ECONSIGS
ECONSIGS
ECONSIGS
OF SIGNAL EXTRACTION
by D.S.G. POLLOCK
University of London
and
GREQAM: Groupement de Recherche en
Economie Quantitative d’Aix–Marseille
Emails: stephen [email protected]
The Wiener–Kolmogorov signal extraction filters, which are widely used in econo-
metric analysis, are constructed on the basis of statistical models of the processes
generating the data. In this paper, such models are used mainly as heuristic de-
vices that are to be specified in whichever ways are appropriate to ensure that the
filters have the desired characteristics. The digital Butterworth filters, which are
described and illustrated in the paper, are specified in this way.
The components of an econometric time series often give rise to spectral
structures that fall within well-defined frequency bands that are isolated from each
other by spectral dead spaces. We find that the finite-sample Wiener–Kolmogorov
formulation lends itself readily to a specialisation that is appropriate for dealing
with band-limited components.
1. Introduction
The econometric methods of signal extraction that are based on linear filters
have attained a high level of sophistication. They have now coalesced into two
distinct categories.
In the first category are the methods that are favoured by the central
statistical agencies of many of the OECD counties. These were developed
in the North American agencies, notably in Statistics Canada (Dagum 1980)
and in the U.S. Bureau of the Census (Findley et al. 1998), and they have
been supported and refined in other agencies throughout Europe and in the
Antipodes.
The methods are based upon a variety of moving-average smoothing filters,
and they are used, principally, in trend extraction and in deseasonalising data
series. Considerable efforts have been devoted to dealing with the end effects
that can arise when such filters are applied to short non-stationary sequences.
The success of the methods of the central statistical agencies has entailed
a widespread acceptance of a set of common conventions and definitions. Thus,
it is commonly agreed that a deseasonalised data series can be fairly and simply
defined as the product of the relevant methods of the central statistical office.
However, the acceptance of such conventions, whilst necessary to ensure
comparability of statistics across nations, inhibits scientific innovation and dis-
covery. Therefore, academic interest has been focused mainly on the so-called
model-based procedures, which constitute the second category of econometric
signal-extraction methodology.
1
D.S.G. POLLOCK: Econometric Signal Extraction
The model-based approaches are derived from the idea that the compo-
nents of an econometric times series, which are its trend, its secular cycles, its
seasonal cycles and its irregular component, can all be modelled by autoregres-
sive integrated moving-average (ARIMA) processes of low orders. There are
several exponents of this genre who have taken slightly different approaches.
In the structural approach, which originated with Harrison and Stevens
(1976) and which has been developed by Harvey (1989) and others, such models
are fitted jointly to the data in a manner that renders their parameters readily
accessible. (These methods have been implemented in the STAMP program of
Koopman et al. 2000.) In the alternative canonical approach, which has been
advocated by Hillmer and Tiao (1982) and by Maravall and Pierce (1987),
amongst others, the models of the individual components must be disentangled
from a fitted ARIMA model that represents their joint effects. (An accessible
implementation is in the SEATS–TRAMO program of Gómez and Maravall,
1994, 1996.)
The model-based procedures have the seeming advantage that they subsist
within a framework that facilitates conventional statistical inference. Thus,
for example, confidence intervals are easily generated that can surround the
estimated data components. However, the validity of such inferences depends
crucially upon the cogency of the linear time-invariant ARIMA models that are
applied to the components. The ability of such models to reflect the underlying
data structures is limited. In particular, the models are liable to be subverted
whenever the structures show any significant tendency to evolve through time.
In this paper, we also use models in deriving the filters that are used to
isolate the data components, but the models will be treated mainly as heuristic
devices that allow us to exploit the mathematical formalisms of the Wiener–
Kolmogorov theory of signal extraction. This theory indicates that the optimal
estimates of the data components are provided by their conditional expectations
that are formed in the light of the observed data and of the models that are
presumed to have generated them.
When the models themselves are to be fitted to the data, it is important
that they should be realistic—otherwise they will provide an insecure basis
for forming the supposedly optimal filters. However, in practice, they rarely
achieve much realism. When the models are used merely as heuristic devices,
they can be specified in whichever ways are appropriate to ensure that the
filters have the desired characteristics. Moreover, the desirable characteristics
will be unaffected, in the main, by the evolutions of the data components that
the resulting filters are designed to isolate.
The original Wiener–Kolmogorov theory was developed under the fictional
assumption that the data are generated by a stationary stochastic process and
that they form a doubly infinite sequence. In econometrics, one has to contend
with short non-stationary sequences; and, to cope with these, it is common to
resort to the Kalman filter.
The Kalman filter and the associated smoothing algorithms are compli-
cated and powerful devices, of which the workings can often seem obscure. The
difficulty can be attributed to the all-encompassing nature of the algorithms.
2
D.S.G. POLLOCK: Econometric Signal Extraction
In this paper, we shall pursue a simpler approach that fulfils the same objec-
tive of obtaining quasi minimum-mean-square-error estimates, but which deals
directly with the specific features of the problem at hand. We shall use the
finite-sample version of the bidirectional Wiener–Kolmogorov filter that has
been expounded in previous papers of the present author (see Pollock, 1997,
2000, 2001, and 2002).
One of the contentions of this paper is that the components of an economet-
ric time series often give rise to spectral structures that fall within well-defined
frequency bands that are isolated from each other by spectral dead spaces. This
leads us to consider the nature of band-limited stochastic processes, which are
characterised by singular dispersion matrices. We find that the finite-sample
Wiener–Kolmogorov formulation lends itself readily to a specialisation that is
appropriate for dealing with band-limited components.
where t ∈ {0, ±1, ±2, . . .} is the index of the discrete-time observations. Ac-
cording to the classical assumptions, which we shall later amend in various
ways, the signal and the noise are generated by stationary stochastic processes
that are mutually independent. It follows that the autocovariance generating
function of the data, defined by
∞
X
γ(z) = γ0 + γτ (z τ + z −τ ), (2)
τ =1
In practice, the available data will form a finite sequence which consti-
tutes a vector y = [y0 , y1 , . . . , yT −1 ]0 with a signal component ξ and a noise
component η such that
y = ξ + η. (4)
The data might owe their stationarity to a prior differencing operation or to
an operation that has involved the extraction of a trend from the original data
and the retention of the residue. For these vectors, the moment matrices are
E(ξ) = 0, D(ξ) = Ωξ ,
E(η) = 0, D(η) = Ωη , (5)
and C(ξ, η) = 0.
3
D.S.G. POLLOCK: Econometric Signal Extraction
y = x + h, (10)
which is to say that the estimated components add up to the data vector y, as
do the true components ξ and η in equation (4). To calculate the estimates,
we may begin by solving the equation
(Ωξ + Ωη )b = y (11)
x = Ωξ b and h = Ωη b. (12)
4
D.S.G. POLLOCK: Econometric Signal Extraction
The solution to equation (11) may be found via a Cholesky factorisation that
sets Ωξ +Ωη = GG0 , where G is a lower-triangular matrix. The system GG0 b = y
may be cast in the form of Gp = y and solved for p. Then G0 b = p can be
solved for b.
The solution via the Cholesky decomposition constitutes a recursive bidi-
rectional filtering process that generates the vector b via two passes running
in opposite directions through the data. The vector p is the product of a pass
that runs forwards in time, and the vector b is generated from p in a reverse-
time pass. Then b is subjected to further non-recursive filtering operations,
described by (12), which produce x and h.
Exactly the same results would be obtained, albeit in a more complicated
way, by using the Kalman filter, in a forwards pass, and an associated smoothing
algorithm, in a backwards pass. (For an exposition of the latter procedures,
see, for example, Pollock 2003c, where the method of Ansley Kohn 1985 for
initiating the recursions is also analysed.)
It is notable that the recursive filter weights that are provided by the rows
of the matrices G and G0 vary as the filter progresses through the sample. As the
sample size increases, the weights in the final rows of G will tend asymptotically
to the set of constant weights that would be derived under the assumption of
a doubly-infinite data sequence.
In some small-sample implementations of the Wiener–Kolmogorov filter,
a set of constant weights has been applied to data samples that have been
extended at either end by backcasting and forecasting. The additional extra-
sample values have been used in a run-up to the filtering process wherein the
filter is stabilised by providing it with a plausible presample history, if it is
working in the direction of time, or with a plausible post-sample future, if it is
working backward in time. The manner of treating the end-of sample problem
that is implicit in constructions presented here has the theoretical sanction that
it conforms with the theory of minimum-mean-square-error prediction in finite
samples.
The Wiener–Kolmogorov principle of signal extraction is the foundation of
the model-based methods of unobserved components analysis that are nowadays
in widespread use. The parameters of the filters are determined in the process
of fitting ARIMA models to the data components. However, the principle also
supports a variety of heuristic filters of which the parameters are determined
by rule of thumb or in view of the desired characteristics of their frequency
response functions.
Amongst such heuristic filters is the digital version of the Butterworth
filter, which has been advocated by Pollock (1997, 1999, 2000, 2001a, 2001b).
The frequency response of the filter is maximally flat in the vicinity of the zero
frequency and it has a transition band, centered on a chosen cut-off frequency,
that can be narrowed by increasing the filter order. (Gómez 2001 has also
advocated the Butterworth filter, but he has widened its definition to include
filters, such as the filter of Hodrick and Prescott 1997, that do not share the
property of maximal flatness.)
The Butterworth filter, which is of the lowpass Wiener–Kolmogorov va-
5
D.S.G. POLLOCK: Econometric Signal Extraction
11.5
11
10.5
10
0 50 100 150
Figure 1. The quarterly series of the logarithms of consumption in the U.K., for the
years 1955 to 1994, with a linear function interpolated by least-squares regression.
0.15
0.1
0.05
0
−0.05
−0.1
0 50 100 150
Figure 2. The residuals obtained by fitting a linear trend through the logarithmic
consumption data of Figure 1.
0.01
0.0075
0.005
0.0025
0
0 π/4 π/2 3π/4 π
Figure 3. The periodogram of the residuals obtained by fitting a linear trend through
the logarithmic consumption data of Figure 1. The gain of the lowpass Butterworth
filter of order n = 6 and with a cut-off frequency of π/4 is represented by the dotted
line. (The gain is unity at zero frequency.)
6
D.S.G. POLLOCK: Econometric Signal Extraction
σξ2 γξ (z)
ψ(z) =
σξ2 γξ (z) + ση2 γη (z)
(13)
(1 + z)n (1 + z −1 )n
= .
(1 + z)n (1 + z −1 )n + λ(1 − z)n (1 − z −1 )n
(Here, the normalised autocovariance functions γξ (z) and γη (z), which are au-
tocorrelation functions in other words, need to be scaled by the factors σξ2 and
ση2 respectively, which stand of the variances of the white-noise processes from
which the signal and the noise components are supposedly derived by linear
filtering.)
The autocovariance generating functions relate to an heuristic statistical
model as opposed to a realistic one. The denominator function γ(z) = σξ2 γξ (z)+
ση2 γη (z) stands in place of that of the data process and the numerator function
σξ2 γξ (z) corresponds to that of the signal. Here, λ = ση2 /σξ2 = {1/ tan(ωC /2)}2n
incorporates the nominal cut-off frequency of ωC . By setting z = LT and
z −1 = FT in the numerator and the denominator of ψ(z), we derive the matrices
Ωξ and Ωξ + Ωη , respectively, which can be entered into equation (8).
To derive the two unidirectional filters, the rational function is factorised
as ψ(z) = β(z)β(z −1 ), where β(z), which relates to the direct-time filter, con-
tains the poles that lie outside the unit circle, and β(z −1 ), which relates to
the reverse-time filter, contains the poles that lie inside the circle. This fac-
torisation is described as the Cramér–Wold decomposition. In the case of the
Butterworth filter, analytic expressions for the roots of both the denominator
and the numerator, i.e the poles and the zeros of the filter, are available. The
roots of the denominator have been given by Pollock (2000).
For most other Wiener–Kolmogorov filters specified in the manner of the
Butterworth filter, it is necessary to use an iterative procedure for finding the
Cramér–Wold decomposition. (See for, example Pollock 2003b.) The algorithm
of Wilson (1969), which is based on the Newton–Raphson procedure, is an
effective way of achieving the factorisation; and versions which are coded in C
and in P ascal have been provided by Pollock (1999). (See, also, Laurie 1980,
1982.)
There can be a reasonable objection to the assumption that the data com-
ponents are generated by ordinary linear stochastic processes that comprise the
full range of frequencies from zero up to the limiting Nyquist frequency of π
radians per period. (In discretely sampled systems, the frequencies in excess of
the Nyquist value will be aliased by frequencies within the interval [0, π].) We
shall illustrate the grounds for questioning the assumption via an analysis of a
leading economic index.
Example 1. Figure 1 show the logarithms of the quarterly consumption data
for the U.K. for the years 1955–1994, through which a linear trend has been
interpolated by least-squares regression. When a quadratic polynomial trend
was fitted, it was discovered that the coefficient associated with t2 was not
7
D.S.G. POLLOCK: Econometric Signal Extraction
significantly different from zero. This implies that, over the years in question,
the underlying growth of the economy was at a constant exponential rate. The
residual deviations from the trend, which are shown in Figure 2, represent a
variable multiplicative factor by which underlying trend is modulated; and the
residuals reveal both secular and seasonal variations in consumption.
The periodogram of the residuals is shown in Figure 3. This has a low-
frequency spectral structure, which extends no further than the frequency value
of π/8. The remainder of the periodogram shows a dead space that is punctu-
ated by tall spikes in the vicinities of the frequencies of π/2 and π. The first
of these spikes corresponds to the fundamental frequency of the seasonal fluc-
tuations that play on the back of the more gentle variations that surround the
ascending line in Figure 1. The spike at π is corresponds to the first harmonic
of the seasonal frequency.
The low-frequency structure of Figure 3, which occupies the frequency
interval [0, π/8], can be isolated successfully by any of a wide variety of filters.
All that is required of such a filter is that its transition from pass band to
stop band occurs within the spectral dead space that stretches from π/8 to the
vicinity of π/2, where the spectral structure of the seasonal fluctuations is first
encountered. The Butterworth filter of order n = 6 with a cut off frequency of
π/4 fulfils this requirement. Its frequency response function is superimposed
on Figure 3.
A more exacting task is the extraction of the low-frequency components
from data that is observed at monthly intervals. In that case, the fundamental
seasonal frequency is at π/6 and the transition of the filter must occur within
a correspondingly reduced interval.
The sharpening of the transition can be achieved by raising the order
n of the filter. However, a sharp transition in the low frequency range can be
achieved with a recursive Wiener–Kolmogorov filter only at the cost of bringing
the poles of the filter into close proximity with the perimeter of the unit circle.
This can lead to problems of filter instability, which include the propagation
of numerical rounding errors and the prolongation of the transient effects of
ill-chosen start-up conditions.
These problems have been addressed within the context of the Wiener–
Kolmogorov specification by Pollock (2003b). Alternative specifications for
recursive filters have been investigated in Pollock (2003a). In the next section,
we shall also deal with the problems of “sharp filtering” within the context of
the Wiener–Kolmogorov theory; but we shall forsake the method of recursive
filtering in favour of a method based on Fourier analysis.
KT = [e1 , . . . , eT −1 , e0 ], (14)
8
D.S.G. POLLOCK: Econometric Signal Extraction
0.15
0.1
0.05
0
−0.05
−0.1
0 50 100 150
Figure 4. The low-frequency component of the consumption residuals of Figure 2.
The component has been extracted by applying a lowpass Butterworth filter of order
n = 6 with a cut off point at ωc = π/4.
0.06
0.04
0.02
0
−0.02
−0.04
−0.06
0 50 100 150
Figure 5. The component extracted from the consumption residuals by applying a
highpass Butterworth filter of order 6 with a cut off point at ωc = π/4.
0.04
0.02
0
−0.02
−0.04
−0.06
0 50 100 150
Figure 6. The seasonal component of the consumption residuals, synthesised from
the Fourier ordinates in the vicinities of π/2 and π .
9
D.S.G. POLLOCK: Econometric Signal Extraction
which is formed from the identity matrix IT by moving the leading vector to
the back of the array.
This operator effects the cyclic permutation of the elements of any (col-
umn) vector of order T . The matrix is T -periodic such that K q+T = K q .
Whereas LT y = [0, y0 , . . . , yT −2 ] is obtained from y = [y0 , y1 , . . . , yT −1 ] by
deleting the final element and placing a zero in the leading position, the vector
KT y = [yT −1 , y0 , . . . , yT −2 ] is obtained from y by moving the final element to
the leading position.
The powers of K form the basis for the set of circulant matrices. In
particular, we may define a matrix of circular autocovariances via the formula
D◦ (y) = Ω◦ = γ(K)
∞
X
= γ0 I + γτ (K τ + K −τ )
τ =1 (15)
X
T −1
= γ0◦ I + γτ◦ (K τ + K −τ ).
τ =1
of which the generic element in the jth row and tth column is
Ū U = U Ū = I, (19)
K = Ū DU = U D̄Ū , (20)
where
D = diag{1, W, W 2 , . . . , W T −1 } (21)
10
D.S.G. POLLOCK: Econometric Signal Extraction
is a diagonal matrix whose elements are the T roots of unity, which are found
on the circumference of the unit circle in the complex plane. Observe also that
D is T -periodic, such that Dq+T = Dq , and that K q = Ū Dq U = U D̄q Ū for
any integer q.
The spectral factorisation of the circulant autocovariance matrix gives
This represents the cosine Fourier transform of the sequence of the ordinary
autocovariances; and it corresponds to an ordinate (scaled by 2π) sampled at
the point ωj from the spectral density function of the linear (i.e. non-circular)
stationary stochastic process. (An account of the algebra of circulant matrices
has been provided by Pollock 2002. See, also, Gray 2002.)
The circulant autocovariance matrices that are the counterparts of the
ordinary autocovariance matrices defined in (5) are
Ω◦ξ = Ū Λξ U, Ω◦η = Ū Λη U,
(24)
Ω◦ = Ū ΛU = Ū (Λξ + Λη )U,
x = Ū Λξ {Λξ + Λη }+ U y = Ū Jξ U y, (25)
h = Ū Λη {Λξ + Λη }+ U y = Ū Jη U y. (26)
11
D.S.G. POLLOCK: Econometric Signal Extraction
12
D.S.G. POLLOCK: Econometric Signal Extraction
observe that · ¸
Q0∗
[ S∗ S] = S∗ Q0∗ + SQ0 = IT , (28)
Q0
and that · ¸ · ¸ · ¸
Q0∗ Q0∗ S∗ Q0∗ S I 0
[ S∗ S]= = d . (29)
Q0 Q0 S∗ Q0 S 0 IT −d
When the difference operator is applied to the data vector y, the first d
elements of the product, which are in g∗ , are not true differences and they are
liable to be discarded: · 0¸ · ¸
Q∗ g
∇T y =
d
y= ∗ . (30)
Q0 g
However, if the elements of g∗ are available, then the vector y can be recovered
from g = Q0 y via the equation
y = S∗ g∗ + Sg. (31)
The columns of the matrix S∗ provide a basis for the set of polynomials of
degree d − 1 defined over the integer values t = 0, 1, . . . , T − 1. Therefore,
p = S∗ g∗ is a vector of polynomial ordinates whilst g∗ can be regarded as a
vector of d polynomial parameters.
We may approach the filtering of a trended data sequence in the following
manner. First, we reduce the data to stationarity by differencing it an ap-
propriate number of times. (We rarely need to difference the data more than
twice.) From the differenced data, viewed in an appropriate manner, we may
discern the nature and the frequency ranges of the various data structures that
we wish to isolate.
Next, the components of the differenced data that correspond to these
structures may be extracted, either by a recursive filtering process—using, for
example, a Butterworth filter—or via the Fourier method described in the
preceding section.
Finally, the components of the differenced data may be integrated, with an
appropriate choice of initial conditions, to provide estimates of the components
of the original trended sequence.
An apparent problem with this procedure is that the act of differencing is
liable to attenuate the components of the low-frequency data structure to such
an extent that they become invisible in the periodogram of the differenced data.
The problem is illustrated in Figure 8, which shows the periodogram of g = Q0 y
in the case of the the once-differenced consumption data.
The problem vanishes when we recognise that we can discern the low-
frequency structure via the periodogram of the residual sequence
13
D.S.G. POLLOCK: Econometric Signal Extraction
Q0 y = Q0 ξ + Q0 η,
(33)
= ζ + κ = g,
and we may assume, by analogy with (5), that ζ and κ are characterised by
their first and second moments, which are
E(ζ) = 0, D(ζ) = Ωζ = Q0 Ωξ Q,
E(κ) = 0, D(κ) = Ωκ = Q0 Ωη Q, (34)
and C(ζ, κ) = 0.
Here, the derived dispersion matrices Ωζ and Ωκ retain the Toeplitz structure
that is a feature of Ωξ and Ωη .
Let the estimates of ζ and κ be denoted by z and k. If x and h are the
estimates of ξ and η respectively, then it is reasonable to require that Q0 x = z
and Q0 h = k so that
Q0 y = Q0 x + Q0 h
(35)
= z + k = g.
The estimates z and k must be integrated to give
This requires that the estimated trend x should adhere as closely as possible
to the data. The minimising value is
z∗ = (S∗0 Ω−1
η S∗ )
−1 0 −1
S∗ Ωη (y − Sz) (38)
Minimise h0 Ω−1 0 −1
η h = (S∗ k∗ + Sz) Ωη (S∗ k∗ + Sk). (39)
k∗ = −(S∗0 Ω−1
η S∗ )
−1 0 −1
S∗ Ωη Sk. (40)
14
D.S.G. POLLOCK: Econometric Signal Extraction
Using
P∗ = S∗ (S∗0 Ω−1
η S∗ )
−1 0 −1
S ∗ Ωη , (41)
we get, from (36), the following values:
x = P∗ y + (I − P∗ )Sz, and h = (I − P∗ )Sk. (42)
The disadvantage in using these formulae directly is that the inverse matrix
Ω−1
η , which is of order T , is liable to have nonzero elements in every location.
(This will be so whenever Ωη has the form of an autocovariance matrix of a
moving-average process, as it does in the case of the Butterworth filter, for
example.)
The appropriate recourse is to use the identity
I − P∗ = I − S∗ (S∗0 Ω−1
η S∗ )
−1 0 −1
S ∗ Ωη
(43)
= Ωη Q(Q0 Ωη Q)−1 Q0
to provide an alternative expression for the projection matrix I − P∗ that incor-
porates the band-limited matrix Ωη instead of its inverse. The equality follows
from the fact that, if Rank[R, S∗ ] = T and if S∗0 Ω−1
η R = 0, then
I − S∗ (S∗0 Ω−1
η S∗ )
−1 0 −1
S∗ Ωη = R(R0 Ω−1
η R)
−1 0 −1
R Ωη . (44)
Setting R = Ωη Q gives the result. Given x = y − h, it follows that we can write
x = y − (I − P∗ )Sk
(45)
= y − Ωη Q(Q0 Ωη Q)−1 k,
where the second equality depends upon Q0 S = I.
So far, we have not specified the precise method by which the estimates z
and k of the differenced components have been obtained. They may be obtained
equally via a recursive filtering method or via the Fourier that has been outlined
in the preceding section. In case we have used the Fourier method, we might
be inclined to use the circulant version of the dispersion matrix Ωη within the
foregoing formulae.
Let us consider, instead, the possibility of obtaining the estimate k via
recursive filtering. Then, with reference to equation (9), we can see that the
assumptions of (34) imply that the estimate should take the form of
k = Q0 Ωη Q(Ωζ + Q0 Ωη Q)−1 Q0 y, (46)
On substituting this in the equation of (45), we get
x = y − Ωη Q(Ωζ + Q0 Ωη Q)−1 Q0 y. (47)
In the case of Butterworth filter, we take the quasi-autocorrelation func-
tions of the nonstationary signal sequence ξ(t) and of the stationary noise se-
quence η(t) to be
(1 + z)n (1 + z −1 )n
γξ (z) = σξ2 and γη (z) = ση2 (1 − z)n−d (1 − z −1 )n−d (48)
(1 − z)d (1 − z −1 )d
15
D.S.G. POLLOCK: Econometric Signal Extraction
11.5
11
10.5
10
1960 1970 1980 1990
Figure 7. The quarterly series of the logarithms of income (upper) and consumption
(lower) in the U.K for the years 1955 to 1994 together with their interpolated trends.
0.3
0.2
0.1
0
0 π/4 π/2 3π/4 π
0.04
0.02
0
−0.02
−0.04
−0.06
0 50 100 150
Figure 9. The bandpass estimates of the fluctuations, within the range of the
business-cycle frequencies, of the logarithmic income series (solid line) and of the
logarithmic consumption series (broken line).
16
D.S.G. POLLOCK: Econometric Signal Extraction
17
D.S.G. POLLOCK: Econometric Signal Extraction
In that case, the corresponding elements of Λη should all have same value, and
so, likewise, should the corresponding elements of Λ+ η.
The remaining elements of Λ+ η , which correspond to zero-valued Fourier
ordinates and which can take arbitrary values, may be set to the same values
as the elements corresponding to the seasonal ordinates. Thus Λ+ η , which needs
to be determined only up to a scalar factor, becomes an arbitrary multiple of
the identity matrix—and it may as well become the identity matrix itself. In
that case, we should have Ω◦η = Ū U = I and (Ω◦η )+ = I.
This simplification allows us to specialise equation (38) to give
The elements of the matrix S∗0 S∗ can be found via the formulae
X
T
1
t2 = T (T + 1)(2T + 1) and
t=1
6
(52)
X
T
1 1
t(t − 1) = T (T + 1)(2T + 1) − T (T + 1).
t=1
6 2
(A compendium of such results has been provided by Jolly 1961, and proofs
of the present results were given by Hall and Knight 1899.) The matrix is
somewhat ill-conditioned. Moreover, when the order of differencing exceeds two
or three, it is necessary, in calculating the polynomial ordinates of p = S∗ z∗ , to
use to an orthogonal basis in place of the monomial basis that is provided by
the columns of S∗ . However, this case is rare.
5. Bandpass Filtering
Econometricians often characterise the business cycle in terms of a sinusoid
that fluctuates around a slow-moving trend. According to the definitions of
Burns and Mitchell (1946), the effects of the business cycle within an economic
index correspond to the sinusoidal elements therein that have periods of no less
than one-and-a-half years and of no more than eight years. A duration of one-
and-a-half years seems too short, and we prefer to set the shortest duration at 2
years—and this seems to be a common preference (see, for example, Christiano
and Fitzgerald 1998).
The business cycle, defined in this manner, is unlikely to correspond to any
self-contained spectral structure that might be discerned by inspecting the rele-
vant periodogram. In the case of quarterly data, the business cycle frequencies
range from π/16 radians per period to π/4 radians per period (corresponding
to a duration of 2 years.) Neither of these values corresponds to a natural break
in the periodogram of the consumption residuals of Figure 3.
18
D.S.G. POLLOCK: Econometric Signal Extraction
Im Im
i i
Re Re
−1 1 −1 1
−i −i
Figure 10. The pole–zero diagrams (left) of the lowpass Butterworth filter of order
n = 12 with a cut-off frequency of ωU = π/4 and (right) of the highpass filter of
order n = 6 with a cut-off frequency of ωL = π/16.
1.25
1
0.75
0.5
0.25
0
0 π/4 π/2 3π/4 π
Figure 11. The gain of the 12th order digital Butterworth lowpass filter with cut-off
frequency of ωU = π/4 and of 6th order highpass filter with cut-off frequency of
ωL = π/16, superimposed on the same diagram.
19
D.S.G. POLLOCK: Econometric Signal Extraction
20
D.S.G. POLLOCK: Econometric Signal Extraction
wherein the condition Q0 S = I has been effective in simplifying the final ex-
pression.
Example 4. Figure 9 shows the business cycle fluctuations that have been
extracted from the quarterly logarithmic income and consumption data for the
U.K. over the period 1955 to 1994. In both cases, a Fourier bandpass filter
has been applied that has a lower cut-off point at π/16 radians per period
(corresponding to a cycle of 8 years duration) and an upper cut-off point of
π/4 radians per period (corresponding to a cycle of 2 years duration).
There is evidence here that the fluctuations in consumption precede those
in income. This contradicts the common supposition that the business cycle
is driven by variations in ‘‘autonomous expenditures”, which do not include
consumption, and in the rate of investment.
One might be doubtful of the comparisons at the beginning and the end of
the sample, where the interpolated functions are not tied down by precedeing
or succeeding data points and where they appear to be heading in opposite
directions. The problem could be overcome by adding a few extrapolated points
at either end of the sample that would serve to tie down the functions.
6. Multiple Components
The problems of econometric signal extraction have been handled, so far, within
the context of a model, described by equation (4), that has only a signal com-
ponent and a noise component. Allowance has been made for a non stationary
signal component. However, it might be required to partition the data amongst
more than two components. Thus, in a classical econometric time-series anal-
ysis, at least four components are identified. These are the trend, the business
cycle, the seasonal cycle and an irregular component.
21
D.S.G. POLLOCK: Econometric Signal Extraction
The two-component model can also serve the purpose of extracting several
components, for the reason that its components are readily amenable, if neces-
sary, to further decompositions. Thus, for example, an initial decomposition of
the data sequence into a trend/cycle component and a residue can be followed
by decomposition of the residue into a seasonal cycle an irregular cycle. If the
data are stationary, it is unnecessary to perform such a multiple decomposition
sequentially—each component can be extracted separately.
If the data are nonstationary and if there are more than one nonstationary
component, then a sequential decomposition might be called for. A typical
model of an econometric time series, described by the equation y = ξ + η =
(µ + ρ) + η, comprises both a trend/cycle component µ and a seasonal compo-
nent ρ that are described by ARIMA models with real and complex unit roots
respectively.
To reduce the data to stationarity, an operator is used that is the product
of the d-fold difference operator ∇dT = (I −LT )d and a deseasonalising operator
ΣT = (I − LsT )(I − LT )−1 . (The operator Σ is used instead of (I − LsT ) because
it can be assumed, without loss of generality, that the seasonal deviations from
the trend have zero mean.) Let the product of the two operators be denoted
by MT = ΣT ∇dT = [Q∗ , Q]0 , where Q0∗ contains the first d + s − 1 rows of the
matrix, and let the inverse operator MT−1 = [S∗ , S] be partitioned conformably
such that S∗ contains the first d + s − 1 columns. The factors of MT−1 are
further partitioned as Σ−1 −d
T = [SΣ∗ , SΣ ] and ∇T = [S∇∗ , S∇ ].
Let the components of the differenced data be denoted by Q0 ξ = ζ, Q0 µ =
ζµ and Q0 ρ = ζρ . Then there is
Q0 y = Q0 ξ + Q0 η
(57)
= Q0 (µ + ρ) + κ = (ζµ + ζρ ) + κ.
Also, let the estimates of µ and ρ be denoted by m and r and those of ζµ and
ζρ by zm and zr . Then, in parallel with equation (57), there is
Q0 y = Q0 x + Q0 h
(58)
= Q0 (m + r) + k = (zm + zr ) + k.
x = (m + r) = S∗ z∗ + Sz
(59)
= S∗ z∗ + S(zm + zr ).
22
D.S.G. POLLOCK: Econometric Signal Extraction
This may be solved uniquely for z∗m and z∗r ; and, for this purpose, only the
first s + d − 1 rows of the system are required. Thereafter, the estimates of µ
and ρ are given by
References
Ansley, C.F., and R. Kohn, (1985), Estimation, Filtering and Smoothing in
State Space Models with Incompletely Specified Initial Conditions, The Annals
of Statistics, 13, 1286–1316.
Baxter, Marianne, and R.G. King, (1999), Measuring Business Cycles: Approx-
imate Band-Pass Filters for Economic Time Series, The Review of Economics
and Statistics, 81, 575–593.
Burns, A.M., and W.C. Mitchell, (1946), Measuring Business Cycles, New
York: National Bureau of Economic Research.
Christiano, L.J., and T.J. Fitzgerald, (1998), The Business Cycle, It’s Still a
Puzzle, Federal Reserve Bank of Chicago, Economic Perspectives, 22, 56–83.
Constantinides, A.G., (1970), Spectral Transformations for Digital Filters, Pro-
ceedings of the IEE, 117, 1585–1590.
Dagum, E.B., (1980), The X-11-ARIMA Seasonal Adjustment Method. Cata-
logue No. 12-564E, Statistics Canada, Ottawa.
Findley, D.F., B.C. Monsell, W.R. Bell, M.C. Otto and B.-C. Chen, (1998), New
Capabilities and Methods of the X-12-ARIMA Seasonal-Adjustment Program,
Journal of Business and Economic Statistics, 16, 127–177.
Gómez, V., (2001), The Use of Butterworth Filters for Trend and Cycle Esti-
mation in Economic Time Series, Journal of Business and Economic Statistics,
19, 365–373.
Gómez, V., and A. Maravall, (1994), Program TRAMO: Time Series Regression
with ARIMA Noise, Missing Observations, and Outliers—Instructions for the
User, EUI Working Paper Eco No. 94/31, Department of Economics, European
University Institute.
Gómez, V., and A. Maravall, (1996), Programs TRAMO and SEATS. Instruc-
tions for the User, (with some updates). Working Paper 9628, Servicio de
Estudios, Banco de España.
Gray, R.M., (2002), Toeplitz and Circulant Matrices: A Review, Information
Systems Laboratory, Department of Electrical Engineering, Stanford Univer-
sity, California, http://ee.stanford.edu/gray/~toeplitz.pdf.
Hall, H.S., and S.R. Knight, (1899), Higher Algebra, Macmillan and Co., Lon-
don.
23
D.S.G. POLLOCK: Econometric Signal Extraction
Harrison, P.J., and C.F. Stevens, (1976), Bayesian Forecasting (With a Discus-
sion), Journal of the Royal Statistical Society, Series B, 38, 205–247.
Harvey, A.C., (1989), Forecasting, Structural Time Series Models and the
Kalman Filter, Cambridge University Press, Cambridge.
Hillmer, S.C. and G.C. Tiao, (1982), An ARIMA-Model Based Approach to
Seasonal Adjustment, Journal of the American Statistical Association, 77, 63–
70.
Hodrick, J.R., and E.C. Prescott, (1997), Postwar U.S. Business Cycles: An
Empirical Investigation, Journal of Money, Credit and Banking, 29, 1–16.
Jolly, L.B.W., (1961), Summation of Series: Second Revised Edition, Dover
Publications: New York.
Koopman, S.J., A.C. Harvey, J.A. Doornik and N. Shephard, (2000), STAMP:
Structural Time Series Analyser, Modeller and Predictor, Timberlake Consul-
tants Press, London.
Laurie, D.P., (1980), Efficient Implementation of Wilson’s Algorithm for Fac-
torising a Self-Reciprocal Polynomial, BIT, 20, 257–259.
Laurie, D.P., (1982), Cramér-Wold Factorisation, Algorithm AS 175, Applied
Statistics, 31, 86–90.
Maravall, A., and D.A. Pierce (1987), A Prototypical Seasonal Adjustment
Model, Journal of Time Series Analysis, 8, 177–193. Reprinted in Hylleberg,
S. (ed.), Modelling Seasonality, Oxford University Press, 1992.
Pollock, D.S.G., (1997), Data Transformations and De-trending in Economet-
rics, Chapter 11 (pps. 327–362) in Christian Heij et al. (eds.), System Dynamics
in Economic and Financial Models, John Wiley and Sons.
Pollock, D.S.G., (1999), A Handbook of Time-Series Analysis, Signal Processing
and Dynamics, Academic Press, London.
Pollock, D.S.G., (2000), Trend Estimation and Detrending via Rational Square
Wave Filters, Journal of Econometrics, 99, 317–334.
Pollock, D.S.G., (2001a), Methodology for Trend Estimation, Economic Mod-
elling, 18, 75–96.
Pollock, D.S.G., (2001b), Filters for Short Nonstationary Sequences, Journal
of Forecasting, 20, 341–355.
Pollock, D.S.G., (2002), Circulant Matrices and Time-Series Analysis, The In-
ternational Journal of Mathematical Education in Science and Technology, 33,
213–230.
Pollock, D.S.G., (2003a), Sharp Filters for Short Sequences, Journal of Statis-
tical Inference and Planning, 113, 663–683.
Pollock, D.S.G., (2003b), Improved Frequency-Selective Filters, Journal of
Computational Statistics and Data Analysis, 42, 279–297.
24
D.S.G. POLLOCK: Econometric Signal Extraction
25