Professional Documents
Culture Documents
Y Hristophe Lanas and Aoul Epoutot
Y Hristophe Lanas and Aoul Epoutot
ADJUSTMENT
1. INTRODUCTION
2. MODEL SPECIFICATION
SðBÞst ¼ hs ðBÞast
ð2:3Þ
r2 nt ¼ hn ðBÞant
For the seasonal adjustment of monthly series, the airline model has an interesting
feature: for some values of ðh1 ; h12 Þ, it yields optimal filters very close to the
historical X-11 ones. Table I reports the couples ðh1 ; h12 Þ yielding the WK signal
extraction filter the closest to the X-11 ones. The loss function used is the average
of the squared differences between the frequency transfer functions; it is reported
in Table I as d, and it will be explicitely defined in Section 3. The 20 seasonal
adjustment filters embodied in recent X-11 releases like X-12-ARIMA are
considered. They are obtained by combining the 3 3; 3 5; 3 9; 3 15 plus
the X-11 default for the seasonal moving average with Henderson trend filters of
TABLE I
AIRLINE MODEL PARAMETERS YIELDING WK FILTERS CLOSEST TO X-11 FILTERS
Henderson MA
X-11 filters Seasonal MA 9-term 13-term 17-term 23-term
33 ðh1 ; h12 Þ (0.599, 0.400) (0.597, 0.384) (0.594, 0.375) (0.589, 0.364)
d 0.008 0.008 0.007 0.006
X-11 default ðh1 ; h12 Þ (0.583, 0.556) (0.583, 0.551) (0.583, 0.549) (0.583, 0.546)
d 0.007 0.007 0.007 0.007
35 ðh1 ; h12 Þ (0.583, 0.563) (0.583, 0.553) (0.583, 0.548) (0.583, 0.543)
d 0.009 0.009 0.010 0.010
39 ðh1 ; h12 Þ (0.583, 0.732) (0.583, 0.728) (0.583, 0.726) (0.583, 0.723)
d 0.012 0.013 0.013 0.014
3 15 ðh1 ; h12 Þ (0.583, 0.828) (0.583, 0.826) (0.583, 0.825) (0.583, 0.824)
d 0.013 0.014 0.014 0.015
length 9,13,17,23. The X-11 default mixes the 3 3 and the 3 5 seasonal moving
averages (Bell and Monsell, 1992).
As can be seen on Table I, the average square differences between frequency
transfer function of the two filters are always less than 0.015. The parameter h1
takes a limited range of values, while the range of variation of h12 is much
larger: it is the parameter which controls the width of the power-vanishing
band around the seasonal harmonics in the WK seasonal adjustment filter.
Figure 1 displays the frequency transfer function of the X-11 adjustment filters
obtained when the length of the Henderson moving average is set to 13, the
most common setting, together with the frequency transfer function of WK
filters obtained with the parameter values ðh1 ; h12 Þ given in the fourth column
of Table I. The frequency transfer function of the X-11 filter with 3 5
seasonal moving average is not displayed, because it is shaped alike the default
one. It can be seen that the airline model can generate signal extraction filters
that are very close to the X-11 ones. Since that model is selected in many
automatic TRAMO-SEATS runs, TRAMO-SEATS and X-12-ARIMA decom-
positions can be very similar. Maravall (1985) found analogous similarities
between decompositions obtained from X-11 and from the basic structural
model.
Notice, however, that X-11 filters cannot be exactly reproduced if attention
focuses on airline models. One reason is that X-11 filters are finite while WK
filters derived from ARIMA ( p, r, q) models with q > 0 are rational filters of
infinite length; see (2.4). Under the invertibility condition, the WK filter
converges, so (2.4) is anyway valid for periods around the centre of usual sample
lengths. For concurrent estimation, though, unknowns future observations must
be forecasted (Maravall, 1996) and, as new observations become available,
correcting the forecast errors will yield some revisions in preliminary estimates.
Hence, the use of a rational filter has an important consequence: in the process of
seasonally adjusting time series, the number of periods needed for a model-based
concurrent estimate to become a final estimator can be relatively large. Some
model-based tools that are discussed in Section 4 can help in anticipating that
delay, but practitioners have no direct control on the length of the revision period.
In Section 3, we develop a procedure that enables practitioners to control the
duration of revisions.
An obvious possibility for reducing the length of the revision period would be to
impose ARIMA ( p, r, 0) models for observed series. Yet such an approach
would limit the flexibility of ARIMA modelling in describing economic time
series properties. We thus choose to approximate infinite model-based filters with
finite ones. The wide use of the finite X-11 filters suggests that such
approximations could indeed be of practical interest. The final estimator will
then be slightly modified, in such a way that the convergence of preliminary
estimates will be faster.
Let aðBÞ denote an infinite symmetric linear filter such that
1
aðBÞ ¼ a0 þ R 1 ak ðBk þ F k Þ
k¼1 2
ð3:1Þ
where the weights ai are real, do not depend on time, and satisfy Rai ¼ 1; ai ¼ aÿi ;
and Ra2i < 1. We consider the problem of finding bm ðBÞ; bm ðBÞ ¼ bm;0 þ
Rm 1 k k
k¼1 2 bm;k ðB þ F Þ, such that
Z p
1
dða; bm Þ ¼ jaðeÿiw Þ ÿ bm ðeÿiw Þj2 dw ð3:2Þ
p 0
bmk
min R ðak ÿ bmk Þ2
;k¼0;...;m k¼0
ð3:3Þ
with respect to
m
C:1 R bmk ¼ 1
k¼0
m ÿ2k‘p
C:2 R bmk cos
k¼0 12 ¼ 0; ‘ ¼ 1; . . . ; 6
The classical analysis of revisions has been developed by Pierce (1980). We briefly
review it here, concentrating on preliminary estimates computed close to the end
of the sample and involving model-based forecasts of the unknowns. Let ^ntjtþk
denote the preliminary estimate of nt computed at time t þ k; k 0. Assuming
that Xtþk ¼ fx1 ; . . . ; xtþk g is available, then ^
ntjtþk ¼ Eð^nt jXtþk Þ. Let
nðBÞ ¼ þ nÿn Bn þ þ n0 þ n1 F þ þ nm F m þ
denote the polynomial which relates the estimator (2.4) to the innovations
at : ^nt ¼ nðBÞat . That polynomial is such that
r2 SðBÞnðBÞ ¼ hðBÞmn ðBÞ
Given (2.4), the weights nk verify
r2 hðF ÞnðBÞ ¼ Vn hn ðBÞhn ðF ÞSðF Þ ð4:1Þ
The invertibility of hðF Þ implies that nðBÞ converges in F . Like for inverted
moving average processes, the forward weights verify
1
lim nk ¼ 0
k!1
and lim
k!1
R n2j < 1
j¼kþ1
with variance
1
nt ÿ ^
V ½^ ntjtþk ¼ R n2i
i¼kþ1
ð4:3Þ
Another tool useful for the monitoring of revisions is given by the theoretical
convergence path. After a number k of additional observations, the convergence
path of the preliminary estimates can be evaluated using
V ½^
ntjtþk ÿ ^ntjt
RCðkÞ ¼ 100
V ½^nt ÿ ^ntjt
k ð4:4Þ
R
¼ 100 i¼1
n2i
1 2
Ri¼1 ni
Given an ARIMA model and its optimal decomposition, expression (4.3) gives
the theoretical revision variance, allowing inference to be made, while (4.4) helps
in anticipating the convergence delay.
Geweke (1978) and Pierce (1980) showed that, given any filter, extending the
series with MMSE forecasts minimizes the total revision error in preliminary
estimates. Hence, when bm ðBÞ is used, the m forecasts needed for concurrent
estimation can be derived from the observed series model. The resulting revisions
can be analysed like previously, the ni weights being derived from the equality
r2 SðBÞnðBÞ ¼ hðBÞbm ðBÞ
Because bm ðBÞ embodies a finite number of weights, the nðBÞ’s forward part is
finite as well and the conditions
1
lim nk ¼ 0
k!1
and lim
k!1
R n2j < 1
j¼kþ1
are verified. The quantities (4.3) and (4.4) involving either mn ðBÞ or bm ðBÞ can thus
be easily derived from any ARIMA model. Empirical estimates will be given in
the application. Notice that, since the WK filter gives the expectation of the
unobserved component conditional on the sample available, on the model selected
and, in particular on the model parameters, the expressions (4.3) or (4.4) and their
empirical counterparts cannot take into account the error in the model parameters
estimates.
After convergence, the signal extraction procedure still yields an estimation
error, called final estimation error and defined as et ¼ nt ÿ ^nt . First, assuming
observed series stationarity, a frequency domain expression for the variance of the
final error in an estimator ^ nt ¼ cðBÞxt is given by
Z p
1
V ½nt ÿ ^
nt ¼ gn ðeÿiw Þ½1 ÿ cðeÿiw Þ2 þ gs ðeÿiw Þcðeÿiw Þ2 dw ð4:5Þ
p 0
where gs ðeÿiw Þ and gn ðeÿiw Þ are the spectra of the seasonal and nonseasonal
components. When cðBÞ is the WK filter (2.4), the variance of the final estimation
error turns out to be the variance of the ARMA process
hðBÞzt ¼ hn ðBÞhs ðBÞbt ð4:6Þ
with V ðbt Þ ¼ Vs Vn (Maravall, 1996). Bell’s (1984) assumption A implies that (4.6)
is still valid in the nonstationary case so that the final estimation error remains
finite. When bm ðBÞ is used, the symmetry of bm ðBÞ and the constraint C1 imply
that ½1 ÿ bm ðBÞ has at least two unit roots at the zero frequency; similarly, C2
implies that bm ðBÞ has at least two unit roots at the seasonal frequencies. For the
airline model, these properties are sufficient to ensure that the integral (4.5) is well
defined and that the final estimation error obtained with constrained filters is still
finite.
truncation lengths 36, 24 and 18; as we shall see, shorter lengths would not be
appropriate.
Applying (2.4) to model (5.1)–(5.3) yields the analytic expression of the WK
filter. The finite approximations are then computed using the results in Appendix
I; in particular, applying expression (7.1) to the coefficients mk of the WK filter
yields
Figure 5a shows the resulting weights together with the original ones. No
striking discrepancy can be seen: the weights of the original filter are just corrected
so that the constraints C1 and C2 are verified for a given number of coefficients.
These corrections have, however, several consequences. First, the gains of the
approximated filters can be above 1; see Figure 5b. Second, as noticed by a
referee, some distortions can be seen in the range ð0; p=6Þ which can affect the
long-term behaviour of the seasonally adjusted series. Ideally, bm ðBÞ should be
subject to the constraint bm ðeÿiw Þ ¼ mn ðeÿiw Þ for every frequency in ð0; p=6Þ.
However, bm ðBÞ is already subject to seven constraints. We found that increasing
the number of constraints turns out to damage the filter band-pass structure: for
example, for m ¼ 18, imposing three further constraints within ð0; p=6Þ for a total
of ten constraints seriously affects the approximation, which is very distant from
the original infinite filter aðBÞ (see Appendix II). In consequence, we decided not
to impose additional constraints. The distortions within ð0; p=6Þ can be seen as a
price to pay for controlling the duration of revisions. As any drawback, it is
important to monitor them so as to check that the output series is not excessively
affected. Since the frequency range ð0; p=6Þ is related to the long-term behaviour
of the series, we suggest to perform that check by scrutinizing trends extracted
from the different outputs. In the case of the distortions being found to be too
important, we advise users to increase m, the duration of revisions considered
when approximating the infinite filter.
Figure 6 show the seasonally adjusted series obtained with bm ðBÞ; m ¼ 24; 18;
together with the WK estimator. The adjustment obtained with m ¼ 36 is not
plotted because it is indistinguishable from the unconstrained one. The
convergence of the preliminary figures is now completely under control: the
estimators become final as, respectively, 36, 24, and 18 observations are added.
Small discrepancies are visible with m ¼ 24; they become slightly more
important with m ¼ 18. The concurrent estimates at the sample end are
computed with respectively 36, 24 and 18 forecasts obtained from the observed
FIGURE 5. FRTP Seasonal Adjustment Filters- (a) weights and (b) frequency transfer function
series model (5.1). Notice that the differences between the seasonal adjustments
cannot be described as a white noise process: the ratio of the difference between
the frequency transfer functions of the unconstrained and approximated filters
to the series spectrum would have to be constant over ½0; p, a condition that, in
general, will not be met.
Figure 7 shows trends extracted from the seasonally adjusted series obtained
with the unconstrained filter and with the three approximations, m ¼ 36; 24; 18.
These long-term trends are outputs of the Hodrick–Prescott filter (Hodrick and
Prescott, 1997) with signal to noise ratio set to 1/1600, mainly to have a large
enough low-pass band. As can be seen in Figure 7, the four trends are
indistinguishable: the discrepancies are less than 10ÿ3 in absolute value. Hence,
in that example, the distortions seen in the filters gain over the frequency range
ð0; p=6Þ do not affect the long-term movements.
Table II displays the distance between WK and the truncated filters, the final
estimation error variance (4.5) and the theoretical revision variance (4.3). It is
seen that dðmn ; bm Þ lies between 10ÿ4 for m ¼ 36 and 0.0172 for m ¼ 18, and that
the truncation increases the final estimation error variance by 2% with m ¼ 36
and by 10% with m ¼ 18. The effect on the theoretical size of total revisions is
less clear, since they can be either amplified to 0.131 with m ¼ 24 instead of
0.107, or reduced to 0.101 with m ¼ 18; that is because our procedure focuses on
the convergence period, not on the revision size. Figure 3 shows that for the
months of year 1992, shorter filters do not amplify the revisions actually
recorded.
TABLE II
FINITE APPROXIMATION VS WK ESTIMATOR
FILTER DISTANCE AND MODEL-BASED VARIANCE OF ESTIMATION ERRORS
FRTP
Filter dðbm ; mn Þ V ½nt ÿ ^
nt V ½^
nt ÿ ^
ntjt
WK – 0.111 0.107
m = 36 0.0001 0.113 0.115
m = 24 0.0013 0.116 0.131
m = 18 0.0172 0.124 0.101
TABLE III
ACCEPTABLE DURATION OF REVISION PERIOD IN AIRLINE MODEL
h12
h1 0.8 0.6 0.4 0.2
0.8 m1 ¼ 72 m1 ¼ 36 m1 ¼ 24 m1 ¼ 12
m2 ¼ 84 m2 ¼ 48 m2 ¼ 24 m2 ¼ 24
0.6 m1 ¼ 72 m1 ¼ 36 m1 ¼ 24 m1 ¼ 15
m2 ¼ 84 m2 ¼ 48 m2 ¼ 28 m2 ¼ 24
0.4 m1 ¼ 72 m1 ¼ 36 m1 ¼ 24 m1 ¼ 15
m2 ¼ 84 m2 ¼ 48 m2 ¼ 28 m2 ¼ 24
0.2 m1 ¼ 72 m1 ¼ 36 m1 ¼ 24 m1 ¼ 13
m2 ¼ 84 m2 ¼ 48 m2 ¼ 27 m2 ¼ 24
m1 = minimum m-value such dðbm ; mn Þ < 0:01
m2 = minimum m-value such dðbm ; mn Þ < 0:005
The problem of giving a signal extraction interpretation to X-11 filters has been
the subject of debate in the statistical literature. For instance, Burridge and
Wallis (1984) derived ARIMA models for unobserved components leading to a
seasonal adjustment signal extraction filter very close to the X-11 default
adjustment filter. That X-11 default is the one obtained by convoluting 3 3
and 3 5 seasonal moving average with the 13-term Henderson moving
average. In this section, we show that the X-11 default filter is actually closer
to some finite approximations of the unconstrained WK filter. Regarding
historical estimation, the model specification proposed by Burridge and Wallis
(1984) was
Vs =Vn ¼ 0:017
Denoting mX ðBÞ the X-11 default adjustment filter, this decomposition leads to
the distance dðmn ; mx Þ ¼ 0:0036, which was a minimum in Burridge and
Wallis’ model specification search. Using (6.1) and (7.1), we computed for
m ¼ 24; . . . ; 100 the filters bm ðBÞ as optimal approximations to mn ðBÞ according to
(3.3). For every m value, we checked the distance between bm ðBÞ and mX ðBÞ and
the ratio rm defined as
dðbm ; mX Þ
rm ¼ 100 ð6:2Þ
dðmn ; mX Þ
We found that an improvement of order 40% in the approximation of the
X-11 default filter can be obtained by truncating the WK filter at length
m ¼ 43, with a final distance at dðb43 ; mX Þ ¼ 0:0023. Figure 8 shows that this
improvement is related to the capacity of truncated filters to have gains higher
than 1.
7. CONCLUSION
In this paper, we have built finite approximations to infinite filters and we have
applied them to the control of the revision period in the seasonal adjustment
problem. These approximations are obtained via simple corrections to the original
signal extraction filters that are given in Appendix I. Our procedure enables
statisticians to produce seasonally adjusted series not subject to revisions after a
given number of additional observations. However, as truncating filters implies
where k and l‘ ; ‘ ¼ 1; . . . ; 6 are the usual Lagrange operators. The solution, b , is such that
6
dLðb Þ
¼ ÿ2ðak ÿ bk Þ þ k þ R l‘ wk‘ ¼ 0
dbk ‘¼1
m
dLðb Þ
¼ R bk ÿ 1 ¼ 0
dk k¼0
m
dLðb Þ
¼ R bk wk‘ ¼ 0 ‘ ¼ 1; . . . ; 6
dl‘ k¼0
m
gives the coefficients of the frequency transfer function bm ðeÿiw Þ ¼ Rk¼0 bk cos kw which
solves (3.3).
Let us consider b18 ðBÞ, the approximation with 18 coefficients of the infinite WK filter
related to models (5.1)–(5.3). That approximation is subject to the seven constraints
implied by C1 and C2 (see (3.3)), and its distance to the unconstrained filter is
dðmn ; b18 Þ ¼ 0:0172: Let us now impose three more constraints at the frequencies
w ¼ 2kp=ð4 12Þ; k ¼ 1; 2; 3; all within ½0; p=6: The 18 coefficients of b18 ðBÞ are now
subject to ten linear constraints. The distance to the original filter becomes
dðmn ; b18 Þ ¼ 0:0397, i.e. more than twice the distance obtained when only C1 and C2 were
considered. A check of the frequency transfer function would show that the resulting
approximation is, as expected, very satisfactory within ð0; p=6Þ, but it suffers of strong
distortions at the other frequencies that make it no longer suitable for seasonal adjustment.
Clearly, the more constraints, the larger the deviation from the original filter and the less
satisfying the decomposition is.
ACKNOWLEDGEMENTS
This paper was written within the framework of a study on Seasonal Adjustment
methods, conducted for EUROSTAT (contract nr. 8221011/6 to the first author)
while the second author was in detachment at EUROSTAT. The ideas expressed
here are those of the authors and do not necessarily reflect the positions of
EUROSTAT and of the JRC. Thanks are due to the Members of EUROSTAT’s
internal task-force on seasonal adjustment for their remarks and suggestions, and
to an anonymous referee for very useful comments that have substantially
improved the paper.
NOTES
Christophe Planas Institute for Systems, Informatics and Safety, Joint Research
Centre of European Commission, TP361, Via E. Fermi, 1, I-21020 Ispra, Italy,
email: christophe.planas@jrc.it. Please send correspondence to first author.
Raoul Depoutot raoul.depoutot@insee.fr, INSEE, France.
REFERENCES
BELL, W. R. (1984) Signal extraction for nonstationary time series. The Annals of Statistics 12(2),
646–64.
—— and MONSELL, B. C. (1992) X-11 symmetric linear filters and their transfer functions. Bureau of
the Census, Research Report n. RR 92/15,Washington.
BOX, G. E. P. and JENKINS, G. M. (1970) Time Series Analysis: Forecasting and Control. San Francisco:
Holden Day.