Working Paper No.

46

Monetary Policy Rules under Uncertainty:

Empirical Evidence, Adaptive Learning and
Robust Control

by

Wenlang Zhang and Willi Semmler

University of Bielefeld
Department of Economics
Center for Empirical Macroeconomics
P.O. Box 100 131
33501 Bielefeld, Germany
 Monetary Policy Rules under Uncertainty:
Empirical Evidence, Adaptive Learning and
Robust Control
Wenlang Zhang∗ Willi Semmler†
March 2003

Abstract
Monetary policy faces great challenges because of many kinds of un-
certainties such as model uncertainty, data uncertainty and shock uncer-
tainty. This paper explores monetary policy rules under model and shock
uncertainties. Facing such uncertainties, a central bank may resort to
different strategies, it can either reduce uncertainty by learning or just
choose a policy rule robust to uncertainty. Empirical evidence of model
and shock uncertainties is explored in a State-Space model with Marked-
switching in both linear and nonlinear Phillips curves. The evidence in-
dicates that there has been too great uncertainty in the U.S. economy to
define accurately monetary policy rules. Moreover, there seem to have
been structural shifts. On the basis of this evidence, we explore monetary
policy rules with the recursive least squares (RLS) learning. The simula-
tions of the RLS learning in a framework of optimal control indicate that
the state variables do not necessarily converge even in a non-stochastic
model, no matter whether the linear or nonlinear Phillips curve is taken as
constraints. This is different from the results of papers which discuss the
RLS learning without optimal control or in the LQ framework. Finally, we
explore monetary policy rules under uncertainty with robust control and
find that the robustness parameter affects the economic variables greatly.
A larger robustness parameter leads to a smaller variance of the state
variable in a stochastic model and faster convergence of the state variable
in a non-stochastic model. An evaluation of those two methods is given
at the end of the paper.
JEL: E17, E19
Keywords: State-Space Model, Markov-Switching, Nonlinear Phillips
Curve, Recursive Least Squares, Robust Control

∗ Center for Empirical Macroeconomics, Bielefeld University, Germany.

† Center for Empirical Macroeconomics, Bielefeld, and New School University, New York.

1
1 Introduction
In the profession it has increasingly been recognized that formal modelling of
monetary policy faces great challenges because of many kinds of uncertainties
such as model uncertainty, data uncertainty and shock uncertainty. Recent liter-
ature dealing with these uncertainties can be found in Isard, Laxton, and Elias-
son (1999), Söderström (1999), Giannoni (2000), Meyer, Swanson and Wieland
(2001), Wieland (2000), Tetlow and von zur Muehlen (2001a), Orphanides and
Williams (2002), Svensson (1999), Martin and Salmon (1999), Hall, Salmon
Yates and Batini (1999) and so on. These papers explore, usually theoretically,
how a certain kind of uncertainty affects the decisions of the central banks or
households and firms. Semmler, Greiner and Zhang (2002a) undertake a time-
varying parameter estimation of the Phillips curve and Taylor rule by way of the
Kalman filter and observe parameter shifts over time. Cogley and Sargent (2001)
study the inflation dynamics of the U.S. after WWII by way of Bayesian Vector
Autoregression with time-varying parameters without stochastic volatility. Sims
(2001b), however, points out that the monetary policy behavior may not have
experienced such a sharp change as shown by Cogley and Sargent (2001). Sims
and Zha (2002) also study parameter shifts of the U.S. economy and find more
evidence in favor of stable dynamics with unstable variance of the disturbance
than of clear changes in model structure. In contrast to Sims (2001b), Cogley
and Sargent (2002) study the drifts and volatilities of the U.S. monetary poli-
cies after WWII through a Bayesian Vector Autoregression with time-varying
parameters and stochastic volatility and claim to have found regime changes.
Facing model and shock uncertainties, economic agents (central banks for
example) may resort to different strategies: They can either reduce uncertainty
by learning or just choose a policy robust to the model uncertainty without
learning. The results of these two strategies may be different. By intuition
we would expect all agents to improve their knowledge of the economy with
all information available. But recently more and more literature is concerned
with bounded-rationality and the assumption of rational expectation is being
increasingly doubted. Therefore, in the research below we will consider the
situation in which agents improve their knowledge of an economic model through
a certain mechanism of learning. Another interesting topic in macroeconomics
since the 1990s is the nonlinearity of the Phillips curve. It is argued that positive
deviations of aggregate demand from potential output are more inflationary than
negative deviations are dis-inflationary. The nonlinearity of the Phillips curve
will also be dealt with below and this turns out to be an important difference
of our paper from the others.
As stated above, central banks may also resort to a monetary policy rule
robust to uncertainty. This is a completely different strategy from adaptive
learning. In this approach a central bank considers the economic model only
as an approximation to another model that it can not specify. With a so-called
robustness parameter it pursues a monetary policy rule in the “worst case”
scenario. While the adaptive learning considers mainly parameter uncertainty,
the robust control might consider more general uncertainty. In spite of some

2
criticisms, the robust control is given much attention in macroeconomics.
The remainder of this paper is organized as follows. In the second section
we present empirical evidence of model and shock uncertainties in the IS and
Phillips curves by way of a State-Space model with Markov-Switching. We con-
sider both linear and nonlinear Phillips curves. In Section 3 we explore monetary
policy rules under model uncertainty with adaptive learning. Section 4 explores
monetary policy rules with the robust control. Section 5 briefly evaluates adap-
tive learning and robust control and Section 6 concludes the paper.

2 Empirical Evidence of Uncertainty: A State-

Space Model with Markov-Switching
Consider an economic model
∞
X
M in E0 ρt L(xt , ut ), (1)
{ut }0
∞
t=0

subject to
xt+1 = f (xt , ut , εt ), (2)
where ρ is the discount factor bounded between 0 and 1, L(xt , ut ) a loss function
of an economic agent (central bank for instance), xt a vector of state variables, ut
a vector of control variables, εt a vector of shocks and E0 denotes the mathemat-
ical expectation operator upon the initial values of the state variables. This kind
of model can be found in many papers on monetary policy, see Svensson (1997
and 1999), Beck and Wieland (2002) and Clarida, Gali and Gertler (1999) for
example, where the constraint equations are usually IS-Phillips curves. Given
the loss function L(x, u) and the state equation (2), the problem is to derive
a path of the control variable to satisfy (1). The question arising is, however,
whether the state equation can be correctly specified in reality. The uncertainty
of the state equation can be caused by the uncertainty in the shock εt and un-
certainty in parameters and data. Svensson (1999) and Semmler, Greiner and
Zhang (2002b) derive an optimal monetary policy rule from an optimal control
problem similar to the model above and find that the optimal monetary policy
rule is greatly affected by the estimated parameters of the model. Therefore,
if the parameters in the model are uncertain, the optimal monetary policy rule
may also be uncertain.
Semmler, Greiner and Zhang (2002a) estimate time varying parameters of
the traditional Phillips curve with the following State-Space model:

yt = Xt βt + εt , εt ∼ N (0, σε2 ) (3)

βt = βt−1 + ηt , ηt ∼ N (0, ση2 ) (4)

with βt being a vector of time-varying parameters. Note that in this model it is

assumed that the shocks have constant variance and only βt is uncertain. Cog-
ley and Sargent (2001) study the inflation dynamics of the U.S. after WWII by

3
way of Bayesian Vector Autoregression with time-varying parameters without
stochastic volatility. Sims (2001b), however, claims that the monetary policy
behavior may not have experienced such a sharp change as demonstrated by
Cogley and Sargent (2001). Sims and Zha (2002) also study the macroeconomic
switching of the U.S. and find more evidence in favor of stable dynamics with un-
stable disturbance variance than of clear changes in model dynamics. Therefore,
Cogley and Sargent (2002) modify the model by considering both time-varying
parameters and stochastic volatility and claim to have found regime switching.
A drawback of the traditional State-Space model such as (3) and (4) is that
the changes of the time-varying parameters may be exaggerated, because the
shocks are assumed to have constant variance. This is the reason why Cogley
and Sargent (2002) assume stochastic volatility. Therefore in the research below
we assume that εt has state-dependent variance. This is similar to the assump-
tion of Cogley and Sargent (2002). But unlike Cogley and Sargent (2002), who
assume the variances of the shocks to change from period to period, we assume
that there are only two states of disturbance variance with Markov property.
This is to some extent similar to the assumption of Sims and Zha (2002) who
assume that there are three states of economy. With such an assumption we
can figure out the probability of regime switching. One more advantage of the
State-Space model with Markov-switching is that, as we will see afterwards, it
can explore not only parameter uncertainty but also shock uncertainty.
Following Kim (1993) and Kim and Nelson (1999), we simply assume that
εt in (3) has two states of variance with Markov property, namely,
2
εt ∼ N (0, σε,S t
), (5)
with
2 2 2 2 2 2
σε,S t
= σε,0 + (σε,1 − σε,0 )St , σε,1 > σε,0 ,
and
P r[St = 1|St−1 = 1] = p,
P r[St = 0|St−1 = 0] = q,
where St = 0 or 1 indicates the states of the variance of εt and P r stands for
probability. In the research below we explore uncertainty in the IS and Phillips
curves, since these two curves form the core of a monetary policy model. We
will consider both linear and nonlinear Phillips curves.

2.1 Linear IS-Phillips Curves

In this subsection we will explore uncertainty in the traditional linear IS-Phillips
curves which have often been taken as constraints in an optimal control model
such as (1) and (2). In order to reduce the dimension of the model, we estimate
simple Phillips and IS curves with only one lag of the inflation rate and output
gap:
πt = α1t + α2t πt−1 + α3t yt−1 + επ,t , (6)
yt = β1t + β2t yt−1 + β3t (rt−1 − πt−1 ) + εy,t , (7)

4
where πt is the inflation rate, yt the output gap, rt the short-term nominal
interest rate, and επ,t and εy,t are shocks subject to Gaussian distributions with
zero mean and Markov-Switching variances.1 Let it denote the real interest
rate, namely, it = rt − πt , the model can be rewritten in a State-Space form as
follows:

Y t = X t φt + ε t , (8)
φt = Φ̄St + F φt−1 + ηt , (9)

where Φ̄St (St =0 or 1) is the drift of φt and F a diagonal matrix with constant
elements to be estimated from the model. ηt has the distribution shown in eq.
(4). εt is now assumed to have the distribution presented in eq. (5).2

Let ψt−1 denote the vector of observations available as of time t−1. In the
usual derivation of the Kalman filter in a State-Space model without Markov-
Switching, the forecast of φt based on ψt−1 can be denoted by φt|t−1 . Similarly,
the matrix denoting the mean squared error of the forecast can be written as

Pt|t−1 = E[(φt − φt|t−1 )(φt − φt|t−1 )0 |ψt−1 ],

where E is the expectation operator.

In the State-Space model with Markov-Switching, the goal is to form a fore-
cast of φt based not only on ψt−1 but also conditional on the random variable
St taking on the value j and on St−1 taking on the value i (i and j equal 0 or
1):
(i,j)
φt|t−1 = E[φt |ψt−1 , St = j, St−1 = i],
and correspondingly the mean squared error of the forecast is
(i,j)
Pt|t−1 = E[(φt − φt|t−1 )(φt − φt|t−1 )0 |ψt−1 , St = j, St−1 = i].

Conditional on St−1 = i and St = j (i, j = 0, 1), the Kalman filter algorithm

for our model is as follows:
1 Forward-looking behaviors have been frequently taken into account in the Phillips curve.

A survey of this problem can be found in Clarida, Gali and Gertler (1999). Because it is quite
difficult to estimate a State-Space model with forward-looking behaviors, we just consider
backward-looking behaviors in this section. In fact a justification of the backward-looking
model can be found in Rudebusch and Svensson (1999).
2 Theoretically, the elements of F and the variance of η may also have Markov property, but
t
since there are already many parameters to estimate, we just ignore this possibility to improve
the efficiency of estimation. Note that if the elements of F are larger than 1 in absolute value,
that is, if the time-varying parameters are non-stationary, the transition equation should be
altered to be the form of eq. (4).

5
 (i,j)
φt|t−1 = Φ̄j + F φit−1|t−1 , (10)
(i,j) i
Pt|t−1 = F Pt−1|t−1 F 0 + ση2 , (11)
(i,j) (i,j)
ξt|t−1 = Yt − Xt φt|t−1 , (12)
(i,j) (i,j)
νt|t−1 = Xt Pt|t−1 Xt0 2
+ σε,j , (13)
(i,j) (i,j) (i,j) (i,j) (i,j)
φt|t = φt|t−1 + Pt|t−1 Xt0 [νt|t−1 ]−1 ξt|t−1 , (14)
(i,j) (i,j) (i,j) (i,j)
Pt|t = (I − Pt|t−1 Xt0 [νt|t−1 ]−1 Xt )Pt|t−1 , (15)
(i,j)
where ξt|t−1 is the conditional forecast error of Yt based on information up to
(i,j) (i,j)
time t−1 and νt|t−1 is the conditional variance of the forecast error ξt|t−1 . In
order to make the above Kalman filter algorithm operable, Kim and Nelson
(i,j) (i,j)
(1999) develop some approximations and manage to collapse φt|t and Pt|t
into φjt|t and Pt|t
j
respectively.3

For the Phillips curve in our model, we have

Yt = πt , Xt = (1 πt−1 yt−1 ), φt = (α1t α2t α3t )0 , εt = επt ,
with
2
επt ∼ N (0, σεπ,S t
),
2 2 2 2 2 2
σεπ,S t
= σεπ,0 + (σεπ,1 − σεπ,0 )St , σεπ,1 > σεπ,0 ,
and
ηt = (ηα1t ηα2t ηα3t )0 ,
ση2 = (ση2α1 ση2α2 ση2α3 )0 ,
Φ̄St = (Φ̄α1 ,St Φ̄α2 ,St Φ̄α3 ,St )0 ,
 
f α1 0 0
F =  0 f α2 0 ,
0 0 f α3
and similarly for the IS curve, we have
Yt = yt , Xt = (1 yt−1 it−1 ), φt = (β1t β2t β3t )0 , εt = εyt ,
with
2
εyt ∼ N (0, σεy,S t
),
2 2 2 2 2 2
σεy,S t
= σεy,0 + (σεy,1 − σεy,0 )St , σεy,1 > σεy,0 ,
3 As for the details of the State-Space model with Markov-Switching, the reader is referred

to Kim and Nelson (1999, ch. 5). The program applied below is based on the Gauss Programs
developed by Kim and Nelson (1999).

6
and

ηt = (ηβ1t ηβ2t ηβ3t )0 ,

ση2 = (ση2β1 ση2β2 ση2β3 )0 ,
Φ̄St = (Φ̄β1 ,St Φ̄β2 ,St Φ̄β3 ,St )0
 
fβ1 0 0
F =  0 f β2 0 .
0 0 f β3

Next we use the U.S. quarterly data 1961.1-1999.4 for the estimation. The
inflation rate is measured by changes in the Consumer Price Index, the output
gap is measured by the percentage deviation of the log value of the Industrial
Production Index (IPI) from its HP filtered trend and rt is the Federal Funds
rate.4 The data source is the International statistics Yearbook 2000. The es-
timates of the hyper-parameters are shown in Table 1. We find significant dif-
ferences between σεπ,0 (0.0021) and σεπ,1 (0.0053), and σεy,0 (1.11×10−7 ) and
σεy,1 (0.0205). The differences between Φ̄β2 ,0 (0.3972) and Φ̄β2 ,1 (0.8460), Φ̄β3 ,0
(0.0273) and Φ̄β3 ,1 (−0.4893), and Φ̄α3 ,0 (0.0046)and Φ̄α3 ,1 (0.0132) are also
significant. The fact that all the elements of F are smaller than 1 indicates that
the time-varying parameters are stationary and therefore justifies the adoption
of eq. (9).

The paths of α2t are shown in Figure 1A.5 We leave aside the paths of the
intercepts in the IS- and Philips curves. In Figure 1A, “Alpha 2t,0” denotes the
path of α2t when [St = 0|Yt ], namely α2t,0 . “Alpha 2t,1” denotes the path of
α2t when [St = 1|Yt ], namely α2t,1 . “Alpha 2t” denotes the weighted average
of α2t,0 and α2t,1 , α2t . That is,

α2t = P r[St = 0|Yt ]α2t,0 + P r[St = 1|Yt ]α2t,1 .

The paths of α3t are shown in Figure 1B. In Figure 1B, “Alpha 3t,0” denotes the
path of α3t when [St = 0|Yt ] (α3t,0 ). “Alpha 3t,1” denotes the path of α3t when
[St = 1|Yt ] (α3t,1 ). Similarly, “Alpha 3t” denotes the weighted average of α3t,0
and α3t,1 (α3t ). Figure 1C represents the weighted average of the forecast errors
ξt|t−1 . In (13) we find that the conditional variance of the forecast errors consists
4 The IPI has also been used by Clarida, Gali and Gertler (1998) to measure the output for

Germany, France, the U.S., the U.K., Japan and Italy. As surveyed by Orphanides and van
Norden (2002), there are many methods to measure the output gap. We find that filtering
the IPI using the Band-Pass filter developed by Baxter and King (1995) leaves the measure
of the output gap essentially unchanged from the measure with the HP-filter. The Band-Pass
filter has also been used by Sargent (1999).
5 In order to eliminate the effects of the initial startup idiosyncracies of the Kalman filter

algorithm, we present the paths of the variables concerned from t=12 on, namely from 1964.3
to 1999.4.

7
Figure 1: Results of the Linear Time-Varying Phillips Curve

8
 Phillips curve IS curve
Parameter Estimate S.D. Parameter Estimate S.D.
−7
σεπ,0 0.0021 0.0004 σεy,0 1.11 × 10 0.0125
σεπ,1 0.0053 0.0009 σεy,1 0.0205 0.0032
σηα1 2.53×10−8 0.0051 σηβ1 0.0061 0.0005
σηα2 0.0698 0.0113 σηβ2 0.0106 0.1951
σηα3 2.63×10−9 0.0146 σηβ3 0.0420 0.0374
Φ̄α1 ,0 0.0009 0.0015 Φ̄β1 ,0 0.0011 0.0009
Φ̄α1 ,1 0.0057 0.0073 Φ̄β1 ,1 0.0011 0.0020
Φ̄α2 ,0 0.3401 0.1133 Φ̄β2 ,0 0.3972 0.1112
Φ̄α2 ,1 0.2733 0.1014 Φ̄β2 ,1 0.8460 0.1496
Φ̄α3 ,0 0.0046 0.0096 Φ̄β3 ,0 0.0273 0.0518
Φ̄α3 ,1 0.0132 0.0249 Φ̄β3 ,1 −0.4893 0.0762
f α1 0.2584 0.9793 f β1 0.6177 0.0890
f α2 0.6536 0.1095 f β2 0.3645 0.1004
f α3 0.8431 0.2936 f β1 0.2129 0.1429
p 0.9687 0.0339 p 0.8187 0.1012
q 0.9867 0.0139 q 0.9442 0.0269
Likelihood -567.50 Likelihood -458.24

Table 1: Estimates of the Hyperparameters in the Linear Time-Varying IS- and

Phillips Curves

of two distinct terms: The conditional variance due to changing coefficients

(i,j)
Xt Pt|t−1 Xt0 and the conditional variance due to the switching of σε,j
2
. In Figure
(i,j)
1D “Var1” denotes Xt Pt|t−1 Xt0 , “Var2” denotes σε,j
2
and “Var” is the sum of
(i,j)
the two terms, νt|t−1 . When there is no switching in the variance of the forecast
2
errors, σε,j is constant. Figure 1E represents the path of P r[St = 1|Yt ]. The
probability that there is regime switching in the Phillips curve around 1982-83,
1992 and 1994-96 is very high. 1983 seems to be a break point for α3t : Before
1983 it has been quite smooth in state 1 and experienced small changes in state
0, but increased suddenly to a much higher value in 1984 in both state 1 and
state 0. α2t has also experienced some changes in 1983, though not so obviously
as α3t .
The result of the IS curve estimation is demonstrated in Figure 2, which has
a similar interpretation as Figure 1. The paths of β2t and β3t are represented
in Figure 2A and 2B. The forecast errors are represented in Figure 2C. The
conditional variance of the forecast errors are represented in Figure 2D and
Figure 2E is the path of P r[St = 1|Yt ]. From Figure 2E we find that the
probability that there is regime switching in the IS curve around 1970, 1983
and 1992 is very high. From Figure 2A and 2B we find similar evidence. β2t
evolves between 0 and 1.4, with β3t between −0.7 and 0.1.

9
Figure 2: The Results of the Time-Varying IS Curve

10
2.2 Nonlinear Phillips Curve
In the previous subsection we have explored uncertainty in the simple IS and
Phillips curves. The 1990s, however, has seen the development of the literature
on the so-called nonlinear Phillips curve. More specifically, according to this
literature, positive deviations of aggregate demand from potential are more in-
flationary than negative deviations are dis-inflationary.6 Dupasquier and Rick-
etts (1998a) survey several models of the nonlinearity in the Phillips curve. The
five models surveyed are the capacity constraint model, the mis-perception or
signal extraction model, the costly adjustment model, the downward nominal
wage rigidity model and the monopolistically competitive model. As mentioned
by Akerlof (2002), the nonlinearity of the Phillips curve has been an important
issue of macroeconomics. Aguiar and Martins (2002), for example, test three
kinds of nonlinearities (quadratic, hyperbole and exponential) in the Phillips
curve and Okun’s law with the aggregate EURO-area macroeconomic data and
find that the Phillips curve turns out to be linear, but the Okun’s law nonlinear.
Many empirical studies have been undertaken to explore the Phillips-curve non-
linearity. Dupasquier and Ricketts (1998a) explore nonlinearity in the Phillips
curve for Canada and the U.S. and conclude that there is stronger evidence in
favor of nonlinearity for the U.S than for Canada. Other studies on the nonlin-
earity of the Phillips curve include Knoppik (2001), Razzak (1997), Gómez and
Julio (2000), Clements and Sensier (2002), Dupasquier and Ricketts (1998b),
Chadha, Masson and Meredith (1992), Laxton, Meredith and Rose (1995) and
Bean (2000). Monetary policy with a nonlinear Phillips curve has also been ex-
plored, see Schaling (1999), Tambakis (1998) and Flaschel, Gong and Semmler
(2001) for example. Since monetary policy with a linear Phillips curve can be
different from that with a nonlinear Phillips curve, we will explore uncertainty
in a nonlinear Phillips curve below. In the following section we will also take
into account the nonlinearity of the Phillips curve in an optimal control model.
As discussed by Aguiar and Martins (2002), there may be different forms of
nonlinearity in the Phillips curve. In the research below we just follow Schaling
(1999) and assume that the nonlinear form of the output gap in the Phillips
curve reads as7
αyt
f (yt ) = , α > 0, 1 > β ≥ 0, (16)
1 − αβyt
where yt denotes the output gap and the parameter β indexes the curvature of
the curve. When β is very small, the curve approaches a linear relationship.
Assuming α=10 and β=0.99, we present f (yt ) with the U.S. quarterly data in
Figure 3. It is obvious that when the actual output is lower than the poten-
tial output, the curve of f (yt ) is flatter. From the figure we see this function
6 There is, of course, also the other issue, that the central bank may react with interest

rate changes more to inflationary than to deflationary pressures, for the Euro-area case, see
Semmler, Greiner and Zhang (2002b).
7 Note that this function is not continuous with a breaking point at y = 1 . When
t αβ
1 1
yt < αβ , f 00 (yt ) > 0 and if yt > αβ , f 00 (yt ) < 0. In the research below we choose appropriate
values of α and β so that with the U.S. output gap data we have f 00 (yt ) > 0.

11
 Figure 3: An Example of f (yt )

describes very well the idea that positive deviations of aggregate demand from
potential are more inflationary than negative deviations are dis-inflationary.
Substituting f (yt ) for yt in the Phillips curve, we have now

πt = α1t + α2t πt−1 + α3t f (yt−1 ) + επ,t . (17)

Following the same procedure in the previous subsection, we present the results
of the State-Space form of eq. (17) in Table 2 and Figure 4.
In Figure 4 we also observe some structural changes in the coefficients. But
the difference between Figure 4 and Figure 1 is obvious. The structural changes
of the coefficients show up mainly between the second half of the 1970s and the
beginning of the 1990s in Figure 4, while they show up between the second half
of the 1980s and the first half of the 1990s in Figure 1.
Above we have explored model and shock uncertainties in the IS-Phillips
curves with the U.S data. We also have explored whether regime changes have
occurred in the U.S. economy since the 1960s. The results are, to some ex-
tent, consistent with the line of research that maintains that there were regime
changes in the U.S. economy.8 Overall, the uncertainty of parameters and
shocks, and their impact on monetary policy rules suggest exploring monetary
policy rules with learning and robust control.

3 Monetary Policy Rules with Adaptive Learn-

ing
Svensson (1999) and Semmler, Greiner and Zhang (2002b) derive an optimal
monetary policy rule in an optimal control model with a quadratic loss function
8 See Cogley and Sargent (2001, 2002) for example.

12
Figure 4: Results of the Time-Varying nonlinear Phillips Curve

13
 Parameter Estimate S.D.
σεπ,0 0.0073 0.0013
σεπ,1 1.470 ×10−9 0.0024
σηα1 0.003 0.0002
σηα2 1.087 ×10−8 0.0104
σηα3 3.258 ×10−9 0.0013
Φ̄α1 ,0 0.0124 0.0018
Φ̄α1 ,1 0.0015 0.0007
Φ̄α2 ,0 0.152 0.051
Φ̄α2 ,1 0.213 0.069
Φ̄α3 ,0 0.012 0.010
Φ̄α3 ,1 0.0065 0.005
f α1 0.169 0.089
f α2 0.778 0.071
f α3 0.174 0.639
p 0.949 0.021
q 0.746 0.101
Likelihood -582.60

Table 2: Estimates of the Hyperparameters in the Nonlinear Time-Varying

Phillips Curve

and the IS- and Phillips curves as constraints. This optimal monetary policy
rule is similar to the Taylor rule (Taylor 1993 and 1999). Yet, it is found that
the optimal monetary policy rule can be greatly influenced by the parameters
in the state equations. The question arising is, therefore, what is the optimal
monetary policy rule in case some parameters or shocks in an economic model
such as eq. (2) are uncertain? Recently numerous papers have been contributed
to this problem. Svensson (1999), Orphanides and Williams (2002), Tetlow and
von zur Muehlen (2001a), Söderström (1999), and Beck and Wieland (2002), for
example, explore optimal monetary policy rules under the assumption that the
economic agents learn the parameters in the model in a certain manner. One
assumption is that the economic agents may learn the parameters using the
Kalman filter. This assumption has been taken by Tucci (1997) and Beck and
Wieland (2002). Another learning mechanism which is also applied frequently
is recursive least squares (RLS). This kind of learning mechanism has been
applied by Sargent (1999) and Orphanides and Williams (2002). By intuition
we would expect that economic agents reduce uncertainty and therefore improve
economic models by learning with all information available. Of course, there
is the possibility that economic agents do not improve model specification but
seek a monetary policy rule robust to uncertainty. This is what robust control
theory aims at.
In this section we will explore monetary policy rules under uncertainty under
the assumption that the central banks reduce uncertainty by way of learning. As
mentioned above, some researchers, Beck and Wieland (2002) and Orphanides

14
and Williams (2002) for example, have explored this problem. Besides the dif-
ference in the learning algorithm, another difference between Beck and Wieland
(2002) and Orphanides and Williams (2002) is that the former do not consider
role of expectations in the model, while the latter take into account expecta-
tions in the Phillips curve. Unlike Beck and Wieland (2002), Orphanides and
Williams (2002) do not employ an intertemporal framework. They provide a
learning algorithm with constant gain but do not use a discounted loss function.
Moreover, Orphanides and Williams (2002) assume that the government knows
the true model, but the private agents do not know the true model and have
to learn the parameters with the RLS algorithm. In their case the government
and the private agents are treated differently. Sargent (1999) employs both a
learning algorithm as well as a discounted loss function but in a linear-quadratic
(LQ) model. This implies that after one step of learning the learned coefficient is
presumed to remain forever when the LQ problem is solved. In our model, how-
ever, both the central bank and the private agents are learning the parameters,
that is, they are not treated differently.
The difference of our model from that of Beck and Wieland (2002) can be
summarized in three points: First, we consider both linear and nonlinear Phillips
curves. Second, we take into account expectations. This is consistent with the
model of Orphanides and Williams (2002). An important characteristic of New-
Keynesian economics is that current economic behavior depends not only on
the current and past policy but also on the expectations of agents. Third, we
employ the RLS learning algorithm instead of the Kalman filter. In fact, Harvey
(1989) and Sargent (1999) prove that RLS is a specific form of the Kalman filter.
Evans and Honkapohja (2001) analyze expectations and learning mechanisms
in macroeconomics in detail. The difference to Sargent (1999) is that we in
fact can allow for both coefficient drift through learning by RLS and solve a
nonlinear optimal control model using a dynamic programming algorithm.

3.1 RLS Learning in Linear Phillips Curve

Orphanides and Williams (2002) assume that the current inflation rate is not
only affected by the inflation lag but also by inflation expectations. Following
Orphanides and Williams (2002), we assume that the linear Phillips curve takes
the following form:

πt = γ1 πt−1 + γ2 πte + γ3 yt + εt , ε ∼ iid(0, σε2 ), (18)

where πte denotes the agents’ (including the central bank) expected inflation
rate based on the time t information, γ1 , γ2 ∈ (0,1), γ3 > 0 and ε is a serially
uncorrelated innovation. In order to simplify the analysis, we further assume
the IS equation to be deterministic taking the following form:9

yt = −θrt−1 , θ > 0, (19)

9 This is the same as Orphanides and Williams (2002), except that they include a noise in

the equation.

15
where
rt = rrt − r∗ ,
with rrt denoting the real interest rate and r ∗ the equilibrium real rate. Sub-
stituting eq. (19) into (18), we have

πt = γ1 πt−1 + γ2 πte − γ3 θrt−1 + εt , ε ∼ iid(0, σε2 ). (20)

In case of rational expectations, namely, πte = Et−1 πt , we get

Et−1 πt = γ1 πt−1 + γ2 Et−1 πt − γ3 θrt−1 ,

that is,
Et−1 πt = āπt−1 + b̄rt−1 ,
with
γ1
ā = (21)
1 − γ2
γ3 θ
b̄ = − . (22)
1 − γ2
With these results we get the rational expectations equilibrium (REE)

πt = āπt−1 + b̄rt−1 + εt . (23)

Now suppose that the agents believe the inflation rate follows the process

πt = aπt−1 + brt−1 + εt ,

corresponding to the REE, but that a and b are unknown and have to be learned.
Suppose that the agents have data on the economy from periods i = 0, ..., t −
1. Thus the time-(t-1) information set is {πi , ri }t−1 i=0 . Further suppose that
agents estimate a and b by a least squares regression of πi on πi−1 and ri−1 .
The estimates will be updated over time as more information is collected. Let
(at−1 , bt−1 ) denote the estimates through time t-1, the forecast of the inflation
rate is then given by
πte = at−1 πt−1 + bt−1 rt−1 . (24)
The standard least squares formula gives the equations
  t
!−1 t
!
at X
0
X
0
= zi zi z i πi , (25)
bt
i=1 i=1

0
 ri−1 .
where zi = πi−1
at
Defining ct = , we can also compute eq. (25) using the stochastic approx-
bt
imation of the recursive least squares equations

ct = ct−1 + κt Vt−1 zt (πt − zt0 ct−1 ), (26)

Vt = Vt−1 + κt (zt zt0 − Vt−1 ), (27)

16
where ct and Vt denote the coefficient vector and the moment matrix for zt
using data i = 1, ..., t. κt is the gain. To generate the least squares values, the
initial value for the recursion must be set appropriately.10 The gain κt is an
important variable. According to Evans and Honkapohja (2001), the assumption
that κt = t−1 (decreasing gain) together with the condition γ2 < 1 ensures the
convergence
  of ct as t → ∞. That is, as t → ∞, ct → c̄ with probability 1, with
ā
c̄ = and therefore πte → REE.
b̄
As indicated by Sargent (1999) and Evans and Honkapohja (2001), if κt is a
constant, however, there might be difficulties of convergence to the REE. If the
model is non-stochastic and κt sufficiently small , πte converges to REE under the
condition γ2 < 1. However, if the model is stochastic with γ2 < 1, the belief does
not converge to REE, but to an ergodic distribution around it. Here we follow
Orphanides and Williams (2002) and assume that agents are constantly learning
in a changing environment. The assumption of a constant gain indicates that
the agents believe the Phillips curve wanders over time and give larger weights to
the recent observations of the inflation rate than to the earlier ones. Orphanides
and Williams (2002) denote the case of κt = 1t as infinite memory and the case
of constant κt as finite memory. As many papers on monetary policy (Svensson,
1997, 1999 for example) we assume that the central bank pursues a monetary
policy by minimizing a quadratic loss function. The problem reads as
∞
X
M in E0 ρt L(πt , rt ), L(πt , rt ) = (πt − π ∗ )2 , (28)
{rt }0
∞
t=0

subject to eq. (20), (24), (26) and (27). π ∗ is the target inflation rate, which
will be assumed to be zero just for the purpose of simplification.
Note that the difference of our model from that of Sargent (1999) is obvious,
although he also applies the RLS learning algorithm and an optimal control
framework with infinite horizon. Yet, Sargent (1999) constructs his results in
two steps. First, following the RLS with a decreasing or constant gain, the
agents estimate a model of the economy (the Phillips curve) using the latest
available data, updating parameter estimates from period to period. Second,
once the parameter is updated, the government pretends that the updated pa-
rameter will govern the dynamics forever and derives an optimal policy from an
LQ control model. These two steps are repeated over and over. As remarked
by Tetlow and von zur Muehlen (2001b), however, there is a problem in the
approach of Sargent(1999). Sargent’s approach is based on two assumptions:
First, the economy is subject to drift in its structural parameters and second,
notwithstanding this acknowledgement, the policymaker takes the estimated pa-
rameters at each date as the truth and bases policy decisions on these values.
It is easy to see that the second assumption is inconsistent with the first one.
Our model, however, treats the changing parameters as endogenous variables in
10 Assuming Z = (z , ...z )0 is of full rank and letting π k denote π k = (π , ..., π )0 , the
k 1 k 1
Pk 0
initial value ck is given by ck = Zk−1 π k and the initial value Vk is given by Vk = k−1 ki=1 zi zi .

17
a nonlinear optimal control problem. This is similar to the methodology used
by Beck and Wieland (2002).
As mentioned above, if the unknown parameters are adaptively estimated
with RLS with a small and constant gain, they will converge in distributions in
a stochastic model and converge to a point in a non-stochastic model. But in an
optimal control problem such as (28) with nonlinear state equation the model
will not necessarily converge even if the state equations are non-stochastic.
Next we undertake some simulations for the model. Though the return
function is quadratic and the Phillips curve linear, the problem falls outside the
scope of LQ optimal control problems, since some parameters in the Phillips
curve are time-varying and follow a nonlinear path. Therefore the problem can
not be solved analytically and numerical solutions have to be employed. In the
simulations below we resort to the algorithm developed by Grüne (1997), who
applies adaptive instead of uniform grids. A less technical description of this
algorithm can be found in Grüne and Semmler (2002). The simulations are
undertaken for the deterministic case. In order to simplify the simulations, we
assume that at is known to be a constant value equal to ā. Therefore only bt
has to be learned in the model. In this case we have ct = bt and zi = ri−1 .
As mentioned by Beck and Wieland (2002), the reason for focusing exclusively
on incomplete information regarding b is that this parameter is multiplicative
to the decision variable rt and therefore central to the tradeoff between current
control and estimation.

Simulation

In the simulations we assume γ1 = 0.6, γ2 = 0.4, γ3 = 0.5, θ = 0.4, ρ = 0.985

and κt = 0.05. The initial values of πt , bt and Vt are 0.2, −0.6 and 0.04. The
paths of πt , bt , Vt and rt are shown in Figure 5A-D respectively. Figure 5E is the
phase diagram of πt and rt . Neither the state variables nor the control variable
converge. In fact, they fluctuate cyclically. We try the simulations with many
different initial values of the state variables and smaller κt (0.01 for example)
and find that in no case do the variables converge. Similar results are obtained
with different values for γ1 (0.9 and 0.3 for example) and γ2 (0.1 and 0.7 for
example).
With the parameters above, we have ā = 1, b̄ = −0.33, therefore the REE is
πt = πt−1 − 0.33rt−1 + εt . (29)
In the case of RLS learning, however, we have
πt = πt−1 + b̃t rt−1 + εt ,
with
b̃t = γ2 bt−1 − γ3 θ.
The path of b̃t is presented in Figure 6. b̃t evolves at a higher level than b̄.
Simulations are undertaken with different initial values of the state variables
and similar results for b̃t are found.

18
Figure 5: Simulations of RLS Learning (solid) and Benchmark Model (dashed)
with Linear Phillips Curve

Figure 6: Path of b̃t (solid) in the Linear Phillips Curve

19
 If there is perfect knowledge, namely, the agents have rational expectation,
πt can converge to its target value π ∗ (zero here), since the model then becomes
a typical LQ control problem which has converging state and control variables
in a non-stochastic model. We define this case as the benchmark model. The
results of the benchmark model are shown in Figure 5A and 5D (dashed line).
Note that in the benchmark model there is only one state variable, namely
πt with dynamics denoted by (29). In the non-stochastic benchmark model
the optimal monetary policy rule turns out to be rt = 3.00πt and the optimal
trajectory of πt is πt = 0.01πt−1 . From Figure 5A and 5D we observe that as
time goes on πt and rt converge to zero in the benchmark model.

3.2 RLS Learning in Nonlinear Phillips Curve

As mentioned in Section 2, the Phillips curve could be nonlinear. Given the
nonlinearity of the Phillips curve, eq. (18) reads as,

πt = γ1 πt−1 + γ2 πte + γ3 f (yt ) + εt , ε ∼ iid(0, σε2 ), (30)

with f (yt ) given by eq. (16). Substituting eq. (19) into eq. (16), and then (16)
into (30), we get the following nonlinear Phillips curve

πt = γ1 πt−1 + γ2 πte − γ3 g(rt−1 ) + εt , ε ∼ iid(0, σε2 ), (31)

where
αθrt
g(rt ) = .
1 + αβθrt
The REE turns out to be

πt = āπt−1 + b̄g(rt−1 ) + εt , (32)

γ3
where ā is defined in (21) but b̄ is changed to be − 1−γ 2
. The forecast of the
inflation rate is now given by

πte = at−1 πt−1 + bt−1 g(rt−1 ). (33)

The RLS learning mechanism is the same as the case of the linear Phillips curve,
except that zi is now modified as

0
zi = πi−1 g(ri−1 ) .

The optimal control problem (28) now turns out to have constraints (31), (33),
(26) and (27).

Simulation

In the simulations we take the same values for the parameters in the model
as in the previous subsection and assume α = 10 and β = 0.99. The simulations
with the same starting values of the state variables as in the previous subsection

20
Figure 7: RLS Learning with Linear (solid) and Nonlinear Phillips Curves
(dashed)

are presented in Figure 7A-D. Figure 7A represents the path of πt , 7B the path
of bt , 7C the path of Vt and 7D the path of rt . The results of this subsection
(nonlinear Phillips curve) are presented by dashed lines, while the results from
the previous subsection (linear Phillips curve) are indicated by solid lines. 11
We find that the state variables also do not converge in the optimal control
problem with the nonlinear Phillips curve. Similar to the case of the linear
Phillips curve, the state and control variables fluctuate cyclically. Simulations
with many different initial values of state variables were undertaken and in no
case are the state variables found to converge. But the difference between the
simulations with linear and nonlinear Phillips curves is not to ignore. Figure 7
indicates that both πt (Figure 7A) and bt (Figure 7B) evolve at a higher level
in the case of a nonlinear Phillips curve than in the case of a linear one. The
mean and standard deviation of πt , bt , Vt and rt from the two simulations are
shown in Table 3. The S.D. and absolute values of the mean of these variables
are larger in the case of nonlinear Phillips curve than when the Phillips curve
is linear.
Next we show the b̃ in the nonlinear Phillips curve in Figure 8. b̃ in the
nonlinear Phillips curve equals γ2 bt−1 − γ3 . The b̃ and b̄ from the simulations
with the linear Phillips curve are also shown in Figure 8, from which we find
11 In order to see the differences of the simulations clearly, we just present the results from

t=6 on.

21
 πt bt Vt rt
L NL L NL L NL L NL
mean 0.0102 0.0135 0.0181 0.0243 0.0037 0.0049 -0.0101 -0.0135
S.D. 0.0016 0.0022 0.0069 0.0077 0.0060 0.0064 0.0174 0.0190

Table 3: Mean and S.D. of State and Control Variables. (L and NL stand for
linear and nonlinear Phillips curves respectively)

Figure 8: Paths of b̃t and b̄ in Linear and Nonlinear Phillips Curves (NL stands
for nonlinear)

that the b̃ evolves at a higher level than b̄ in both linear and nonlinear Phillips
curves.
Above we have explored optimal monetary policy rules with adaptive learn-
ing. The simulations indicate that the state variables do not converge no matter
whether the linear or nonlinear Phillips curve is employed as constraint in the
optimal control problem. But the state and control variables seem to experi-
ence larger changes in the nonlinear Phillips curve than in the linear one. The
results are different from those of Sargent (1999), since in his model the state
variables should converge in a non-stochastic model, as explored by Evans and
Honkapohja (2001).

22
4 Monetary Policy Rules with the Robust Con-
trol
Facing uncertainties, economic agents can improve their knowledge of economic
models by learning with all information available. This is what has been ex-
plored in Section 3. A disadvantage of the adaptive learning analyzed in the
previous section is that we have considered only parameter uncertainty. Other
uncertainties such as shock uncertainty explored in the first section, may also
exist. Moreover, as studied in some recent literature, there is the possibility
that economic agents resort to a strategy robust to uncertainty instead of learn-
ing. This problem has recently been largely explored with the robust control
theory. Robust control induces the economic agents to seek a strategy for the
“worst case”. The robust control theory assumes that there is some model
misspecfication–not only the uncertainty of the parameters like αt and βt es-
timated in the IS- and Phillips curves in Section 2, but also other kinds of
uncertainties. Therefore, the robust control might deal with more general un-
certainty than the adaptive learning. The robust control is now given more
and more attention in the field of macroeconomics, because the classic optimal
control theory can hardly deal with model misspecification. On the basis of
some earlier papers (see Hansen and Sargent 1999, 2001a, 2001b and 2001c),
Hansen and Sargent (2002) explore robust control in macroeconomics in details.
Cagetti, Hansen, Sargent and Williams (2001) also employ the robust control in
macroeconomics. Svensson (2000) analyzes the idea of robust control in a sim-
pler framework. Giordani and Söderlind (2002), however, extend robust control
by including forward-looking behavior.
In this section we will also explore monetary policy rules using the robust
control. Before starting empirical research we discuss briefly the framework of
robust control, following Hansen and Sargent (2002). Let the one-period loss
function be L(y,u)=−(x0 Qx + u0 Ru), with Q and R both being symmetric and
positive semi-definite matrices. The optimal linear regulator problem without
model misspecification is
∞
X
M ax E0 ρt L(xt , ut ), 0 < ρ < 1, (34)
{ut }t=0
∞
t=0

subject to the so-called approximating model

xt+1 = Axt + But + Cˇ
t+1 , x0 given, (35)
where {ˇ} is an iid Gaussian vector process with mean zero and identity con-
temporaneous covariance matrix. If the decision maker thinks there is some
model misspecification, he will not regard the model above as true but as a
good approximation to another model that he can not specify. To represent
a dynamic misspecification which can not be depicted by ˇ because of its iid
nature, Hansen and Sargent (2002) take a set of models surrounding eq. (35) of
the form (the so-called distorted model)
xt+1 = Axt + But + C(t+1 + ωt+1 ), (36)

23
where {t } is another iid Gaussian process with mean zero and identity covari-
ance matrix and ωt+1 a vector process that can feed back in a general way on
the history of x:
ωt+1 = gt (xt , xt−1 , ...), (37)
where {gt } is a sequence of measurable functions. When eq. (36) generates the
data, the errors ˇ in (35) are distributed as N (ωt+1 , I) rather than as N (0,I). To
express the idea that eq. (35) is a good approximation when eq. (36) generates
the data, Hansen and Sargent (2002) restrain the approximation errors by
∞
X
0
E0 ρt+1 ωt+1 ωt+1 ≤ η0 . (38)
t=0

In order to solve the robust control problem (34) subject to eq. (36) and
(38), Hansen and Sargent (2002) consider two kinds of robust control problems,
the constraint problem and the multiplier problem, which differ in how they
implement the constraint (38). The constraint problem is
∞
X
M ax M in∞ E0 ρt U (xt , ut ), (39)
{ut }t=0 {ωt+1 }t=0
∞
t=0

subject to eq. (36) and (38). Given θ ∈ (θ, +∞) with θ > 0, the multiplier
problem can be presented as
∞
X
0
M ax M in E0 ρt {U (xt , ut ) + ρθωt+1 ωt+1 }, (40)
{ut }∞
t=0 {ωt+1 }t=0
∞
t=0

subject to eq. (36). Hansen and Sargent (2002, ch. 6) prove that under cer-
tain conditions the two problems have the same outcomes. Therefore, solving
one of the two problems is sufficient. The robustness parameter θ reflects the
agents’ preferences of robustness and plays an important role in the problem’s
solution. If θ is +∞, the problem collapses to the traditional optimal con-
trol without model misspecification. In order to find a reasonable value for θ,
Hansen and Sargent (2002, ch. 13) design a detection error probability function
by a likelihood ratio. Consider a fixed sample of observations on the state x t ,
t = 0, ..., T −1, and let Lij be the likelihood of that sample for model j assuming
that model i generates the data, the likelihood ratio is
Lii
ri ≡ log , (41)
Lij
where i 6= j. When model i generates the data, ri should be positive. Define

pA = P rob(mistake|A) = f req(rA ≤ 0),

pB = P rob(mistake|B) = f req(rB ≤ 0).

Thus pA is the frequency of negative log likelihood ratios rA when model A is

true and pB is the frequency of negative log likelihood ratios rB when model

24
B is true. Attach equal prior weights to model A and B, the detection error
probability can be defined as
1
p(θ) = (pA + pB ). (42)
2
When a reasonable value of p(θ) is chosen, a corresponding value of θ can be
determined by inverting the probability function defined in (42). Hansen and
Sargent (2002, ch. 7) find that θ can be defined as the negative inverse value of
the so-called risk-sensitivity parameter σ, that is θ = − σ1 .
Note the interpretation of the detection error probability. As seen above,
it is a statistic concept designed to spell out how difficult it is to tell the ap-
proximating model apart from the distorted one. The larger the detection error
probability, the more difficult to tell the two models apart. In the extreme case,
when it is 0.5 (θ = +∞), the two models are the same. So a central bank can
choose a θ according to how large a detection error probability it wants. If the
detection error probability is very small, that means, if it is quite easy to tell
the two models apart, it does not make much sense to design a robust rule.
As stated by Anderson, Hansen and Sargent (2000), the aim of the detection
error probability is to eliminate models that are easy to tell apart statistically,
since it is not plausible to set the robustness parameter to be so small that we
tailor decisions to be robust against alternatives that can be detected with high
confidence with a limited amount of data. Note that the higher the θ, the lower
the robustness, not the opposite. In the research below we can see that a larger
detection error probability corresponds to a larger θ.

Next we present the solution of the multiplier problem. Define

D(P ) = P + P C(θI − C 0 P C)−1 C 0 P, (43)
0 −1 0
F(Ω) = ρ[R + ρB ΩB] B ΩA, (44)
0 −1 0


T (P ) = Q + ρA P − ρP B(R + ρB P B) B P A. (45)
Let P be the fixed point of iterations on T ◦ D:
P = T ◦ D(P ),
then the solution of the multiplier problem (40) is
u = −F x, (46)
ω = Kx, (47)
with
F = F ◦ D(P ), (48)
−1 −1 0 −1 0
K=θ (I − θ C P C) C P [A − BF ]. (49)
It is obvious that in case θ = +∞, D(P ) = P and the problem collapses into
the traditional LQ problem.

25
 Simulations

With the same U.S. data as in Section 2 we get the following OLS estimates
of the backward-looking IS- and Phillips curves (t-statistics in parentheses) for
1962.1-1999.4:

πt = 0.002 + 1.380 πt−1 − 0.408 πt−2 + 0.214 πt−3 − 0.221 πt−4

(1.961) (17.408) (2.967) (1.570) (2.836)
2
+ 0.045 yt−1 , R = 0.970, (50)
(3.024)

yt = 0.002 + 1.362 yt−1 − 0.498 yt−2 − 0.074 (Rt−1 − πt−1 ), R2 = 0.843.

(1.050) (19.486) (7.083) (1.360)
(51)

Let A11 be the sum of the coefficients of the inflation lags in the Phillips curves
(0.965) and A22 be the sum of the coefficients of the output gap lags in the IS
curve (0.864), we define
     
0.965 0.045 0 πt
A= , B= , xt = .
0.074 0.864 −0.074 yt

The problem to solve turns out to be

∞
X
0
ρt −(πt2 + λyt2 ) + ρθωt+1
 
M ax M in∞ E0 ωt+1
{Rt }t=0 {ωt+1 }t=0
∞
t=0

subject to
xt+1 = Axt + BRt + C(t+1 + ωt+1 ).
With the parameters above and  the starting  values of π0 and y0 both being
0.01 0
0.02, λ = 1, ρ = 0.985 and C = , we present the detection error
0 0.01
probability in Figure 9.12 If we want a detection error probability of about 0.15,
σ = −33, that is θ = 0.03. With θ = 0.03, we get
 

5.291 0.247
F = 10.462 12.117 , K = ,
4.737 × 10−7 5.486 × 10−7

and the value function turns out to be V(π,y) = 16.240 π 2 +1.033y 2 +1.421πy+0.113.
If we want a higher detection error probability, 0.40 for example, σ = −11 (θ =
0.091) and we get
 

1.173 0.055
F = 7.103 11.960 , K = ,
1.072 × 10−7 1.805 × 10−7
2 2
and V(π,y) = 11.134

π +1.022y +0.945πy+0.080. In case θ = +∞, we have
F = 6.438 11.929 and V(π, y) = 10.120π 2 +1.020y 2 +0.850πy+0.073. From
12 T (number of periods) is taken as 150. 5000 simulations are undertaken here.

26
 Figure 9: Detection Error Probability

θ S.D. of πt S.D. of yt S.D. of Rt

0.03 0.038 0.028 0.223
0.09 0.032 0.017 0.186
+∞ 0.030 0.015 0.179

Table 4: Standard Deviations of the State and Control Variables with Different
of θ

the results above we find that the lower the θ is, the higher the coefficients of the
inflation and output gaps in the interest rate. That is, the farther the distorted
model stays away from the approximating one, the stronger the reaction of the
interest rate to the inflation and output gaps. We also find that the lower the
θ is, the higher the parameters in the value function.
We present the paths of the inflation rate deviation, output gap, the nominal
interest rate and the value function with different θ in Figure 10. Figure 10A
presents the paths of the value function. It is obvious that the value function
with θ = 0.03 (namely σ = −33) evolves at a higher level than those with θ
being 0.09 (σ = −11) and +∞ (σ = 0). Figure 10B, 10C and 10D present the
paths of the inflation deviation, output gap and interest rate. We find that the
lower the θ is, the larger the volatility of the state and control variables. The
standard deviations of the state and control variables are shown in Table 4,
which indicates that the standard deviations of the state and control variables
increase if θ decreases and therefore the value function also increases with the
decrease of θ.
Next we come to a special case, namely the case of zero shocks. What do the
state and control variables look like and how can the robustness parameter θ
affect the state variables and the objective function? According to the certainty
equivalence principle, the optimal rules of the robust control with zero shocks
are the same as when there are non-zero shocks. That is, F and K in eq.

27
 Figure 10: Simulation of the Robust Control with π0 = 0.02 and y0 = 0.02

(48) and (49) do not change no matter whether there are shocks or not. The
difference lies in the value function. The simulations for zero shocks and with
the same parameters as the case of non-zero shocks are shown in Figure 11.
Figure 11A, 11B, 11C and 11D represent the paths of the value function, the
state and control variables with different θ. In Figure 11 we find that the state
variables converge to their equilibria zero as time tends to infinity no matter
whether the robustness parameter is small or large. But in case the robustness
parameter is small, the state variables evolve at a higher level and converge more
slowly to zero than when the robustness parameter is large. The value function
also evolves at a higher level in the case of small robustness parameters. It is
interesting that the interest rate also converges to zero as time tends to infinity,
this seems inconsistent with the fact that the nominal interest rate should be
bounded by zero. This would not be surprising if we note that the interest rate
turns out to be a linear function of the inflation and output gaps in the model.
If we consider a long run equilibrium level of interest rate R̄ as in the simple
Taylor rule (Taylor 1993), the interest rate will converge to something around R̄
rather than zero, though the state variables converge to zero. The simulations
tell us that the larger the robustness parameter θ, the lower πt , yt , Rt and the
value functions are, and moreover, the faster the state variables converge to
their equilibria. And in case θ = +∞, the state variables reach their lowest
values and attain the equilibria at the highest speed. This is consistent with
the conclusion from the simulations with non-zero shocks.

28
 Figure 11: Results of the Robust Control with Zero Shocks

5 Monetary Policy Rules under Uncertainty: An

Evaluation
We have explored two strategies of monetary policy making under uncertainty:
Adaptive learning and robust control. The difference of the two strategies is
clear. In the former case the central bank is assumed to improve its knowledge
of economic models by learning from the information available. In the latter
case, however, the central bank accepts model misspecification as a permanent
state of affairs and directs its efforts to designing robust controls, rather than
to using data to improve model specification over time. As mentioned before,
while the adaptive learning considers mainly parameter uncertainty, the robust
control considers not a specific kind of uncertainty and might deal with more
general uncertainty than the adaptive learning does.
The simulations of these two strategies show much difference. In the learning
strategy the state variables do not converge in both the linear and nonlinear
Phillips curves even if the model is non-stochastic. With the robust control,
however, the state variables may not converge in the stochastic model, but
converge in the non-stochastic model.
We should, however, note two problems here. First, the robust control seeks
a monetary policy rule in the so-called “worst case”, which may not occur, and
moreover, how to specify the “worst case” is a problem. The “worst case” and
therefore the robust monetary policy rule are greatly influenced by the robust-
ness parameter θ, as shown in our simulations. How to specify θ is a problem.
As mentioned by Anderson, Hansen and Sargent (2000), if the robustness pa-

29
rameter is too small, it does not make sense to design a policy rule for a model
which is very easy to tell apart. On the other hand, if θ is very large, the
difference between the approximating model and the distorted model is very
small and the robust rule may not be of much help. Although one can choose a
robustness parameter with the help of the detection probability, the uncertainty
is not really eliminated. In contrast to the robust control, the approach of learn-
ing, however, assumes agents reduce uncertainty through a specific algorithm
of learning. The problem of this strategy is that there are many ways to spec-
ify how the parameters are learned. Different learning algorithms may lead to
different policy rules. The learning algorithms discussed, for example, include
the RLS, the Kalman filter, and the stochastic gradient learning. Second, in the
empirical research of robust control we have assumed that the Phillips curve is
linear. This leads to the convergence of the state variables in the non-stochastic
model. The state variables may not converge if the nonlinear Phillips curve is
used instead of the linear one.
Some researchers have even casted doubt on the strategy of robust control.
Chen and Epstein (2000) and Epstein and Schneider (2001), for example, criti-
cize the application of the robust control for time-inconsistency in preferences.
Hansen and Sargent (2001b), therefore, discuss variants in which alternative
representations of the preferences that underlie robust control theory are or are
not time consistent. An important criticism of the robust control comes from
Sims (2001a). He criticizes the robust control approach on conceptual grounds.
The robust control imposes also neutrality properties of the model which will
not be removed by better local approximations. There are major sources of
a more fundamental type of uncertainties that the robust control theory does
not address.13 Sims (2001a) points out that more important uncertainties are
ignored in such an approach. One major uncertainty is to what extent there is
a medium run trade-off between inflation and output. Sims (2001a) shows that
long run effects of inflation on output may not need to be completely permanent
in order to be important. On the other hand, a deflation may have strong desta-
bilizing effects while interest rates are already very low. Thus, there may in fact
be a long-run non-vertical Phillips curve.14 Yet, the robust control approach
follows the neutrality postulate, implying a vertical long-run Phillips curve.

6 Conclusion
This paper is concerned with monetary policy rules under uncertainty. We first
present the evidence of uncertainty using a State-Space model with Markov-
Switching. Our empirical model using the U.S. data indicates that there have
been regime changes in both parameters and shocks. We have considered not
only the traditional IS and linear Phillips curve, but also a nonlinear Phillips
curve.
13 Moreover, steady states might not be optimal, if multiple steady states exist, see Greiner

and Semmler (2002).

14 See Graham and Snower (2002) and Semmler, Greiner and Zhang (2002a, ch. 7).

30
 Based on the evidence of uncertainty in monetary policy, we explore two
kinds of strategies to deal with uncertainty. The first strategy is adaptive learn-
ing of unknown parameters in models. Both linear and nonlinear Phillips curves
are considered. In contrast to previous models with adaptive learning (see Sar-
gent, 1999 and Orphanides and Williams, 2002), where LQ control models have
been used, our simulations with a nonlinear optimal control model indicate that
the state variables do not converge in either case, but the state variables have
larger means and variances in the nonlinear Phillips curve than in the linear
one.
The second strategy we have considered is the robust control which may deal
with more general uncertainty than the adaptive learning. Using the robust
control the central bank resorts to a monetary policy rule robust to uncertainty
instead of learning. The empirical research indicates that the robustness pa-
rameter plays an important role in the policy design. It influences not only the
means and variances of the state and control variables but also the speed of con-
vergence. Yet, as Sims (2001a) has argued, other major sources of uncertainty
should also be considered.

References
[1] Aguiar, Alvaro and Manuel M.F. Martins (2002), “Trend, cycle, and non-
linear trade-off in the Euro-area 1970-2001”, CEMPRE.

[2] Akerlof, George A. (2002), “Behavior macroeconomics and macroeconomic

behavior”, American Economic Review, Vol. 92, pp411-433.

[3] Anderson, Evan W., Lars Peter Hansen and Thomas J. Sargent (2000),
“Robustness, detection and the price of risk”, manuscript.
[4] Baxter, Marianne and Robert G. King (2002), “Measuring business cycles:
Approximate Band-Pass filters for economic time series”, NBER working
paper, No. 5022.

[5] Bean, Charles (2000), “The convex Phillips curve and macroeconomic pol-
icymaking under uncertainty”, manuscript, London School of Economics.

[6] Beck, Günter W. and Volker Wieland (2002), “Learning and control in
a changing economic environment”, Journal of Economics Dynamics and
Control, Vol. 26, pp1359-1377.

[7] Cagetti, Hansen, Sargent and Williams (2001), “Robustness and pricing
with uncertain growth”, manuscript, February.

[8] Chadha, B., P. Masson and G. Meredith (1992), “Models of inflation and
the costs of disinflation”, IMF Staff Papers, Vol. 39, No. 2, pp395-431.

[9] Chen, Z. and L.G. Epstein (2000), “Ambiguity, risk and asset returns in
continuous time”, manuscript.

31
[10] Clarida, Richard, Jordi Gali and Mark Gertler (1998), “Monetary policy
rules in practice: Some international evidence”, European Economic Re-
view, Vol. 42, 1033-1067.
[11] − − − (1999), “The science of monetary policy: A new Keynesian perspec-
tive”, Journal of Economic Literature, XXXVII, 1661-1707.

[12] Clements, Michael P. and Marianne Sensier (2002), “Asymmetric output-

gap effects in Phillips curve and mark-up pricing models: Evidence for the
U.S. and the U.K.”, manuscript.

[13] Cogley, Timothy and Thomas J. Sargent (2001), “Evolving post-world war
II U.S. inflation dynamics”, manuscript, June.

[14] − − − (2002), “Drifts and volatilities: Monetary policies and outcomes in

the post WWII U.S.”, manuscript, March.

[15] Dupasquier, Chantal and Nicholas Ricketts (1998a), “Nonlinearities in the

output-inflation relationship”, manuscript.

[16] Dupasquier, Chantal and Nicholas Ricketts (1998b), “Nonlinearities in the

output-inflation relationship: Some empirical results for Canada”, Bank of
Canada working paper 98-14.

[17] Epstein, L. and M. Schneider (2001), “Recursive multiple priors”,

manuscript.

[18] Evans, George W. and Seppo Honkapohja (2001), “Learning and Expecta-
tions in Macroeconomics”, The Princeton University Press.

[19] Flaschel, Peter, Gang Gong and Willi Semmler (2001), “Nonlinear Phillips
curve and monetary policy in a Keynesian macroeconometric model”, work-
ing paper 18, CEM, University of Bielefeld.

[20] Giannoni, Marc P. (2000), “Does model uncertainty justify caution? Ro-
bust optimal monetary policy in a forward-looking model”, manuscript,
September.

[21] Giordani, Paolo and Paul Söderlind (2002), “Soluation of macromodels

with Hansen-Sargent robust policies: Summary and some extensions”,
manuscript, February.

[22] Gómez, Javier and Juan Manuel Julio (2000), “An estimation of the non-
linear Phillips curve in Colombia”, manuscript.
[23] Graham, Liam and Dennis J. Snower (2002), “The return of the long-run
Phillips curve”, Institute for the Study of Labor (IZA), discussion paper,
No. 646.

32
[24] Greiner, Alfred and Willi Semmler (2002), “Monetary policy, non-
uniqueness of steady states and hysteresis effects”, in Reinhard Neck (ed.),
Modeling and Control of Economic Systems, Elsevier Science Ltd, Oxford.
[25] Grüne, Lars (1997), “An adaptive grid scheme for the discrete Hamilton-
Jacobi-Bellman equation”, Numerische Mathematik, Vol. 75, 319-337.

[26] Grüne, Lars and Willi Semmler (2002), “Using dynamic programming with
adaptive grid scheme for optimal control problems in economics”, working
paper, www.wiwi.uni-bielefeld.de/∼semmler/cem.

[27] Hansen, Lars Peter and Thomas J. Sargent (1999), “Robustness and com-
mitment: A monetary policy example”, manuscript, March.

[28] − − − (2001a), “Robust control and model uncertainty”, manuscript, Jan-

uary.

[29] − − − (2001b), “Time inconsistency of robust control?”, manuscript, Oc-

tober.

[30] − − − (2002), “Robust Control and Filtering for Macroeconomics”, book

manuscript.
[31] Harvey, A. C. (1989), “Time Series Models”, Phillip and Allan Publishers
Limited.

[32] Heinemann, Maik (2000), “Convergence of adaptive learning and expec-

tational stability: The case of multiple rational-expectations equilibria”,
Macroeconomic Dynamics, Vol.4, 263-288.

[33] Isard, Peter (1999), Douglas Laxton and Ann-Charlotte Eliasson, “Simple
monetary policy rules under model uncertainty”, International Tax and
Public Finance, Vol.6, 537-577.

[34] Kasa, Kenneth (1999), “Will the Fed ever learn?”, Journal of Macroeco-
nomics, Vol. 21, No 2, 279-292.

[35] Kiefer, Nicholas and Yaw Nyarko (1989), “Optimal control of a linear pro-
cess with learning”, International Economic Review, Vol. 30, No. 3, 571-87.

[36] Kim, Chang-Jin (1993), “Sources of monetary growth uncertainty and eco-
nomic activity: The time-varying-parameter model with heteroskedastic
disturbances”, Review of Economics and Statistics, Vol. 75, 483-92.

[37] Kim, Chang-Jin and Charles R. Nelson (1999), “State-Space Models with
Regime Switching”, The MIT Press.

[38] Knoppik, Christoph (2001), “The long-run Phillips curve of Akerlof, Dick-
ens and Perry (1996), and the role of limits to downward nominal wage
rigidity”, Dept. of Economics, University of Regensburg, Germany.

33
[39] Laxton, D., G.Rose and D. Tambakis (1997), “The U.S. Phillips curve: The
case for asymmetry”, manuscript.

[40] Ljung, Lennart (2000), “Recursive least squares and accelerated conver-
gence in stochastic approximation schemes”, manuscript, October.

[41] Martin, Ben and Chris Salmon (1999), “Should uncertain monetary policy-
makers do less?”, manuscript, Bank of England.

[42] Meyer, Laurence H., Eric T. Swanson and Volker W. Wieland (2001),
“NAIRU uncertainty and nonlinear policy rules”, manuscript, January.

[43] − − − (2000), “Robust monetary policy under model uncertainty: Incor-

porating rational expectations”, manuscript, November.

[44] Orphanides, Athanasios and Simon van Norden (2002), “The unreliabil-
ity of output gap estimation in real time”, The Review of Economics and
Statistics, 84(4): 569-583.

[45] Orphanides, Athanasios and John C. Williams (2002), “Imperfect knowl-

edge, inflation expectations and monetary policy”, manuscript, November.

[46] Razzak, Weshah (1997), “The inflation-output trade-off: Is the Phillips

curve symmetric? A policy lesson from New Zealand”, Reserve Bank of
New Zealand.

[47] Rudebusch, Glenn D. and Lars E. O. Svensson (1999), “Policy rules for
inflation targeting”, in John B. Taylor, ed., Monetary Policy Rules, The
University of Chicago Press.

[48] Sargent, Thomas J. (1999), “The Conquest of American Inflation”, Prince-

ton University Press.

[49] Schaling, Eric (1999), “The nonlinear Phillips curve and inflation forecast
targeting”, Bank of England.

[50] Semmler, Willi, Alfred Greiner and Wenlang Zhang (2002a), “The Macroe-
conomy and Monetary and Fiscal Policies in the Euro-Area”, book
manuscript, at www.wiwi.uni-bielefeld.de/∼Semmler/cem.

[51] − − − (2002b), “Monetary policy rules in the Euro-area: Was it too tight
in the 1990s?”, Atlantic Economic Journal, Vol. 30, 283-297.

[52] Sims, Christopher A. (2001a), “Pitfalls of a minimax approach to model

uncertainty”, manuscript, Princeton university.

[53] −−− (2001b), “Comment on Sargent and Cogley’s “Evolving U.S. postwar
inflation dynamics”, manuscript, Princeton University.

[54] Sims, Christopher A. and Tao Zha (2002), “Macroeconomic switching”,

manuscript, Princeton University.

34
[55] Söderström, Ulf (1999), “Monetary policy with uncertain parameters”,
manuscript, March.

[56] Svensson, Lars E.O. (1997), “Inflation forecasting targeting: implementing

and monitoring inflation targets”, European Economic Review, Vol. 41,
1111-46.

[57] −−− (1999), “Inflation targeting: Some extensions”, Scandinavian Journal

of Economics, Vol. 101(3), 337-61.

[58] − − − (2000), “Robust control made simple”, manuscript, October.

[59] Tambakis, Demosthenes N. (1998), “Monetary policy with a convex Phillips

curve and asymmetric loss”, working paper 98/21, IMF.

[60] Taylor, John B. (1993), “Discretion versus policy rules in practice”,

Carnegie-Rochester Conference Series on Public Policy, Vol. 39, 1993, 195-
214.

[61] − − − (1999), “Historical analysis of monetary policy rules”, in Taylor,

John B., ed., Monetary Policy Rules, Chicago, IL: University of Chicago
Press.
[62] Tetlow, Robert J. and Peter von zur Muehlen (2001a), “Simplicity versus
optimality: The choice of monetary policy rules when agents must learn”,
Journal of Economic Dynamics and Control, Vol. 25, 245-79.

[63] − − − (2001b), “Avoiding Nash inflation: Bayesian and robust responses

to model uncertainty”, Board of Governors of the Federal Reserve System,
manuscript.

[64] Tucci, Marco, P. (1997), “Adaptive control in the presence of time-vaying

parameters”, Journal of Economic Dynamics and Control, Vol. 22, 39-47.

[65] Wieland, Volker (2000), “Monetary Policy, parameter uncertainty and op-
timal learning”, Journal of Monetary Economics, Vol. 46, 199-228.

35