Вы находитесь на странице: 1из 18

Fractional Cointegration

Willa W. Chen and Cliord M. Hurvich

Abstract We describe a variety of seimparametric models and estimators


for fractional cointegration. All of the estimators we consider are based on
the discrete Fourier transform of the data. This includes the ordinary least
squares estimator as a special case. We make a distinction between Type I and
Type II models, which dier from each other in terms of assumptions about
initialization, and which lead to dierent functional limit laws for the partial
sum processes. We compare the estimators in terms of rate of convergence.
We briey discuss the problems of testing for cointegration and determining
the cointegrating rank. We also discuss relevant modeling issues, such as the
local parametrization of the phase function.

1 Introduction

A collection of two or more time series, observed at equally-spaced time


points, is fractionally cointegrated if there exists a non-null contemporane-
ous linear combination of the series, based on deterministic time-invariant
weights, such that the linear combination is less persistent than any of the
individual series, where persistence is measured in terms of the memory pa-
rameter, assumed here to be the same for all of the series. (In Section 3, we
survey various generalizations to the case of unequal memory parameters of
the observed series, which often arises in practice.) The most thoroughly-
studied case to date is standard cointegration, in which the memory param-

Willa W. Chen
Department of Statistics, Texas A&M University, College Station, TX 77843, USA, e-mail:
wchen@stat.tamu.edu
Cliord M. Hurvich
New York University, 44 W. 4th Street, New York, NY 10012, USA, e-mail:
churvich@stern.nyu.edu

T.G. Anderson et al., Handbook of Financial Time Series, 709


DOI: 10.1007/978-3-540-71297-8_31, Springer-Verlag Berlin Heidelberg 2009
710 W.W. Chen and C.M. Hurvich

eter is reduced from 1 to 0 by the linear combination. In general, there is


no integer constraint, and no assumption of pre-knowledge, on the memory
parameters of the original series or of the linear combination.
The seminal paper of Engle and Granger (1987) allowed for the fractional
case, though fractional and standard cointegration have largely been devel-
oped separately in the literature. In this chapter, we will present a selective
survey of the literature on representation, estimation and testing in frac-
tional cointegration. As this literature is growing rapidly, we do not attempt
to summarize it in its entirety, but rather we focus selectively on certain
themes.
We will start by describing two dierent denitions of the order of integra-
tion of a time series, called the Type I and Type II denitions, which lead to
dierent limit laws for partial sums. Both have been used in the fractional-
cointegration literature. Next, we describe a variety of semiparametric models
for fractional cointegration. Next, we consider estimation of the cointegrating
parameters i.e., the linear combinations that reduce the memory parameter.
The various estimators can be classied according to the type of cointegra-
tion model assumed, the use of tapering and dierencing, assumptions on
the phase angle, the use of a xed or increasing bandwidth, and the nature
of the estimator: either a direct estimator of the cointegrating parameters
alone, or a joint estimator of the cointegrating parameters together with the
memory and other parameters. Finally, we discuss estimation of the cointe-
grating rank, and describe some results on testing for cointegration, focusing
on residual-based tests and joint tests.
It should be emphasized that in this chapter, as in general, there is a
link between the strength of assumptions and the strength of the theoretical
results that can be established under these assumptions. Remembering this
fact helps to put the existing results into an appropriate context, and may
serve to prevent invidious comparisons in subsequent work.

2 Type I and Type II Denitions of I(d)

2.1 Univariate series

There are a variety of ways in which long memory can be dened. We start
by considering univariate series, and then move on to consider vector series.
Most denitions of long memory for a stationary series involve an asymp-
totic power law for some quantity, e.g., the spectral density near zero fre-
quency, the autocorrelation at large lags, the variance of partial sums with
increasing aggregation, the moving average coecients in an M A() repre-
sentation of the process. If the process {xt } is weakly stationary and invert-
ible with spectral density f , autocovariance sequence {cr }
r= , and moving
Fractional Cointegration 711

average representation xt = k=0 ak tk where {t } are white noise, the
long-memory assumption may be written in terms of the memory parameter
d (1/2, 1/2) {0} in the following generally non-equivalent ways:

f () K1 2d as 0+ ,

cr K2 r2d1 as r ,

n
var xt K3 n2d+1 as n ,
t=1

ak K4 k d1 as k ,
where K1 > 0, K2 , K3 > 0 and K4 are constants. Connections between these
properties are given in Robinson (1995b) and Taqqu (2003). All of these
properties hold for fractional ARIMA models (Adenstedt 1974, Granger and
Joyeux 1980, Hosking 1981), with d (1/2, 1/2) {0}, though semipara-
metric specications of long memory typically assume one of these properties
at the potential expense of the others. We will say that a weakly-stationary
series has short memory if its spectral density satises the property given
above with d = 0, i.e., the spectral density tends to a positive constant as
the frequency tends to zero.
We will say that a process is integrated of order d, denoted by I(d), if it
has memory parameter d. Thus, for a stationary and invertible series, I(d)
can be dened by any of the asymptotic power-law relations given above.
Since in fractional cointegration the original series may be nonstationary, it
is essential to be able to dene I(d) even for d > 1/2. It is also convenient to
be able to dene I(d) in the non-invertible case, d 1/2.
Hurvich and Ray (1995) worked with the spectral denition of long mem-
ory as given above, which remains valid in the non-invertible case. For all
d < 1/2, they dene a weakly stationary process to be I(d) if its spec-
tral density satises f () K1 2d as 0+ , where K1 > 0. In the
non-stationary case d > 1/2, they dene a process to be I(d) if there ex-
ists a positive integer k such that the kth ordinary dierence of the series
is I(d k), where d k (1/2, 1/2). For example, if {xt } is a random
walk, xt = xt1 + t where {t } is white noise, the ordinary rst dierence
is xt xt1 = t , which is I(0), so {xt } is I(1). In a semiparametric context,
the denition (henceforth referred to as the Type I denition) has been used
in papers on estimation of the memory parameter by Velasco (1999a, 1999b),
and Hurvich and Chen (2000) and in papers on fractional cointegration by
Chen and Hurvich (2003a, 2003b, 2006), and Velasco (2003).
Marinucci and Robinson (1999) present an alternative denition of I(d)
based on truncation, which we will call the Type II denition. Given the
choice of a time origin, t = 0 (which plays an important role here), for any
d > 1/2, the (Type II) I(d) series is represented for t 1 as
712 W.W. Chen and C.M. Hurvich


t1
xt = k tk , (1)
k=0

where {t } is a weakly stationary, zero-mean short-memory series, and for


d = 0

k = (k + d)/[ (k + 1) (d)] Ck d1 as k (C > 0),

while for d = 0, we dene 0 = 1, k = 0 (k = 0). A more general formulation


for k , with the same limiting behavior as above, can also be considered,
subject to suitable regularity conditions. Even for nonzero d (1/2, 1/2)
the series {xt } is not stationary, but it is asymptotically stationary in the
sense that there exists a function g such that for t > u, cov(xt , xu ) g(t u)
if u/(t u) 0.
A representation equivalent to (1) is

xt = (1 B)d t (t 1), (2)

where B is the backshift operator, and t = t for t 1, t = 0 for t 0.


In (2), (1 B)d is dened for all d through the formal expansion (see, e.g.,
Brockwell and Davis 1991, p. 522),


d
(1 B) = k B k .
k=0

It follows that for d > 1/2, the Type-II I(d) process, which is neither station-
ary nor asymptotically stationary in this case, can be written as the partial
sum
xt = u1 + u2 + + ut t 1,
where

t1
ut = (1 B)1d t = ak tk ,
k=0

with ak = (k + d 1)/[ (k + 1) (d 1)], so that {ut } is Type-II I(d 1). In


this case (d > 1/2), Marinucci and Robinson (2000), following earlier work
of Akonom and Gourieroux (1987) and Silveira (1991), derived the weak
limit of suitably normalized partial sums of {ut } assuming that {t } has a
linear representation with respect to an iid or absolutely regular sequence (see
Pham and Tran 1985). The limiting process, which Marinucci and Robinson
(2000) called Type II fractional Brownian motion, also known as Riemann-
Liouville fractional Brownian motion, WH () (with H = d + 1/2), is dened
for any H > 0, whereas the so-called Type I fractional Brownian motion
BH () considered by Mandelbrot and Van Ness (1968) and others, is only
dened for 0 < H < 1, and even in this case the two processes are not
equivalent.
Fractional Cointegration 713

Limit theory has also been developed for statistics based on I(d) processes
under the Type I denition. For example, Sowell (1990) obtained the limiting
distribution of the OLS estimator in a regression of xt on xt1 for a process
{xt } such that (1 B)d xt = t , where t is iid with zero mean, with d
(0.5, 1.5). The limit distribution is a functional of Type-I fractional Brownian
motion. (See the remark on Sowells results in Marinucci and Robinson 1999).
Implications of the Type I and Type II denitions of I(d) on properties
of the discrete Fourier transform (DFT), semiparametric log-periodogram re-
gression estimates of the memory parameter, and other statistics of interest
were studied by Velasco (2007). There are noticeable dierences in the prop-
erties of the DFTs even for d < 1/2, but these do not have any important
impact on the properties of the memory parameter estimates unless d 1/2.
It is notable in this regard that, for the region d [1/2, 3/4), assuming a
Gaussian process, the (suitably normalized) log-periodogram regression esti-
mate of d has been shown to be asymptotically normal in the Type-I case
(Velasco 1999a), but no such result has been established as of yet in the
Type-II case, owing perhaps to the much stronger asymptotic correlations of
the normalized DFTs under Type-II compared to Type I I(d) when d 1/2.

2.2 Multivariate series

Here, we present a Type I denition of I(d) for a q-dimensional vector process


{xt }. We focus on the stationary case, in which the memory parameters of
the entries of the vector are all less than 1/2. Extension to the nonstationary
case is then done similarly as for the univariate Type I I(d) case.
Recall that for any weakly stationary real-valued q-vector process {xt }
with spectral density f , we have f () = f () for all [, ], and the
lag-h autocovariance matrix is


(h) = E[xt+h xt ] = f ()eih d ,

so that (h)  (h), where the superscript prime denotes transposition.


A weakly stationary q-vector process {xt } is (Type I) I (d1 , . . . , dq ) if its
spectral density matrix f satises

f () G , 0+ , (3)

where G is a nonnegative denite real symmetric matrix with nonzero diag-


onal entries not depending on ,
 
d d
= diag ei1 || 1 , . . . , eiq || q
714 W.W. Chen and C.M. Hurvich

with dk < 1/2, for k = 1, . . . , q, is the conjugate transpose of , 1 = 0,


and the phase angles k (k = 2, . . . , q) are real. The standardization 1 = 0
is made for the sake of identiability, since f in (3) remains unchanged if
is multiplied by a complex number with modulus 1. The series {xt } may or
may not be cointegrated, but if it is then G must be singular.
Several special cases of the model have been considered in the literature.
Case 1: The case where all the phase angles are zero (1 = 2 = =
q = 0) was assumed by Christensen and Nielsen (2006). One way this could
happen would be if the spectral density f () were real for all [, ]. In
this case, we would have (h) = (h), so that the lag-h cross-covariance of
two entries of {xt } would be the same regardless of which entry leads. This
would clearly entail a loss of generality. In the case of an ARFIMA model,
one would obtain 1 = 2 = = q = 0 if and only if d1 = d2 = = dq .
Case 2: The case k = (/2)dk was considered in Robinson and Yajima
(2002), Shimotsu (2006) and Lobato (1999) (though for convenience of esti-
mation, Lobato (1999) subsequently changed to the assumption that all phase
angles are zero). The assumption k = (/2)dk is satised by the fractional
ARIMA model. In view of the identiability problem discussed above, the
assumption is equivalent to k = (/2)(dk d1 ).
Case 3: A more general case is k = (/2 k )(dk d1 ) where 2 , . . . , q
are real numbers satisfying the constraints that the resulting values of k are
in (, ). Note that these constraints depend on the memory parameters.
This case was considered by Robinson (2006) in a bivariate setting.
Chen and Hurvich (2003a,b, 2006) assumed a model that is ultimately
equivalent to (3), but they made an additional assumption, which was not
needed for their theoretical results, that the phase functions in the trans-
fer function for the linear representation of the fractionally dierenced series
are continuous (and therefore zero) at zero frequency. If this assumption is
removed, then the model becomes equivalent to (3). Such a model is more
general than those obtained in Cases 1, 2 and 3 above, since the k may be free
of, or depend nonlinearly on (d1 , . . . , dq ). Chen and Hurvich (2003a,b, 2006)
obtained the limiting asymptotic normality of the discrete Fourier transform
of a multivariate Type I I(d1 , . . . , dq ) series for a xed set of Fourier frequen-
cies, assuming that the series has a linear representation with respect to an
iid sequence.
A multivariate Type II denition for I (d1 , . . . , dq ) was given by Robinson
and Marinucci (2001, 2003), and Marinucci and Robinson (2000), also used by
Marmol and Velasco (2004) and Nielsen and Shimotsu (2007), as diag((1
B)d1 , . . . , (1 B)dq )Xt = ut I{t 1} for t = 1, 2, . . . where {Xt } is the
observed series, {ut } is a stationary short-memory series with zero mean,
and dk > 1/2, k = 1, . . . , q. Limit theory for partial sums of such processes
was developed by Marinucci and Robinson (2000).
Fractional Cointegration 715

3 Models for Fractional Cointegration

As stated in the introduction, a (q1) I(d0 , . . . , d0 ) series {xt } is cointegrated


if there exists a vector = 0, such that ut =  xt is I(du ) where du < d0 .
Any such vector is called a cointegrating vector. There may be up to q 1
linearly independent cointegrating vectors, and the corresponding contem-
poraneous linear combinations may have dierent memory parameters. The
number of linearly independent cointegrating vectors, r, with 1 r < q,
is called the cointegrating rank. It is assumed in the above denitions, as
in the work of Engle and Granger (1987), and Chen and Hurvich (2003a,
2003b, 2006), among others, that all entries of the observed series have the
same memory parameter. For a general I(d 1 , , d q ) series, the concept of
cointegration requires a more careful denition, and several dierent ones
have been provided, as described in Robinson and Yajima (2002). By any of
these denitions, a necessary condition for cointegration is that at least two
of d 1 , . . . , d q are equal. Robinson and Yajima (2002, Denition 2) start by
partitioning the entries of the observed series into blocks with equal mem-
ory parameters within each block, and unequal memory parameters across
blocks. (They also provide a data-driven procedure for constructing such a
partition, which is successful at this task with probability approaching 1).
The (q 1) series is cointegrated if for some block the entries of that block
are cointegrated in the sense dened earlier. The cointegrating rank of the
entire (q 1) series is the sum of the cointegrating ranks of the blocks. Robin-
son and Marinucci (2003) use a dierent denition, under which the (q 1)
series {xt } is cointegrated if there exists = 0 such that  xt = ut is I(du )
with du < min(d 1 , . . . , d q ).
We will present some semiparametric Type I and Type II models for frac-
tional cointegration. We start with the Bivariate Type I model yt = xt + ut ,
where the observed series {xt } and {yt } are both I(d), the unobserved series
{ut } is I(du ), and du < d, so that {(xt , yt ) } is fractionally cointegrated. We
do not require that {xt } and {ut } be mutually independent. The stationary,
positive-memory-parameter case d, du (0, 1/2) was considered by Robinson
(1994). Chen and Hurvich (2003a) assumed that the original series {Xt }, {Yt }
{Ut } are potentially nonstationary, but after p 1 ordinary dierences they
yield the stationary (but potentially noninvertible) series {xt }, {yt }, {ut }
satisfying yt = xt + ut , with d, du (p + 1/2, 1/2), p 1. For example,
if p = 1, the range would be (1/2, 1/2), i.e., the stationary invertible case.
If p = 2, the range would be (3/2, 1/2), which allows for noninvertibility
induced by (potentially unintentional) overdierencing.
Chen and Hurvich (2006) proposed the Type I fractional common com-
ponents model for the (q 1) observed series {yt } with cointegrating rank r
(1 r < q), and s cointegrating subspaces (1 s r), given by
(0) (1) (s)
yt = A0 ut + A1 ut + + As ut , (4)
716 W.W. Chen and C.M. Hurvich

where Ak (0 k s) are unknown q ak full-rank matrices with a0 =


q r and a1 + + as = r such that all columns of A0 , . . . , As are linearly
(k)
independent, {ut } k = 0, . . . , s, are unobserved (ak 1) processes with
memory parameters {dk }sk=0 with p + 1/2 < ds < < d0 < 1/2. Thus,
each entry of {yt } is I(d0 ), and the space Rq can be decomposed into a direct
sum of orthogonal cointegrating subspaces of dimension ak (k = 0, . . . , s)
such that any vector in the kth cointegrating subspace yields a linear
combination  yt which is I(dk ) with dk < d0 if k > 0. The series {yt }
is stationary, the result of p 1th dierencing of an original, potentially
nonstationary series {Yt }, with p 1. The case of standard cointegration
could be obtained, for example, by taking p = 2, s = 1, d0 = 0, d1 = 1, and
1 r q 1. Equation (4) can be written as

yt = Azt , (5)
   
(0) (s)
where zt = vec ut , . . . , ut and A = A0 As . Chen and Hurvich
(2006) make additional assumptions on the spectral density of {zt }. These
assumptions guarantee that {zt } is not cointegrated, and that the spectral
density of {yt } satises (3) with G singular, which is in turn a necessary
and sucient condition for cointegration in a stationary process such that all
components have the same memory parameter. The methodology presented
in Chen and Hurvich (2006) does not require either r or s to be known.
Robinson and Marinucci (2003) considered a Type-II model with cointe-
grating rank 1 for a non-dierenced q 1 series {Zt } which is partitioned
as {Zt } = {(Yt , Xt ) } with {Yt } 1 1 and {Xt } (q 1) 1 such that
Yt =  Xt + Ut , where is an unknown (q 1) 1 cointegration param-
eter, and {Ut } has a smaller memory parameter than the minimum memory
parameter of the entries of {Zt }. Thus, (1,  ) is a cointegrating vector for
{Zt }. The need to specify one of the entries of {Zt } as the response variable
may cause diculties when q 3 since there is no guarantee in general that
all entries of {Zt } appear in a cointegrating relationship with at least one of
the other entries. Thus if a randomly chosen component of {Zt } is labeled as
the response variable, there is no guarantee that the regression model above
will hold, and in any case regression-type estimators of the parameter will
not be invariant to this choice.

3.1 Parametric models

The models for cointegration considered above are all semiparametric, in that
the spectral density is only specied in a neighborhood of zero frequency.
Parametric models are also of interest, in which the time-series dynamics
of the series is fully determined by a nite set of xed, unknown parame-
ters, although the distribution of the innovations may not be parametrically
Fractional Cointegration 717

specied. For reasons of brevity, we will not present a detailed discussion


of parametric models for cointegration, or of the properties of estimators in
such models, but we provide some references here. Gaussian Maximum like-
lihood estimation of a cointegrated fractional ARIMA model was considered
by Dueker and Startz (1998). Maximum likelihood estimation of the coin-
tegrating parameter in a multivariate Type-I model with a known but not
necessarily Gaussian distribution was considered by Jeganathan (1999). Es-
timation of the cointegrating parameter in a bivariate nonstationary Type-
II model was considered using generalized least-squares by Robinson and
Hualde (2003). Here, the rate of convergence for the estimator of the cointe-
grating parameter is n , where the degree of cointegration (i.e. reduction in
the memory parameter) is assumed to be greater than 1/2. Properties of
ordinary least-squares estimators were considered for Type-I autoregressive
models with fractional errors by Chan and Terrin (1995). A parametric coin-
tegration model of Granger (1986) was followed up in a fractional context by
Breitung and Hassler (2002), and applied by Davidson (2002) and Dolado,
Gonzalo and Mayoral (2003).

4 Tapering

Tapering is the multiplication of an observed series by a sequence of constants


(the taper) prior to Fourier transformation, in order to reduce bias in the pe-
riodogram. A cosine bell taper was used in Hurvich and Ray (1995), and
was found to be especially helpful in both the noninvertible case (d < 1/2)
and the nonstationary case (d > 1/2). A class of tapers due to Zhurbenko
was used by Velasco (1999a, 1999b) for estimation of d in the nonstation-
ary case. Hurvich and Chen (2000) chose to dierence the data p 1 times
to remove deterministic polynomial trends and to render the series sta-
tionary, and then proposed applying the pth order taper {hp1 t }, where
ht = (1/2)[1 exp{i2(t 1/2)/n}], to handle the potential noninvertibility
of the dierenced series. They showed that this allows for Gaussian semipara-
metric estimation of d with a smaller variance ination than incurred by the
methodology of Velasco (1999b). Variance ination in estimation of d can be
removed entirely using the class of tapers of Chen (2006), and the resulting
estimator is therefore comparable to the non-tapered exact Local Whittle es-
timator of Shimotsu and Phillips (2005). In a cointegration context, tapering
was used by Chen and Hurvich (2003a, 2003b, 2006), who used dierencing
together with the tapers of Hurvich and Chen (2000), and Velasco (2003).
Since the taper of Hurvich and Chen (2000) is complex-valued, care must be
taken in using fast Fourier transform software, since hp1
t exp(i) is not the
conjugate of hp1
t exp(i).
718 W.W. Chen and C.M. Hurvich

5 Semiparametric Estimation of the Cointegrating


Vectors

We rst consider direct estimators of the cointegrating parameter, i.e., estima-


tors which can be constructed without estimating other nuisance parameters,
such as memory parameters and phase parameters. The limiting distributions
of these estimators have been shown to be non-Gaussian in most cases, which
is somewhat undesirable from the point of view of practical usefulness. We
focus here on rates of convergence. We start with the bivariate model

yt = xt + ut ,

in the Type-I stationary positive-memory case d, du (0, 1/2), du < d, as


considered by Robinson (1994). Suppose we have n observations on the re-
sponse and explanatory variables, {yt }nt=1 , {xt }nt=1 . Consider the ordinary
least squares (OLS) estimator OLS of obtained from linear regression of
{yt } on {xt }. If it is known that {ut } has zero mean, the regression can be
performed without an intercept. Robinson (1994) pointed out that if OLS
converges in probability, the probability limit must be

+ cov(xt , ut )/var(xt ).

Thus, as long as there is a nonzero contemporaneous correlation between xt


and ut , OLS will be an inconsistent estimator of .
Robinson (1994) proposed a modication of OLS that remains consistent
even if cov(xt , ut ) = 0. To motivate the modication, it is helpful to consider
the source of the inconsistency of OLS from a frequency-domain perspective.
Note that
 
cov(xt , ut )/var(xt ) = fxu ()d/ fx ()d,

where fx is the spectral density of {xt } and fxu is the cross-spectral density
of {(xt , ut ) }. By the Cauchy-Schwartz inequality, fxu () = O( (d+du ) ) as
0+ , so that fxu ()/fx () 0 as 0+ . Thus, if lies in a neighbor-
hood which shrinks to zero as the sample size n increases, the contribution
to cov(xt , ut ) from frequency is negligible compared to the contribution
to var(xt ) from frequency . This motivates the narrowband least-squares
estimator of , given by Robinson (1994) as


mn 
mn
N BLS = Re{Jx,j Jy,j }/ |Jx,j |2 ,
j=1 j=1

a frequency-domain regression of y on x, where mn is a bandwidth parameter,


mn , mn /n 0 as n ,
Fractional Cointegration 719

1 
n
Jz,j = zt exp(ij t)
2n t=1

for any series {zt }, and j = 2j/n. It then follows from Robinson (1994)
that N BLS as n , under an additional assumption on the errors
p

in the M A() representation of {xt }, and assuming that d, du (0, 1/2).


If, in violation of the assumption mn /n 0, we take mn = "n/2# (so that
N BLS is no longer in fact narrowband), then N BLS = OLS , where OLS is
computed from a regression that includes an intercept. If for this same choice
of mn the lower index of the summations in N BLS were changed from 1
to 0, then we would again have N BLS = OLS , but in this case the OLS
regression would not contain an intercept.
Robinson and Marinucci (2001) derived asymptotic properties of N BLS
(with mn , mn /n 0) and OLS for the bivariate model Yt = Xt +Ut ,
assuming that the observed series {Xt } and {Yt } are nonstationary Type-II
I(dX ) with dX 1/2, and the unobserved errors {Ut } are Type-II I(dU )
with dU < dX . They show that OLS and N BLS based on the observed non-
dierenced data {Xt }nt=1 , {Yt }nt=1 are consistent, with a rate of convergence
that depends on dX and dU , but not simply on dX dU .
Chen and Hurvich (2003a) considered a Type-I bivariate model, and pro-
posed CH , a modied version of N BLS computed from tapered, dierenced
data. The original series {Xt }, {Yt } and {Ut } are dierenced p 1 times, re-
sulting in stationary series {xt }, {yt } with memory parameter d, and {ut }
with memory parameter du , du < d, and d, du (p + 1/2, 1/2). The dier-
encing transforms the original model Yt = Xt + Ut into yt = xt + ut . The
estimator is then given by


m0 
m0
CH = T T
Re{Jx,j Jy,j }/ |Jx,j
T 2
| ,
j=1 j=1

where m0 1 is a xed bandwidth parameter,



n
T
Jz,j = hp1
t zt exp(ij t),
t=1

and
ht = (1/2)[1 exp{i2(t 1/2)/n}]. (6)
Chen and Hurvich (2003a) showed that subject to suitable regularity condi-
tions the rate of convergence of CH is always nddu (the same as ndX dU ),
in the sense that nddu (CH ) converges in distribution. This uniformity
of the rate of convergence of CH (i.e. the dependence on d du alone), as
contrasted with the nonuniformity of the rate of OLS or N BLS , is due to
the combination of tapering and the use of a xed bandwidth m0 in CH .
720 W.W. Chen and C.M. Hurvich

Though it may at rst seem surprising that CH can be consistent for
even though m0 is xed (after all, the use of a xed number of frequencies
in semiparametric estimation of the memory parameter would result in an
inconsistent estimator), the consistency of CH follows since it can be shown
that each term of

m0 
m0
CH = T T
Re{Jx,j Ju,j }/ |Jx,j
T 2
|
j=1 j=1

is Op (ndu d ), and since there are a xed number of terms.


We present here some comparisons of the rate of convergence of CH with
those of OLS and N BLS , keeping in mind that the theory for the former
estimate is based on a Type-I model, while that for the latter estimates is
based on a Type-II model. Let dX and dU represent the memory parameters
of the nondierenced observed and error series, {Xt } and {Ut }. All of our
comparisons below are based on the assumption dX 1/2 (the observed series
are nonstationary), as this is assumed in the results we will use from Robinson
and Marinucci (2001). The rates of convergence for OLS and N BLS remain
the same whether or not the zero frequency is excluded from the sums in
their denitions. Since dX 1/2, OLS is consistent. This follows from the
Cauchy-Schwartz inequality and the fact that in this case

n 
n
p
Ut2 / Xt2 0.
t=1 t=1

For a given value of the dierence of the memory parameters (i.e., the strength
of the cointegrating relationship), CH , which is based on the tapered dier-
ences of order p 1 (and is invariant to additive polynomial trends of order
p 1 in the levels) will be ndX dU -consistent. The comparison is separated
into several cases, due to the non-uniform rates of convergence for OLS
and N BLS . All three estimators converge at the same rate when dU > 0,
dU + dX > 1, and also when dU = 0, dX = 1. The estimators CH and
OLS converge at the same rate when dU = 0, dX > 1, though Robinson
and Marinucci (2001) do not present a rate of convergence for N BLS in this
case. In the remaining cases, CH has a faster rate of convergence than the
other two estimators. We present the comparisons in terms of the improve-
ment factor, given as the ratio of the rates of convergence. For example, if
another estimator is n -consistent with < dX dU then the improvement
factor for CH relative to the other estimator is ndX dU . For the case
dU > 0, dU + dX = 1, the improvement factor for CH relative to OLS is
log n, and the improvement factor for CH relative to N BLS is log mn . For
the case dU 0, dU + dX < 1, the improvement factor for CH relative to
OLS is n1dU dX , and the improvement factor for CH relative to N BLS
U dX
is m1d
n . In the latter two cases, the slower the rate of increase of mn ,
Fractional Cointegration 721

the less inferior is the performance of N BLS compared to that of CH . This
helps to justify the use of a xed bandwidth m0 in CH . It can be seen that
the spectral density and periodogram matrices at the very low frequencies
(e.g., 1 , . . . , m0 with m0 xed) play a key role in fractional cointegration.
For a multivariate series, Robinson and Marinucci (2003) considered prop-
erties of NBLS and OLS estimators of the cointegrating parameter in the
multiple regression model
Yt =  Xt + Ut
described above. The estimators are

N BLS = FXX (1, mn )1 FXY (1, mn )

and OLS given by a similar formula with (1, mn ) replaced by (1, "n/2#) or
(0, "n/2#), where

k

Fab (, k) =
Re(Ja,j Jb,j ),
j=

the superscript denotes conjugate transpose, and for any time series of
column vectors {ct }nt=1 ,

1 
n
Jc,j = ct exp(ij t).
2n t=1

Again, the bandwidth mn satises mn , mn /n 0. They rst consid-


ered the Type-I stationary case, in which OLS is inconsistent, and showed
that

N BLS i = Op ((n/mn )dU di )
i

for i = 1, . . . , q 1, where iN BLS is the ith entry of N BLS , d i is the memory
parameter of the ith entry of {Xt }, and dU is the memory parameter of
{Ut }. They next considered a Type-II model in which the observable series
are nonstationary. Here, the convergence rates (also reported in Marinucci
and Robinson 2001) for iN BLS i and iN BLS i are analogous to those
obtained for bivariate series in Robinson and Marinucci (2001).
Chen and Hurvich (2006) considered estimation in the Type-I fractional
common components model (4), (5), based on the averaged periodogram
matrix,

m0
T
Im0 = Re{Jy,jT
Jy,j },
j=1

T
where m0 is a xed positive integer, Jy,j is the (q 1) tapered DFT vector


n
T
Jy,j = hp1
t yt exp(ij t),
t=1
722 W.W. Chen and C.M. Hurvich

and the taper {ht } is given by (6). Note that Im0 , and statistics based on
it, are equivariant to permutation of the entries of {yt }. The eigenvalues
1 2 q of Im0 satisfy j = Op (n2dk ), for j Nk , where
N0 = {1, . . . , a0 } and Nk = {(a0 + + ak1 ) + 1, . . . , (a0 + + ak )}
for k = 1, . . . , s. The eigenvectors corresponding to the jth largest eigenval-
ues for j Nk converge in distribution to (random) vectors lying in the kth
cointegrating subspace, k = 0, . . . , s. Although these eigenvectors do not con-
sistently estimate xed population vectors, the estimated space nevertheless
converges to the kth cointegrating subspace in the sense that the norm of
the sine of the angle between the true and estimated cointegrating subspaces
is Op (nk ) where k is the shortest gap between the memory parameters
corresponding to the given and adjacent subspaces. Gaussian semiparametric
estimators constructed from the residuals, i.e., the contemporaneous linear
combination of yt with weights given by an eigenvector in the estimated
kth cointegrating subspace, are asymptotically normal and consistent for dk ,
where the bandwidth mn , with an upper bound that becomes more
restrictive when min = mink (k ) = min(d0 d1 , . . . , ds1 ds ) decreases.
If the short-memory component of the spectral density of {zt } is suciently
smooth, then the upper bound is determined by (m2+1 n /n2 ) log2 mn 0,
1/2
where = min(min , 2). If min > 1/2, then the estimator is mn -consistent.
1/2
If min 1/2 then the estimator is no longer mn -consistent. The restriction
on the rate of increase of mn arises because a linear combination of series with
slightly dierent memory parameters will typically have an irregular short-
memory component in its spectral density. Given an a priori lower bound
on min it is possible to use the estimates of the dk to consistently estimate
the dimensions a0 , . . . , as of the cointegrating subspaces, as well as s itself
and the cointegrating rank r = a1 + + as . This can be accomplished by
setting the group boundaries at the points where the sorted estimates of the
dk dier by a sucient amount. While the need for a lower bound on min
is unfortunate since such a quantity would rarely be known in practice, we
note that such lower bounds (assuming s = 1) arise implicitly or explicitly
in other works on semiparametric fractional cointegration. See Robinson and
Yajima (2002), Assumption D, and Velasco (2003), Theorems 2 and 4, as well
as Nielsen and Shimotsu (2007).
The estimators for the cointegrating parameter considered above are all
direct in the sense that they do not require estimation of memory param-
eters or other nuisance parameters. An alternative promising approach was
proposed by Robinson (2006), in the context of a Type I stationary bivariate
system. Under this approach, a bivariate local Whittle estimator is used to
jointly estimate the memory parameters du1 < du2 < 1/2, the parameter
(which is a cointegrating parameter if it is nonzero) and the phase pa-
rameter in the bivariate system (yt , xt ) where yt = xt + u1,t , xt = u2,t
and the spectral density of (u1,t , u2,t ) satises (3) with phase parameters
1 = 0, 2 = (/2 )(du2 du1 ). If = 0, then both {xt } and {yt } have
memory parameter du2 but the linear combination yt xt has memory pa-
Fractional Cointegration 723

rameter du1 < du2 . On the other hand, if = 0 there is no cointegration


and the memory parameters of {xt } and {yt } are unequal. Assuming su-
cient smoothness on the spectral density of (u1,t , u2,t ) , the resulting local
Whittle estimator of is asymptotically normal, and, to within logarithmic
terms, is n2/5+(du2 du1 )/5 -consistent, a faster rate than attained by any of the
semiparametric estimators considered above. Furthermore, as du2 du1 ap-
proaches 1/2, the local Whittle estimator of becomes nearly n1/2 -consistent.
The vector of standardized parameter estimates is asymptotically 4-variate
normal (the feature of asymptotic normality is not typically shared by the
direct estimators of the cointegrating parameter), and the estimate for
converges at a faster rate than the other components. The case = 0 is
allowed, and tests can be constructed for this hypothesis, which implies no
cointegration, but still entails the assumption that du1 < du2 . Although it is
semiparametric, the estimator is nevertheless sensitive to misspecication of
the phase parameter. If 2 is indeed of the form given above, but an incorrect
value for , say , is xed in the local Whittle objective function, which is
then minimized with respect to the other parameters, then if = 0 mod
(2/(du2 du1 )) the estimated parameter vector is still asymptotically nor-
mal with the same standardization as above, though the limiting covariance
matrix becomes more complicated, while if = 0 mod (2/(du2 du1 )) the
estimates of du1 and du2 are rendered inconsistent and the estimate of is
still asymptotically normal, but with the inferior rate (n/mn )du2 du1 , where
mn is the bandwidth used in the estimator. If the form of 2 were
actually other than that assumed, for example 2 = (/2 )(du2 du1 )2 ,
then all variants considered above of the local Whittle estimator may yield
inconsistent estimates of .

6 Testing for Cointegration; Determination of


Cointegrating Rank

Although this chapter has focused mainly on models for fractional cointegra-
tion and estimation of the cointegrating parameter, there remains the more
basic question of testing for the existence of fractional cointegration. There
has been some progress in this direction, which we describe briey here. A
method proposed by Robinson (2006) based on joint local Whittle estima-
tion of all parameters was described in Section 5, although the bivariate
model considered there rules out the possibility that the two series have the
same memory parameter if there is no cointegration. Marinucci and Robin-
son (2001) proposed a Hausman-type test in which the memory parameters of
the observed series are estimated by two dierent methods: (1) a multivariate
Gaussian semiparametric estimator, imposing the constraint that the mem-
ory parameters are the same, and (2) a univariate Gaussian semiparametric
estimator of a particular entry. The component of (1) for this entry would
724 W.W. Chen and C.M. Hurvich

be consistent if and only if there is no cointegration, but the estimator (2)


would be consistent in either case. Comparing the standardized dierences
between these two estimators leads to a test for the null hypothesis H0 of no
cointegration versus the alternative hypothesis of cointegration. Marmol and
Velasco (2004), who assumed a nonstationary Type-II model, use regression-
based estimators of the cointegrating/projection parameter, and exploit the
fact that under H0 the regressions will be spurious and therefore the true
projection parameter is not consistently estimated by OLS. This, together
with another estimator of the projection parameter that is consistent under
H0 , leads to a Hausman-type test of H0 , under the assumption that if there is
cointegration, the cointegrating error is also nonstationary. This assumption
would be dicult to maintain if the observed series were volatilities, which
are usually considered in the literature to be stationary. Chen and Hurvich
(2006) use a statistic based on the dierence between the largest and smallest
residual-based Gaussian semiparametric estimator of the memory parameter
to construct a conservative test of H0 .
If there is indeed cointegration, another question then arises: what is the
cointegrating rank? Unlike in the parametric classical cointegration case it is
not possible here to handle the problem by multivariate unit root testing as
in Johansen (1988, 1991). Robinson and Yajima (2002) proposed a model-
selection type method to consistently estimate the cointegrating rank, given
an a priori lower bound on the degree of memory parameter reduction. A
similar method was employed by Chen and Hurvich (2003b), who considered
just a single cointegrating subspace, requiring an upper bound on the mem-
ory parameter of the cointegrating error and a lower bound on the memory
parameter of the observed series. A related method for nonstationary sys-
tems based on the exact local Whittle estimator was considered by Nielsen
and Shimotsu (2007). A residual-based method for determining not only the
cointegrating rank but also the dimensions of the fractional cointegrating
subspaces was briey mentioned in Section 5, and described in more detail
in Chen and Hurvich (2006), and in even more detail in a pre-publication
version of that paper.

References

Adenstedt, R. K. (1974): On large-sample estimation for the mean of a stationary random


sequence. Annals of Statistics 2, 10951107.
Akonom, J. and Gourieroux, C. (1987): A functional central limit theorem for fractional
processes. Preprint, CEREMAP, Paris.
Breitung, J. and Hassler, U. (2002): Inference on the cointegration rank in fractionally
integrated processes. Journal of Econometrics 110, 167185.
Brockwell, P. J. and R. A. Davis (1991): Time series: theory and methods: Second Edition.
Springer, New York.
Chan, N. H. and Terrin, N. (1995): Inference for unstable long-memory processes with
applications to fractional unit root autoregressions. Annals of Statistics 23, 1662
1683.
Fractional Cointegration 725

Chen, W. (2006): Eciency in estimation of memory Preprint, SSRN e-Library.


Chen, W. and Hurvich, C. (2003a): Estimating fractional cointegration in the presence of
polynomial trends. Journal of Econometrics 117, 95121.
Chen, W. and Hurvich, C. (2003b): Semiparametric estimation of multivariate fractional
cointegration. Journal of the American Statistical Association 98, 629642.
Chen, W. and Hurvich, C. (2006): Semiparametric estimation of fractional cointegrating
subspaces. Annals of Statistics 34.
Christensen, B. J. and Nielsen, M. O. (2006): Asymptotic normality of narrow-band least
squares in the stationary fractional cointegration model and volatility forecasting.
Journal of Econometrics 133, 343371.
Davidson, J. (2002): A model of fractional cointegration, and tests for cointegration using
the bootstrap. Journal of Econometrics 110, 187212.
Dolado, J.J., Gonzalo, J. and Mayoral, L. (2003): Long range dependence in Spanish po-
litical opinion poll data. Journal of Applied Econometrics 18, 137155.
Dueker, M. and Startz, R. (1998): Maximum-likelihood estimation of fractional cointegra-
tion with an application to U.S. and Canadian bond rates. The Review of Economics
and Statistics 80, 420426.
Engle, R. and Granger C. W. J. (1987): Co-integration and error correction: representation,
estimation and testing. Econometrica 55, 251276.
Granger, C. W. J. and Joyeux, R. (1980): An introduction to long-memory time series
models and fractional dierencing. Journal of Time Series Analysis 1, 1529.
Granger, C. W. J. (1986): Developments in the study of cointegrated economic variables.
Oxford Bulletin of Economics and Statistics 48, 213228.
Hosking, J. R. M. (1981): Fractional dierencing. Biometrika 68, 165176.
Hurvich, C. M. and Chen, W. (2000): An ecient taper for potentially overdierenced
long-memory time series. Journal of Time Series Analysis 21, 155180.
Hurvich, C. M. and Ray, B. K. (1995): Estimation of the memory parameter for nonstation-
ary or noninvertible fractionally integrated processes. Journal of Time Series Analysis
16, 1741.
Jeganathan, P. (1999): On asymptotic inference in cointegrated time series with fractionally
integrated errors. Econometric Theory 15, 583621.
Johansen, S. (1988): Statistical analysis of cointegration vectors Journal of Economic Dy-
namics and Control 12, 231254.
Johansen, S. (1991): Estimation and hypothesis testing of cointegration vectors in Gaussian
vector autoregressive models. Econometrica 59, 15511580.
Lobato, I. N. (1999): A semiparametric two-step estimator in a multivariate long memory
model. Journal of Econometrics 90, 129153.
Mandelbrot, B. and Van Ness, J. (1968): Fractional Brownian motions, fractional noises
and applications. SIAM Review 10, 422437.
Marinucci, D. and Robinson, P.M. (1999): Alternative forms of fractional Brownian motion.
Journal of Statistical Planning and Inference 80, 111122.
Marinucci, D. and Robinson, P.M. (2000): Weak convergence of multivariate fractional
processes. Stochastic Processes and their Applications 86, 103120.
Marinucci, D. and Robinson, P.M. (2001): Journal of Econometrics Journal of Econometrics
105, 225247.
Marmol, F. and Velasco, C. (2004): Consistent testing of cointegrating relationships. Econo-
metrica 72, 18091844.
Nielsen, M. O. and Shimotsu, K. (2007): Determining the cointegrating rank in nonstation-
ary fractional systems by the exact local Whittle approach. Journal of Econometrics
to appear.
Pham, T. D., and Tran, L. T. (1985): Some mixing properties of time series models.
Stochastic Processes and their Applications 19, 297303.
Robinson, P. M. (1994): Semiparametric analysis of long-memory time series. Annals of
Statistics 22, 515539.
726 W.W. Chen and C.M. Hurvich

Robinson, P. M. (1995b): Gaussian semiparametric estimation of long range dependence.


Annals of Statistics 23, 16301661.
Robinson, P. M. (2006): Multiple local Whittle estimation in nonstationary systems.
Preprint.
Robinson, P. M. and Hualde, J. (2003): Cointegration in fractional systems with unknown
integration orders. Econometrica 71, 17271766.
Robinson, P. M. and Marinucci, D. (2001): Narrow-band analysis of nonstationary pro-
cesses. Annals of Statistics 29, 947986.
Robinson, P. M. and Marinucci, D. (2003): Semiparametric frequency domain analysis
of fractional cointegration. In: Robinson, P. (Ed.): Time series with long memory,
Advanced Texts in Econometrics. Oxford University Press, Oxford.
Robinson, P. M. and Yajima, Y. (2002): Determination of cointegrating rank in fractional
systems. Econometrics 106, 217241.
Shimotsu, K. (2006): Gaussian semiparametric estimation of multivariate fractionally in-
tegrated processes. Journal of Econometrics, forthcoming.
Shimotsu, K. and Phillips, P. C. B. (2005): Exact local Whittle estimation of fractional
integration. Annals of Statistics 33, 18901933.
Silveira, G. (1991): Contributions to Strong Approximations in time series with applica-
tions in nonparametric statistics and functional central limit theorems. Ph.D. Thesis,
University of London.
Sowell, F. (1990): The fractional unit root distribution. Econometrica 58, 495505.
Taqqu, M. S. (2003): Fractional Brownian motion and long-range dependence. In: Doukhan,
P., Oppenheim, G. and Taqqu, M. (Eds.): Theory and Applications of Long-Range
Dependence. Birkhuser, Boston.
Velasco, C. (1999a): Non-stationary log-periodogram regression. J. Econometrics 91, 325
371.
Velasco, C. (1999b): Gaussian semiparametric estimation of non-stationary time series
Journal of Time Series Analysis 20, 87127.
Velasco, C. (2003): Gaussian semi-parametric estimation of fractional cointegration. Jour-
nal of Time Series Analysis 24, 345378.
Velasco, C. and Marmol, F. (2004): Consistent testing of cointegration relationships. Econo-
metrica 72, 18091844.
Velasco, C. (2007): The periodogram of fractional processes. Journal of Time Series Anal-
ysis, forthcoming.

Вам также может понравиться