Journal of International Money and Finance 41 (2014) 46–64

Contents lists available at ScienceDirect

Journal of International Money

and Finance
journal homepage: www.elsevier.com/locate/jimf

Macroeconomic fundamentals and the exchange

rate dynamics: A no-arbitrage macro-finance
approachq
Weiwei Yin a, Junye Li b, *
a
International School of Economics and Management, Capital University of Economics and Business, Beijing
100070, China
b
ESSEC Business School, Paris, France; Singapore 188064, Singapore

a b s t r a c t

JEL classification: In this paper, we propose an arbitrage-free international macro-

F31 finance model that links the exchange rate dynamics to macro-
G12 economic fundamentals. Jointly using data on exchange rates,
E43
yields of zero-coupon bonds, and macroeconomic variables of the
C32
US and the Euro area, we find a close link between macroeconomic
Keywords: fundamentals and the exchange rate dynamics. The model-implied
Exchange rate dynamics monthly exchange rate changes can explain about 57% variation of
Macroeconomic fundamentals the observed data. The macroeconomic innovations can help cap-
Stochastic discount factor ture large variation of exchange rate changes. Robustness checks
Term structure of interest rates
show that the results also hold for other major exchange rates.
Unscented Kalman filter

2013 Elsevier Ltd. All rights reserved.

1. Introduction

The nominal floating exchange rate has often been regarded as an asset price in exchange rate
modeling. According to the standard asset pricing theory, its current price should reflect market’s
expectations concerning present and future economic conditions (Frenkel and Mussa, 1985; Obstfeld
and Rogoff, 1996; Cochrane, 2005). However, a long-standing puzzle in international economics and
finance is the disconnection between exchange rate movements and macroeconomic fundamentals.

q We would like to thank Carlo Favero, Francesco Giavazzi and M. Hashem Pesaran for very helpful advices and comments.
Comments of Lucio Sarno, Donald Robertson, Claudo Tebaldi, Andrew Meldrum, and participants of Bocconi Economics
workshop, ESSEC Business School Finance seminar, Bocconi Alesina seminar, Cambridge Macroeconomic workshop, and CES
Annual Conference 2010 are also appreciated.

* Corresponding author. ESSEC Business School, Paris-Singapore, Singapore 188064. Tel.: þ65 64139462.
E-mail addresses: weiweiyin.phd@gmail.com (W. Yin), li@essec.edu (J. Li).

0261-5606/$ – see front matter
2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.jimonfin.2013.10.004
 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64 47

Since 1970’s, a variety of models have been proposed and tried to link the exchange rate dynamics to
macroeconomic fundamentals. Monetary models (Frenkel, 1976, 1978; Mussa, 1976; Bilson, 1978;
Dornbusch, 1976) state existence of a long-run equilibrium relationship among relative money sup-
plies, relative income levels and the nominal exchange rates. New open economy macroeconomics
models (Obstfeld and Rogoff, 2003) attempt to explain exchange rate movements by incorporating
imperfect competition and nominal rigidities in a general equilibrium open economy. However, these
models have not found empirical evidence on a close relationship between short-run exchange rate
movements and macroeconomic fundamentals (Meese, 1990; Frankel and Rose, 1996; Engel and West,
2005). Furthermore, they fail to capture the volatile foreign exchange risk premium, implied by the
well documented forward premium anomaly in foreign exchange markets (Fama, 1984; Hodrick, 1989;
Backus et al., 1993; Bansal et al., 1995; Bekaert, 1996).
In this paper, we investigate interactions between the exchange rate dynamics and macroeconomic
fundamentals by proposing an arbitrage-free stochastic discount factor model that jointly prices bond
yields and exchange rates. Under a two-country world, the exchange rate of these two economies is
governed by the ratio of their stochastic discount factors, which are modeled by a factor representation
under the no-arbitrage condition. We take outputs, inflations and short-term interest rates as
fundamental macroeconomic factors. Real output growth directly governs the aggregate consumption
of an economy and should be a key element of the stochastic discount factor. Inflation can also enter
into the stochastic discount factor via its dynamic interactions with the real production (Piazzesi and
Schneider, 2006). The short-term interest rate is typically viewed as a macroeconomic variable
reflecting monetary policy (Duffee, 2007). We extend macro-finance term structure models (Ang and
Piazzesi, 2003; Diebold et al. 2006; Ang et al., 2007) to a two-country framework in order to improve
identification of the time-varying market prices of risks. This is important since ignoring risk premia or
assuming constant market prices of risks may mislead to a conclusion that exchange rates are not
linked to macroeconomic fundamentals.
Under the above modeling set-ups, the exchange rate has a nonlinear relation with macroeconomic
fundamentals. In contrast to uncovered interest parity (UIP), our model indicates that the expected
exchange rate changes are determined by both the interest rate differential of two countries and the
foreign exchange risk premium. Moreover, in our model, the unexpected exchange rate changes are
also driven by the fundamental innovations, whose roles are amplified by the time-varying market
prices of economic risks.
Using monthly data of the US and the Euro area (EA) ranging from January, 1999 to December, 2008,
we find a close connection between macroeconomic fundamentals and the exchange rate dynamics.
The model-implied monthly exchange rate changes can explain 57% variation of the observed data. We
also find that both economies are highly interdependent. These findings are in stark contrast to pre-
vious studies using monetary and new open economy macroeconomics models and to a recent study of
Dong (2006), which follows a similar modeling approach to this paper. The former finds that the
models can only explain at most 10% variation of the data (Lubik and Schorfheide, 2005; Engel and
West, 2005), and the latter, by using latent factors and assuming no impact of the foreign country
on the home country, finds that his model can explain about 38% variation of exchange rate changes
between the US dollar and the German Mark.
The time-varying foreign exchange risk premium plays an important role in explaining the forward
premium anomaly and in remedying the failure of uncovered interest parity. For example, if we run a
regression of exchange rate changes only on the interest rate differential, the estimated coefficient is far
away from unity, and the R2 is very tiny. However, if we introduce the risk premium term, as suggested
by our model and Fama (1984), this estimate becomes closer to unity. The parameter estimate of the
risk premium term is positive and highly significant. The R2 is improved dramatically. Note that the risk
premia generated by Brennan and Xia (2006) and Sarno et al. (2012) can fully explain the forward
premium anomaly. However, given that our model is fairly simple and intuitive and we explicitly take
into account macroeconomic variables, the results above are still remarkable.
Macroeconomic fundamentals enter into the exchange rate dynamics through the time-varying
market prices of risks, and their shocks have time-varying effects on exchange rate movements. This
is in contrast to the model with the constant market prices of risks, in which the exchange rate changes
become time-homogeneous, and whose performance is dramatically deteriorated. For example, the
48 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64

exchange rate changes implied from the model with the constant market prices of risks can only
explain 25% variation of the observed data, less than half of the explanation power of our model. By
decomposing the exchange rate changes into three parts: the interest rate differential, the foreign
exchange risk premium, and the macro-shock driving part, we find that the first two components are
much smoother and that a large fraction of variation of exchange rate changes is explained by the last
component. For example, the last component, which is the product of the differential of market prices
of risks and the fundamental innovations, can explain 37% variation of the observed data, taking 76% of
the total explanation power of the model.
To deeply understand the interaction between the exchange rate movements and macroeconomic
fundamentals, we implement robustness checks using the longer time period of exchange rate data
and corresponding yield and macro data on other major exchange rates, i.e., USD/DEM, USD/GBP, and
USD/JPY. We find that our results hold in all these three cases and are quite robust. For example, our
model can explain 52%, 50%, and 30% variations of the above three exchange rates, respectively.
The paper is related to previous studies on international affine term structure models such as
Brennan and Xia (2006), Sarno et al. (2012), Dong (2006) and Anderson et al., (2010), among others.
These studies employ latent-factor models of the term structure of yields and exchange rates. Some of
them are quite successful at explaining the forward premium anomaly. Our paper is different as we
explicitly allow for macroeconomic variables in the model. Furthermore, we do not impose any re-
strictions on interdependence between two economies and use a more flexible approach. As a result,
our study is more economically meaningful. Our work is also related to the work on explaining the
forward bias in the cross-sectional asset pricing perspective (Lustig et al., 2011; Burnside et al., 2011;
Menkhoff et al., 2012). These papers find that the foreign exchange risk premia are able to price the
cross-section of currency returns and to explain the UIP puzzle. Our paper is different in that it focuses
on the time-series aspects of the forward bias and we explicitly use the macro state variables. More-
over, the paper jointly models and prices the term structure of bond yields and exchange rates. Our
study sheds new lights on the relationship between macroeconomic fundamentals and the exchange
rate dynamics. Last, we propose an efficient likelihood-based estimation method by using the un-
scented Kalman filter, which is recently developed in the field of engineering and can efficiently handle
highly nonlinear state-space models (Julier and Ulmann, 1997, 2004).
The rest of paper is organized as follows. Section 2 introduces a no-arbitrage macro-finance
approach in modeling the exchange rate dynamics. Section 3 proposes a likelihood-based estimation
method relying on the unscented Kalman filter. Section 4 presents the data used in the paper, discusses
the empirical results and their economic implications, and implements robustness checks. Section 5
concludes the paper.

2. The modeling framework

Consider a two-country world, a home country and a foreign country, each with its own currency.
Under the absence of arbitrage, the exchange rate between these two currencies is governed by the ratio
of their stochastic discount factors. In this section, we first discuss how to model stochastic discount
factors in subsection 2.1. We then proceed to model the exchange rate dynamics in subsection 2.2.
Subsection 2.3 introduces a two-country affine term structure model for pricing zero-coupon bonds.

2.1. Macroeconomic fundamentals and stochastic discount factors

~ and the short-term interest rate ~r as main

In each country, we take the output gap ge, the inflation p
macroeconomic fundamentals. Putting the home and foreign factors together, we have a state vector Xt
in a two-country open economy,
h i
ðhÞ ðhÞ ðhÞ ðf Þ ðf Þ ðf Þ 0
~ t ; ~rt ; g~t ; p
Xt ¼ g~t ; p ~ t ; ~r t ; (1)

where the home factors are denoted with superscripts (h), and the foreign factors with superscripts (f).
Tildes are used upon factors to distinguish them from the market observed macroeconomic variables,
 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64 49

which are assumed to be collected with measurement errors. We assume that the state Xt determines
the dynamics of two-country economy and follows a Gaussian vector autoregression process,

Xt ¼ m þ FXt1 þ Sεt ; (2)

where m is a constant 6  1 vector, F a constant 6  6 matrix, εt an i.i.d Gaussian white noise N(0,I6), and
S a low-triangular matrix such that SS0 captures the variance-covariance of εt.
In this two-country world, assume that no-arbitrage holds. Then, in each country, there exists at least
one almost surely positive process Mt with M0 ¼ 1 such that the discounted gains process associated
with any admissible trading strategy is a martingale (Harrison and Kreps, 1979). Mt is called the sto-
chastic discount factor (SDF). In a Lucas-type exchange economy (Lucas, 1982), the stochastic discount
factor is also often interpreted as the representative agent’s intertemporal marginal rate of substitution.
ðhÞ ðhÞ ðf Þ
We denote the home SDF as Mt and the forelt ign one as Mt . In the following, whenever a relation
holds for both countries, we suppress the superscript (h) or (f) unless otherwise specified.
For absence of a generally accepted equilibrium model for asset pricing, many studies use flexible
factor models under the no-arbitrage condition (Cochrane, 2005). In this paper, we also use a factor
representation for the SDF’s, based on which the exchange rate and the term structure of interest rates
ðhÞ ðf Þ
are modeled. For each of the home and foreign stochastic discount factors (Mt and Mt ), assume that
it has an exponential form

Mtþ1 ¼ expðmtþ1 Þ
 0 0  (3)
¼ exp  ~r t  12lt lt  lt εtþ1 ;

where ~r t is the short-term interest rate of that country, lt is the time-varying market prices of risks
assigned by investors in that country, and εt is the shock to the state Xt, which is the only common term
ðhÞ ðhÞ ðhÞ ðf Þ ðf Þ ðf Þ
~ p~ ~r ~ ~ ~r
for both the home and foreign SDFs. Correspondingly, we have lt ¼ ðlt ; lt ; lt ; lt ; lt ; lt Þ0 , the
g g p

market prices of risk for each factor in the state vector Xt.
Denote the market prices of risks assigned in the home country as and those assigned in the foreign
ðf Þ
country as lt . We use the state Xt to summarize uncertainties in this two-economy world and assume
that market prices of risks assigned in each country are affine functions of Xt (Dai and Singleton, 2000;
Duffee, 2002)

lt ¼ l0 þ l1 Xt ; (4)
where l0 is a constant 61 vector, and l1 a constant 66 matrix. The specification (4) implies that
investors of each country may assign different market prices for these risks contained in the state Xt if
l0 and l1 are different across these two countries and that if ltðhÞ and lðft Þ comove tightly, the two SDFs
could be highly correlated.

2.2. Exchange rate dynamics and forward premium anomaly

Define the nominal spot exchange rate S t at time t as the domestic currency price of one unit of the
foreign currency. No-arbitrage dictates that the ratio of the stochastic discount factors between the
home and foreign economies determines the dynamics of their exchange rate (Backus et al., 2001;
Bekaert, 1996; Brandt and Santa-Clara, 2002; Brandt et al., 2006). We thus have
ðf Þ
S tþ1 M
¼ tþ1 : (5)
St ðhÞ
Mtþ1

The above relation formally defines the link between the stochastic discount factors of two econ-
omies and exchange rate movements between them. In complete markets, the stochastic discount
factors in both economies are unique, and they uniquely determine the dynamics of their exchange
rate. When markets become incomplete, there may exist many different stochastic discount factors
that can guarantee the absence of arbitrage. In this case, introduction of extra instruments, i.e. zero-
coupon bonds, can help identify market prices of risks.
50 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64

Taking natural logarithms for both sides of Equation (5) and using specifications of the SDF’s (3), we
obtain the following exchange rate dynamic equation

Dstþ1 hstþ1  st ¼ mðftþ1

Þ ðhÞ
 mtþ1
      (6)
ðf Þ 0
0 0
¼ ~r ðhÞ ~ðf Þ þ 12 lðhÞ
t  rt
ðhÞ ðf Þ ðf Þ
t lt  lt lt
ðhÞ
þ lt  lt εtþ1 ;

which shows that macroeconomic fundamentals Xt are imparted to the exchange rate dynamic via
market prices of risk in a nonlinear form and that shocks to output gaps, inflations and interest rates
also drive variation of the exchange rate changes. This is in contrast to the traditional models that often
assume linear relation between the exchange rate and macroeconomic fundamentals or/and that only
use latent factors and do not have this economically meaningful 0 interpretations.
ðhÞ ðf Þ ðhÞ ðhÞ ðf Þ0 ðf Þ
The time-varying conditional mean, mst hð~r t  ~r t Þ þ 12 ðlt lt  lt lt Þ, captures predictable
variation of returns in foreign exchange markets. Equation (6) shows that market prices of risks not
only are important in determining the conditional mean of exchange rate changes, but also directly
ðhÞ ðf Þ
affect the conditional volatility of exchange rate changes through sst hlt  lt . Exposure of the ex-
change rate to macroeconomic innovations is amplified by the difference of the time-varying market
prices of risks between two economies. Therefore, the exchange rate changes are heteroskedastic in
our model.
Equation (6) clearly shows that uncovered interest parity does not hold in our model. Uncovered
interest parity states that the currency with higher interest rate is expected to depreciate against the
one with lower interest rate and thus the expected change of exchange rate is equal to the interest rate
differential of two countries. However, in our model, the expected change of exchange rate is composed
of two parts, the interest rate differential and a term called the foreign exchange risk premium rpt,

1  ðhÞ0 ðhÞ 0
ðf Þ ðf Þ

rpt h lt lt  lt lt : (7)
2
The importance of the time-varying foreign risk premium is also argued by Fama (1984) who points
out that the departure from uncovered interest parity should be attributed to a time-varying risk
premium.

2.3. Term structure of interest rates

Having specified the stochastic discount factors for the home and foreign countries, we can model
the short rates and price zero-coupon bonds. Introduction of bonds in our modeling framework is
important for identifying market prices of risks. Because short rates of the home and foreign countries
have been included in the state Xt as factors, the affine short rate equations can be easily specified as

~r t ¼ d0 þ d01 Xt ; (8)
ðhÞ 0 ðf Þ 0
where d0 ¼ 0, and d1 ¼ ð0; 0; 1; 0; 0; 0Þ ¼ ð0; 0; 0; 0; 0; 1Þ for the foreign
for the home country and d1
country.
No-arbitrage guarantees that a zero-coupon bond with maturity n-year in each country can be
priced at time t by using the following Euler equation

ðnÞ
h ðn1Þ
i
P~ t ¼ Et Mtþ1 P~ tþ1 (9)

with the initial condition ~ ð0Þ

¼ 1. Again, tilde indicates the true value. Under specifications of the
P t
state (2), the short rate (8), and the SDF (3), we can show that the bond price is an exponential linear
function of the state Xt

ðnÞ  
P~ t ¼ exp An þ B0n Xt ; (10)

where An and Bn solve the following difference equations

 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64 51

1
Anþ1 ¼ An þ B0n ðm  Sl0 Þ þ B0n SS0 Bn  d0 ; (11)
2
 X 0
Bnþ1 ¼ F l1 Bn  d1 ; (12)

with A1¼d0 and B1¼d1 being the initial conditions. Accordingly, the yield is also an affine function of
the state

ðnÞ
log Pt
~tðnÞ h 
y ¼ an þ b0n Xt ; (13)
n
where an¼An/n and bn¼Bn/n.
From the difference Equations (11) and (12), we can see that the constant market price of risk
parameter l0 only affects the constant yield coefficient an, whereas the parameter l1 affects the factor
loading bn. This implies that the parameter l0 affects average term spreads and average expected bond
returns, whereas the parameter l1 governs time variation in term spreads and expected bond returns.

3. Econometric methodology

Because we assume that the real macroeconomic factors are unobservable and that the econo-
metrician observed macroeconomic variables are contaminated with measurement errors, we first cast
the model into a state-space representation and then use a filtering approach to estimate the model.
At each period t, we can observe the exchange rate change Dst, the yields of zero-coupon bonds in
ðhÞ ðf Þ
the home and foreign countries (yt and yt ), and the output gaps and inflations of the home and
ðhÞ ðhÞ ðhÞ ðf Þ ðf Þ ðf Þ
foreign countries ¼ (Vt and ðgt ; pt Þ0
¼ ðgt ; pt Þ0 ). We assume that each of these variables is
Vt
collected with the normal i.i.d measurement errors. Thus, we have the following measurement
equations
   0 0   0 0 
Dst ¼ ~rðhÞ ~ðf Þ 1 lðhÞ lðhÞ  lðf Þ lðf Þ
t1  r t1 þ 2 t1 t1 t1 t1 þ
lðhÞ lðf Þ 1 hDs
t1  t1 S ðXt   Xt1 Þ þ t
m F
ðhÞ 0
yðhÞ
yt ¼ aðhÞ þ bðhÞ Xt þ ht ;
ðf Þ 0 ðf Þ
yt ¼ aðf Þ þ bðf Þ Xt þ hyt ;
ðhÞ ðhÞ
Vt ¼ ð I2 024 ÞXt þ hVt ;
ðf Þ ðf Þ
Vt ¼ ð 023 I2 021 ÞXt þ hVt ;
(14)
ðhÞ
where we use εt¼S1(XtmFXt1) in the exchange rate dynamic equation, is a 71 vector con- yt
ðf Þ
taining yields of all maturity considered in the home country, yt is a 51 vector containing yields of all
maturity considered in the foreign country, and ht ’s capture measurement errors with distinct vari-
ances for different variables/series and are assumed to be mutually independent.
We have the latent factor Xt, which follows a first-order VAR with its dynamic (2). From the mea-
surement equations, we note that observations depend on the current and lagged macroeconomic
factors Xt and Xt1, both of which should be taken as states and can be written in the following compact
form
        
Xt m F 066 Xt1 S
¼ þ þ εt : (15)
Xt1 061 I6 066 Xt2 066

Given the state-space model representation (14) and (15) with Gaussian noises, we can implement
model estimation using filtering approaches. We have noted that the exchange rate dynamic equation
is a highly non-linear function of states, which makes the standard Kalman filter inapplicable. Instead,
52 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64

we can use the nonlinear Kalman filters. The usually used nonlinear Kalman filter is the extended
Kalman filter, which linearizes the nonlinear system around the current state estimate using the first-
order Taylor approximation. However, for the highly nonlinear system, the extended Kalman filter is
computationally demanding and performs very poorly. An alternative is the unscented Kalman filter
(UKF), recently developed in the field of engineering (Julier and Ulmann, 1997, 2004). The idea behind
this approach is that in order to estimate the state information after a nonlinear transformation, it is
favorable to approximate the probability distribution directly instead of linearizing the nonlinear
functions. The unscented Kalman filter overcomes to a large extent pitfalls inherent to the extended
Kalman filter and improves estimation accuracy and robustness without increasing computational cost
(see Appendix for the detailed algorithm).
The unscented Kalman filter provides us with the predictive observations, Y b  , and the corre-
t

sponding variance-covariance matrix, PYt . Assuming that the predictive errors are normally distributed,
we can construct the log likelihood function at time t as follows
 
1   1   
b  0 P  Yt  Y
Lt ðQÞ ¼  lnPYt   Yt  Y b ; (16)
t Yt t
2 2
where Q is a vector of model parameters. Parameter estimates can be obtained by maximizing the joint
log likelihood

X
T
b ¼ arg max
Q Lt ðQÞ; (17)
Q˛X t ¼ 1

where X is a compact parameter space, and T is the length of the observations. Because the log like-
lihood function is misspecified for the non-Gaussian model, a robust estimate of variance-covariance of
parameter estimates can be obtained using the approach of White (1982)
h i1
b Q ¼ 1 AB1 A
S ; (18)
T
where
     
b
T v2 Lt Q T vLt Qb vL Q b
1X 1 X t
A ¼  ; B ¼ : (19)
T t ¼ 1 vQvQ0 T t ¼ 1 vQ vQ 0

b , the latent macroeconomic factors X

With these parameter estimates Q b t can be extracted using the
unscented Kalman filter.

4. Empirical results and discussions

4.1. Data

We use monthly data of the United State (US) and the Euro area (EA). The US is taken as the home
country, and the EA as the foreign country. There are three types of data: the macroeconomic data
(outputs and inflations), the yields data, and the exchange rate data. Data range from January 1999 to
December 2008 in monthly frequency, in total, 120 months.
The home and foreign output gaps and inflations are proxied by their own Industrial Production
Index’s and CPI’s, respectively. The US data are downloaded from the Federal Reserve, St. Louis, and the
EA data are downloaded from the European Central Bank (ECB). The raw macroeconomic data are
seasonally adjusted. Inflation rates are measured by the 12-month changes of the log CPI, and output
gaps are constructed by applying the HP filter (Hodrick and Prescott, 1997) to Industrial Production
Index with the smoothing parameter being set to 129,600 (Ravn and Uhlig, 2002).
The yields of US zero-coupon bond data are those with maturity 1, 3, 12, 24, 36, 48, and 60 months.
The 1-month and 3-month rates are directly from the CRSP Fama-Bliss rate file. While the yields of 12-
month and longer maturity are derived from the respective zero-coupon bond prices, which are
 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64 53

ðnÞ ðnÞ
obtained from the CRSP Fama-Bliss discount bond file, by using the formula yt ¼ logðPt =100Þ=n.
ðnÞ ðnÞ
Here Pt is the zero-coupon bond price at time t with maturity n-year, and yt is the corresponding
yield. Because the yields of euro area zero-coupon bond data from the ECB are available only from
December 2006 and are not long enough to implement estimation, they are proxied by German data,
which are obtained from BBK Statistics-Deutsche Bundesbank. The EA yields have maturity 1-month,
3-month, 12-month, 36-month, and 60-month. The exchange rate data are those of end-of-period US-
Dollar/Euro spot exchange rates in each month, downloaded from the Federal Reserve, St. Louis.
The upper panels of Fig. 1 plot the annualized macroeconomic data used in estimation. We can see
that two series of inflation rates are very persistent and output gaps are more volatile. During the
sample period, the US output gap reaches a high level in the middle of 2000, just before the collapse of
the dot-com bubble, and has a dramatic decline during the 2008–2009 financial crisis. Similar
movements can also be observed in the EA output gap. The lower panels of Fig. 1 present the annu-
alized yields of zero-coupon bonds of the US and the EA. The US yields are more volatile than the EA
ones.
We implement robustness checks for other major exchange rates. First, we use longer sample period
of the German mark to proxy the euro such that the sample period starts in January 1985 and ends in
May 2009. Second, we also test our model using the exchange rates of USD/GBP and USD/JPY. Both
exchange rates and macroeconomic data come from the International Financial Statistics (IFS) data-
base, provided by the International Monetary Fund (IMF). The zero-coupon bond data on the US,

Fig. 1. Macroeconomic and yield data. Note: The upper panels of the figure plot the annualized macroeconomic data used in the
estimation in the upper panels. The output gaps are in solid lines and inflation rates in dish lines. The lower panels plot the
annualized yields. The US yields have maturity 1-month, 3-month, 1-year, 2-year, 3-year, 4-year and 5-year, and the EA yields have
maturity 1-month, 3-month, 1-year, 3-year, and 5-year. The left panels are for the US and the right ones for the EA. The unit on the
vertical axis is in percentage.
 54
W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64
Table 1
Parameter estimates of the state dynamics.

m(103) F S(103)
~ðhÞ
g ~ ðhÞ
p ~r ðhÞ ~ðf Þ
g ~ ðf Þ
p ~rðf Þ ~ðhÞ
g ~ ðhÞ
p ~r ðhÞ ~ðf Þ
g ~ ðf Þ
p ~r ðf Þ
~ðhÞ 0.298 (0.94)
g 0.834 (3.49) 0.000 (0.65) 0.003 (0.50) 0.010 (2.19) 0.016 (0.83) 0.002 (0.53) 0.637 (3.77) 0 – 0– 0– 0 – 0 –
~ ðhÞ
p 0.149 (2.50) 0.001 (1.24) 0.918 (4.48) 0.023 (1.81) 0.007 (1.16) 0.001 (1.75) 0.011 (0.38) 0.018 (1.43) 0.142 (1.94) 0– 0– 0 – 0 –
~rðhÞ 0.083 (0.35) 0.048 (2.05) 0.051 (1.98) 0.941 (2.56) 0.043 (1.02) 0.007 (0.61) 0.002 (0.95) 0.101 (1.21) 0.241 (0.44) 0.344 (1.96) 0– 0 – 0 –
~ðf Þ 0.331 (2.14)
g 0.025 (2.03) 0.014 (1.07) 0.008 (1.41) 0.776 (2.28) 0.008 (1.92) 0.036 (0.40) 0.424 (1.99) 0.166 (0.48) 0.014 (1.95) 0.796 0 – 0 –
(2.03)
~ ðf Þ
p 0.369 (0.83) 0.006 (1.15) 0.015 (1.06) 0.018 (1.11) 0.008 (1.65) 0.930 (2.67) 0.002 (2.02) 0.008 (0.81) 0.005 (0.54) 0.041 (2.16) 0.060 0.146 0–
(2.15) (2.27)
~rðf Þ 0.029 (1.85) 0.008 (1.73) 0.011 (0.49) 0.022 (2.07) 0.050 (1.97) 0.095 (2.12) 0.896 (2.84) 0.113 (1.11) 0.080 (1.05) 0.252 (1.23) 0.042 0.192 0.220
(0.95) (3.46) (2.35)

Note: The table reports the parameter estimates for the state dynamics that follow a VAR(1) process. In parentheses, the absolute value of t-ratio of each estimate is reported. The sample
period for estimation is from February 1999 to December 2008 and the data is in monthly frequency.
 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64 55

Fig. 2. Macro factors and market prices of risks. Note: Each panel of the figure plots the extracted macroeconomic factor (dark solid
line) and the corresponding time-varying market price of risk assigned by the US market (dashed line) and the EA market (bold line).

Germany, the UK, and Japan are taken from the International Zero Coupon Yield Curve Dataset used by
Wright (2011).

4.2. Estimation and economic implications

4.2.1. State dynamics and macroeconomic factors

Table 1 References presents the estimates of the state dynamics and corresponding t-ratios (in
brackets). The diagonal elements of the matrix F determine the persistence of the macroeconomic
factors, and the off-diagonal ones in F govern their dynamic interactions. We note that the diagonal
estimates are all larger than 0.77 and highly statistically significant, indicating that macroeconomic
factors are very persistent. In particular, for both the US and the EA, the inflation rate is more persistent
than the output gap. This can also be observed from estimates of S, where the output gap estimates are
larger than the inflation rate estimates and in Fig. 1, where the inflation rate is much smoother than the
output gap. The US output gap is more persistent than the EA one, whereas its inflation rate is slightly
less persistent than that of the EA. The short-term interest rates in both economies are also highly
persistent.
From the off-diagonal estimates, we observe a weak link between output gap and inflation in both
countries. For each country, we find that the short rate responds positively to shocks to output gap and
 56
W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64
Table 2
Parameter estimates of market prices of risks.
ðhÞ
US l0 lðhÞ
1 EA l0
ðfÞ
lðfÞ
1
ðhÞ ðf Þ
~ðhÞ
g ~
p ~ðhÞ
r ~ðf Þ
g ~
p ~ðf Þ
r ~ðhÞ
g ~ ðhÞ
p ~r ðhÞ ~ðf Þ
g ~ ðf Þ
p ~r ðf Þ
~ðhÞ
g 0.361 105.1 0.015 0.001 0.031 0.012 0.006 0.369 95.49 0.004 0.009 0.091 0.012 0.013
(1.65) (6.25) (0.35) (0.86) (2.03) (1.74) (0.83) (1.65) (4.36) (0.60) (1.22) (1.99) (1.05) (1.78)
~ ðhÞ
p 0.205 0.006 30.66 0.013 0.004 0.002 0.005 0.223 0.004 27.10 0.153 0.076 0.005 0.004
(1.06) (0.70) (3.79) (3.05) (0.60) (1.12) (0.72) (2.06) (0.53) (2.83) (1.12) (0.49) (1.76) (0.79)
~r ðhÞ 0.313 0.015 0.047 106.6 0.009 0.018 0.008 0.306 0.011 0.098 107.5 0.050 0.026 0.011
(2.37) (1.07) (1.48) (3.36) (1.20) (1.89) (2.42) (2.03) (2.12) (0.44) (3.82) (0.70) (1.80) (0.75)
~ðf Þ
g 0.211 0.008 0.011 0.078 78.25 0.013 0.009 0.222 0.015 0.035 0.086 84.60 0.017 0.045
(0.69) (0.41) (0.48) (1.65) (4.04) (1.62) (0.45) (0.74) (2.31) (0.65) (1.00) (3.74) (0.50) (1.80)
~ ðf Þ
p 0.360 0.022 0.018 0.022 0.082 46.49 0.007 0.414 0.012 0.101 0.022 0.029 6.511 0.015
(1.63) (0.91) (0.43) (0.45) (1.05) (4.45) (1.99) (1.91) (2.60) (1.34) (0.51) (0.59) (1.91) (0.91)
~r ðf Þ 0.349 0.006 0.004 0.055 0.011 0.002 111.5 0.310 0.012 0.076 0.027 0.008 0.017 124.4
(1.74) (0.42) (0.99) (2.04) (1.87) (1.99) (2.87) (2.04) (1.53) (0.80) (2.04) (1.11) (0.60) (4.48)

Note: The table presents the parameter estimates for affine equation of market prices of risk. In parentheses, the absolute value of t-ratio of each estimate is reported. The sample period for
estimation is from February 1999 to December 2008 and the data is in monthly frequency.
 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64 57

inflation, suggesting that the short-term interest rate increases with both inflation and real output
growth. We split the matrix F into four 33 sub-matrices. The off-diagonal sub-matrices control the
interdependence between two countries. We can see that both economies are mutually dependent
since a number of elements in the off-diagonal sub-matrices are statistically significant. In particular,
the US output gap has a positively significant impact on the EA output gap, and vice versa. This implies
a co-movement of the business cycles in these two countries. This finding is in contrast to previous
studies that assume that the US economy has a leading impact on the foreign economy, but not vice
versa.
The dark solid lines in Fig. 2 plot the macroeconomic factors extracted from the data using the
unscented Kalman filter. The left panels are for the US factors, and the right panels for the EA factors. To
compare with the observed data, the estimated factors evolve similarly to the observed ones, but they
have smaller variations, indicating that the real data are noisy and contaminated by measurement
errors.

4.2.2. Market prices of economic risks

Table 2 References reports the parameter estimates of market prices of macroeconomic risks. Most
ðhÞ ðf Þ ðhÞ ðf Þ
of estimates in l0 and l0 and in l1 and l1 not only have the same signs, but also are very close each
other. This implies that the SDFs of two countries should be highly correlated. Indeed, the correlation
between the model-implied SDFs is as high as 99%. Brandt et al. (2006) show that based on the data
from the volatility of the exchange rate and volatility of the SDFs in asset markets the two SDFs must be
highly correlated across countries. Dong (2006) also finds a very high correlation between the US SDF
and the German SDF. Fig. 2 also plots the market price of risk of each macroeconomic factor assigned by
the home and foreign investors (dashed line and bold line, respectively). We can see that market prices
between these two markets are almost indistinguishable and highly correlated. If we think of risks as
goods, the prices of these goods should be highly equalized in these two markets.
Because the estimate of m in the state dynamics is very small, the estimate of l0 approximately
captures the average market prices on the macroeconomic factors. The matrix l1 measures how the
ðhÞ ðf Þ
market price varies with respect to the risk level. All estimates in l0 and l0 are negative, but some of
ðhÞ ðf Þ
them are not statistically significant. We note that all the diagonal estimates of l1 and l1 are highly
statistically significant and that a number of off-diagonal estimates there are not statistically different
from zero, indicating that in each country, the market price on each macroeconomic factor varies
mainly with its own risk level.
For each country, the diagonal output gap estimates in l1 are positive, whereas those for inflations
and interest rates are negative, indicating that market prices of output gaps become less negative when
the real outputs are high, but market prices of inflation factors and short-rate factors become more
negative when the inflation rates and the short-rates are high. From Fig. 2, we can clearly see these
tendencies. The figure also shows that all the market prices of risks in these two economies are
negative. Consistent to parameter estimates in l1, the market price of each risk is highly correlated with
the corresponding macroeconomic factor. Investors demand higher compensation during the recession
and during the time when inflation is high. When the short interest rate goes higher, its market price
becomes even negatively smaller. Previous studies attribute the upward-sloping mean interest rate
term structure to the negative market price of the interest rate risk (Backus et al., 1998).

Table 3
Standard deviations of measurement errors.

Yields n¼1 n¼3 n¼12 n¼24 n¼36 n¼48 n¼60

sðhÞ
h 5.76 (4.83) 4.30 (3.10) 2.58 (3.33) 0.58 (1.95) 3.13 (3.44) 2.37 (3.08) 4.63 (2.03)
sðfh Þ 2.47 (1.91) 3.27 (3.07) 1.11 (2.02) –– 2.26 (2.16) –– 5.27 (2.06)
Macro sgh
h sph h sgf
h sph f – – sDh s
& exchange rate 3.50 (3.52) 1.19 (1.97) 2.25 (2.56) 0.69 (2.07) –– –– 238 (3.07)

Note: The table reports the parameter estimates of standard deviations of measurement errors (in basis point). In parentheses,
the absolute value of t-ratio of each estimate is reported. The sample period for estimation is from February 1999 to December
2008 and the data is in monthly frequency.
58 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64

4.2.3. Model performance analysis

Table 3 presents estimates of standard deviations of the observation measurement errors. For both
the US yields and the EA yields, their standard deviations of the measurement errors are very small,
ranging from 0.6 basis point to 5.8 basis point. The standard deviations of the measurement errors for
the macroeconomic data are also small and are from 0.7 basis point to 3.5 basis point. These results are
comparable with those obtained by Ang and Piazzesi (2003), Duffee (2007), Dong (2006), among
others, and indicate that the model can effectively capture the term structure of interest rates and the
dynamics of macroeconomic fundamentals.
The standard deviation of the exchange rate measurement errors is large (238 basis point) with
comparison to those of yields and macro variables as shown in Table 3. However, the correlation be-
tween the data and the model-implied exchange rate changes is as high as 76%! If we run a regression
of the data (Ds) on the model-implied values (Dbs ) with a constant, we have the following result

Dst ¼  0:0001 þ 1:23 Dbs t þ et ; (20)

ð0:08Þ ð12:47Þ

where in brackets presents the absolute values of t-ratios. The constant term is very small and not sta-
tistically significant, and the coefficient of Db
s is about 1.2 and highly statistically significant. The resulted R2
of this regression is 57%. Therefore, a reasonable proportion of exchange rate movements can be explained
by our model. In contrast, empirical studies based on monetary models and/or new open economy
macroeconomics models can only explain at most 10% variation of exchange rate changes. For example,
Lubik and Schorfheide (2005) investigate the USD/Euro exchange rate and find that their estimated model
explains 10% variation of the one-quarter exchange rate changes in data. A recent study by Dong (2006),
which follows a similar approach to this paper, finds that 38% variation of the data can be explained.
As comparisons, we also investigate the other two nested models. One assumes that the market
price of risk parameter l1 is diagonal, and the other assumes that it is zero. Table 4 presents explained
variances by the model-implied exchange rate changes and correlations between the observed and
model-implied values. We note that our general model has the largest explained variance and corre-
lation, whereas the constant case has the smallest explained variance and correlation. The likelihood
ratio tests reject these two nested models.

4.2.4. Foreign exchange risk premium and forward premium anomaly

One of the most notable puzzles in foreign exchange markets is the forward premium anomaly,
which finds the tendency for high interest rate currencies to appreciate. Fama (1984) attributes this
departure from uncovered interest parity (UIP) to a time-varying risk premium. Our model also sug-
gests that the expected exchange rate change is equal to the sum of the interest rate differential and the
risk premium. Table 2 shows that the diagonal estimates of l(h) and l(f) are huge, and this results in a
substantial foreign exchange risk premium, which can also been seen in Fig. 3. Here we study an
augmented UIP, which also takes into account the foreign exchange risk premium,
 
Dstþ1 ¼ a0 þ a1 rtðhÞ  rtðf Þ þ a2 rpt þ etþ1 ; (21)

where rpt is the foreign exchange risk premium defined in Equation (7), and etþ1 is a noise term. Using
our data and the estimated risk premium, we obtain an estimate of a1 0.63 with t-ratio 0.17 and an

Table 4
Model performance comparison.

Explained variance Correlation (Ds, Ds^) LR test

Full 57% 76% –
Diagonal 34% 58% 92
Constant 25% 50% 173

Note: The table reports the explained variance of the model-implied value and correlation between the model-implied values
and the observed data for each of four models examined in the paper. The Full refers to our general model where l1 is a full
matrix. The Diagonal refers to the model where l1 is a diagonal matrix. The Constant represents the model where l1 is a zero
matrix. Explained variance is the adjusted R2 of a regression of the observed exchange rate changes on the model-implied values
with a constant.
 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64 59

estimate of a2 0.81 with t-ratio 1.97. The resulted R2 is 0.097. a2 is statistically significant and its value is
not far away from unity, and although a1 is not significant, its value is positive. However, if we impose
zero on a2 and estimate the UIP regression, the estimated a1 is negative and far away from unity (1.88
with t-ratio 0.97) and R2 is only 0.008. The above regressions indicate that the UIP puzzle can be
reasonably alleviated by introducing the foreign exchange risk premium term. Note that completely
focusing on the foreign risk premia, Brennan and Xia (2006) and Sarno et al. (2012) both generate the
foreign risk premia that can explain the forward premium anomaly. However, given that our model is
fairly simple and intuitive and we explicitly take into account macro variables, the results above are
reasonably remarkable.
Fama (1984) argues that the implied risk premium should be negatively correlated with and have
larger variance than the interest rate differential, in order to explain the forward premium anomaly.
They are usually termed as Fama’s conditions. Our model implied risk premium (rpt) does negatively
correlate with the interest rate differential (r(h)r(f)) with a correlation about 46% and have a larger
variance (0.996 vs. 0.02). The top panel of Fig. 3 plots the foreign exchange risk premium and the
interest rate differential. It clearly shows a negative correlation between them and a greater variance of
the risk premium.
The middle panel of Fig. 3 presents the output gap differential (g(f)g(h)). We find that the risk
premium is positively correlated to the output gap differential with a correlation about 31%. This
positive correlation implies that when the foreign output gap is higher than the domestic one, people

Fig. 3. The foreign exchange risk premium. Note: The top panel of the figure plots the foreign exchange risk premium and the
interest rate differential (r(h)r(f)); The middle panel plots the foreign exchange risk premium and the output gap differential
(g(f)g(h)). The bottom panel plots the foreign exchange risk premium and the inflation rate differential (p(f)p(h)). The risk premium
is in dark line and the others are in shallow line.
60 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64

Fig. 4. Components of model-implied exchange rates. Note: The figure plots the model-implied exchange rate changes (bold solid
ðf Þ 0 0 0 0
ðhÞ c 1 b
ðhÞ ðhÞ
b bðf Þ bðf Þ bðhÞ  b ðf Þ
line) and its components: Dbs 1;tþ1 ¼ ðb e
~r t  ðrÞ t Þ, D s 2;tþ1 ¼ 2 ð l t l t  l t l t Þ, and D s 3;tþ1 ¼ ð l t
b b l t Þbε tþ1 . The first
component is in dashed line, the second in dark solid line and the third in dashed line with mark.

in the market anticipate the foreign currency to appreciate while the domestic currency to depreciate.
When one country is in a better economic situation than the other, the market becomes more confident
to that country’s currency and thus people would like to hold it, leading to its currency to appreciate.
The lower panel depicts the risk premium and the inflation rate differential (p(f)p(h)). They also have a
positive correlation (80%). If the current inflation of the foreign country is high, people may expect the
central bank to increase its interest rate in the future. This results in a decreased interest rate differ-
ential and an increased risk premium.

4.2.5. Macroeconomic shocks and the exchange rate dynamics

Previous studies find that exchange rate movements are largely disconnected to macroeconomic
fundamentals. In monetary models and/or new open economy macroeconomic models, the exchange
rate is a linear function of contemporaneous macroeconomic variables. Since the residuals are usually
serially correlated in these models, the estimation is always implemented using the first-order dif-
ferences of relevant variables

Dst ¼ b0 þ bðhÞ
1
DrtðhÞ þ bðf1 Þ Drtðf Þ þ b2ðhÞ DgtðhÞ þ bðf2 Þ Dgtðf Þ þ bðhÞ
3
Dpt þ bðf3 Þ Dpðft Þ þ vt ; (22)
ðhÞ ðf Þ
where vt is a noise term. In these models, coefficients are typically constrained by bi ¼ bi , for
i¼1,2,3. When estimating this linear model on the data used in this paper, we find a R2 of 4.7% for the
unconstrained regression and a R2 of 2.8% for the constrained regression. Even though macroeconomic
factors in our model can account for 57% variation of exchange rate movements, the linear model (22)
cannot capture this link between macroeconomic factors and exchange rates.
What exact roles do macroeconomic fundamentals play in our model? We rewrite the exchange
rate dynamic Equation (6) as follows
   0 0   0 0 
Dbs tþ1 ¼ br ðhÞ bðf Þ þ 12 lðhÞ
t  rt
ðhÞ ðf Þ ðf Þ
t lt  lt lt
ðhÞ ðf Þ
þ lt  lt bε tþ1
(23)
hDbs 1;tþ1 þ Dbs 2;tþ1 þ Dbs 3;tþ1 ;
 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64 61

ðhÞ ðf Þ
where Db
s 1;tþ1 ¼ b
rt  b
r t is the estimated differential of short term interest rates between the US and
ðhÞ0 ðhÞ ðf Þ0 ðf Þ
the EA, Dbs 2;tþ1 ¼ 1 ðb l t bl t  bl t bl t Þ is the estimated foreign exchange rate risk premium, and
2
0
ðhÞ ðf Þ0
Dbs 3;tþ1 ¼ ðbl t  b
l t Þbε tþ1 is the estimated unexpected exchange rate changes related to macroeco-
nomic shocks.
Fig. 4 presents these three components of the exchange rate changes. The first component (Dbs 1;tþ1 )
is very smooth. The second one (Dbs 2;tþ1 ) becomes volatile in comparison to the first one, but it still has
much smaller variation than the model-implied exchange rate changes. This implies that the third
component (Dbs 3;tþ1 ) must be more volatile and should play more important role in explaining ex-
change rate movements. This is true from the figure that Dbs 3;tþ1 is very volatile and mimics fluctua-
tions of exchange rate changes. The regression of the data on the unexpected exchange rate changes
(Dbs 3;tþ1 ) and a constant results in a R2 of 37%, taking 76% of the total explained variance.
The unexpected
ðhÞ0 ðf Þ0
exchange rate change Dbs 3;tþ1 is a product of the differential of market prices of risks
(b
l t  bl t ) and the macro innovations (bε tþ1 ), both of which are macro-dependent. When we regress
the data on the model-implied macroeconomic innovations b ε tþ1 with a constant, the R2 is 15%.
However, in VAR approach, macroeconomic shocks on exchange rate changes are time-homogeneous.
In our model, the role of the macro innovations is further amplified by the time-varying differential of
market prices of risks, and hence the exchange rate dynamic is heteroskedastic. The importance of
heteroskedasticity can also be see from the third row in Table 4, where by setting the market prices of
risks constant, only 25% variation of the data can be explained. Macroeconomic innovations are always
regarded as “news” to macroeconomic fundamentals. Their importance has also been investigated by
Engel et al. (2007) and Andersen et al. (2007).

4.3. Robustness checks

Because of restrictions of data availability, the EA data only span 120 months in the previous
analysis. We may be concerned about the results above being specific to this short sample of data. In
this subsection, we implement robustness checks. First, by proxing the USD/EURO exchange rate data
with the USD/GEM, we extend the sample period ranging from January 1985 to May 2009, in total, 293
observations. Second, we also test our model using the exchange rates of USD/GBP and USD/JPY.
Using the same econometric method described in Section 3, the robustness checks support the
previous findings. We find the full model performs much better than the other two restricted nested
models in all three cases. The economic implications of parameter estimates are very similar to those
obtained previously. As for the model performance, we find that the standard deviations of mea-
surement errors for the macro variables and yields are quite small, indicating that the full model can
effectively capture the term structure of interest rates and the dynamics of macroeconomic
fundamentals.
When we move to the measurement errors for exchange rates, we find that the standard deviation
is 309 basis point (bp) for the USD/DEM, 307 bp for the US–UK, and 307 bp for the US–Japan. These
values are relatively large with comparison to measurement errors of macro variables and yields.
However, if we run regressions of the observed values of exchange rates on the model-implied values,
we obtain the following results for the USD/DEM, the USD/GBP, and the USD/JPY, respectively,

Ds1;t ¼ 0:0003 þ 1:17 Dbs 1;t þ e1;t ; R21 ¼ 52%; (24)

ð0:26Þ ð17:83Þ

Ds2;t ¼ 0:0012 þ 0:96 Dbs 2;t þ e2;t ; R22 ¼ 50%; (25)

ð0:88Þ ð17:05Þ

Ds3;t ¼ 0:0002 þ 0:74 Dbs 3;t þ e3;t ; R23 ¼ 33%; (26)

ð0:14Þ ð11:93Þ

where t-ratios are presented in brackets. All constants are quite small and not statistically significant.
The coefficients of Dbs t are very close to one and highly statistically significant. The resulted R2 is the
62 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64

largest for the US–Germany case (52%), whereas it is the smallest for the US–Japan case (33%). All these
values are much larger than those obtained using monetary models and new open economy models.
The model implied foreign exchange risk premia from the above three cases satisfy the Fama (1984)
conditions, which indicate validity of the risk premia in explaining the forward premium anomaly.
First, the correlations between the foreign risk premia and the interest rate differentials are all
negative. They are 21%, 12%, and 28% for the US–Germany, the US–UK, and the US–Japan,
respectively. Second, the variances of the foreign risk premia are much larger (0.65, 1.96, 1.21) than
those of the interest rate differentials (0.02, 0.03, 0.01).

5. Conclusion

This paper investigates relationship between the short-run nominal exchange rates changes
and macroeconomic fundamentals by adopting a no-arbitrage macro-finance approach under a
two-country framework, where macroeconomic information enters into the exchange rate dy-
namics through different channels in a non-linear form. Based on empirical analysis using an
enriched dataset including exchange rates, yields of zero-coupon bonds, and macroeconomic
variables of the US and the Euro area, the paper finds a close link between macroeconomic fun-
damentals and exchange rate dynamics. The model-implied exchange rate changes can explain
about 57% variation of the observed data. This is in stark contrast to previous studies using
monetary and new open economy macroeconomics models, which can explain only around 10%
variation of exchange rates. Having been amplified by the time-varying market prices of risks, the
innovations of macroeconomic fundamentals are the driving engine for generating large volatility
of exchange rate changes. Similar results hold when using longer sample period and other major
exchange rates.
The model in this paper has a fairly good fit to monthly exchange rate changes. However, there is
still nearly 40% variation that cannot be explained. This is because there may be other missing factors
such as current account (Hooper and Morton, 1978), market incompleteness (Brandt and Santa-Clara,
2002), government deficit, “news” from the stock market, default risk and so on. Therefore, it would be
interesting to investigate these factors and to see whether they can help explain the exchange rate
dynamics in the future.

Appendix. The unscented Kalman filter

To implement the unscented Kalman filter, we firstly concatenate the state variables
0
xt1¼[Xt1,Xt2] , the observation noise ht1, and the state noise εt1 at time t1

xet1 ¼ x0t1 h0t1 ε0t1 0 ; (27)

whose dimension is L¼LxþLhþLε and whose mean and covariance are

2
x 3
Pt1 0 0
b e
x t1 ¼ ½ E½xt1  0 0 0 ; Pt1
e
¼ 4 0 S2h 0 5:
0 0 I6
We then form a set of 2Lþ1 sigma points
 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
cet1 ¼  bx et1 bx et1 þ ðL þ lÞPt1 e b e
x t1  ðL þ lÞPt1 e  (28)

and the corresponding weights

l l  
ðmÞ ðcÞ
w0 ¼ ; w0 ¼ þ 1  a2 þ b ; (29)
Lþl Lþl

ðmÞ ðcÞ 1
wi ¼ wi ¼ ; i ¼ 1; 2; .; 2L; (30)
2ðL þ lÞ
 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64 63

where superscripts (m) and (c) indicate that weights are for construction of the posterior mean and
covariance, respectively, l ¼ a2 ðL þ kÞ  L is a scaling parameter, the constant a determines the spread
of sigma points around x and is usually set to be a small positive value, k is a second scaling parameter
with value set to 0 or 3L, and b is a covariance correction parameter and is used to incorporate prior
knowledge of the distribution of x.
With these sigma points, we implement the UKF as follows: for the time update

  
X
2L
ðmÞ x
cxtjt1 ¼ F cxt1 ; cεt1 ; b
xt ¼ wi ci;tjt1 ;
i¼0

X
2L   0
Pxt ¼ wi
ðcÞ
cxi;tjt1  bx 
t cxi;tjt1  bx 
t ;
i¼0

and for the measurement update

   X
2L
h b ¼
Y tjt1 ¼ H cxtjt1 ; ctjt1 ; Y
ðmÞ
wi Y i;tjt1 ;
t
i¼0

X
2L   0
PYt ¼ wi
ðcÞ b
Y i;tjt1  Y b ;
Y i;tjt1  Y
t t
i¼0

X
2L   0
Pxt Yt ¼ wi
ðcÞ
cxi;tjt1  bx  b ;
Y i;tjt1  Y
t t
i¼0

 
b  b ;
x t þ Pxt Yt ðPYt Þ1 Yt  Y
xt ¼ b t

Pxt ¼ Pxt  ðPxt Yt ðPYt Þ1 ÞPYt ðPxt Yt ðPYt Þ1 Þ0 ;

where Yt is the observation vector containing all the observed variables, Y b  its predicted values, P  its
t Yt
conditional variance-covariance matrix, b x t the filtered state vector, and Pxt its variance-covariance
matrix.
The number of parameters in our model is large. Maximization of the likelihood (16) may involve a
large number of likelihood evaluations. Therefore, we adopt a sophisticated quasi-Newton approach
with the inverse Hessian of the likelihood function updated by the BFGS algorithm. The initial values
are carefully selected by the following way. We first run the Nelder-Mead optimization algorithm for
100 feasible sets of starting values and stop them after 100 iterations. Then the best 10 parameter
estimate sets (in terms of the likelihood) are selected among these 100 runnings as the initial values for
the quasi-Newton algorithm. The parameter estimates are those resulting in the largest likelihood
among these 10 runnings of the quasi-Newton method.

References

Andersen, T., Bollerslev, T., Diebold, F.X., Vega, C., 2007. Real-time price discovery in stock, bond and foreign exchange markets.
J. Int. Econ. 73, 251–277.
Anderson, B., Hammond, P., Ramerzani, C.A., 2010. Affine models of the joint dynamics of exchange rates and interest rates. J.
Financial Quantitative Anal. 45, 1341–1365.
Ang, A., Piazzesi, M., 2003. A no-arbitrage vector autoregression of term structure dynamics with macroeconomic and latent
variables. J. Monetary Econ. 50, 745–787.
Ang, A., Dong, S., Piazzesi, M., 2007. No-arbitrage Taylor Rules (Working Paper).
Backus, D.K., Gregory, A., Telmer, C.I., 1993. Accounting for forward rates in markets for foreign currency. J. Finance 48, 1887–
1908.
Backus, D.K., Foresi, S., Telmer, C.I., 1998. Discrete-time models of bond pricing (Working Paper).
64 W. Yin, J. Li / Journal of International Money and Finance 41 (2014) 46–64

Backus, D.K., Foresi, S., Telmer, C.I., 2001. Affine term structure models and the forward premium anomaly. J. Finance 56, 279–
304.
Bansal, R., Gallant, A.R., Hussey, R., Gauchen, G., 1995. Nonparametric estimation of structural models for high frequency cur-
rency market data. J. Econom. 66, 251–287.
Bekaert, G., 1996. The time variation of risk and return in foreign exchange markets: a general equilibrium perspective. Rev.
Financial Stud. 9, 427–470.
Bilson, F. O. John, 1978. The current experience with floating exchange rates: an appraisal of the monetary approach. Am. Econ.
Rev. 68, 392–397.
Brandt, M.W., Santa-Clara, P., 2002. Simulated likelihood estimation of diffusions with an application to exchange rate dynamics
in incomplete markets. J. Financial Econ. 63, 161–210.
Brandt, M.W., Cochrane, J.H., Santa-Clara, P., 2006. International risk sharing is better then You think, or exchange rates are too
smooth. J. Monetary Econ. 53, 671–691.
Brennan, M.J., Xia, Y., 2006. International capital markets and foreign exchange risk. Rev. Financial Stud. 19, 753–795.
Burnside, C., Eichenbaum, M., Kleshchelski, I., Rebelo, S., 2011. Do Peso problems explain the returns to the carry trade? Rev.
Financial Stud. 24, 853–891.
Cochrane, J., 2005. Asset Pricing (Revised). Princeton University Press, Princeton.
Dai, Q., Singleton, K.J., 2000. Specification analysis of affine term structure models. J. Finance 55, 1943–1978.
Diebold, F.X., Rudebusch, G.D., Aruoba, S.B., 2006. The macroeconomy and the yield curve: a dynamic latent factor approach. J.
Econom. 131, 309–338.
Dong, S., 2006. Macro Variables Do Drive Exchange Rate Movements: Evidence from a No-arbitrage Model (Working Paper).
Dornbusch, R., 1976. Expectations and exchange rate dynamics. J. Polit. Econ. 84, 1161–1176.
Duffee, G.R., 2002. Term premia and interest rate forecasts in affine models. J. Finance 57, 406–443.
Duffee, R.G., 2007. Are Variations in Term Premia Related to the Macroeconomy? (Working Paper).
Engel, C., West, K.D., 2005. Exchange rates and fundamentals. J. Polit. Econ. 113, 485–517.
Engel, C., Mark, N., West, K.D., 2007. Exchange Rate Models Are Not as Bad As You Think. In: NBER Macroeconomics Annual, vol.
22, pp. 381–441.
Fama, E., 1984. Forward and spot exchange rates. J. Monetary Econ. 14, 319–338.
Frankel, A.J., Ross, A.K., 1996. A panel project on purchasing power parity: mean reversion within and between countries. J. Int.
Econ. 40, 209–224.
Frenkel, A.J., 1976. A monetary approach to the exchange rate: doctrinal aspects and empirical evidence. Scand. J. Econ. 78, 200–
224.
Frenkel, A.J., 1978. Purchasing power parity: doctrinal perspective and evidence from the 1920s. J. Int. Econ. 8, 169–191.
Frenkel, A.J., Mussa, M.L., 1985. Asset markets, exchange rates, and the balance of payments. Handbook Int. Econ. II, 679–747.
Harrison, M., Kreps, D., 1979. Martingales and arbitrage in multiperiod securities markets. J. Econ. Theor. 20, 381–408.
Hodrick, R.J., 1989. Risk, uncertainty, and exchange rates. J. Monetary Econ. 23, 433–459.
Hodrick, R.J., Prescott, E.C., 1997. Postwar US business cycles: an empirical investigation. J. Money, Credit Banking 29, 1–16.
Hooper, P., Morton, J.E., 1978. Fluctuations in the dollar: a model of nominal and real exchange rate determination. J. Int. Money
Finance 1, 39–56.
Julier, J.S., Uhlmann, J.K., 1997. A new extension of the Kalman filter to nonlinear systems. In: Proceedings of AeroSense: the 11th
International Symposium on Aerospace/Defense Sensing, Simulation and Controls, pp. 182–193.
Julier, J.S., Uhlmann, J.K., 2004. Unscented filtering and nonlinear estimation. Proc. IEEE 92, 401–422.
Lubik, T., Schorfheide, F., 2005. A Bayesian Look at the New Open Economy Macroeconomics. In: NBER Macroeconomics Annual,
vol. 20, pp. 313–382.
Lucas, R., 1982. Interest rates and currency prices in a two-country world. J. Monetary Econ. 10, 335–359.
Lustig, H., Roussanov, N., Verdelhan, A., 2011. Common risk factors in currency markets. Rev. Financial Stud. 24, 3731–3777.
Meese, A.R., 1990. Currency fluctuations in the Post-Bretton woods era. J. Econ. Perspect. 4, 117–134.
Menkhoff, L., Sarno, L., Schmeling, M., Schrimpf, A., 2012. Carry trades and global foreign exchange volatility. J. Finance 67, 681–718.
Mussa, L.M., 1976. The exchange rate, the balance of payments and monetary and fiscal policy under a regime of controlled
floating. Scand. J. Econ. 78, 229–248.
Obstfeld, M., Rogoff, K.S., 1996. Foundations of International Macroeconomics. MIT Press, Cambridge, MA.
Obstfeld, M., Rogoff, K.S., 2003. Risk and exchange rates. In: Helpman, E., Sadka, E. (Eds.), Economic Policy in the International
Economy: Essays in Honor of Assaf Razin. Cambridge University Press, Cambridge.
Piazzesi, M., Schneider, M., 2006. Equilibrium Yield Curves. In: NBER Macroeconomics Annual, pp. 389–442.
Ravn, M.O., Uhlig, H., 2002. Notes on Adjusting the HP Filter for the Frequency of Observations (CEPR Working Paper).
Sarno, L., Schneider, P., Wagner, C., 2012. Properties of foreign exchange risk premiums. J. Financial Econ. 105, 279–310.
White, H., 1982. Maximum likelihood estimation of misspecified models. Econometrica 50, 1–25.
Wright, J.H., 2011. Term premia and inflation uncertainty: empirical evidence from an international panel dataset. Am. Econ.
Rev. 101, 1514–1534.