

2
C
4

C
2

_
2 n
_
when ||
n

2C
2

,
and by independence, normality of the noise, the law of iterated expec-
tations, and (A.27), we have
E
_
exp
_

a
n
n

i=1

X
i
_

X
i1
+
Y
i1
_
_
I
{
C
=1}
_
= E
_
exp
_
n

i=1
_

a
n

X
i
_
2
_

2
X
+
2
Y
__
2
_
I
{
C
=1}
_

1
1
_

2
X
+
2
Y
_

2
C
2

/ na
2
n

n/2
exp
__

2
X
+
2
Y
_
C
2

2
/a
2
n
_
,
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

Fan, Li, and Yu: Vast Volatility Matrix Estimation Using High-Frequency Data 427
when ||

na
n
C

_
2C

(
2
X
+
2
Y
)
.
Similar results can be found for the other terms above, with the same
techniques.
The second termin

n
K
R
1
works similarly and has the same bound.
The other terms in

n
K
R
1
(with
2
X
replaced by
2
X
+
2
Y
) and the whole
term of

n
K
R
2
are of order o
p
(1) and have good tail behaviors. Again,
by using a sequence a
n
, we can conclude that their exact bounds will
not matter in our choice of the constants and we only need to show
that their moment generating functions are appropriately bounded as
in (A.2). The arguments needed to prove the inequalities of form (A.2)
for each elements in these terms are similar to those presented in the
above proofs, and are omitted here.
Hence, by still letting w = 8 and redening
C
1,x
=
1
4(2C

)
2
C

and
C
2
= max
_
2(2C

)
4
C
2

, 2
_

2
X
+
2
Y
_
2
_
= 32C
4

C
2

for the typical case when C

X
,
Y
,
we have, when 0 x c

n
1/6
,
P
__
n
1/6

Z
+
, Z
+

_
1
0

Z
+
t
2
dt

> x
_
_
{
C
= 1}
_
4 exp(C

x
2
),
and
P
__
n
1/6

, Z

_
1
0

t
2
dt

> x
_
_
{
C
= 1}
_
4 exp(C

x
2
),
where c

= 4

2C
1,x
C
2
and C

= (32C
2
w
2
)
1
.
Finally, for the TSCV, when 0 x c n
1/6
,
P
__
n
1/6

X, Y
1

_
1
0

(X)
t

(Y)
t

(X,Y)
t
dt

> x
_
_
{
C
= 1}
_
P
__
n
1/6

Z
+
, Z
+

_
1
0

Z
+
t
2
dt

> 2x
_
_
{
C
= 1}
_
+P
__
n
1/6

, Z

_
1
0

t
2
dt

> 2x
_
_
{
C
= 1}
_
8 exp(Cx
2
),
where c = c

/2 = 2

2C
1,x
C
2
and C = 4C

= (8C
2
w
2
)
1
. For big
enough n and the typical case when C

1 and C

X
, we have
c = 4

2C
2

and C =
_
256w
2
C
4

C
2

_
1
. (A.29)
This completes the proof of the rst half of the statement of Theorem
2, when
C
1.
For the second half of the theorem, we have
P
_
n
1/6

X, Y
1

_
1
0

(X)
t

(Y)
t

(X,Y)
t
dt

> x
_
= P
__
n
1/6

X, Y
1

_
1
0

(X)
t

(Y)
t

(X,Y)
t
dt

> x
_
{
C
< 1}
_
+P({
C
< 1})
< 8 exp
_

_
256w
2
C
4

C
2

_
1
x
2
_
+2k

exp
_
aC
b

_
,
when 0 x 4

2C
2

n
1/6
.
Let C

= (
x
2
256w
2
aC
2

)
1
4+b
, the above inequality becomes
P
_
n
1/6

X, Y
1

_
1
0

(X)
t

(Y)
t

(X,Y)
t
dt

> x
_
(8 +2k

) exp
_
(256w
2
)
b
4+b
a
4
4+b
C
2b
4+b

x
2b
4+b
_
, (A.30)
when 0 x 2
5b12
2b
C

(w
2
a)
2
b
n
4+b
6b
.
Remark 4. Note that the argument is not restricted to TSCV based
on the pairwise-refresh timesit works the same (only with n re-
placed by n

, the observation frequency of the all-refresh method)

for the case when the synchronization scheme is chosen to be the all-
refresh method, as long as the sampling conditions (Conditions 34) are
satised.
[Received November 2010. Revised October 2011.]
REFERENCES
At-Sahalia, Y., Fan, J., and Xiu, D. (2010), High Frequency Covariance Esti-
mates With Noisy and Asynchronous Financial Data, Journal of the Amer-
ican Statistical Association, 105, 15041517. [412]
At-Sahalia, Y., Mykland, P. A., and Zhang, L. (2005), How Often to Sample a
Continuous-Time Process in the Presence of Market Microstructure Noise,
Review of Financial Studies, 18, 351416. [412]
Barndorff-Nielsen, O. E., Hansen, P. R., Lunde, A., and Shephard, N. (2008),
Multivariate Realised Kernels: Consistent Positive Semi-Denite Estima-
tors of the Covariation of Equity Prices With Noise and Non-Synchronous
Trading, Journal of Econometrics, forthcoming. [412,414,417]
(2009), Realized Kernels in Practice: Trades and Quotes, Economet-
rics Journal, 12, 132. [412]
(2011), Subsampling Realised Kernel, Journal of Econometrics, 160,
204219. [412]
Bickel, P. J., and Levina, E. (2008), Regularized Estimation of Large Covari-
ance Matrices, The Annals of Statistics, 36, 199227. [412]
Fan, J., Fan, Y., and Lv, J. (2008), Large Dimensional Covariance Matrix
Estimation Via a Factor Model, Journal of Econometrics, 147, 186197.
[412]
Fan, J., and Wang, Y. (2007), Multi-Scale Jump and Volatility Analysis for
High-Frequency Financial Data, Journal of the American Statistical Asso-
ciation, 102, 13491362. [412]
Fan, J., Zhang, J., and Yu, K. (2011), Asset Allocation and Risk Assessment
With Gross Exposure Constraints for Vast Portfolios, Journal of American
Statistical Association, forthcoming. [412,413,414,421]
Hautsch, N., Kyj, R. M., and Oomen, L. C. A. (2009), A Blocking and Regu-
larization Approach to High Dimensional Realized Covariance Estimation,
Journal of Applied Econometrics, forthcoming. [412,414]
Hayashi, T., and Yoshida, N. (2005), On Covariance Estimation of Non-
Synchronously Observed Diffusion Processes, Bernoulli, 11, 359379.
[412]
Jacod, J., Li, Y., Mykland, P. A., Podolskij, M., and Vetter, M. (2009), Mi-
crostructure Noise in the Continuous Case: The Pre-Averaging Approach,
Stochastic Processes and Their Applications, 119, 22492276. [412]
Jagannathan, R., and Ma, T. (2003), Risk Reduction in Large Portfolios: Why
Imposing the Wrong Constraints Helps, Journal of Finance, 58, 16511683.
[412,413]
Karatzas, I., and Shreve, S. E. (2000), Brownian Motion and Stochastic Calculus
(2nd ed.), New York: Springer. [422,426]
Kinnebrock, S., Podolskij, M., and Christensen, K. (2010), Pre-Averaging
Estimators of the Ex-Post Covariance Matrix in Noisy Diffusion Mod-
els With Non-Synchronous Data, Journal of Econometrics, 159, 116
133. [412]
Lam, C., and Fan, J. (2009), Sparsistency and Rates of Convergence in Large
Covariance Matrices Estimation, The Annals of Statistics, 37, 42544278.
[412]
Ledoit, O., and Wolf, M. (2004), Honey, I Shrunk the Sample Covariance
Matrix, Journal of Portfolio Management, 30, 110119. [413]
Li, Y., and Mykland, P. (2007), Are Volatility Estimators Robust With Respect
to Modeling Assumptions?, Bernoulli, 13, 601622. [412]
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2

428 Journal of the American Statistical Association, March 2012
Markowitz, H. (1959), Portfolio Selection: Efcient Diversication of Invest-
ments, New York: Wiley. [412]
Markowitz, H. M. (1952), Portfolio Selection, Journal of Finance, 7, 7791.
[412]
Rothman, A. J., Levina, E., and Zhu, J. (2009), Generalized Thresholding of
Large Covariance Matrices, Journal of the American Statistical Associa-
tion, 104, 177186. [412]
Tao, M., Wang, Y., Yao, Y., and Zou, J. (2011), Large Volatility Ma-
trix Inference via Combining Low-Frequency and High-Frequency Ap-
proaches, Journal of the American Statistical Association, 106, 1025
1040. [412]
Wang, Y., and Zou, J. (2010), Vast Volatility Matrix Estimation for High-
Frequency Financial Data, The Annals of Statistics, 38, 943978. [412]
Xiu, D. (2010), Quasi-Maximum Likelihood Estimation of Volatility With
High Frequency Data, Journal of Econometrics, 159, 235250. [412]
Zhang, L. (2006), Efcient Estimation of Stochastic Volatility Using Noisy
Observations: A Multi-Scale Approach, Bernoulli, 12, 10191043. [412]
(2011), Estimating Covariation: Epps Effect and Microstructure
Noise, Journal of Econometrics, 160, 3347. [412,414,415]
Zhang, L., Mykland, P. A., and At-Sahalia, Y. (2005), A Tale of Two
Time Scales: Determining Integrated Volatility With Noisy High-Frequency
Data, Journal of the American Statistical Association, 100, 13941411.
[412,415]
Zheng, X., and Li, Y. (2011), On the Estimation of Integrated Covariance Ma-
trices of High Dimensional Diffusion Processes, The Annals of Statistics,
39(6), 31213151. [412]
D
o
w
n
l
o
a
d
e
d

b
y

[
P
r
i
n
c
e
t
o
n

U
n
i
v
e
r
s
i
t
y
]

a
t

0
5
:
4
1

1
9

S
e
p
t
e
m
b
e
r

2
0
1
2