Journal of International Economics 91 (2013) 18–26

Contents lists available at SciVerse ScienceDirect

Journal of International Economics

journal homepage: www.elsevier.com/locate/jie

On the unstable relationship between exchange rates and

macroeconomic fundamentals☆
Philippe Bacchetta a,b,c,⁎, Eric van Wincoop d,e
a
University of Lausanne, Switzerland
b
Swiss Finance Institute, Switzerland
c
CEPR, United Kingdom
d
University of Virginia, United States
e
NBER, United States

a r t i c l e i n f o a b s t r a c t

Article history: Survey evidence shows that the relationship between the exchange rate and macro fundamentals is perceived
Received 11 May 2010 to be highly unstable. We argue that this unstable relationship naturally develops when structural parameters
Received in revised form 10 April 2013 in the economy are unknown. We show that the reduced form relationship between exchange rates and fun-
Accepted 11 June 2013
damentals is then driven not by the structural parameters themselves, but rather by expectations of these
Available online 21 June 2013
parameters. These expectations can vary significantly over time as a result of perfectly rational “scapegoat”
JEL classification:
effects. These effects can be expected to hold more broadly in macro and finance beyond the application to
F31 exchange rates in this paper.
F37 © 2013 Elsevier B.V. All rights reserved.
F41

Keywords:
Imperfect information
Time-varying parameters
Parameter uncertainty

1. Introduction evidence suggests a time-varying relationship between exchange

rates and macro fundamentals. Fratzscher et al. (2012) confirm this
The weight that foreign exchange traders attach to different macro by regressing changes in exchange rates on the macro fundamentals
fundamentals as drivers of exchange rates fluctuates considerably as well as the fundamentals interacted with the survey weights. The
over time. This was first documented by Cheung and Chinn (2001) latter are statistically significant, implying that the derivative of the
through a survey of U.S. foreign exchange traders. More extensive exchange rate with respect to a fundamental depends on its survey
evidence of these time-varying weights of macro fundamentals was weight. The time-varying ranking of the macro fundamentals there-
recently reported by Fratzscher et al. (2012) based on data from Con- fore implies a time-varying relationship between exchange rates and
sensus Economics for 12 currencies over a period of 9 years. Forty fundamentals.
six foreign exchange market participants from large countries are The main goal of this paper is to show that large and frequent
asked on a monthly basis to “rank the current importance of a range variations in the relationship between the exchange rate and macro
of different factors in determining exchange rate movements.” The fundamentals occur naturally when there is uncertainty about the
rankings, on a scale from 0 to 10 for six key macroeconomic indica- structural parameters in the economy. We show that with parameter
tors, vary significantly over time for all six indicators. The survey uncertainty, the relationship between a forward looking variable like
the exchange rate and macro fundamentals is determined not by the
structural parameters themselves, but rather by the expectations of
☆ We would like to thank two referees, Raf Wouters, Harald Uhlig, Luca Dedola,
structural parameters. Moreover, we show that these expectations can
Charles Engel, and other seminar participants at Breugel, Toulouse, Tilburg, Bocconi,
Dutch Central Bank, ECB, University of Lausanne and University of Zurich for useful vary significantly over time, giving rise to a highly unstable reduced
comments and suggestions. We also thank Toni Beutler for able research assistance. form relationship between exchange rates and fundamentals. This hap-
We gratefully acknowledge the financial support from the Bankard Fund for Political pens even though agents are perfectly rational Bayesian learners.
Economy, the National Science Foundation (grant SES-0649442) (van Wincoop), the In addition to the usual learning process, whereby parameter
National Centre of Competence in Research “Financial Valuation and Risk Management”
(NCCR FINRISK), and the Hong Kong Institute for Monetary Research (Bacchetta).
expectations converge to their actual values, expectations can exhibit
⁎ Corresponding author at: University of Lausanne, Switzerland. large short-term fluctuations. These are due to a mechanism that we
E-mail address: [email protected] (P. Bacchetta). refer to as a “scapegoat” effect. Some information about the nature of

0022-1996/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.jinteco.2013.06.001
 P. Bacchetta, E. van Wincoop / Journal of International Economics 91 (2013) 18–26 19

structural parameters can be derived by analyzing macroeconomic macro fundamentals is not detected by standard parameter instability
data and exchange rates. But these data are also driven by shocks to tests based on regressions of exchange rates on fundamentals.
unobserved fundamentals. Such unobserved fundamentals can generate We show that knowing the time-varying weights improves the
considerable confusion in the short to medium run. When the exchange explanatory power of fundamentals for exchange rates. While this
rate fluctuates as a result of an unobserved macroeconomic shock, it can improvement is not large in a statistical sense, it may be important
be optimal for agents to give more weight to an observed macro funda- in an economic sense since small increases in predictive power can
mental and therefore making it a “scapegoat”.1 For example, if the dollar generate substantial economic gains. This point is illustrated by the
depreciates it may be natural to attribute it to a large current account def- existence of a large carry trade industry that is based on the statisti-
icit, even when the depreciation is unrelated to this deficit. cally small predictive power of interest differentials (in terms of R2s).
When we refer to a variable as becoming a scapegoat, we mean that The next section presents the model. It shows that the relationship
the expectation of the structural parameter associated with that vari- between the exchange rate and fundamentals depends on expecta-
able becomes much larger than the structural parameter itself, so that tions of structural parameters. We then go on to derive how these
the variable temporarily has a large weight in the reduced form ex- expectations are updated with a Kalman filter. Section 3 calibrates
change rate equation. As in the example above, this may be caused the model based on data on interest rates and exchange rates and
by an unrelated change to an unobserved fundamental. More generally, presents numerical results for the relationship between exchange
we will show that even when the true structural parameters are con- rates and fundamentals based on simulations. Section 4 discusses the
stant the expectation of structural parameters is affected by changes empirical relevance of the scapegoat effect, time-varying structural
in both observed and unobserved fundamentals. In that context we parameters and the broader role of the mechanism in macroeconomics
also think of the term scapegoat more broadly as referring to a situation and finance. Section 5 concludes.
where macroeconomic news that is unrelated to changes in structural
parameters nonetheless leads to rational changes in beliefs about struc- 2. A model with unknown parameters
tural parameters. We show that this leads to a time-varying relation-
ship between the exchange rate and fundamentals. The underlying framework is a standard exchange rate model with
In illustrating the importance of such scapegoat effects, and their constant parameters. We first describe the model when parameters are
role in the unstable reduced form relationship between exchange known. Then we show how the exchange rate is affected by parameter ex-
rates and fundamentals, we slightly generalize the “canonical” ex- pectations when parameters are unknown. In deriving parameter expecta-
change rate model. There is a broad class of exchange rate models tions, we show that in addition to a standard learning process, there is a
that can be reduced to a single stochastic difference equation, derived mechanism that we call a scapegoat effect. This mechanism leads to an un-
from an interest rate parity equation and an equation that relates stable relationship between fundamentals and exchange rates.
the interest differential to observed fundamentals. The latter can be
obtained either from monetary policy specifications or money market 2.1. Basic framework
equilibrium in a standard monetary model. We assume that the un-
derlying parameters in this second equation, such as monetary policy We consider the class of fundamental-based exchange rate models
or money demand parameters, or the relationship between policy that can be reduced to a single stochastic difference equation. The
targets and observed fundamentals, are not perfectly known. More- equilibrium value of the exchange rate in these models depends on
over, we assume that some macro fundamental is not observable. the present value of expected future fundamentals. We start with
Our paper is not the first to introduce parameter uncertainty and the usual case of known parameters. We follow Engel and West
Bayesian learning in a standard exchange rate model. Lewis (1989) (2005) and slightly rewrite their Eq. (1):
assumes the existence of a one-time change in the constant term of 2 3 2 3
the money demand equation. Kaminsky (1993) assumes that money X
∞   X
∞
st ¼ ð1−λÞ4F t þ bt þ λ Et F tþj þ btþj 5−λ4ϕt þ λ Et ϕtþ j 5 ð1Þ
j j
growth is equal to a drift term that can switch between two values
j¼1 j¼1
based on a Markov process. In both cases agents learn about the
unknown parameters through Bayesian updating. Tabellini (1988) em-
where st is the log nominal exchange rate (domestic per foreign cur-
phasized that such a framework can lead to increased exchange rate
rency), Et is the expectation of the representative investor, ϕt is the
volatility relative to the case where parameters are known.2 However,
risk premium and 0 b λ b 1. We denote by Ft a linear combination of
these papers do not consider uncertainty about parameters multiplying
observed macro fundamentals: F t ¼ ft′ β where ft = (f1t,f2t, …,fNt)′
fundamentals and the role of unobserved fundamentals. Consequently,
is the vector of N observed macroeconomic fundamentals and β =
the scapegoat effect is not present in this prior literature.
(β1,β2, …,βN)′ is the vector of associated parameters. Finally, bt repre-
To examine the quantitative relevance of the effects we describe,
sents unobserved macro fundamentals.
we calibrate the model to data for 5 industrialized countries, matching
Engel and West (2005) present several models that lead to this equa-
moments related to interest rates and exchange rates and the explana-
tion. A core element of these models is an interest rate parity condition:
tory power of observed fundamentals. We find that the derivative of the
exchange rate with respect to fundamentals can be very volatile when 
Et stþ1 −st ¼ it −it þ ϕt ð2Þ
structural parameters are unknown. However, we show that this insta-
bility in the reduced form relationship between exchange rates and where it and i∗t represent the domestic and foreign nominal one-period
interest rates. The other basic element is an equation that relates the
interest rate differential to the exchange rate and to observed and
1 unobserved fundamentals. This equation represents for example the
In a previous short paper, Bacchetta and van Wincoop (2004), we developed the
idea of such a scapegoat effect in the context of a simple static noisy rational expecta- reduced form of a monetary model or the differential in interest rate
tions model in which some parameters are unknown. However, a static model does rules. This equation can be written in the following form:
not allow us to address the unstable dynamic relationship between exchange rates

and fundamentals and its implications. Apart from the dynamic setup, the model in this it −it ¼ μst −μ ðF t þ bt Þ: ð3Þ
paper also differs in that there is no private information. Scapegoat effects naturally de-
velop as long as there is incomplete information about parameters.
2
There is also a vast literature on learning and exchange rates with deviations from
Combining Eqs. (2) and (3), integrating forward and assuming no
rational expectations (e.g., Goldberg and Frydman, 1996; Gourinchas and Tornell, bubble gives Eq. (1), where λ = 1/(1 + μ). Notice that in practice we
2004, or Lewis and Markiewicz, 2009). know that λ is very close to 1 (e.g., see Engel and West, 2005).
20 P. Bacchetta, E. van Wincoop / Journal of International Economics 91 (2013) 18–26

Eq. (3) is a source of information for the agents. Agents know the Since agents know the value of ft′ β þ bt , for a given expectation
value of the sum Ft + bt as they can observe the exchange rate and Etβ of the parameters we have
the interest rate. When the parameter β are known, agents can
then infer the value of bt. When the parameters β are unknown, Et bt ¼ ft′ β þ bt −f t′ Et β ¼ f t′ ðβ−Et βÞ þ bt : ð9Þ
knowing Ft + bt gives an imperfect signal about β as a result of
the unobserved fundamental bt. However, if the observed and
The exchange rate can then be written as
unobserved fundamental followed the same AR process, this lack
of knowledge of the parameters β would not matter. We would
have Et(Ft + j + bt + j) =ρj(Ft + bt), with ρ as the AR parameter. Ig- st ¼ ft′ ðθβ þ ð1−θÞEt βÞ þ θbt : ð10Þ
noring the risk premium term, the exchange
 rate in Eq. (1) then
would become st ¼ ½ð1−λÞ=ð1−λρÞ f t′ β þ bt . It is only in this The only difference in comparison to Eq. (5) is that the coefficients
knife-edge case, where the processes for the observed and multiplying the fundamentals are not equal to the structural parame-
unobserved fundamentals are exactly identical, that the exchange ters β, but to a weighted average of the structural parameters and of
rate depends on the actual parameters β and not on their the expectations of these parameters. This significantly changes the
expectations. relationship between the exchange rate and fundamentals as the
We therefore assume that ft and bt follow different processes. We weight on the true parameters is small: since λ is close to one, θ is
will assume that the unobserved fundamental bt follows an AR process: close to zero.5 Moreover, as we will see, even though the parameters
themselves are constant, the expectations of parameters can change
b
bt ¼ ρb bt−1 þ εt ð4Þ significantly over time. This affects the impact of fundamentals on
the exchange rate.
where the innovation has mean zero and variance σ2b. Because of exten- The derivative of the exchange rate with respect to fundamentals
sive evidence of a unit root in the nominal exchange rate, as well as in is now
many macro variables driving the exchange rate, we will assume that
the observed fundamentals ft follow a random walk.3 ∂st ∂E β
Assuming for now that the risk premia ϕt are zero, when the pa- ¼ θβn þ ð1−θÞEt βn þ ð1−θÞf t′ t : ð11Þ
∂f nt ∂f nt
rameter β are known these assumptions imply that Eq. (1) becomes

st ¼ ft′ β þ θbt ð5Þ First consider the sum of the first two terms, which is the weighted
average of the parameter βn and its expectation. Again, since θ is close
where to zero almost all the weights is on the expectation rather than the
parameters themselves. As we will see below, changes in fundamen-
1−λ tals also change the expectations of parameters. In the last term of
θ¼ : ð6Þ Eq. (11) the derivative of the expected parameters with respect to
1−ρb λ
the fundamental interacts with the level of fundamentals.
It is immediate in this case that the derivative of the exchange rate
with respect to fundamentals is 2.3. Expectation of parameters

∂st We now turn to the computation of the expected parameters. For

¼ βn : ð7Þ illustrative purposes we describe the case where N = 1, so that there
∂f nt
is only one fundamental. After that we briefly describe how the re-
sults generalize to multiple fundamentals, leaving the derivation to
The derivative of the exchange rate with respect to fundamentals is
Appendix A.
therefore equal to a known constant, which is equal to the structural
In each period agents observe yt = ftβ + bt. In period 1 the pa-
parameter βn.
rameter β is drawn from a distribution with mean β and standard
deviation σ. We assume that f1 = b1 = 0, so that in period 1 itself
2.2. Exchange rate when parameters are unknown
the signal y1 is simply zero and provides no information about the
value of β. Starting in period 2, agents gradually learn through the sig-
We now turn to the case where the parameters β are unknown.4
nals ftβ + bt. It is useful to rewrite the new information that comes
Specifically, assume that during an initial period 1 all parameters are
available at time t, starting at t = 2, as
drawn from a distribution with mean β and standard deviation σ.
Agents can learn over time about the value of the parameters from
b
the observation Eq. (3), which allows them to observe F t þ bt ¼ Δ̃yt ¼ yt −ρb yt−1 ¼ Δ̃f t β þ εt ð12Þ
f t′ β þ bt . Agents use these observations for each period to update their
beliefs about the parameters and bt. where Δ̃f t ¼ f t −ρb f t−1 . Written in this way, the signal Δ̃yt depends
When parameters are unknown, Eq. (1) no longer implies Eq. (5). on the unknown parameter β, multiplied by the fundamental Δ̃f t ,
Instead, we have plus an idiosyncratic noise εbt .
In period 1 the expectation of β is equal to β and the perceived stan-
st ¼ ð1−λÞf t′ β þ λf t′ Et β þ ð1−λÞbt þ λθρb Et bt : ð8Þ dard deviation is equal to σ. After that expectations can be updated
using the Kalman filter. It is useful to first consider the Kalman updating
formulas for uncertainty about the parameter. Let pt be the variance of
3
The key results do not depend on this assumption. They will also hold when the the expectation of β at time t.
observed and unobserved fundamentals are all stationary. We only need to assume
that Ft and bt follow a different process, for example in the form of a different AR
coefficient.
4 5
The only source of imperfect knowledge is about the parameter vector β. We could The only exception is again the limiting case where the process of the observed and
have assumed that μ and ρb are unknown as well, but that complicates the signal ex- unobserved fundamentals become identical, which here implies ρb → 1, so that θ → 1.
traction problem in learning about these parameters as the observed signals then in- But this is only relevant when we are very close to the knife-edge case where these
volve products of unknown parameters. processes are identical.
 P. Bacchetta, E. van Wincoop / Journal of International Economics 91 (2013) 18–26 21

Starting with p1 = σ2, the Kalman updating formulas give (see influenced by β than by β for a long time. Moreover, the expectation
Eq. (24) in Appendix A): of the parameter remains disconnected from its actual value because
of the last term, which reflects the scapegoat effect. It depends on the
pt ¼ pt−1 α t ð13Þ product of current and past values of Δ̃f t and εbt . While the fundamen-
tals and the noise shocks have nothing to do with the value of the
where parameter itself, the scapegoat effect that results from their current
and past products can have a significant effect on the expectation of
σ 2b the parameter.
αt ¼  2 : ð14Þ
Δ̃f t pt−1 þ σ 2b Finally, to determine the last term in Eq. (11) we need to compute
the derivative of Etβ with respect to ft. From Eq. (17) we see that:
Note that 0 b αt b 1 and in fact αt is close to 1. The difference from  
 2 ∂Et β ∂α t ∂zt b
one is second-order as Δ̃ f t pt−1 is fourth order. This means that ¼ ðEt−1 β−βÞ þ Δ̃f t þ zt ε t : ð20Þ
∂f t ∂f nt ∂f t
agents learn slowly. While uncertainty about the parameter declines
over time, this happens at a slow rate. The logic behind this is that in The last term on the right hand side depends on the innovation εbt ,
the signal Δ̃ f t β þ εbt , the parameter β is multiplied by a small first- which is a transitory shock. This term therefore leads to a transitory
order variable Δ̃ f t , especially when ρb is close to one. component in the derivative of the exchange rate with respect to
Using this result we have fundamentals.
Appendix A derives analogous expressions for the case of N funda-
2 mentals. For uncertainty, the extension to the multivariate case gives
pt ¼ g t σ ð15Þ
2
where Pt ¼ Gt σ ð21Þ

t
gt ¼ ∏ αi : ð16Þ where Pt is now the variance of the entire vector β of parameters. Gt is
i¼2 defined in Appendix A, representing cumulative learning in a way
that is analogous to gt for the single fundamental case.
Here gt represents cumulative learning. It is the remaining uncer- The generalization of Eq. (19) to the multivariate case is
tainty about the parameter as a fraction of the initial uncertainty σ2.
As αt is close to 1, gt declines only slowly with time. X
t
−1 b
Next consider the updating formula for the parameter β. Starting Et β ¼ Gt β þ ðI N −Gt Þβ þ Gt Gi Zi Δ̃f i εi ð22Þ
i¼2
with E1 β ¼ β, for t > 1 we have (see Eq. (28) in Appendix A):

b where IN is an identity matrix of size N and Zt is a multivariate gener-

Et β ¼ α t Et−1 β þ ð1−α t Þβ þ zt Δ̃f t εt ð17Þ
alization of zt that is defined in Appendix A. Notice that the scapegoat
where effect is again associated with the interaction between current and
past fundamentals Δ̃f t and noise innovations εbt . In the multivariate
p case the extent to which a fundamental becomes a scapegoat depends
zt ¼  2 t−1 : ð18Þ b
Δ̃f t pt−1 þ σ 2b on the product Δ̃f nt εt for other fundamentals as well, which deter-
mines the plausibility of other fundamentals as scapegoats.
Finally, in analogy to Eq. (20), Appendix A derives an expression
The first two terms on the right hand side of Eq. (17) reflect the of the derivative of Etβ with respect to fundamentals that depends
speed of learning. If the last term is zero, the expectation of β at on Et−1β − β and a term that is proportional to εbt . The latter again
time t is a weighted average of the expectation at time t − 1 and leads to an entirely transitory component in the derivative of the ex-
the actual value of the parameter β. We have already seen that agents change rate with respect to fundamentals.
learn slowly when αt is close to 1. Here this is reflected in a high
weight of the current expectation on the last period's expectation. 3. Numerical analysis
The last term on the right hand side of Eq. (17) is the key to our
results. It reflects what we call a scapegoat effect. It depends on the To evaluate the potential significance of the scapegoat effect, we
product Δ̃f t and εbt . In order to understand this result, recall that calibrate the model to monthly data for exchange rates, interest
through the exchange rate and the interest rate differential agents rates and observed fundamentals. We then compute the derivative
observe in each period Δ̃f t β þ εbt . Assume that εbt > 0 and Δ̃f t > 0. of the exchange rate with respect to fundamentals, as well as the ex-
Agents do not know whether the larger than expected value of the pectation of parameters, during a simulation of the model. We show
signal is a result of a positive innovation εbt or Δ̃f t > 0 combined that both can be very volatile, while the learning process about the
with a larger than previously expected value of β. They will at least structural parameters is slow.
give some weight to the latter by raising their expectation of β. In
this case the fundamental ft becomes the scapegoat when the true 3.1. Calibration
shock is in the noise εbt . Making the fundamental the scapegoat for
the noise shock is perfectly rational. We use data on 5 currencies relative to the U.S. dollar: Swiss Franc,
Integrating Eq. (17) forward, we have British pound, Canadian dollar, Japanese yen and German mark (Euro
since 1999) with monthly data from September 1975 to September
X
t
z b
Et β ¼ g t β þ ð1−g t Þβ þ g t i
Δ̃f i ε i : ð19Þ 2008. The macro fundamentals are the differential of money supply
i¼2
gi growth, industrial production growth and unemployment rate growth
relative to the U.S., the growth in the oil price and the lagged interest
The first two terms are a weighted average of the initial estimate rate differential relative to the U.S. A description of the data can be
β and the actual parameter itself. As learning is slow, gt remains found in Bacchetta et al. (2010). Consistent with the data, we treat
close to 1 for a long time, so that the expectation of β remains more one period in the model as one month.
22 P. Bacchetta, E. van Wincoop / Journal of International Economics 91 (2013) 18–26

Table 1 3.2. Results

Benchmark parameter assumptions.

Parameters 3.2.1. Learning

Before considering a particular simulation, it should be noted that
σb 0.022
ρb 0.9 agents learn very slowly. The uncertainty about parameters, as mea-
σv 0.0284 sured by their standard deviation, is 2.5 in period 1. At the end of
Ψ1 0.1 the 10-year history that we generate, it is on average 2.3. This is the
Ψ2 0.1
uncertainty at the start of the sample of 397 months. At the end of
N 5
σf 0.000252 the sample the standard deviation is on average 1.9. Overall the re-
λ 0.97 duction in risk is therefore small. While we only simulate the model
β 5 over 33 years, we find that it takes 145 years for this uncertainty to
σ 2.5 be reduced by half.

3.2.2. Unstable relationship between exchange rate and fundamentals

Fig. 1 reports the derivative of the exchange rate with respect to
For calibration purposes we re-introduce time-varying risk premia
the 5 fundamentals for a particular simulation of a 397 month period.
in order to match the observed properties of exchange rates. Let vt be
Again the simulation starts with a history of 10 years prior to the
the present discounted value of the risk premium in the last term of
reported results.
Eq. (1):
Fig. 1 shows a highly unstable derivative of the exchange rate with
respect to each of the fundamentals. If the parameters were known,
X
∞
vt ¼
k
λ Et ϕtþk : the derivative would be constant and equal to 5 for each of the
k¼0 parameters. Instead we see substantial fluctuations in the derivative.
There are both large high frequency fluctuations and lower frequency
fluctuations. The weight attached to fundamental 1 is mostly below
To match the observed volatility and autocorrelation of Δst, we as-
5, while that of fundamental 5 is mostly above 5. For the other fun-
sume that vt follows the process
damentals it fluctuates between levels well above and well below
5. The range over which the weights vary is about 3 to 4 for all
v
vtþ1 −vt ¼ ψ1 ðvt −vt−1 Þ−ψ2 vt þ εtþ1 ð23Þ fundamentals.
To provide further insight into what is driving this unstable rela-
tionship between exchange rates and fundamentals, Fig. 2 shows the
where εvt+1 ~ N(0,σ2v ).
expectation of the 5 parameters. Not surprisingly, this expectation
In calibrating the model, we first generate a history of 10 years.
is also very volatile, with both low and high frequency fluctuations.
This is more realistic than to start from time 1 as expectations of
One striking difference between Figs. 1 and 2 is that the very sharp
the parameters, interest rates and exchange rates depend on the his-
high frequency fluctuations in Fig. 1 are absent from Fig. 2. The deriv-
tory of past innovations in observed and unobserved fundamentals.
ative of the exchange rate with respect to fundamentals does not just
As we discuss below, a history of 10 years has little impact on uncer-
depend on the expectation of the parameters, but also on the deriva-
tainty about the unknown parameters. After the 10 year history we
tive of expected parameters with respect to the fundamentals. As we
simulate the model for 397 months (just over 33 years), the same
saw in Eq. (20), the latter has a transitory component that depends
sample as in the data. We repeat this 1000 times and compute the
on εbt . This generates the large high frequency fluctuations in Fig. 1
average moments across these simulations.
that are absent from Fig. 2.
Table 1 reports the parameters adopted for the benchmark param-
Fig. 3 further illustrates how this unstable relationship between
eterization. The first 5 parameters are associated with the process for
the exchange rate and the fundamentals is a result of the uncertainty
bt and vt. These are set to match the average across the 5 currencies of
about the level of the parameters. It shows the average standard
four moments related to exchange rates and interest rates: the stan-
deviation of ∂st/∂fnt and Etβn as a function of the standard deviation
dard deviation of Δst, the standard deviation of it − i∗t , the first-
σ that represents the initial uncertainty about the parameters. The
order autocorrelation of Δst and the first-order autocorrelation of
average is computed over 1000 simulations and over the 5 funda-
it − i∗t . The next two parameters relate to the fundamentals. Consis-
mentals. We vary σ from 0 to 5 (twice the value in the benchmark).
tent with the data, we set the number of fundamentals equal to
As expected, the unstable relationship between the exchange rate and
N = 5. The standard deviation σf of fundamental innovations is set
fundamentals, as measured by the standard deviation of ∂st/∂fnt, is
to match the explanatory power of the fundamentals. This is the average
closely related to the initial uncertainty about the parameters. Except
across the currencies of the R2 of a regression of Δst on the fundamen-
for very low levels of uncertainty, the relationship between the volatil-
tals. It is a very low 0.023 in the data, consistent with the well-known
ity of ∂st/∂fnt and σ is almost linear.
limited explanatory power of macro fundamentals for the exchange
rate. Finally, we set λ equal to 0.97, consistent with values for λ reported
in Engel and West (2005) that are close to 1. 3.2.3. Parameter instability test
We set βn ¼ β ¼ 5 for all parameters. We set the initial uncer- Even though we have documented the very unstable relationship
tainty at σ = 2.5 in the benchmark parameterization.6 This means between exchange rates and fundamentals, this does not mean that
that a 2 standard deviation confidence band lies between 0 and 10. the instability can be easily detected using regressions of exchange
We do not calibrate this uncertainty, but will consider the impact rates on fundamentals. The econometrician who conducts such tests
of changing the degree of uncertainty about the parameter. has far less information than the agents, who know the entire model
and therefore know the time-varying weights documented in Fig. 1.
For example, if one of the reduced-form weights is high one period
and low the next, we would clearly not have sufficient data to detect
6
The important parameter is the ratio σ=β . The levels of σ and β are just a matter of
this based on regressions of exchange rates on fundamentals. The lim-
standardization: e.g., doubling them leads to identical results if we half the volatility of ited explanatory power of macro fundamentals further complicates
fundamentals. the detection of time-varying weights on the fundamentals.
 P. Bacchetta, E. van Wincoop / Journal of International Economics 91 (2013) 18–26 23

8 Variable 1 8 Variable 2 8 Variable 3

7 7 7

6 6 6

5 5 5

4 4 4

3 3 3

2 2 2
0 100 200 300 400 0 100 200 300 400 0 100 200 300 400

8 8
Variable 4 Variable 5
7 7

6 6

5 5

4 4

3 3

2 2
0 100 200 300 400 0 100 200 300 400

Fig. 1. Derivative Δst with respect to Δfnt.

To illustrate this, we conduct a Quandt–Andrews breakpoint test. well below what is needed to reject the null of a constant relationship
The test statistic, due to Andrews (1993), is the average F statistic between the exchange rate and parameters. The 5% critical value is
over multiple breakpoint tests. All breakpoints from 60 months to 6.13 (see Andrews and Ploberger, 1994). When we apply the same
338 months into the sample are considered. The average F statistic test to the data, the average over the 5 currencies for the average
is 1.0 in the model under the benchmark parameterization. This is F statistics is 1.3, which also cannot reject stability and is close to

Variable 1 Variable 2 Variable 3

7 7 7

6 6 6

5 5 5

4 4 4

3 3 3
0 100 200 300 400 0 100 200 300 400 0 100 200 300 400

Variable 4 Variable 5
7 7

6 6

5 5

4 4

3 3
0 100 200 300 400 0 100 200 300 400

Fig. 2. Expectation of parameters.

24 P. Bacchetta, E. van Wincoop / Journal of International Economics 91 (2013) 18–26

Standard Deviation st/ fnt Standard Deviation Et n

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

Fig. 3. Role of parameter uncertainty.

what we find in the model. The results do not change if we consider This industry exists even though the R2 of regressions of Δst on lagged
multiple breakpoints at once.7 interest differentials is on average only about 0.01 for monthly data. As
exchange rates are so hard to predict, small improvements in predic-
3.2.4. Additional predictive power of fundamentals tive power can be important to investors who actively trade based
The explanatory power of fundamentals should be higher when on such predictive power.
their weights are known. This is indeed what Fratzscher et al. (2012)
find when using survey measures of the weights. They regress Δst on 4. Discussion
the fundamentals Δft as well as the survey weights times the funda-
mentals and find the latter to be highly significant. For our purposes In this section we address three issues. First, we consider the empir-
we use ∂st/∂fnt in Eq. (11) as the weights. We use the adjusted R2 as ical relevance of the scapegoat theory. Second, we discuss how results
a measure of explanatory power. are affected when the structural parameters are not just unknown but
The explanatory power of fundamentals is low, both in the data also time-varying. Finally, we discuss broader implications for parame-
and the model. Under the benchmark parameterization the R2 of a ter instability in macroeconomics.
regression of Δst on the five fundamentals is on average 0.023 (see
Table 2). The adjusted R2 is only 0.010. This low predictive power
4.1. Empirical relevance
is behind the well-known Meese–Rogoff puzzle that it is hard to
outperform a pure random walk. We find that adding the interaction
The survey data used by Fratzscher et al. (2012) may not necessar-
term of the fundamentals and the weights ∂st/∂fnt raises the adjusted
ily be the result of scapegoat effects. First, survey respondents could
R2 to 0.014. While this is a 40% increase, it is clearly not impressive as
simply give weights based on the relative volatility that different fun-
the predictive ability remains small. This is also consistent with
damentals recently displayed. Second, time-varying weights could
Table 2, which reports that the standard deviation of the exchange
also result from a known variation in structural parameters. However,
rate is virtually the same for the benchmark model as for the version
Fratzscher et al. (2012) run two types of regressions that are inconsis-
of the model where the parameters are known (σ = 0).
tent with these interpretations. First, they find that, separately from
Fratzscher et al. (2012) report larger adjusted R2s. This is partially
the fundamentals themselves, the interaction between fundamentals
a result of different macro fundamentals, some of which are usually
and survey weights has a highly statistically significant impact on ex-
not included in exchange rate models (e.g., equity flows). We can
change rate changes. If respondents had simply based their answers
give a somewhat more important role to the macro fundamentals
on fundamentals' behavior, this would already be captured by the
by doubling the innovations t b and t f . Doubling σb and σf leaves
fundamentals themselves in the regression. Second, Fratzscher et al.
parameter expectations unchanged. We find that in that case the ad-
(2012) regress the survey weights on an interaction term similar to
justed R2 is 0.036 when regressing on the fundamentals alone and
Δ̃f nt εbt in our model, which captures the scapegoat effect. They mea-
rises to 0.048 when also regressing on the interaction term of funda-
sure the unobserved macro shock using order flow data. They present
mentals and weights ∂st/∂fnt. Tripling the standard deviations gives
strong evidence that the survey weight on a fundamental rises with
respectively 0.066 and 0.087.
Δ̃f nt εbt (Eq. (17)). They interpret this as supporting the scapegoat
While the improvement in predictive power due to known time-
theory developed in this paper.
variation in the weights on the fundamentals is small from a statistical
point of view, it may well be economically important. For example,
Abhyankar et al. (2005) report non-negligible utility gains for inves- 4.2. Time-varying structural parameters
tors who manage a portfolio using exchange rate forecasts based on
the monetary model of exchange rates as opposed to using a random We have not taken a stand on what generates parameter uncer-
walk forecast. Moreover, there is a huge industry that manages FX tainty. One possibility is that structural parameters change slowly
portfolios based on interest differentials, known as the carry trade.

Table 2
7
Rossi (2006) provides some evidence of parameter instability in reduced-form ex- Moments: data and model.
change rate models, but the evidence is not very strong. Of the four models she con-
siders, only two provide some limited evidence of instability. Apart from structural Data Benchmark σ = 2.5 σ=0 σ=5
break tests, she also considers out-of-sample test where parameters are allowed to
Standard deviation Δst in % 2.91 2.91 2.90 2.93
vary gradually over time. The latter includes a random walk time-varying parameter
Corr(Δst, Δst − 1) 0.04 0.04 0.05 0.04
(TVP) model, where the parameters are assumed to follow a random walk process.
Standard deviation it − i⁎t in % 0.22 0.22 0.22 0.22
However, for all currencies, the model remains unable to outperform the random walk
Corr(it − i⁎t , it − 1 − it − ⁎1 ) 0.92 0.90 0.90 0.90
exchange rate forecast, even when parameters are explicitly allowed to change each
R2 regression Δst on fundamentals 0.023 0.023 0.022 0.029
period in the TVP specification.
 P. Bacchetta, E. van Wincoop / Journal of International Economics 91 (2013) 18–26 25

over time but can change significantly over a very long period of identity matrix. Constant underlying parameters also imply that
time. Such a setup is considered in a previous version of this paper, Pt+1|t = Pt|t ≡ Pt. The observation equation is the analog of Eq. (12)
Bacchetta and van Wincoop (2009). This delivers very similar results so that Ht ¼ Δ̃f t , wt = εbt , and R = σ2b.
to those in this paper. The only difference is that even in the very The uncertainty is updated according to Eq. (13.8.7) in Hamilton,
long-run agents will never learn the true values of parameters as which gives
they are chasing a moving target.
Parameter uncertainty can also be generated by infrequent very Pt ¼ At Pt−1 ð24Þ
large jumps in parameters, perhaps because of a revolution, a techno-
logical breakthrough or a fundamental change in the nature of the where
government or monetary policy. Our results suggest that even if such  
2 −1
changes occur only once in a century, the uncertainty about the magni- At ≡ IN −Pt−1 Δ̃f t Δ̃f t′ Pt−1 Δ̃f t þ σ b Δ̃f t′ : ð25Þ
tude of the parameters may remain very large at all times.
In period 1, we start from P1 = σ2IN. It follows that
4.3. Parameter instability in macro and finance
2
Pt ¼ Gt σ ð26Þ
While the focus of this paper is on exchange rates, our explanation
for the unstable reduced-form relationship could apply similarly to where
other forward looking financial or macroeconomic variables. As first
shown by Stock and Watson (1996), and since then by many others, Gt ¼ At At−1 …A2 : ð27Þ
the phenomenon of reduced form parameter instability in macroeco-
nomic data is widespread.8 There has also been great interest in the Applying Eq. (13.8.6) in Hamilton, we have
impact of parameter or model uncertainty on optimal monetary pol-
icy.9 The same is the case for financial data. In a survey, Pastor and Et β ¼ At Et−1 β þ ðIN −At Þβ þ Zt Δ̃f t εt
b
ð28Þ
Veronesi (2009) point out that “parameter uncertainty is ubiquitous
in finance” and “many facts that appear baffling at first sight seem where
less puzzling once we recognize that parameters are uncertain and
 
subject to learning”.10 2 −1
Zt ¼ Pt−1 Δ̃f t′ Pt−1 Δ̃f t þ σ b : ð29Þ
The application to exchange rates in this paper is just one illustra-
tion of how uncertainty about structural parameters can generate sig-
nificant instability in reduced form parameters. More generally, the Starting from E1 β ¼ β, it follows that
mechanism that we have highlighted occurs whenever agents have
difficulty distinguishing between unobserved macro fundamentals and X
t
−1 b
Et β ¼ Gt β þ ðI N −Gt Þβ þ Gt Gi Zi Δ̃f i εi : ð30Þ
unobserved structural parameters. This applies not just to other asset i¼2
prices, such as equity prices, but to macro models in general in which
agents need to learn about the structure of the model or its parameters. We finally compute ∂Etβ/∂fnt. Define ηn as a vector of size N with a
1 in space n and zeros otherwise. Using Eq. (28), we have:
5. Conclusion
 
∂Et β ∂At ∂Zt b
¼ ðEt−1 β−βÞ þ Δ̃f t þ Zt ηn εt : ð31Þ
Survey evidence suggests that the relationship between the ex- ∂f nt ∂f nt ∂f nt
change rate and macro fundamentals is highly unstable. In order to
explain this, we have developed a model where structural parameters Defining the scalar:
are unknown. We have shown that the relationship between a for-
ward looking variable like the exchange rate and macro fundamentals πt ¼ Δ̃f t′ Pt−1 Δ̃f t þ σ b
2
ð32Þ
is determined not by the structural parameters themselves, but rather
by the expectations of these structural parameters. These expectations we can write:
can vary significantly over time due to perfectly rational scapegoat
effects as agents have difficulty distinguishing unobserved funda- Pt−1
Zt ¼ : ð33Þ
mentals and unobserved structural parameters. Such effects can be πt
expected to hold more broadly in macro and finance models beyond
the particular application to exchange rates discussed here. This implies that:

Appendix A. Kalman filter with N fundamentals ∂Zt Δ̃f ′ P η P

¼ −2 t t−12 n t−1 : ð34Þ
∂f nt πt
While in Expectation of parameters section we focused on a single
fundamental, in this Appendix A we describe the expectation of Similarly we can write:
parameters Etβ in the case of N fundamentals. We follow Hamilton
(1994), particularly Section 13.8. Since underlying parameters are Pt−1 Δ̃f t Δ̃f t′
constant, the state equation is trivial and, using Hamilton's notation At ¼ I N − : ð35Þ
πt
(Section 13.2), we have Q = 0, vt = 0, and F = IN, where IN is the
Therefore:
8
Recent contributions include Cogley and Sargent (2005), Del Negro and Otrok (2007),  
or Primiceri (2005). ∂At Pt−1 ηn Δ̃f t′ þ Δ̃f t η ′n Δ̃f ′ P η P Δ̃f Δ̃f ′
9
See for example contributions by Hansen and Sargent (2008), Onatski and Williams ¼− þ 2 t t−1 n 2t−1 t t : ð36Þ
∂f nt πt πt
(2003) or Levin et al. (2006).
10
For example, Cogley (2005) and Piazzesi and Schneider (2007) introduce uncer-
tainty about time-varying parameters to explain the term spread. We can just substitute Eqs. (34) and (36) into Eq. (31).
26 P. Bacchetta, E. van Wincoop / Journal of International Economics 91 (2013) 18–26

References Goldberg, Michael D., Frydman, Roman, 1996. Imperfect knowledge and behaviour in
the foreign exchange market. The Economic Journal 106, 869–893.
Abhyankar, Abhay, Sarno, Lucio, Valente, Giorgio, 2005. Exchange rates and fundamentals: Gourinchas, Pierre-Olivier, Tornell, Aaron, 2004. Exchange rate puzzles and distorted
evidence on the economic value of predictability. Journal of International Economics beliefs. Journal of International Economics 64 (2), 303–333.
66 (2), 325–348. Hamilton, James D., 1994. Time Series Analysis. Princeton University Press, Princeton,
Andrews, Donald W.K., 1993. Tests for parameter instability and structural change with New Jersey.
unknown change point. Econometrica 61 (4), 821–856. Hansen, L., Sargent, T., 2008. Robustness. Princeton University Press.
Andrews, Donald W.K., Ploberger, Werner, 1994. Optimal tests when a nuisance Kaminsky, Graciela, 1993. Is there a peso problem? Evince from the dollar/pound exchange
parameter is present only under the alternative. Econometrica 62 (6), 1383–1414. rate 1976–1987. American Economic Review 83, 450–472.
Bacchetta, Philippe, van Wincoop, Eric, 2004. A scapegoat model of exchange rate Levin, A., Onatski, A., Williams, J., Williams, N., 2006. Monetary policy under uncertainty
determination. American Economic Review, Papers and Proceedings 94, 114–118. in micro-founded macroeconometric models. In: Gertler, M., Rogoff, K. (Eds.), NBER
Bacchetta, Philippe, van Wincoop, Eric, 2009. On the unstable relationship between Macroeconomics Annual 2005. MIT Press, Cambridge, MA.
exchange rates and macroeconomic fundamentals. NBER Working Paper No. Lewis, Karen K., 1989. Can learning affect exchange-rate behavior? The case of the dollar
15008. in the early 1980's. Journal of Monetary Economics 23, 79–100.
Bacchetta, Philippe, van Wincoop, Eric, Beutler, Toni, 2010. Can parameter instability Lewis, Vivien, Markiewicz, Agnieszka, 2009. Model misspecification, learning and the
explain the Meese–Rogoff puzzle? NBER Seminar on International Macroeconomics exchange rate disconnect puzzle. The B.E. Journal of Macroeconomics 9 (1) (Topics,
2009, 61(1), pp. 125–173. Article 13).
Cheung, Yin-Wong, Chinn, Menzie David, 2001. Currency traders and exchange rate Onatski, A., Williams, N., 2003. Modeling model uncertainty. Journal of the European
dynamics: a survey of the US market. Journal of International Money and Finance Economic Association 1, 1022–1087.
20 (4), 439–471. Pastor, Lubos, Veronesi, Pietro, 2009. Learning in financial markets. Annual Review of
Cogley, Timothy, 2005. Changing beliefs and the term structure of interest rates: cross- Financial Economics 1, 361–381.
equation restrictions with drifting parameters. Review of Economic Dynamics 8, Piazzesi, Monika, Schneider, Martin, 2007. Equilibrium yield curves. In: Acemoglu,
420–451. Daron, Rogoff, Kenneth, Woodford, Michael (Eds.), NBER Macroeconomics Annual
Cogley, Timothy, Sargent, Thomas J., 2005. Drifts and volatilities: monetary policies 2006. MIT Press.
and outcomes in the post WWII U.S. Review of Economic Dynamics 82 (2), Primiceri, Giorgio E., 2005. Time varying structural vector autoregressions and mone-
262–302. tary policy. Review of Economic Studies 72, 821–852.
Del Negro, Marco, Otrok, Christopher, 2007. Dynamic Factor Models with Time-varying Rossi, Barbara, 2006. Are exchange rates really random walks? Some evidence robust
Parameters (mimeo). to parameter instability. Macroeconomic Dynamics 10 (1), 20–38.
Engel, Charles, West, Kenneth D., 2005. Exchange rates and fundamentals. Journal of Stock, James H., Watson, Mark W., 1996. Evidence on structural instability in macroeconomic
Political Economy 113, 485–517. time series relations. Journal of Business and Economic Statistics 14, 11–30.
Fratzscher, Marcel, Sarno, Lucio, Zinna, Gabriele, 2012. The scapegoat theory of exchange Tabellini, Guido, 1988. Learning and the volatility of exchange rates. Journal of International
rates: the first tests. ECB Working Paper No. 1418. Money and Finance 7, 243–250.