PROGRAM TSW

REFERENCE
MANUAL

Gianluca Caporello, Agustin Maravall

and Fernando J. Sanchez

Banco de España - Servicio de Estudios

Documento de Trabajo nº 0112
 PROGRAMTSW
REFERENCE
MANUAL

Gianluca Caporello, Agustin Maravall

and Fernando J. Sanchez

Banco de Espana· Servicio de Estudios

Documento de Trabajo n2 at 12
 In publishing this series the Banco de Espana seeks to disseminate
studies of interest that will help acquaint readers better
with the Spanish economy.

The analyses, opinions and findings of these papers represent

the views of their authors; they are not necessarily those
of the Banco de Espana.

The Banco de Espana disseminates its main reports and most

of its publications via the INTERNET at the following website:
http://www.bde.es

]SSN: 0213-2710
ISBN: 84-7793-761-3
Dep6sito legal: M. 37356-200]
lmprenta del Banco de Espana
 Abstract

The user instructions for Program TSW are provided.

TSW is a Windows version, developed by G.Caporello

and AMaravall, of Programs TRAMO and SEATS (Gomez

and Maravall, 1996), that incorporates several

modifications and new facilities.

TABLE OF CONTENTS

INSTALLATION 5

1. BRIEF DESCRIPTION OF THE PROGRAMS 6

1.1 PROGRAM TRAMO 6

1.2 PROGRAM SEATS 10

1.2.1 Brief description 1 0

1.2.2 Decomposition of a n ARIMA Model 13

1.3 PROGRAM TERROR 17

1.4 PROGRAM TSW 17
1.5 REFERENCES 17

2. USER INSTRUCTIONS 23

2.1 MAIN WINDOW 23

2.2 LOADING A SERIES 24
2.3 MODEL SPECIFICATION 27
Checking Input Parameters 29

2.4 REGRESSION VARIABLES 30

2.5 EXECUTION OF THE PROGRAM AND OUTPUT FILE 35

2.5.1 Execution 35
2.5.2 Main output files 36
2.5.3 Out-tables ( output series ) 38
2.5.4 Summary output and Out-Matrix 39
a) summaryt.txt
b) summarys.txt
c) out-matrix

2.6 GRAPHS 42
2.7 SAVE/LOAD 44
2.8 MANY SERIES ANDIOR MODELS: THE ITER PARAMETER 45
2.9 DATA BASE FACILITY: DBXPLORE 46
2.10 PROGRAM TERROR 49

3. INPUT PARAMETERS 50

3.1 AUTOMATIC PROCEDURE 50

3.2 TRAMO PARAMETERS 50

3.2.1 ARIMA model 50

3.2.2 Calendar effects 52
3.2.3 Outliers 52
3.2.4 Automatic model identification 53
3.2.5 Other parameters 53
3.2.6 Regression Variables 53

-7-
 3.3 SEATS PARAMETERS 54
3.4 TERROR PARAMETERS 55

4. INDEX OF INPUT PARAMETERS 56

-8-
 INSTALLATION

A) Local installation.

For local installation of TSW, create a directory, say TEMP, with six subdirectories, DISK1,
DISK2 ..... DISKS. Copy the file Disk1.zip in DISK.1. the file Disk2.zip in DISK2..... and the
file DiskS.zip in DISKS. Then. extract the six files and execute SETUP. EXE in DISK1. The
program autoinstalls itself. For future versions, you do not have to uninstall the program.
Simply follow the same installation steps; the old version will be automatically replaced. The
program is installed in the directory PROGRAM FILESITSW; the output file of the program
wilt be deposited in PROGRAM FILESITSw\OUTPUT. and the arrays for the graphs in
PROGRAM FILESITSw\GRAPH. To execute the pnogram simply follow the steps: "Start­
>Programs->Seats Tramc Windows->TSW",
Note: Everytime the program is restarted, the OUTPUT directory is emptied.

B) Network installation.

The directory NETINSTALL contains the file "netinstall.exe" which is a small program for
installation of TSW in a network. To do this, first, TSW should be installed in the server, and
then each user should execute the program netinstall.exe from their own PC (the program
resides in the server). The only information the user should supply is the name of the local
destination directory of TSW on the user PC. In this PC several directories will be created
(OUTPUT. GRAPH. BIN....) where the output files of TSW ";11 be deposited (the programs
will remain in the server).

- 9 -
 1. BRIEF DESCRIP TION OF THE PROGRAMS

TRAMO and SEATS are two programs developed by Victor Gomez and Agustin Maravall for
applied time series analysis. Program TERROR is a particular application of TRAMO, and
program TSW is a Windows version of TRAMO-SEATS developed by Gianluca Caporello
and Agustin Maravall. The programs are briefly summarized in the following pages; the first
two sections are a revised version of sections 1.1,2.1, and 2.2 in Gomez and
Maravall (1996)

1.1 PROGRAM TRAMO

TRAMO ("Time Series Regression with Alima Noise. Missing Observations and Outliers") is a
program for estimation, forecasting, and interpolation of regression models with missing
observations and ARIMA errors, in the presence of possibly several types of outliers (no
restriction is imposed on the location of the missing observations in the series). The program
can be run in an entirely automatic manner.

Given the vector of observations:

z= ( ZII' . . .• ZI.. )' (1)

where 0 < t, <.. . < 41, the program fits the regression model

Z,=y',j3+X,. (2)

where p;: ( 131, . .. , Pn )' is a vector of regression coefficients, y, ;: ( y 1 t , . . . , y n

I
) denotes

n regression variables, and XI follows the general ARIMA process

¢(B)6(B)x, = B(B) a"

(3)

where B is 1he backshilt operalor, � (B ), Ii ( B ) and 9 ( B ) are finite polynomials in B, and

aI is assumed a n.i.i.d (0, cr�) white-noise innovation.

The polynomial Ii (B ) contains 1he unit roots associated .,,;th differencing (regular and
seasonal), ¢ ( B ) is the polynomial "';1h 1he stationary autoregressive roots, and 9 (B )
denotes the (invertible) moving average polynomial. In TRAMO, they assume the following
multiplicative form:

' o
o(B) = (1- B ) (I- B')
�(B) = (I +�, B+ . + �,B')(1 +$, B ' +...+$, B "')
..

9(B)=(1 +e, B+ ... +e, B')(1 +0, B'+...+00B"0) •

- 10 -
where s denotes the number of observations per year. The model may contain a constant /J,
equal to the mean of the differenced series 5 (8 ) z t In practice, this parameter is estimated
•

as one of the regression parameters in (2).

The regression variables can be input by the user (such as a variable capturing holidays, or an
economic variable thought to be related with zJ, or generated by the program. The variables
that can be generated are trading day, easter effect and intervention variables (Box and liao,
1975) of the type:

a) dummy variables;
b) any possible sequence of ones and zeros;
c) 1/(1-66) of any sequence of ones and zeros, where 0< 6 s1;
d) 1/(1-6,8') of any sequence of ones and zeros, where 0< 6, :5:1;
e) 1/(1-6)(1-6') of any sequence of ones and zeros.

The program:

1) estimates by exact maximum likelihood (or unconditional/conditional least

squares) the parameters in (2) and (3);
2) detects and corrects for several types of outliers;
3) computes optimal forecasts for the series, together with their MSE;
4) yields optimal interpolators of the missing observations and their associated
MSE; and
5) contains an option for automatic model identification and automatic outlier
treatment.

The basic methodology followed is deserbed in G6mez and Maravall (1992, 1994, 1996,
2001a), and G6mez, Maravall and Pena (1999). The program is aimed at monthly or lower
frequency data, the maximum number of observations is 600 and the minimum depends on the
periodidty of the data (in particular, 16 for quarterly and 36 for monthly data).

Estimation of the regression parameters (including intervention variables and outliers, and the
missing observations among the initial values of the series), plus the ARIMA model parameters,
can be made by concentrating the former out of the likelihood (default), or by joint estimation.
Several algorithms are available for computing the likelihood or more precisely, the nonlinear
sum of squares to be minimized. When the differenced series can be used, the algortthm of
Marl, Sidhu and Kailath (1974) is employed, with a simplification similar to that of Melard
(1984), but also extended to multiplicative seasonal moving average models. For the
nondifferenced series, it is possible to use the ordinary Kalman filter (default option), or its
square root version (see Anderson and Moore, 1979). The latter is adequate when numerical
difficulties arise; however it is markedly slower. 8y default. the exact maximum likelihood
method is employed. Nonlinear maximization of the likelihood function and computation of the
parameter estimates standard errors is made using Marquardfs method and first numerical
derivatives.

Estimation of regression parameters is made by using first the Cholesky decomposition of the
inverse error covariance matrix to transform the regression equation (the Kalman filter provides
an efficient algorithm to compute the variables in this transformed regression). Then, the
resulting least squares problem is solved by applying the QR algOrithm, where the Householder
orthogonal transformation is used. This procedure yields an efficient and numerically stable
method to compute GLS estimators of the regression parameters, which avoids matrix
inversion.

- 11 -
For forecasting, the ordinary Kalman filter or the square root filter options are available. These
algorithms are applied to the original series; see G6mez and Maravall (1993) for a more
detailed discussion on how to build initial conditions on a nonstationary situation.

Missing observations can be handled in two equivalent ways. The first one is an extension to
nonstationary models of the skipping approach of Jones (1980), and is described in G6mez and
Maravall (1994). In this case, interpolation of missing values is made by a simplified Fixed Point
Smoother, and yields identical results to Kohn and Ansley (1986). The second one consists of
assigning a tentative value and specifying an additive outlier to each missing observation. If this
option is used, the interpolator is the difference between the tentative value and the estimated
regression parameter and coincides with the interpolator obtained with the skipping approach
(the likelihood is corrected so that it coinddes with that of the skipping approach; see G6mez,
Maravall and Peria (1999) for more details.) When concentrating the regression parameters out
of the likelihood, mean squared errors of the forecasts and interpolations are obtained following
the approach of Kohn and Ansley (1985).

When some of the initial missing values are unestimable (free parameters), the program
detects them, and flags the forecasts or interpolations that depend on these free parameters.
The user can then assign arbitrary values (typically, very large or very small) to the free
parameters and rerun the program. Proceeding in this way, all parameters of the ARIMA model
can be estimated because the function to minimize does not depend on the free parameters.
Moreover, it will be evident which forecasts and interpolations are affected by these arbitrary
values because they will strongly deviate from the rest of the estimates. However, if all
unknown parameters are jointly estimated, the program may not flag all free parameters. It may
happen that there is convergence to a valid arbitrary set of solutions (i.e., that some linear
combinations of the initial missing observations, induding the free parameters, are estimable).

The program has a facility for detecting outliers and for removing their effect; the outliers can be
entered by the user or they can be automatically detected by the program, using an original
approach based on those of Tsay (1986) and Chen and Liu (1993). The outliers are detected
one by one, as proposed by Tsay (1986), and multiple regressions are used, as in Chen and
Liu (1993), to detect spurious outliers. The procedure used to incorporate or reject outliers is
similar to the stepwise regression procedure for selecting the "best" regression equation.

In brief, regression parameters are initialized by OLS and the ARIMA model parameters are
first estimated with two regressions, as in Hannan and Risannen (1982). Next, the Kalman filter
and the OR algorithm provide new regression parameter estimates and regression residuals.
For each observation, t-tests are computed for several types of outliers. If there are outliers
whose absolute t-values are greater than a pre-selected critical level C, the one with the
greatest absolute t-value is selected. Otherwise, the series is free from outlier effects and the
algorithm stops.

If some outlier has been detected, the series is corrected by its effect and the ARMA model
parameters are re-estimated. Then, a multiple regression is performed using the Kalman filter
and the OR algorithm. If there are some outliers whose absolute t-value is removed from the
regression residuals provided by the last multiple regression, t-tests are computed for the
different types of outliers and for each observation. If there are outliers whose absolute t-values
are greater than the critical level C, the one with the greatest absolute t-value is selected and
the algOrithm goes on to the estimation of the ARMA model parameters to iterate. Otherwise,
the algorithm stops. A notable feature of this algorithm is that all calculations are based on
linear regression techniques, which reduces computational time. By default, three types of
outliers are considered: additive outlier, level shift, and tranSitory change.

- 12 -
The program also contains a facility to pretest for the log-level specification (based on a
comparision of the BIC using both specifications) and, if appropriate, for the possible
presence of Trading Day and Easter effects (the pretests are made with regressions using
the default model for the noise and, if the model is subsequently changed, the test is
redone); it further performs an automatic model identification of the ARIMA model. This is
done in two steps. The first one yields the nonstationary polynomial 6(B) of model (3). This is
done by iterating on a sequence of AR and ARMA(1,1) models (with mean), which have a
multiplicative structure when the data is seasonal. The procedure is based on results of Tiao
and Tsay (1983), and Tsay (1984). Regular and seasonal differences are obtained, up to a
maximum order of !1 2!1s' where !1 = 1- B and !1s = 1-BS .

The second step identifies an ARMA model for the stationary series (corrected for outliers and
regression-type effects) following the Hannan-Rissanen procedure, with an improvement which
consists of using the Kalman filter instead of zeros to cala.Jlate the first residuals in the
computation of the estimator of the variance of the innovations of model (3). For the general
multiplicative model

$,(B)<I>,(B')x,=9,(B)0Q(B')a, •

the search is made over the range 0,; ( p q ) '; 3 0,; ( P Q ) ,; 2 . This is done sequentially
• • •

(for fixed regular polynomials, the seasonal ones are obtained, and vice versa). and the final
orders of the polynomials are chosen according to the BIC criterion. with some possible
constraints aimed at increasing parsimony and favouring "balanced" models (similar AR and
MA orders).

Finally, the program combines the facilities for automatic detection and correction of outliers
and automatic ARIMA model identification just described in an efficient way, so that it can
perform automatic model identification of a nonstationary series in the presence of outliers and
missing observations (perhaps with some regression effects).

The default model in TRAMO is the so-called Airline Model, popularized by Box and Jenkins
(1970). The model is given by the equation

(4)

with -1 � (9,,9,) � 1. It is often found appropriate for many series (see the large-scale study in
Fischer and Planas (2000», and displays many convenient features (see, for example, Maravall
(1998»; in particular it encompasses many other models, including models with close to
deterministic trend or seasonality. or models without seasonality. For very short series, for
which the automatic model identification is unreliable, TRAMO relies heavily on the Airline
model specification.

Although TRAMO can obviously be used by itself, for example, as an outlier detection an
interpolation or a forecasting program, it can also be seen as a program that polishes a
contaminated "ARIMA series". That is, for a given time series, it interpolates the missing
observations, identifies outliers and removes thei� effect, estimates Trading Day and Easter
Effect, etc... , and eventually produces a series tha,t can be seen as the realization of a linear
stochastic process (i.e., an ARIMA model). Thus, jTRAMO, can be used as a pre-adjustment
program to SEATS (see below), which decomposes then the "linearized series" and its
forecasts into its stochastic components.
I

- 13 -
Both programs can handle routine applications to a large number of series and provide a
complete model-based solution to the problems of forecasting, outlier correction, interpolation
and signal extraction for nonstationary time series.

1.2 P ROGRAM SEATS

1.2.1 Brief Description

SEATS ("Signal Extraction in ARIMA Time Series") is a program for decomposing a time series
into its unobserved components (Le., for extracting from a time series its different signals),
following an ARIMA-model-based method. The method was developed from the work of
Cleveland and Tiao (1976). Box. Hillmer and Tiao (1978). Burman (1980), Hillmer and Tiao
(1982), Bell and Hillmer (1984), and Maravall and Pierce (1987), in the context of seasonal
adjustment of economic time series. In fact, the starting point for SEATS was a program built by
Burman for seasonal adjustment at the Bank of England (1982 version).

In the standard case, SEATS receives from TRAMO the original series, the deterministic effects
TRAMO has estimated (outliers, trading day or easter effects, and in general regression
variable effects), the interpolated series with the deterministic effects removed (i.e., the
"linearized" series x, in (2», and the ARIMA model identified and estimated for these series,
given by (3). The model can be written in detailed form as

(5)

and, in concise form, as

(6)

where <I>(B) � � (B)o(B) represents the complete autoregressive polynomial, including all unit
roots. Notice that, if p denotes the order of �(B) and q the order of 8(B), then the order of
<I>(B) is P � p+d+ D x s .

The program decomposes a series that follows model (5) into several components. The
decomposition can be multiplicative or additive. Since the former becomes the second by taking
logs, we shall use in the discussion an additive model, such as

X!= L Xi! , (7)

where x it represents a component. The component that SEATS considers are:

x pI = the TREND·CYC LE component,
x $t the SEASONA L component,
;;

x ct;; the TRANSITORY component,

X ut;; the IRREGULAR component.

Broadly, the trend-cycle component captures the low-frequency variation of the series and
displays a spectral peak at frequency 0 the seasonal component, in tum, captures the spectral
peaks at seasonal frequencies; and the irregular component captures erratic, white-noise
behavior, and hence has a flat spectrum. The tranSitory component is a zero-mean stationary

-14-
component that picks up transitory fluctuations that should not contaminate the trend-cyde or
seasonal component and are not white-noise (see next section). The components are
determined and fully derived from the structure of the (aggregate) ARIMA model for the
observed series, which can be directly identified from the data. Like TRAMO, SEATS is aimed
at monthly or lower frequency data and has the same restrictions on the maximum and
minimum number of observations.

The deoomposition assumes orthogonal components, and each one will have in tum an ARIMA
expression. In order to identify the components, we will require that (except for the irregular
one) they be clean of noise. This is called the "canonical" properly, and implies that no additive
white noise can be extracted from a component that is not the irregular one. The variance of the
latter is, in this way, maximized, and, on the contrary, the trend-cycle and seasonal component
are as stable as possible (compatible with the stochastic nature of model (6)). Although an
arbitrary assumption, since any other admissible component can be expressed as the
canonical one plus independent white-noise, lacking a priori information on the noise variance,
the assumption seems rather sensible.

The model that SEATS assumes is that of a linear time series with Gaussian innovations. In
general, SEATS is designed to be used with the companion program TRAMO, which removes
from the series special effects, such as Trading Day, Easter, holiday, and intervention or
regression variable effects, identifies and removes several types of outliers, and interpolates
missing observations. TRAMO passes to SEATS the linearized series and the ARlMA model
for this series, perhaps obtained through the automatic facility. When no outliers or deterministic
effects have to be removed and there are no missing values, SEATS can be used by itself
because it also contains an ARIMA estimation routine. This routine is also used when the
TRAMO model should be modified in order to decompose the series (such is the case, for
example, when the TRAMO model does not accept an admissible decomposition). In either
case, SEATS performs a control on the AR and MA roots of the model. When the modulus of a
root converges within a preset interval around 1, the program automatically fixes the root. If it is
an AR root, the modulus is made 1; if it is an MA root, it is fixed to the lower limit (by default,
. 99). This simple feature, we have found, makes the program very robust to over- and under­
differencing.

SEATS computes new residuals for the series in the following way. The TRAMO residuals are
obtained with the Kalman filter and are equal in number to the number of observations in the
series minus the sum of the number of observations lost by differencing and the degrees of
freedom lost by estimation of outliers and other deterministic effects. SEATS uses the ARIMA
model to filter the linearized series and estimates by maximum likelihood the residuals that
correspond to the observations lost by differencing. In this way, SEATS assigns a residual for
each period spanned by the Original series. The SEATS residuals are called "extended
residuals·; for the overlapping periods, they are extremely close to the TRAMO ones. The
extended residuals are subject to the same diagnostics as the ones in TRAMO (presence of
autocorrelation, presence of seasonality, randomness of Signs, skewness, kurtosis, normality,
and nonlinearity). The program proceeds then to decompose the ARIMA model. This is done in
the frequency domain. The spectrum (or pseudospectrum) is partitioned into additive spectra,
associated with the different components. (These are determined, mostly, from the AR roots of
the model.) The canonical condition identifies a unique deoomposition, from which the ARIMA
models for the components are obtained (including the component innovation variances).

For a particular realization [x I ,x 2

•... X T]
• • the program yields the Minimum Mean Square
Error (MMSE) estimators of the components, computed with a Wiener-Kolmogorov-type of filter
applied to the finite series by extending the latter with forecasts and backcasts (use is made of

- 1 5-
the efficient Burman�Wilson algorithm; see Burman, 1980). For i = 1 , ... , T , the estimate :R iliT ,
equal to the conditional expectation E(x it I X l, ... , X T) , is obtained for all components.

For a large enough series and values of t not close to 1 or T, the estimator x it l T becomes the
"final" or "histOrical" estimator, which we shall denote x it . (In practice, it is achieved for large
enough k = T � t , and the program indicates how large k can be assllmed to be.) For t = T ,
the concurrent estimator, :R ,TIT , is obtained, i.e., the estimator for the last observation of the
series. The final and concurrent estimators are the ones of most applied interest. When
T - k < t < T,:R iliT yields a preliminary estimator, and, for t > T , a forecast. Besides their

estimates, the program produces several years of forecasts of the components, as well as
standard errors (SE) of all estimators and forecasts. For the last two and the next two years, the
SE of the revision the preliminary estimator and the forecast will undergo is also provided. The
program further computes MMSE estimates of the innovations in each one of the components.

The joint distribution of the (stationary transformation of the) components and of their MMSE
estimators are obtained; they are characterized by the variances and auto- and cross­
correlations. The comparison between the theoretical moments for the MMSE estimators and
the empirical ones obtained in the application yields additional elements for diagnosis (see
MaravaU, 1987). The program also presents the Wiener-Kolmogorov filter for each component
and the filter which expresses the weights with which the different innovations a j in the series
contribute to the estimator :R il IT ' These weights directly provide the moving average
expressions for the revisions. Next, an analysis of the estimation errors for the trend and for the
seasonally adjusted series is performed. Let

dit = Xii-Xii,

denote the final estimation error, the preliminary estimation error, and the revision error in the
preliminary estimator :R iliT . The variances and autocorrelation functions for d it , d itll ,rilll are
displayed. (The autocorrelations are useful to compute SE of linearized rates of growth of the
component estimator.) The program then shows how the variance of the revision error in the
concurrent estimator rilll decreases as more observations are added, and hence the time it
takes in practice to converge to the final estimator. Similarly, the program computes the
deterioration as the forecast moves away from the concurrent estimator and, in particular, what
is the expected improvement in Root MSE associated with moving from a once-a�year to a
concurrent seasonal adjustment practice. Finally, the SE of the estimators of the linearized
rates of growth most closely watched by analysts are presented, for the concurrent estimator of
the rate and its successive revisions, both for the trend and seasonally adjusted series. The
SEs computed assume that the ARIMA model for the observed series is correct. Further details
can be found in Maravall (1988, 1993, 1995), Gomez and Maravall (1992, 2001b), and Maravall
and Planas (1999). For a basic introduction to the time series analysiS concepts and tools in
oonnection with TRAMO-SEATS, see Kaiser and Maravall (2000).

- 16-
As in TRAMO, the default model in SEATS is the Airline Model, given by (4), which provides
very well behaved estimation filters for the components. The implied components have models
of the type

(8)

(9)

where S = 1 + B + ... + 8S-1, is the annual aggregation operator and e p ( B) and 9 s ( B) are
both of order 2 and (s-1) , respectively. Compared to other fixed filters, the default model
allows for the observed series to estimate 3 parameters: 91 I
related to the stability of the
trend-cycle component; 9s ' related to the stability of the seasonal component; and a2a ' a
measure of the overall predictability of the series. Thus, to some extent, even in this simple
fixed model application, the filters for the component estimators adapt to the specific structure
of the series. Notice that model (8) implies a locally linear trend, that becomes quadratic when
model (4) oontains a constant, and that model (9) implies that the sum of the seasonal
component over a one-year period will, on average, be zero. The fact that, for a particular year,
the seasonal component does not exactly cancel out implies that the annual averages of the
original and seasonally adjusted series will not be equal. This is a typical feature of stochastic
models with stochastic (or "moving�) components.

Programs TRAMO and SEATS provide a fully model-based method for forecasting and
signal extraction in univariate time series. (The relation between them is somewhat similar to
the one between the programs REGARIMA and the revised X11 ARIMA that form the new
method X12 ARIMA; see Findley et ai, 1998.) The procedure is flexible, yet robust and
reliable. Due to the model-based features, it becomes a powerful tool for detailed analysis of
important series in short-term policy making and monitoring, yet TRAMO-SEATS can also be
effiCiently used for routine application to a very large number of series due to the automatic
procedures available. The standard automatic procedure pretests for the log-level
specification and, if appropriate, for the possible presence of Trading Day and Easter effects;
it further performs an automatic model identification and outlier detection and correction
procedures (for several types of outliers), interpolates the missing values if any, and
decomposes the series net of the previous (deterministic) effects into a seasonal, trend­
cycle, transitory, and irregular components. (If the identified ARIMA model does not accept
an admissible decomposition, it is automatically replaced by a decomposable
approximation). Finally, the components (and forecasts thereof) estimated by SEATS are
modified to incorporate the deterministic effects that were estimated by TRAMO and
removed from the series in order to linearize it. As a general rule, additive outliers are added
to the irregular component, transitory changes to the transitory component, and level shifts
to the trend. Trading Day and Easter effects are added to the seasonal component, as well
as Holiday effect; their sum is called Calendar effect. RegreSSion variables can be added to
any one of the components, or (by default) form a separate component. When added to the
seasonal component, SEATS checks that the effect is properly centered.

1.2.2 Decomposition of the ARIMA Model

Let the total AR polynomial <I>(B) of the ARIMA model (6) be factorized as

<I>(B) � �,(B) �,(B') 8d 8,0.

-17 -
The roots of <I> (B ) are assigned to the unobserved components as follows.

Roots of dd = 0 : Assigned to trend-cycle component.

Roots of 6.s0=0: Factorizing it as (LlS)D=O, where

S=l+B+ ... + B s , ]
the root of d = 0 goes to the trend-cycle.
•

the roots of S = 0 go to the seasonal component.

•

If $,(B)=I+$,B+... + $ , B' and $,(B)=I+$,B', let z=B-1 and consider the roots of the
polynomials
$r{Z)= zP +$]ZP-I + ..+$p.
� ,(z)= z' +�,.

Roots of �,(z ):
Real positive roots:
• If modulus � k, assigned to trend-cycle.
• If modulus < k, assigned to transitory component.

Real negative roots:

• If s -;c 1, and modulus � k assigned to seasonal component (root implies a
period of 2).
•If s -;c 1 and modulus < k, assigned to transitory component
•If s = 1 (annual data), assigned to transitory component.

Complex roots: Let w denote the frequency of the root.

If <0 E [a seasonal frequency ±e], assigned to seasonal component.
• Otherwise, assigned to transitory component.

Roots of�, ( z'), Letting �denote the real positive root of(- �,)1f' , the polynomial �,(z)
can be rewritten as (Z_$XZS-1 +$ZS-2 + $2ZS-3 + ... + $S-1 ).
• when �, k , the AR root (1- $B) is assigned to the trend; the other (s-
1) roots to the seasonal component.
• when 4> < k , roots are assigned to the transitory component.

(Note: The parameters k and E can be controlled by the user.)

The factorization of cD(B) can be rewritten as

where $, (B), $, (B) and $, (B) are the AR polynomials with the trend, seasonal, and transitory
roots, respectively. Let P and Q denote the orders of the polynomials <l>(B) and 6(B) in (6):

-18-
a) Consider firsllhe case P ;, Q. A polynomial division of the spectrum (or pseudospeclrum)
of model (6) yields a first decomposition of the Iype

8(B) ll(B)
1l>(B) at = cI>(B) all + VI •

where the order of 1l(B) is min (0, P-1), and v, is a co nst ant (0 � P>O).

A partial fraction expansion of the spectrum of [ll (B) I <I> (B)] a" yields the decomposition

where, letting j � p, s,c, we have order (e J)S order( � J) . If gj(Ol) denotes the spectrum of
[llj (B) I � j (B) hr,,, let

Imposing the canonical condition

gj (00) �g, (00) Vj , j �p,5,C,

-

the spectrum of the final components are obtained. Factorizing these spectra, the models for
the components:

$, (B) p, � 8, (B) a"

Ut = white noise (0. v)

are obtained (the spectral factorization algOrithm is described in Maravall and Mathis (1994».
All components have balanced models, in the sense that t he order of the AR polyn omial equals
thai of the MA one.

b) When (Q>P), the decomposition proceeds as follows.

A first decomposition is performed, whereby

ARIMA(P,Q)�ARIMA(P,P-l)+MA(Q-P) .

- 19 -
The first component falls under case a), and hence can be decomposed in the previous way.
Let this decomposition be, in general,

ARlMA(P, P-l)=p, +" +c, +u,

where PI' S I ,c I and ut denote the trend-cyde, seasonal, transitory, and irregular
component. The MA(Q-P) component, which represents stationary short-term deviations, is
added to the transitory component. The series is decomposed then, into a balanced trend-cycle
model, a balanced seasonal model, a top-heavy transitory model, and a white-noise irregular.
The first three components are made canonical (i.e., noise free).

As a general rule, it is recommended that balanced models be favoured, since they tend to
display good decomposition properties. Models for which Q is much larger than Pare
discouraged because the excess MA structure may provoke awkward, on occasion
nonadmissible, decompositions.

Example: The monthly model

( 1 +.4B-.32B')t.t.02x,=9 (B)a, ,

with 9(B) of order Q=16 (> P = 15), would deoompose as follows. Factorizing the AR(2), tt can
be rewritten as (1- .4B)(1 +.8 B), and hence the first factor goes to the transitory oomponent
(by default, k = .5), and the seoond factor to the seasonal oomponent (a peak for co = . ).
Therefore, the models for the components will be of the type:

Trend:

Seasonal: (I + .8B)S" = 9, (B)a"

Transitory: (1-.4 B) c, = 9, (B)a"

Irregular: White noise.

The orders of the polynomials 9, (B), 9, (B), and 9, (B) would be 2, 12, and 2, respectively.

Examples of straightforward TRAMO-SEATS applications can be found in Kaiser and

Maravall (2001 a), Maravall and Sanchez (2000), and Maravall (2000); extensions of
TRAMO-SEATS to the problem of quality control of data and business-cyde estimation are
described in Luna and Maravall (1999) and Kaiser and Maravall (2001b).

Final Remark

TRAMO and SEATS are intensively used at (and reaammended by) Eurostat (see Eurastat
(1996, 1998, 1999, 2000» and at the European Central Bank, (see European Central Bank
(1999, 2000» together with Xl2ARIMA. They are used at many central banks, statistical
offices, and other economic agencies in and outside Europe, both for in-depth treatment and
analysis of important series (see, for example, European Central Bank (2000), Banco de
EspaM (1994), Banca d'italia (1999), or Banaa de Reserva de EI Salvador (1998), for careful

-20-
 treatment of groups of (most often) economic indicators (some Spanish examples, in the public
and private sectors, are found in Ministerio de Economia y Hacienda (2000), Instituto Nacional
de Estadistica (1997), Analistas Financieras Intemacianales (1998), Expansion (1998), Agencia
Tributaria (1999), or Banaa Santander Central Hispana (2000», or for relatively large scale use
(see, for example, Eurastat (1995, 1997), ISTAT (2000), Statistics Sweeden (2000), or National
Bank of Belgium (2001». The main applications are short-telTTl forecasting and monitiring,
seasonal adjustment, trend-cycJe estimation, interpolation, detection and correction of outliers,
detection of errors and anomalies in data, and estimation of special effects.

1.3 PROGRAM TERROR

,
TERROR ( 'TRAMO for errors"). is an application to quality control of data; in particular, to the
detection of errors in reported (time series) data. Program TERROR is designed to handle large
sets of time series with a monthly or lower frequency of observation, and specifies a particular
conflQuration of TRAMO, that will be applied to each time series, mostly based on the automatic
model identification and outlier detection and correction procedures.

For each series, the program automatically identifies an ARIMA mooel and detects and corrects
for several types of outliers. (It also interpolates missing observations if there are any.) Next, the
one-period-ahead forecast of the series is computed and compared with the new observation
(this new observation is not used for estimation). In brief, when the forecast error is, in absolute
value, larger than some a priori specified limit, the new observation is identified as a possible
error. More details are provided in Caporello and Maravall (2000), and Luna and
Maravall(1999).

1.4 PROGRAM TSW

Program TSW is a Windows version of TRAMO-SEATS, with a few modifications, developed

by Gianluca Caparella and Agustin Maravall. Besides the usual text format for the input file,
the program can access and deposit files in EXCEL, and contains several new facilities, in
particular, summary tables with the main series that are output of the two programs
(deterministic corrections and stochastics components with their forecasts), summary results
for both programs, and a small data base facility to help routine applications. The program
contains an on-line help facility with the meaning of all parameters.

1.5 REFERENCES

ANDERSON, B. and MOORE, J. (1979), Optimal Finern

i g, New Jersey: Prentice Hall.

AGENCIA TRIBUTARIA (1999), Cayuntura Economica y Recaudaci6n Tributara i en el Tercer

Trimestre de 1999, Nota de Coyuntura 4/99, Madrid:.Agencia Tributaria.

ANALISTAS FINANCIEROS INTERNACIONALES (1998), Informe Mensual, Junia 1998, Madrid:

Analistas Financieros Intemacionales (monthly bulletin).

BANCA D'ITALIA (1999), Supplementi al Bol/ettino Statistica; Note metodo/ogiche e infarmazioni

statistiche, IX, 35, July 1999; Rama: Banca d'italia.

BANCO DE ESPAf<A (1993), "La Desestacianalizaci6n de los Agregados Manetarias en 1993",

Bo/etJn Econ6mico, Febrero 1993, Madrid: Banco de Espana.

BANCO DE ESPAf<A (1999), "Indicadares Econ6micos", Boletin Economico, Madrid: Banco de

Espana (monthly bulletin).

-21-
BANCO DE RESERVA DE EL SALVADOR (1998). Boletin Econ6mico (mon1hly bullelin).
November 1998.

BANCO SANTANDER CENTRAL HISPANO (2000). La Economla de Cantabria. Sanlander:

Gobierno de Cantabria (monthly bulletin).

BANQUE NATIONALE DE BELGIQUE (2001). "TRAMO-SEATS: Developpemenl par

Composantes Reutilisables", by Jean Palate, Cellule R&D. Dept. de Statistique, May 2001.

BELL, W.R. and HILLMER, S.C. (1984). "Issues Involved with the Seasonal Adjustment of Economic
Time Series·. Journal of Business and Economic Statistics 2, 291-320.

BOX, G.E.P., HILLMER, S.C. and TIAO, G.C. (1978). "Analysis and Modeling of Seasonal Time
Series", in Zellner. A. (ed.), Seasonal Analysis of Economic Time Series, Washington, D.C.:
U.S. Dept. of Commerce - Bureau of the Census, 309-334.

BOX, G.E.P. and JENKINS, G.M. (1970), Time Series Analysis: Forecasting and Control, San
Francisco: Holden-Day.

BOX, G.E.P. and TIAO, G.C. (1975). "Intervention Analysis with Applications to Economic and
Environmental Problems", Journal of the American Statistical Association 70, 71-79.

BURMAN, J.P. (1980), "Seasonal Adjustment by Signal Extraction", Journal of the Royal Statistical
Society A. 143, 321-337.

CAPORELLO, G. and MARAVALL, A. (2000). "TERROR: An Applicalion of TRAMO to the Detection of

Errors; User Instructions", mimeo.

CHEN, C. and LlU, L.M. (1993), "Joint Estimation of Model Parameters and Outlier Effects in Time
Series", Journal of the American Statistical Association 88, 284-297.

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

EUROPEAN CENTRAL BANK (1999), Monthly Buffetin, Frankfurt am Main: ECB.

EUROPEAN CENTRAL BANK (2000), Seasonal Adjustment of Monetary Aggregates and HICP for
the Euro Area, Fra nkfu rt am Main: ECa.

EUROSTAT (1995), Industrial Trends; Monthly Statistics, Luxembourg: Office des publications
officielles des Communautes europeennes (monthly bulletin)

EUROSTAT (1996), Methodology of Industrial Shorl-tenn Indicators; Rules and Recommendations,

Luxembourg: Office for Official Publications of the European Communities.

EUROSTAT (1997), Monthly Panorama of European Industry, Luxembourg: Office for Official
Publications of the European Communities (monthly bulletin).

EUROSTAT (1998), Seasonal Adjustment Methods: A Comparison, Luxembourg: Office for Official
Publications of the European Communities.

EUROSTAT (1999a), Eurostat Suggestions Concerning Seasonal Adjustment Policy. result of the
Eurostat task force on seasonal adjustment policy.

-22-
EUROSTAT (1 999b), Monthly Panorama of European Business, Luxembourg: Office for Official
Publications of the European Communities.

EXPANSIO N ( 1 998), "Indicador Expansion de Actividad Economica�, Madrid: Recoletos, January

12, 1998 (monthly report).

FINDLEY, D.F., MaNSEll, B.C., BEll, W.R., ana, M.C., and CHEN, B.C. (1996), "New
Capabilities and Methods of the X-12-ARIMA Seasonal Adjustment Program" (with discussion),
Journal of Business and Economic Statistics, 12, 127-177.
I
FISCHER, B. and PLANAS, C. (2000), "Large Scale Fitting of Regression Models with ARIMA Errors,"
Journal of Official Statistics, 16, 173-184.

GOMEZ, V. and MARAVALL, A. (1 992), "Time Series Regression with ARIMA Noise and Missing
Observations - Program TRAM", : EUI Working Paper ECO No. 92181 , Department of
Economics, European University Institute.

GOMEZ, V. and MARAVALL, A (1 993), "Initializing the Kalman Filter with Incompletely Specified Initial
Conditions", in Chen, G.R. (ed.), Approximate Kalman Filtering (Series on Approximation and
Decomposition), London: World Scientific Pub!. CO.

GOMEZ, V. and MARAVALL, A. (1 994), "Estimation, Prediction and Interpolation for Nonstationary
Series with the Kalman Filter" , Journal of the American Statistical Association 89, 6 1 1 -624.

GOMEZ, V. and MARAVALL, A. (2001a), "Automatic Modelling Methods for Univariate Series", Ch.7 in
Pena D., Tiao G.C. and Tsay, R.S. (eds.) A Course in Time Series Analysis. New York: J. Wiley
and Sons.

GOMEZ, V. and MARAVALL, A (2001b), "Seasonal Adjustment and Signal Extraction in Eoonomic
Time Series", Ch.8 in Pena D., Tiao G.C. and Tsay, R.S. (eds.) A Course in Time Series
Analysis, New York: J. Wiley and Sons.

GOMEZ, V. and MARAVALL, A (1996), "Programs TRAMO (Time series Regression with Arima nOise,
Missing observations, and Outliers) and SEATS (Signal Extraction in Arima Time Series).
Instructions for the User" , Working Paper 9628, Servicio de Estudios, Banco de Espana.

GOMEZ, V., MARAVALL A, and PE NA, D. ( 1 999), "Missing Observations in ARIMA Models:
Skipping Approach Versus Additive Outlier Approach", Journal of Econometrics, 88, 341-364.

HANNAN, E.J. and RISSANEN, J. (1 982), "Recursive Estimation of Mixed Autoregressive-Moving

Average Order" , Biometrika 69, 81 -94.

HILLMER, S.C., BELL, W.R. and TIAO, G.C. (1983), "Modeling Considerations in the Seasonal
Adjustment of Economic Time Series", in Zellner, A (ed.), Applied Time Series AnalYSis of
Economic Data, Washington, D.C.: U.S. Department of Commerce - Bureau of the Census,
74-100.

HILLMER, S.C. and TIAO, G.C. (1 982). "An ARIMA-Model Based Approach to Seasonal Adjustment",
Journal of the American Statistical Association 77, 63-70.

INSTITUTO NACIONAL DE ESTADISTICA (1 997), �Indicadores de Coyuntura Ir. Fuentes

Estadfsticas, 30 Noviembre 1 997.

- 23 -
ISTAT (1 999), wLa nuova slralegia di destagionalizzazione degli indicatori congiunturaJi� Note
Rapide, 1 5 April 1999, Roma: Istituto Nazionale di Statistica.

1STAT (2000), wGuida all' utilizzo di TRAMO-SEATS per la destagionalizzazione delle serie
stroriche", Documento # 4/2000 (several authors).

JONES, R. (1980) "Maximum likelihood Fitting of Anna Models to Time Series With Missing
Observations·, Technometrics 22, 389-395.

KAISER, R. and MARAVALL, A. (2000), MNotes on Time Series Analysis, ARIMA Models, and Signal
Extraction" Documento de Trabajo 0012, Servicio de Estudios, Banco de Espana.

KAISER, R. and MARAVAlL. A. (2oo1a), WAn Application of TRAMO-SEATS: Changes in Seasonality

and Current Trend-Cycle Assessmenf, in Proceedings of the II International Conference on
Establishment Sunreys (invited papers), Arlington. VA American Statistical Association. (A,
more detailed version is contained in Documento de Trabajo 001 1 . Banco de Espana.)

KAISER, R. and MARAVALl, A. (2oo1b). Measuring Business Cycles in Economic Time Series,
Lecture Notes in Statistics, 154. NY: Springer-Verlag.

KAISER, R. and MARAVALL, A. (2001c), "Seasonal outliers in time series", Estadfstica (Joumal of the
Inter-American Statistical Institute). special issue on Time Series. forthcoming.

KOHN, R. and ANSLEY. C.F. (1985), "Efficient Estimation and Prediction in Time Series Regression
Models", Biometrika 72. 694-697.

KOHN, R. and ANSLEY, C.F. (1 986), "Estimation, Prediction and Interpolation for ARIMA Models with
Missing Data", Joumal of the American statistical Association 81. 751-761.

LUNA, C. and MARAVALL, A. (1999), wUn Nuevo Metoda para el Control de Calidad de los Datos en
Series Temporales", Boletin Econ6mico. Banco de Espana, May 1999, 3744.

MARAVALL, A. (1 987). "On Minimum Mean Squared Error Estimation of the Noise in Unobserved
Component Models", Journal of Business and Economic Statistics, S, 115-120.

MARAVAlL. A. (1 988), "The Use of ARIMA Models in Unobserved Components Estimation", in

Barnett. W., Berndt, E. and White, H. (eds.), DynamiC Econometric Modeling, Cambridge:
Cambridge University Press.

MARAVALL, A. (1 993), �Stochastic Linear Trends-, Joumal of Econometrics, 56, 5-37.

MARAVALL, A. (1995), "Unobserved Components in Economic Time Series", in Pesaran, H., Schmidt,
P. and Wickens, M. (eds.), The Handbook of Applied Econometrics, vol. 1. Oxford: Basil
Blackwell.

MARAVALL. A. (1 998), "Comments on the Seasonal Adjustment Program Xl2ARIMA", Joumal of

Business and Economic Statistics, 16, 155-160.

MARAVALL. A. (1999), "Short-term Analysis of Macroeconomic Time Series ", in Economics: Beyond
the Millenium, Oxford: Oxford University Press.

MARAVALL, A. (2000), "An application ofTRAMO and SEATS", in Annali di Statistica, special issue on
Seasonal Adjustment Procedures. Experiences and Perspectives, X, 20, 271-344.

- 24 -
MARAVALL, A and GOMEZ, V. (1 992), Signal Extraction in ARIMA Time Series - Program SEATS",
H

EUI Working Paper ECO No. 92165, Department of Economics, European University Institute.

MARAVALL, A and MATHIS, A (1 994), "Encompassing Univariate Models in Multivariate Time Series:
A Case Study", Journal of Econometrics 6 1 , 197-233.

MARAVALL, A. and PIERCE, DA (1 987), "A Prototypical Seasonal Adjustment Model-, Journal of
Time Series Analysis 8, 177-193.

MARAVALL, A. and PLANAS, C. (1 999), "Estimation Error and the Specification of Unobserved
Component Models", Journal of Econometrics, 92, 325-353.

MARAVALL, A. and SANCHEZ, F.J. (2000), "An Application of TRAMO-SEATS: Model Selection and
Out-of-Sample Performancew, Proceedings in Computational Statistics. COMPSTAT 2000,
Heidelberg, Physica-Verlag. (A more detailed version is contained in Documento de Trabajo
0014, Banco de Espana.)

M ELARD, G. (1 984), "A Fast Algorithm for the Exact LikelihoOO of Autoregressive-Moving Average
Models", Applied Statistics 35, 104-114.

MINISTERIO DE ECONOMjA Y HACIENDA (1 998), Sintesis de Indicadoras Econ6micos, Madrid:

Ministerio de Economia y Hacienda (monthly bulletin).

MORF, M., SIDHU, G.S. and KAILATH, T. (1 974), "Some New Algorithms for Recursive Estimation on
Constant, Linear. Discrete-Time Systems", IEEE Transactions on Automatic Control, AC - 19,
315-323

TIAO, G.C. and TSAY, RS. (1 983), "Consistency Properties of Least Squares Estimates of
Autoregressive Parameters in ARMA Models", The Annals of Statistics 1 1 , 856-871 .

TSAY, R.S. (1 984), "Regression Models with Time Series Errors", Journal of the American Statistical
Association 79, 1 1 8-124.

TSAY, R.S. (1 986), "Time Series Model Specification in the Presence of OuUiers", Journal of the
American Statistical Association 81, 1 32-141.

Note: Some relevant additional references concerning the development and first applications of
the model-based approach to seasonal adjustment are the following

f\JERLOVE, M., GRETHER, D.M. and CARVALHO, J.L. (1 979), Analysis of Economic Time Series: A
Synthesis, New York: Academic Press.

HYLLEBERG, S. (ed.) (1 992), Modelling Seasonality, New York: Oxford University Press.

DEN BUTTER, FAG. and FASE, M.M.G. (1991), Seasonal Adjustment as a Practical Problem,
Amsterdam: North Holland.

PICCOLO, D. and VITALE, C. (1981), Metodi Statistic; perl'Analisi Economica, Bologna: II Mulino.

ZELLNER, A. (ed.) (1 978). Seasonal Analysis of Economic Time Series, Washington D.C.: US Dept. of
Commerce-Bureau of the Census.

ZELLNER, A (ed.) ( 1 983), Applied Time Series Analysis of Economic Data, Washington D.C.: US
Dept. of Commerce-Bureau of the Census.

- 25 -
 2. USER INSTRUCTIONS

2.1. MAIN WINDOW

TOOls Bar

The TSW main window is composed by:

It! 1'''1 ..
,.... " , -- " .'..

Navigation Tree
I
��
"" " N�'
:::=:,
FlnolF>.rlod
�=

....,.y.� I

I lOb./Pitnodl
-l
t"s
/
enestnbutes
At Field

j
.,-1 r
r

:;:;;:; "��;�'-.i1---Iter
..,.,

SB�-"::'=��:;:s- - p:,..."'�t"T
Run on Ir-...aI
F,rstOb.r- lJutOb, r-
l
� Program
Ontinn

Tools Bar
Navigation Tree
Series Attributes fields
Iter Parameter
Programs Option

The Tools Bar contains the following buttons:

./ Series permits to load a single series file or a list of files.

./' + Model permits 10 specify an input model fof the selected series.

./' ++ Model permits to specify an input model (the same) for all the series loaded in the navigation tree.
./' Run executes SealsfTramo
../ Output visualizes the standard SeatsITramo output files.
./' Graph shows the graphs computed by the programs.

./' Save permits to freeze the navigation tree saving i l o n a binary proprietary output file (·,gbt).
../ Load loads a WOI'king tree saved.
../' DbXplore Is the manager of a small data base (Db ) facility.

../' About shows the release and authors information of the program.

- 26 -
Series Attributes

N,me

#of Observations (NZ; it includes missing values)

Starting Year

First Observation Period (ex: for monthly series, 1 if Jan., 2 if Feb, ..)

#obsIperiod (12 if monthly, 4 if Quarterly,...)

Iter ParameteG

Iter One series, one model specification (usual case).

One series, several model specifications.
Many series, one model specification common to all of them (the specification can simply be an automatic
procedure).
Several series, one model specification for each series.

The last 3 cases will be explained below; for now we proceed with Iter = 1.

ProoldmOPtion:

Seals: Only SEATS will be exeCIJted

Tramo: Only TRAMO will be executed
TramoiSeats: Both programs will be exeCIJted (the usual case).
Terror: Program TERROR will be exeCIJted

Run on Interval:

FOf a series with observations t = 1 ,.. .. ,NZ, it is possible to select an interval of the sample periOd, and apply TSW only to
the interval. The interval starts at observation - First Cbs-, and ends at - Last Obs -.

2. 2. LOADING A SERIES

Series Button

Oicking on the button series the program opens a standard dialog window in order to select an input file v.1'1lch contains the
series. The following screen is displayed

.. .. • � ,- '!"1 .. � g '* 0
,- 11-.1.-1 .� "- 0 ... a,rlo' •..,., ..... '- .,, - ....

s.... l.ooI·

]
�'Od�j
.t!one.e deI ..cI-ov<x -
1., - ,
�,-l
r
r
]
1
...._del-' I.... n.. '·' 3 �
0-

T.rro, r

F"otStObsr- LestOb.
r-- J

II 1]
- 27 -
The list that appears is the names ofthe series in the directory PROGRAM FILES\TS'N\SER1ES.

TSW accepts two series file fOfmat: text or Excel. The first line of the file should contain the name of the series. The second
lIne should contain four numbers: Number of observations, starting year, starting period of the year, and number of
observations per year (format free). The following lines should contain the numerical values of the series (format free and read
from left to right). Missing values are entered as -99999.

One or many series can be loaded. We shall look first at the case of ONLY ONE SERIES.

If the series of Interest appears in the screen,

- dicking with the right mouse button (r.m.b), the file can be opened and edited.
- d idl.ing with the left mouse button (I.m.b), the series Is loaded to the Navigation Tree.

(In what follows, when no button is specified, it refers to the I.m.b)

If the series of interest is in some other directory, by dicking ., SERIES one can move to the other directories in the usual
WWldows manner, and select the series by dlcklng on it

- 28 -
The selected series is incorporated to the Navigation Tree, the Main window shO'NS

__
_
"�D
9 • " "" Ii!

[series
,... .... "- .Ilc(po.C C!.II '*'-', OlJt,,,,,,,, ij,�

8 t_�e;��� IIPHS�
NemtI

Serie Aliribute..-

IObseMlbon, fii1 StartingYeOf [i983

FirstPeriod r IOb./Period �
I

[ �er - 0 '" ller-' r

a1s
Iter-2 r Iter- 3 r
.., ...�..., �
S'.""�

f
- �oiSe a I

.
Seats ('" T
""""', Trsmo ('" T.= r
n, .... 132) Run on lnt8rvo1
... -...

. .1
n. - Fi/slObs 11 LestOb• 1161

�; �
�
. . . . .
'" ., -- - " - - - , .. I
OJ -
,
rn -- --
--
OJ .. - -
' . .
"
, '" '" '"
,

- 29 -
2.3. MODEL SPECIFICATION

Having selected a series. one proceeds to enler the model.

+Model Button

if
Clicking the button (It is active only a series the navigation tree has been selected) ltIe program shaNs a Tabsheel Sel Window
on
strucllJre (it has the appearance of notebook dividers) whidl permits to set the SeatsfTramo Input parameters.
The window contains three pages with the input p.rameters. For their meaning. dick In the parameter entry, then use F1 for
Help

'" Ii iii 01 .. �
G,"" Lood S... ..- ....

- ] The progf,m 1.,1, lor Ihe logllevel x

sp.cificl1ion. interpolaln mining I 0IheI,1 T_ I
observations (If Iny). and performs
automatic model identification and outlier
deteetion. Thr.. tyPI8 of outliers ara
considered: .ddiriV't outliers. transitory er j1983
thangn Ind IIVIII ihiftl; the leve! m
,igni6c.nc. is sel by Ihe program and
depends OtIlh. lenglh mlhe series. The AWImIbc F'Iood..
.o P.....
riodP;--
... mod'i is nlimMed by ..act
maximum �ktlihood. and foretasls mlhe
sanel up \0 I two-ynr horizon ,If'
'SA •
�
computf'd. The model is decomposed and r
optimal 15timalol1 ind forecasls mlhe
ir r
:���0:���� ��b. ��� 3�:�;����.I.h.e �
o r ta .•

···
': m
.IIl!< .. . r-
r
110 ! • ..... 1

• •.... 1 �
Fl foHeb.
ro , .. . -
..,......,...- �------.....I
""F
i'""iFf r-r
=r=T"
" ,
m'"------"-�
�c-------cc
, ooc-------c,'
�"

The complete description of the parameters is contained In the next section.

The Cancel Sutton permits to exit from the form without saving the model. The OehIult button sets the parameter values to
their Default. The OK button exits and saves !tie model associating il to the selected series.

The first page contains the purely Automatic Procedure controlled by !tie parameter RSA.

- 30 -
 x
.....""
..

0 • .. * ,-' <ll Ii! /il " .. �

,...

==1
."'" ...... H� n_ ,o' d:J", (i.� t',.',. .'.. "'" .... "' - .....

8 ]< ���;"J Aaomatic

"
Pu.udn I Ar_ Modell Ohn I T.,gr I
x

er [i9a3
AW:mabc f'!oo:.:1n p•..--
riodli'""

J
--

'" . . . . _-. -
... ....
"� I � fJ1.I
lC __ 1
�J
.. r-
r

H" -

/ill�
f""""""i'61
" lC ·-1

I ! I .I I. J I I j I
" Fl forfl�
"
!O : . . .. .

., ,
, " '" , .,

The second page contains the ARtMA Model parameters

P . K!M3 Q - f1--::J 0 _ �
81'· ro--3 BQ - � 80 - r;---::J
INIT _ �
AlmoP.,omeI.eu
I'HI//� r:o:r fflI(.iJo r-:o:;- flIP). r-:o:;­
Ffl/J� r--:o:T r� r-:o:;- TH/.'I}- � ,,­

...EAN . [i""""--::::J
$FHI/ll" I:oT pr�» r-:o:;-
WI . [i""""--::::J
Arrna f"",edP",ameI"..
FCT . �
i .lFR{1J> � Dl'>/""" �
� ro-3
.HlR(J-
} ro-::::J .� ro---J � � TYPE . ro--3
.lPS(1J- � JfJ!�'» ro---::J UNITS . ro--3
Fl lO'fleip.

- 31 -
 �Ptoc..b.I __ o__ l r_ 1
-"'-,� 0_. -
OUT - --=
= ==-,
ro---3
I(AST � ro---3 IN'C.
� �

l'-'--'
TNTEIIP .
I'IlAO - ro---3 ro---::::l
lOlF ·
THEG _ r--o
lOUR � rs---:::I TSIG _ � N8ACIt - r--o
NPflEO _ r--o

"...
�llP_ XL

I
� _ � MAXIllAS _ .---0:5
�O •
� ,..- . ,-----, .�.". - I' :3
I"'Vl< �
�
VA • � IINOO - .-----n5 IQM � �
INri � �
INT2 _ r-16i
F l ,.. H"",

The fourth page contains the parameters for a TERROR application.

When aD parameters (different from the default option) have been set, dick on the button �

No
teon the Auloma�cprocedure:
TI'I8 automatic configurations associated with the RSA parameter are described belOW; however they can be modified: after
setting the RSA parameter, enter the desired values of the modified parameters (If the value desired is the default one, you still
have to reenter the parameter).

Cht(;.king Input Parameters

Once the input parameters specifying the model have been entered, by selectlng · L Model 0 • in the Navigation Tree and
cllcklng the right mouse button, one has the option Show Parameters Ust.

iii J
$._
•
1:1._
"
"-
r
I "
"
"

J
... P....- .
... - 0 P' ... . , r
�•• � 2 r ... . , r

S....lA/Tramo
s._ r r.
I
Tram�
T,arng r T_ r
Run on �C- --- - - --

I
-;
F.... Ob. � l.P.Obe �

- 32 -
Clicking on this op�on, a SCleen with only the parameters that h ave been entered is displayed. ( the other par ameters, th at
remain at their default values, are not included ).

0.7

Edit.. De/sUI

2.4 REGRESSION VARIABLES

When, on page 3 ("Others .. ), " IREG " k > 0, a new window is displayed that will set the regression variables and their
associated parameters.

S ..... Name
101 :E IP!(S

'USE R _ �
R""....,.,Eftoct
REGEFF_ �
IIV..-.
"SER _ r---o

- 33 -
 When IUSER = 1 the variable is entered bythe user, observation by observation. REGEFF determines to which
component In SEATS the regression variable will be assigned, NSER = 1, and, as In all cases,
ILONG = NZ + FORECAST HORIZON
(the regression variable should cover the forecasting period). Clicking Inside the "RegreSSion- field. the cells for
entering the variable become visible.

When IUSER '" • 1 . the regression variable(s) is (are) read from a file. The file should be a matrix with ILONG

is contained.
rows and k columns. Each column represent a regression variable. Setting NSER = k and the values of REGEFF and ILONG.
elieking with the r.m.b. inside the "Regression" field. and then on the 'OpenFile" command, a window is opened thai allows us
to load the file from the directory where it

r l....
v. Voriablo

ISEQ .

DELTAS . r---c

f"·....
Ij
HSER_ �

F"1 forHetL

- 34 -
 When IUSER = 2, k outliers are fixed (k = 1, 2, , ..). Only NSER = k needs to be entenKl, and CliCking m ille blank field,
IIle following screen appears.

;-;11'1l!i�1

[-�
I ntecv V.....
--,

rROIIJ
.....",EJIect
I� r-J ISEQ ..

J
,---0 D"" . ,---o

REGEFF_ � DELTAS • ,---0 10105 . r-3

C-_
D....
NSER ,----, .... -
" D�T�
['- - �
T. IC
IUlN6 .. r---o '" "

.... Ok )( Crooj
n b H.

Each outlier is entered in a row containing two numbers. The first one indicates the position of the outlier (observation
number). and the second one the type of outlier (AO: Additive OUUier; Te: Transiloly Change: LS: Level shift).

When IUSER ,. • 2 the regression variable contains an array with hOlidays, that will be combined with the Trading Day
variable. NSER and ILONG need 10 be set, and dicking in the blank field, the holidays can be entered by the user, or, if
the r.m.b. is clicked, read from a file in a directOl)'.

When IUSER • 0 the regression variable will be an intervention variable built by the program. Each intervention variable
has to be entered as a separate regression variable. After setting REGEFF, NSER=1, and ILONG. the parameter ISEQ:: k
Indicates that the intervention variable will contain k sequences of ones. DELTA :: d would Indicate that the operator 1/(1·
d B) wiU be apptied to these sequences of ones, DELTAS :: d. that the operator 1/(1 - do B") wiU be applied to IIle
sequences of ones, and 1010S = 1 that the operator 1/vV. will be applied to the sequences of ooes, Clicking inside the
blank area, the sequences of ones can be entered, The first column contains the starting position of the sequence of ones,
and the second column the length 01 the sequence.

EXAMPLE: Assume a monthly series of a 161 observations, Three intervention variables are induded as regressors. For
each intervention variable, NSER = 1 . and ILONG = 161 + 24 :: 185 (24 is the default number of forecasts lot' monthly
series).

The SCfeen with the inp!.lt data lor the first variable is set as

- 35 -
 - ...

[R""
ro----::J

[-
IUSfR ..
-
lSEQ .. ,------; DELTA. .. r---;-
...... EK.d
,------0 101DS _ �
REGEFf .. r-3
DELTAS

E
RegeW::oI

�
�
-- - .....
"SER _
•

E
85

IlON G .. 11'85

" .... HoIp,

"- ,, x """, 1

II Indicates that the variable presents a level shift at observation 21, and that. starting at period 85, there is a ramp effect lasting
5 periods. The variable will be assigned to the trend-cycle component in SEATS (REGEFFa1).

The screen for the second variable is set as

'"
OELTA.
-
IrteN "....

1'0:: -
[R"" IO-::J I
I
� ISEQ .. ,-----, .. r---o

•
...... EIIOCI
REGEFF .. f2""---:::l
DELTAS
- ,-----, IOIDS .. ,,-----::]

IF
-Aeo-euiorl

SIriIg PotO:In ...."

--
"

E nDHG _ �_

- 36 -
It indicates that the intervention variable consists of isolated spikes every 12 months, starting at period 96, It will be
centered by SEATS and assigned to the seasonal compooent(REGEFF=2); the mean effect will go to the trend-cyde.

The saeen for the third variable is set as

r=""'- -
�������-
r'-V-
UJ5(I'I - ra-:::J
ISEIl • r--T O(UA r--:o:r

E
_

DEUAS - � ID1D5 - rn---3

I'I[lH'f _ �

It indicates that, starting at period 155, there will be a transitory effect, slmilar to a transitory change but with alteming signs.
In SEATS It will be assigned to the irregular component (REGEFF=3).

The fuR Navigation Tree for the examplewould be

iii • or;, 09 ...

,... ,� .Q� o.,, '.t.... o· " .� Jiroph

�f-2 r
,.. - ,

I
tIer - 3
r
r

��.ot. � 1
s.olt{f,OIIID

I �,:: �
RunDn � �'-������-
F"ntOb.� Le.a�
I

- 37 -
 In general, the Navigation Tree has the Classical Windows explore structure. It is possible to remove series/models (Cane
Key), select a series (Clicking on it), expand or collapse the tree structure, rename a series (selecting and clicking on il),
show the associated model (double dick on a model tree node) and perhaps modify it, show the regression variables
(double click on a Regression variable tree mode) and perhaps modify or remove it.

2.5. EXECUTION OF THE PROGRAMS AND OUTPUT FIL.ES

2.5.1 Execution

Once the model has been specified, to execute TRAMO and SEATS, man.: the name of the series in the
Navigation ·Tree. and click in the RUN icon (when running. the program shows an Hour Glass). When estimation is
finished, the (expanded) Navigation Tree looks as foIlO'NS.

The part above " e - L: Preadjusted Series" refers to TRAMO; the part below refers to SEATS. The first time the
selleS name appears it refers to the Original series; the second time it appears it refers to the preadjusted (or
linearized) series, and the series in the graph at the bottom of the main window changes (if TRAMO has made some
correction). The first Model Q) contains the Input file for TRAMO, the secood Model (l) contains the input file thai
has been created for SEATS (with only 2 pages and a reduced number of parameters).

- 38 -
2 5 2 Main Output Files

When the first series name is marked, clicking 00 the OUTPUT icon, the ovtput file of TRAMO is obtained

;:i�, ,I,ll 'dmG Oulpoul f,l� lPtfSp.om) ..fJ E3

TIME SERIES REGRESSION MODELS WIlH ARIHA ERRORS. MISSING VALU'fS Aml au
BETA VERSION ( - )

Iff

VICTOR GOMEZ 6. AGUSTIN MAR/WALL

with the progromming assistonce of G. CAPORELlO

(- ) Copyrigbt : V. GOMEZ. A. MARAVALl ( 1994.1996)

Cl EOd I.
(Spain)
.:.l.E'M I
SERIES TITIE-IPI

� SINCE lONGER fORECAST fUNCTION IS REQUIRED

BY SEATS . NPRED CHANGED TO (24)

ICJCf§JI INITIAL MISSING OBSERVATION �ER

INITIAL MISSING OBSERVATION NUMBER

ISSING OBSERVATION NUMBER IS

ISSING OBSERVATION NUMBER 16

ISSING OBSERVATION mlMBER 138

ORIGINAL SERIES

''' ....0..'' ..... ""<'

..n"�...'''''� .
•

It is possible to navi�ate on the file using the scroll bar, to search for a ""'Clrd in the file (button Find), 10 obtain a
preview or to print the full file.

When the series name appearing the second time is marked, clicking on the OUTPUT Icon, the output file of SEATS
Is obtained. The same facilities are available; the followil1g screen shows the preview.

- 39 -
 IIGII.u aXU..lCtIOIf Ilf ....an... . TDQI SlUIl lin... vauulIu (.,

v. COIlll% . .... IlLllV.u.�.

nUT ,...aT:
AIUIIA IUlllAnC>I

nUll! TITn: HItS' ....,

PIll�Jv.nD "IlK T1U.I!O : Til
IU:TKOD: K.UCIJIVR LIIOIUKOOD

)
<1.. ow c ..........nOH. "161

The two output files can also be fourK! in the directories PROGRAM FllEs\TSW\OUTPUT\TRAMO arK! PROGRAM
FllES\TSVv'\OUTPUl\SEATS, both under the same name: "seriesname.out".

rlIm!.!!g; Every time TSW is initialized. the files in the OUTPUT directory are erased. If they are to be used in later
sessions, they should be stored in some other directory before exiting.

Finally. by dicking on Preadjustment Component the following screen appears

-40-
 The top part contains the columns:
Original Series
Level Shift Outliers
Transitory Outliers (sum of AD and TC outliers)
Easter effect
Trading Day Effect

The bottom part contains the Regressioo Variable Effects classified according to the component they will be assigned
to in SEATS.

2.5.3Out·Tables ( output series)

• When the first series name is mar1<;ed, clicking on the OUT·TABLES Icon, a file is diSplayed that contains the
variables that are prOduced by TRAMO. They cover the sample period (1: HZ) plus the forecasting period (NZ "'" 1: NZ
• FH), where FH = forecast horizoo. The columns contain the following variables
,. column; Date ofobseNalion
2"" column: Original series
3"' column: Interpolated series
4'" column: Linearized series
5'" column: Deterministic mean
6'" column: Trading day effect
7'" column: Easlereffect
8'" column: Additive Outliers
9'" column: Transitory Changes
'"
1 0 column: Level shifts

IPI (Spain) �
DATE Xorig Xint
1-1983 8 3 . 600000000 8 3 . 600000000 86 . 2
2-1983 8 3 . 300000000 8 3 . 300000000 84.3
3-1983 -99999. 000000000 8 8 . 502317985 88 . 4
4-1983
��:�-
-99999 .000000000 8 5 . 173501293
5- 198 3 8 9 . 300000000 8 9 . 300000000
6- 1983 8 6 . 900000000 8 6 . 900000000 87.1
7-1983 8 2 . 800000000 8 2 . 800000000 85.3
8-1983 4 9 . 900000000 4 9 . 900000000 49.4
�S_ I 9-1983
10-1983
8 7 . 800000000
8 6 . 800000000
8 7 . 800000000
8 6 . 800000000
87.8
89.9
89.0
r'r@find .. I
11-1983 8 8 . 900000000 8 8 . 900000000
12-1983 8 6 . 600000000 8 6 . 600000000 87.7

�fri.- !
1-1984 8 6 . 500000000 8 6 . 500000000 87.6
66 . 1
��:���������
2-1984 6 7 . 200000000

,.. --!
3-1964 -99999. 000000000 90.2
4-1984 -99999. 000000000 8 2 . 9 62008317 67. 4
5-1984 69 . 800000000 8 9 . 800000000 89 . 4

! llQo !..
6-1984
7-1984
85 . 600000000
8 5 . 800000000
8 5 . 600000000
8 5 . 800000000
87 . 1
87.2
8-1984 5 3 . 300000000 5 3 . 300000000 53.0
9-1984 B S . 400000000 B 5 . 400000000 89.1
10-19B4 9 1 . 600000000 9 1 . 600000000 90.8
11-1964 9 0 . 200000000 9 0 . 200000000 B9.8
12-1984 8 2 . 600000000 8 2 . 600000000 85.2
1-1965 8 8 . 600000000 6 8 . 600000000 87.6
2-1985 8 4 . 600000000 8 4 . 600000000 65 . 4
3-1985 8 9 . 200000000 8 9 . 200000000 89.7
4- 198 5 8 4 . 800000000 86 . 9

�
8 4 . 800000000
5- 1 9 85 9 0 . 600000000 9 0 . 600000000 89 . 8
83 . 800000000 8 3 . 800000000 86
, 0 c""nnnnnn on cn... ... ...... ... ... ...
6-1985
]
•
,"
0 "

I
0 00 "
•

- 41 -
 When the series name marked is the one appearing the second time, did<ing on the OUT-TABLES ieon, a file is
displayed with the series produced by SEATS, extended over the forecasting period. The COlumns contain the
following series:

, . column: Original series

� column: Final Trend-cycle
�
3 column: Final Seasonally Adjusted series
4'" column: Final seasonal component (or factor)
5'" column: Calendar effect (Trading Day effect + Easter effect + Holiday effect)
6'" column: Transitoty-irregular component (exxnbined effect of transitory and irregular exxnponeots or factors)
.
.,. . column: Preadjustment component
8 column: Extended residuals (exxnputed
'"
by SEATS).

IPI (SplI.iD)
DATE SERIES TRENDCYClE SASERI
1-1983 83.6000000 8 3 . 8369565 83.015
2-1983 83.3000000 8 3 . 9489700 82.278
3-1983 8 8 . 5023180 8 4 . 0997075 83.125
4-1983 8 5 . 1735013 8 4 . 2771294 83.37
5-1983 89.3000000 84.4096315 84.075
6-1983 86.9000000 84.4300615 83.666
7-1983 8 2 . 8000000 8 4 . 3717359 82.992
8-1983 4 9 . 9000000 8 4 . 3 936881 8 1 . 94 7

,, -
9-1983 8 7 . 8000000 8 4 . 5665536 83.726
10-1983 8 6 . 8000000 8 4 . 8 182525 8 3 . 187
11-1983 8 8 . 9000000 8 5 . 1506335 83.336
Qfind_. 12-1983 8 6 . 6000000 8 5 . 5057196 85.433
1-1984 8 6 . 5000000 8 5 . 7461818 84.419
e.EM<. 2-1984 8 7 . 2000000 8 5 . 8478348 85.515

" -I
3-1984 9 1 . 0806465 8 5 . 8232749 84.253
4-1984 8 2 . 9620083 8 S . 7363735 84.289
5-1984 89 . 8000000 8 5 , 6768516 83.706

ILll ..g 0
6-1984
7-1984
8-1984
8 S . 6OOOOOO
8 5 . 8000000
53.3000000
8 5 . 7011273
8 5 . 8284170
85.9039360
83.851
8 4 . 255
86.996
9-1984 B 5 . 4 0000oo 8 5 . 7472564 B4.837
10-1984 9 1 . 6000000 8 5 . 5075344 83.239
11-1984 9 0 . 2000000 85 . 4 157432 83.ln8
12-1984 8 2 . 6000000 8 5 . 4634261 8 4 . 36 1
1-1985 8 8 . 6000000 8 5 . 5063889 85.661
2-1985 8 4 . 6000000 8 5 . 4967017 83.770
3-1985 89. 2000000 8 5 . 5316533 84.628
4-1985
5-1985
8 4 . 8000000
90.6000000
85.6413410
8 5 . 7927572
84.429
85.o �
IniDiD "8�0e""'.cIIo M"". 1 8U_IOJ" Tr_Se«I···IJ�T� I!:WICIOd\llOld j
l$\INa.. ["iCt24

The two files can be found In the directories PROGRAM FILE$\TSW\OUTPunTRAMO and PROGRAM
FILES\TSw\oUTPunsEATS, under the names table-t.out and tab/e-s.out. They can also be saved as Excel files in
which case they are deposited in the directory SAVED.

2.5.• SummaryOutput and Out-Matrix

In the case ITER:O ( one series, one input file) the files Summaryt.bct and Summanrys.bct are available in
OUTPUT. They contain the following summary of the TRAMO and SEATS results.

a) Summaryt.txt ( results from TRAMO):

�:
So" Execution time (in seconds).
"" Number of observations in series.
Lam: o if logs have been taken; 1 if levels.
Mean: o if model has no mean; 1 if it has a mean.
p,d,q,bp,bd,bq: orders (P, D, 0) (BP, BD. BQ). of the fitted ARIMA model.
SE(res): Standard Error of Residuals.
Q..val: LjOng-Box-Piert:e 0 statistics for residual autocorrelation.
N-test Bowman-Shenton test for Normality of the residuals.
SK(t): t-value for Ho: Skewness of residuals " 0

-42-
Kur(t): t·value for Ho: Kurtosis of reslduals:z 3
as: Pierce Os-test fof seasonal autocorrelation in residuals. (O)
02: a-statistics for autocorrelation in squared residuals.
Runs: I-test for runs (randOmness) In signs of residuals.
(-) when the lag·1 2 autocorrelation is negative, as is unrelated to seasonality and the value -99.99 is printed.

ARMA Parameters

The order is the following.

Estimate of the regular AR polynomial ( 1 +4118�B2+4I)B3 )
Estimate of the seasonal AR polynomial ( 1 +4I,B' )
Estimate ofthe regular MA polynomial ( 1+9IB+9282+9}Bl )
Estimate of the seasonal MA polynomial ( 1 +0,81 )

The associated t-values are also given

Deterministic Effect(total)

m Number of Trading Day variables.

EE: Presence I absence of Easter effect.
# OUT: Total number of Outliers
AO: Number of Additive OuUiers
TC: Number of Transitory Change outliers.
LS: Number of Level Shift outliers.
REG: Number of ( additional) regression variables.
MO : Number of missing observations

Caltndar Effect

TD1, .. .. , T06: Estimators of Trading Day variable effects.

LY: Estimator of Leap-Year effect
EE: Estimator of Easter effect

The associaled I·values are also provided.

Detected and corrected outliers are listed; firsl Additive Outliers, then, Transitory Changes, and finally, Level Shifts. For
each outlier, the date and assoc. t·value are given.

Regression variables

The regression variables ( their total number equal 10 IREG ) are listed in the order in which they were entered. The
coefficient estimators and assoc. t-values are printed.

b) Sa.mmarys.txt ( results from SEATS):

.
Preadj. Preadjusted with TRAMO ( Y I N )
Model Changed; Model passed by TRAMO has been changed by SEATS ( Y I N )
Approx. to NA: The model used to decompose the series is an approximaUon to an original model that
provided a non-admissible decomposition. ( Y I N ).
New Model: When the model from TRAMe has been changed by SEATS, the new model orders are
printed in these columns.
SO{at): Standard Deviation of the ( recomputed and ) extended SEATS residuals.
Specl factor: Spectral Factorization that provides the model decomposition (0 • OK I E = ERROR).
Check on ACF: Check on the comparison of variances among the theorical components, the theorical
estimators, and the empirical estimates ( 0 I E ).
Determ. Compon. Mod . . The stochastic SEATS component is modified by some of the deterministic effects
captured by TRAMO ( Y I N ).

SO{ innov): Standard deviation of the component innovation. Expressed In units of the series ( logs if lAM=O ).The
components are: TC • Trend-cyde; S = Seasonal component; Tran: Transitory component; U:
Irregular component; SA: Seasonally Adjusted Series.

SE est ( cone.) : Stalldard Error of the concurrent estimator (TC and SA series ).

- 43 -
SE rev ( conc.) : Standard EITOf ofthe tolal revision error In the concurrent estimator ( TC and SA series ).

Convergence ( In % ): % reduction in the variance ofthe revision error ofthe concurrent estimator after 1 and 5 years of
additional data are available (TC and SA series ).

Signif. Seaso. (95 % ): number of periods per year for which seasonality is significantly different form a ( at the
95 % !eve! ). Given that the estimation errors vary, significance is assese s d for:
Historical estimatiOn (Hlsl)
Last observed year ( Prel.)
Qne.year·ahead Forecast Function ( Fore.)

fm.!:!!J...!I
SE: r. ofg. : Standard Error of the rates of growth of the estimaled component (in per cent points).
T11: Pefiod..to-period rate of growth (TC and SA series )
T1 Met: Annual rate of growth, centered at the last available observation ( TC, SA, and Or\gil"l3l
Series ).

( For an additive decomposition, • rate..of..growth· should be rep!aced by "growth ·, expressed In the series units ).

Diff. annual means: Average of the absolute value of the differences between the annual means of the original
series, SA series, and TC ( in % ).

AA example of a summary file is the fOllowing:

IPI(Spoin)
NZ -161; PERIOD-01-1983/05-1996; MQ-12;

Model fit
Sec, Nz [.0111 Meon P D Q BP BD BQ Sf(ns) Q-val N-tea·
3 . 17 161 a a a 1 1 a 1 1 0 . 02 2 4 . 78 2.6

ARMA Pi5rarneters
PHIl It } PHI2 It } PHI3 It} BPHI
O . 0000 I 0 . 00 ) o . 0000 0 . 00) 0 . 0000 0 . 00) 0 . 0000 O.

Deterministic Effect (total)

TD EE /IOIJI AO 1'0 LS REG MO
Q find _. 1 1 1 0 0 1 , 5

I!. Eri"·· Colendor Effect

TD'
.. --I
TO! It} T02 It} It } TD.
o . 0059 I 12 . 76) 0 . 0000 0 . 00) 0 . 0000 0 . 00) 0 . 0000

I l\.l;l,.. I
Ou·tliers
LSOl(1192. -3.83)

Regression Variables
Reg 0 1 ( -0 . 0242. -5.86) ReS02 ( - 0 . 0 1 6 1 . - 3 . 69) ReS 0 3 ( - 0 . 0004. -0.32

- 44 -
c) Out-matrix:

When TSW is run on an input containing many series and/or many models ( cases ITER:#: 0 ) the summary.b:t files are
not produced. Instead, the summary results are stored under the icon Out Matrix.

Each matrix corresponds to one of the rows of the Summaryt and Summarys files, with the rows of the matrix referring to
one of the series I models in the input. The following is an example.

-!MiF 'al...1
T'_!Sd... 1
f"oled M_ !AlmoP.._... 1 o__ E_1 c•....;. E_l 0..-. 1 FI�I
Ha <_ !I . _ SI,¥ ••• Q_ V d
U. 0 0 O . DUSU zO . n
D. n Z34 O . D03UO 3�.?'
D . 23 le O
O . uuu

.. =-

The matrices can be stored in Excel. For the case of a simple series. by setting ITER-2, the summary.txt files are replaced by
the matrices, each one containing a simple rcNI.

2.6. GRAPH

Clicking on the GRAPH icon a graph utility form is run, from which it is possible to visualize. print, Of" save the graphs produced
by SealslTramo.
The window is divided into a Navigation Tree (similar to the one of the Main Window) and a Plotting Area. It Is possible to
expand or collapse the tree by clicking on it. The graphs are divided in sub-trees: SERIES, ACF, FILTERS, SPECTRA,
FORECAST, and REGOUTSE-

Clicking on + SERIES, for example, the graphs available show up in the Navigation Tree. The main graph rundonalities are the
following:

- 45 -
 •

It!
fJ.TERS
� FORECAST IPII� UtlEARIZED SERIES
IiJ-REGOUTSE
s.SERIES
NOI.oIlecclocl",.
:- e>d."d",I<.<iduoh
;- 1inaI ... .....
;.. 1inaI ...... lec:I<n
: IinaI ht>Oqde
:" -.w �
:- ..... oHII""lcu
;.. ...... fIogo)
� ....ieo; ......w.w '_
:- --- �
;- .... fI'IO<"II:ht; ,_uI
:- atoeha.lic io,.g<./.oIl..,
atocN.Iic ......
otCt;haatoo;: ....oonoIl.
; atoc:t'IMIic �enok)dt
:- _d>gdo,joeo'l«:ll«to
;.. -� .... Ifog
: �� ..
:..- nndqdoo; """""'" '.
:-�eI1ec>t. """,ocI �. . .� . ..... :
:... ruopalal:ed terie.
: !M.tad ..... . . . + .. .. +. ······· i ·

:-:::� .....
!B
;- .......
SPECtRA
� :-"-�.
C---�-- . --�
. -C:- .
-C�.��
.-.i;---�
-
�-C.+.
,.
iC---::
-
.

./ Plot Select a node on the tree and double-dick on it to plot the graph. A new graph will be plotted in the same way
(cleaning the plotting area).

./ OVerlay: Selecl a node and dick with the r.m.b.; it Will show a menu with !he Item Add. Click on it in order to overlay the
selected graph with the one in the plot area,

. . I
�·F1lTERS
!!i"FtIRECAST IPI/SI*>I UNEARIZEO SER IES
� �:���TSE H 1Nl\l. lREND -CVCl.E

:�� I
� -""'*'- '"
: _ _ .".... ,_.
'"
,�
'00
� ,
:'- "' '''' l\agol
!- ,.",,: � ,•• 00

" J/1 ,f..

,
;- - .....- .,
..... ...... ,_01,
_�I« .. � ..
_u_
_ _ I. "
-­ ro
�..tt.g dll'oIfec>tI.od'
;.- �-"'---(Iog .. ,. -+- ......
!- �_..,. cooroonoo
� �-.""".: � .. ..
..-- .tlecloCCll<'"
; �od...n.. ..
: -'-I .... ..
:-.,.;g;.,. .....

;;;, ,;;, J
: •
�
:- ...- ....- -.
'-
It! SPECTAA �I
,­

- 46 -
./' Zoom In: Drag a rectangle on the plot area starting from the left-top comer.
./' Zoom Out: Drag a rectangle on the plot area starting from the right-bottom comer.
./' (x,y) coordinate: It's possible to visualize the (x, y) coordinate of a point on !he graph using !he Shift+Left mouse button
combination.

./' Graph Panning: Clicking the r.m.b. (with the cursor on the plot area) and moving the mouse pointer will produce an
horizontal or vertical saoll of the graph according to the mouse pointer movemenl

./' Save: Clicking on the button it is possible to save the plot area in two different formats: Bitmap r.bmp) or Windows Meta
File (o.wmf). Both are standard format and it is possible then to include the graph in your Word, Excel, .. . doruments.

./' Options: the option button gives the possibility to personalize the look of the graphs. The following options are ava lable
for each single graph In the plot area

Add points on the graph in different format: Square, Cirde, Triangle etc.
Change the line foema!: Solid, Dash, Dot, DashDot, etc.
Change the �ne COlor.
Change the line width.

For the entire graph It is possible to:

Transfotm it into a 3D graph.

Change the fonts and the color of the title.
Change the printer options. It is possible to print in PortraitlLandscape and define the quantity of subplots to
put in a single sheet (2 Max on the Xsheet·axis; 4 Max on the Ysheet-axls).
Change the X·Y Scale.
Backup the graphs in order to overlay them with those of a new model. To do that, click on the option
Backup. Then run the same series with another model, dick on graph and on the SQuare Overfay with Old
Models. Then dooble click on the selected graph. The graph tOt this and the previous model are plot.

The arrays of the graphs are stored in the directory PROGRAM FILES\TSW\GRAPH

2.1. SAVE I LOAD

Clicking 00 the SAVE Icon, the navigation tree is save<l on a binary proprietary ootput file (o.gbf). In this way, series with their
models can be save<l. (The fite is stored in the directory PROGRAM FILES\TSWlSAVEO).

Clicking on the LOAD Icon, the files in the directory SAVED are displayed and can be restOted In the Navigation Tree. (If the
saved file was moved to another working directory, it can be accessed in the usual manner).

When saving as Excel files the files in out·Tables or out-Matrix, the .xls files are also stored In the SAVED directory.

-47-
 ��
��-�
, � ModeI O
e-� �e<lS.",
B IiI IPI(s�1

I
&r�lo4ode10

I
I:I.CIIfboe del ordw<:r .I_;;;;_""�"----"- �
,,-, - ,
�f-3
r
r
1
1
G_darcomo� "
1_"
_
::;•
'•:;0
, ·""
..------,,
3 � [
erno,ISe otl r-
T.emo r T_ r

Run on Irt\eoWI
FirslOb. r---;-

2.B. MANY SERIES ANDIOR MODELS: ITER parameter

The previous pages refer the case iter = 0, in which a single series is treated with a single model specification.

• ITER .. 1 One series; several specifications. Having set Iter .. 1 and selected a series, the models are entered by
Clicking on the MODEL++ button. CNcking on RUN, all cases are estimated. The extended Navigation Tree looks like

• " • '" <il I!<

.- .- .� Q� OIlT_, Oo.CM� Ilr,p,

1;1 SeriaL.iOl

II""''''''
b
: s Iil 1P1fS""",,

, 'I>: e
1:5'"' '''' .' -
:: :;. MocW 3
...... ,
: . Mod,,.. IOb'.M'1IIOnl � SUubngY.$I' fi9s3
9-� Pr�ed S";..
F,l-Ii! MODELl
: e.@ MocW O
FlfltPlirioo r IOb./Pllriod�
: ' ei �c--.t,
� MOOEL2
: 9 � ModeIO

[seelSeab
sfTreJTlO
: · -ei � c-·
B Ii! 1oI00[lJ
:. 9 @ ModeIO

FntObl1l
" t9l Pre� Component,
9-Iil IolODEl4 r
: s-� Modli O Tremo r
: - ei Pre��,
e Iil IoIOOEL5 Run on InttIMIl
8-� loIodo10
. lIIJ! .....-.- �

- 48 -
Glicking on OUTPUT, all TRAMO-SEATS output files can be opened. Clicking on OUT·TABLES, the tables are lisled In a
single file, following the sequence MODEL 1, MODEL 2,.,. Clicking on OUT-MATRIX each matrix contains the summary results
for aU models, Clicking on GRAPH provides a selection of graphs: for each model, only the original series, final seasonally
adjusted series, and final trend-cyde, as well as their forecasts, can be plOt.

• ITER ; 2 Several series; one model specification

Having set Iter = 2, pressing the CtrI key and dicking on the series, several Series can be selected. Clicking on
MODEL++ a model spedfication is enlered thai will be common to all series (this specification can be. for example, RSA
= 4; that is. automatic treatment for all). Oicking on RUN. aJl series are treated and the following Navigation Tree is
""","'"

• I&l G3 Ii:! &!

.- g.q,... OI/lT_. D'4�"''' .li.1.p,

!;I s .... u..

$ !il IF'I('5�
·$ "'",,", 0
9 � EXP-IMP
S.n. Allritwt..
6t "'''''"' 0
8 rill CA ('5-.1
- Sf "'",,", 0 # Ob••"-'OIIO". � S1orbngV .,.. �
•

(;I1i!l "'_�AI..P ('5......:l

El I: �s_
_$ "'",,", 0 Ftt.t�nod r- #Ob./�riod �
t;l IiI IF'I('51WW11
8·8 "'",,", 0 !t.,PDnlmew,
c 6i! �....... �.
€Iiii EXP.j�p 1ItI , - O r ",-, - , r
h••• � r

S.a.ttamO
lSeDl-tT,
, (;I til 6i!
"'",,"
,0 lie,. 3 p
; "'.�_ c-..
<"
iii iii CA ('5"""1
8'& �"""' 0
, , .8i "'�_�. T.erno <"
15 ::E "'_�.Al..Pts�
Run on Inteoval

I FirstOb. 11
13 8 "'",,", 0
' 5i ��

.>

eo 1 00 , 20 , 4.0 H
I
ll
·�I
, eo 200 :no

Selecting one of the series and dicking on OUTPUT, the corresponding outp\Jt file can be accessed. The button OUT.
TABLES, woks as in the case Iter = 1 (one single file, listing firsl the table for the first series, then, for the second series,
and so on), and OUT-MATRIX yields matrices of the type already described,

GRAPH yields, for each series, the same selection of graphs as in the case Iter ;1.

• ITER • 3 Several series, each one with a different model

Having set Iler ,. 3, several series can be selected as In the previous case. Choosing a sene in the Navigation Tree,
MODEL sets the modal specificatiOn for thai series. RUN eXeaJtes TRAMO-SEATS; selecting a series and clicking on
OUTPUT provides the COfresponding outpul file. OUT·TABLES, OUT-MATRIX and GRAPH are as in the previous case.

2.9 DATA BASE FACILITY:" DBXPLORE

The dataDase facility is intended to help in routine treatment of groups of series. The set\es are stored together with the
model specification ( orders of ARIMA model, dale and type of oudiers, type of TOJEE variables, set of regression variables),
Then. a new observation can be added and the coefficients of the model saved updated.

Oicking in the button ObXplore, the user can access a DataBase screen. Three windows appear in it. The first one,
SeriasSet, is related to the directories or records in which the DataBase is organized, When the tJSe( selects one of them, the
second window, Series, will show the series in that directory. In the window SeriasSet several lil1le buttons are available. The
firsl button takes the user 10 the firsl directory of the DataBase. The second one takes the user 10 the previous directory, the
third one, takes the user to the next directory, the fourth one lakes the user to the lasl directory or record, the fifth one Is used
10 insert a new direclory, the sixth one is used to erase a directory, the seventh one permits the user 10 edit (change the name
of the directory, the eighth one saves the change In a directory thai has been edited, the ninth one cancels edition of Ihe
directory, and the last one refreshes the series of the directories.

- 49 -
 r-
S�__ -- --
-- --
-- --
-- --
-- -
--
"""-S...-­
"""
........
1<o40rw,0 agg.et end nerest rates
Ph;etrdeloer
Stock PICa
u...�c.tIegonet:

M .. t· " + _

'-----'--

Sarrde SIZe Sla!li100bs So. Yea CataP

,'" '978 "
'55 ""
252 ' 970

J
"
EXPORTS [SPIin) '55 "" "
"
__ "5 �7 ••

• • , " ('

The second window, Series, shows the name of the series, Its sample size, starting observation . starting year and
data perlodicy. The small buttons in this window are similar to those In the window SeriesSet.

On the two tables you can navigate (scroll up,down). add reCOfds (categories or Series). or remove and update
1tIem. It is also possible to add new series values. Some hidden tables are also defined (models, regs) v.tIich contain the model
and regression namelist associated to the series. It is possible to move/done a series to a different SeriesSet (rigth-button
mouse dick on Series Grid).

The third window of DbXplOf8 Is called SeriesValue.For the selected series. it shows the values of the series and
the associated dale.

When the user wants to run in TSW a series in the DataBase by pressillg the button + Series List and selecting the
series, clicking on the left button of the mouse, an option called Add to SeriesUst will appear in the screen. This option will
lake the series to the navigation tree of TSW. To select more than one series. use tile Control key.

""":Add Jo �"""£ I'£1 1III[oiJ iE3

- 50 -
 r es
Alternatively, in order to add series in the Series Ust to the DataB ase select the series (or the entire Serieslist) and right­
mouse-dick. A small window shows up with two options: Save to DataBase and Update DateBase. In the first case, the user
can s ave a new se i in the D ataB ase, and in the second case a serles in the D ataB ase will be upd ated.

iii
....
•
..... M
� I �
B�
�
Q- OlOt
m!D
Tables
<Sl
O�M&ltjw
Ii!
Q,..,
t'il
....
Ia
S..
..
Qb_
II
"""

1
I? Se\ie$lJ$I Series Nlmle
e-1iJ PAOO Of MANUFAC METAl. [ITALy) I�A!S (.IAPANI
'--@ ModeiO

rSerie Attributes

# Observ&ions I2S2 Sterting Year f"i""97O

�� I
�SAJ
,.... ,
s-IHJ CPI [Spairl)
r lOb. I Period rz-
'....� ModeiO
8-:L f'I�edSerie*
IterParamete
PAOOOF MANUFACMETAL OTALy)
$ Her - O r
II EXP4MP

[ �':� r.l
I±l- EXPOATS�) Her - :! r
1fI- GNP [USA)
I!J CPI[Spairl) S'

[
T,"",""" ...

2S��
Tramo r TelTOf r_

;:
Run on Interv!l.l == ======:
FirstOb!l 1l lall Ob �
EXPORTS (JAPAN)

::�::ttno 20 40 60 90 100 120 140 160 190 200 220 240

After selecting the option Save to DataBase a new small window appears indicating the D ataB ase directory in which
the user wants to save the series. The DataBase gives the user several predefined directories; the user can of course create
addition al ones.

Q • "" '" IIiJ .. .. Q

,.- .- � .." " .., ..., ,� Qb)(pb• """'
s."•• N"",.
iCf'l ('Spe;-.J
l

15_0 ...--S...o5.

tI.r - ' r
OZ!!!l;l"! �_ �r·3
"- r
10.4.,.,.,. ��o. _ ""'..�"' ',""eo
Pric. ,,,-,,,
Stod<..-ice. TrlllYlO/SeOlt Co"
U...-do_ C"'�

��::_
�
T.rror r
J
----L-
____
____
____
__

First 010• .----;- L"5'lOb. I 1ao "]

- 51 -
 TSW otfe� to the userthe possibility of editing a series in the DataBase. On the bottom of the screen there are several
buttons. To edit a series press the ( fourth from the right I Edit Record button. then the values of the series can be modified in
the SeriesValue screen by cliking on the left buttom of the mouse. A new value can be added by diking in the +AddValue
button. The Changes can be saved by clicking the Post Record button. which is to the right of Edit Record.

2.10 PROGRAM TERROR

The program is an aplication of TSW to the detection of errors in new data reported to sets of time series. The program is
also available in the DOS version of TRAMO. with more complete documentation.

Errors in the new data are detected as abnormally large forecast errors in the l-period-ahead forecast computed ignoring
the new Observation with TSW run in an automatic: manner.

TERROR is run by:

a) Choosing · Terror· as the program option in the Main Window of TSW;

b) entering the set of series ( pertlaps from an Excel file I:
c) clicking in ++ MODEL, and then selecting the · Terror· sheet:

AlAometicProceO..e l MMMocWI Other: Tenor I

rT- POO'",::" -
'-1 _ r---r

N..... ..

Fl la Help.

Setting the approplate parameter values ( see next section ) and dickill9 in OK. the program is executed. The only outpul
produced is the file listout in OUTPUT. II contains the position in the file of the series that have produced abnormally large
forecast errors ( according to the values of Kl and K2 selected), as well as the associated t· statistics.

- 52 -
 3. INPUT PARAMETERS

3.1 AUTOMATIC PROCEDURE

RSA =1 As RSA=3 below, but using always the default (" Airline ") model ( no automatic
model identification ).
=3 The program tests for the log/level specifICation, interpolates missing
observations (if any), and periorms automatic model identification and outlier
detection and correction. Three types of outliers are considered: additive
outliers, transitory changes and level shifts; the level of significance is set by the
program and depends on the length of the series. The full model is estimated by
exact maximum likelihood, and forecasts of the series up to a two-year horizon
are computed. The model is decomposed and optimal estimators and forecasts
of the components are obtained, as well as their mean squared error. These
components are the trend-cycle, seasonal, irregular and (perhaps) transitory
component. If the model does not accept an admissible decomposition, it is
replaced by a decomposable one.
=4 As before, but a pretest is made for the presence of Trading Oay and Easter
effects, with the first effect using a one parameter specification (working I non­
working days).
=5 As RSA=4, but a pretest for the presence of Leap Year effect is added.
=6 As RSA=4, but the Trading Day specification uses 6 parameters (each working
day, plus weekend).
=8 As RSA=6, but the leap year effect is added.

3.2 TRAMO PARAMETERS

3.2.1 ARIMA model

P =0 (Default) order of regular autoregressive polynomial.

= 1, 2, 3

Q =1 (Default) order of regular moving average polynomial.

= 0, 2, 3

0 =1 (Default) order of regular differences.

= O. 2

BP =0 (Default) order# of seasonal autoregressive polynomial.

=1

BQ =1 (Default) order of seasonal moving average polynomial.

=0

BD =1 (Default) order of seasonal differences.

=0

- 53 -
INIT =0 (Default) All ARIMA parameters will be estimated.
=1 Some parameters are fixed.The location of fixed parameters is entered setting :
JQR(I)=1 ; JQS(I)=1 ; JPR{I)=1 ; JPS(I)=1 ; The fixed values of the parameters
are entered as TH{i)=fixedvalue, PHI(i)=fixedvalue, . .
=2 Values for all parameter input and n o parameter estimation i s done. Parameters
entered in TH, 8TH, PHI, BPH1.

PHI Estimates of regular autoregressive parameters (Default: All -.1). Not input if
INIT=O. If ( IN IT=2 ) or ( INIT= 1 , JPR(I)=1 ), PHI(I)=k fixes the I-th regular AR
parameter.

TH = Estimates of regular moving average parameters (Default All -.1). Not input if
INIT=O. If ( I NIT=2 ) or ( INIT=1, JQR(I)=1 ), TH(I)=k fixes the Hh regular MA
parameter.

BPHI Estimates of seasonal autoregressive parameters (Default: All -.1). Not input if
INIT=O. If ( IN IT=2 ) or ( INIT=1, JPS(I)=1 ), BPHI(I)=k fixes the seasonal AR
parameter.

BTH Estimates of seasonal moving average parameters (Default All -.1). Not input if
INIT=O. If ( INIT=2 ) or ( INIT=1 ,JQS(I)=1 ), BTH(I)=k fixes the seasonal MA
parameter.

JPR(I) =1 When INIT=1 parameter number I in the regular autoregressive polynomial fixed
to the value set in PHI(I) (it is not estimated).
=0 (Default) Parameter not fixed.

JQR(I) =1 When IN1T=1 parameter number I in the regular moving average polynomial fixed
to the value set in TH(I) (it is not estimated).
=0 (Default) Parameter not fixed.

JPS(I) =1 When IN1T=1 parameter number I in the seasonal autoregressive polynomial

fixed to the value set in BPHI(I) (it is not estimated).
=0 (Default) Parameter not fixed.

JQS(I) =1 When INIT=1 parameter number I in the seasonal moving average polynomial
fixed to the value set in 8TH(I) (it is not estimated).
=0 (Default) Parameter not fixed

IMEAN =0 No mean correction.

=1 (Default) Mean correction.

LAM = 0 Takes logs of data.

= 1 (Default) No transformation of data.
= ·1 The program tests for the log-level specification.

FCT =1 ( Default ) Real value. Controls the bias in the log/level pretest.
>1 Favors levels;
<1 Favors logs.

TYPE =0 (Default) Exact Maximum Likelihood (for SEATS and TRAMO).

=1 Least Squares (conditional for SEATS, unconditional for TRAMO).

- 54 -
UNITS =0 (Default) The units of the original series are preserved.
=1 If the series units are too small ( min Zt :2:: 1 04 ) or too large (max Zt � 10-3 ), the
series is rescaled.

3.2.2 Calendar Effects

lEAST =0 (Default) No" Easter effect.

=1 Easter effect adjustment.
=-1 The program pretests for Easter effect.

ITRAO =0 (Default) No Trading Day effect is estimated.

=1 # of (M, T, W, Th, F) - # (Sat, Sun) x 5/2.0ne parameter specification.
=2 Iv;:; the previous case, but with leap-year effect correction.
=6 # M - # Sun, # T # Sun, ..... , # Sat # Sun. Six parameter specification.
- . -

=7 As the previous case, but with leap-year correction.( Seven parameter

specification . )
= -1 As ITRAD =1, but a pretest is made.
= -2 As ITRAD =2, but a pretest is made.
= -6 As ITRAD =6, but a pretest is made.
= -7 As ITRAD =7, but a pretest is mad�.

lOUR =6 (Default) Duration of period affected by Easter (# of days).

=k a positive integer.

3.2.3 Outliers

IATIP =0 (Default) No correction for outliers.

=1 Automatic detection and correction for outliers.

AIO =1 All outliers are treated as addilive' outliers or transitory changes (in this way
the level of the series is preserved).
=2 (Default) Additive outliers, transitory changes and level shifts are considered.
=3 Only level shifts and additive outliers are considered.

Two integer parameters, INT1 and INT2, can be used to define the interval (INT1. INT2) over which
outliers have to be searched. By default

INT1 = 1 ; INT2 = NZ ( number of observa�ions in series )

When INT2 = -k < 0, outliers are automatically detected and corrected in the interval (INT1, NZ-k).
Then, the detection procedure is applied to the last k ob' servations, and if some outlier is detected a
warning is printed, but no correction is made.

IMVX =0 (Default) The fast method of Hannan-Rissanen is used for parameter

estimation in the intermediate steps of the automatic detection and correction of
outliers.
= 1 Maximum likelihood estimation is used.

VA = 1< A positive real number. Sets the critical value for outlier detection.
The default value depends on NZ:

if (NZ $= 50) then VA = 3.0

if (50< NZ < 450) then VA = 3.0+0.0025·(NZ-50)
else VA = 4.0

- 55 -
INT1, INT2 See parameter : AIO.

3.2.4 Automatic Model Identification

INIC =0 (Default) No automatic model identification is performed for stationary model;

=3 The program searches for regular polynomials up to order 3, and for seasonal
polynomials up to order 1 ( stationary model );

IDIF =0 (Default) No automatic model identification for non·stationary roots.

=3 The program searches first for regular differences up to order 2 and for
seasonal differences up to order 1 .

TSIG =1 (Default) Minimum t for significant mean.

=k a real number 0 < k < 2.

3.2.5 Otherparameters

INTERP =0 No interpolation of missing observations.

=1 Interpolation of missing observations with the fixed-point smoother.
=2 (Default) Interpolation of missing observations is made through regression
("Additive Outlier Approach").

When automatic model identification is simultaneously performed, missing values are interpolated
using the addillve outlier approach.

NBACK =0 (Default) No out-of-sample forecast test.

= k<O K a negative integer, then [k[ observations are omitted from the end of the
series. The model is estimated for the shorter series, one-period-ahead
forecast errors are sequentially computed for the last k periods (without
reestimation of the model), and an F-test is performed that compares the
out-of-sample forecasts errors with the in-sample residuals.

NPRED =k a positive integer, # of multistep forecasts to compute when only TRAMO is

used; When used with SEATS, NPRED is automatically set to MAX(2MQ,8),
where MQ is the number of observations per year.
=0 (Default).

3.2.6 RegreSSion Variables

IREG =0 (Default) No regression variable.

=k A positive integer.
k = # of regression variables entered by the user (regvariables with IUSER
1) + NSER for the (NSER) variables entered as a matrix ( with NSER
columns lin an external file (regvariables with IUSER = -1) + # of "a priori"
specified outliers ( NSER in regvariables with IUSER = 2) + # intervention
variables built by the program ( regvariables with IUSER = 0, ISEQ > 0 ).

ILONG Length of regreSSion variable.

= NZ + MAX(2MQ, 8) if SEATS will be used after TRAMO.
= NZ + NPRED if only TRAMO is used.

IUSER =1 The user will enter a series X(I), 1=1 .. ILONG for this regression variable.
=-1 The program will read NSER series from a file. There must be NSER columns
of lenght ILONG in this file separated by blanks, containing the NSER series.

- 56 -
 =0 (Default) No regression variable when ISEQ=O. When ISEQ>O the program
will generate the regression variable.
=2 The user specifies the presence of some outliers; he J she will provide a
sequence of NSER pairs of number-string: (11, j1 )..( tNSER, jNSER), where t
denotes the position of the outlier and j denotes the type of outlier according
to the following code:

10 Innovation Outlier
AO Additive Outlier
LS Level Shift
TC Temporary Change

= ·2 The program will read the holidays series X(I), 1=1 .ILONG from a file. The
holidays are incorporated to the Trading Day variable.

REGEFF = 0 (Default) The regression effect is a separate additional component; it is not

included in the seasonally adjusted series.
=1 Regression effect assigned to trend.
=2 Regression effect assigned to seasonal component.
=3 Regression effect assigned to irregular component.
=4 Regression effect assigned to the seasonally adjusted series, but as an
additional separate component.
=5 Regression effect assigned to transitory component.
=6 Regression effect assigned to seasonal component as part of the calendar
effect.

ISEQ =k (k a positive integer) only when IUSER=O. The program will generate one
intervention variable of length ILONG consisting of k-sequences of ones
separated by zeroes. The user will provide k-pairs of numbers; the j-th pair
indicates the time index where the j-th sequence of ones is to begin and its
length, respectively.
=0 (Default) The program will generate no regression variable.

DELTA =d (0 $ d $ 1); the filter 1 J(1-d8) will be applied to the k sequences of ones
generated by the program.
=0 (Default).

DELTAS = ds (0 $ ds $ 1 ); the filter 1/(1-ds8s), s=MQ, will be applied to the k sequences of

ones generated by the program.
=0 (Default).

IDlOS =1 The program will apply the filter 1/(1-8)(1-81), s=MQ, to the k sequences of
ones generated by the program.
=0 (Default).

3.3 SEATS PARAMETERS

XL = .99 (Default) When the modulus of an estimated root falls in the range (XL, 1), it is
set equal to 1 if root is in AR polynomial. If root is in MA polynomial, it is set
equal to Xl.

=k A real number, .5 < k < 1 .

- 57 -
EPSPHI =3 (Default).

=k A real number. When the AR polynomial +(B) contains a complex root, this
root is allocated to the seasonal if its frequency differs from one of the
seasonal frequencies by less than EPSPHI (measured in degrees).
Otherwise, it goes to the transitory component.

RMOD = .5 (Default)
=k (0 < real number < 1 ) Cutting point for the modulus of an AR real root. If
modulus <k it goes to the transitory component ; if >k, to the trend-cycle.

NOADMISS = 0 (Default) When model does not accept an admissible decomposition,no

approximation is made.
=1 When model does not accept an admissible decomposition, it is automatically
replaced with a decomposable one.
IQM =k Number of autocorrelations used in computing ljung-Box Q-statistics.The
default value depends on MO. For MQ=12 it is equal to 24; for MQ=2, 3, 4, 6
it is equal to 4MQ; for MQ=1 it is equal to 8.

3.4 TERROR PARAMETERS

SENS =0 Low sensitivity

=1 Medium sensitivity ( Default )
=2 High sensitivity

The parameter SENS sets two parameters, k, and kz. Let t = out-of-sample forecast
error/standard deviation of in-sample residuals.Then, for a particular series,
If It1 > k2. the new observation in the series is dassified as "likely" to contain an error.
If k ) < Iti S k2 • the new observation is dassified as containing a "possible"
error.
If I t I .$ k ) . the new observation is accepted as without error.

The values of k, and kz for the different levels of sensitivity are as follows:

SENS = 0 k\ = 5
SENS = 1 k, = 4
SENS = 2 k\ = 3

These values can be changed by setting

SENS . 3,

one can then enter the new values of k, and/or k2.

NMATRIX =1 ( Default ) The matrices that summarize the results, described in 2.5.4 c), are
computed.
=0 The matrices in Out-Matrix are not computed.

- 58-
4 INDEX OF INPUT PARAMETERS

AIO 52 53 ISEQ 32 53 54
BD 39 50 ITER 3 23 24 39 42 45 46
BP 39 50 ITRAD 52
BPHI 51 IUSER 31 32 53 54
BQ 39 50 51 JPR 51
BTH 51 JPS 51
0 39 50 JQR 51
OELTA 32 54 LAM 39 40 51
DELTAS 32 54 M Q 4 1 53 54 55
EPSPHI 55 NBACK 53
FCT 51 NMATRIX 55
K1 49 55 NOAOMISS 55
K2 49 55 NPREO 53
IATIP 52 NSER 31 32 53 54
10105 32 54 NZ 24 31 38 39 52 53
10lF 53 P 39 50
lOUR 52 PHI 51
lEAST 52 Q 39 50
ILONG 31 32 53 54 REGEFF 31 32 33 34 35 54
IMEAN 51 RMOO 55
IMVX 52 RSA 27 29 46 50
INIC 53 SENS 55
INIT 51 TH 51
INT1 52 53 TSIG 53
INT2 52 53 TYPE 51
INTERP 53 UNITS 52
IQM 55 VA 52
IREG 30 40 53 XL 54

- 59 -
 WORKING PAPERS (1)

9901 Jose Ramon Martinez Resano: Instrumentos derivados de los tipos Overnight: calJ money
swaps y futuros sobre Condos federales.

9902 J. Andres, J. D. LOpez-Salido and J. Valles: The liquidity effect in a small open economy
model.

9903 Olympia Dover and Ramon Gomez: Another look at unemployment duration: long-term
unemployment and exit to a permanent job. (The Spanish original of this publication has
the same number.)

9904 Ignacio Hernando y Josep A. Trib6: Relaci6n entre contratos laborales y financieros: Un
estudio te6rico para el caso espanol.

9905 Cristina Mazon and Soledad Nuiiez: On the optimality of treasury bond auctions: the Spa­
nish case.

9906 Nadine Watson: Bank Lending Channel Evidence at the Firm Level.

9907 Jose VUials: El marco general de 1a poiftica monetaria unica: racionalidad, consecuencias y
cuestiones pendientes.

9908 Olympia Bover and Manuel Arellano: Learning about migration decisions from the mi­
grants: an exercise in endogenous samplig and complementary datasets.

9909 Olympia Bover and Pilar Velilla: Migrations in Spain: Historical background and current
trends.

9910 Fernando Restoy: Los mercados financieros espafioles ante la Uni6n Monetaria.

9911 Luis J. Alvarez and M: de los Llanos Matea: Underlying inflation measures in Spain.

9912 Regina Kaiser and Agustin MaranH: Estimation of the business cycle: a modified Hodrick­
Prescott filter.

9913 Enrique Albcrola and Jose Manuel Marques: On the relevance and nature of regional infla­
tion differentials: The case of Spain.

9914 Agustin MaranU: An application of TRAMO and SEATS. Report for the ((Seasonal Ad­
justment Research Appraisal» project.

9915 Regina Kaiser and Agustin MaranU: Seasonal outliers in time series.

9916 Enrique Alberola and Humberto L6pez: Internal and external exchange rate equilibrium in
a cointegration framework. An application to the Spanish Peseta.

9917 Jose Viiials and Javier Valles: On the real effects of monetary policy.

9918 Regina Kaiser and Agustin MaravaU: Short-term and long-term trends, seasonal adjustment,
and the business cycle.

99/9 J. Andres, J. D. LOpez-Salido and J. Valles: lntertemporal substitution and the liquidity
effect in a sticky price model.

9920 J. Andres, I. Hernando and J. D. Lopez-Salido: The role of the financial system in the
growth-inflation link: The OECD experience.

9921 Angel Estrada and Ignacio Hernando: Microeconomic price adjustments and inflation: Evi­
dence from Spanish sectoral data.

9922 Concha Artola and Una-Louise Bell: Identifying Labour Market Dynamics using Labour
Force Survey Data.
9923 Juan Ayuso and Roberto Blanco: Has financial market integration increased during the
nineties?

9924 Ignacio Fuentes and Teresa Sastre: Merkers and acquisitions in the Spanish Banking in­
dustry: some empirical evidence.

0001 Georges Siotis: Market power, total factor productivity growth, and structural change. An
illustration for Spain, 1983-1996.

0002 Alberto Cabrero: Seasonal adjustment in economic time series: The experience of the Banco
de Espana (with the model-based method).

0003 Luis Gordo and Pablo Hernandez de Cos: The financing arrangements for the regional (au­
tonomous) governments for the period 1997-2001. (The Spanish original of this publication
has the same number.)

0004 J. Andres, F. BaUabriga and J. Valles: Monetary Policy and Exchange Rate Behavior in the
Fiscal Theory of the Price LeveL

0005 Michael Binder, Cheng Hsiao and M. Hashem Pesaran: Estimation and Inference in Short
Panel Vector Autoregressions with Unit Roots and Cointegration.

0006 Enrique Alberola and Luis Molina: Fiscal discipline & Exchange Rate Regimes. A case for
currency Boards?

0007 Soledad Nunez y Miguel Perez: La rama de setvicios en Espana: un analisis comparado.

0008 Olympia Bover and Nadine Watson: Are There Economies of Scale in the Demand for
Money by Firms? Some Panel Data Estimates.

0009 Angel Estrada, Ignacio Hernando and J. David Lopez-Salido: Measuring the NAIRU in
the Spanish Economy.

0010 Eva Ortega and Enrique Alberola: Transmission of shocks and monetary policy in the curo
area. An exercise with NIGEM. (The Spanish original of this publication has the same num­
ber.)

0011 Regina Kaiser and Agustin Maranll: An Application of TRAMO-SEATS: Changes in

Seasonality and Current Trend-Cycle Assessment. The German Retail Trade Turnover Series.

0012 Regina Kaiser and Agustin Maravall: Notes on Time Series Analysis, ARIMA Models
and Signal Extraction.

0013 Jordi Gali, J. David LOpez-Salido and Javier Valles: Technology Shocks and Monetary
Policy: Assessing the Fed's Performance.

0014 Agustin Maravall and Fernando J. Sanchez: An Application of TRAMO-SEATS: Model

Selection and Out-of-sample Performance. The Swiss CPI series.

0015 Olympia Bover, Samuel Bentolila and Manuel Arellano: The Distribution of Earnings in
Spain during the 1980s: The Effects of Skill, Unemployment, and Union Power.

0016 Juan Ayuso and Rafael Repullo: A Model of the Open Market Operations of the Euro­
pean Central Bank.

0017 Francisco Alonso, Roberto Blanco, Ana del Rio and Alicia Sanchis: Estimating liquidity
premia in the Spanish government securities market.

0018 Santiago Fernandez de Lis, Jorge Martinez Pages and Jesus Saurina: Credit growth,
problem loans and credit risk provisioning in Spain.

0019 Pablo Hernandez de Cos, Isabel Argimon and Jose Manuel Gonzilez-Paramo: Does pu­
blic ownership affect business performance? Empirical evidence with panel data from the
Spanish manufacturing sector. (The Spanish original of this publication has the same
number.)

0020 Jordi Gall, Mark Gertler and J. David LOpez-Salido: European inflation dynamics.
0101 Silvio Rendon: Job Creation under Liquidity Constraints: The Spanish Case.

0102 Ravi Balakrishnan: The interaction of firing costs and on-the-job search: an application of
a search theoretic model to the Spanish labour market.

0103 Francisco de Castro, Jose Manuel Gonz3J.ez-paramo and Pablo Hernandez de Cos: Eva­
luating the dynamics of fiscal policy in Spain: patterns of interdependence and consistency
of public expenditure and revenues.

0104 Carsten Krabbe Nielsen: Three Exchange Rate Regimes and a Monetary Union: Determi­
nacy, Currency Crises, and Welfare.
0105 Juan Ayuso and Rafael Repullo: Why Did the Banks Overbid? An Empirical Model of the
Fixed Rate Tenders of the European Central Bank.
0106 Francisco J. Ruge-Murcia: Inflation targeting under asymmetric preferences.
0107 Jose Viiials: Monetary policy issues in a low inflation environment.
0108 Agustin Maravall and Ana del Rio: Time aggregation and the Hodrick·Prescott filter.
0109 Jordi GaIi and J. David LOpez-Salida: A New Phillips Curve for Spain.
0110 Angel Estrada and David Lopez-Salida: Accounting for Spanish productivity growth using
sectoral data: New evidence.
0111 Francisco Alonso, Roberto Blanco and Ana del Rio: Estimating inflation expectations
using french government inflation-indexed bonds.
0112 Gianluca Caporello. Agustin MaravaD and Fernando J. Sanchez: Program TSW. Reference
manual.

(1) Previously published Working Papers are listed in the Banco de Espana publications catalogue.

Queries should be addressed to: Banco de Espana

Seccion de Publicaciones. Negociado de Distribucion y Gestion
Telephone: 91 338 5180
Alcala, 50. 28014 Madrid