Generalized Sp ectral Estimation
Jeremy Berkowitz
June 12, 1996
Abstract
This pap er provides a framework for estimating parameters in a wide class
of dynamic rational exp ectations mo dels. The framework recognizes that dy-
namic RE mo dels are often meant to match the data only in limited ways.
In particular, interest may fo cus on a subset of frequencies. Thus, this pap er
designs a frequency domain version of GMM. The estimator has several advan-
tages over traditional GMM. Aside from allowing band-restricted estimation, it
do es not require making arbitrary instrumentorweighting matrix choices. The
general estimation framework also includes least squares, maximum likeliho o d
and band restricted maximum likeliho o d as sp ecial cases.
Key Words: Estimation, Frequency Domain, Missp eci cation
JEL Classi cation: C13, C22
Federal Reserve Board, Mail Stop 61A, Washington, D.C. 20551. Telephone: (202) 736-5581.
Email: [email protected]. I would like to thank Frank Dieb old, Lee Ohanian, Valentina Corradi
and Jin Hahn for helpful suggestions. The view presented are solely those of the author and do not
necessarily represent those of the Federal Reserve Board or its sta .
1 Intro duction
This pap er develops frequency domain techniques for estimating dynamic rational
exp ectations mo dels. This approach allows, mo dels to be estimated and tested over
a subset of frequencies, such as business cycle frequencies, seasonal frequencies, or
long horizons. The techniques describ ed in this work are also particularly useful in
allowing researchers to deal squarely with high frequency measurement error.
It is natural for researchers interested in avoiding high frequency noise or in match-
ing particular cyclical b ehavior to carry out estimation and evaluation of such mo dels
in the frequency domain. The frequency domain provides an orthogonalization of the
uctuations in the observed data. Engle (1974) intro duced band sp ectral regression as
a means to assess the relationship b etween economic variables at sp eci c frequencies.
In that work, the criterion b eing minimized was restricted to linear least squares.
Generalized sp ectral estimation (GSE) allows for a much wider class of minimiza-
tion criteria than was previously p ossible. The GSE framework includes a new class
of estimators which I will call 'whitening estimators', as well as least squares, band
sp ectrum regression, and 'Whittle likeliho o d' estimation (which is asymptotically
1
maximum likeliho o d).
This pap er builds on Dieb old, Ohanian and Berkowitz (1995), who prop ose tech-
niques for estimating and evaluating dynamic rational exp ectations mo dels in the
frequency domain. Their framework allows for parameter estimation and mo del as-
sessment inavery general setting. Parameters are estimated by minimizing distance
between sp ectra of observed data and mo del-generated data. Distance may b e de ned
by the user and may fo cus on any relevant subset of frequencies. A shortcoming of
this framework is that in all but the simplest cases, the mo del must b e approximated
and simulated in order to carry out the estimation.
Since, in general, analytic solutions are not available for nonlinear dynamic equi-
librium mo dels, one must cho ose an approximate solution metho d. Furthermore, in
order to pro ceed with estimation, the approximation to the mo del must be carried
out at each parameter con guration. That means the mo del must b e simulated hun-
dreds if not thousands of times. Solution metho ds which can be chosen arbitrarily
'close' to the true mo del (such as discretizing the parameter space) are generally pre-
cluded b ecause of the extreme computational intensity asso ciated with simulating the
mo del only once. For estimation, then, faster and less accurate solution metho ds are
required. To the extent that the approximate solution di ers from the true mo del,
the Dieb old, Ohanian and Berkowitz (1995) estimated parameters will di er from the
parameters which minimize loss. It is dicult to make general statements regarding
1
For a discussion of band-restricted maximization of the Whittle likeliho o d function see Engle
(1980), Dieb old, Ohanian, and Berkowitz (1995). 2
this sort of approximation error. However, Taylor and Uhlig (1990), in a compari-
son of 14 approximation metho ds applied to the sto chastic growth mo del, concluded
that "the simulated sample paths generated by the di erent solution metho ds have
signi cantly di erent prop erties."
The generalized sp ectrum estimator will allow for mo del estimation, inference and
evaluation without requiring an approximate solution of the mo del. We accomplish
this by imp osing moment conditions given by the mo del and then minimizing devi-
ations from these conditions in the frequency domain. It is thus very much in the
spirit of generalized metho d of moment and other minimum distance estimators. For
whitening estimators, we imp ose moment conditions on the residuals which require
that the residuals are 'close' to white noise.
Section 2 de nes the generalized sp ectral estimator and presents some sp ecial
cases for illustration. Section 3 illustrates GSE estimation by presenting the results
of a Monte Carlo exp eriment. In the exp eriment, I maintain the realistic assumption
that the true mo del is unknown. Section 4 concludes.
2 Generalized Sp ectral Estimation
The Euler equation implied by a typical rational exp ectations mo del can be written
as
E (g (y ; ) j )=0; (1)
t 0 t
where g(; ) is a function given by mo del's rst order conditions, y is an Tx1 vector of
t
observable data, isavector of parameter values, and is the ( algebra de ned
0 t
by) agent's time t information set. g (y ;) is sometimes called the Euler residual.
t
Equation 1says that the Euler residual has a zero conditional mean. It implies that
for any rx1 instrument x ,inthe agent's time t information set,
t 1
" #!
1
E g (y ; ) = 0: (2)
t 0
x
t 1
Equation 2 is the familiar basis for GMM estimation. The notion here is that economic
agents should be unable to forecast the mo del residual. Each element of x , each
t 1
instrument, gives us an additional residual whose mean should be zero.
x is generally taken to be some number of lags of the endogenous variables,
t 1
y : However, consideration of lags of g (y ; ) as instruments will lead to some very
t t 0
interesting results. Note that it is always true that all lags of g (y ; ) are in agents'
t 0
time t information set. The motivation for doing this is that equation 1 says that the
Euler residuals form a martingale di erence sequence. Since martingale di erence 3
sequences are also white noise, equation 2 says that the Euler residual g (y ; ) is
t 0
2
multivariate white noise .
This fact provides the motivation for a class of GSE estimators, which I will call
'whitening estimators.' The reasoning is as follows. The mo del residual by design
should contain only the unexplainable part of the mo del in question. So, if we wish to
estimate ,wewould like to eliminate as much of the predictable dynamics in g (y ; )
0 t 0
as p ossible. These dynamics should b e incorp orated into the mo del. Another way of
saying this, is wewould like to estimate the parameter con guration in suchawayas
to make the residuals as 'close' as p ossible to white noise. We are, in e ect, trying to
whiten the g (y ; ): Some examples of whitening estimators will b e provided b elow.
t 0
Rewriting equation 3, we have that
E (g (y ; ) g(y ; )) = 0; =1;2; ::: (3)
t 0 t 0
For notational simplicity, let u = g (y ; ), so that u is a single Euler residual. Then
t t 0 jt
equation 3 is equivalenttothe statement,
2l T
f (! )= k; ! = ;l =0; ::: :
u l
jt
T 2
The sp ectral density of each u ; is constantover the entire supp ort, [0, ]: Then given
jt
T
a nite realization of the residuals, fu g ; we can calculate a consistent estimate of
jt
1
^
p
the sp ectrum, such that f (! ) ! k; 8!:
u
j
In reality, though, u = g (y ; ); and we do not know : So we never observe
t t 0 0
T
a realization of the true Euler residualsfu g ; but rather at b est we can observe
jt
1
T
fu ()g ; where is chosen somehow by the user.
jt
1
2.1 A Simple Example
In order to x ideas, consider a simple linear example, an AR(1),
y = y + u :
t 0 t 1 t
The y are observed and assume u (0; 1): In GSE notation, u ( ) = g (y ; ) =
t t t 0 t 0
(1 L)y ; where L is the lag op erator. If we do not know , we may lo ok at the
0 t 0
sp ectrum of the residual implied by any arbitrary ; f (! ); to see whether it is
u()
constant across all frequencies. For this arbitrary ; we can generate a length T
residual,
u ( )=(1 L)y
t t
2
By this, I mean that, if g (y ; ) isavector sequence, each elementof g(y ; ) is white noise.
t 0 t 0 4
=(1 ( + )L)y
0 t
1
=u y = u u = A(L)u :
t t 1 t t 1 t
1 L
0
And so the sp ectrum of this residual u ( ) is,
t
2
i!
f (! )= 1 e
u( )
i!
1 e
0
Then f (! ) 6= k; across frequencies, unless =0 (that is, at = ):
u( ) 0
It is in this sense natural to estimate , by setting the sp ectrum f (! ) 'close'
0
u( )
to the sp ectrum of white noise. In other words, nd
X
^
=arg min C f (! ) ;
GS E
u( )
!
T
where the loss function C() is a measure of divergence of the sp ectrum from a at
line.
This approach has some signi cant advantage relative to Maximum Likeliho o d
and GMM estimation. To continue with the example, ML estimation would pro ceed
T
by minimizing the distance b etween the mo del data, fy ( )g ; and the observed data,
1
T
fy g .
1
The problem is that in the context of general dynamic equilibrium mo dels the data
generating pro cess (the p olicy function) is not known analytically. The innovation
T
pro cess, fu ( )g ; must rst be simulated and then the rest of the mo del data must
t
1
be generated via an approximate p olicy function. MLE thus requires an additional
layer of approximation.
The advantage of GMM is that one never needs to construct an approximate p olicy
function. Parameters can be estimated using only the observed data. The problem
with GMM is that there is no such thing as the GMM estimator. The researcher is
required to make achoice regarding b oth instruments and the weighting matrix.
GSE, like GMM, can pro ceed without requiring an approximate solution of the
mo del. But unlike GMM, the instruments and weighting are not arbitrarily chosen.
Rather, the user cho oses a frequency band of interest. For example, if the mo del is
designed to explain business cycle uctuations, the user may sp ecify cycles between
2to8 years.
De nition: The generalized sp ectral estimator is de ned as:
^
X
^
=arg min C f (! ) ;
GS E
u( )
! 2$
T 5
^
where C() may include a wide variety of appropriate loss functions, f (! ) is a
u( )
consistent estimate of the sample sp ectral density matrix (or p ossibly the p erio dogram
matrix ) of u ( )=g(y ;): The summation is taken over the frequencies of interest,
t t
^
represented by the set $: Wemay appropriately think of the f (! ) as a triangular
l
u( )
2l T T
array with ! = ;l =0; ::: : The maximum frequencies p er sample size are . This
l
T 2 2
loss function allows for fo cusing on a subset of frequencies which are judged to be
economically relevant. This is not p ossible in the time domain. To get a b etter feel
for this estimation strategy,wenow describ e some examples of the loss function C():
For what follows, we consider only a univariate u ( ):
t
2.2 Whitening Estimators
We might sp ecify,
2
X
^ ^
=arg min (! ) ! ; (4)
F
u( )
!
^
(! )isthe sp ectral distribution or 'cumulative sp ectrum' of u : We have, where
F
t
u( )
3
in e ect, inverted the Cramer-Von Mises statistic to obtain an estimator . A few
comments are in order here.
1) Rational exp ectations implies that u is white noise, so our moment conditions
t
are that E F (! ) = !: In words, we have that, in p opulation, the sp ectral
u( )
distribution function of white noise is the 45 line. This estimator minimizes the
2
L distance between the sample sp ectral distribution and the p opulation sp ectral
distribution of white noise.
2) This estimator is very much like GMM in that it minimizes the squared devi-
ations from moment conditions.
3) There are an in nite numberofvalid functions, C (), with unique minima where
u is white noise. Durlauf (1991), for example, considers a numb er of sp eci c distance
t
functions that measure deviations of the cumulative p erio dogram in the context of
testing for white noise.
4) We suppress dep endence of the set of ! on the sample size for the remainder
of the pap er for notational simplicity.
A particularly interesting example of a whitening estimator, which I will call
"Sp ectral-GMM", arises when we take C ()as follows,
R
1
2
3
(BB(z)) dz ; where BB(z) is a Brownian Bridge. The Cramer-Von Mises statistic is actually
0 6
2
^
X
^
=arg min f ; (5) f (! )
u( ) u ( )
1
! 2$
where f is the average of the estimated sp ectra (or p erio dograms).
u( )
This estimator displays some interesting prop erties. We will show that it coincides
with a time domain GMM estimator, with the choice of $ corresp onding to the choice
of a GMM weighting matrix. First we need to de ne some notation.
1
i! j
i 1
p
Set W be the Fourier matrix. It has typical element [W] = [e ], note
ij
2T
y y
that this matrix has the prop erty that W W = I, where W is the conjugate transp ose
of W:
1
p
De ne W*, [W*] = [cos (! j )]k (j ), where k(j) is a smo othing window so
mj m
2T
that k(j)=0 for j>B ;m 2 $: So the number of rows of this matrix is equal to the
T
number of frequencies included in $: Call the number of frequencies included in the
band M, so that W* is an M x T matrix. Let,
2 3
P
1
u u
t t 1
T 1
P
6 7
1
6 7
u u
t t 2
T 2
^ 6 7
G = (6)
6 7
:::
4 5
P
1
u u
t
t (T 1)
T (T 1)
be the vector of moment conditions implied by the mo del. Also de ne,
3 2
1 1
:: 1
T T
7 6
1
:: 1 ::
(7) A =
5 4
T
1 1