Optimal Monetary Poli y Rules
Eri
T. Swanson
Board of Governors of the Federal Reserve System
eswansonfrb.gov
Abstra
t
A standard result in the literature on monetary poli
y rules is that of
er
tainty equivalen
e: given the expe
ted values of all the state variables of the
e
onomy, poli
y should be set in a way that is independent of all higher mo
ments of those variables. Some ex
eptions to this rule have been pointed out
by Smets (1998), who restri
ts poli
y to respond to only a limited subset of
state variables, and by Orphanides (1998), who restri
ts poli
y to respond to
estimates of the state variables that are biased. In
ontrast, this paper stud
ies unrestri
ted, fully optimal poli
y rules with optimal estimation of state
variables. The rules in this framework exhibit
ertainty equivalen
e with re
spe
t to estimates of an unobserved, possibly
ompli
ated, state of the e
on
omy X , but are not
ertaintyequivalent when 1) a signalextra
tion problem
is involved in the estimation of X , and 2) the optimal rule is expressed as
a redu
ed form that
ombines poli
ymakers' estimation and poli
ysetting
stages. In general, I show that it is optimal for poli
ymakers to attenuate
their rea
tion
oeÆ
ient on a variable about whi
h un
ertainty has in
reased,
while responding more aggressively to all other variables, about whi
h un
ertainty hasn't
hanged.
JEL Classi
ation: E52
Version 0.5.1
May 30, 2000
I thank Athanasios Orphanides, Glenn Rudebus
h, Brian Sa
k, David Small, Volker
Wieland, and parti
ipants at the San Fran
is
o Fed/Stanford SIEPR Conferen
e for
helpful dis
ussions,
omments, and suggestions. The views expressed in this paper,
and all errors and omissions, should be regarded as those solely of the author, and are
not ne
essarily those of the individuals listed above, the Federal Reserve System, or
its Board of Governors.
1
1. Introdu
tion
In
reased un
ertainty about the
urrent growth rate of produ
tivity, potential output,
and the natural rate of unemployment has led to questions about how monetary poli
y
should be altered in the fa
e of this un
ertainty. The question is extremely important
from a pra
ti
al point of view (What should the Federal Reserve do today?), as well as
being of parti
ular theoreti
al interest (What are the various types of un
ertainty fa
ed
by poli
ymakers, and what ee
ts should ea
h of these have on optimal poli
y?)
A natural pla
e to begin the investigation of these diÆ
ult questions is within the
framework of monetary poli
y rules, su
h as those put forward by Taylor (1993). The
advantage of these rules is that they are expli
it, welldened, and simple fun
tions of
variables within a
ompletely spe
ied e
onomi
model. Dierent types of un
ertainty
within the model
an then be spe
ied and their ee
ts studied. The fa
t that Taylortype
rules have mat
hed the histori
al behavior of the Federal Reserve for the past 15{20 years
also lends support to their use as an analyti
al laboratory for the investigation of these
issues.
A typi
al setup involves an e
onomy that is linear in all of its variables, and poli
y
makers who minimize an expe
ted dis
ounted sum of squared deviations of goal variables
from their respe
tive targets. The following ba
kwardlooking model serves as an illustra
tive example:1
(yt y ) = '(yt 1y ) (rt r ) + "t (1a)
t = t 1 + (yt y ) + t (1b)
where yt , t , and rt refer to output, in
ation, and the real interest rate in period t, y
and r denote the levels of potential output and the \natural" rate of interest
onsistent
with longrun equilibrium, respe
tively, and rt is set by poli
ymakers at the beginning
of period t based on information available through the end of period t 1. A typi
al
1 I will also
onsider forwardlooking models in the main body of the paper, below.
2
spe
i
ation of poli
ymakers' preferen
es is:
1
X
min (1 Æ ) Et Æs t
(s )2 +
(ys y )2
(2)
s=t
a simple dis
ounted sum of expe
ted squared deviations of output from potential and
in
ation from its target, , with weight
pla
ed on the output gap. The
ase
= 0
orresponds to pure in
ation targeting by poli
ymakers, but does not ne
essarily prevent
urrent or past values of the output gap from entering poli
ymakers' optimal rea
tion
fun
tion, sin
e these variables may help fore
ast future values of in
ation. Note that it is
assumed in the model for simpli
ity that poli
ymakers have
ontrol over the shortterm
real interest rate, rt .
This is a simple, dis
retetime dynami
programming problem with quadrati
ob
je
tive and linear
onstraints, the solution of whi
h is well known (Sargent (1987)):
rt r = a (yt 1 y ) + b ( t 1 ) (3)
where a and b are
onstants, determined by the parameters of the model. The form of
equation (3) and the values of a and b are the same no matter what the varian
es of "
and poli
ymakers behave in a
ertaintyequivalent fashion regardless of the varian
e
asso
iated with the laws of motion of the e
onomy.
A major impli
ation of this nding is that un
ertainty about the level of potential
output, y, should have no ee
t on the monetary poli
y rule. If at time t we have
y = y^ + , where is a meanzero sto
hasti
disturban
e and y^ Et y , then the form
of equations (1a) and (1b) is essentially un
hanged, with y repla
ed by y^, "t repla
ed
by "~t "t (1 '), and t repla
ed by ~t t . Then, as before, the varian
es
of "~t and ~t , and hen
e the varian
e of , have no ee
t on the optimal poli
y ruleone
simply repla
es y with y^ in equation (3) and a
ts as if this estimate y^ were known with
ertainty. This result is emphasized by Estrella and Mishkin (1998), and is standard in
the literature on monetary poli
y rules.
To be sure, not all types of un
ertainty are sterile when it
omes to their ee
ts
on optimal poli
y. For example, if poli
ymakers are unsure about the ee
ts of their
3
ontrol variable rt on the e
onomy, so that is sto
hasti
, the optimal
hoi
e of rt will
depend on the degree of un
ertainty underlying the parameter (Brainard (1967), Sa
k
(1998)). Here, the un
ertainty is multipli
ative with respe
t to the
ontrol variable, rather
than being simply additive.2 In this paper, however, I abstra
t away from multipli
ative
un
ertainty for a number of reasons. First, it is well known that
ertainty equivalen
e does
not hold in the presen
e of multipli
ative un
ertainty. Se
ond, the ee
ts of multipli
ative
un
ertainty on poli
ymakers' optimal response fun
tion (3) are ambiguous in sign (when
more than one
oeÆ
ient is sto
hasti
), and found by several authors to be quantitatively
small in simulations (Estrella and Mishkin (1998), Rudebus
h (1999a)).3 Finally, and
most importantly, the un
ertainty whi
h poli
ymakers appear to be most
on
erned with
today is primarily additive in nature: Has there been a stru
tural break in the growth
rate of produ
tivity, potential output, and the natural rate of unemployment? These are
questions about the state variables of the e
onomy themselves, not about the ee
ts of
hoi
e variables on the e
onomi
state.
The main point of this paper is that additive un
ertainty
an have important ee
ts
on optimal poli
y, even within a linearquadrati
framework. Suppose, for example, that
the state of the e
onomy is Xt, an unobserved, possibly large ve
tor.4 Suppose further that
output and in
ation are observable fun
tions of this unobserved state of the e
onomy Xt.
Poli
ymakers will behave in a
ertaintyequivalent fashion with respe
t to Et Xt , the ex
pe
ted value of Xt, but will use observations of output and in
ation to help infer what the
value of EtXt is. This inferen
e stage of the problem, be
ause it is one of signal extra

tion, is sensitive to the amount of un
ertainty that is present in the observable variables
of the system. In parti
ular, as the noise in the data on the output gap, yt y, in
reases
relative to the signal about Xt , it is optimal to de
rease the
oeÆ
ient on yt y in the
estimation of Et Xt , and in
rease the
oeÆ
ient on in
ation in that estimation. Thus, the
optimal poli
y, when expressed as a redu
edform fun
tion of output and in
ation, rather
2 Note that un
ertainty about r , yt 1, and t 1 are also all additive with respe
t to the
ontrol
variable rt, and thus also have no ee
t on the optimal poli
y in (3).
3 Sa
k (1998), however, nds larger ee
ts of parameter un
ertainty within a monthly VAR framework.
4 Some
omponents of X may be observed without altering the results.
t
4
than Et Xt, is not invariant to the degree of additive un
ertainty in the former variables.
It is important to note that both the rule itself and poli
ymakers' estimate EtXt are fully
optimal in this frameworkat every time t, poli
ymakers' interestrate
hoi
e and estimate
of state variables are globally optimal, and
annot be improved upon in any way. This
is one of the main respe
ts in whi
h the present paper diers from previous work in this
area.
Previous studies of non
ertaintyequivalen
e with additive un
ertainty have followed
either Smets (1998) or Orphanides (1998). Smets (1998) points out that
ertainty equiva
len
e fails to hold when poli
ymakers are
onstrained to respond to only a limited subset of
the state variables of the system, so that the poli
y rule is a
onstrained optimum, rather
than a global optimum. (This restri
tion makes sense, in parti
ular, when the size of the
state spa
e is large, as in Orphanides et al. (2000), who work with the Federal Reserve
Board's ma
ro model.) Un
ertainty about any of the variables in the poli
ymakers' simple
rule will then lead the optimal
oeÆ
ients of the simple rule to
hange.5
Orphanides (1998) fo
uses on the fa
t that the data, su
h as the output gap and
in
ation rate, are observed only imperfe
tly in real time. By the
ertainty equivalen
e
prin
iple, poli
ymakers' optimal response in this
ase is to formulate best estimates of the
output gap, Et (yt yt), and in
ation, Et t , and a
t as if these estimates were known
with
ertainty. Orphanides and others (Aoki (1999), Rudebus
h (1999a,b)) bring about
non
ertaintyequivalen
e in this framework by
onstraining poli
y to rea
t to the a
tual
realtime data, rather than to the best estimates above. This naturally raises the question
as to why the a
tual realtime data are not (realtime) best estimates of the true values.
For example, if yt denotes the level of potential output, and ytjt the realtime esti
mate of the level of potential output, one would normally expe
t that:
yt = ytjt + t (4)
where t is a meanzero random variable. In this
ase, ytjt = Et yt, so the realtime data is
5 Un
ertainty about any state variables in the poli
ymakers' simple rule will have no ee
t; this
not
will be obvious from the dis
ussion of the results in Se
tion 2, below.
5
the realtime best estimate! The
onstraint that poli
ymakers rea
t only to the realtime
data is then not really a
onstraint at all, and the
ertainty equivalen
e prin
iple holds.
It is only be
ause Orphanides formulates the realtime data problem as one of signal
extra
tion, with
ytjt = yt + t (4)0
where yt (the true value underlying the data, or signal) and t (the noise) are orthogonal
random variables, that the
ertainty equivalen
e prin
iple is
ir
umvented. Note that
in (4)0, the realtime data ytjt has the property that ytjt 6= Et yt, so that it is no longer
an unbiased estimate. When the problem is modied in this manner, an in
rease in the
varian
e of t now does have an ee
t on the best estimate Et ytin parti
ular, if yt and t
are normally distributed around zero, then Et yt = y2 =(y2 + 2) ytjt in the univariate
ase. Although the optimal poli
y is still a
ertaintyequivalent fun
tion of poli
ymakers'
best estimate Et yt, expressing the poli
y as a fun
tion of the realtime ytjt now leads to
ee
ts of additive un
ertainty t on the
oeÆ
ients in the optimal rule. It is in this respe
t
that Orphanides (1998) nds an ex
eption to the
ertainty equivalen
e prin
iple.
It should be
lear from the above analysis that the use of realtime data per se has
nothing to do with the non
ertaintyequivalen
e demonstrated in Orphanides (1998). In
stead, it is the introdu
tion of a signal extra
tion problem into the poli
ymakers' inferen
e
step that drives the result. The general impli
ations of formulating poli
ymakers' inferen
e
problem in this way is the domain of the present paper. In addition, the poli
ies that are
emphasized in Orphanides (1998), Orphanides et. al. (2000), and Rudebus
h (1999a,b) are
not fully optimal, so that it is diÆ
ult in those papers to distinguish between the ee
ts
of un
ertainty proper and the ee
ts of un
ertainty intera
ted with a substantially
on
strained poli
y response fun
tion. In the present paper, all poli
y rules and all estimation
will be fully optimal. Again, this is one of the main points of departure from previous
work.
The paper pro
eeds as follows. Se
tion two develops the relationship between signal
extra
tion and non
ertaintyequivalen
e in a simple des
riptive model of the e
onomy,
under both naive and rational expe
tations. Se
tion three extends these results to the
6
general linearquadrati
Gaussian framework and proves the
oeÆ
ient attenuation result
for the general
ase. Se
tion four extends the basi
model to allow for dynami
evolution of
un
ertainty, and proves that the results of se
tion two are robust to this extension. Se
tion
ve dis
usses the results and
on
ludes.
2. The Basi Model
Poli
ymakers have preferen
es over in
ation and output of the form:
1
X
min (1 Æ ) Et Æs t
(s )2 +
(ys y )2
(5)
s=t
where denotes poli
ymakers' longrun target for the in
ation rate and y denotes the
level of \potential" output
onsistent with longrun equilibrium. The
ase
= 0
orre
sponds to pure in
ation targeting, but does not ne
essarily pre
lude
urrent or past values
of the output gap from entering poli
ymakers' optimal rea
tion fun
tion, sin
e they may
help fore
ast in
ation.
For the purposes of this se
tion, the unobserved \signal" variable Xt will be taken to
be a s
alar. The interpretation in this
ase is that the true underlying state of the e
onomy
is s
alar, or alternatively that Xt is an index of in
ationary pressures (whi
h one
ould
think of as
orresponding to the
on
ept of \ex
ess demand"), this index itself being an
amalgam of a great many variables underlying the e
onomy. In either
ase, Xt is assumed
to evolve a
ording to:
Xt = 'Xt 1 (rt r ) + "t (6a)
while the output gap and in
ation are observable fun
tions of this unobserved state:
(yt y) = Xt + t (6b)
t = te + Xt + t (6
)
Here r denotes the \natural" rate of interest,
onsistent with longrun equilibrium, and
, , and ' are known positive parameters with ' < 1. The sto
hasti
disturban
es "t , t ,
and t are independent of ea
h other, over time, of
urrent and prior values of X , y, and ,
7
and are normally distributed with
onstant varian
es "2, 2, and 2 , respe
tively. The
variable Xt (and its past values) are never observed by poli
ymakers and must be inferred
from previous observations of output and in
ation. Equations (6b) and (6
) represent the
signal extra
tion aspe
t of the problem, with t and t denoting noise disturban
es that
are orthogonal to the underlying signal Xt.
One may assume that y is observed with
ertainty, or alternatively that it is sto
has
ti
( and r are presumed to be known with
ertainty, although in prin
iple these re
stri
tions, too,
ould be dropped). Letting y = y^ + , where is a random variable,6 one
an rewrite equation (6b) as:
(yt y^ ) = Xt + ~t (6b)0
where ~t t + . From the point of view of poli
ymakers' de
ision at time t, this
is equivalent to simply in
reasing the varian
e of t in equation (6b), so the analysis is
simplied, without loss of generality, by restri
ting attention to that equation, and studying
the ee
ts of an in
rease in 2 .7
Equation (6
) also in
orporates e
onomi
agents' prior expe
tation of the in
ation
rate, te, whi
h is known and xed at the beginning of period t. In prin
iple, this
ould
be a rational expe
tation (te Ett , where Et denotes the expe
tation at the beginning
of period t, before sho
ks are realized), a naive adaptive expe
tation (te t 1 ), or a
fore
ast derived by some other means. In pra
ti
e, the spe
ial linearquadrati
stru
ture
of the problem is lost unless te is either a rational expe
tation or a xed linear
ombination
of nitely many lags of observable variables; thus, that assumption will be maintained here.
In addition, for the purposes of this se
tion, it will be assumed that te is either rational
or purely naive (te t 1 ), as this keeps the number of state variables in the model to a
6 I assume here that the error is orthogonal to
urrent and past values of X and all other variables
of the system. This is the
ase, for example, when the un
ertainty surrounding y is due to a regime
hange that has been known to o
ur in the previous period, so that y , whi
h was previously known with
ertainty, is suddenly un
ertain. The more general
ase, where may be
orrelated with Xt, is not diÆ
ult,
and is solved in Se
tions 3 and 4, below. The results there are essentially identi
al to those presented here.
7 There is one
ompli
ation in that un
ertainty about y diminishes over time (in the absen
e of new
sho
ks to y ) as poli
ymakers learn from observations of y and . This poses no problem, however,
be
ause the separation of estimation and
ontrol in this linearquadrati
framework implies that this
learning pro
ess (whi
h takes pla
e via Kalman ltering) has no ee
t on the optimal poli
y (see, for
example, Bertsekas (1987), p. 292).
8
minimum.8
The timing of poli
ymakers' observations and a
tions is as follows. At the beginning
of period t, poli
ymakers update their beliefs about Xt 1 based on observations of yt 1 ,
t 1 , and the earlier
hoi
e of rt 1 . Based on these updated beliefs, poli
ymakers then
hoose a value of rt that minimizes the expe
ted loss fun
tion (5). Sho
ks to the e
onomy
("t , t , and t ) are then realized and the values yt and t observed. Thus, poli
ymakers'
information set at time t is:
It f; ;
; Æ; '; "2; 2; 2 ; E0X0; Var0 X0; ; r; y; te; se; s; rs; ys j s < tg (7)
where E0X0 and Var0 X0 denote the mean and varian
e of poli
ymakers' prior (time 0)
distribution on X0, whi
h is assumed to be normal.9
Poli
ymakers update beliefs about Xt 1 via Kalman ltering. Be
ause ("t ; t ; t ) is
multivariate normally distributed, this is the optimal inferen
e pro
edure (minimizing the
meansquared error of the estimate), and is equivalent to Bayesian updating.10
The optimal solution to poli
ymakers' problem (5), given the stru
ture of the e
on
omy (6) and information set (7), is:
rt = r + a Et Xt 1 + b (te ) (8)
where a and b are positive
onstants determined by the parameters of the system.11 Note
that a and b are
ompletely invariant to "2, 2, 2 , and Vart Xt 1 (poli
ymakers' (time t)
prior varian
e on Xt 1, derived re
ursively from Var0 X0 by the Kalman ltering algo
rithm). In this respe
t, the linearquadrati
problem with signal extra
tion displays
er
tainty equivalen
e.
8 It ise not diÆ
ult to show (by introdu
ing a
oeÆ
ient on te in (6
) and letting it tend to one) that
when t is rational, poli
ymakers
an set Ett = and Et yt = y in every period, and thus te = in
every period. See Appendix A. Also note that the exa
t pro
ess by whi
h agents' expe
tations are formed
has no ee
t on poli
ymakers' period t estimate of Xt. It is only be
ause of the
ontrol aspe
ts of the
problem that assumptions about agents' expe
tation formation are required.
9 The assumption of normality is not required if we restri
t poli
ymakers to linear estimates of the
unobserved state Xt. Under the assumption that all disturban
es and priors are Gaussian, the optimal
estimate Et Xt is linear, so the need to make this distin
tion does not arise.
10 Without any assumptions regarding the distribution of (" ; ; ), Kalman ltering is still the optimal
t t t
linear inferen
e pro
edure.
11 This is most easily seen by solving the poli
ymakers' problem expli
itly. See Appendix A.
9
In forming the optimal estimate Et Xt 1 , however, poli
ymakers do respond to the
amount of un
ertainty in the problem. Their prior (time t 1, i.e., before values of yt 1
and t 1 are observed) distribution on (Xt 1; yt 1 y; t 1 te 1 ) is given by:
(Xt 1 ; yt 1 y ; t 1 te 1 )
2 3 2 2 3 !
' Et 1 Xt 2 (rt 1 r ) x x2 x2
N 4 ' Et 1 Xt 2 (rt 1 r ) 5 ; 4 x2 + 2 x2 5
' Et 1 Xt 2 (rt 1 r ) x2 + 2
2
where I have let x2 denote Vart 1 Xt 1 , poli
ymakers' prior (time t 1) varian
e on Xt 1.
Their posterior distribution on Xt 1 (after observing yt 1 and t 1 ) then has mean:
Et X t 1 = 'Et 1 Xt 2 (rt 1 r )
2
x2 2 x2 + 2 x2
+ (yt y ) 'Et 1 Xt 2 + ( rt r ) (9)
1 1
x4 + x2 x2 + 2
+ (t te 1) 'Et 1 Xt 2 + (rt r )
1 1
where (x2 + 2)( 2 x2 + 2 ) (x2 )2 = x2 2 + 2 x2 2 + 22 . Equation (9)
is analogous to the simpler formula for signal extra
tion with one
observable variable,
Et Xt 1 = Et 1 Xt 1 + x2 =(x2 +2 ) (yt 1 y ) Et 1 (yt 1 y ) , with additional terms
in the
oeÆ
ients that take into a
ount the
ovarian
e between output and in
ation.12
Grouping terms in (9) yields:
=
2 2
Et Xt 1 'Et 1 Xt 2 (rt 1 r )
x2 2
+ (yt 1 y ) (10)
+ x (t 1 te 1 )
2 2
12 Given the normality assumption, the formula for the best predi
tor Et Xt 1 is the theoreti
al regres
sion/proje
tion:
Et Xt 1 = Et 1 Xt 1 + Covt 1 (Xt 1 ; Zt 1 )(Vart 1 Zt 1 ) 1 (Zt 1 Et 1 Zt 1 )
where Zt 1 [ tyt 1 1 yte 1 ℄. This yields:
2 1
x + 2 x2
Et Xt 1 = Et 1 Xt 1 + [ x x ℄ 2 2 (Zt 1 Et 1Zt 1)
2 x2 + 2
whi
h is the expression given in (9).
10
where the Et 1 Xt 2 term
an be
as
aded ba
kward and expressed as a fun
tion of lagged
observations of y, , and r to the point where the original distribution on X0 is negligible,
as it is multiplied by a large power of '.
Note that, even though poli
ymakers' optimal rea
tion fun
tion (8) is
ertainty
equivalent in terms of the estimate Et Xt 1, when expressed as a fun
tion of present and
past observable variables (y, , and r),
ertainty equivalen
e in the redu
ed form no longer
holds. The varian
es of the additive disturban
e terms enter into the
oeÆ
ients of (10).
For example,
onsider the ee
ts of an in
rease in 2 on the
oeÆ
ients in equa
tion (10).13 As 2 in
reases, the quantity in
reases, but less than proportionately. This
implies that the
oeÆ
ient on yt 1 y in (10) de
reases in magnitude, so that poli
ymakers
pla
e less weight on the observation of the un
ertain output gap in forming their inferen
e
about the underlying state of the e
onomy. Thus, we have an example of
oeÆ
ient atten
uation on the noisy or un
ertain variable. In addition, the
oeÆ
ients on ea
h of the other
variables in equation (10) in
rease in absolute value: poli
ymakers pla
e more weight on
those variables about whi
h they are relatively more
ertain. In this sense, poli
ymakers
be
ome \less proa
tive and more rea
tive," responding less for
efully to the
urrent value
of the output gap, and more for
efully to past observations of output and in
ation, and the
urrent in
ation rate, be
ause these variables provide more reliable information about the
urrent state of the e
onomy and future values of poli
ymakers' goal variables, in
ation
and output.14
These
on
lusions are not idiosyn
rati
to an in
rease in the varian
e of the additive
disturban
e t . For example, an in
rease in 2 , instead of 2, leads to a de
rease in the
oeÆ
ient on (t 1 te 1 ) in equation (10), and an in
rease in the
oeÆ
ients on the
other variables in that equation. Exa
tly analogous results (lower
oeÆ
ient on the noisy
13 As noted above in footnotes 6 and 7, this
an be thought of as
orresponding to a stru
tural break in
y whi
h has been known to o
ur in the previous period. This
auses 2 to rise for the
urrent period
but not for previous periods, and hen
e leaves x2 unae
ted.
14 This \less proa
tive" result
ould be emphasized by setting up the main model with an additional lag
in equation (6
), so that:
t = te + Xt 1 + t (6
)0
However, this
ompli
ates the analysis by making the signal extra
tion problem a fun
tion of updates of
two lags of the variable Xt instead of only one. For this reason, this approa
h was not taken here.
11
variable, higher
oeÆ
ients on the others) obtain in response to an in
rease in x2 , the
poli
ymakers' prior varian
e on Xt 1. Thus, the result des
ribed above is robust, and
derives not from any spe
ial assumptions surrounding the model, but rather from the
general prin
iple that in linear regression, or statisti
al proje
tion, less weight is given to
observations that have higher varian
e.
Obviously, as with linear regression and statisti
al proje
tion, the
ovarian
es be
tween the dierent variables matters for the
oeÆ
ients in (10). So far, I have abstra
ted
away from this problem by assuming that the disturban
es are orthogonal to ea
h other,
and to poli
ymakers' prior for the variable Xt 1, but I show below for the general
ase, with
arbitrary
ovarian
es, that the basi
oeÆ
ient attenuation result still holds: an in
rease
in the un
ertainty surrounding a given variable
auses poli
ymakers to assign less weight
to that variable in forming their best estimate of the underlying state of the e
onomy,
Et Xt 1 . Moreover, the ampli
ation of the
oeÆ
ients on in
ation and its lags in (10)
is also quite robust, and holds for models more general than that of the present se
tion.
For example, se
tion 4 proves this result for the
ase where in
reased un
ertainty about
potential output extends ba
kward any number of periods, is
orrelated a
ross time, and
is
orrelated with poli
ymakers' priors about the unobserved state of the e
onomy Xt 1,
in a manner whi
h is
onsistent with poli
ymakers learning about potential output over
time.
Finally, the dire
tion that the
oeÆ
ients in equation (10) move
an be ambiguous
when one in
reases the varian
e of more than one variable at a time. For example, in
reas
ing 2 and x2 simultaneously
auses the
oeÆ
ient on the unae
ted variable (t 1 te 1 )
to in
rease unambiguously, but the ee
t on the other two
oeÆ
ients is not
lear.15 I will
return to this issue in se
tion 4, below, but will remove the ambiguity by making use of
the expli
it relationships between x2 and past values of 2 that arise from the Kalman
ltering algorithm.
15 Although if 2 in
reases by more than x2 , it is easy to show that the
oeÆ
ient on (yt 1 y ) must
de
rease in magnitude, while if x in
reases by more than , then the
oeÆ
ient on Et 1Xt 2 must
2 2
de
rease in magnitude.
12
2.1 Signal Extra
tion vs. Imperfe
t Observation of State Variables
It is important to note that the above results hinge
ru
ially on setting up model (6) as
one involving signal extra
tion, rather than one simply involving imperfe
t observation of
state variables. For example, repla
ing (6b) with:
X t = (yt y ) + t (6b)y
or, with an un
ertain potential output:
Xt = (yt y^ ) + ~t (6b)z
where y y^ + and ~t t , the non
ertaintyequivalen
e results des
ribed above
are
ompletely eliminated. In both (6b)y and (6b)z, the poli
ymakers' optimal estimate
of the underlying state of the e
onomy, Et Xt 1 , is simply (yt 1 y ) in the rst
ase,
and (yt 1 y^) in the se
ond. Plugged into the
ertaintyequivalent stru
tural response
in equation (8), the redu
edform poli
y response thus retains the
ertainty equivalen
e
property.
Note that this analysis implies that it is not realtime data per se that justies
aution on the part of poli
ymakers in Orphanides (1998) and Rudebus
h (1999a). If the
realtime data are unbiased fore
asts of the true values, analogous to (6b)y or (6b)z, then
ertainty equivalen
e holds, and the amount of un
ertainty surrounding these realtime
estimates (2) is
ompletely irrelevant for optimal poli
y. Certainty equivalen
e only fails
to hold in their framework if the realtime data are realizations of true values plus noise
(and thus are biased estimates of the true values), so that estimation of the true values
involves a signal extra
tion problem, as in (6b).
This naturally raises the question as to whether the realtime data are better mod
eled as rational estimates or as noise
ontaminated observations. Orphanides (1998, 1999)
presents gures demonstrating that the output gap, in parti
ular, has been badly mis
measured by poli
ymakers in real time. However, both the size of these errors and their
serial
orrelation were not evident until several years after the fa
t, so it is not
lear that
13
these realtime estimates of the output gap were not rational at the time. To take the
position that poli
ymakers were deliberately irrational in their realtime estimates seems
unwarranted without a more rigorous analysis to support this point of view.
Rigorous analysis of the performan
e of realtime data and oÆ
ial fore
asts has been
undertaken by a number of authors, albeit with data that is more readily observable than
the output gap, su
h as real GDP and in
ation. Mankiw and Shapiro (1986) analyze
whether the realtime real GNP data produ
ed by the BEA is better modeled as a rational
fore
ast or as a realization with noise of the \true" (i.e., \nal release") value. They nd
that the realtime data appear to be unbiased and eÆ
ient rational fore
asts. M
Nees
(1995) looks at the oÆ
ial fore
asts of real GNP/GDP and in
ation published by the
CBO and Federal Reserve System (as presented in \HumphreyHawkins" reports) and
nds that they perform at least as well, if not better, than privatese
tor fore
asts in terms
of meansquared error. Romer and Romer (2000) nd that not only are the Federal Reserve
Board's internal \Greenbook" fore
asts of output and in
ation unbiased and eÆ
ient, they
ompletely dominate private se
tor fore
asts, in the sense that the private se
tor fore
ast
should be thrown out entirely if the Board's fore
asts were to be made publi
.
These results might at rst seem to
ontradi
t Rudebus
h's (1999a) nding of a
signi
ant, irrational \noise"
omponent in the realtime in
ation data (as measured by
either the GNP/GDP de
ator or xedweight pri
e index). However, as in Orphanides
(1998, 1999), Rudebus
h's \nal" data is from the perspe
tive of the late 1990's, and thus
in
ludes denitional revisions to GNP and
hanges in base year. In
ontrast, all of the
papers
ited above take parti
ular
are to evaluate the performan
e of the fore
asts with
respe
t to a nal measure of the statisti
on a denitionally
onsistent basis. For example,
it seems unfair to evaluate the rationality of poli
ymakers' 1970 estimate of real GNP and
in
ation using today's estimates of 1970 GDP in
hainweighted 1996 dollars, yet this is
exa
tly what Orphanides and Rudebus
h do. Thus, it is likely that the \noise" found by
Rudebus
h (1999a) re
e
ts nothing other than denitional
hanges in the data rather than
deviations from rationality in poli
ymakers' estimates.16
16 The problem of denitional revisions to output and in
ation does raise an interesting theoreti
al issue:
14
The
ase for biased realtime data is thus somewhat un
ompelling from an empiri
al
as well as a theoreti
al standpoint. This implies that models in
orporating only realtime
data un
ertainty should be
ertaintyequivalent. However, framing poli
ymakers' inferen
e
problem about the state of the e
onomy as one of signal extra
tion more generally, as is
done in this paper, is still quite plausible. The interpretation of poli
ymakers' estimation
pro
ess in this
ase is one of an unobserved, possibly
ompli
ated state of the e
onomy
that must be inferred using (possibly a large number of) observable e
onomi
indi
ators.
For example, Chairman Greenspan is renowned for looking at a wide variety of e
onomi
statisti
s in an attempt to infer the
urrent state of the e
onomy and its future
ourse.
Alternatively, one
ould think of the unobserved e
onomi
state X as being an aggregate
on
ept su
h as \ex
ess demand," with poli
ymakers performing signal extra
tion using a
number of utilization and in
ation measures in an attempt to infer what the level of ex
ess
demand is. When poli
ymakers fa
e signal extra
tion problems like these, as in the model
above and the general framework below, a strong
ase for
aution in the fa
e of un
ertainty
an still be made.
2.2 CoeÆ
ient Attenuation, Simple Rules, and Robust Control
The impli
ations of this paper
ontrast in an interesting way with those from the literature
on \simple rules" and \robust
ontrol." In parti
ular, the signal extra
tion framework of
the previous se
tion found that an in
rease in un
ertainty surrounding a given indi
ator
variable should be met with an attenuation in poli
ymakers' response
oeÆ
ient to that
variable, and an ampli
ation of their response
oeÆ
ients on all other indi
ators, about
whi
h un
ertainty hasn't
hanged.17
While it is fairly
lear that the rationality of a fore
ast should be evaluated using denitionally
onsistent
realizations of the statisti
, it is not so obvious that this is the
orre
t approa
h when we start to
onsider
poli
ymakers' welfare. For example, it is possible that the postdenitional
hange statisti
is a better
measure of the loss poli
ymakers, and so
iety, a
tually experien
ed.
I would, however, take the following view. Poli
ymakers' losses are a fun
tion of the true underlying
state of the e
onomy. The various aggregate statisti
s are generally poor re
e
tions of this state. A
denitional
hange in one of the aggregate statisti
s
hanges the relationship between the true state of
the e
onomy and the given aggregate statisti
, but does not
hange the true state of the e
onomy itself.
Thus, the optimal
oeÆ
ients in poli
ymakers' signal extra
tion problem should
hange in response to the
denitional revision, but there is not any irrational \noise" in the realtime data or in poli
ymakers' losses.
17 In the
ompletely general setup of the model (se
tion 3), only the
oeÆ
ient attenuation result on
15
The literature on \simple rules," in
ontrast, generally nds that poli
ymakers
should attenuate their response
oeÆ
ients on all variables in their rea
tion fun
tion,
even if the in
rease in un
ertainty surrounds only a single variable.18 This literature,
typied by Smets (1998),
onsiders optimal poli
y within a
lass of rules that rea
t to
only a limited subset of state variables in the model. Be
ause these simple rules are not
globally optimal, they typi
ally do not possess the
ertainty equivalen
e property. Thus,
observation error on a variable, even of the type in (6b)y and (6b)z, will generally lead to
non
ertaintyequivalent behavior in these
onstrainedoptimal rules.
For example, Smets (1998) restri
ts poli
ymakers to rules involving only one lag of
output, the fourquarter average in
ation rate, and one lag of the interest rate as argu
ments, and nds that the optimal
oeÆ
ients on all of these variables are attenuated by
an in
rease in un
ertainty surrounding the output gap. Orphanides et al. (2000), investi
gating a similarly
onstrained
lass of rules within the Federal Reserve Board's FRB/US
model, also nd that attenuating the
oeÆ
ients on every variable in the rule is the best
response to in
reased un
ertainty surrounding the output gap.19 The reason that these
ndings dier from those of the present paper
an be explained as follows. The optimal
rule in all of these models is a fun
tion of multiple lags of the output gap, in
ation, and
interest rates (and, in the FRB/US model, many other variables as well). To the extent
that these variables are omitted, those that enter the simple rule serve partially as proxies
for the variables that have been ex
luded. If the fourquarter average in
ation rate enters
negatively into an estimation equation for some of these other terms (su
h as past values
the un
ertain variable holds unambiguously. However, both in the simple framework (6) above and its
generalization to un
ertainty extending ba
kward any number of periods below (se
tion 4), ampli
ation
of
oeÆ
ients on the other variables, about whi
h un
ertainty hasn't in
reased, holds generally.
18 Drew and Hunt (2000) is one ex
eption to this rule, but their model and ndings are idiosyn
rati
for a
number of reasons. First, the RBNZ's ma
ro model is signi
antly nonlinear, so that
ertainty equivalen
e
would fail to hold even with a fully optimal rule. Se
ond, and more importantly, poli
ymakers' estimate
of potential output is irrational, using an HPlter that yields estimates that are signi
antly
orrelated
with the business
y
le. Poli
ymakers thus have a strong in
entive to rea
t aggressively to the output gap
estimated in this way, sin
e it will generally be smaller than the true output gap that enters their loss
fun
tion.
19 Orphanides et al. do nd that under some
ir
umstan
es an ampli
ation in the
oeÆ
ient on the
in
ation rate is optimal. However, this only o
urs when they in
rease the relative weight on the output
gap in poli
ymakers' loss fun
tion to high levels (0.75 or above).
16
of the interest rate), then the desired ampli
ation in
oeÆ
ients on
urrent and past in
ation that I nd will be oset by the desired ampli
ation in the
oeÆ
ients on these
additional lagged output, in
ation, and interest rate terms, making the overall ee
t on
the in
ation
oeÆ
ient ambiguous.20
Finally, the emerging literature on \robust
ontrol" arrives at just the opposite
on
lusion: poli
ymakers ought to respond more aggressively to every variable in their
rea
tion fun
tion when fa
ed with model un
ertainty. This literature, typied by Onatski
and Sto
k (2000),
hooses
oeÆ
ients of a poli
y rule to minimize the maximum loss over
all possible values for a given parameter within a given range; thus, the poli
ymaker is
guaranteed not to make mistakes that are extremely
ostly for parameters within this
range. This approa
h is
learly very dierent from the maximization of expe
ted value
approa
h I have taken here, so it is not surprising that the results dier. Intuitively, their
ndings are driven by the fa
t that a bad draw on the ee
tiveness of the poli
y tool (the
parameter in my model)
an result in very large losses if the rule's responsiveness is not
suÆ
iently great. However, it is not
lear that an in
rease in additive un
ertainty about
potential output would lead to the same
on
lusions.
3. Signal Extra tion in the General LQG Framework
In this se
tion, I solve the poli
ymakers' signal extra
tion and optimization problem for
the general linearquadrati
Gaussian framework, and provide the relevant proofs.
As before, I denote the underlying state of the e
onomy by Xt, whi
h may now be a
ve
tor. Xt evolves a
ording to a linear fun
tion of one lag of itself and a ve
tor of poli
y
instruments rt . Thus,
Xt = AXt 1 + Brt + "t (11)
20 There are some other minor dieren
es at work as well. For example, Smets evaluates the ee
ts
of in
reases in the varian
e of the output gap that go ba
k into the innite past, while the experiment
onsidered in this se
tion has been more
losely related to a
hange in regime or stru
tural break at a
given point in time t. However, I show in se
tion 4, below, that the results of this paper
ontinue to
hold when un
ertainty about the output gap extends ba
kward any number of periods. Also, Smets and
Orphanides et al.
onstrain their simple poli
y rules to be fun
tions of in
ation deviations from target (the
fourquarter moving average of (t 1 )) rather than the in
ation surprise (t 1 te 1) whi
h would
be
loser to the optimal inferen
e pro
edure I derive in equation (10).
17
where A and B are known matri
es of the appropriate dimensions. Any
onstants
an
be in
orporated by dening one
omponent of Xt to be a ve
tor of ones. I denote the
observable variables of the system by Zt . These may be a subset of the variables in Xt,
noisy realizations of a linear fun
tion of variables in Xt , or some
ombination of the two.
Thus,
Zt = CXt + t (12)
where C is a known matrix of appropriate dimension, with every observable relationship
among the elements of Xt
orresponding to a row in (12). The noise ve
tor t may have
some
omponents that are always zero,
orresponding to elements of Xt that are a
tually
observed. Other
omponents of Xt, that are not dire
tly observed, must be inferred from
observations of Zt .21 Note that (12) has been set up as a signal extra
tion problem rather
than one of imperfe
t observation (in whi
h
ase Xt would be a fun
tion of the observable
variables Zt plus disturban
e terms). The latter would exhibit
ertainty equivalen
e; the
former does not (with respe
t to the observable variables).
The sto
hasti
disturban
es "t and t are assumed to be independent of ea
h other,
over time, of
urrent and past values of r, X , and Z , and are (multivariate) normally
distributed with
onstant varian
e
ovarian
e matri
es " and , respe
tively.22 In pra

ti
e, these assumptions are not as restri
tive as it might seem, be
ause serial
orrelation
and
ross
orrelation of t and "t
an be introdu
ed by in
luding lags of these variables as
elements of Xt , and redening the disturban
es in (11) and (12) to be orthogonal innova
tions to these pro
esses.23 Correlation between t and Xt
an be introdu
ed in a similar
fashion.
21 It should be noted that equations in (12) that are redundant, or are not informative about Xt, in the
sense that poli
ymakers' prior varian
es on the
orresponding elements of Zt are zero, should be dropped
from (12). Intuitively, realizations of these
omponents of Zt
ontain no new information, and thus are
irrelevant for0 updating poli
ymakers' beliefs about Xt. Mathemati
ally, this ensures that the matrix
C t 1jt 1C + is nonsingular in the updating equations below.
22 No diÆ
ulties arise when one allows and to vary over time, so long as this variation is inde
"
pendent of the poli
y instrument.
23 Aoki (1967), for example, makes this observation (pp. 38{39).
18
Poli
ymakers minimize a quadrati
loss fun
tion:
1
X
min (1 Æ ) Et Æ s t Xs0 DXs (13)
s=t
where D is a positive semidenite matrix. Note that this spe
i
ation does not pre
lude
poli
ymakers' preferen
es from depending on observables Z , sin
e X
an be expanded to
in
lude elements of Z as needed. Past values of r
an also be in
orporated into X and Z .
Poli
ymakers
hoose a value for the ve
tor of instruments rt at the beginning of ea
h
period t,
onditional on all information available through the end of period t 1. After
rt is
hosen, the sho
ks "t and t are realized, and the value of the ve
tor Zt is observed.
Poli
ymakers' information set at the beginning of period t is thus:
It fA; B; C; D; Æ; "; ; E0X0; Var0 X0; Zs j s < tg (14)
where E0X0 and Var0 X0 denote the mean and varian
e of poli
ymakers' prior (time 0)
distribution on X0, whi
h is assumed to be normal.
Poli
ymakers update beliefs about Xt via Kalman ltering, whi
h is the optimal
inferen
e pro
edure given the assumptions of normality above. Letting sjt denote Vart Xs,
the varian
e of Xs
onditional on information available at the beginning of period t, we
have the re
ursive equations:
Et 1 X t 1 = A Et 1Xt 2 + Brt 1 (15a)
Et 1 Z t 1 = C Et 1 Xt 1 (15b)
Et X t 1 = Et 1 Xt 1 + t 1jt 1C
0 C t 1jt 1 C 0 + 1
Zt 1 Et 1 Z t 1
(15
)
t 1jt 1 = At 2jt 1A0 + " (16a)
t 1jt = t 1jt 1 t 1jt 1 C 0 C t 1j t 1 C
0 + 1 C t 1jt 1 (16b)
Note that the varian
e tjt evolves deterministi
ally over time, as is typi
al in the LQG
framework. In parti
ular, the varian
es of poli
ymakers' future estimates are unae
ted by
19
their
hoi
e of the
urrent instrument rt .24 This leads to separability between the estima
tion and
ontrol stages of poli
ymakers' problem, and hen
e to the
ertainty equivalen
e
result of the following proposition.
Proposition 1: The optimal solution to poli
ymakers' problem (13), subje
t to the law
of motion (11), observation equation (12), and information set (14) is given by:
rt = (B 0 V B ) 1 B 0 VA Et Xt 1 (17)
where V is the \value" matrix, dened to be the unique negative semidenite solution to
the Ri
ati equation
V = D + ÆA0 VA ÆA0 V B (B 0 V B ) 1 B 0 VA (18)
Proof: Bertsekas (1987), pp. 292{293.
Equation (17) is
ertaintyequivalent with respe
t to the state variable Xt 1. How
ever, as should be
lear from the previous se
tion,
ertainty equivalen
e generally will not
hold with respe
t to the observable variables Zt 1 . The following proposition demonstrates
this fa
t by proving the
oeÆ
ient attenuation result from the previous se
tion for the gen
eral LQG framework. Note that by holding t 1jt 1 xed in what follows, the model is
onsistent with the interpretation that a stru
tural break in the degree of un
ertainty
surrounding the indi
ator variables has o
urred in the previous period.
Proposition 2: Suppose that the varian
e of the rst
omponent of t 1 in (12) is in
reased, in the sense that element (1; 1) of is in
reased while all other elements of ,
and all elements of t 1jt 1 in (15) and (16), are held xed. Then poli
ymakers' optimal
response to observables, obtained by substituting (15) into (17), exhibits an attenuation in
the response of all elements of poli
ymakers' instrument rt to the rst
omponent of Zt 1 .
Proof: Re
all that rt is a ve
tor of instruments, hen
e the proposition states that the
optimal setting of ea
h of these is attenuated with respe
t to the rst
omponent of Zt 1 .
This is intuitive be
ause the ordering of the elements of rt is arbitrary.
24 The poli
ymaker's
hoi
e of rt does not ae
t the signal extra
tion aspe
ts of the problemneither
the varian
e of s nor the varian
e of Xs for any s tbe
ause it is assumed that the
oeÆ
ient matri
es
A and C are known with
ertainty. This is in marked
ontrast to the \experimentation" motive that
is present in Wieland (1998), where poli
ymakers'
hoi
e of rt helps to resolve the Brainard un
ertainty
about the multipli
ative parameters of the model.
20
Let M denote the positive denite matrix C t 1j t 1 C
0 + in equation (15), and
partition M into:
M11 M12
M21 M22
where M11 is a s
alar, M21 a
olumn ve
tor, and M12 = M210 . Letting N denote M 1 and
partitioning N in a
ordan
e with M , we have:
jM22j=jM j M111 M12 N22
N= (19)
M221 M21 N11 M221 + M221 M21 N11 M12 M221
Let M11 be multiplied by a fa
tor > 1,
orresponding to the in
rease in . Then
jM j in
reases in magnitude be
ause, expanding along the rst row or
olumn, jM j =
M11 jM22 j + S , where S is a sum of element
ofa
tor produ
ts not involving M11 , and M11 ,
jM22j, and jM j are positive. Thus, N11 is attenuated, and it follows from (19) that N21 is
attenuated in the same proportion, say by the fa
tor < 1. Thus the rst
olumn of N is
attenuated by the fa
tor .
Inspe
tion of (18) reveals that V is invariant to the
hange in , and by the in
varian
e of the other parameters in (15), (16), and (17), it follows that the rst
olumn of
(B 0V B ) 1B 0VA t 1jt 1C 0 C t 1jt 1 C 0 + 1 is attenuated by the same fa
tor .
These are exa
tly the
oeÆ
ients in question,
ompleting the proof.
One would like to be able to in
rease the
ovarian
es among the
omponents of t
as well, but unfortunately,
ompletely general statements in this
ase
annot be made.
4. Signal Extra tion Dynami s in the Basi Model
The analysis of the pre
eding se
tions has been essentially stati
in nature, in that a
stru
tural break in un
ertainty was known to have o
urred in the previous period. It
is not
lear, then, that the results still apply if the in
rease in un
ertainty o
urs several
periods earlier, parti
ularly when we take into a
ount the fa
t that in
reased un
ertainty
about potential output feeds through to in
reased un
ertainty about subsequent estimates
of X , the unobserved state of the e
onomy. Moreover, errors in poli
ymakers' estimates
of potential output and X are serially
orrelated, as poli
ymakers learn about mistakes in
21
their estimates only slowly over time. A proper treatment of the basi
model in se
tion 2
would take these dynami
aspe
ts of poli
ymakers' un
ertainty into
onsideration.
Thus, the model here will be essentially the same as in se
tion 2. Poli
ymakers'
preferen
es are assumed to be of the form:
1
X
min (1 Æ ) Et Æs t s2 +
ys2 (20)
s=t
where, for ease of notation, s now denotes the deviation of in
ation from poli
ymakers'
target and ys denotes the output gap, both at time s. The e
onomy is assumed to follow:
Xt = 'Xt 1 rt + "t (21a)
yt = Xt + t (21b)
t = te + Xt + t (21
)
where Xt denotes the unobserved state of the e
onomy, rt the deviation of the real interest
rate from its \natural" value, and te agents' expe
tation of in
ation, as before.
Unlike the earlier model, I now allow the error term to be persistent:
t = t 1 + t (21d)
where is the degree of persisten
e of sho
ks to potential output.25 The normally dis
tributed disturban
es "t , t , and t are assumed to be orthogonal to ea
h other, a
ross
time, and to all other variables in the system.
The timing of poli
ymakers' observations and a
tions is the same as in se
tion 2,
with information set
It f; ;
; Æ; ; '; "2; 2; 2; E0X0 ; Var0 X0; te; se; s; rs; ys j s < tg (22)
25 This interpretation is
lear if we repla
e (21b) with
yt = Xt + t + t (21b)0
where t represents sho
ks to potential output with persisten
e , and t represents whitenoise sho
ks
to the output gap relative to the state of the e
onomy. Clearly, poli
ymakers' problem in this
ase is
essentially identi
al to that given in the text, so I have retained the more parsimonious model there.
22
Poli
ymakers' optimal solution to (20), subje
t to (21) and (22), is given by:
rt = aEt Xt 1 + bEtt 1 +
te (23)
where a, b, and
are
onstants invariant to the un
ertainty surrounding X and (see
Appendix A). In
ontrast to se
tion 2, poli
ymakers now
are about past values of as
well as X . Note that be
ause yt 1 = Xt 1 + t 1 , we
an rewrite (23) as:
rt = a~Et Xt 1 + byt 1 +
te (24)
where a~ a b.
As values of y and are observed, poli
ymakers update their beliefs about X and
by Kalman ltering. Letting x2t , 2t , and xt denote VartXt , Vartt , and Covt(Xt ; t ),
respe
tively, poli
ymakers' best estimate of Xt 1 at time t is given by:
Et X t = 2 2t 1+ xt 1
Et 1 X t
1
t 1 1
+ 2 x2t1+ xt 1
yt (25)
t 1 1
x2t 1 2t x
2
+ t
1 t 1 t 1 te 1
1
where t 2 x2t 2t x2 t + 2 x2t+ 2xt+ 2t for all t. Equation (25) was derived
exa
tly as was poli
ymakers' estimation equation (10) in se
tion 2. The Et 1Xt 1 term
an be written as 'Et 1Xt 2 rt 1 and
as
aded ba
kward, as before.
Given an exogenous in
rease in un
ertainty about s at the beginning of period
s < t, we must begin by tra
ing out its ee
ts on subsequent values of 2 , x2 , and x .
The interpretation of the exogeneity of the in
rease in 2s is that of a stru
tural break in
un
ertainty surounding potential output in period s. The
ase where un
ertainty about
in
reases exogenously in several periods s1 ; s2; : : : ; sk , while perhaps more interesting, is
simply a positive linear
ombination of the ee
ts given below, and thus does not need to
be
onsidered separately.
23
A straightforward
omputation (using equations (16) from the previous se
tion)
shows that poli
ymakers' varian
es evolve a
ording to:26
2
xt+1 xt+1
2 (x2t 2t x2 )
' 2 '
2
0
xt+1 2t+1
= t
t
' 2 + 0 2
" (26)
For notational
onvenien
e, dene
2 ( 2 2 2 )
Nt xt t xt
t (27)
whi
h is the key term in (26). Totally dierentiating Nt with respe
t to 2s , using (26),
yields
dNt 4 2 2 dN
=
ds t
2 2 (xt+ xt ) + ' (t+ xt )
2 t 1
d2s
(28)
whi
h is a re
ursive sequen
e terminating with:
dNs Ns
d2s
= 2
s (29)
2
4
= ( + xs )2
2 xs
s
The rst equality in (29) follows from the assumption that the in
rease in un
ertainty about
s in period s is exogenous, while the
ovarian
e xs and varian
e x2s , whi
h derive from
un
ertainty about and X in prior periods, are held xed.
Note that equations (28) and (29) imply dN =d2s 0 for all s. This fa
t,
together with the relations implied by (26), helps us assess the ee
ts of the break in
un
ertainty about s on the
oeÆ
ients in the periodt estimation equation (25).
For example, the
oeÆ
ient on the most re
ent in
ation surprise, (t 1 te 1 ),
unambiguously in
reases in (25). This follows from the fa
t that:
d (x2t 1 2t 1 x
2
t 1
) = dNt 1 > 0
ds
2 t 1 2 d2s
and the
oeÆ
ient on the in
ation surprise in (25) is positive. Thus, the nding of an
ampli
ation of the
oeÆ
ient on in
ation in se
tion 2 is robust to extending the in
rease
in un
ertainty about the output gap ba
kward any number of periods.
26 Note that Xt and t are
orrelated, sin
e t 1 = yt 1 Xt 1 . Hen
e Covt (Xt 1 ; t 1 ) = Vart Xt 1
and Covt(Xt; t ) = ' Vart Xt 1.
24
Similarly, the
oeÆ
ient on the most re
ent output gap, yt 1 , is ne
essarily attenu
ated, under the assumption that > ' (whi
h
orresponds to assuming sho
ks to potential
output are more persistent than movements in the output gap):
d 2 (x2t 1+ xt 1 )
d2s t 1
2
= 2 (' )2 2 (x2t 1+ 'xt 1 ) x2t 1+ ( + ')xt 1+ '2t 1 dN t 2
d 2
t 1 s
2
= ( ')2 + 2 "2 "2 + '2 dN
2
t 2
d2s
<0
t 1
Thus, this nding from se
tion 2 is also robust, no matter when the stru
tural break in
un
ertainty o
urred.27
The
oeÆ
ient on the lagged estimate of the state variable, Et 1Xt 1 , may either
in
rease or de
rease in (25), a
ording to:
d 2 (2t 1+ xt 1 )
d2s t 1
2 dN
= 2 ( ') ('t 1+ xt 1 ) xt 1+ ( + ')xt 1+ 't 1 dt2 2
2 2 2 2 2
t 1 s
2
= ( ')2 2 '2 "2 + '2 dN
2
t 2
d2s
t 1
whi
h is positive if and only if ( ') > 2 '2 =2 .28 CoeÆ
ients on lags of in
ation
and unemployment, obtained by
as
ading the expe
tational term ba
kward in (25), may
thus also go either way, depending on the sign of the derivative above. For example, if
( ') > 2 '2 =2 , then the
oeÆ
ient on E X in
reases for every s. It then
follows that the
oeÆ
ient on the lagged in
ation surprise ( e ) also in
reases for
every s (although the
hange in
oeÆ
ient on lags of the output gap, y , is not
lear
for s t 2). If ( ') < 2 '2=2 , then the
oeÆ
ient on E X de
reases for every
s, the
oeÆ
ient on y also de
reases for every s, and the
hange in
oeÆ
ient on
( e ) now be
omes analyti
ally un
lear for s t 2.
27 Te
hni
ally, this is only attenuation under the assumption that poli
ymakers' optimal response to the
output gap, b + 2 (x2t 1+ xt 1 )=t 1, is positive. This assumption seems warranted (it
an be shown
that b > 0, for instan
e).
28 The
oeÆ
ient on E
t 1 Xt 1 in (25) is ne
essarily positive, assuming again that > ' (and one
an
show that a~ > 0 in any
ase).
25
Note that I use the terminology \not
lear," rather than \ambiguous," here be
ause the
hange in
oeÆ
ients may be theoreti
ally unambiguous, but
omputation of the
analyti
al derivatives and a further sign
he
k of the result would be required to as
er
tain this fa
t, and these
omputations qui
kly be
ome very burdensome. For example, it
an be shown analyti
ally that the
oeÆ
ient on the se
ond lag of the in
ation surprise,
(t 2 te 2 ), also ne
essarily in
reases for > ' (no matter what the value of 2 '2=2 ),
further
orroborating the nding that poli
ymakers should rea
t more aggressively to in
ation.
5. Con lusions
A standard result in the literature is that optimal monetary poli
y in a linearquadrati
framework is
ertaintyequivalent. It was emphasized in this paper that optimal poli
y
is not
ertaintyequivalent with respe
t to observable variables when poli
ymakers fa
e a
signal extra
tion problem in their estimation of the e
onomi
state. For example, the state
of the e
onomy may be very
ompli
ated, and poli
ymakers may have only a relatively
small number of aggregate statisti
s to help them infer what the state of the e
onomy
is. Alternatively, the state of the e
onomy
ould be regarded as an unobserved aggregate
on
ept su
h as \ex
ess demand," with poli
ymakers using various measures of utilization
and in
ation to help them infer what the level of ex
ess demand is. In both
ases, in
reased
un
ertainty about an indi
ator variable
auses poli
ymakers' optimal rea
tion
oeÆ
ient
on that variable to be attenuated. I also nd that, in the fairly standard model above,
it is optimal for poli
ymakers to amplify their rea
tion
oeÆ
ients on all other e
onomi
indi
ators, about whi
h un
ertanity hasn't
hanged.
It is important to note that signal extra
tion is
ru
ial to the above results. If
instead of a signal extra
tion problem, poli
ymakers fa
e simply imperfe
t observation of
state variables, so that the state is perfe
tly observed up to a whitenoise error term, then
the
ertainty equivalen
e property holds, even with respe
t to observable variables.
The realtime data literature of Orphanides (1998), Rudebus
h (1999a,b), and others
should be regarded as a spe
ial
ase. In those papers, if the realtime data are rational
26
estimates of the underlying true values, then poli
y should be
ertaintyequivalent, even
with respe
t to the realtime data. However, if the realtime data are known to
ontain
an irrational noise
omponent, then poli
ymakers fa
e a signal extra
tion problem in their
estimation of the true values, and
ertainty equivalen
e will no longer hold with respe
t to
the observable, realtime data.
In general, I nd that poli
ymakers should take into a
ount the unobserved na
ture of the e
onomy when applying the
ertainty equivalen
e prin
iple. In parti
ular, it
should be kept in mind that the prin
iple does not apply to poli
ies that are expressed in
terms of observable e
onomi
variables when a signal extra
tion problem is involved in the
estimation of the e
onomi
state.
27
Appendix A: Solution to the Basi
Model
Here I solve the basi
model of se
tion 2 under both naive expe
tations (te t 1 ) and
rational expe
tations (te Et t). Re
all that the model's basi
equations are given by:
Xt = 'Xt 1 (rt r ) + "t (A1)
(yt y) = Xt + t (A2)
t = te + Xt + t (A3)
The solution under naive expe
tations is standard (e.g., Sargent (1987)), and is given by:
rt = r + a Et Xt 1 + b (t 1 ) (A4)
where
a = (B 0 V B ) 1B 0 V A; A 0
' (A5)
b 1 ; B
'
and V is the unique negative semidenite solution to the Ri
ati equation:
0 0 0
V = D + ÆA V A ÆA V B (B V B ) B V A; 1 0
D 0 1
0 (A6)
In parti
ular, the solution (A4) is
ertaintyequivalent, in that a and b are invariant to the
se
ond and higher moments of the sto
hasti
parameters of the system.
For the solution under rational expe
tations, it is easiest to think of repla
ing (A3)
with
t = te + Xt + t (A3)0
and
onsider the limit as tends to one. Under both dis
retion and
ommitment, the
dynami
s of the problem in this
ase are trivial, be
ause it has no persisten
e (poli
ymakers
an set Xt up to a sto
hasti
disturban
e term, and t is a jump variable).
Poli
ymakers' problem in period t thus redu
es to minimizing the periodt loss fun
tion
Et (yt y )2 +
(t )2 .
When poli
ymakers are shortsighted (\dis
retionary"), taking te as xed, it is easy
to show that optimization and rational expe
tations lead to:
2 (1 )
Et t = 2 ; Et (yt y ) = 2 (A7)
+
(1 ) +
(1 )
28
whi
h
onverge to and 0, respe
tively, as ! 1.
Alternatively, when poli
ymakers are farsighted (\
ommitted"), optimization and
rational expe
tations lead to:
2 (1 )
Et t = 2 ; Et (yt y ) = 2 (A8)
+
(1 ) 2 +
(1 )2
whi
h likewise
onverge to and 0 as ! 1.29
Thus, under rational expe
tations, we
an regard poli
ymakers as solving the fol
lowing more standard linearquadrati
problem:
Xt = 'Xt 1 (rt r ) + "t (A9)
(yt y) = Xt + t (A10)
t = + Xt + t (A11)
for whi
h the optimal solution is:
rt = r + a Et Xt 1 + b ( ) (A12)
where a and b are given by equation (A5), exa
tly as before. Obviously, the ( ) term
an be dropped, but leaving it fa
ilitates
omparison to (A5) and the single solution given
in the main body of the text,
rt = r + a Et Xt 1 + b (te ) (A13)
It is then not hard to solve expli
itly for V in (A6) and show that a and b are
ne
essarily positive.
The same methods
an be applied to the slightly more
ompli
ated model of se

tion 4.
29 This is not surprising, sin
e setting t = and (yt y ) = 0 in expe
tation is the global optimum,
and this was a
hieved even under dis
retion.
29
Referen
es
Aoki, Kosuke. \On the Optimal Monetary Poli
y Response to Noisy Indi
ators," unpublished
manus
ript (1999), Prin
eton University.
Aoki, Masanao. Optimization of Sto
hasti
Systems: Topi
s in Dis
reteTime Systems (A
a
demi
Press: New York, 1967).
Bertsekas, Dimitri. Dynami
Programming (Prenti
eHall: Englewood Clis, NJ, 1987).
Brainard, William. \Un ertainty and the Ee tiveness of Poli y," AEA Papers & Pro eed
ings 57 (1967), pp. 411{425.
Drew, Aaron and Benjamin Hunt. \EÆ
ient Simple Poli
y Rules and the Impli
ations of
Potential Output Un
ertainty," Journal of Business and E
onomi
s 52 (2000), pp. 143{159.
Estrella, Arturo and Frederi
Mishkin. \Rethinking the Role of NAIRU in Monetary
Poli
y: Impli
ations of Model Formulation and Un
ertainty," in Monetary Poli
y Rules, ed.
John Taylor (1999), pp. 405{430.
Mankiw, N. Gregory and Matthew Shapiro. \News or Noise: An Analysis of GNP Revi
sions," Survey of Current Business (1986), pp. 20{25.
M
Nees, Stephen. \An Assessment of the `OÆ
ial' E
onomi
Fore
asts," New England E
o
nomi
Review (1995), pp. 13{23.
Onatski, Alexei and James Sto
k. \Robust Monetary Poli
y Under Model Un
ertainty in a
Small Model of the U. S. E
onomy," NBER Working Paper 7490 (2000).
Orphanides, Athanasios. \The Quest for Prosperity Without In
ation," unpublished manu
s
ript (1999), Federal Reserve Board.
Orphanides, Athanasios. \Monetary Poli
y Evaluation with Noisy Information," Federal Re
serve Board Finan
e and E
onomi
s Dis
ussion Series 1998{50 (1998).
Orphanides, Athanasios, Ri
hard Porter, David Reifs
hneider, Robert Tetlow, and
Frederi
o Finan. \Errors in the Measurement of the Output Gap and the Design of Mon
etary Poli
y," Journal of Business and E
onomi
s 52 (2000), pp. 117{142.
Romer, Christina and David Romer. \Federal Reserve Private Information and the Behavior
of Interest Rates," NBER Working Paper 5692 (1996); a revised version is forth
oming,
A.E.R. (2000).
Rudebus
h, Glenn. \Is the Fed Too Timid? Monetary Poli
y in an Un
ertain World," Federal
Reserve Bank of San Fran
is
o Working Paper 99{05 (1999a).
Rudebus
h, Glenn. \Assessing Nominal In
ome Rules for Monetary Poli
y with Model and
Data Un
ertainty," unpublished manus
ript (1999b), European Central Bank Conferen
e.
Sa
k, Brian. \Does the Fed A
t Gradually? A VAR Analysis," Federal Reserve Board Finan
e
and E
onomi
s Dis
ussion Series 1998{17 (1998); a substantially revised version is forth
om
ing, J.M.E. (2000).
Sargent, Thomas. Dynami
Ma
roe
onomi
Theory (Harvard University Press: Cambridge,
MA, 1987).
30
Smets, Frank. \Output Gap Un
ertainty: Does it Matter for the Taylor Rule?" unpublished
manus
ript (1998), European Central Bank.
Svensson, Lars and Mi
hael Woodford. \Indi
ator Variables for Optimal Poli
y," unpub
lished manus
ript (1999), European Central Bank Conferen
e.
Taylor, John. \Dis
retion versus Poli
y Rules in Pra
ti
e," CarnegieRo
hester Conferen
e Se
ries on Publi
Poli
y 39 (1993), pp. 195{214.
Wieland, Volker. \Monetary Poli
y and Un
ertainty About the Natural Unemployment Rate,"
Federal Reserve Board Finan e and E onomi s Dis ussion Series 1998{22 (1998).