Sie sind auf Seite 1von 17

Seasonal Adjustment by Signal Extraction

Gerhard Thury*)

Zusammenfassung

In der vorliegenden Arbeit wird untersucht, ob die auf einem ARIMA-Modell basierenden
Answ der Saisonbereinigung in der praktischen Anwendung zu verlw Ergebnis-
sen f0hren als empirische Verfahren, wie etwa CENSUS X-11. Um diese Frage zu kl~iren,
wird eine Saisonbereinigung von 6sterreichischen Arbeitsmarktdaten mit beiden Gruppen
von Verfahren durchgefQhrt. Es werden dafQr bewuSt Zeitreihen gew~ihlt, die eine stark un-
terschiedliche Stabilit~it der Saisonfigur aufweisen. Bei diesem Vergleich zeigt sich, dab auf
einem Modell basierende Ans~tze bei Zeitreihen mit relativ stabiler Saisonfigur zu eindeutig
besseren Ergebnissen fL~hren als empirische Verfahren. Diese Uberlegenheit kann jedoch
verlorengehen, wenn das als Ausgangspunkt dienende ARIMA-Modell schlecht spezifiziert
ist. FQr Zeitreihen mit stark schwankender Saisonfigur fQhren beide Gruppen von Saison-
bereinigungsverfahren zu w Ergebnissen.

Seasonally adjusted data play a prominent role in assessing the current state of economic
activity, but many users of seasonally adjusted data do not understand the methods by
which those data are produced. It is probably unreasonable to expect a statistically unso-
phisticated person to understand the methods, but even trained statisticians are often mys-
tified by these procedures. For example, CENSUS X-11, the most widely used method, ap-
plies moving averages in order to obtain seasonally adjusted figures. The basic idea behind
moving averages is rather simple; they will smooth irregular fluctuations in the data. But the
way in which they are used in X - l l (see S h i s k i n - - Y o u n g - - M u s g r a v e , 1967) is extremely
complex. Above all, the theoretical statistical underpinnings for this procedure are not un-
derstood at all. The only justification for its use is the fact that, in the majority of cases,
X-11 produces relatively satisfactory results.

1. Theoretical foundations of seasonal adjustment

The objective of a seasonal adjustment method is to decompose an observed series Xt


into trend (~), seasonal (St), and irregular component (It), i. e.,

(1) x , = T , + S , + L .

*) Financial support by the Jubil~umsfonds der Oest~rreichischen Nationalbank under grant No. 2203
is acknowledged. I would like to thank Peter Burman, Johannes Ledolter, Fritz Schebeck and Erich
StreiBler for their helpful comments.

191
We concentrate here on the additive model, because a nonadditive decomposition can
usually be converted into an additive one by an appropriate power transformation of the
raw data. It is surprising, however, that this decomposition into trend, seasonal, and irregu-
lar component has been so widely used without defining these components in a rigorous
manner. Statements in the literature about these components are vague, being descrip-
tions rather than definitions. Because of this lack of precision the statistical properties of
these methods are difficult to assess from a theoretical point of view. Therefore, mathemat-
ical statisticians have made numerous attempts to replace these empirically developed ap-
proaches to seasonal adjustment by theoretically well defined model-based procedures. A
good survey of this development is given in Bell -- Hillmer (1984). In the present paper, we
only deal with one of these approaches, namely the ARIMA model-based approach, which
gained considerable popularity in recent years. The publication of the book by Box - - Jen-
kins (1970) was a catalyst for the development of this approach, because it offered rela-
tively simple and yet efficient techniques for modelling seasonal time series. Pierce (1978)
was one of the first to start using ARIMA models for purposes of seasonal adjustment.
Wecker (1978) suggested an extension to Pierce's approach. Box -- Hillmer -- Tiao (1978)
started with the model

(1--B)(1--B ~2) Xt = ( 1 , - - O ~ B ) ( 1 - - O ~ 2 B ~2) at

which is a very common model for monthly data, and derived models for the trend, sea-
sonal, and irregular components that are consistent with this overall model. To achieve this
goal certain assumptions had to be made including~one that calls for the maximization of
the variance of the irregular component. This implies the minimization of the variance of the
trend and the seasonal component. This approach was later extended to a more general
class of ARIMA models by Burman (1980) and Hillmer - - Tiao (1982). Here we summarize
the approach by Hillmer and Tiao, because it is easier to understand for the statistically un-
sophisticated reader,

The objective of the ARIMA model-based seasonal adjustment approach is also to achieve
a decomposition of the form (1). But here, quite in contrast to the empirical procedures
such as X-11, the trend, seasonal, and irregular components are rigorously defined. It is as-
sumed that they follow ARIMA models, i. e.,

(2) ~ r (B) Tt = Or (B) bt,

~s (8) s, = Os (B) c,,

~1(B) L = O~(8)d,,

where B is the backshift operator such that B ~ = Tt-t, each of the pairs of polynomials
{ ~ r (B), Or (B)}, {~s (B), Os (B)}, and { ~ i (B), 0 i (B)} have their roots on or outside the unit
circle and have no common roots, and bt, ct, and #t are three mutually independent white

192
noise processes, identically and independently distributed as N(O, a~), N(O, o~), and
N(O, o~), respectively. The assumption that these component models are of the ARIMA
type is absolutely harmless if the original series Xt follows an ARIMA model. After substitut-
ing the relations (2) into equation (1), one finds that Xt follows the ARIMA model

(3) ~ (B) x, = O (~) a,,

where

r = ~r(l~)~s(B)~1(B)

and where O (B) and a~ can be obtained from

O(B)O(r) o~ = Or(B)Or(F) o~ + Os(8)Os(F) O~(B)O~(F)


(4) r +r(B)~'r(F) ~s(B)~s(F) ~ + r ~167

where F = B -1. This result can be shown via the autocovariance generating function (see
Granger -- Newbold, 1977, p. 27). Now, if the different polynomials in (2) were known, it
would be easy to obtain estimates of Tt, St, and It using signal extraction methods sug-
gested by Whittle (1963) and Cleveland -- Tiao (1976). Unfortunately, however, these com-
ponent models are not known. Therefore, it is impossible to calculate the estimates of Tt,
St, and /t directly. We can, however, estimate the ARIMA model for the observable Xt
series, and use this model to determine the models for the component series. But, in order
to achieve this goal, it is necessary to make certain prior assumptions about these compo-
nent models. Thus, the question is whether it is possible to find realistic prior assumptions
which still allow an acceptable decomposition. A brief discussion of these assumptions is
given next.

1.1 Prior assumptions about the trend

Economic data often exhibit strong time trends. For relatively short time segments such a
trend might be adequately modelled by a polynomial in time. However, since trends usually
change over time, a fixed polynomial time function is inappropriate over a longer time span.
Consequently, only a stochastic trend model will be adequate in such a situation. Hillmer --
Tiao (1982) assume that the trend component, Tt, follows the nonstationary model

(5) (1 - - B) d T t = OT(B)b,,

where OT(B) is a polynomial in B of degree at most d, and bt is i. i. d. N(O, o~). Equation


(5) can be viewed as a polynomial trend model with stochastic coefficients.

193
An analysis of the properties of this model in the frequency domain is especially informa-
tive. Intuitively, the spectral density function of a trend component should be large for low
and small for higher frequencies, because a trend is defined to be a longer term phenom-
enon. Since equation (5) is nonstationary, the spectral density function is strictly speaking
not defined, but we can define a pseudospectral density function (PSDF). Writing equation
(5) in its moving average representation

Or(B) b,,
7", = (1L-B)"

this PSDF is given by


2

(1--e"O)d
eT(ei~~ OT(e--ic~ d2b,
= (l__ei(o)d (l__e--i(o)d

0_< oo_<~.

Now, it can be seen immediately that the PSDF (6) is infinite at ( o = 0 and very large for
small co. Thus, its shape is consistent with what could be regarded as a stochastic trend
component.

1.2 Prior assumptions about the seasonal component

A deterministic seasonal component of period s has the property that it repeats itself every
s periods and that the sum of any s consecutive terms is a constant, i. e.,

(7) s , = s,_,

and

U ( B ) St = const.,

where U ( B ) = 1 + B + . . . + B "-+ and, since seasonal factors are deviations from an


overall trend line, the constant can be taken as zero without loss of generality. For eco-
nomic time series, however, a fixed seasonal pattern is a far too restrictive assumption. We
need a seasonal component with stochastic properties. On the one hand, it must be ca-
pable of evolving over time. But, on the other hand, a regular seasonal pattern should be
preserved locally. In other words, U ( B ) St should be random but cluster about zero. Hill-
m e r - - Tiao (1982) use the following nonstationary model for the seasonal component:

(8) U (B) S, = e s (B) c,,

194
where Os(B) is a polynomial in B of degree at most s - - l , and ct is i. i. d. N(O, o~). In other
words, the consecutive sum of s seasonal components follows a moving average model of
order (at most) s - - l . The restriction on the order of the model is necessary to make sure
that the model of St does not allow for predictable changes in the seasonal pattern; such
changes should be absorbed in the trend component. We have also

E [ U ( B ) SI ] = OS(B) E[c,] = 0 i

Thus, a seasonal component obeying model (8) will locally follow a fixed pattern, but as
long as o~ > 0 the forecasting function of (8) will be updated continually as the time origin
advances; thus, the pattern in St will evolve over time.

Again, an inspection of the PSDF, fs (co), of the model (8) is informative. This PSDF is given
by

(9) fs (co) = ] ~ (ei~~ [ z a


OS (e i~~ OS (e -i~ _2
= U ( e . O ) U ( e -~') ~'~,

0_< co _ :T.

It can be shown that fs (co) has infinite power at the seasonal frequencies and relatively
small power away from the seasonal frequencies.

1.3 Restrictions implied by the data

The information available from the data, embodied in the known ARIMA model for Xt, puts
additional restrictions on the possible models for Tt, S;, and /t. The assumptions made in
the previous sections imply that 9~ (B) in {3) contains the factor ( l - - B ) # from the trend
component and the factor U(B) from the seasonal component. Moreover, it is required
that ~1 (B), the autoregressive polynomial of /t in (2), has no common roots with either
( l - - B ) d or U(B). Otherwiese,/t would still contain systematic factors which should be ab-
sorbed into Tt and St, respectively. Thus, 99 (B) must be of the form

(1o) r = (1 - B) '~ U ( B ) # , ( B ) ,

where the three factors on the right hand side have no common roots. Since we can esti-
mate 9~ (B) from the data, it is possible - - assuming that a decomposition exists - - to de-
termine uniquely the autoregressive polynomials for Tt, St, and/t. A simple example may il-
lustrate this. Let the estimated ARIMA model for Art be

( 1 - - B 12) Xt= ( 1 - - 0 1 2 B 12) at.

195
Thus, we have

r (B) = r (B) r (B) ~i (B)


= (1 - - B TM)
= (1 - - B ) (1 + B A- . . . + B 1')

or

~T(B) = (1 - - B),

~s(B) = (1 + B + . . . + B 11) = U(B),

~l(O) = 1.

The autoregressive polynomials of the three component models can indeed be determined
uniquely. Additionally, making use of our above assumptions about the trend and the sea-
sonal components, the relation (4) can be written as

(11) 0 (B) 0 (F) ~2 OT (B) Or (V) o2b + Os (") es (F) _~ + 01 (8) Oi (F) ~
~ ( ~ ) ~ ( v ) v" = (1--8)~ (1--F)" V(B)-O-(~ oc ~,(8---j-~,~u~.

It is considerably more difficult to determine the moving average polynomials of the compo-
nent models and to calculate the corresponding innovation variances a~, a2c, and aw Bur-
man (1978) suggests that the left hand side of (11) should be expressed as a polynomial
quotient and a remainder. The polynomial quotient is then identified with the irregular com-
ponent, while the remainder can be partitioned into two partial fractions which are identified
with the trend and seasonal components, respectively. Thus, assuming that ~ (B) takes the
form (10), we may perform a partial fraction decomposition of the left hand side of (11) to
yield

0 (B) 0 (F) a2 = Qr (B, F) Qs (B, F) QI (B, F)


(12) 9~(-~-~) ) ~ (1--B)a(1--F) d+ U(B) U(F) + ~ z ( B ) ~ I ( F ) '

where QI (B, F) is the polynomial quotient and Qr(B, F) and Qs (B, F) are the partial frac-
tions of the remainder. They are given by

d--1
QT(B,F) = q o r + ~ q i r ( B i + F i)
i=1

and

s--2
Q s ( B , F ) = qos + ~, q i s ( B i + Fi) 9
i=1

196
1.4 A canonical decomposition

Unfortunately, however, the information embodied in the known model for Xt will generally
not be sufficient to uniquely determine the models for Tt, St, and It. To achieve this goal,
an additional assumption is necessary. Intuitively, it seems desirable to extract as much
variability as possible from the trend and the seasonal component, because these compo-
nents should be as deterministic as possible while remaining consistent with the informa-
tion in the data. Thus, Hillmer -- Tiao (1982) suggest to choose a decomposition that maxi-
mizes the innovation variance o~,. They call such a decomposition that maximizes o3 sub-
ject to the restrictions in (11) a canonical decomposition. Such a canonical decomposition
has several attractive properties:

- - It is unique.

-- It also minimizes the innovation variances o~ and O'2c.Since the randomness in Tt and St
arises from the sequences {bt} and {ct}, respectively, minimizing their variances makes
the trend and seasonal components as deterministic as possible while remaining con-
sistent with the information in the data.

- - It can be shown that the trend and seasonal components corresponding to any other
choice of dw are just the canonical trend and seasonal components plus some white
noise sequences added to them. Thus, the choice of a decomposition other than the ca-
nonical one would produce trend and seasonal components with unnecessary random
fluctuations.

--Also the variance of U(B)St is minimized for the canonical decomposition. Since
E [ U ( B ) St] = 0 a small value for the variance of U(B) St will guarantee that the sum of s
consecutive seasonal component terms will remain close to zero.

Given the stochastic structure of the component models, standard signal extraction meth-
ods (see Whittle, 1963, Cleveland - - Tiao, 1976) can be applied to obtain the minimum
mean square error estimates of the trend and seasonal components. They are given by

(13) ~, = w~(8? x~

and

~, = ws (B) X,,

where

Wr(B) = og 0 (B) O (F) Or(B) Or(F)


a~ 9~(BJ qp(FJ ~r(B) ~ ( V )

197
and

Ws (B) = aac O (B) O (F) Os :B) Os :F)


O2aqo (B) qo (F) ~s (B) ~s (F)"

The irregular component, /t, finally, follows from

i,

2. Application and comparative evaluation


In this section, we apply signal extraction for the seasonal adjustment of three Austrian la-
bour market series. We use two versions of this model-based approach, namely that o f
B u r m a n (1980) and that of H i l l m e r - - Tiao (1982), and compare the results with the out-
come from X-11 ARIMA(1). The series under analysis are the total number of unemployed,
the unemployment rate, and the number of unemployed in agriculture. The first two series
play a prominent role in the assessment of the current state of the Austrian economy, while
the third is of negligible importance in that respect. Economic significance, however, is only
of secondary importance for the purposes of this study. Our choice of series is based on
the structure of the ARIMA models, and especially on the size of the parameters in the sea-
sonal moving average polynomials of these models. The estimated ARIMA models for the
above three series are given by

(1--B)(1--B ~2) w l , = ( 1 - - 0 . 4 6 9 B 12) a~t,

(1--B)(1--B 12) w2t = ( 1 - - 0 . 6 9 1 B 12) a2t,

( 1 - - 0.836 B) ( 1 - - B 12) w3t = ( 1 - - 0 . 7 9 6 B 12) a3t,

where

wl t = total number of unemployed with transformation power p--0.25,


w2 t = natural logarithm of the unemployment rate, and
w3 t = number of unemployed in agriculture with transformation power p = 0.50.

All estimated coefficients are highly significant, and the ait's a r e definitely white noise se-
quences. The different values for the transformation power p are necessary in order to
achieve an additive decomposition into trend, seasonal, and irregular component(2). We
note that all three ARIMA models have similar and simple structure. Only the size of the co-
efficient of B 12 on the right hand side varies from model to model. However, it is exactly this
coefficient, that is of greatest importance in seasonal adjustment. It contains information

198
about the stability of the underlying seasonal pattern. A value close to unity indicates a
stable seasonal pattern. In such a situation, it is appropriate to apply long filters in order to
extract the seasonal component. It is shown in C l e v e l a n d - - Tiao (1976) that the filters ap-
plied in X-11 (and, accordingly, also those in X - l l ARIMA) are valid for a series with
812 = 0.4. From our above estimates of ~712we would expect, therefore, that in the case of
the total number of unemployed the ARIMA model-based approach and the X - l l proce-
dure should lead to similar results. But, for the unemployment rate and especially for the
number of unemployed in agriculture, the differences should increase substantially. The ra-
tionale for analysing this third series in the present study is to investigate if this theoretical
result can be verified empirically. Therefore, we turn now to an analysis of the estimates for
the trend, seasonal, and irregular components obtained by the three above mentioned sea-
sonal adjustment procedures.

2.1 The trend

The trend component in the decomposition (1) should not exhibit random fluctuations.
Thus, relative smoothness of the extracted trend is a requirement which a good seasonal
adjustment procedure must meet. A widely used measure of smoothness is the root mean
square of the first differences of the extracted trend. The larger this measure, the rougher
the extracted trend. Table 1 contains this measure for the trends of the total number of un-
employed, the unemployment rate, and the total number of unemployed in agriculture de-
rived by two variants of signal extraction (Burman and Hillmer-Tiao) and by X-11 ARIMA.

Following K u i p e r (1978), the measures of smoothness are calculated for the full period
1970 to 1984, the historical period (i. e., after deleting the first three and the last three

Table 1
Measures of smoothness for the trend
Method Full period Historical period Current period
Total number of
unemployed
Burman 0.1027 0.0871 0.1309
Nillmer-Tiao 0.1140 0.0945 0.1395
X-11 ARIMA 0.1014 0.0860 0.1296
Unemployment rate
Burman 0.0325 0.0290 0.0376
Hillmer-Tiao 0.0323 0.0290 0.0370
X-11 ARIMA 0.0278 0.0127 0.0113
Number of unemployed
in agriculture
Burman 0.0375 0.0275 0.0360
Hillmer-Tiao 0.0248 0.0214 0.0273
X-11 ARIMA 0.0127 0.0108 0.0157

199
years) for which the estimates may be considered final in a statistical sense, and the cur-
rent period (i. e., the last three years). The results in Table 1 fully agree with our theoretical
expectations. For the total number of unemployed, signal extraction methods and X - l l
ARIMA give practically identical results. This is simply a consequence of the fact that for
this series the seasonal moving average parameter @12 is approximately equal to 0.4. For
such a value of ~12, the filters applied in signal extraction and in the standard version of
X - l l ARIMA are rather similar. The small observed differences between the two variants of
signal extraction stem probably from a different treatment of outliers and from differences
in the estimation routines for the underlying ARIMA models. For the unemployment rate
and, especially, for the number of unemployed in agriculture, where the seasonal moving
average parameter ~ 2 is significantly larger than 0.4, the results of signal extraction and of
X-11 ARIMA begin to diverge substantially. Somewhat perplexing is perhaps the fact that,
at first glance, these results seem to diverge in the wrong direction. The trend estimates
derived by signal extraction methods are significantly rougher than those of X - l l ARIMA.
Moreover, these differences in smoothness are largest for the current period, which is the
most interesting one for the business cycle analyst. Some reflection reveals that this. sur-
prising result is consistent with our theoretical expectations. Apparently, if ~ 2 > 0 . 4 , the
standard version of X-11 ARIMA exhibits a tendency to oversmooth the trend component.
The larger @~2 becomes, the stronger is this tendency to oversmooth. The explanation for it
is quite trivial. In situations, where ~ 2 is significantly larger than 0.4, X-11 ARIMA creates a
too flexible seasonal pattern, because the applied filters are not long enough. As a conse-
quence of this, the seasonal component contains predictable fluctuations, which should be
absorbed by the trend. Or in more technical terms, the order of the moving average poly-
nomial of the seasonal component, P)s (B), is higher than s - - l .

The fact that oversmoothing of the trend can occur, has important implications for applied
research work. In comparative studies of seasonal adjustment procedures (for example,
Newbold - - Thury, 1984) it is standard practice to use measures of smoothness as a kind
of test statistic. Thereby, it is assumed that a method with a smoother trend estimate is to
be preferred to a method with somewhat rougher estimates. The possibility of oversmooth-
ing, however, casts certain doubts on the usefulness of such comparisons.

2.2 The seasonal component

As already briefly mentioned in the previous section, the observed differences in the out-
comes of signal extraction and X - l l ARIMA are primarily due to the varying degrees in the
rigidity of the extracted seasonal patterns. Therefore, we turn now to a detailed analysis of
these seasonal components.

Table 2 displays inequality coefficients for the different seasonal components extracted by
the various methods. These coefficients are defined as

200
n
1 I Sit--Sltl . 100,
ICq = n ~ Xt '
t=l

where Sit and Sit are the estimated seasonal components coming from methods i and j
respectively, and Xt denotes the original series. It can be seen immediately that the differ-
ences between signal extraction and X-11 ARIMA are minimal for the total number of unem-
ployed. For the unemployment rate and especially for the number of unemployed in agricul-
ture, however, the discrepancies between the different approaches become substantial,
exactly as one would have expected it to happen from theoretical considerations. In the tat-
ter two cases, systematic movements, which belong to the trend, are attributed to the sea-
sonal component leading to an unnecessarily large standard deviation of the monthly sea-
sonal factors.

Table 3 shows the standard deviations of the seasonal factors calculated for each month
separately for the period 1970 to 1984. The fact that the variability of the extracted seasonal
factors decreases for series with a more stable seasonal pattern, is self-evident. As ex-
pected, we also observe differences in the variability of the monthly seasonal factors. For
the unemployment rate and the number of unemployed in agriculture the seasonal compo-
nent extracted by the two signal extraction approaches exhibits much less variation than
that of X-11 ARIMA. We apply the WlLCOXON Signed Rank Test in order to investigate
whether these differences are significant. We obtain a WILCOXON T § statistic of 55.0 for
the unemployment rate and 78.0 for the number of unemployed in agriculture. The critical
value of this statistic at the 5 percent significance level is 48.0. Thus, the differences be-
tween model-based approaches and empirical methods are highly significant for series with
a relatively stable seasonal pattern. Finally and somewhat surprisingly, we also find signifi-
cant differences in the variability of the extracted seasonal components for the number of

Table 2
Inequality coefficients for the seasonal components
Method Full period Historical period Current period
Total number of unemployed
Burman 0.000 0.000 0.000
Hillmer-Tiao 0.186 0.000 0.141 0.000 0.178 0.000
X-11 ARIMA 0.248 0.148 0.000 0.186 0.145 0.000 0.240 0.110 0.000
Unemployment rate
Burman 0.000 0.000 0.000
Hillmer-Tiao 0.476 0.000 0.550 0.000 0.297 0.000
X-11 ARIMA 1.841 1.792 0.000 1.891 1.851 0.000 1.276 1.269 0.000
Number of unemployed in agriculture
Burman 0.000 0.000 0.000
Hillmer-Tiao 0,489 0.000 0.459 0.000 0.340 0.000
X-11 ARIMA 1.089 1.172 0.000 0.680 0.745 0.000 2.173 1.915 0.000

201
Table 3
Standard deviations of monthly seasonal factors
Total number of
unemployed
Burman 0.162 0.112 0.106 0.128 0.098 0.106 0.052 0.094 0.095 0.069 0.068 0.095
Hillmer-Tiao 0.100 0.061 0.068 0.104 0.074 0.079 0.030 0.067 0.065 0.039 0.046 0.074
X-ll ARIMA 0.061 0.058 0.082 0.128 0.060 0.091 0.029 0.076 0.071 0.042 0.052 0.100
Unemployment
rate
Burman 0.013 0.016 0.010 0.021 0.010 0.041 0.020 0.015 0.012 0.006 0.011 0.012
Hillmer-Tiao 0.015 0.016 0.009 0.022 0.011 0.043 0.021 0.015 0.011 0.006 0.012 0.013
X-11ARIMA 0.016 0.024 0.005 0.044 0.002 0.056 0.033 0.022 0.017 0.007 0.023 0.017
Number of
unemployed in
agriculture
Burman 0.007 0.010 0.020 0.010 0.003 0.005 0.008 0.010 0.012 0.014 0.002 0.006
Hillmer-Tiao 0.008 0.009 0.015 0.011 0.000 0.005 0.009 0.009 0.011 0.014 0.005 0.007
X-11ARIMA 0.013 0.014 0.032 0.018 0.008 0.008 0.013 0.015 0.021 0.024 0.007 0.012

unemployed. Here, however, the differences do not occur between the two signal extrac-
tion methods and X - l l ARIMA, but between the Burman programme on the one side and
the Hillmer-Tiao approach and X - l l ARIMA on the other. This fact might be interpreted as
indication that the Burman programme has some difficulties in extracting the correct sea-
sonal component for a series that exhibits a somewhat erratic seasonal pattern. The im-
portance of this problem should not be overrated, however. First, the differences between
the various programmes, although significant, are relatively small. Second and more impor-
tant, one should not draw general conclusions from only one example. One would have to
analyse several series with unstable seasonal patterns before one can reach a final conclu-
sion on this question(3).

2.3 The irregular component

The irregular component which is extracted in a seasonal adjustment run must be white
noise or a low order moving average process. Otherwise, it would contain systematic fluc-
tuations which should be captured by the trend or the seasonal component. Our test for
autocorrelation of the residuals is based on two portmanteau test statistics suggested by
L j u n g - - B o x (1978) and P i e r c e (1978), respectively.

The Ljung-Box Q-statistic reveals immediately that the irregular component extracted by
X-11 ARIMA is definitely not white noise for any of the variables under analysis. Moreover,
an inspection of the individual autocorrelation coefficients shows that it is not a low order
moving average process either. On the contrary, it is a rather complicated high order mov-
ing average process, indicating the existence of additional systematic factors which should

202
Table4
Test for autocorrelation and residual seasonality in the irregular component
Method Ljung-Box (2 Pierce Qs
Total number of unemployed
Burman 28.90 19.55" *
Hillmer-Tiao 26.40 17.46* *
X-11 ARIMA 53.28"* 9.22*
Unemployment rate
Burman 37.41 * 7.69*
Hillmer-Tiao 39.07* 8.91 *
X-11 ARIMA 42.39* * 7.73*
Number of unemployed in agriculture
Burman 19.30 2.57
Hillmer-Tiao 18.03 1.72
X-11 ARIMA 44.51 * * 3.08
. . . Significant at the 5 percent level,
9* . . . significant at the 1 percent level.
24
Ljung-Box Q = n ( n + 2 ) f , r ~ ( k ) / ( n - - k ) distributed as 2"2 with 23 degrees of freedom,
k=l
3
Pierce Qs = n ( n + 2 ) f. r ~ ( 1 2 j ) / ( n - 1 2 j ) distributed as 2"2 with 2 degrees of freedom.
j=l

be absorbed into the trend or the seasonal component. The irregular component extracted
by the model-based approaches corresponds to the expectations implied by statistical the-
ory. It is white noise for the total number of unemployed and for the number of unemployed
in agriculture. For the unemployment rate, it is a first order moving average process with
the lag 1 autocorrelation coefficient amounting to --0.293 for the Hillmer-Tiao approach and
to --0.324 for the Burman method.

The Pierce Qs-statistic, which is an indicator for the presence of residual seasonality in the
irregular component, also provides valuable information. We obtain significant Qs-values
for the total number of unemployed and for the unemployment rate indicating residual sea-
sonality in the adjusted figures of these two series. An inspection of the individual auto-
correlation coefficients, which are used for the calculation of this statistic, reveals that the
majority of them is negative. This implies that the significant Qs-values are all caused by
"overadjustment". This phenomenon, however, is an unavoidable consequence even of
optimal seasonal adjustment procedures. Therefore, it cannot and should not be used as
argument against a specific seasonal adjustment procedure. For a detailed discussion of
this problem refer to Grether - - Nerlove (1970). However, it is interesting to observe that
the degree of overadjustment apparently varies with the rigidity of the seasonal pattern
present in the series to be adjusted. If this seasonal pattern is relatively unstable, overad-
justment is rather severe. For series with a stable seasonal pattern, the problem of overad-
justment loses considerably in importance.

203
3. Conclusions

In the present paper, model-based approaches to seasonal adjustment are compared with
the standard version of X - l l ARIMA. It~is well known that X-11 ARIMA is an optimal proce-
dure for series where the seasonal moving average parameter 012 is approximately 0.4. The
size of this seasonal moving average parameter is therefore applied as selection criterion
for the series which are included into this investigation. We use three series from the Aus-
trian labour market, namely the total number of unemployed (0,2-~ 0.5), the unemployment
rate (012---0.7), and the number of unemployed in agriculture (0,2=0.8). The latter series
is selected because of its stable seasonal pattern.

Our empirical findings are in close agreement with the expectations from statistical theory.
For the total number of unemployed, the model-based approaches and X - l l ARIMA pro-
duce very similar results. This simply follows from the fact that the underlying model for
X-11 is

(1--B)(1--B '2) Xt = (1-- O~ B) (1-- O,2 B ~2) at

with 0,2=0.4, which is not very different from our estimated ARIMA model for this series.
Only the irregular :component deviates substantially between the model-based approaches
and X - l l ARIMA. I t i s definitely not white noise or a low order moving average process in
the case of X-11 ARIMA. For the unemployment rate, a series with a more stable pattern,
the outcomes from the model-based approaches and X-11 ARIMA are significantly differ-
ent. The cause for these differences lies in the length of the filters, which are applied to ex-
tract the seasonal component. X-11 ARIMA uses here filters which are definitely too short.
Therefore, this programme leads to a seasonal component that varies too much over time
and captures factors, which would belong into the trend or irregular component. As a con-
sequence, this results either in oversmoothing of the trend component or in nonrandom-
ness of the irregular component, what is reflected clearly in our calculated measures of
smoothness and randomness. The results for the number of unemployed in agriculture, a
series with a slightly more stable seasonal pattern than the unemployment rate, reveal this
rather problematic property of X-11 ARIMA even more clearly.

In concluding, we would like to point out some implications of these empirical results for
applied seasonal adjustment. One of the great advantages of model-based approaches to
seasonal adjustment is the fact that the underlying statistical assumptions are specified ex-
plicitly. In contrast, X-11 and other empirical methods are not based upon statistical theory,
so that it is difficult to judge them on theoretical grounds. Consequently, :improvements are
difficult to achieve with ~mpirical methods. One of the central problems of seasonal adjust-
ment is its inherent arbitrariness. As discussed in the theoretical part of this paper, it is
necessary to make some arbitrary assumptions in order to obtain a unique decomposition
into trend, seasonal, and irregular component. Unfortunately, every seasonal adjustment
procedure, be it a model-based approach or an empirical method, is faced by this problem.
But again, model-based approaches are at an advantage here. The assumptions made to

204
overcome this inherent arbitrariness are specified explicitly, and great efforts are made to
justify this particular choice. Thus, if someone disagrees with these assumptions, they may
either make their own assumptions or refrain from seasonal adjustment. With empirical
methods like X-11, the situation is extremely unsatisfactory in this respect. They also have
to deal somehow with the same kind of arbitrary choices, but never make it clear how this
is actually done. Therefore, it is impossible to judge whether these assumptions are rea-
sonable.

Finally, a seasonal adjustment procedure must be consistent with the data, and it should be
relatively flexible. Model-based approaches use models for the trend, the seasonal, and the
irregular component, which are constrained to satisfy equation (4) above. Consequently,
provided that the ARIMA model for Xt is appropriate, the estimates of the trend, the sea-
sonal, and the irregular component are consistent with the information in the data. It may
happen, however, that for certain models no feasible decomposition exists. In general, it is
possible to find another ARIMA model, for which a decomposition exists. If not, this implies
that the series in question should not be seasonally adjusted. Model-based approaches,
moreover, exhibit a high degree of flexibility, because for each different series an ARIMA
model of appropriate form can be identified and estimated using the relevant data. In X-11
ARIMA, one of the most advanced empirical methods, this is also done. However, this esti-
mated model is not used to obtain a decomposition, which is consistent with the data, but
it is only applied to provide forecasts and backcasts of the series to be adjusted. This al-
lows the use of symmetric filters at both ends of the series. Consequently, X-11 ARIMA,
like the other empirical methods, wilt be consistent with the information in the data only by
chance. For our analysed series, this is only the case for the total number of unemployed.
For the unemployment rate and the number of unemployed in agriculture, however, the
standard version of X-11 ARIMA is inconsistent with the information in the data. This fact
has substantial negative effects on the quality of the resulting component estimates. It
might be argued, however, that the X-11 programme offers a variety of options that can be
adjusted by analysts. Admittedly, the application of nonstandard options in X-11 ARIMA
may increase the chances that the outcome of an adjustment run will be consistent with
the data for some series. But, one should be aware of the fact that the use of nonstandard
options constitutes no general solution of this problem. The number of available options is
necessarily limited, and they are selected subjectively, not objectively based on the data.

All in all, the present study provides clear evidence that model-based approaches to sea-
sonal adjustment are superior to empirical methods like X-11 ARIMA in several respects.
But these advantages may be lost if the starting point is an inadequate ARIMA model for
Xt. Hence, diagnostic checking of this model is extremely important in this context.

4. References

Bell, W., Hillmer, S., "Issues Involved with the Seasonal Adjustment of Economic Time Series", Journal
of Business and Economic Statistics, 1984, 2, pp. 291-320.

205
Box, G. E. P., Hillmer, S. C., Tiao, G. C., "Analysis and Modeling of Seasonal Time Series", in Zellner, A.
(Ed.), Seasonal Analysis of Economic Time Series, U. S. Department of Commerce, Bureau of the Cen-
sus, Washington, D. C., 1978, pp. 309-335.

Box, G. E. P., Jenkins, G. M, Time Series Analysis: Forecasting and Control, Holden Day, San Fran-
cisco, 1970.

Burman, J. P., "Comment on 'A Survey and Comparative Analysis of Various Methods of Seasonal Ad-
justment' by J. Kuiper", in Zellner, A. (Ed.), Seasonal Analysis of Economic Time Series, U, S. Depart-
ment of Commerce, Bureau of the Census, Washington, D. C., 1978, pp. 77~84.

Burman, J. P., "Seasonal Adjustment by Signal Extraction", The Journal of the Royal Statistical Society,
Series A, 1980, t43(3), pp. 321-337.

Cleveland, W. P., Tiao, G. C., "Decomposition of Seasonal Time Series: A Model for the X-11 Program",
Journal of the American Statistical Association, 1976, 71 (335), pp. 581-587.

Dagum, E. B., The X-11-ARIMA Seasonal Adjustment Method, Time Series Research and Analysis
Staff, StatisticsCanada, Ottawa, 1983.

Granger, C. W. J., Newbold, P., Forecasting Economic Time Series, Academic Press, New York, 1977.

Grether, D. M., Nerlove, M., "Some Properties of 'Optimal' Seasonal Adjustment", Econometrica, 1970,
38, pp. 682-703.

Hillmer, S. C., Tiao, G. C., "An ARIMA Model-Based Approach to Seasonal Adjustment", Journal of the
American Statistical Association, 1982, 77(377), pp. 63-70.

Kuiper, J., "A Survey and Comparative Analysis of Various Methods of Seasonal Adjustment", in Zell-
ner, A. (Ed.), Seasonal Analysis of Economic Time Series, U. S. Department of Commerce, Bureau of
the Census, Washington, D. C., 1978, pp. 59-76.

Ljung, G. M., Box, G. E. P., "On a Measure of Lack of Fit in Time Series Models", Biometrika, 1978, 65,
pp. 297-303.

Newbold, P., Thury, G., "Seasonal Adiustment of Austrian Labour Force Series", Empirica, "~984, 1t (2~,
pp. 147-204.

Pierce, D.A., "Seasonal Adjustment When Both Deterministic and Stochastic Seasonality Are Pres-
ent", in Zellner, A. (Ed.), Seasonal Analysis of Economic Time Series, U. S. Department of Commerce,
Bureau of the Census, Washington, D. C., 1978, pp. 242-269.

Shiskin, J., Young, A. H., Musgrave, J. C., "The X-11 Variant of the Census Method II Seasonal Adjust-
ment Program", U S. Department of Commerce, Bureau of the Census, Technical Paper, 1967, (15)
(revised).

206
Wecker, W. E., "Comment on 'Seasonal Adjustment When Both Deterministic and Stochastic Season-
ality Are Present' by D. A. Pierce", in Zellner, A. (Ed.), Seasonal Analysis of Economic Time Series,
U. S. Department of Commerce, Bureau of the Census, Washington, D. C., 1978, pp. 274-279.

Whittle, P., Predictions and Regulations by Linear Least-squares Methods, Van Nostrand, Princeton,
1963.

5. Notes
(1) The following thee computer programmes are used in this comparison:

Signal Extraction Seasonal Adjustment Program of the Bank of England. (This programme is based on
the approach of Burman, 1980.)

ARIMA Model-Based Seasonal Adjustment Program written by S. Hillmer. (This programme is based
on the approach of Hillmer - - Tiao, 1982.)

X-11-ARIMA Program of Statistics Canada. (This programme is documented in Dagum, 1983.)

(2) The possibility incorporated in the seasonal adjustment programme of Bell Laboratories, SABL, is
exploited to derive these values of p.

(3) Peter Burman believes that these differences are due to different estimates of ~12~

Correspondence:

Gerhard Thury
Osterreichisches Institut for Wirtschaftsforschung
Arsenal, Objekt 20
Postfach 91
A-1103 Wien

207

Das könnte Ihnen auch gefallen