Vous êtes sur la page 1sur 15

Electronic copy available at: http://ssrn.

com/abstract=1301031
Dynamic portfolio management: an application of Fourier
method for covariance estimation
Maria Elvira Mancino

, Elena Rapini

and Simona Sanfelici

Abstract
The economic benet of applying the Fourier covariance estimation methodology over
other estimators in the presence of market microstructure noise is studied from the perspec-
tive of an asset-allocation decision problem. We nd that using Fourier methodology yields
statistically signicant gains.
JEL: G11, C14, C22.
Keywords: nonparametric covariance estimation, non-synchronicity, microstructure, Fourier
analysis, optimal portfolio choice.
1 Introduction
The recent availability of large high frequency nancial data sets potentially provides a rich
source of information about asset price dynamics. Specically, nonparametric variance/covariance
measures constructed by summing intra-daily return data (i.e. realized variances and covari-
ances) have the potential to provide very accurate estimates of the underlying quadratic vari-
ation and covariation and, as a consequence, accurate estimation of betas for asset pricing,
index autocorrelation, lead-lag patterns. These measures, however, have been shown to be sen-
sitive to market microstructure noise inherent in the observed asset prices. Moreover, it is well
known from [Epps, 1979] that the non-synchronicity of observed data leads to a bias towards
zero in correlations among stocks as the sampling frequency increases. Motivated by these dif-
culties, some modications of realized covariance type estimators have been proposed in the
literature: [Martens, 2004], [Hayashi and Yoshida, 2005], [Voev and Lunde, 2007], [Large, 2007],
[Barndor-Nielsen and al., 2008].
A dierent methodology has been proposed in [Malliavin and Mancino, 2002], which is
explicitly conceived for the multivariate analysis. This method is based on Fourier analysis
and does not rely on any data synchronization procedure but employs all the available data.
[Mancino and Sanfelici, 2008a] show that the univariate Fourier estimator is robust to market
microstructure eects. The analysis is extended to the multivariate case in [Mancino and Sanfelici, 2008b],
where both the non-synchronicity issue and the eect of (dependent) microstructure noise are
taken into account.
Most of the works concerning the comparison of the eciency of dierent variance/covariances
estimators consider only simple statistics such as bias and mean squared error (MSE). In this

DiMaD, University of Firenze, Italy, mariaelvira.mancino@dmd.uni.it

Centro Leasing Banca, Spa, Firenze, Italy, elenarapini@libero.it

Dept. of Economics, University of Parma, Italy, simona.sanfelici@unipr.it


1
Electronic copy available at: http://ssrn.com/abstract=1301031
regard, among the most recent papers [Voev and Lunde, 2007] and [Grin and Oomen, 2006]
investigate the properties of three covariance estimators, namely realized covariance, realized co-
variance plus lead- and lag-adjustments, and the covariance estimator by [Hayashi and Yoshida, 2005],
when the price observations are subject to non-synchronicity and contaminated by (i.i.d.) mi-
crostructure noise. They conclude that the ranking of the covariance estimators in terms of
eciency depends crucially on the level of microstructure noise. [Gatheral and Oomen, 2007]
compare twenty realized variance estimators using simulated data and nd that the best vari-
ance estimator is not always the one suggested by theory. The theoretical properties of the
Fourier estimator are studied by [Mancino and Sanfelici, 2008a, Mancino and Sanfelici, 2008b],
who show that the Fourier estimator of covariance is not signicantly aected by the microstruc-
ture noise.
Nevertheless this approach to the comparison of covariance estimators does not have an
economic basis and treats overestimates and underestimates of volatility of the same magnitude
as equally important. In this paper we consider the gains oered by the Fourier estimator
over other covariance measures from the perspective of an asset-allocation decision problem,
following the approach of [Fleming et al., 2001], [Engle and Colacito, 2006], [Bandi et al., 2006]
and [De Pooter at al., 2008] who study the impact of volatility timing versus unconditional
mean-variance ecient static asset allocation strategies and of selecting the appropriate sampling
frequency or choosing between dierent bias and variance reduction techniques for the realized
covariance matrices. A preliminary result we prove here concerns the positive semi-deniteness
of the estimated covariance matrix using Fourier methodology, when the Fejer kernel is used.
This property has important consequences in the asset allocation framework. An investor is
assumed to choose his/her portfolio to minimize variance subject to a required return constraints.
Investors with dierent covariance forecasts will hold dierent portfolios. Correct covariance
information will allow the investor to achieve lower portfolio volatility. Therefore we study the
forecasting power of the Fourier estimator and of other alternative realized variance measures in
the context of an important economic metric, namely the long-run utility of a conditional mean-
variance investor rebalancing his/her portfolio each period. We show that the Fourier estimator
carefully extracts information from noisy high-frequency asset price data for the purpose of
realized variance/covariance estimation and allows for non-negligible utility gains in portfolio
management.
Inspired by [Fleming et al., 2001, Bandi et al., 2006], we construct daily variance/covariance
estimates using the Fourier method and the method proposed by [Hayashi and Yoshida, 2005],
as well as estimates obtained by using conventional (in the existing literature) 1-, 5- and 10-
minute intervals and MSE-based optimally sampled continuously-compounded returns for the
realized measures. From each of these series, we derive one-day-ahead forecasts of the vari-
ance/covariance matrix. A conditional mean-variance investor can use these forecasts to opti-
mally rebalance his/her portfolio each period. We compare the investors long-run utility for
optimal portfolio weights constructed from each forecast. Our simulations show that the gains
yielded by the Fourier methodology are statistically signicant and can be economically large,
although the realized covariance with one lead-lag bias correction and suitable sampling fre-
quency can be competitive. The analysis is conducted through Monte Carlo simulations, using
the programming language Matlab.
The paper is organized as follows. In section 2 we describe the Fourier estimation methodol-
ogy and we prove the positive semi-deniteness of the Fourier covariance matrix. In section 3 we
explain the asset allocation framework and metric to evaluate the economic benet of dierent
covariance forecasts. Section 4 presents several numerical experiments to value the gains oered
2
by Fourier estimator methodology in this context. Section 5 concludes.
2 Some properties of the Fourier estimator
The Fourier method for estimating co-volatilities was proposed in [Malliavin and Mancino, 2002]
having in mind the diculties arising in the multivariate setting when applying the quadratic
covariation theorem to the true returns data, given the non-synchronicity of observed prices for
dierent assets. In fact the quadratic covariation formula is unfeasible when applied to estimate
cross-volatilities, because it requires synchronous observations which are not available in real
situations. Being based on the integration of all data, the Fourier estimator does not need any
adjustment to t non-synchronous data. We briey recall the methodology.
Assume that p(t) = (p
1
(t), . . . , p
k
(t)) are Brownian semi-martingales satisfying the following
Ito stochastic dierential equations
dp
j
(t) =
d

i=1

j
i
(t) dW
i
+b
j
(t) dt j = 1, . . . , k, (1)
where W = (W
1
, . . . , W
d
) are independent Brownian motions. The processes are observed on
a xed time window, which can be always reduced to [0, 2] by change of origin and rescaling,
and

and b

are adapted random processes satisfying hypothesis


(H) E[
_
2
0
(b
i
(t))
2
dt] < , E[
_
2
0
(
j
i
(t))
4
dt] < i = 1, . . . , d, j = 1, . . . , k.
Moreover, we assume that the observed prices are aected by microstructure noise in the
form
p
i
(t) = p
i
(t) +
i
(t) i = 1, . . . , k (2)
where the noise process is i.i.d. and the following assumptions hold:
M1. p and are independent processes, moreover (t) and (s) are independent for s = t
and E[(t)] = 0 for any t.
M2. E[
i
(t)
j
(t)] =
ij
< for any t, i, j = 1, . . . , k.
From the representation (1) we dene the volatility matrix, which in our hypothesis depends
upon time

ij
(t) =
d

r=1

i
r
(t)
j
r
(t).
The Fourier method reconstructs
,
(t) on [0, 2] using the Fourier transform of dp

(t).
The main result in [Malliavin and Mancino, 2005] relates the Fourier transform of
,
to the
Fourier transforms of the log-returns dp

. More precisely the following result is proved: compute


the Fourier transform of dp
j
for j = 1, . . . , k, dened for any integer z by
F(dp
j
)(z) =
1
2
_
2
0
e
izt
dp
j
(t)
and consider the Fourier transform of the cross-volatility function dened for any integer z by
F(
ij
)(z) :=
1
2
_
2
0
e
izt

ij
(t)dt,
3
then the following convergence in probability holds
F(
ij
)(z) = lim
N
2
2N + 1

|s|N
F(dp
i
)(s)F(dp
j
)(z s).
As a particular case (by choosing z = 0) we can compute the integrated covariance, given
the log-returns of stocks, as the following limit in probability
_
]0,2[

ij
(t) dt = lim
N
(2)
2
2N + 1

|s|N
F(dp
i
)(s)F(dp
j
)(s). (3)
From this convergence result, we can derive a suitable estimator for the integrated co-
variance matrix. We assume that the price process for asset j (j = 1, . . . , k) is observed at
high-frequency intra-daily times {t
j
l
, l = 1, . . . , n
j
}, which may be dierent on each daily trading
period normalized to length 2. Set
F(dp
j
n
j
)(s) :=
1
2
n
j
1

l=1
exp(ist
j
l
)
I
j
l
(p
j
),
where
I
j
l
(p
j
) := p
j
(t
j
l+1
)p
j
(t
j
l
). The Fourier estimator of the integrated covariance
_
2
0

ij
(t)dt
is then

ij
N,n
i
,n
j
:=
(2)
2
2N + 1

|s|N
F(dp
i
n
i
)(s)F(dp
j
n
j
)(s) =
n
i
1

u=1
n
j
1

l=1
D
N
(t
i
u
t
j
l
)
I
i
u
(p
i
)
I
j
l
(p
j
), (4)
where D
N
(x) =
1
2N+1
sin[(N+
1
2
)x]
sin
x
2
is the rescaled Dirichlet kernel.
The construction of the estimator (4) can be modied by considering the Fejer summation,
therefore in the sequel we will consider the variant obtained through the Fejer kernel

ij
N,n
i
,n
j
:=
n
i
1

u=1
n
j
1

l=1
F
N
(t
i
u
t
j
l
)
I
i
u
(p
i
)
I
j
l
(p
j
) (5)
where F
N
(x) =
_
sin Nx
Nx
_
2
. This estimator has the advantage to preserve positivity of the covari-
ance matrix, as it is stated by the following
Proposition 2.1 The Fourier estimator

N
is positive semi-denite.
Proof. Using Bochner theorem (see [Malliavin, 1995] pg 255) it suces to prove that
_

0
sin
2
t
t
2
e
itx
dt 0 x R.
As the Fourier transform of
[
1
2
,
1
2
]
(t) is the function
sinx
x
,
_

0
sin
2
t
t
2
e
itx
dt =
2
_

0

[
1
2
,
1
2
]
(x t)
[
1
2
,
1
2
]
(t)dt 0 x R.
4
For the sake of completeness we now recall the denition of the other estimators of covari-
ance which will be considered in our analysis.
The realized covariance-type estimators are based on the choice of a synchronization pro-
cedure, which gives the observations times {0 =
1

2

n
2} for both assets. The
quadratic covariation-realized covariance estimator is dened by
RC
ij
:=
n1

u=1

u
(p
i
)
u
(p
j
),
where
u
(p

) = p

(
u+1
) p

(
u
). It is known that the realized covariance estimator is not
consistent under asynchronous trading, [Hayashi and Yoshida, 2005].
The realized covariance plus leads and lags estimator is dened by
RCLL
ij
:=

u
L

h=l

u+h
(p
i
)
u
(p
j
). (6)
The estimator (6) has good properties under microstructure noise contaminations of the prices,
but it is still not consistent for asynchronous observations. This is due to the fact that all
the realized covariance type estimators need a data synchronization procedure, because of the
denition of the quadratic covariation process. Nevertheless, the introduction of one lead and
one lag appears to provide a correction for the downward bias by non-synchronous trading.
The [Hayashi and Yoshida, 2005] All-Overlapping (AO) estimator is
AO
ij
n
1
,n
2
:=

l,u

I
i
l
(p
i
)
I
j
u
(p
j
)I
(I
i
l
I
j
u
=)
, (7)
where I
(P)
= 1 if proposition P is true and I
(P)
= 0 if proposition P is false. It is unbiased in
the absence of noise. From the practitioners point of view both this estimator and the Fourier
estimator are easy to implement as they do not rely on any choices of synchronization methods
and sampling schemes. However, in [Grin and Oomen, 2006, Voev and Lunde, 2007] the AO
estimator is proved to be not ecient in the presence of microstructure noise.
The theoretical properties of the Fourier estimator are studied by [Mancino and Sanfelici, 2008a,
Mancino and Sanfelici, 2008b], who show that the bias of the covariance estimator is not aected
by the presence of i.i.d. noise. The Fourier estimator of covariance under microstructure noise is
asymptotically unbiased, as in the case of the univariate Fourier estimator under microstructure
noise. Moreover, if the number of the Fourier coecients N is conveniently chosen, the mean
squared error of the Fourier estimator does not diverge and it is not signicantly aected by
the microstructure noise. In contrast, the realized covariation and the AO estimator are not
biased by i.i.d. noise; nevertheless both realized covariation and AO estimator are inconsis-
tent under i.i.d. noise, because the MSE diverges as the number of observations increases, see
[Mancino and Sanfelici, 2008b] for details on this point.
3 Forecasting and asset allocation
We use the methodology suggested by [Fleming et al., 2001] and [Bandi et al., 2006] to evaluate
the economic benet of the Fourier estimator of integrated covariance in the context of an asset
allocation strategy. Specically, we compare the utility obtained by virtue of covariance forecasts
5
based on the Fourier estimator to the utility obtained through covariance forecasts constructed
using the more familiar realized covariance and other recently proposed estimators. In the fol-
lowing, we adopt a notation which is common in the literature about portfolio management. It
will not be dicult for the reader to match it with the one in the previous section.
Let R
f
and R
t+1
be the risk-free return and the return vector on k risky assets over a day
[t, t +1], respectively. Dene
t
= E
t
[R
t+1
] and
t
= E
t
[(R
t+1

t
)(R
t+1

t
)

] the conditional
expected value and the conditional covariance matrix of R
t+1
. We consider a mean-variance
investor who solves the program
min
w
t
w

t
w
t
,
subject to
w

t
+ (1 w

t
1
k
)R
f
=
p
,
where w
t
is a k-vector of portfolio weights,
p
is a target expected return on the portfolio, and
1
k
is a k 1 vector of ones. The solution to this program is
w
t
=
(
p
R
f
)
1
t
(
t
R
f
1
k
)
(
t
R
f
1
k
)

1
t
(
t
R
f
1
k
)
. (8)
We estimate
t
using one-day-ahead forecasts

C
t
given a time series of daily covariance
estimates, obtained using the Fourier estimator, the realized covariance estimator, the realized
covariance plus leads and lags estimator and the AO estimator. The out-of-sample forecast is
based on a univariate ARMA model.
Given sensible choices of R
f
,
p
and
t
, each one-day-ahead forecast leads to the deter-
mination of a daily portfolio weight w
t
. The time series of daily portfolio weights then leads
to daily portfolio returns. In order to concentrate ourselves on volatility approximation and to
abstract from the issues that would be posed by expected stock return predictability, for all
times t we set the components of the vector
t
= E
t
[R
t+1
] equal to the sample means of the
returns on the risky assets over the forecasting horizon. Finally, we employ the investors long-
run mean-variance utility as a metric to evaluate the economic benet of alternative covariance
forecasts

C
t
, i.e.
U

=

R
p


2
1
m
m

t=1
(R
p
t+1


R
p
)
2
,
where R
p
t+1
= R
f
+ w

t
(R
t+1
R
f
1
k
) is the return on the portfolio with estimated weights w
t
,

R
p
=
1
m

m
t=1
R
p
t+1
is the sample mean of the portfolio returns across m n days, and is a
coecient of risk-aversion.
Following [Bandi et al., 2006], in order to avoid contaminations induced by noisy rst mo-
ment estimation, we simply look at the variance component of U

, namely
U =

2
1
m
m

t=1
(R
p
t+1


R
p
)
2
, (9)
see [Engle and Colacito, 2006] for further justications of this approach. The dierence between
two utility estimations, say U
A
U
B
, can be interpreted as the fee that the investor would be
willing to pay to switch from covariance forecasts based on estimator A to covariance forecasts
based on estimator B. In other words, U
A
U
B
is the utility gain that can be obtained by
investing in portfolio B, with the lowest variance for a given target return
p
.
6
4 Valuing the economic benet by simulations
In the following sections we show several numerical experiments to assess the gains oered by the
Fourier estimator over other estimators in terms of in-sample and out-of-sample properties and
from the perspective of an asset-allocation decision problem. In Section 4.1 our attention is fo-
cused mainly on covariance estimation, since in this respect eects due to both non-synchronicity
and microstructure noise become eective. Nevertheless, the results in Sections 4.2, 4.3 and 4.4
can be fully justied only by considering the properties of the dierent estimators for both the
variance and the covariance measures.
Following a large literature, we simulate discrete data from the continuous time bivariate
Heston model
dp
1
(t) = (
1

2
1
(t)/2)dt +
1
(t)dW
1
dp
2
(t) = (
2

2
2
(t)/2)dt +
2
(t)dW
2
d
2
1
(t) = k
1
(
1

2
1
(t))dt +
1

1
(t)dW
3
,
d
2
2
(t) = k
2
(
2

2
2
(t))dt +
2

2
(t)dW
4
,
where corr(W
1
, W
2
) = 0.35, corr(W
1
, W
3
) = 0.5 and corr(W
2
, W
4
) = 0.55. The other pa-
rameters of the model are as in [Zhang et al., 2005]:
1
= 0.05,
2
= 0.055, k
1
= 5, k
2
= 5.5,

1
= 0.05,
2
= 0.045,
1
= 0.5,
2
= 0.5. The volatility parameters satisfy the Fellers condition
2k
2
which makes the zero boundary unattainable by the volatility process. Moreover, we as-
sume that the additive logarithmic noises
1
(t),
2
(t) are i.i.d. Gaussian, contemporaneously cor-
related and independent from p. The correlation is set to 0.5 and we assume (E[
2
])
1/2
= 0.002,
i.e. the standard deviation of the noise is 0.2% of the value of the asset price. From the simu-
lated data, integrated covariance estimates can be compared to the value of the true covariance
quantities.
We generate (through simple Euler Monte Carlo discretization) high frequency evenly sam-
pled ecient and observed returns by simulating second-by-second return and variance paths
over a daily trading period of h = 6 hours, for a total of 21600 observation per day. In order to
simulate high frequency unevenly sampled data, we extract the observation times in such a way
that the durations between observations are drawn from an exponential distribution with means

1
= 6 sec and
2
= 8 sec for the two assets respectively. Therefore, on each trading day the pro-
cesses are observed at a dierent discrete unevenly spaced grid {0 = t
l
1
t
l
2
t
l
n
l
2}
for any l = 1, 2.
For the realized covariance type estimators, we generate equally-spaced continuously-compounded
returns using the previous tick method. We consider 1, 5 and 10-min sampling intervals or opti-
mally sampled realized covariances. [Bandi et al., 2006] provide an approximate formula for opti-
mal sampling, which holds for uniform synchronous data. Given our general data setting, the op-
timal sampling frequency can be obtained by direct minimization of the true mean squared error.
In order to preserve the positive deniteness of the covariance matrices, we use a unique sampling
frequency for realized variances and covariances, given by the maximum among the three optimal
frequencies. For the Fourier and AO estimators, we employ all the available data set. In imple-
menting the Fourier estimator

12
N,n
1
,n
2
, the smallest wavelength that can be evaluated in order
to avoid aliasing eects is twice the smallest distance between two consecutive prices (Nyquist
frequency). Nevertheless, as pointed out in the univariate case by [Mancino and Sanfelici, 2008a]
and conrmed in the bivariate case in [Mancino and Sanfelici, 2008b], smaller values of N may
provide better variance/covariance measures. More specically, the optimal cutting frequencies
for the various volatility measures can be obtained independently by minimizing the true MSE.
7
Although the positivity result of Proposition 2.1 is ensured only when the same N is used for
all the entries of the covariance matrix, numerical experiments show that the use of dierent
cutting frequencies N for variances and covariances still preserves positive deniteness of the
covariance matrix, both in the sample and in the forecasting horizon.
4.1 Covariance estimation and forecast
As a rst application, we perform an in-sample analysis in order to shed light on the properties
of the dierent estimators in terms of dierent statistics of the covariance estimates, such as
bias, MSE and others. More precisely, we consider the following relative error statistics
= E
_

C
12

_
2
0

12
(t)dt
_
2
0

12
(t)dt
_
, std =
_
V ar
_

C
12

_
2
0

12
(t)dt
_
2
0

12
(t)dt
__
1/2
,
which can be interpreted as relative bias and standard deviation of an estimator

C
12
for the
covariance. The Fourier and RC
opt
estimators have been optimized by choosing the cutting
frequency N of the Fourier expansion and the sampling interval on the basis of their MSE. The
results are reported in Table 1. Within each table, entries are the values of , std, MSE and bias,
using 750 Monte Carlo replications which roughly correspond to three years. Rows correspond
to the dierent estimators. The sampling interval for the realized covariance-type estimators
is indicated as a superscript. The optimal sampling frequency for RC
opt
is obtained by direct
minimization of the true MSE and corresponds to 2.33 min.
When we consider covariance estimates, the most important eect to deal with is the Epps
eect. The presence of other microstructure eects represents a minor aspect in this respect. On
the contrary, it may in some sense even compensate the eects due to non-synchronicity, as we
can see from the smaller MSE of 1-minute realized covariance estimator with respect to 5-minute
estimator. We remark that the corresponding 1-minute estimator for variances is more aected
by the presence of noise, since it is not compensated by non-synchronicity. As any estimator
based on interpolated prices, the realized covariance-type estimators suer from the Epps eect
when trading is non-synchronous, but the lead-lag correction reduces such an eect, at least in
terms of bias to the disadvantage of a slightly larger MSE. Note that the lead/lag correction
contrasts the Epps eect, thus producing occasionally positive biases. On the other hand, the
presence of noise strongly aects the AO estimator. This is due to the Poisson trading scheme
with correlated noise. In fact, the AO remains unbiased under independent noise whenever
the probability of trades occurring at the same time is zero, which is not the case for Poisson
arrivals. The Fourier estimator provides good covariance measures, both in terms of bias and
MSE. Therefore, we can conclude that contrary to the AO estimator the Fourier covariance
estimator is not much aected by the presence of noise, so that it becomes a very interesting
alternative especially when microstructure eects are particularly relevant in the available data.
Before turning to asset allocation, we evaluate the forecasting power of the dierent esti-
mators. In the tradition of [Mincer and Zarnowitz, 1969], we regress the real daily integrated
covariance over the forecasting period on one-step-ahead forecasts obtained by means of each co-
variance measure. More precisely, following [Andersen and Bollerslev, 1998], we consider a larger
sample path of 1000 days and we split it into two parts: the rst one containing 20% of total
estimates is used as a burn-in period to t a univariate AR(1) model for the estimated covari-
ance time series and then the tted model is used to forecast integrated covariance on the next
8
Method std MSE bias
RC
1min
-0.07805275472999 0.22595032210904 0.00000085621669 -0.00032576060834
RC
5min
-0.01130593672881 0.39847063473932 0.00000280603760 -0.00001894057172
RC
10min
0.01355081842814 0.54479459409222 0.00000571706029 0.00008501424105
RCLL
1min
0.01338414896851 0.32281801149180 0.00000173043743 0.00009896376793
RCLL
5min
-0.00869870282055 0.62833287597643 0.00000696972045 0.00003089840943
RCLL
10min
-0.02385652069431 0.89431477251656 0.00001491460726 -0.00003947110061
RC
opt
-0.02732364713452 0.29983219944312 0.00000150028381 -0.00008496777460
AO 0.56243149971616 0.35295341500361 0.00000366031965 0.00176446312515
Fourier -0.06518255776964 0.17089219158470 0.00000049500437 -0.00026715723226
Table 1: Relative and absolute error statistics for the in-sample covariance estimates for dierent
estimators.
day. The choice of the AR(1) model comes from [At-Sahalia and Mancini, 2007], who consider
the univariate Heston data generating process. The total number of out-of-sample forecasts m
is equal to 800. Each time a new forecast is performed, the corresponding actual covariance
measure is moved from the forecasting horizon to the rst sample and the AR(1) parameters
are re-estimated in real time. For each time series of covariance forecasts, we project the real
daily integrated covariance on day [t, t +1] on a constant and the corresponding one-step-ahead
forecast

C
t
_
t+1
t

12
(s) ds =
0
+
1

C
t
+error
t
,
where t = 1, 2, . . . , m. Alternatively, we can regress simultaneously the real daily integrated
covariance over the forecasting period on various one-step-ahead forecasts obtained by means of
dierent covariance measures. The regression now takes the form
_
t+1
t

12
(s) ds =
0
+

C
t
+

C
t
+error
t
,
where

C and

C are dierent covariance forecasts on day [t, t +1]. The R
2
from these regressions
provides a direct assessment of the variability in the integrated covariance that is explained
by the particular estimates in the regressions. The R
2
can therefore be interpreted as a simple
gauge of the degree of predictability in the volatility process and hence of the potential economic
signicance of the volatility forecasts.
The results are reported in Tables 2, 3 and 4, using a Newey-West covariance matrix. We
remark that in this simulation the Fourier estimator is not optimal in the MSE sense, but
we set N
1
= 155, N
2
= 123, N = 271 from the previous experiment. When we consider a
single regressor, the R
2
is the highest for the Fourier estimator while RCLL
10min
explains less
than ve percent of the time series variability. Moreover, for the Fourier estimator we can not
reject the hypothesis that
0
= 0 and
1
= 1 using the corresponding t tests. In contrast, we
reject the hypothesis that
0
= 0 and
1
= 1 for all the other estimators except RC
5min
and
RCLL
1min
. When we include alternative forecasts besides Fourier estimator in the regression,
the R
2
improves very little relative to the R
2
based solely on Fourier. Moreover, the coecient
estimates for
FE
are generally close to unity, while for the other estimators are near zero except

5min
for the realized covariance which diers signicantly from zero at the 5% level. Therefore,
we can conclude that the higher accuracy and lower variability of Fourier covariance estimates
9
Method R
2
F p
0

1
Fourier 0.210362 212.589597 0.000000 0.000251 0.996739
(0.000251) (0.068361)
RC
1min
0.186222 182.611691 0.000000 0.000004 1.091056
(0.000288) (0.080739)
RC
2min
0.159487 151.420599 0.000000 0.000646 0.872224
(0.000265) (0.070882)
RC
5min
0.158765 150.605610 0.000000 0.000653 0.839401
(0.000265) (0.068399)
RC
10min
0.107694 96.311707 0.000000 0.000654 0.842651
(0.000328) (0.085863)
RCLL
1min
0.155834 147.311660 0.000000 0.000824 0.781413
(0.000254) (0.064382)
RCLL
5min
0.067863 58.096974 0.000000 0.000705 0.811731
(0.000413) (0.106497)
RCLL
10min
0.041545 34.590271 0.000000 0.002408 0.364288
(0.000248) (0.061940)
AO 0.186796 183.303989 0.000000 -0.001162 0.888731
(0.000373) (0.065642)
Table 2: Regression of real integrated covariance on each covariance forecast over the forecasting
horizon. Standard deviations are listed in parenthesis.
translate into superior forecasts of future covariances and this explains the superior performance
of the Fourier forecasts.
4.2 Dynamic portfolio choice and economic gains
In this section, we consider the benet of using the Fourier estimator from the perspective of
the asset-allocation problem of Section 3.
Given any time series of daily variance/covariance estimates, we split our sample of 750
days into two parts: the rst one containing 30% of total estimates is used as a burn-in
period, while the second one is saved for out-of-sample purposes. The out-of-sample forecast is
based on univariate ARMA models. More precisely, following [At-Sahalia and Mancini, 2007],
the estimated series of 225 in-sample covariance matrices is used to t univariate AR(1) models
Method R
2
F p
0

FE

1min

5min

10min
RC 0.218 55.438 0.000 0.000064 0.762475 0.052218 0.322030 -0.098934
Std (0.000338) (0.162943) (0.199279) (0.143406) (0.145376)
T-statistics (0.188921) (4.679382) (0.262034) (2.245592) (-0.680538)
RCLL 0.212 53.448 0.000 0.000319 0.895311 0.147186 -0.052503 -0.017447
Std (0.000405) (0.119065) (0.118386) (0.177799) (0.089995)
T-statistics (0.788629) (7.519509) (1.243270) (-0.295294) (-0.193873)
Table 3: Regression of real integrated covariance on Fourier (FE) and RC/RCLL -type estimators
over the forecasting horizon. The rst panel refers to the realized covariance estimator with 1,
5, 10-min sampling, the second one to its lead/lag bias correction.
10
Method R
2
F p
0

FE

2min

AO
Others 0.2110 70.962 0.000 -0.000025 0.852102 0.030607 0.121691
Std (0.000434) (0.194413) (0.136529) (0.170537)
T-statistics (-0.056499) (4.382940) (0.224183) (0.713576)
Table 4: Regression of real integrated covariance on Fourier (FE), RC
2min
and AO estimators
over the forecasting horizon.
for each variance/covariance estimates separately. The total number of out-of-sample forecasts
m for each series is equal to 525. Each time a new forecast is performed, the corresponding
actual variance/covariance measure is moved from the forecasting horizon to the rst sample
and the AR(1) parameters are re-estimated in real time. Given sensible choices of R
f
,
p
and
t
,
each one-day-ahead variance/covariance forecast leads to the determination of a daily portfolio
weight w
t
. The time series of daily portfolio weights then leads to daily portfolio returns and
utility estimation.
We implement the criterion in (9) by setting R
f
equal to 0.03 (converted to daily values by
dividing by 250) and considering three targets
p
, namely 0.09, 0.12, 0.15. In order to concentrate
on volatility timing and abstract from issues related to expected stock return predictability, for
all times t we set the components of the vector
t
= E
t
[R
t+1
] equal to the sample means of the
returns on the risky assets over the forecasting horizon. For all times t, the conditional covariance
matrix is computed as an out-of-sample forecast based on the dierent variance/covariance
estimates.
We interpret the dierence U

C
U
Fourier
between the average utility computed on the basis
of the Fourier estimator and that based on alternative estimators

C, as the fee that the investor
would be willing to pay to switch from covariance forecasts based on estimator

C to covariance
forecasts based on the Fourier estimator. Table 5 contains the results for three levels of risk-
aversion and three target expected returns. Due to the presence of microstructure noise eects
which spoils the sum of squared high-frequency intra-day returns, besides the All-overlapping
estimator we consider the AO + RCLL
1min
estimator which is based on the AO estimator for
the covariances and on the 1-minute RCLL estimator for the variances. When the target is
0.09, the investor would pay between 0.52% (when = 2) of his portfolio return and 2.59%
(when = 10) per year to use the Fourier estimator versus the RC
1min
estimator. When the
target is 0.12, the investor would pay between 0.92% (when = 2) of his portfolio return
and 4.60% (when = 10). Finally, when the target is 0.15, the investor would pay between
1.44% (when = 2) of his portfolio return and 7.19% (when = 10). The same investor
would pay marginally less to abandon RC
5min
, according to the better in-sample properties
of this estimator for the whole covariance matrix which translate into more precise forecasts.
The remaining part of the table can be read similarly. Strikingly, the utility gain of the Fourier
estimator over the AO is very large, but this is due to the presence of microstructure eects. Even
when considering the AO + RCLL
1min
estimator, the Fourier estimator is superior, while only
a very small utility loss is encountered when considering the RCLL
1min
estimator. Notice that
the optimally sampled realized covariance estimator cannot achieve the same performance. In
particular, this evidence partially contradicts the conclusions of [De Pooter at al., 2008] about
the greater eects obtainable by a careful choice of the sampling interval rather than by bias
correction procedures.
11
Method
p
= 0.09
p
= 0.12
p
= 0.15
2 7 10 2 7 10 2 7 10
RC
1min
0.52 1.81 2.59 0.92 3.22 4.60 1.44 5.03 7.19
RC
5min
0.28 0.97 1.38 0.49 1.72 2.46 0.77 2.70 3.85
RC
10min
0.62 2.18 3.12 1.11 3.88 5.55 1.73 6.07 8.67
RCLL
1min
-0.30 -1.07 -1.52 -0.54 -1.90 -2.71 -0.85 -2.97 -4.24
RCLL
5min
1.50 5.26 7.52 2.68 9.37 13.38 4.18 14.64 20.91
RCLL
10min
1.76 6.14 8.78 3.12 10.93 15.61 4.88 17.08 24.40
RC
opt
0.004 0.01 0.02 0.007 0.02 0.04 0.01 0.04 0.06
AO 1.60 5.60 7.99 2.84 9.96 14.22 4.45 15.56 22.23
AO +RCLL
1min
0.38 1.33 1.89 0.67 2.36 3.37 1.05 3.69 5.26
Table 5: Annualized fees U

C
U
Fourier
that a mean-variance investor would be willing to pay
to switch from

C to Fourier estimates.
4.3 The statistical signicance of the economic gains
One way to assess the statistical signicance of the economic gains resulting from Table 5 is to
perform the following joint statistical test. For any target
p
and any estimator, one can dene
alternative covariance forecasts

C
t
and portfolio returns R
p(

C)
t+1
. Dene
a

C
t+1
= (R
p(Fourier)
t+1


R
p(Fourier)
)
2
(R
p(

C)
t+1


R
p(

C)
)
2
.
Assessing the statistical signicance of the economic gains of the Fourier estimate over alternative
forecasts can be conducted by testing whether the mean of a

C
t+1
is larger than (or equal to) zero
against the alternative that the mean is smaller than zero.
Following [Bandi et al., 2006], for any target return d = 0.09, 0.12, 0.15, we dene the vector
A
d
t+1
=
_
a

C
1
t+1
, a

C
2
t+1
, a

C
3
t+1
_

,
where the triple of estimators (

C
1
,

C
2
,

C
3
) is given by (RC
1min
, RC
5min
, RC
10min
), (RCLL
1min
,
RCLL
5min
, RCLL
10min
) and (RC
opt
, AO, AO+RCLL
1min
) respectively. We write the regression
model
A
d
t+1
=
d
1
3
+
t+1
,
where
d
is a scalar parameter. We perform the one-sided test H
0
:
d
0 against H
A
:
d
< 0.
The parameter
d
is estimated by GMM using a Bartlett HAC covariance matrix. The t-statistics
of all the tests imply rejection of the null and hence statistical signicance of the economic gains
at 5% level.
4.4 A small-sample Monte Carlo experiment
Another way to asses the superiority of the Fourier estimator over the others is based on the
work of [Diebold and Mariano, 1995] and consists in examining each a

C
t
time series separately in
a Monte Carlo experiment. By regressing on a constant, the null hypothesis is simply a test that
the mean of a

C
t
is zero. Therefore, a negative number is evidence in favor of better performance of
the Fourier estimator over

C. A similar approach is used also by [Engle and Colacito, 2006]. We
12
RC
1min
RC
5min
RC
10min
86 69 68
RCLL
1min
RCLL
5min
RCLL
10min
54 76 75
RC
2min
AO AO +RCLL
1min
76 99 86
Table 6: Number of times (out of 100 Monte Carlo trials) that the mean of a

C
t
has a signi-
cant negative value, i.e. the asset allocation based on the Fourier estimator has a statistically
signicant benet over the others.
explore the signicance of the proposed methods on a sample of 1000 days, with m = 800 out-
of-sample forecasts, and simulate a total of 100 samples. We allocate assets according to (8) and
run the one-sided Diebold-Mariano test in each Monte Carlo trial. In Table 6 we list the times
that the asset allocation based on the Fourier estimator has a statistically signicant benet over
the others at a 5% signicance level. We remark that in this automatic Monte Carlo experiment,
the Fourier and RC estimators have not been optimized with respect to MSE. On the contrary,
we arbitrarily x the sampling period for RC
opt
at 2 min and N
1
= 155, N
2
= 123, N = 271 for
the Fourier estimator for the two variances and the covariance respectively. The table reveals a
superiority of the Fourier procedure over all the other estimators, with a percentage of success
between 54% and 99%.
5 Conclusions
We have analyzed the gains oered by the Fourier estimator from the perspective of an asset-
allocation decision problem. The comparison is extended to realized covariance-type estimators,
to lead-lag bias corrections and to the All-Overlapping estimator.
We show that the Fourier estimator carefully extracts information from noisy high-frequency
asset price data and allows for non-negligible utility gains in portfolio management. Specically,
our simulations show that the gains yielded by the Fourier methodology are statistically signif-
icant and can be economically large, while only the realized covariance with one lead-lag bias
correction and suitable sampling frequency can be competitive.
Analyzing the in-sample and out-of-sample properties of dierent covariance measures,
we nd that the Fourier estimator provides more precise variance/covariance estimates which
translate into more precise forecasts.
References
[At-Sahalia and Mancini, 2007] At-Sahalia, Y. and Mancini, L. (2007) Out of sample forecasts
of quadratic variations. Working paper.
[Andersen and Bollerslev, 1998] Andersen, T. and Bollerslev, T. (1998) Answering the skeptics:
yes, standard volatility models do provide accurate forecasts. International Economic Review,
39/4, 885905.
[Bandi et al., 2006] Bandi, F.M., Russel, J.R. and Y. Zhu (2006) Using high-frequency data in
dynamic portfolio choice. Econometric Reviews, forthcoming .
13
[Barndor-Nielsen and al., 2008] Barndor-Nielsen, O.E., Hansen, P.R., Lunde, A. and Shep-
hard, N. (2008) Multivariate Realised kernels: consistent positive semi-denite estimators of
the covariation of equity prices with noise and non-synchronous trading. Working paper.
[De Pooter at al., 2008] De Pooter, M., Martens, M. and Dick van Dijk. (2008) Predicting the
daily covariance matrix for S&P100 stocks using intraday data: but which frequency to use?
Econometric Reviews, 27, 199229.
[Diebold and Mariano, 1995] Diebold, F.X. and Mariano, R.S. (1995) Comparing predictive
accuracy. Journal of Business & Economic Statistics, 13, 253263.
[Engle and Colacito, 2006] Engle, R. and Colacito, R. (2006) Testing and valuing dynamic
correlations for asset allocation. Journal of Business & Economic Statistics, 24(2), 238253.
[Epps, 1979] Epps, T. (1979) Comovements in stock prices in the very short run. Journal of
the American Statistical Association, 74, 291298.
[Fleming et al., 2001] Flaming, J., Kirby, C. and Ostdiek, B. (2001) The economic value of
volatility timing. The Journal of Finance, LVI, 1, 329352.
[Gatheral and Oomen, 2007] Gatheral, J. and Oomen, R.C.A. (2007) Zero-intelligence realized
variance estimation. Working Paper.
[Grin and Oomen, 2006] Grin, J.E. and Oomen, R.C.A. (2006) Covariance Measurement in
the Presence of Non-Synchronous Trading and Market Microstructure Noise. Working Paper.
[Harris et al., 1995] Harris, F.H deB., McInish, T.H., Shoesmith, G.L. and Wood, R.A. (1995)
Cointegration, error correction, and price discovery on informationally linked security markets.
Journal of Financial and Quantitative Analysis, 30, 563579.
[Hayashi and Yoshida, 2005] Hayashi, T. and Yoshida, N. (2005) On covariance estimation of
nonsynchronously observed diusion processes. Bernoulli, 11(2), 359379.
[Large, 2007] Large, J. (2007) Accounting for the Epps eect: Realized covariation, cointegration
and common factors . Working paper.
[Malliavin, 1995] Malliavin, P. (1995). Integration and Probability, Springer Verlag.
[Malliavin and Mancino, 2002] Malliavin, P. and Mancino, M.E. (2002). Fourier series method
for measurement of multivariate volatilities. Finance and Stochastics, 4, 4961.
[Malliavin and Mancino, 2005] Malliavin, P. and Mancino, M.E. (2005). A Fourier transform
method for nonparametric estimation of multivariate volatility. Annals of Statistics, forth-
coming.
[Mancino and Sanfelici, 2008a] Mancino, M.E. and Sanfelici, S. (2008) Robustness of Fourier
Estimator of Integrated Volatility in the Presence of Microstructure Noise. Computational
Statistics and Data Analysis, 52(6), 29662989.
[Mancino and Sanfelici, 2008b] Mancino, M.E. and Sanfelici, S. (2008) Estimating covariance
via Fourier method in the presence of asynchronous trading and microstructure noise. Sub-
mitted to J. Financial Econometrics.
14
[Martens, 2004] Martens, M. (2004). Estimating Unbiased and Precise Realized Covariances.
EFA 2004 Maastricht Meetings Paper No. 4299.
[Mincer and Zarnowitz, 1969] Mincer, J. and Zarnowitz, V. (1969). The evaluation of economic
forecasts. In: Mincer J. (Ed.) Economic forecasts and expectations. National Bureau of Eco-
nomic Research, New York.
[Voev and Lunde, 2007] Voev, V. and Lunde, A. (2007) Integrated Covariance Estimation Using
High-Frequency Data in the Presence of Noise. Journal of Financial Econometrics, 5.
[Zhang et al., 2005] Zhang, L., Mykland, P. 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.
15