\\244\\

The generalized dynamic factor model:

identification and estimation

by

Mario Forni*
Mare Hallin * *
Lucrezia Reichlin**
Marco Lippi***

September 1998

* Università degli Studi di Modena

Dipartimento di Economia Politica
Viale Berengario, 51
41100 Modena (Italia)
e-mail: [email protected]

** Université Libre de Bruxelles,

ECARE and ISRO

*** Università degli Studi di Roma

andECARE
 l. Introduction 1

Dynamic factor models have been used extensively in finance for the analysis of asset
prices, and in macroeconomics to study the business cycle. The assumption under-
lying these models is that the dynamics of multivariate time series can be modeled
as the effect of a small number of common factors and of idiosyncratic components.
This approach is particularly useful when the dimension of the system to be analysed
is large, since it provides a parsimonious dynamic representation, whereas traditional
VARMA models would require the estimation of too many parameters. In fact, many
problems in economics and finance require the dynamic analysis of large number of
assets and many sectoral, regional or individuai variables. In economics in particu-
lar, the researcher has to face data sets which typically contain many cross-sectional
information (n large) and a relatively short time period (T small). In such context,
•
the use of factor models seems to be particularly appropriate, as shown by a small
but growing recent literature in macroeconomics (Quah and Sargent, 1993, Forni
and Reichlin, 1996, 1997, 1998, Forni an Lippi, 1997, Stock and Watson, 1998).
In this paper, as in Forni and Lippi (1998), we propose a very generai factor
mode!, which is nove! to the literature. Since we allow the idiosyncratic components
to be cross-correlateci, individuai shocks may have dynamic effects to other sectors.
This is a more realistic assumption than the traditional orthogonality assumption
underlying the traditional exact factor representation; it is particularly interesting
for the analysis of that class of business cycle models in which local shocks propagate
throughout the economy because of spillovers and complementarities (e.g., Cooper
and Haltiwanger, 1990 and Davis and Haltiwanger, 1992), or because ofinput-output
relations (e.g., Long and Plosser, 1982). The non-orthogonal case is commonly
ABSTRACT considered in the financialliterature, where, however, it is restricted t o stati c models.
Our mode! encompasses, as a special case, the static approximate factor mode! of
Chamberlain (1983) and Chamberlain and Rothschild (1983), and generalizes the
dynamic factor mode! of Sargent and Sims (1977) and Geweke (1977) to the case of
This. paper analyzes identificat.io~ conditions, and proposes an estimator, for a dy- non-orthogonal idiosyncratic components.
namJc factor model.where the 1d10syncratic components are allowed to be mutually Our representation differs from Stock and Watson (1998), where factors (apart
non-ort~ogonal. ThJS mode!, which we cali generalized dynamic fador model is nove! from time-varying coefficients) are static, but growing in number with the cross-
to. the hterature,. an d generalizes the stati c approximate ·factor mode! of Chamber- sectional dimension. An important feature of our mode! is that the common com-
lam and Rothschild (1983), as well as the exact factor mode! à la Sargent and Sims ponent is allowed to have an infinite Moving Average (MA) representation, so as
(197~). We prove rr:ean-s~uare convergence of our estimator to the common factor as to accommodate for both autoregressive (AR) and MA factors. It is more generai
t~e t1me cross-~ect10nal d1~ensions go to infinity at appropriate rates. Simulations than the finite dynamic factor mode!, which can be analyzed as a static factor mode!
yJe!d encouragmg results m small samples. An empirica! example on the out t where lagged factors are treated as additional static factors. The infinite dimensionai
growth of US states illustrates the method. pu dynamic mode! is a relevant generalization, since AR factors are likely to arise in

JEL classification nos.: C13, C33, C43. 1 This research has been supported by an A.R.C. contract of the Communauté française de
Belgique. We would -like to thank Christine De Mol for generously giving us her t ime, Dag Tj~stheim
Keywords: dynamic factor models, pane! data, dynamic principal components. for helpful discussions and Jorge Rodrigues for very valuable research assistantship.

l
macroeconomic data with business cycle features. 2. The Mode!
We provide both identification and estimation results. We show that, although
the mode! for finite n is not identified in genera!, under certain conditions, it is All stochastic variables under study are members of the standard Hilbert space
identified for n tending to infinity. We provide these conditions and we use them as L 2 (0., :F, P), where (O, :F, P) is some given probability space. W e will consider a
a basis for a heuristic method for the identification of the number of common factors. double sequence x= {xit, i E N; t E Z}, whose mode! is
For estimation, we propose a method which works well in situations where the
cross-sectional dimension is large and traditional estimation methods, based on max- (l)
imum likelihood, are no t appropriate. The basic idea of the estimator as an aggregate
is a development of Forni and Reichlin (1998), who show uniform consistency far an where the following Assumptions 1-4 are made.
estimator constructed from cross-sectional averages. The estimator proposed here
is based on principal components, i.e., on a weighted average of observations. We AssuMPTION 1.
show that the projection of the variables onto the leads and lags of the dynamic (I) The q-dimensiona! vector process {( ult, Uzt, · · ·, Uqt )', t E Z} is orthonor-
principal components converges to the common factor space for both n and T going mal white noise, i.e. var(Ujt) = l for any j and t, Ujt _l Ujt-k far any j, t and k cf O,
to infinity. • Ujt _l Ust-k for any s cf j, t and k;
In the static case, a principal component estimator has been used by Connor (II) ~ = {~it, i E N, t E Z} is a double sequence such that, firstly,
and Korajczyk (1986), who build on results in Chamberlain and Rothschild (1983)
to show convergence for n going to infinity and T fixed. Stock and Watson (1998) ... '
use the same estimator far a more genera! mode! and provide consistency results for
n and T going to infinity at some rate. is a wide-sense stationary vector process for any n, and, secondly, ~it _l Ujt-k for
This paper is closely related to Forni and Lippi (1998), who develop the repre- any i, j, t and k;
sentation theory for the same mode! we discuss in this paper. (III) the filters bij(L) are square summable and bilatera!.
Mode! (1) is a factor analytic mode!. The variables Ujt and Xit = Xit - ~it,
j = 1, ... , q, will be called the common factors and the common component of Xit,
respectively. The variable ~i t will be called the idiosyncmtic component of Xi t.
Note that Assumption l implies that the n-dimensiona! vector process
1
••• 1 Xnt ) , t E Z}

is stationary for any n. Note also that the filters bij (L) may contain negative powers
of the lag operator L, so that in genera! bi_i(L)ujt contains both lags and leads of
Ujt· Clearly if (1) is interpreted as a structural representation, where the common
factors have an economie meaning, then it would be appropriate to assume unilat-
eral impulse-response functions. Since in this paper we are not concerned with the
structural interpretation, we allow for genera! bilater filters.
The following restriction, though not strictly necessary, will considerably sim-
plify the proofs of our results.
AssUMPTION 2. Both the process ~n and X n = {( Xlt, Xzt, · · ·, Xnt )', t E Z}
have rational spectral density for any n (thus, the filters bij(L) are rational).
Obviously, Assumption 2 implies that the spectral density of Xn is rational for
any n, and therefore defined everywhere and continuous on [-7r, 7r].

2 3
 The main features of the mode! are the following. First, it is dynamic as in diverge, as n--> oo, a. e. in [-7!', 7!'], whereas the (q+ 1)-th one is uniformly (with
Geweke (1977) and Sargent and Sims (1977). Second, in contrast with the tradì- respect to (}) bounded.
tiana! dynamic factor mode!, the cross-sectional dimension is infinite. This feature is The importance of Proposition l lies in the fact that it transforms statements
the same as in the static factor mode! of Chamberlain (1983) and Chamberlain and on the dynamic eigenvalues of the unobservable Xn and ~n into st~te~ents on the
Rothschild (1983). An infinite cross-section, together with Assumptions 3 and 4 be- dynamic eigenvalues of the Xn· Thus, if the analysis of the d.ynamic e1ge~values of
low, is crucial for identification of our mode!: indeed, an d this is the third distinctive the observed process leads to the conclusion that the first q e1genvalues d1verge.a.e.
feature of (l), which differentiates it from the traditional dynamic factor mode!, we in [-7!', 7!'], whereas the (q+ 1)-th one is uniformly bounded, then the hypothes1s of
are not assuming mutua! orthogonality ofthe idiosyncratic components çit· Without a mode! with q factors is plausible. . .
orthogonality, for fixed n, reasonable assumptions allowing for identifièation of the We call mode! (1), under Assumptions l to 4, the generalzzed dynamzc factor
idiosyncratic and the common component would be very hard to find. model (GDFM).
Let :E~ be the spectral density matrix of Xn, and derrate by À~.i the function We will show that, under Assumptions l through 4, the common components
associating with any (} E [-7!', 7!'] the rea! non-negative j-th eigenvalue of :E~((}) X and the idiosyncratic components çit are asymptotically identified and can be es-
in descending arder of magnitude. The functions À~.i will be called th~ dynamic ti~ated. On the other hand, the identification and estimation of the common factors
eigenvalues of :E~. 2 In the same way, using obvious notation, À~.i and À;.i derrate the Ujt and the filters bij(L), while being obviously of great interest whe n representation
4
dynamic eigenvalues of :E~ and :E~, respectively. The latter will be called common (l) is interpreted as structural, are beyond the scope of th1s paper.
and idiosyncratic eigenvalues respectively.
AssU!vlPTION 3: The first idiosyncratic dynamic eigenvalue 1
À;
IS uniformly 3. A veraging Sequences and Asymptotic Aggregates
bounded, i.e., there exists a rea! A such that À;
1((}) ~ A for any (} E [-7!', 7!']
and any n. In the next section we prove that the common components Xit can be obtained
as limits of linear dombinations of the observations. To get an idea of the linear
ASSUMPTION 4. The first q common dynamic eigenvalues diverge almost everywhere combinations we would like to use, consider the simple example
in [-7!', 7!'], i.e., limn~oo À~1 ((}) = oo for j ~q, a.e. in [-7!', 7!'].
Assumption 3 is clearly satisfied if the x's are mutually orthogonal at any lead
Xit = Ut + t;it, i E N, t E Z, (2)

and lag and have uniformly bounded spectral densities, but is more genera! as it where ~n is assumed to be orthonormal white noise. This im?lies ~hat the spectral
allows, so to speak, for a limited amount of dynamic cross-correlation. Similarly, As- density matrix :E~ is the n x n identity matrix, so that Assumptwn 3 IS fulfille,d. In the
sumption 4 guarantees a minimum amount of cross-correlation between the common sequel, we will derrate by Xnt the vector of variables (X !t, X2t, ... , Xnt) (so that
components. With a slight oversimplification, Assumption 4 implies that each u.i is for the vector process Xn defined in the previous section, we have Xn = { Xnt, t E Z}) ·
present in infinitely many cross-sectional units, with non-decreasing importance. On Analogous meanings are to be given to Xnt and ~nt·
the contrary, Assumption 3 implies that idiosyncratic causes of variation, although Consider the arithmetic mean of Xnt:
possibly shared by many (even all) units, have their effect concentrated on a finite
l n l n
number of units an d tending to zero as i tends to infinity. 3 - """Xit = Ut +-n.Lt;it·
These assumptions have the following crucial consequence, the proof of which nu
i=l ~=l
is given in the Appendix.
When n tends to infinity, the variance of the second term on the right hand side,
PROPOSITION l. Under Assumptions 1 through 4, the first q eigenvalues of :E~ which equals 1/n, tends to zero, so that
2 l n
We use the terminology 'dynamic eigenvalues' to insist on the difference between the func-
tions À and the eigenvalues of the variance-covariance matrix employed in the static principal
lim -
n--co n
LXit = Ut
component analysis. A standard reference for eigenvalues and eigenvectors of the spectral density i= l
matrix is Brillinger (1981), Chapter 9.
3 4On the identification and estimation of the comrnon factors in a related rnodél, see Forni
As a further illustration of Assumptions 3 and 4 see the example in Remark 2, Section 3
below. and Reichlin (1998).

4 5
in the mean square. The same effect, asymptotic canceling of the idiosyncratic the proof of Lemma l in the Appendix fora more generai result). Using Assumption
component, is obtained by taking any system of weighting vectors {an, n E N}, 3, we obtain
with an = ( an1, anz, ann), instead of the uniform weights employed for
computing the standard average
llan(L)çntll 2 =i: an(e-i ):E~(e)an(e-i )de
8 8

provided that
W 11 = ( 1/n, 1/n,

n
1/n),
:; i: .\~1 (e) (t lan;(e-i
8
)1 2 ) de:; A t i: 8 2
lan;(e-i )1 de.

n-+oo
2
lim lanl = lim
n-+oo
"'a;_
L....t 1
=O. QED
j=l
Proposition 2 generalizes to variables fulfilling Assumption 3 the statement that
Going back to our generai dynamic mode!, we not only will allow for generai can be proved by elementary considerations when the components t;;t are strictly
weights, but also for linear combinations involving leads and lags of the x:.S. idiosyncratic, i. e. mutually orthogonal at any lead and lag and of uniformly bounded
spectral densities. Assumption 3 thus provides a good motivation for calling çit the
DEFINITION l. A dynamic averaging sequence, DAS henceforth, is a sequence idiosyncratic component of Xit· As already noted in the Introduction, Assumption
{an(L), n E N} where an(L) is the row vector 3 allows for very interesting economie cases that lie between common and strictly
idiosyncratic components. Forni and Lippi (1998) show that Assumption 3 is also
ann(L)), necessary far (3) to hold for any DAS.
ani (L) being a square summable bilatera! filter, with the condition Given a subset Y of L 2 (r!, F, P), Jet us denote by span(Y) the minimum closed
subspace of L 2 (r!,F,P) containing Y, and set C = span({Xit, i E N, t E Z}).
Moreover, Jet us denote by Q(x) the set of ali the asymptotic aggregates of the x's.
Proposition 2 implies that Q(x) ç C, i.e. that if an(L)Xnt converges, then the limit
is an element of C. In Proposition 3 we show that Xit E Q(x) for any i E N, so
that Q(x) = C. Moreover, the proof provides a constructive procedure, based on
DEFINITION 2. Suppose that {an(L), n E N} is a DAS and that an(L)Xnt converges the spectral density matrices :E~, leading to a DAS {Kn;(L), n E N} such that
in the mean square. We say that Yt = limn~oo an (L )Xnt is an asymptotic aggregate Kn;(L)Xnt converges to Xit, for any i E N.
of the x's. The construction will employ the dynamic principal components of the vector
Xnt· Let us recall that there exists an n-tuple of functions P~.i : [-1r, 1r] r-t c n,
j = l, ... , n, such that
4. Recovering the Common Components (i) Pj,n(e) is a row eigenvector of :E~(e) corresponding to .\~ 1 (e), i.e.,
p~ (e):E~(e) = .\~ 1 (e)p~ 1 (e)
The following result shows that, by averaging with any DAS, the itliosyncratic com- 1 for any e E [-1r, 1r];

ponent cancels asymptotically. (ii) IP~jcew =l for any j ande E [-7r, 7r];
(iii) p~1 (e)p~ 8 (e) = 0 for j oF sand any e E [-1r, 1r];
PROPOSIT!ON 2. If Assumptions 1, 2 and 3 are fulfilled, then
(iv) P~.i is measurable on [-1r, 1r];
(see Brillinger, 1981, Chapter 9, and Forni and Lippi, 1998).
(3)
An n-tuple fulfilling properties (i) through (iv) will be called a set of dynamic
eigenvectors of :E~. A consequence of (ii) and (iv) is that the Fourier expansion
in the mean square for any DAS {an(L), n E N}.
PROOF. Denote by A the transposed complex conjugate of a matrix A. Moreover,
we recall that À~ 1 (e) is the maximum ofb:E~(e)b' under the constraint lbl =l (see

6 7
converges in the mean square. Defining .. is the i-th component of P~j· Let us set
where pxn,7,t

x -x x + Pn2,iPn2
-x x + ... +p-x ·Px (4)
px.(L)=~
-n.J
~
27r .L...,;
k=-oo
[j" Pn·(e)e-ikBde]Lk
.1 '
Kni = Pn!,iPnl nq,z nq'

while K~i(L) is the corresponding filter.

-1r

px. (L) is square summable.

the fiJter -nJ For j = l, ... , n, the scalar process PROPOSITION 3. Suppose that Assumptions 1 through 4 hold. Then {K~i(L), n E
{P~/L)xnt, t E Z}, whose spectral density is N} is a DAS and
lim Xit,n = lim K~;(L )Xnt = Xi t (5)
n-too n-+oo
x :Ex -x Àx
Pnj nPn.i = nj1
in mean square far alli and t.
will be called the j-th dynamic principal component of Xn. A consequence of (iii) is PROOF. See the Appendix.
that if j =F k then the j-th and the k-th principal components are orthogonal at any REMARK 1. Note that the filters K~.i(L) resuJt from a simpJe ruJe invoJving the
lead and lag. • dynamic eigenvectors of the matrices :E~, with no intervention of ;he unobser:able
In the sequel, given a function f : [-1r, 1r] >--> C , whose components belong to
8
x's and ç's, and therefore can be estimated from the observed x s (see Sect10n 7
L 2([-1r, 1r], C), we will denote by f(L) the fiJter
beJow).
Now, Jet U = span( {u.it, j = l, ... , q, t E Z} ). The inclusion g(x) C U
immediateJy follows from Proposition 2. Proposition 4 proves the converse.
PROPOSITION 4. Under Assumptions 1 through 4,
(a particular case is that ofthe eigenvectors p~.i and the filters E.~./ L) defined above).
g(x) = U. (6)
Now, Jet

PROOF. See the Appendix.

where Xit,n is the orthogonal projection of Xit onto the subspace of L 2(1J, :F, P)
spanned by the first q dynamic principal components of Xnt, i.e. The following is an immediate impJication of Propositions 3 and 4.
COROLLARY. Suppose that Xit can be represented as in (1), and that Assumptions 1
Xit,n = Proj (xitlspan ( {E.~.i(L)Xnr, j =l, ... , q, T E Z})), through 4 are fulfilled. Suppose that Xit admits the alternatwe representatwn

and f.it,n the residua!. To obtain an explicit expression for Xit,n, Jet us define In as Xit = b; 1 (L)u~t + b; 2 (L)u;t + · · · + b;q,(L)u~'t + ç;t, (7)
the n x n identity matrix and observe that since
and that Assumptions 1 through 4 are also fulfilled far (7). Then, x:t = Xit-t;;t = Xit,
I -x x -x x -x x
n = Pn!Pnl + Pn2Pn2 + · · · + PnnPnn so that ç;t = f.it· Moreover, U = span({ujt, j = 1, ... , q'; t E Z}), and therefore
q'= q.
(the vectors p~.i are an orthonormal system for In), PROOF. Uniqueness of the common component follows from (5) and the fact that
, depends onJy on the x's. Equality (6) impJies U = span({ujt, j =l,···, q', t E
X,t,n QED
Xnt = p~ 1 (L )p~ 1 (L )Xnt + p~ 2 (L )p~ 2 (L )xnt + · · · + p~n (L )p~n (L )Xnt· Z} ).
REMARK 2. An important consequence of the Corollary is that representation (1)
Since different principaJ components are orthogonal at any lead and lag then is non-redundant, i.e. no other representation fuJfilling Assumptions l through 4
is possibJe with a smaller number of factors. In the following e~ampJe we h~ve a
common-idiosyncratic representation with one factor. However, smce Assumpt10n 4

8 9
 is not fulfilled, a representation with zero factors fulfilling Assumptions l through 4 can be avoided for eigenvalues and eigenvectors (as they are constant): for exam-
is possible. Consider the mode!
ple, Proposition 1 in the static case simply states that the first ~ eigenvalues of the
variance-covariance matrix diverge, and that the (q+ 1)-th one JS bounded.
However apart from the extreme white-noise case, there are specifications of
(l) for which' a "static" analysis may be tempting. Consider for instance
where çn is orthonormal white noise. Now suppose that the sequence of coefficients
bi, i E N, is square summable, i.e., that 2::~ 1 bf < oo. If {an(L), n E N} is a DAS, Xit = Ut + D<iUt-1 + /;it, (9)
then
f
with ai = 1 for i even, ai = O for i odd, and çn orthonormal white noise. Defining
V t -- u t-1, mode! (9) might be thought of as isomorphic to . mode! (8) with q =k 2,
with the consequence that only the variance-covariance matnx of Xn would be ta en
The second term on the right hand side tends to zero by Proposition 2. The first into consideration. The first two eigenvalues of this matrix diverge, while the third
term has variance
• one is bounded (this results from direct analysis of the va:iance-covarian~e ma~rices),
which is consistent with two static factors. However, th1s strategy JS m1sleadmg, as
shown by the fact that variance-covariance matrices do not reveal any distinction
between (9) and
(lO)
and therefore tends to zero by the assumption on the coefficients bi and the definition
of a DAS. Thus, neither Ut, nor the common component Xit = biut, can be recovered where w and W2t are orthogonal at any lead and lag. By contrast, dynamic analysis,
lt . 'Id
in this case. On the other hand, Assumption 4 does not hold since the first eigenvalue i.e. analysis of the eigenvalues of the spectral density matnces, y1e s:
of ~~ is 2::7= 1 bf, which is not divergent. Rather, the variables Xit fulfill Assumption (A) The process Yn, generateci by (10), has co~stant spectral densit!, so that
3 and therefore, in spite of the non-zero correlation among different Xit 's, they are dynamic and static analysis coincide. By Propo~1t10n l, t~e first two e1genvalues
idiosyncratic. In conclusion, a non-redundant common-idiosyncratic representation diverge, whereas the third one is bounded, cons1stently w1th two facto_rs. .
of Xi t, i.e. a representation fulfilling Assumptions l through 4, must have the trivial (B) By Lemma 1, Appendix, the first eigenvalue of the spectral dens1ty matnx
common component Xit = O for al! i. of Xn is not smaller than

n 2

5. Dynamic versus Static Analysis 2:(1 + aie-i 0

) = [n(l + ane-i0)[ 2 ,
i= l

The possibility of recovering the common components by aggregation has been stud-
ied in Chamberlain and Rothschild (1983) and Chamberlain (1983) for the mode! where an = 2::7= 1 a;/n, and therefore diverges for any ~ E [-~, 7r], w~ile the
~ second dynamic eigenvalue is uniformly bounded; th1s !S. cons1stent w1th one
dynamic factor. Thus, the difference between (9) and (10) IS fully revealed.
(8)
Moreover, as soon as a model as simple as
which has no time dimension. Mode! (8) is "isomorphic" to mode! (1) under the
assumptions that bi.i (L) is constant and that çn is a white noise process. If this is Xit =
the case, the spectral density of Xn, its eigenvalues an d eigenvectors do not depend
on (), and coincide with the variance-covariance matrix of Xn, its eigenvalues and is considered, with ai drawn from the uniform distribution ~ver [O, 1], dynamic anal-
eigenvectors, respectively (which are indeed the tools employed in Chamberlain an d ysis reveals that the first eigenvalue diverges everywhere m [-7r, 7r], whereas the
Rothschild's analysis). In this "static" case, our Assumptions 3 and 4 and Proposi- nd one is uniformly bounded. By contrast, static analysis leads to the conclu-
tions l through 4 have a simpler form, in which reference to the frequency domain seco d Th. . . t
sion that al! eigenvalues of the variance-covariance matrix iverge. IS IS cons1sten

lO 11
 with an infinite number of static common factors, but also with completely different Static mode!, one factor:
models, as for example
M la
Xit = Pit,
Static mode!, two factors:
where the variables Pit, i= l, ... , n are mutually orthogonal (at any leads and lags)
white noises such that var(pit) = i. Mlb
Static mode] with delay, one factor:
(
6. The Choice of q
Xit = a;ult + ,fiçit fori even
M2a
So far, we have assumed that q, the number of non-redundant common factors, is Xit = a;Ult-1 + ,fit:,it fori odd.
known. In practice of course, q is not predetermined, and also has to be selected
Static with delay, two factors:
from the data. Proposition l links the number of factors in (l) t o the 'lligenvalues
of the spectral density matrix of Xn: precisely, if the number of factors is q and ç
is idiosyncratic, then the first q dynamic eigenvalues of I:~ diverge a.e. in [-1r, 1r]
Xit = a;U!t + b;U2t + ,fj,çit fori even
M2b
whereas the (q+ 1)-th one is uniformly bounded. Forni and Lippi (1998) prove that Xit = a;Ult-1 + b;U2t-l + ,fit:,it fori odd.
the converse is also true: if the first q eigenvalues of I:~ diverge a. e. in [-Jr, 1r] an d
MA(l) mode!, one factor:
the (q+ l)- th is uniformly bounded, then the x's admit a representation of the form
(l) with q factors (and ç idiosyncratic). M3a
No forma! testing procedure can be expected for selecting the number q of factors
in finite sample situations. Even letting T-> oo does not help much. The definition MA(l) mode], two factors:
of the idiosyncratic component indeed is of an asymptotic nature, where asymptotics
are taken as n-> oo, and there is no way a slowly diverging sequence (divergence, X;t = ao;Ult + a 1;ult-l + bo;U2t + bliU2t-l + 2f;it· M3b
under the mode!, can be arbitrarily slow) can be told from an eventually bounded
sequence (for which the bound can be arbitrarily large). Practitioners thus have to AR(l) common component, one factor:
rely on a heuristic inspection of the eigenvalues against the number of series n.
More precisely, if T observations are available for a large number n of variables M4a
xit, the spectral density matrices I:;'T, r ::; n, can be estimated, and the resulting
empirica] dynamic eigenvalues >.;J computed for a grid of frequencies. The following
AR(l) common component, two factors:
two features of the eigenvalues computed from I:';T, r = l, ... , n, should be consid-
ered as reasonable evidence that the data have been generated by fl), with q factors
and that ç is idiosyncratic: xit
a
= .,---'--::-ult + 1-b;d; LU2t + mc
· ~it· M4b
1- c;L
(a) The average aver B of the first q empirica] eigenvalues diverges, whereas the
average of the (q+ 1)-th one is relatively stable. In al! these models, U!t, U2t, a;, ao;, a1;, b;, b?"· .b h · and C,·t
' are i.i.d. standard
.
(b) Taking r = n there is a substantial gap between the variance explained by norma! deviates, while c; and d; are uniformly d1st:Ibuted o:er [-0.8, 0.8], m .or~er
the q-th principal component and the variance explained by (q+ 1)-th one. A .
to ensure co-st a t wnan"ty of the x's . Note that the idwsyncrat1c
. shocks are mult!phed
preassigned minimum, such as 10%, for the explained variance, could be used as b a constant so that, an average, the cross section~l ~mts have the same co:nmon-
a practical criterion for the determination of the number of common factors to irosyncratic variance ratio in all models. This rat10 1s 1/2 m the models w1th one
retain. The lO% limi t is used in the empirica] exercise of Section 8. factor and 1 in the models with two factors.
To illustrate criteria (a) and (b), we have generated the following eight different We have generated data from the above models with n :- 100 and T .= 200.
factor models.
Then, we have estimated the spectral density matrix for a gnd of frequencws and

12 13
 .
F 1gure 6 .1. Dynamic eigenvalues averaged over frequencies, models (Ml)
computed the true spectral density matrix for the same frequencies. 5 Lastly, we have
through (M4)
computed the eigenvalues of the upper-left r x r submatrix, r = l, ... , n, both for
theoretical and estimated spectral density matrices. iOO iSO /

/ '
Figure 6 .l below reports the pio t of the aver age over frequencies of the theo- BO l
/
/
/

retical and estimated eigenvalues. On the horizontal axis we indicate the number of l
l ' iOO
cross-sectional units r, which obviously is maximum when the whole sample n = 100 60

is considered. Features (a) and (b) emerge quite clearly for al! models: the first q 40 l
/
50
averaged eigenvalues exhibit an approximately constant positive slope, while the re- f
maining ones are rather flat; moreover, the variance explained by the q-th principal 20
----iiiiilii~ii
component is substantially larger than the variance explained by the (q+ 1)-th, even 20 40 60 80 100
20 40 60 80 100
for small r.

iSO
• i20 /

7. Estimation of the Common Component iOO l

/

l '
BO
l iOO
l
7.1 Theory
60
50
Proposition 3 shows how the common component Xit can be recovered, asymptoti- 40
l
cally, from the sequences Kni(L)Xnt, where the filters Kni(L) are obtained as func-
tions of the eigenvectors Pni(il), j = l, ... , q, associated with the spectral density W -~=~aliiiiiJiii
0 20 40 60 80 100
matrices ~n (il) (for notational simplicity, from now on, we will drop the superscript o 20 40 60 BO iOO

x when indicating spectral density, eigenvalues and eigenvectors associated with the 300
250
x's). In practice, of course, these characteristics of the observed process are no t avail- /
250
able, and have to be replaced with empirica! counterparts, based on finite realizations 200
l
/
l
l
l

of the form l 200

l
/

i 50
X~= (xn1 1 Xn2, ... ' XnT) · /
iSO
iOO
Denote by ~;;_" (il) an estimator of the spectral density ~n (il), based on the data in
iOO

x;;. 50
~="'"'iii"iiili~~i
ASSUMPTION 5. Let O"ij,n(il) and O"D,n(il) denote the i,j entries of ~n(il) and ~;;_"(il) o
o 20 40 60 80 100 20 40 60 BO iOO

respectively. We suppose that CTD,n(il) converges in probability to O"j:j,n(il) uniformly

in [-71', ?r] for T--+ oo, i.e. 200
200

lim P{ max
T-HX> BE[-?T, 7r]
IO"Dn(il)-
,
O"ij,n(il)l} =O, iSO i 50
'
/
l
iOO iOO
for any i, j = l, ... , n. /
/

l
50
U nder Assumption 5, the estimat ed counterpart of Kni (il) allows for a consistent 50
reconstruction of the factor space. More precisely, we prove that the projection of _-_-i i j 'i i i~ i i~
20 40 60 80 100
5
The spectral density has been estimateci using the method described in Section 7.2, Bartlett Horizontal axis: r; vertical axis: variance/27r. So l'd
l
l'me: theoretical eigenvalues; dotted lines:
lag-window, size 14.
estimateci eigenvalues.
15
14
 xit onta the space spanned by the first q empirica! principal components converges PROPOSITION 5. Assume that Assumptions 1 thmugh 5 are satisfied. Then, far all
to the common component Xit· E> O and ry >O, there exists No(E, ry) such that
Assumption 5 is fulfilled under quite genera! conditions by lag-window or
periodogram-smoothing spectral estimators (see, e.g., Grenander and Rosenblatt, P [IK::(L)Xnt- Xitl >El S: 1J
1957, p. 262).
X;: p;:
Now, denote by 1 (e) and 1 (e), respectively, the eigenvalues and eigenvectors far all t= t*(T) satisfying {11}, all n 2: No and all T larger than some To(n, E, ry).
of the matrix 2J;:(e). Since eigenvalues and eigenvectors are continuous functions of PROOF. For any t S: T, we have
the entri es of the corresponding matrix, Assumption 5 implies t hat :>..;: (e) an d
1
p;:
1
(e)
converge to ÀnJ(e) and Pn.i(e), respectively, in probability, uniformly in e E [-1r, 1r],
for T--> oo. P [IK::(L)Xnt- Xitl >El
Moreover, considering
S: P[i(K::(L)- Kn;(L))Xntl > E/2l +P [IKn;(L)Xnt- Xitl > E/2]

= R;:i + Rnz, say.

i.e., the empirica! counterpart of Kn;(e), i S: q, :K;:;(e) converges to Kn;(e) in Proposition 3 ensures the existence of an No(E, ry) such that, for n 2: No, Rnz =S: ~­
e
probability, uniformly in E [-7r, 7r], for T --> 00. Thus, for all E > o and 1) > o, As for RTt
nl1
it follows from Chebyshev's theorem that, for T 2: Tr (n, o, ~),
t h ere exists T1 = Tr (n, E, 1)) such t hat, for all T 2: T1 ,

P[ sup Ji:K;;i(e)- Kni(e)l >E] S: 7). R;:i S: P [I(K::(L)- Kn;(L))Xntl > E/2 and sup IK;:f(e)- Kn;(e)l s: o]
BE[-n, n] BE[-n, n]

N ow, observe t hat, in principle, given the estimateci spectral density matrix
2J;:(e), :K;:;(e) can be computed for any e, so that each of the coefficients of the +P [ sup
BE[-n, n]
IK;:f(e)- Kn;(e)l >o]
corresponding bilatera! filter
<i
- E2
E [I(KT\L)- Kn;(L))Xntl 2
-m
1 sup
BE[-n, n]
I(K;:f(e)- Kni(e))l s: o]

Tt
If the filter Kni (L) and the observation Xnt were independent, t hen, m
. vrew
. of
·T the classica! properties of eigenvalues, the above expression would reduce to
can be obtained. However, in practice, the projection Kni(L)Xnt of Xit onta the space
spanned by the first q empirica! principal components cannot be computed, since,
for t S: O and t > T, Xnt is not available; Kni (L )xnt actually has to be truncated
at lag t -1 and lead T- t, respectively, yielding the finite-arder filter K::(L). Due
to this truncation, the common component Xit, for fixed t, never can be recovered,
even as n and T tend to infinity: indeed, part of its variance is lost because of the
non-observability of Xnt, t S: O and t> T. We therefore restrict our attention to the
"centra! part" of the observed seri es, i. e., to values of t of the form t = t* (T), with

t*(T) t*(T)
O< a< liminf-- < limsup-- < b <l. (ll)
- T~= T - T~oo T -
The following result then provides the empirica! counterpart of Proposition 3.

16 17
 Thus, for n 2 No(E, ry), and T 2 T 1(n, o, ry/4) with . o2 =e 2 9 . It must be pointed out that an estimator of the common components could be
Tt Tt ' s J::~ >-ni(O)do, we would obtam A Tt
Rnl + f!n2 ~ TJ f?r a!l t= t*(T) satisfying (11). The proposition follows
obtained by computing Kni (L), as suggested by (4), then applying this filter to Xnt·
This reasonmg IS essentially correct, and it carries the basic idea o.f th f This estimator however gives poor results in small samples. The alternative proce-
b ] · h d Tt e proo dure we suggest here, i.e., the 018 estimates ofthe observations regressed on present,
K . (L) and Xnt vams
A

t e ow, smce t e ependence between -nz · h es as T - t oo A fo 1 past and future of the estimated dynamic principal components, seems preferable in
reatment, however, requires a slightly more elaborate argument: see the Appe:::. practice. Note that, for M= O, p;;:J(L) is simply the j-th eigenvector of the (esti-
mated) variance-covariance matrix of Xnt: the dynamic principal components then
7.2 The proposed estimator reduce to the static principal components.
In order to render our procedure operational, we need a rule for fixing M as well
as the number of leads (s) and lags (g) of the principal components to be retained
In practice, the estimator of the common component b
· 0 S . s can e constructed by in the regressions. We propose the following rule. First, take a maximum value for
nmg n L regresswns (equation by equation) of the ob t" run-
and futu:e of the estimate~ dynamic principal componen::r;~ 7~Jx::,p~eset, -~as~
M, s and g, say Mo(T), such that Mo(T) - t DO and Mo(T)/T - t O as T - t oo.
Second, estimate all of the specifications with O ~ M ~ Mo(T), O ~ s ~ Mo(T),
Propos1twn 5 appl!es prov1ded that the estimated eigenvect;~J fulfill As~ • 'r '~ O~ g ~ M 0 (T), and choose the one minimizing some dynamic specification criterion.
We proceed as follows. For a fixed inte er M ump wn ·
variance matrix rTk of x t and x for k -gO M,we dcompute the sample co- Here we propose
n. n nt-k - , ... , an the 2M+ l po' t d" T n
crete Fourier transform of the truncated bilatera] sequence rT rT m s T IS- n Llog&i + 2q(s + g + 1), (13)
where r n,-k = r~.k· This implies computing n,- M• ... , nO• ... , r nM• i= l

M
where &i is the estimated variance of the residuals of equation i. This criterion is the
cross-sectional average of AICs. Note that we cannot use the multivariate AIC, since
~;;:_(Bh)= k~Mr;;:_kwke-ikO,, Bh=2Jrhj(2M+l), h=0, ... ,2M, the determinant of the covariance matrix of the residuals is very poorly estimated
for large n (the estimate isO for n> T). Neither do we propose BIC, since we found
where l - that a richer dynamic specification such as the one implied by the AIC criterion gives
.
Wk -
- are th e we1g
k
_(M+l)
· hts corresponding to the Bartlett Jag window
of s1ze lVI. Prov1ded that lVI - t DO an d JVI /T - t 0 as T - t T .
better results for our simulated models.
Assumption 5. oo, ~n (Bh) sat1sfies
Then we compute the first q eigenvectors pT J. = 1 f ...,T e
O 2 '1' Th nJ ' ' · .. ' q, 0 z.., ( h) for h - 7.3 Simulation results
'···, m · e proposed estimator of the filter p (L) · _ 1 '!· ' -
f th · d' -nJ ' J - , · · · , q, IS constructed
rom e mverse 1screte Fourier transform of the vector In order to evaluate the performance of our estimation procedure for finite val-
ues of n and T, we have carried out Monte Carlo experiments on models (Mlb),
(M2b), (M3b) and (M4b) of Section 6. We generated data from each mode! with
i.e., from the computation of n = 10, 20, 50,100 and T = 20, 50,100,200 and applied the estimation procedure
described in Section 7.2 with Mo(T) = round['7]. Each experiment was replicated
l 2M 400 times.
PT = ' \ ' PT (B ) ikO,
-n.7,k 2M + l L_. nj h e We measured the performance of our estimator, Xit, by means of the criterion
h=O
2
L::i,t(Xit- Xit)
for k =-lVI,··., M. The estimator of the filter is given by
A ) -

Rx,x-
(
"L.Ji,t Xit2

Table 7.1 reports the average and the standard deviation (in brackets) of this statistic
(12)
across the experiments.

18 19
 Table 7.2. Average R*
For ali models, we see that the fit improves as both n and T increase. To better
appreciate the results, we added aline reporting R(x, x), where Xit is the unfeasible T-20 T= 50 T -100 T -200
estimate of the common components obtained by using the unobservable true com- Mode! (M1b)
mon factors U.it in piace of the dynamic principal components as the regressors; Xit n= 10 0.477 0.317 0.263 0.247
0.301 0.182 0.145 0.123
is computed only for n= 100. The average AIC criterion (13) is used for dynamic n=20
0.057
n= 50 0.212 0.106 0.073
specification. Note that for the autoregressive mode! (M4b), the results obtained 0.184 0.083 0.052 0.036
n= 100
with n 2: 50 are comparable with those obtained with the true factors.
Mode! (M2b)
n= 10 0.611 0.440 0.367 0.304
Table 7.1. Average and standard deviation (in brackets) n=20 0.514 0.335 0.242 0.191
of R(x, x) across 400 experiments n= 50 0.436 0.264 0.169 0.120
n= 100 0.416 0.243 0.148 0.094
T=20 T=50 T= 100 T= 200 Mode! (M3b)
Mode! (M1b) n= 10 0.595 0.442 0.356 0.313
n= 10 0.636 (0.380) 0.344 (0.173) n=20 0.506 0.340 0.259 0.191
0.273 (0.125) 0.2"49 (0.095)
n=20 0.462 (0.273) 0.188 (0.070) n= 50 0.432 0.271 0.188 0.133
0.145 (0.042) 0.124 (0.033)
n= 50 0.363 (0.216) 0.106 (0.027) 0.073 (0.014) n= 100 0.416 0.253 0.163 0.111
0.057 (0.010)
n= 100 0.320 (0.214) 0.084 (0.018) 0.052 (0.008) 0.036 (0.005)
R(x, x) with n = 100 0.197 (0.105) 0.062 (0.012) 0.031 (0.005) 0.015 (0.002) Mode! (M4b)
n= 10 0.527 0.385 0.312 0.269
Mode! (M2b) 0.402 0.262 0.199 0.165
n= 20
n= 10 o. 754 (0.466) n=50 0.333 0.182 0.128 0.096
0.502 (0.206) 0.406 (0.169) 0.329 (0.130)
n=20 0.651 (0.273) n= 100 0.305 0.158 0.106 0.072
0.383 (0.141) 0.259 (0.081) 0.198 (0.063)
n=50 0.569 (0.190) 0.291 (0.085) 0.177 (0.049) 0.122 (0.032)
n= 100 0.563 (0.167) 0.269 (0.071) 0.151 (0.041) 0.096 (0.027)
R(x, x) with n = 100 0.347 (0.113) 0.103 (0.019) 0.052 (0.008) 0.025 (0.004)
s. Empirica! Illustration
Mode! (M3b)
n= 10 O. 718 (0.263) 0.500 (0.151) 0.384 (0.110) 0.329 (0.087) In this Section, we illustrate our method by estimating the generalyzed dynamic
n= 20 0.637 (0.225) 0.389 (0.115) 0.277 (0.064) 0.208 (0.043)
n= 50 0.568 (0.173) 0.306 (0.077) 0.196 (0.046) 0.136 (0.032) factor model for a panel of annual real output growth of 49 US states from 1948
n= 100 0.556 (0.154) 0.283 (0.071) 0.170 (0.043) 0.113 (0.028) through 1993. The objective is to estimate the "national component" of the US
R(c, c) with n = 100 0.362 (0.112) 0.106 (0.019) 0.052 (0.007) 0.026 (0.003) business cycle, and to extract information on its dynamic structure.
Mode! (M4b) We proceed in two steps.
n= 10 0.684 (0.393) 0.440 (0.214) 0.334 (0.137) 0.281 (0.106)
n= 20 0.549 (0.222) 0.289 (0.095) 0.208 (0.057) 0.169 (0.044)
Step 1. vVe first identify q using the procedure outlined in Section 6. Figure 8.1 is
n= 50 0.485 (0.193) 0.196 (0.053) 0.132 (0.025) 0.097 (0.016) a three-dimensional pio t of the first ten averaged (aver frequencies) eigenvalues for
n= 100 0.456 (0.156) 0.165 (0.029) 0.107 (0.014) 0.072 (0.009) cross-sectional groups of different size (random arder).
R(x, x) with n = 100 0.405 (0.118) 0.183 (0.028) 0.112 (0.015). 0.066 (0.007)
Note that two eigenvalues are increasing with n, while the remaining ones have a
flat shape. When the whole cross-section is considered, the first principal_ component
Finally, in arder to evaluate the ability of the average AIC criterion (13) in explains on aver age 52% of the variance, the second o~e 20%: and the th1rd one, 7%.
choosing the best dynamic specification, we computed R*, i. e., the minimum of R Thus, according to both informai criteria suggested m Sectwn 6, we may conclude
aver k, gand s, for each of the experiments of Table 7.1. The average R* is reported for a two common-factor mode!.
in Table 7.2. The results are very good: comparing Table 7.2 with Table 7.1, we see Step 2. Having identified q = 2, we then proceed to the estimation of the com~on
that, for T 2: 50 and n 2: 20, R is very close to R* far ali models. component following the methodology outlined in Section 7.2. The lag selectwn
criterion suggests s = l, g = 2. Regressions on leads and lags of the first two prm-
cipal components produce an average R 2 of 59.3%, with a standard deviation of
15.7%. Figure 8.2 reports the estimates for six large states. Due to the heterogenous

20 21
 Figure 8.2 Common component of six representative US states 1951-1992
Figure 8.1. Ten largest dynamic eigenvalues (increasing n-averaged over
frequencies)

Cali!ornia Florida

60 70 80 90 60 70 80 90
o
Illinois Michigan

0.05 0.05 ..
PCs (!rom the 1st to the 1oth)
States (total of 49)
o o
propagation mechanism, the common components have different variance~ and dif-
ferent turning points. However, all states have downturns corresponding to the two -0.05 -0.05 .
major oil shocks. A detailed analysis of turning points is beyond the scope of this
p a per, but the empirica! illustration conducted h ere indicates potentially interest- 60 70 80 90 60 70 80 90
ing applications of our method for the analysis of regional business cycles and their
synchronization.
From the estimateci common components we also computed the average spectral New York Texas
densities of the national and state-specific component of output changes. These are
presented in Figure 8.3. - 0.05 .
From this last exercise, we can observe that output fiuctuations in US states
have a large common component with a clear peak at cycles of period ranging from
6 to 9 years. The idiosyncratic component is not only small, but also has no cyclical o
shape.
-0.05 ....

9. Summary and Discussion 60 70 80 90 60 70 80 90

Horizontal axis: years.
The generalized dynamic factor mode! analyzed in this paper is nove! to the !iter-
23
22
 Figure 8.3 Aver age spectra of the common and idiosyncratic components , ut ut rowth of US states. This analysis can be exploite.d to ~t~dy
com~onen~ of. o p. ~l business cycles and it is a first necessary step for ldentifymg
10-5
16'rx~----r-----,-----~------r-----,-----~, turnmg p01~ts m reg~~~cks driving the components of the business cycle (the factors
the factors, 1.e., the h tatl'on) · This analysis is beyond the scope
1 ·d t'fi d up to an ort ogona1 ro .
are on y l en l e . lts on estimation of the components, it can be easlly
14
of this paper, but, giVen our resu d b F: . d Reichlin (1998).
cond~c~~d
0
a;~~; t~e~~~:t::J~~= estfma~~~~ :~ategy propos~d ~ere cannot ~e used
Ice. 'thout imposing some further restnct1ve assumptwns on
12 for forecastmg purposes Wl d d'f 1. th bilatera! filters used for estimation.
~~:e~~:~:~ci: ~~:~~f:~:;:r: :ta r:c:n~~ ~!tedepaper by Stock and Watson (1998).
f

10

8 ··············································-··

----:-

o 0.5 1.5 2 2.5 3

Horizontal axis: frequencies. Solid line: national component; dotted lines: idiosyncratic component.

ature since it allows for an infinite moving average representation of the common
component and for non-orthogonal idiosyncratic components. We have shown that,
although for a finite cross-sectional dimension this mode! is not identified in generai,
asymptotic identification conditions can be established as the cross-sectional dimen-
sion goes to infinity. These identification conditions are given on the spectral density
matrix of the process and therefore allow us to distinguish between static and lagged
factors.
For large cross-sections, dynamic factor models cannot be estimateci on the basis
of traditional likelihood based methods. We have proposed a méthod of estimation
which is appropriate in such situations and simple to implement. This method
allows for consistent estimates of the components as the cross section and the time
dimension go to infinity at some rate. The common components are computed as
the projections of the observations onto the leads and lags of the dynamic principal
components of the observations and the idiosyncratic components are derived as the
orthogonal residuals.
Consistent estimation of the components is essential for "historical analysis"
and for the identification of the factors driving common and idiosyncratic dynamics.
In the empirica! illustration we show, for example, the estimates of the "national

24 25
 REFERENCES Sargent , T . J . a nd Sims , C. A. (1977) "Business
.
cycle modelling without pretending
, . . ( ) N M h d
to have too much a priori economie theory m S1ms, C.A. ed. . ew _et o s
in Business Research, Minneapolis: Federa! Reserve Bank of Mmneapohs.
Stock, J.H. and Watson, M.H. (1998) "Diffusion indexes", manuscript, July.
Billingsley, P. (1995), Probability and Measure (3rd ed.), New York: J. Wiley.
Brillinger, D. R. (1981), Time Series Data Analysis and Theory, Holt, Rinehart and
Winston Inc.
Chamberlain, G. (1983), "Funds, factors, and diversification in arbitrage pricing
models", Econometrica 51, 1281-1304.
Chamberlain, G. and Rothschild, M. (1983), "Arbitrage, factor structure and mean-
variance analysis in large asset markets", Econometrica 51, 1305-1324.
Connor, G. and Korajczyk, R.A. (1988) "Risk and return in an equiliboium APT.
Application of a new test methodology", Journal of Financial Economics 21,
255-289.
Cooper, R. and Haltiwanger, J. (1990) "Inventories and propagation of sectoral
shocks", American Economie Review 80, 170-90.
Davis, S.J. and Haltiwanger, J. (1992) "Gross job creation, job destruction, an em-
ployment reallocation", Quarterly Journal of Economics 107, 819-63.
Forni, M. and Lippi, M., (1997) Aggregation and the Microfoundations of Dynamic
Macroeconomics, Oxford: Oxford University Press.
Forni, M. and Lippi, M., (1998) "Representation theorems for dynamic factor models
with large cross-sections", Université Libre de Bruxelles, manuscript, January.
Forni, M. and Reichlin, L., (1996) "Dynamic common factors in large cross-sections",
Empirica[ Economics 21, 27-42.
Forni, M. and Reichlin, L. (1997a) "National policies and local economies", CEPR
working paper no. 1632.
Forni, M. and Reichlin, L. (1998) "Let's get rea!: a factor analytic approach to
disaggregateci business cycle dynamics", Review of Econom~c Studies, 65, 453-
473.
Geweke, J. (1977) "The dynamic factor analysis of economie time series", in D.J.
Aigner and A.S. Golberger (eds.) Latent Variables in Socio-Economie Models,
Amsterdam, North-Holland, Ch. 19.
Grenander, U. and Rosenblatt, M. (1957), Statistica[ Analysis of Stationary Time
Series, Almqvist & Wiksell: Stockholm.
Long, J. and Plosser, C. (1983) "Rea! business cycles", Journal of Politica/ Economy
91, 39-69.
Rozanov, Yu. A. (1967), Stationary Random Processes,San Francisco: Holden Day.

26 27
APPENDIX with S,, orthogonal to ( Vnlt-k, Vn2t-k, , Vnqt-k )' for any k E Z. Then, the
spectral density of Snt cor~;verges to zero a. e. in [-1r, 1r].. . .
LEMMA l. The jollowing inequalities hold for any BE [-1r, 1rj, j and n: 8 8
PROOF. Since An(e-i )An(e' ) is Hermitian non-negat1ve defimte, there ex1sts a
À~J(B) _:; À~J(B) + À~i(B); unitary matrix Mn(ciB) be unitary and such that
À~j(e):::::: À~j(B);
À~i(B) 2:: À~i(B). M,(e-iB)An(e-i 8 )An(ei8 )Mn(ei8 ) = diag ( Vn! (B), v,z(B), · · ·, Vnq(B)),
8 8
PROOF. Recai! that the eigenvalues Àk, k = i, ... , n of a complex non-negative the functions vn.i being the non-negative eigenvalues of An(e-i )An(é ). From (14),
definite n x n hermitian matrix q_; salve the following minimax problem: f

Àk = c,,IIl,~~- 1 max{bq_ibjlbl =l, b _l c 1 , ... , b j_ Ck-d M,(L) ( Vnlt, Vnzt, · · ·, Vnqt )' (16)
= Mn(L)A,(L) ( UJt, Uzt, Uqt )
1
+ Mn(L)Rnt·
where CJ, ... 'Ck-1 and b belong to cn (see Brillinger, 1981, p. 84, Exercise 3.10.16).
Then the lemma follows immediately from the fact t hat I:~ = I:~ + I:~.. QED Since the entries of Mn(e-i 8 ) are bounded uniformly in n, the spectral density of
PROOF OF PROPOSITION l. The statement on the first q eigenvalues I:~ follows oF Mn(L)R,, tends to zero a.e. in [-1r, 1r]. Thus, taking the spectral densities of both
from the second inequality of Lemma l. The statement on the (q+ 1)-th one follows sides in (16),
from the first inequality and the fact the (q+ 1)-th eigenvalue of I:~ vanishes at any lim (1- Vni(B)) =O
n~oo

frequency. QED 8
a.e. in [-7r, 1r] for j =l, ... , q. Uniform boundedness of the entries of Mn(e-i )
To prove Propositions 3 and 4 we need some intermediate lemmas.
implies that
LEMMA 2. Let p~.i,i be the i-th coordinate of p~J· For j _:;q, limnP~J,i(B) =O for B
a.e. in [-1r, 1r].
8
PROOF. Let P~. be the n x n matrix having the eigenvectors p~i' j = l, ... , non a.e. in [-1r, 1r]. Moreover, sincethe eigenvalues of An(e-i8 )An(ei ) coincide with
the rows. From the identity P~diag (À~ 1 , À~ 2 , À~n) P~ . I:~, we obtain the eigenvalues of An(ei8 )A,(c' 8 ) (see Brillinger, 1881, Theorem 3.7.2, p. 72),

q n
lim [rq- An(ei )An(e-i
8 8
)] =O
L IP~j,JB)I À~j(B) + L
.i=l
2 2
IP~j,i(B)I À~j(B) = O'f(B), n~oo

J=q+l
a.e. in [-1r, 1r], and therefore
where O'f is the spectral density of Xit· By Proposition l, À~j(B) diverges, e a.e. in
[-1r, 1r], for j .:; q, so that, for j .:; q ande a.e. in [-1r, 1r], IP~J.JB)I converges to (17)
zero. QED
LEMMA 3. Suppose that {( VnJt Vnzt Vnqt )'' t E Z} is q-dimensionai or- a a.e. in [-1r, 1r]. From (15), using (14),
thonormal white noise for any n E N, and that, given the orthogonal projection,
[rq- An(L- )An(L)] ( UJt,
1
Uzt,
l
Uqt) - An(L
- -1
)Rnt = Snt·
( V,Jt, Vnzt, · · ·, Vnqt )' = An(L) ( UJt, Uzt, · · ·, Uqt )' + Rnt, (14)
The spectral density of the first term on the left hand side t~~ds to ~ero a.e. in
with Rnt orthogonal to ( UJt-k, Uzt-k, Uqt-k )' for any k E Z, the q x q [-1r 1r] by (17). Uniform boundedness of the entnes of An(e ) 1mpl!es that the
spectral density matrix of R,, converges to zero a. e. in [-1r, 1r]. Consider the spe~tral density of An(L- 1 )Rnt tends to zero a.e. in [-1r, 1r]. Thus, the spectral
orthogonal projection density of S,, tends to zero a.e. in [-1r, 1r] as well. QED
l - -1 1 With no loss of generality we can assume that
Uqt) = An(L ) ( VnJt, Vnzt, Vnqt ) + Snt, (15)
29
28
 AsSUMPT!ON A. À~j(e) 2: l for any j, n ande E [-7r, 7r]. PROOF OF PROPOSITIONS 3 AND 4. We have
For, possibly by embedding Lz(Sì, :F, P) into a larger space, we can assume
that Lz(Sì,:F,P) contains a double sequence {q\it, i E N, t E Z} such that tPit j_ (20)
span( {çit, i E N, t E Z} ), tPit _l l:f for any i, var(<Pit) =l for any i, and tPit j_ tP.it-k
and
for any t, k and i cl j. Defining çit = çit + tPit, and
(21)
Yit = Xit + çit, (18) Let us show that the spectral density matrix of K~;(L)f.nt tends to zero a.e. m
[-1r, 1r]. By (4) and Lemma 2,
for i E N and t E Z, we obviously have:
q
(l) Mode! (18) fulfills Assumptions l through 4, with I:~(e) = I:~(e) +In, JK~i(e)J =L IPn.i,i(eW
2

I:K(e) = I:~(e)
y
+In, and therefore À!. .1
(e) = Àçn{; (e)+ 1 ' ÀY'nJ (e)= X"
.,nJ
.(e)+ 1. .i=1
M oreover, Pnj = P~.i for any n and J, so that K~i = K~i for any n and i.
(2) As a consequence, if we prove that {K~i(L), n E N} is a DAS and converges t o zero a. e. in [-1r, 1r]. Thus the result follows from Assumption 3 an d
that limn~ooK~i(L)ynt = Xit, then, since limnK~i(L)<Pnt = O, we have
!imn~ooK~;(L)Xnt = Xnt·

Under Assumption A the function f.L~.i(e) = [À~.i(e)J- 1 1 2 is defined for any Moreover, since JK~i(e)J 2 :S:: q for any e E [-1r, 1r], then the Lebesgue dominateci
e E [-1r, 1r], is bounded an d therefore has a mean-square convergent Fourier repre- convergence theorem applies and
sentation. Let us denote by !!.}n (L) the corresponding square-summable filter.

LEMMA 4. Under Assumptions 1 through 4 and Assumption A, setting Vnjt =

1:!:.~./L)P.~.i(L)xnt, far j = l, ... , q, the vector ( Vnlt, Vnzt, · · ·, Vnqt )1 is or-
thonormal white noise. M oreover, given the orthogonal decomposition so that {K~i(L), n E N} is a DAS.
Now, from (20) and (21)
1
Uqt) +Rnt, (19)
(22)
the spectral density of Rnt converges to zero a. e. in [-1r, 1r].
PROOF. The first statement follows from the definition of dynamic eigenvectors. Consider the spectral density of the left-hand side of (22). Since çit is orthogonal
Then note that, since the x's belong to U and the ç's are orthogonal to U, to al! x's at any leads and lags, this is the sum of the spectral densities of the
two terms, minus two times the rea! part of the cross spectrum between çit,n and
Vnjt = f:!:.~/L)p~/L)Xnt = f:!:.~j(L)p~)L)Xnt + f:!:.~;(L)p~j(L)f.nt [Xit- K~;(L )xntl· Consider firstly the cross spectrum between çit,n and Xit· Setting
bi(L) = (bi 1(L), biz(L), · · · ,biq(L) ), we have
is the j-th orthogonal projection in (19). The spectral density of the second term on
Uqt )' = bi(L).Àn(L- 1) ( Vnlt,
1
the right hand side is Vnzt, .. ·, Vnqt )
+ bi(L)Snt,
where ( U!t, Uzt, · · ·, Uqt) l = An(L-
- 1) ( Vn1t, Vnzt, · · ·, Vnqt ) ' + Snt 1S
•
the
which converges to zero a.e. in [-1r, 1r] by Assumptions 3 and 4. QED orthogonal decomposition analyzed in Lemma 3 and Vnjt has the definition given in
Incidentally, note that Assumption 4, together with f.L~;(e) :S:: l, i.e., Assumption Lemma 4. Since Xit ' n is orthogonal to the terms -nJ
px .Xnt, for j = l, ... , q, at any
A, implies, by the Lebesgue dominateci convergence theorem, that f.Lx .(L)px .(L) is lead an d lag, it also orthogonal at any lead an d lag t o the terms Vnjt. Thus the
a DAS. -nJ -n1
cross spectrum between çit,n and Xit is equa! to the cross spectrum between çit,n

30 31
and bi(L)Snt· The squared modulus of the latter is bounded by the product of the everywhere in [-1!', 7rj, the smallest eigenvalue of ~~ vanishes everywhere in [-1!', 7rj,
spectral density of /;it,n, which is dominateci by the spectral density of Xit, and the and this, by Assumption 4, cannot occur for any n. Thus, there exists a ii such that
spectral density of bi(L)Snt, i.e., by for all n > ii, the first q eigenvalues of ~~ are positive, except for a su bse t of [-7r, 7r]
ofmeasure zero. We have proved that, for n > n, the rank of the spectral density
matrix of Xnt is q a. e. in [-1!', 7r]. As a consequence, Cn contains a q-dimensiona!
orthonormal white noise (Rozanov, 1967, pp. 39-43), so that Cn = U. On the other
By Lemma 3 , all the entries of ~~(e) tend to zero a.e. in [-7r, 7rj, so that the hand, by Proposition 3, Cn ç Q(x), so that U = Q(x) and Proposition 4 is proved.
cross spectrum between /;it,n and Xit tends to zero a.e. in [-1!', 7r]. Using the same QED
f
argument, considering the cross spectrum between /;it,n and K~i(L)Xnt, we end up
Tt A

with the cross spectrum between l;it,n and K~i(L)Bn(L)Snt, where Bn(L) is the PROOF OF PROPOSITION 5. For fixed T (and n), the T random variables (Kni (L)-
n x q matrix having the vectors bi (L), j = l, ... , q, on the rows. As for the spectral Kni(L))Xnt, t= l, ... , T, are not identically distributed, due to two reasons: the
density of K~i(L)Bn(L)Snt, first observe that, since ~~(e)= Bn(e-ie)Bn(eie) and truncation of the filters, which depends on t, and the boundary effects, which imply
~~(e)= ~~(e)+ ~~(e), A T A T
that the joint distributions of (Kn;(L),xnl) and (Kni(L),xnT) are not the same as,
•
A A T
e.g., those of (Kni(L),xn,Tj 2). However, the filters Kni(L) and Kni(L) are both
K~;(e)Bn(e-i )Bn(ei )K~i(e) = K~i(e)~~(e)K~i(e)::; K~i(e)~~(e)K~i(e)
0 0

q
square-summable, and {xnt; t E Z} is stationary, so that the infiuence of truncation
is asymptotically negligible, as T--+ oo, for central values of T, i.e., for sequences of
= I>~i,i(e)I À~j(e),
2

.i=l
the form t = t(T) such that
aT::; t(T)::; bT (23)
which is bounded by the spectral density of Xit (see Lemma 2). Next, observe that
the maximum eigenvalue of E~ (e), which is a continuous function of the entri es,
(automatically satisfying (11)).
tends to zero a.e. in [-7r, 7r].
As a consequence, since we have proved that the spectral density of the right Tt
Thus, for T large enough, the difference between E !(Kni (L)- Kni(L))xnt! [
A l
hand side in (22) tends to zero a.e. in [-7r, 7rj, so must the spectral densities of
Xit- K~i(L)Xnt and çit- /;it,n· Moreover, since both spectral densities are obvi- andE [i(K~i(L))- (Kni(L))Xnt!] is arbitrarily small for any sequence t= T( t) sat-
ously bounded by integrable functions for any n, then, by the Lebesgue dominateci isfying (23); note that expressions such as Kni(L)Xnt typically involve the complete,
convergence theorem the variance of bot h terms t end to zero ' so t hat ":.~,n
c 't converges non observed, process {xnt; t E Z}. Similarly, the boundary effects affecting the
t o /;i t an d Xit,n to Xi t. This completes the proof of Proposition 3. QED joint distribution of Xnt and K~i(L) are asymptotically nil as T--+ oo.
Let us now prove Proposition 4. Define
Piecing these two facts together, for all1) > O, t h ere exists a T2 = T2 (n, 1)) sueh
that, for all sequences t 1 (T) and tz(T) satisfying (23), and all T 2: Tz,
Cn=span({Xit, i=l, ... ,n, tEZ}).

Obviously, Cn ç U. We want to show that there exists a n such that , for n > ii, ,
Cn = U. For, observe that the first q eigenvalues of_ ~~(e) = Bn(e-ie)Bn(e-iB)
are equa! to the eigenvalues of the q x q matrix Bn(e-i 0 )Bn(e-i0 ) (see again
l
E [I(K~:' (L)- Kn;(L))xnt,!l sup
BE[-11',11']
IK;;:i(ii)- Kni(li)i:::: o]
~rillinger, 1981,. Theorem 3.7.2, p. 72). Therefore, if dn(e-i 0 ) is the determinant of

[I(K~:2 (L)- Kni(L))xnt il

Bn(e-' 0 )Bn(e-' 0 ), then
-E 2 sup
BE[-11',11']
IK;;:i(e)- Kni(e)! ::0 o] l ::0 1J.

On the other hand, dn is a rational function of e-iO. Therefore, either dn vanishes

e
for any E [-7r, 7rj, or fora subset of Lebesgue measure zero. But when dn vanishes It follows that, for T 2: T2(n, ~ ~),

32 33
 quantity as an empirica! variance, under fourth-order moments assumptions, has
uniformi~ bounded moment of arder two. It follows (cf. Billingsley, 1995, p. 338)
that J"
, Tt (L)- K,;(L))Xntl
E [ I(Kni ' l sup 'T
IKni(e)- Kni(e)j ~ 8]
-1r
)\T
nl (e)de is uniformly integrable (as T-+ oo), so that

BE[-,-,nj

l
)~E [J.: "5\~1 (e)de] [J.: À~1 (e)de] J.: À~1 (e)de.
=E =
~E [ (bT- aT)- 1 L;t=aTI(Kni(L)-
bT ' T Kn;(L))xntl sup 'T
IKni(e)- Kni(e)j ~ 8]
BE[-,-,,-j Thus, there exists some T3(n) such that, forali T 2 T3,

where L;~;aT stands fora sum over t running from the smallest integer laTJ larger
t han or equa! to aT t o the largest integer rbTl smaller than or equa! to bT (a window Summing up, for n ::,:: No(~, 'f!) and T 2 To, with To = To(n,~,f!) =
max ( T 1 (n, 8, ~), T2(n, ~), T3(n, f!), T3(n)) where u<2 -
2
of width UbTl- laTJ). For simplicity, we write (bT- aT)-l for crbTl ;- laTJ)- 1 . < -'7 w e have
- 128 J~,. >..n, (B)dB'
Hence,
proved that Rrf ~ ~· The proposition follows. QED

wh ere "5\~1 (e) denotes the first dynamic eigenvalue associated whith the pseudo-
empirical cross-covariance function r;~ = (bT- aT)- 1 L;~aT(xn,t-kX~.,t-tl· In
this pseudo-empirical cross-covariance structure, each cross-covariance matrix is es-
timated on the basis of a window of lenght (bT- aT). Moreover, ali covariances
are "estimated" using the same number of "observations"; as a- consequence, we
can apply the argument used by Grenander and Rosenblatt, 1957 p. 262 to show
that, under the assumption of linearity of the observed process, we have convergence
(in the me an square) of the corresponding estimated spectral density ]~ (e) t o its
theoretical counterpart, uniformly in [-1!', 1!'], as T-+ oo.
Since eigenvalues and the components of eigenvectors are continuous functions
of the entri es of the corresponding spectral densities, f:" "5\~1 ( e)de converges in prob-
ability, as T-+ oo, to j:," Àn 1 (e) de. Moreover, the sequence J:." "5\~1 (e)de is bounded
by r,. trace[f~ (e)]de = trace[I';;~] = (bT- aT)-l L;~aT L~=l x;,t,i· This latter
35
34
 29. Maria Cristina Marcuzzo (a cura di) [1988] "Richard F. Kahn. A
l.
Maria Cristina Marcuzzo [1985] "Yoan Violet Robinson (1903- discipline ofKeynes", pp. 118
1983)", pp. 134
30. Paolo Bosi {1988] "MICROMOD. Un modello dell'economia
2. Sergio Lugaresi [1986] "Le imposte nelle teorie del sovrappiù", pp. italiana per la didattica della politica fiscale", pp. 34
26
31. Paolo Bosi [1988] "Indicatori della politica fiscale. Una rassegna e
Massimo D'Angelillo e Leonardo Paggi [1986] "PCI e lUl confronto c01Ì l'aiuto di MICROMOD", pp. 25
3.
socialdemocrazie europee. Quale riformismo?", pp. 158
32. Giovanna Procacci ( 1988] "Protesta popolare e agitazioni operaie in
4. Gian Paolo Caselli e Gabriele Pastrello [1986] ''Un suggerimento Italia 1915-1918", pp. 45
hobsoniano su terziario ed occupazione: il caso degli Stati Uniti
1960/1983", pp. 52 33. Margherita Russo [1988] "Distretto Industriale e servizi. Uno studio
dei trasporti nella produzione e nella vendita delle piastrelle", pp.
5. Paolo Basi e Paolo Silvestri [1986] "La distribuzione per aree 157
disciplinari dei fondi destinati ai Dipartimenti, Istituti e Centri""
dell'Università di Modena: una proposta di riforma", pp. 25 34. Margherita Russo [ \988] "The effect of technical change on skill
requirements: an empirica! analysis", pp. 28
6. Marco Lippi [1986] "Aggregations an d Dynamic in One--Equation
Econometrie Models", pp. 64 35. Carlo Grillenzoni {1988] "ldentification, estimations ofmultivariate
transfer functions", pp. 33
7. Paolo Silvestri [1986] "Le tasse scolastiche e universitarie nella
Legge Finanziaria 1986", pp. 41 36. Nerio Naldi [1988] "'Keynes' concept of capitai". JlP· 40

8. Mario Forni {1986] "Storie familiari e storie di proprietà. Itinerari 37. Andrea Ginzburg [1988] "locomotiva Italia?", pp. 30
sociali nell'agricoltura italiana del dopoguerra". pp. 165
38. Giovanni Mottura [1988] "La 'persistenza' secolare. Appunti su
9. Sergio Paba [1986 ] "Gruppi strategici e concentrazione agricoltura contadina ed agricoltura familiare nelle società
nell'industria eurppea degli elettrodomestici bianchi", pp. 56 industriali", pp. 40

IO. Nerio Naldi (1986] "L'efficienza marginale del capitale nel breve 39. GiovaMi Mottura [1988] "L'anticamera dell'esodo. I contadini
periodo", pp. 54 italiani della 'restaurazione contrattuale' fascista alla riforma
fondiaria", pp. 40
IL Fernando Vianello [1986] "LabourTheory of Value", pp. 31
40. Leonardo Paggi [1988] "Americanismo e riformismo. La
12. Piero Ganugi [ 1986] "Rispamùo forzato e politica monetaria negli socialdemocrazia europea nell'economia mondiale aperta", pp. 120
economisti italiani tra le due guerre", pp. 40
41. Annamaria Simonazzi (1988] "Fenomeni di isteresi nella
13. Maria Cristina Marcuzzo e Annalisa Rosselli [1986] "The Theory spiegazione degli alti tassi di interesse reale", pp. 44
ofthe Gol d Standard and Ricardo's Standard Comodity", pp. 30
42. Antonietta Bassetti [1989] "Analisi dell'andamento e della casualità
14. Giovanni Solinas (1986] "Mercati del lavoro locali e carriere di della borsa valori", pp. l 2
lavoro giovanili", pp. 66
43. Giova!Ula Procacci [1989] "State coercion and worker solidarity in
15. Giovanni Bonifati (1986] "Saggio dell'interesse e domanda ltaly (1915-1918): the mora\ and politica\ content of social unrest",
effettiva. Osservazioni sul cap. 17 della Generai Theory", pp. 42 pp.41

16. Marina Murat [1986] "Betwin old and new classica\ 44. Carlo Alberto Magni [1989] "Reputazione e credibilità di una
macroeconomics: notes on Lejonhufvud's notion of full information minaccia in un gioco bargaining", pp. 56
equilibrium", pp. 20
45. Giovanni Mottura [1989] "Agricoltura familiare e sistema
17. Sebastiano Brusco e Giovanni Solinas (1986] "Mobilità agroalimentare in Italia", pp. 84
occupazionale e disoccupazione in Emilia Romagna", pp. 48
46. Mario Forni [1989] "Trend, Cycle and 'Fortuitous cancellation': a
18. Mario Forni [1986] "Aggregazioneedesogeneità", pp. 13 Note on a Paper by Ne\son an d Plosser", pp. 4

19. Sergio Lugaresi [1987] "Redistribuzione del reddito, conswni e 47. Paolo Bosi , Roberto Golinelli • Anna Stagni [ 1989] "Le origini del
occupazione", pp. 17 debito pubblico e il costo della stabilizzazione", pp. 26

20. Fiorenzo Sperotto [1987] ''L'immagine neopopulista di mercato 48. Roberto Golinelli [1989] "Note sulla struttura e sull'impiego dei
debole nel primo dibattito sovietico sulla pianificazione", pp. 34 modelli macroeconometrici", pp. 21

21. M. Cecilia Guerra [ \987] '~Benefici tributari nel regime misto per i 49. Marco Lippi [1989] "A Shorte Note on Cointegration and
dividendi proposto dalla commissione Sarcinelli: lUla nota critica", Aggregati o n", pp. l l
pp. 9
50. Gian Paolo Caselli e Gabrielè Pastrello [1989] "The Linkage
22. Leonardo Paggi [1987] "Contemporary Europe and Modero between Tertiary and Industriai Sector in the Italian Economy: 1951-
America: Theories of Modemity in Comparative Perspective", pp. 1988. From an Extemal Dependence to an Intemational One", pp.
38 40

23. Fernando Viane\lo [1987] "A Critique of Professor Goodwin's 51. Gabriele Pastrello [1989] "Francois quesnay: dal Tableau Zig-zag
'Critique ofSraffa'", pp. 12 al Tableau Formule: una ricostruzione", pp. 48

24. Fernando Viane\lo [1987] "Effective Demand and the Rate of 52. Paolo Silvestri [1989] "Il bilancio dello stato", pp. 34
Profits. Some Thoughts on Marx, Kalecki and Sraffa", pp. 41
53. Tim Mason [1990] "Tre seminari di storia sociale contemporanea",
25. AlUla Maria Sala [1987] "Banche e territorio. Approccio ad un tema pp. 26
geografico-economico", pp. 40
54. Michele Lalla [1990] "The Aggregate E.<Jcape Rate Analysed
26. Enzo Mingione e Giovanni Mottura [1987] "Fattori di throught the Queueing Mode!", pp. 23
trasformazione e nuovi profili sociali nell'agricoltura italiana:
qualche elemento di discussione", pp. 36 55. Paolo Silvestri [1990] "Sull'autonomia fmanziaria dell'univer.>ità",
pp. Il
27. Giovanna Procacci [1988] "The State and Social Contro\ in ltaly
During theFirst World War", pp. 18 56. Paola Bertolini, Enrico Giovannetti [1990] "Uno studio di 'filiera'
nell'agroindustria. Il caso del Parmigiano Reggiano", pp. 164
28. Massimo Matteuzzi e Annamaria Simonazzi [1988] "Il debito
pubblico", pp. 62
 Claudio Grimaldi, Rony Hamaui. Nicola Rossi [1991] "Non 137. David Avra Lane, Irene Poli, Michele Lalla, Alberto Roverato
Paolo Basi, Roberto Golinelli, Anna Stagni [1990] "Effetti 85. Paolo Bosi [1995] "Regionalismo fiscale e autonomia tributaria: [1996]"Lezioni di probabilità e inferenza statistica -Esercizi svolti •
57. Marketable assets and hauseholds' Portfolio Choice: a Case ofStudy 1\1.
macroeconomici, settoriali e distributivi dell'armonizzazione l'emersione di un modello di consenso", pp. 38 "pp. 302
ofltaly", pp. 38
dell'IV A", pp. 24
1\2. Massimo Baldini [\995] "Aggregation Factors and Aggregation Barbara Pistoresi [1996] "Is an Aggregate Errar Correction Mode\
86. Giulio Righi, Massimo Baldini, Alessandra Brambilla [1991] "Le 138.
Michele Lalla [1990] "Modelling Employment Spells from Emilia misure degli effetti redistributivi delle imposte indirette: confronto Bias in Consumer Demand", pp. 33 Representative ofDisaggregate Behaviours? An example" pp. 24
58.
LabourForceData", pp. 18 tra modelli alternativi", pp. 47
l 13. Costanza T orricelli [199 5] "The informatian in the term structure of Luisa Malaguti e Costanza Torricelli (1996] "Monetary policy and
139.
Andrea Ginzburg [1990] "Politica Nazionale e commercio interest rates. Can stocastic models help in resolving the pu.zzle?" the term structure of interest rate s" , pp. 30
59. 87. Roberto Fanfani, Luca Lanini [1991] "IIUloVazione e servizi neUo
internazionale", pp. 22 sviluppo della meccanizzazione agricola in Italia", pp. 35 pp. 25
140. Mauro Dell'Amico, Martine Labbé, Francesco Maffioli {1996]
Andrea Giommì [1990] "La probabilità individuale di risposta nel 114. Margherita Russo [1995] "Industriai complex, pO\e de "Exactso\ution ofthe SONET Ring Loading Problem", pp. 20
60. 88. Antonella Caiumi e Roberto Golinelli [1992] "Stima e applicazioni
trattamento dei dati mancanti", pp. 13 développement, distretto industriale. Alcune questioni sulle unità di
di un sistema di domanda Almost Ideai per l'economia italiana", pp.
indagine nell'analisi dello sviluppo." pp. 45 Mauro Dell'Amico, R.J.M. Vaessens [1996] "Flow and open shop
34 141.
61. Gian Paolo Caselli e Gabriele Pastrello [1990] "The service sector scheduling 6n two machines with transportation times and machine·
in planned economies. Past experiences and future prospectives", 1\5. Angelika Moryson [1995] "50 Jahre Deutschland. 1945- 1995" pp...
Maria Cristina Marcuzzo [1992] "La relazione salari-occupazione independent processing ti mes in NP-hard, pp. l O
89. 21
pp. 32 tra rigidità reali e rigidità nominali", pp. 30
142. M Dell'Amico, F. Maffioli, A. Sciomechen (1996) "A Lagrangean
Giovanni So\inas [1990] "Competenze, grandi industrie e distretti !16. Paolo Basi [1995) "Un punto di vista macroeconomico sulle Heuristic for the Pirze Collecting Travelling Salesman Problem", pp.
62. Mario Biagioli · [1992] "Employ"ee fuiancial participation in
industriali,. ll caso Magneti Marelli", pp. 23 · 90. caratteristiche di lungo periodo del nuovo sistema pensionistico
enterprise results in Italy", pp. 50 14
italiano." pp. 32
63. Andrea Ginzburg [1990] "Debito pubblico, teorie monetarie e Massimo Baldini (1996] "Inequality Decomposition by lncome
91. Mario Biagioli [1992] "Wage structure, relative prices and Gian Paolo Caselli e Salvatore Curatolo (1995] "Esistono relazioni 143.
tradizione civica nell'Inghilterra del Settecento". pp. 30 \l7. Source in ltaly- 1987- \993", pp. 20
international competitiveness", pp. 50 f stimabili fra dimensione ed efficienza delle istituzioni e crescita
64. ·Mario Forni [1990] ..Incertezza, infonnazione e mercati produttiva? Un esercizio nello spirito di D.C. North." pp. 11 Graziella Bertocchi [1996] "Trade, Wages, and the Persistence of
92. Paolo Silvestri e Giovanni Solinas {1993] "Abbandoni, esiti e 144.
assicurativi: lUla rassegna", pp. 37 Underdevelopment" pp. 20
carriera scolastica. Uno studio sugli studenti. iscritti alla Facoltà di Mario Forni e Marco Lippi [1995] "Pennan"ent incarne,
Econol).lÌa e Commercio dell'Università di Modena nell'aMo 118.
65. Mario Forni [1990] "MisspecificationinDynamicModels", pp. 19 heterogeneity and the errar correction mechanism." pp. 21
accademico 199011991", pp. 30 145. Graziella Bertocchi and Fabio Canova [1996] "Did Colonization
matter far Growth? An Empirica! Exploration into the Historical
66. Gian Paolo Caselli e Gabriele Pastrello [1990] "Service Sector 119. Barbara Pistoresi. [1995] "Co·movements and convergence in
Causes of Africa.'s Underdevelopment" pp. 32 ·
Growth in CPE's: An UnsolvedDilemma", pp. 28 93. Gian Paolo Caselli e Luca Martinelli (1993] "Italian GPN growth international output. A Dynamic Principal Compon~:;nts Analysis"
1890-1992: a unit root or segmented trend representatin?", pp. 30 pp. 14
146. Paola Bertolini (1996] "La modernization de l'agricolture ita!ieiUle et
67. Paola Bertolini [1990] "La situazione agro-alimentare nei paesi ad
94. Angela Politi [1993] "La rivoluziOne fraintesa. I partigiani emiliani Mario Forni e Lucrezia Reichlin (1995} "Dynamic common factors le eas de I'Emilie Romagne" pp. 20
economia avanzata", pp. 20 120.
tra liberazione e guerra fredda, 1945-1955", pp. 55 in large cross-section" pp. 17
147. Enrico GiovaiUletti [ 1996] "Organ!sation industrielle et
68. Paola Bertotini [1990] "Sistema agro-alimentare in Emilia développement local: le cas de l'agroindutrie in Emilie Romagne"
Romagna ed occupazione", pp. 65 95. Alberto Rina!di [1993] "Lo sviluppo dell'industria metalmeccanica Giuseppe Marotta [1995]''Il credito commerciale in Italia: una nota
121. pp. 18
in provincia di Modella: 1945-1990", pp. 70 su alcuni aspetti strutturali e sulle implicazioni di politica monetaria"
69. Enrico Giovannetti [1990] "Efficienza ed innovazione: il modello pp. 20 Maria Elena Bontempi e Roberto Golinelli [ 1996) "Le determinanti
"fondi e flussi" applicato ad una f11iera agro-industriale", pp. 38 96. Paolo Emilio Mìstrulli [\993] "Debito pubblico, intennediari 148.
fmanziari e tassi d'interesse: il caso italiano", pp. 30 del leverage delle imprese: una applicazione empirica ai setlori
122. Giovanni Bonifati [ 1995] ''Progresso tecnico, concorrenza e
decisioni di investimento: una analisi delle determinanti di lungo industriali dell'economia italiana" pp. 31
70. Margherita Russo [ \990] "Cambiamento tecnico e distretto
industriale: una verifica empirica", pp. 115 97. Barbara Pistoresi (1993] "ModeHing disaggregate and aggregate periodo degli investimenti" pp. 25
labour demand equations. Cointegration analysis of a labour 149. Paola Bertolini [1996) "L'agriculture et la politique agricole
demand function for the Mai n Sectors of the Ital ian Eeonomy: 1950- Giovanni Bonifati [1995] "Cambiamento tecnico e crescita ita\ieiUle face aux recents seenarios", pp. 20
71. Margherita Russo [1990] "Distretti industriali in teoria e in pratica: 123.
una raccolta di saggi", pp. 119 1990", pp. 45 endogena: una valutazione critica delle ipotesi del modello di
\50. Enrico Giovannetti [1996] "Il grado di utilizzo della capacità
Romer" pp. 21 produttiva come misura dei costi di transazione: una rilettura di
Paolo Silvestri [1990] " La Legge Finanziaria. Voce 98. Giovarmi Bonifati [1993] "Progresso tecnico e accumulazione di
72. 'Nature ofthe Finn' di R. Coase", pp. 75
dell'en~iclopedia Europea Garzanti'', pp. 8 conoscenza nella teoria neoclassica della crescita endogena. Una 124. Barbara Pistoresi e Marcello D'Amato (1995] "La riservatezza del
analisi critica del modello di Romer", pp. 50 banchiere centrale è un bene o un male? ,Effetti dell'informazione
incompleta sul benessere in un modello di politica monetaria." pp. 32 151. Enrico Giovannetti {1996) "Il I0 ciclo del Diploma Universitario
73. Rita Paltrinieri [1990] "La popolazione italiana: problemi di oggi e Economia e Amministrazione delle Imprese", pp. 25
di domani", pp. 57 99. Marcello D'Amato e Barbara Pistoresi [1994) "The relationship(s)
among Wages, Prices, Unemployment and Productivity in Italy". pp. \25. Barbara Pistoresi [1995] "Radici unitarie e persistenza: l'analisi
30 152. Paola Bertolini, Enrico Giovannetti, Giulia Santacaterina [1996) "Il
74. Enrico GiovaiUletti [1990] "Illusioni ottiche negli andamenti delle univariata delle fluttuazioni economiche." pp. 33
Settore del Verde Pubblico. Analisi della domanda e valutazione
Grandezze distributive: la scala mobile e !"appiattimento' delle
\00. Mario Forni (1994] "Consumption Vo\atility and Incarne Barbara Pistoresi e tvlarceUo D'Amato [1995) "Co-movements in economica dei benefici", pp. 35
retribuzioni in una ricerca", pp. 120 126.
Persistence in the Permanent Income Mode\", pp. 30 European rea\ outputs" pp. 20
153. Giovanni Solinas [1996] "Sistemi produttivi del Centro-Nord e del
75. Enrico Giovannetti [1990] "Crisi e mercato del lavoro in un
101. Barbara Pistoresi [1994] "Using a VECM to eharacterise the Antonio Ribba (1996] "Ciclo economico, modello lineare-stocastico, Mezzogiorno. L'industria delle calzature", pp. 55
distretto industriale: il bacino delle ceramiche. Sez I'', pp. \50 127.
relative importance ofpermanent and transitority componeni.S", pp. forma dello spettro del!e variabili macroeconomiche" pp. 31
!54. Tindara Addabbo [1996] "Married Women's Labour Supply in Ita!y
76. Enrico GiovaiUletti (\990) " Crisi e mercato del lavoro in un 28
Carlo Alberto Magni [ \996) "Repeatable an d una tantum rea! options in a Regional Perspective", pp. 85
distretto industriale: il bacino delle ceramiche. Sez. II", .pp. 145 128.
102. Gian Paolo Caselli and Gabriele Pastrello [1994] "Polish recovery a dynamic programming approach" pp. 23
form the slump to an old dilemma", pp. 20 !55. Paolo Silvestri, Giuseppe Catalano, Cristina Bevilacqua [1996) "Le
78. Antonietta Bassetti e Costanza Torricelli [1990] "Una
Carlo Alberto Magni [1996] "Opzioni reali d'investimento e tasse universitarie e gli intetventi per il diritto allo studio: la prima
riqua!ificazione dell'approccio bargaining alla selezioni di 129.
Sergio Paba [1994].,. "Imprese visibili, accesso al mercato e interazione competitiva: programmazione dinamica stocastica in fase di applicazione di una nuova normativa" pp. !59
portafoglio", pp. 4 l03.
organizzazione della produzione", pp. 20 optimal stopping" pp. 26
!56. Sebastiano Brusco, Paolo Bertassi, Margherita Russo [1996)
77. Antonietta Bassetti e Costanza Torricelli {1990] ''Il portafoglio
l04. Giovanni Bonifati [1994] "Progresso tecnico, investimenti e Carlo Alberto Magni [1996) "Vaghezza e logica fuzzy nella "L'industria dei rifiuti urbani in Italia", pp. 25
ottimo come soluzione di un gioco bargaining", pp. \5 \30.
capacità produttiva", pp. 30 valutazione di un'opzione reale" pp. 20
!57. Paolo Silvestri, Giuseppe Catalano [1996) "Le risorse del sistema
79. Mario Forni [\990) ~una nota sull'errore di aggregazione", pp. 6
105. Giuseppe Marotta [1994) "Credit view and trade credit: evidence Giuseppe Marotta [1996] "Does trade credit redistribution thwart universitario italiano: fmanziamento e governo" pp. 400
JJI.
80. Francesca Bergamini [\99\] "Alcune considerazioni sulle from ltaly", pp. 20 monetary policy? Evidence from Italy" pp. 20
!58. Carlo Alberto Magni [1996] "Un semplice modello di opzione di
soluzioni di un gioco bargaining", pp. 21
Margherita Russo [1994] "Unit of investigation for Iaea\ economie Mauro Dell'Amico e Marco Trubian [1996] "Almost-optimal differimento e di vendita in ambito discreto". pp. IO
106. 132.
81. Michele Grillo e Michele Polo [ 1991) "Politica! Exchange an d the development policics", pp. 25 solution oflarge weightcd equieut problems" pp. 30
!59. Tito Pietra, Paolo Siconolfi {1996) "Fully Revealing Equi\ibria in
allocation of surplus: a Mode! of Two-party competition", pp. 34
107. Luigi Brighi [ 1995] "Monotonicity an d the demand theory of the Carlo Alberto Magni [1996]"Un esempio di investimento industriale Sequentia1 Economies with Asset Markets" pp. 17
133.
82. Gian Paolo Caselli e Gabriele Pastrello [l99lj "The 1990 Polish weak axioms", pp. 20 con interazione competitiva e avversione al rischio" pp. 20
160. Tito Pietra, Paolo Siconolfi [1996] "Extrinsic Uncertainty and the
Recession: a Case ofTruncated Multiplier Process", pp. 26
108. Mario Forni e Lucrezia Reichlin (1995] "Modelling the impact of Margherita Rus~o. Peter BOrkey, Emilio Cube!, François Lév€que, Informational Role ofPrices" pp. 42
134.
Gian Paolo Caselli e Gabriele Pastrello [1991] "Polish firms: techno!ogical change across sectors an d aver ti me in manufactorìng", Francisco Mas [\996]"Local sustainability and competitiveness: the
83. Paolo Bertella Fametti [1996]"11 negro e il rosso. Un precedente non
Pricate Vices Pubblìs Virtues", pp. 20 pp. 25 case ofthe ceramic tile industry" pp. 66 161.
esplorato dell'integrazione afroamericana negli Stati Uniti" pp. 26

84. Sebastiano Brusco e Sergio P ab a [1991] "Connessioni, competenze 109. Marcello D'Amato and Barbara Pi~toresi (1995) "Modelling wage 135. Margherita Russo (1996] "Camionetta tecnico e relazioni tra
growth dynamics in Italy: 1960-1990", pp. 38 162. David Lane [1996J"Is what is good far each best far ali? Learning
e capacità concorrenziale nell'industria della Sardegna", pp. 25 imprese" pp. 190
from others in the irUormation contagion mode!" pp. 18
IlO. Massimo Baldini [1995] "IND]}.AOD. Un modello di 136. David Avra Lane, Irene Poli, Michele Lalla, Alberto Roverato
microsimulazione per lo studio delle imposte indirette", pp. 37 [1996] "Lezioni di probabilità e inferenza statistica" pp. 288
163. Antonio Ribba [1996] "A note on the equivalence of long-run and 189. Fabio Canova [ 1997] "Does Detrending Matter for the
short-run identifying restrictions in cointegrated systems" pp. IO 216. Tindara Addabbo [1998] "Probabilità di occupazione e aspettative 243. Gian Paolo Caselli [ 1998] The future of mass consumptìon society in
Determinati an of the Reference Cycle and the Se!ection of Tuming individuali" pp 36
Points?" pp. 35 the former plarmed economies: a macro approach pp 21
164. Antonio Ribba (1996] "Scomposizioni permanenti-transitorie in
sistemi cointegrati con una applicazione a dati italiani" pp. 23 :!17. Lara Mago ani [ 1998] "Transazioni, contratti e organizzazioni: una
190. Fabio Canova e Gianni De Nicolò [1997]"The Equity Premium and chiave di lettura della teoria economica dell'organizzazione pp 39
the Risk Free Rate: A Cross Country, Cross Maturity Examination"
165. Mario Forni, Sergio Paba [ 1996] "Economie Growth, Social pp. 41
Cohesion and Crime" pp. 20 218. Michele Lall:L, Rosella Molinari e Maria Grazia Modena (1998] "La
progressione delle carriere: i percorsi in cardiologia" pp 46
191. Fabio Canova e Angel J. Ubide [1997] "Intemational Business
166. Mario Forni, Lucrezia Reichlin [1996} "Let's get real: a factor Cycles, Financial Market and Household Production" pp. 32 219. Lara Magnani [1998] ''L'organizzazione delle transazioni di
analytical approch to disaggregated business cycle dynamics" pp. 25
subfomitura nel distretto industriale" pp 40
192. Fabio Canova e Giaruù De Nicolò [1997} "Stock Returns, Term
167. Marcello D'Amato e Barbara Pistoresi [1996] "So many Italies: Structure, Inflation and Real Activity: An Intemational Perspective"
Statistica[ Evidence on Regional Cohesion" pp. 31 220. Antonio Ribba [!998]"Recursive VAR orderìngs and identification
pp. 33
of permanent an d transitory shocks" pp 12
168. Elena Bonfiglioli, Paolo Bosi, Stefano Toso [1996] ~L'equità del 193. Fabio Canova e Morten Ravn [1997} "The Macroeconomic Effects
contributo straordinario per l'Europa" pp. 20 221. Antonio Ribba (1998] "Granger-causality and exogeneity in
of German Uni:fication: Rea! Adjustmentsand the Welfare State" pp.
cointegrated Var models" pp 5
34
169. Graziella Bertocchi, Michael Spagat [1996] "D ruolo dei licei e delle
scuole tecnico-professionali tra progresso tecnologico, conflitto 222. Luigi Brighi e Marcello D'Amato (1998] "Optimal Procurement in
194. Fabio Canova [1997] "Detrending and Business Cycle Facts" pp. 40
sociale e sviluppo economico" pp. 37 Multiproduct Monopoly" pp 25
195. Fabio Canova e Morten O. Ravn [1997] "Crossing the Rio Grande:
17Ò .. ·.·, Giarma Boero, Costanza Torricelli [1997] "The Bxpectations f 223. Paolo Bosi, Maria Cecilia Guerra e Paolo Silvestri [1998} "La spesa
Migratioris, Business Cycle·and the·welfare State" pp. 37
Hypothesis of the Term Structure of Interest. Rates: Evidence for sociale nel comune Modena" Rapporto intermedio pp 37
Gennany" pp. 15 196. Fabio Canova e Jane Marrinan [1997] "Sources and Propagation of 224 Mari.o Forni· e Marco Lippi [1998] "On the Microfoundations of
Intemational Output Cycles: Common Shocks or Transmission?" pp. Dynamic Macroeconomics" pp 22
171. ·Mario Forni, Lucrezia Reichlin [1997]"National Policies and Local 41
Economies: Europe.and the US" pp. 22
225. Roberto Ghise!li Ricci [ 1998] "Nuove Proposte di Ordinamento di
197. Fabio Canova e Albert Marcet [1997] "The Poor Stay Poor: Non- Numeri Fuzzy.Una Applicazione ad un Problema di Finanziamento
172. Carlo Alberto Magni [1997] "La trappola del Roe e la Convergence Across Countries and Regions" pp. 44
tridimensionalità del Van in un approccio sistemico", pp. 16 pp 7
198. Carlo Alberto Magni [1997] "Un c"fiterio Strutturalista per la 226.
173. Mauro Del!'Amico [1997} "A Linear Time Algorithm for Tommaso Minerva [1998] "Internet Domande e Risposte" pp 183
Valutazione di Investimenti" pp. 17
Scheduling Outforests with Communication Delays on Two or Three
Processor"pp. 18 221 Tommaso ~nerva [ 1998] "Elementi di Statistica Computazione.
199. Stefano Bordoni [1997] "Elaborazione Automatica dei Dati" pp. 60
Parte Prima: .Il Sistema Operativo Unix ed il Linguaggio C" pp. 57
174 Paolo Bosi [1997] "Aumentare l'eta pensionabile fa diminuire la 200. Paolo Berteli a Fafnetti [ 1997} "The United States an d the Origins of
spesa pensionistica? Ancora sulle caratteristiche di lungo periodo 228. Tommaso Minerva and !rene Poli [1998} "A Germetic Algorithms
European Integration" pp. 19
della riforma Dini" pp. 13 Sclection Method for Predictive Neural Nets and Linear Modenls"
pp. 60
201. Paolo Bosi [1997] "Sul Controllo Dinamico di un Sistema
175. Paolo Bosi e Massimo Matteuzzi [1997] "Nuovi .strumenti per Pensionistico a rupartizione di Tipo Contributivo" pp 17
l'assistenza sociale" pp 31 229. Tommaso Minerva and Irene Poli [1998] "Building an ARlv!A
Mode! by using a Genetic Algorithm" pp. 60
202. Paola Bertolini [1997] "European Union Agricultural Policy:
176. Mauro Dell'Amico, Francesco Maffioli e Marco Trubian [1997] Problems an d Perspectives" p p 18
"New bounds for optium traffic assignment in satellite 230. Mauro Dell'Amico e Paolo Toth [1998] "Algoritluns and Codes for
communication" pp. 2 l Dense Assignment Problems: the State of the Art" p p 35
203. Stefano Bordoni [1997] "Supporti Informatici per la Ricerca delle
soluzioni di Problemi Decisionali" pp30
177. Carlo Alberto Magni [1997] "Paradossi, inverosimiglianze e 231. Ennio Cavazzuti e Nicoletta Pacchiarotti [ 1998] "Ho w to play an
contraddizioni del Van: operazioni certe" pp. 9 hotel!ing game in a square town" pp 12
204. Carlo Alberto Magni (1997} "Paradossi, Inverosimig1ianze e
Contraddizioni del Van: Operazioni Alea_torie" pplO 232
178. Barbara Pistoresi e Marcello D'Amato [1997] "Persistence of Alberto Roverato c Ircne Poli (1998] "Un algoritmo Senetico per la
relative unemployment rates across italian regions" pp. 25 selezione di modelli grafici" pp Il
205. Carlo Alberto Magni [ 1997] ''Tir, Roe e Van: Distorsioni
linguistiche e Cognitive nella Valutazione degli Investimenti" pp 17 233.
179. Margherita Russo, Franco Cavedani e Riccardo Planesani (1997] "Le Marcello D'Amato e Barbara Pistoresi [1998] "Delegation of
Monetary Policy to a Centrai Banker with Private Infonnation" pp
spese ambientali dei Comuni in provincia di Modena, 1993-1995" 206. GiseUaFacchinetti, Roberto Ghisdli Ricci e Silvia Muzzioli [1997}
pp. 23 15.
"New Methods For Ranking Trìangu!ar Fuzzy Numbers: An
lnvestment Choice" pp 9 234 Graziella Bertocchi e Michacl Spagat [1998] "The Evolution of
180. Gabriele Pastrello [!997] '·Time and Equilibrium, Two Elisive
Modem Educational Systcms. Technical vs. Generai Education,
Guests in the Keynes-Hawtrey-Robertson Debate in the Thirties" pp. 207. Mauro Dell'Amico e Silvano Martello [ 1997] "Reduction of the
25
Distributional Con11ict. and Growth" pp 31
Three-Partition Problem" ppl6

181. 235. Andrè Duntas [1998] "Le systeme monetaire Europeen" pp 24.
Luisa Malaguti e Costanza TorriceUi [1997] "The Interaction 208. Carlo Alberto Magni [1997] "IRR, ROE and NPV: a Systemic
Between Monetary Policy and the Expectation Hypothesis of the Approach" pp. 20
Tenn Structurc of Intercst rates in a N-Period Rational Expectation 236. Giarma Boero, Gianluca Di Lorenzoe Costanza Torricelli [l998)
Mode!" pp. 27 "The influence of short rate predictability and monetary policy on
209. Mauro Del!'Amioo, Andiea Lodi e Francesco Maffioli [1997] tests ofthe expectations hypothesis: some comparative evidence" pp
"Solution of the cumulative assignment problem with a we\1-
182. Mauro Dell'Amico [1997] "On the Continuous Relaxation of 30
structured tabu search method" pp. 25
Packing Problems- Technical Note" pp. 8
237. Carlo Alberto Magni [1998] "A systemìc rule for invcstmcnt
210. Carlo Alberto Magni [1997] "La definìzwne di investimento e
183. Stefano Bordoni [ 19971 "Prova di Idoneità di Informatica Dispensa decisions: generalizations of the traditional DCF criteria and new
criterio del Tir ovvero: la realtà inventata" pp.l6
Esercizi Exce! 5" pp 49 conceptìons" pp 30
211. Carlo Albero Magni [1997] "Critica alla definizione classica di
184. Francesca Bergamini e Stefano Bordoni [1997] "Una verifica 238. Marcello D'Amato e Barbara Pistoresi [1998] "Intcrcst Rate Spreads
investimento: un approccio sistemico" pp\7
empirica di un nuovo metodo di selezione ottima di portafoglio" pp. Between Italy and Germany: 1995-1997'' pp 16
22 212. Alberto Roverato (1997] "Asymptotic prior to posterior analysis !Or 239 Paola Bettolini e Alberto Bertacchini [1998] "I! distretto di
graphical gaussian mode!s" pp.8
\85. Gian Paolo Caselli e Maurizio Battini [1997] "Following the tracks lavorazioni carni suine in provincia di Modena" pp 29
of atk.inson and micklewright the changing distribution of income 213. Tindara Addabbo [1997] ''Povertà nel 1995 analisi stati ca e dinamica
and eamings in poland from 1989 to 1995".pp 21 2-10. Costanza Torricelli e Gianluca Di Lorenzo [1998] "Una nota sui
sui redditi familiari" pp 64
fondamenti matematico-!inanziari della teoria delle aspettative della
186. Mauro Dell'Amico e Francesco Maffioli [1997} "Combining Linear struttura della scadenza" pp. 15
214. Gian Paolo Caselli e Franca Manghi [1997] "La transizione da piano
and Non-Linear Objectives in Spanning T ree Problems" pp. 21 a mercato e il modello di Ising" ppl5
241. Christophe Croux, Mario Forni e Lucrezia Reichlin [1998J "A
187. Gianni Ricci e Vanessa Debbia [ 1997] "Una soluzione evolutiva in Measure of Comovement for Economie Indicators: Theory and
215. Tindara Addabbo [1998] "Lavoro non pagato e reddito esteso:
un gioco differenziale di lotta di classe" pp.l4 Empirics"pp 23.
un'applicazione alle famiglie italiane in cui entrambi i coniugi sono
lavoratori dipendenti" pp 54
188. Fabio Canova e Eva Ortega [ 1997] "Testing Calibrated Generai 2-+2. Carlo Alberto Magni [1998] "Note sparse sul dilemma del
Equilibrium Mode!" pp 34 prigioniero (e non solo) pp 13