Federal Reserve Bank of Dallas

Globalization and Monetary Policy Institute

Working Paper No. 207
http://www.dallasfed.org/assets/documents/institute/wpapers/2014/0207.pdf

Can Interest Rate Factors Explain Exchange Rate Fluctuations? *

Julieta Yung
Federal Reserve Bank of Dallas
October 2014
Abstract
This paper explores whether interest rate factors, derived from the yield curve, can explain
exchange rate fluctuations at different horizons. Using a dynamic term structure model
under no-arbitrage, exchange rates are modeled as the ratio of two countries stochastic
discount factors. Key to this framework is that factors are observable, which allows the
model to be estimated by Maximum Likelihood. Results show that interest rate factors can
explain half of the variation in one-year exchange rates and up to ninety percent of five-year
movements, for free-floating currencies from 1999 to 2014. These findings suggest that yield
curves contain important information for modeling exchange rate dynamics, particularly at
longer horizons.
JEL codes: G15, F31, E43

Julieta Yung, Federal Reserve Bank of Dallas, Research Department, 2200 N. Pearl Street, Dallas, TX
75201. 214 922-5443. [email protected]. I am grateful to Tom Cosimano, Jun Ma, Mark
Wohar, and many others for helpful comments and suggestions. This paper also benefited from comments
made by the participants at the Macroeconomics Seminar and the Mathematical Finance Research Group at
the University of Notre Dame. The views in this paper are those of the author and do not necessarily reflect
the views of the Federal Reserve Bank of Dallas or the Federal Reserve System.

Introduction

The motivation for this paper arises from the idea that a few common factors drive most of the movement
in asset prices. Investors deciding how to allocate their wealth consider portfolios with different risk profiles,
assets with various time horizons, and investment opportunities in different countries. We can therefore
think of a unique pricing mechanism that simultaneously determines prices of assets, such as bonds and
exchange rates, given a limited set of information available at any point in time. In this paper, I explore
whether interest rate factors extracted from the yield curve can explain exchange rate fluctuations across a
number of different horizons.
Exchange rates are modeled as the ratio of two countries stochastic discount factors, derived from a
dynamic term structure model with observable interest rate factors. In this context, exchange rate movements
are driven by three components: (i) the interest rate differential, which is linear in the factors, (ii) a nonlinear
risk premium that varies throughout time, and (iii) the difference in shocks to the prices of risk. The interest
rate factors driving exchange rate movements, the level, slope, and curvature, are principal components from
the yield curve that capture over 99% of cross-sectional variation within a country. This paper contributes to
the international and macro-finance literature by modeling movements in the exchange rate from one month
to five years, through a dynamic no-arbitrage term structure model, using only factors from the yield curve.
The results suggest that yield curves contain important information for modeling exchange rate dynamics
during the 1999-2014 period, particularly at longer horizons. For example, three observable interest rate
factors from each country can explain about half of the variation in one-year exchange rates for countries
with free-floating currencies relative to the U.S. dollar: Australia (57%), Canada (45%), Japan (61%),
Norway (49%), Sweden (70%), Switzerland (42%), and the U.K. (61%). As the horizon increases, between
55% (Swiss franc) and 92% (British pound) of the five-year exchange rate fluctuations are accounted for
by these observable factors alone. Results are robust to correcting for small-sample bias with different
simulation-based bootstrapping methods and adjusting for potential heteroskedasticity in the error term,
suggesting that interest rate factors can better explain fluctuations in exchange rates at longer terms. These
findings are in line with work by Mark (1995), who finds that exchange rate fluctuations become more
predictable at longer horizons. Key to my paper is that factors are observable, which allows the no-arbitrage
term structure model to be estimated by Maximum Likelihood, as in Joslin, Singleton, and Zhu (2011).
The main driver of exchange rate fluctuations in this model is a time-varying risk premium, nonlinear in
the factors, derived from the countries stochastic discount factors. This component is consistent with the
Fama (1984) conditions: it is negatively correlated with the interest rate differential and exhibits a significantly higher volatility. I therefore study the model-implied risk premium in the context of the uncovered

interest parity condition, which links expected exchange rate movements to the interest rate differential between two countries with a coefficient of one. The puzzle in this literature arises when the coefficient is not
only different from one, but also negative. I project log exchange rate changes in the data onto the interest
rate differential and the risk premium components implied by the model. After including the risk premium,
the coefficient of the interest rate differential becomes consistently positive (except for Sweden at the oneand three-month horizons), approaches one, and is statistically significant at longer horizons. The coefficient
of the risk premium is always positive, approaches one, and is statistically significant for all currencies at
horizons greater than one month. This suggests that the risk premium implied by the model is not only
consistent with the theory, but also important in explaining exchange rate dynamics.
The ability of interest rate factors to explain cross-sectional variation in the yield curve decreases with
the number of factors. The yield curves third factor, or curvature, can explain less than 1% of interest
rate movements for most countries. Interestingly, the model finds that the curvature factor significantly
contributes to explaining exchange rate fluctuations despite its low explanatory power for the cross-section
of interest rates. For example, a model with two factors (level and slope) can only account for 28% of
variation in the one-year Japanese yen, compared to 61% with a three-factor model, and this is true for most
currencies studied. One exception is the Canadian dollar, since one-year fluctuations are well captured by
two factors alone, and incorporating a curvature factor only improves the fit of the model from 42% to 45%.
The findings in this paper are not driven by the U.S. dollar or any one country in particular. I explore
exchange rates in terms of the British pound, instead of the U.S. dollar, which implies U.K. interest rate
factors are extracted from the yield curve. In general, results are very similar regardless of the chosen
base, and the fit of the model improves at longer horizons, consistent with the baseline results. For some
countries, for example Canada, modeling one-year exchange rate fluctuations with respect to the British
pound significantly improves the models fit from 45% to 64%. This suggests that the U.S. term structure
does not contain more information on exchange rate movements than the U.K. Regardless of the country base,
exchange rate fluctuations are well captured in the long run by the same factors that explain cross-sectional
interest rate variation.
My model also allows one to explore a variety of questions; for example, which countries under the euro
can better explain exchange rate movements? I extract interest rate factors from countries under the euro
regime: Austria, Belgium, Finland, France, Germany, Ireland, Italy, the Netherlands, Portugal, and Spain.
In the baseline framework, each countrys interest rate factors capture the dynamics of the exchange rate at
different horizons, relative to the U.S. dollar. In the case of the euro, however, individual countries exhibit
very different yield curves, and it is unclear which countrys term structure contains more information on
the path of the euro. The model yields evidence that different European countries have roughly the same
3

ability to explain one-year euro-dollar fluctuations, from a stable country such as Germany (46%) to a much
more volatile one like Italy (58%). Moreover, at the five-year horizon, every country can explain more than
83% of exchange rate movements. This suggests that expectations on euro fluctuations are well captured by
the cross-sectional movement of the yield curves in all European countries.
The literature has mixed evidence on whether the information contained in both the domestic and
foreign yield curves alone is useful in accounting for exchange rate movements. For example, Inci and Lu
(2004) extend a quadratic term structure model in which unobservable interest rate factors feed into the
stochastic discount factor of different countries. They provide evidence that interest rates factors alone may
be insufficient to determine exchange rate movements for the U.S.-U.K. and U.S.-Germany country pairs. Li
and Yin (2014) reach similar conclusions in a no-arbitrage macro-finance framework, in which unobservable
yield curve factors alone can only explain 13% of the one-month euro-U.S. dollar exchange rate. They find
that incorporating macro variables helps explain exchange rate movements, although 76% of the variation
they find is driven by innovations to the factors.
Diez de Los Rios (2009) develops a two-country term structure model in continuous time, where the
exchange rate is determined by each countrys risk-free rate. This framework is able to provide statistically
better forecasts than those produced by a random walk for the British pound and the Canadian dollar,
although the results are negative for the German mark/euro and the Swiss franc, which are explained by
a rejection of the restrictions imposed by the term structure model. Clarida and Taylor (1997) estimate
a vector-error-correction model for the term structure of the forward premium (interest rate differentials)
and conclude that interest rates from two countries contain useful information for predicting exchange rates.
More recently, work by Ang and Chen (2010) finds that changes in the interest rates and slope of the yield
curve are significant predictors of foreign exchange returns.1 Although close to my work, these studies only
focus on very short-term exchange rate fluctuations.
My work also builds upon those who have studied the model-implied risk premium in the context of
the uncovered interest parity puzzle, such as Frachot (1996), Backus, Foresi, and Telmer (2001), Dewachter
and Maes (2001), Ahn (2004), and Alexius and Sellin (2012), among many others. I also explore the euro
dynamics as in Egorov, Li, and Ng (2011), although I pose a very different question: Which European
countries are able to better explain changes in the euro by only using their interest rate factors?
This paper contributes to this rapidly growing literature that studies the joint dynamics of exchange
1 These papers are not alone in making significant contributions to the literature, but the number of studies devoted to this
area is too large to summarize in this introduction alone. Some recent studies on the field are Dong (2006), who develops a
no-arbitrage two-country term structure model to explain exchange rate fluctuations between the U.S. and Germany; Benati
(2006), who studies U.S.-U.K. exchange rate fluctuations with unobservable factors; Chabi-Yo and Yang (2007), who introduce
a New Keynesian model to study the Canadian dollar with macroeconomic variables; Pericola and Taboga (2011), who use
observable and unobservable variables to explore exchange rate dynamics with German data; and more recently, Bauer and Diez
de Los Rios (2012), who introduce a multicountry framework with interest rate, macroeconomic, and exchange rate variables.

rates and yield curves through term structure models in the following dimensions. First, it only introduces
observable interest rate factors extracted from the countries yield curves. There are no other sources of risk
(such as macro risk or exchange rate risk), other than the ones spanned by the yield curve, which allows
the model to estimate all parameters by Maximum Likelihood. That is because the goal of this paper is
to explore whether the cross-sectional variation of the term structure alone can help explain movements in
other assets, such as the exchange rate. Second, this paper focuses on a more recent period, 1999 to 2014,
which is particularly interesting given the difficulty of modeling exchange rates during different financial
crises, severe changes to the term structure of interest rates, the creation of the euro, and the more recent
financial integration of asset markets across the world. Third, I consider not only one country pair but
eight different currencies relative to the U.S. dollar or the British pound. The model successfully explains
exchange rate fluctuations for every currency using only their interest rate factors. Fourth, I model exchange
rate changes at different horizons, from one month to five years, which has been largely unexplored in the
context of no-arbitrage term structure models. My results are in line with work by Mark (1995) that finds
short-term fluctuations are very difficult to explain but at longer horizons, there are significantly predictable
components driving exchange rate fluctuations.
The reminder of the paper is organized as follows. Section 2 describes the two-country term structure
framework in which exchange rates are modeled as the ratio of the countries stochastic discount factors.
The model is then estimated by Maximum Likelihood. Section 3 introduces the data and the interest rate
factors used to model both interest rates and exchange rates simultaneously. Section 4 presents the results
of the model as well as measures of its ability to explain exchange rate movements at different horizons,
along with a series of robustness checks. I also evaluate the model in the context of the uncovered interest
parity puzzle and study the models implications for the euro with a new set of European countries. Section
5 concludes.

Term Structure Model and Estimation

I derive a dynamic term structure model in discrete time in the spirit of Joslin, Singleton, and Zhu (2011).
This type of model was originally developed to explain the cross-sectional variation in yields.2 I expand its
purpose to not only account for the term structure of two countries simultaneously, but to also account for
movements in their exchange rate at different horizons.
2 Term structure models originated with Vasicek (1977) in the finance literature, who presumed that the short-term interest
rate was a linear function of a vector that followed a Gaussian diffusion (joint normal distribution). This has been the building
block of a large family of term structure models, formalized by Duffie and Kan (1996), characterized by Dai and Singleton
(2000, 2003), and expanded by many others. Comprehensive surveys on the literature can be found in Diebold, Piazzesi, and
Rudebusch (2005), Piazzesi and Schneider (2007), Piazzesi (2010), and Grkaynak and Wright (2012).

The key assumption is that there are no arbitrage opportunities in the bond markets. This assumption
implies that investors receive the same risk-adjusted compensation for bonds of different maturities. Noarbitrage is implemented in the form of cross-equation restrictions that make the yield curve consistent at
every point in time. Hence, the yields are modeled as linear functions of the factors, in which the coefficients
are restricted so that no-arbitrage holds, as in Duffie and Kan (1996). Although in reality this assumption
may not hold at every point in time, given the liquidity of well-developed bond markets, violations to the
arbitrage-free assumption are usually short lived, since bank holding companies and other large investors are
able to immediately take advantage of those opportunities, acting as arbitrageurs in the economy.

2.1

General Framework

I extend the term structure framework to allow for a domestic and a foreign country, where i = {d, f }. Let
Pti be a set of N observable factors for each country. These factors are constructed as N linear combinations
of zero coupon yields, yti , such that

Pti = W i yti ,

(1)

where W is an (N J) matrix containing the weights that factors have on the J maturity yields. The
model assumes that Pt and yt are priced perfectly, but the observed yields, ytobs , are measured with error;
such that eyt = ytobs yt . The errors are assumed to be conditionally independent of their lagged values and
satisfy consistency conditions as specified in Joslin, Singleton, and Zhu (2011). Each country is affected by
their own set of observable interest rate factors, Ptd and Ptf . Factors are extracted by principal component
analysis, as it is standard in the literature, for which they are orthogonal linear combinations of the country
yields.3
Models based on a reduced set of factors have proven to be a good representation of interest rate movements. Hence the assumption that a small number of factors drive the entire cross-section of interest rates
is not only parsimonious but also effective. Typically, three factors alone are enough to explain almost the
entire cross-sectional variation of interest rates, with the first -and most persistent- factor having the largest
explanatory power over interest rate movements. In this context these same factors will also drive exchange
rate movements at different horizons between the countries.
As in Vasicek (1977), interest rates are linear in the factors: Each country has its own one-period
annualized risk-free rate or short rate,
3 More

details on what the factors are and how they are extracted from the interest rate data can be found in section 3.

rti = i0P + i1P Pti ,

(2)

and its own set of yields at different m maturities, which are linear in the factors as well,

m,i
i
ytm,i = Am,i
P + BP P t ,

(3)

where coefficients AP and BP are restricted to satisfy internal consistency conditions, so that Am,d
Wd =
P
m,d
m,f
Am,f
W f = 0 and BP
W d = BP
W f = I. When dealing with observable factors constructed from
P

the cross-section of yields, Dai and Singleton (2003) argue that internal consistency conditions need to be
satisfied in order for the model to correctly price the state variables at each point in time.

2.2

Bond Pricing and the Stochastic Discount Factor

The stochastic discount factor (SDF) is the pricing mechanism that investors in each country develop in
order to price assets in the local currency. Let Ptm be the price at time t of an m maturity zero coupon bond
(no intermediate cashflows). Asset prices are risk-adjusted expected values of their future payoffs, so that



m1
Ptm = EtP Mt+1 Pt+1
,

(4)

where Mt+1 is the SDF that accounts for the adjustment of risk. The investors demand this risk premium
to compensate for unaccounted fluctuations or uncertainty in the economy. The SDF is also called pricing
kernel in the finance literature and can be specifically linked to preferences through the intertemporal rate of
substitution. SDFs are also related to the beta in conditional versions of the classical Capital Asset Pricing
Model (CAPM) and its multifactor extensions.4 The basic intuition of the CAPM still holds true: the factors
that determine asset prices also determine the risk premium, which compensates investors for weathering
losses during bad times. However, the main goal of CAPM is to express expected excess returns in terms
of a securitys beta with a benchmark portfolio without taking a stand on the characteristics of the pricing
relationship. Term structure models parametrize the SDF directly as a function of the state vector implied
by the market prices of risk, relating the risk premium to bonds across the yield curve without relying on
assumptions about preferences and technology. Let P represent the physical or actual process, such that EtP
represents realized expectations of a historical process.
Let Q represent the risk-neutral process, such that asset prices are expected values of their future payoffs
discounted at the risk-free rate:
4A

detailed discussion on the discount factor can be found in Singleton (2006).

 m1 
Ptm = ert EtQ Pt+1
.

(5)

Equations (4) and (5) are equivalent representations of the asset pricing mechanism. However, under riskneutrality, the prices have already been adjusted for risk and there is no need for a pricing kernel. The
existence of a risk-neutral distribution is guaranteed by the absence of arbitrage. Moreover, if Q is unique
then bond markets are complete.
The evolution of each countrys factors is expressed under the two equivalent probability measures P and
Q:

P,i
P,i
i
Pti = K0P
+ K1P
Pt1
+ iP P,i
t ,

(6)

Q,i
Q,i
i
Pt1
+ iP Q,i
+ K1P
Pti = K0P
t .

(7)

Under P the factors move by an unrestricted vector auto-regressive (VAR) process that describes the evolution
of the yields. Under Q, its equivalent Martingale measure, the process has already been adjusted for risk,


Q
P
, K0P
are (N 1)
allowing for a risk-neutral representation of the movement of factors over time. K0P


Q
P
constant vectors that can be interpreted as the long run average of the factors, K1P
, K1P
are (N N )
matrices that account for the speed of mean reversion in the process towards its long run unconditional
mean, and the shocks that disturb the state vector from moving back to its mean are i.i.d Gaussian, Pt , Q
t
N (0, IN ). P is a lower triangular matrix such that P 0P is the (N N ) variance-covariance matrix of
innovations to Pt . Changing distributions from P to Q removes the drift from the stochastic process, which
P
5
is essentially a shift in the mean and not the variance of P, for which Q
P = P = P .

The risk-free rate equation, together with the equation of motion for the factors under P, help recover the
historical moments of all bond returns. The equation for the short rate, along with the equation of motion for
the factors under Q, price all the assets in the economy. The reason it is necessary to specify the distribution
of the state variables under both measures, is because what lays between the P and Q distributions of the
state vector are adjustments for the market prices of risk: terms that capture agents attitudes toward risk.
Term structure models evaluate the evolution of the factors under both distributions to determine more
information on the pricing kernel as a function of a known state vector. As in Duffee (2002), the price of
risk is determined by the difference in the coefficients of the equations of motion,

P,i
Q,i
i0 = K0P
K0P
,

(8)

5 This is an application of the diffusion invariance principle, which describes the dynamics of stochastic processes when the
physical measure is changed to an equivalent martingale measure but only the drift changes (Girsanov, 1958).

P,i
Q,i
i1 = K1P
K1P
.

(9)

The price of risk coefficients measure how much the investor has to be compensated for fluctuations in the
state vector. t , the market price of risk, takes a linear functional form with respect to the factors,

it = i1
i0 + i1 Pti ,
P

(10)

which captures the prices of shocks to the risk factors per unit of volatility. Since it is independent of cashflow patterns of the security being priced, it is common to all securities with payoffs that are functions of the
factors, and is implied by the fact that factors are Gaussian under both distributions. Duffee (2002) and Dai
and Singleton (2002) show that this equation is able to match many features of the historical distribution of
yields really well. 0 represents the average prices of the factors assigned by investors, whereas 1 measures
how the market prices vary with respect to the risk levels summarized in Pt . Finally, 1
P captures the
interaction of the shocks to the factors.
The stochastic discount factor is then specified under the absence of arbitrage as an exponential quadratic
function that depends on each countrys set of factors through their risk-free rate and price of risk equations,

i
Mt+1



1  i0 i 
i0 i
i
t t t t+1 ,
= exp rt
2

(11)

where t+1 is the vector of shocks to the factors. Notice that each countrys stochastic discount factor
depends on its own interest rate factors alone.

2.3

Exchange Rates

By no arbitrage, the exchange rate between two countries is governed by the ratio of their stochastic discount
factors (Bekaert, 1996). Let St be the nominal spot exchange rate at time t; i.e., the domestic price of one
unit of the foreign currency. We thus have:
Mf
St+1
= t+1
;
d
St
Mt+1

(12)

Mf
St+k
= t+k
,
d
St
Mt+k

(13)

or, if extended to k horizons (in months),

where

i
Mt+k
=

k
Y

i
Mt+j
.

(14)

j=1

i
By taking logarithms, so that st+1 = log (St+1 ) log (St ) and log Mt+1
= rti

1
2


0
0
it it it it+1 ,

the log exchange rate change can be expressed as


  0
 1 0

0
0
st+1 = rtd rtf +
dt dt ft ft + dt dt+1 ft ft+1 .
2

(15)

This relationship formally describes the log exchange rate change as the sum of three different components:
(i) the interest rate differential between the countries, (ii) the exchange rate risk premium, and (iii) the
difference in the shocks to the price of risk assigned by investors from each country. Recall that the risk-free
n
o
rate and the prices of risk are linear functions of the factors Ptd , Ptf , from equations (2) and (10). In this
context, the same factors that drive variation in the cross-section of interest rates also drive the exchange
rate changes between them, but they do so in a nonlinear way through the risk premium component. Also,
this specification for exchange rates yields a time-varying risk premium, which is consistent with earlier
findings from Campbell and Shiller (1991) and Backus et. al (2001), among others, that provide evidence
from yield and forward rate regressions for a time-varying risk premium. Time-variance is also important to
match the dynamics of yields in the data and deviations from the Expectations Hypothesis, as shown in Dai
and Singleton (2002).
n
o
Finally, the model assumes that Pt and ytd , ytf , st are priced perfectly, but the observed variables,
o
n
, are measured with error; such that the vector of errors is
ytd,obs , ytf,obs , sobs
t

d,y
d,obs
e
ytd
t yt

f,y
f,obs y f
et =
et = yt
t

s
obs
et
st st

2.4

(16)

Estimation: Maximum Likelihood

If the factors are latent, estimates governing the physical distribution necessarily depend on those of the
risk-neutral distribution. In this case, the factors are observable variables captured from the yield curve
of each country. I estimate the model following Joslin, Singleton, and Zhu (2011), who show that when
factors are observable, a separation property holds: the parameters that govern the physical and risk-neutral
distributions are different, except for the variance-covariance matrix of the innovations to the factors. The
separation argument holds since the Maximum Likelihood estimates of the parameters are independent of
the variance-covariance matrix, as shown in Zellner (1962). Their work has simplified the computational
10

burden and provided state of the art estimation techniques that allow for more sophisticated specifications
of term structure models.
The first step in the estimation strategy is to optimize over
Q = [ d,Q

P , Q , where P = {dP , fP } and

0
f,Q ] . P is parametrized by the lower triangular Cholesky factorization of the variance-

covariance matrices from the vector auto-regressions under P for both countries. I do not impose any
restrictions between factors of different countries since their dynamics are modeled separately; Joslin, Le,
and Singleton (2013) argue that these initial estimations of P will be very close to the optimal values. Q is
Q
a vector of ordered eigenvalues in the K1P
matrix. It can be interpreted as the speed of mean reversion of the

factors under risk-neutrality and is assumed to be real, parametrized by the difference in ordered eigenvalues.
Q
Therefore, the model is Q stationary because the difference in eigenvalues is negative by construction. K0P
Q,i
Q,i Q,i
Q,i
= r
1 , where r
is the mean of the short rate for country i.
can be expressed in terms of k


The optimal estimates of P , Q are obtained by nonlinear optimization, in which the eigenvalues are

bounded so that Q < 0. P is unconstrained except for the diagonal elements in each country, which are
bounded from below by 0 to preserve singularity. Forcing a positive diagonal for the state transition matrix
implies the stability criteria for arbitrage free models and does not conflict with the trends of dynamic modn
o
P,i
P,i
Q,i
els.6 To summarize, the parameters to be estimated are = K0P
, K1P
, P , Q , k
and the observable

0
(interest rate) factors are Pt = Ptd Ptf .
The conditional log likelihood function can be factorized into the risk-neutral and the physical conditional
densities of the observed data:

obs
f ztobs | zt1
; = f Q ztobs | Pt + f P (Pt | Pt1 ) ,

(17)

where ztobs is the vector of observed yields for the domestic and foreign countries and the observed log

0
obs
d,obs
f,obs
obs
exchange rate changes at different horizons between the countries: zt = yt
. The
yt
st
P conditional density of the observed vector captures the time series properties of the factors, whereas the
Q conditional density captures the cross-sectional properties of yields.
Using the assumption that Pt is conditionally Gaussian, the log likelihood under P can be expressed as

f (Pt | Pt1 ) =

T

NT
1
1 X
1
log (2) + log | 0P P | +
(Pt Et1 [Pt ]) (0P P ) (Pt Et1 [Pt ]) ,
2
2
2 t=1
(18)

where T is the total number of time observations, N T = N d + N f is the total number of factors in the
6 Details

of the model can be found in the appendix.

11

model,



 
i
,
Et1 Pti = K0P,i + I + K1P,i Pt1

0P P = 

0
dP dP
0

dP dP

  0 0
fP fP

  0 0
dP dP
.
0
fP fP

0
fP fP

(19)

(20)

Imposing zeros on the off-diagonal elements of the joint variance-covariance matrix of the factors instead


does not lead to different results; these numbers are actually very small. The parameters K0P,i , K1P,i that
maximize the log likelihood function f conditional on t = 0 information are given by the OLS estimates of
the conditional mean of Pt .
The log likelihood under Q can be expressed as
T

1X






1 T
1 T
f ztobs | Pt =
e2t /2e +
J N T log (2) +
J N T log 2e ,
2 t=1
2
2

(21)

where e2t is the vector with

errors between the actual and the model-implied yields and exchange
r square
P
rates as in (16) and e =

T
e2
t=1 t
T
T (J N T )

, where J T is the number of observed variables.

Data and Interest Rate Factors

3.1

Interest Rate and Exchange Rate Data

Monthly interest rate data from January 1999 to January 2014 are obtained from Bloomberg, L.P. 2014,
for a total of 181 observations. This time period is particularly interesting since it experiences an increase
in the relevance of asset prices given the integration of financial markets, as well as drastic changes in
interest rate behavior, the creation of the Euro, and different financial crises around the globe. The dataset
contains interest rates of 12 different maturities: 3 months, 6 months, and 1 to 10 years for countries
with free-floating currencies for which interest rate data are available: Australia, Canada, Japan, Norway,
Sweden, Switzerland, the United Kingdom, and the United States. Yields are in percentage, end-of-month,
annualized, and expressed in the local currency. Bloomberg uses coupon bonds, bills, swaps, or a combination
of these instruments to derive the discount factors for each country using standard bootstrapping. The zerocoupon yields are then calculated step by step using the estimated discount factors with a minimum of four
instruments at different tenors for each observation.
Figure 1 shows the cross-sectional variation of interest rates throughout time for each country. In the
U.S., short-term interest rates increased steadily starting April 2004 to reach around 5% by the beginning

12

of 2007. In September 2008 the short rate dropped below 1% and has remained close to the zero lower
bound ever since. This pattern of low short-term rates is observed in most other countries as well during
the past few years. Australias rates tend to be slightly higher on average and Norway is the only country
to experience yield curve inversion in 1999. However, Japan is the country that exhibits the most dissimilar
term structure, since interest rates have been close to the zero lower bound for the entire sample.
Table 1 shows summary statistics for one-year yields. On average, Australia has the highest one-year
yields (4.82%) and Japan has the lowest (0.20%). The well established stylized facts on U.S. yield curves
also apply for the rest of the countries in the sample.7 A few features of interest rates during this period
that are worth highlighting are:
1. The yield curves are typically upward sloping. For the most part of the sample, low-maturity yields
are lower than high-maturity yields, except for times of crisis in which the yield curve flattens and
even tends to revert - the magnitudes depending on each countrys vulnerability to financial contagion,
among other factors. During 2008, for example, Switzerlands and Norways short-term rates not only
increased significantly, but the yield curves even inverted, being short-term rates higher than long-term
yields for subsequent months.
2. Standard deviation tends to be lower for long-term yields than short-term yields. This is particularly
true during the earlier years in the sample, where short-term yields exhibit high volatility. During the
last few years, however, yields have stayed close to the zero lower bound, exhibiting less variability
at the short end of the yield curve. Also, some countries exhibit more volatility than others. Japans
one-year yield, for example, has a standard deviation of 0.21 compared to 2.05 for the U.S. during the
same period.
3. Yields are very persistent. The first order autocorrelation is higher than 90% for all countries, with
the most persistent countries being Canada, Norway, the U.K., and the U.S., with an autocorrelation
higher than 65% after one year, as seen in table 1. The least persistent country is Australia, with a
one-year autocorrelation of 35%.
4. Most yields are close to a normal distribution over the sample period. This is consistent with contemporary work that suggests U.S. yields have become more Gaussian in recent years - see for example,
Piazzesi (2010). For most countries, skewness is around zero, except for Japan (1.34) and kurtosis
ranges from 1.5 (U.K.) to 2.5 (Australia). This suggests that modeling yields under the assumption of
a Gaussian distribution is quite reasonable, especially during the time frame considered in the sample.
7 For

analysis on stylized facts on U.S. yield curves refer to Piazzesi and Schneider (2007).

13

5. Yields are highly correlated across countries. Table 2 shows the correlation between one-year interest
rates of different countries. Interest rates of the same maturity tend to be very correlated across
countries. Although the Japanese yield curve is quite different from other countries term structure,
particularly during the earlier years, it is most correlated with Australias (50%) and Switzerlands
(61%). The U.K., the U.S., and Canada share a correlation above 89%.
6. Yields of near maturity are highly correlated. Table 3 shows the correlation of the 3-month interest
rates of each country with respect to different maturities from 6 months to 10 years. For all countries,
the closest the maturity range, the highest the correlation between interest rates of the same country.
For example, 3-month and 6-month rates are highly correlated, moving almost one-to-one, but 3-month
and 10-year rates tend to differ by country, ranging from 37% (Japan) to 84% (the U.K. and Norway).
The exchange rate data are also from Bloomberg L.P, 2014. There are a total of 181 monthly observations
per currency from January 1999 to January 2014. All the end-of-month exchange rates are expressed in
terms of U.S. dollars (USD) for the following countries: Australia (AUD), Canada (CAD), Japan (JPY),
Norway (NOK), Sweden (SEK), Switzerland (CHF), and the United Kingdom (GBP).
Table 4 summarizes the main characteristics of exchange rates during this period. Exchange rates, like
interest rates, are very persistent, with an autocorrelation between 57% (GBP) and 79% (CAD) after one
year. They are also highly related among themselves, sharing a correlation of 88% or higher, except for
the Japanese yen and the British pound, which are the most dissimilar currencies in the sample. The most
volatile currencies during this period are the Australian dollar, the Swiss franc, and the British pound with a
standard deviation above 0.17, and the least volatile currency is the Japanese yen (0.00). All the currencies
exhibit a skewness around zero, from -0.39 (SEK) to 0.83 (JPY), and kurtosis ranges from 1.5 (CAD) to 2.6
(JPY). For the estimation, I calculate annualized log exchange rate changes at different horizons: 1, 3, and
6 months, and 1 to 5 years, in 6-month increments.

3.2

Interest Rate Factors

Given the high correlation of yields within each country, the literature has taken advantage of dimensionalityreduction techniques to find the underlying drivers of interest rates. I extract the three main principal
components or factors of each country by the eigenvalue-eigenvector decomposition of the variance-covariance
matrix of the yields. Typically, the first two or three factors are enough to explain most of the cross-sectional
variability in all of the yields. In general, the factors for each country tend to have similar shapes, with
Japans being the most different. This corresponds to the fact that Japans interest rates have been very
low for the past decade relative to the rest of the sample.
14

The first factor is downward sloping, capturing the drop in interest rates that most countries have
experienced in the last few years. Table 5 presents summary statistics on the country-specific components.
In every country, the first factor extracted from the data is responsible for more than 81% of the crosssectional variation and the first three factors combined account for more than 99% of total variability. Since
the first factor captures most of the variation in all yields within a country, it fluctuates more and has higher
standard deviation (around 3.4) than the second (0.8) and third (0.2) factors. Interest rate factors are also
very persistent, with the first factor being more persistent than the second and third ones. In term structure
models, yields are usually modeled as linear functions of the factors. In this type of setting, that persistence
of yields comes from persistent factors, as observed in the data.
Figure 2 shows the loadings or weights that each factor places on the different maturities for every country.
The factor loadings exhibit the same pattern: horizontal for the first factor (around 0.3), downward sloping
for the second factor, and U-shaped for the third factor. Litterman and Scheinkman (1991) attributed the
labels level, slope, and curvature to these three factors given the effect they have on the yield curve. For
example, the first factor affects all yields the same, hence shocks to the first factor generate parallel shifts
on the curve, changing the level or average of yields. The second factor loads positively for short-term
maturities and negatively for long-term maturities, thus shocks to the second factor move long and short
yields in opposite directions, changing the slope of the yield curve. Finally, the third factor loads positively
on short- and long-term maturities, but negatively on mid-range maturities, thus shocks to the third factor
affect the curvature of the yield curve. The Litterman and Scheinkman (1991) interpretation of the factors
is very common in the literature, although it remains agnostic on links to the macroeconomy.
Recent work provides evidence that the level factor is partly correlated with inflation (Diebold et. al,
2006), aggregate supply shocks from the private sector (Wu, 2006), and shocks to preferences for current
consumption and technology (Evans and Marshall, 2007). The slope factor seems to be related to monetary
policy shocks (Evans and Marshall, 1998, 2007; Wu 2006) and real activity (Diebold et. al, 2006). However,
the curvature factor has not yet been linked to any specific variables. Through my model, the curvature factor
helps explain exchange rate fluctuations despite its low contribution to explaining cross-sectional variability
in yields. I find some evidence that higher order variables of the yield curve, such as the curvature factor,
might be useful in explaining exchange rate movements.

Results

The baseline model is estimated by Maximum Likelihood using three observable interest rate factors from
each country: level, slope, and curvature. The domestic country is the U.S., therefore exchange rates are

15

expressed in terms of USD for the 1999-2014 period. First, I present the results for the base line model
parameters in annual terms, as well as the term structure fit for each country pair. Next, I evaluate the
ability of the model to capture exchange rate movements at different horizons by only using interest rate
factors. I explore different model specifications, such as a model with less factors (only the countries level
and slope) and a model with a different base country instead of the U.S. (the U.K.), as well as several
robustness checks. I also explore the implications the model has on the uncovered interest parity puzzle, one
of the most relevant puzzles in the exchange rate literature. Finally, I study the implications of my model for
ten other countries under a common currency, the euro, and their ability to predict euro-USD fluctuations.
Table 6 shows the parameter estimates with asymptotic standard errors in parentheses that correspond
to the evolution of the factors under the physical distribution as an unrestricted VAR(1) in equation (6). All
P
100. The coefficient of the first factor is always
parameters are annualized. The first column shows K0P

positive, although only sometimes significant, with magnitudes ranging from 0.14 (Canada) to 0.94 (U.S.).
The coefficient of the second factor is negative for Australia, Japan, the U.K., and the U.S. and positive for
the rest, although not statistically significant. The coefficient of the third factor is statistically significant
for all the countries and always positive, except for Switzerland (-0.08). Table 6 also reports estimates of
P
+ I. As expected, the diagonal elements are close to one and statistically significant at the 99% level,
K1P

capturing the fact that interest rate factors are very persistent. The off-diagonal elements of the first and
second columns in the VAR are small and insignificant, although the estimates in the third column that
correspond to the first factor are larger and negative (around -0.40), ranging from -0.65 (Japan) to -0.14
(Canada), except for Switzerland (0.41), which is positive. Finally, the table reports the lower triangular
Cholesky factorization of the variance-covariance matrix, P 100, which contains large, positive, and
significant diagonal elements.
Table 7 reports the parameters of the risk-free rate equation (2), 0P and 1P , along with the largest
Q
eigenvalues of the K1P
+ I matrix, and a measure of goodness of fit of model-implied yields. 0P is always

negative between -0.017 and -0.002, 1P is always positive, and each factors estimate of 1P increases, being
the third factor estimates close to one for all countries, except for Switzerland (-1.054). The eigenvalues of
the transition matrix under risk-neutrality are large and close to one, hence shocks to these factors have
permanent effects in their evolution over time. As a measure of goodness of fit, the last column reports root
mean square errors (RMSE) in basis points for the cross-section of yields. The RMSE is around 10 basis
points, with the best fit for the Japanese yen (4.7) and Australian dollar (4.9) and the worst fit for the Swiss
franc (11.8). Overall, the goodness of fit for interest rates is very good, as it is standard in term structure
models of interest rates, which suggests that the incorporation of exchange rates into the model does not
compromise the fit of the yield curve.
16

4.1

Exchange Rate Fit

In order to obtain a general measure of goodness of fit for exchange rates, I project the exchange rate data
on the model-implied exchange rates. Hence, given the model equation describing annualized log exchange
rate changes at different k horizons,

f
d
st+k = rt+k1
rt+k1
+

i
0
0
0
1 h d0
t+k1 dt+k1 ft+k1 ft+k1 + dt+k1 dt+i ft+k1 ft+i ,
2

(22)

I perform an ordinary least squares (OLS) regression of the data on the model-implied exchange rate equations
for each currency i,

i
i
i
i
sobs,i
t+k = + st+k + t+k .

(23)

Table 8 displays the results from regression (23), which account for the goodness of fit of the annualized
model-implied exchange rate changes at different k horizons, from 1 month to 5 years. This model determines
the exchange rate changes as a function of six total factors (the level, slope, and curvature of each country).
The domestic country is chosen to be the U.S. so that the exchange rates are in terms of USD prices.
Therefore, an increase in st represents a depreciation of the U.S. dollar with respect to the foreign currency,
i. The results for i , along with the corresponding standard errors and the adjusted R2 , are reported for
selected horizons: 1 , 3 , and 6 months, and 1 , 1.5 , 2 , 3 , 4 , and 5 years.
At the one-month horizon, the fit of the model, as measured by the R2 , ranges from 1% and insignificant
(Switzerland) to 7% and statistically significant at the 99% confidence level (Japan). Very short exchange
rate fluctuations are quite difficult to predict. Meese and Rogoff (1983) claimed that the best forecasting
model is obtained under the assumption that the exchange rate follows a random walk. However, at longer
horizons the predictability of exchange rates improves significantly. This type of behavior is observed in other
assets as well. Although short-term movements are not well understood, it seems that at longer horizons
the factors are better able to capture the signal that drives movements in asset returns. Figure 3 shows the
annualized one-month log changes in exchange rates for all the currencies relative to the USD as well as the
model-implied expected exchange rate fluctuations. Overall, the expected exchange rate changes implied by
the model are not as volatile; however, they capture the data movements on average quite well.
In general, the regression coefficients increase in magnitude and significance as the horizons increase. At
the one-year horizon and thereafter, the coefficients of all currencies are close to one. Approximately half of
the variation in one-year exchange rate movements can be explained by the model using interest rate factors

17

alone, for all currencies relative to the USD. Figure 4 shows the data and model-implied one-year expected
fluctuations in the exchange rate. The fit of the model at this horizon ranges from 41.8% for the Swiss franc
to 70.3% for the Swedish krona. The model is even more successful at capturing five-year exchange rate
movements, depicted in figure 5, ranging from 55% (Swiss franc) to 91.6% (British pound).
One of the potential concerns within this framework is small-sample bias. Because the closer to a unit root,
the larger the small-sample bias in Maximum Likelihood estimates, Joslin, Singleton, and Zhu (2011) enforce
a high degree of persistence under the physical distribution of the factors to ensure a stationary process, by
which they constrain all the eigenvalues that govern the speed of mean reversion to be strictly positive. Smallsample bias arises due to the high persistence of the factors, and becomes particularly problematic when the
sample period is short. To this end, I estimate robust (White) standard errors and Newey-West standard
errors to account for small-sample bias and for potential correlation and heteroskedasticity in the innovation
vector. This generates larger standard errors that specify the precision of the estimated coefficients. Similar
to Lustig, Roussanov, and Verdelhan (2011), I follow Andrews (1991) to choose the number of lags for the
Newey-West estimation to be the horizon k (in months) plus one. After the three-month horizon, all of the
estimates are statistically significant at the 99% confidence level regardless of the method used to calculate
standard errors, except for the Norwegian krone under Newey-West standard errors that is significant at the
95% level.
Bauer, Rudebusch, and Wu (2012, 2014) argue that this method still leads to large biased estimates
that may yield an inconsistent risk premium. Given the separation argument, only the parameters that
characterize the physical density of yields are affected. They propose to replace the initial estimates of
the factors vector auto-regressive process under P with simulation-based bias-corrected estimates and then
proceed to the Maximum Likelihood estimation. I therefore implement their suggested inverse bootstrap biascorrected method: For each of the 5,000 bootstrap iterations (after discarding the first 1,000), I simulate a set
of 50 bootstrap samples using some trial parameters and calculate the mean or median of the OLS estimator
to be equal to the original OLS estimate. The bias-corrected estimates then reflect more persistent factor
dynamics, as measured by the largest eigenvalues of the transition matrix. More details on the estimation
of bias-corrected factor dynamics and the consequences of OLS bias in term structure models can be found
in Bauer, Rudebusch, and Wu (2012, 2014).
Table 9 shows the exchange rate fit for the model after adjusting for mean and median bias-corrected
estimates for selected maturities with the coefficients from regression (23), Newey-West standard errors in
parenthesis, and R2 adjusted for degrees of freedom. As expected, the VAR physical dynamics are more
persistent after correcting for bias, either by mean or median correction, although estimates for P remain
close to the estimates of the unrestricted VAR. After estimates are corrected for bias (either by the mean
18

or the median) interest rate factors are still able to explain as much variation in exchange rates. The same
patterns as the base model can be observed: at the one-month horizon, the best fit is the Japanese yen (6%)
and the worst is the Swiss franc (0%), and as the horizon increases, the fit of the model improves for all
currencies. Exchange rate changes at the one-year horizon and thereafter have coefficients around one that
are statistically significant at the 99% confidence level, and about 50% of one-year exchange rate movements
can be explained by the model, from 43% (Swiss franc and Canadian dollar) to 60% (Japanese yen and
Swedish krona). For some currencies, correcting for bias reduces the fit of the exchange rate model, but for
some currencies (Norwegian krone and Swiss franc), it improves it. Most importantly, the results are still
quite similar, and at the five-year horizon, the model accounts for 57% to 91% of exchange rate movements
using interest rate factors alone.
I next evaluate the performance of the model with a different number of factors. Table 10 reports the
estimates for selected maturities of a linear projection of the data on a model with only one factor, the level,
or two factors, the level and the slope, for each country. With fewer interest rate factors, the root mean
square errors that measure the term structure fit of each country increase relative to the three-factor model.
The two-factor model has RMSEs that are from 1.4 basis points (Switzerland) to 17.1 basis points (Canada)
higher than the three-factor model. In contrast to the base model, the two-factor model exhibits coefficients
of 0P that are positive for all countries, except for Sweden, and the coefficients of 1P are positive, except
for Japans first factor, with the second coefficient being larger than the first. The physical distribution
parameters do not exhibit significant differences in signs or magnitudes. The one-factor model exhibits
larger RMSEs than the other models, as expected, ranging from 43.4 basis points for Norway to 103.9 for
Canada.
In general, the same pattern emerges for all models: the exchange rate fit improves at longer horizons.
In every case, the two-factor model outperforms the one-factor model, although they are both well below the
performing levels of the three-factor model. That is, adding a third factor to the model, the curvature factor,
significantly improves the models ability to explain exchange rates, despite having low-content information
to explain cross-sectional interest rate variation.
A few highlights of these lower-factor models are worth mentioning. For Norway, Sweden, and the U.K.,
the one-factor model can explain less than 5% of exchange rate movements at horizons one-year or lower,
so for these countries most of the valuable information in explaining short-term exchange rate fluctuations
actually comes from the other components, the slope and curvature. However, the one-factor model does
remarkably well for Australia and Japan, over 20% of one-year changes are captured by the countries level
alone, which suggests that Japans and Australias level factor may contain a lot of information about their
exchange rate movements. Finally, the two-factor model explains exchange rate changes extremely well for
19

Canada relative to the three-factor model, which suggests that incorporating the curvature factor does not
seem to significantly contribute to modeling the CAD-USD exchange rate path. Given the relevance of the
curvature factor in modeling most currency movements in the long run, the third yield curve component
could be capturing long run expectations on exchange rate fluctuations or other fundamentals relative to
international markets that influence the exchange rate.
I also re-estimate the model considering a different base country, the U.K., so that exchange rates are
in terms of the British pound, instead of the U.S. dollar. Table 11 reports the coefficients, robust and
Newey-West standard errors, and the adjusted R2 for all countries in which the log exchange rate changes
are modeled in terms of the British pound. The model parameters are very similar, for example, 0P and
1P have the same sign and similar magnitude regardless of the base country, including 13P for Switzerland,
which is negative one as well. Notice that modeling the USD-GBP exchange rate is the same as modeling
the GBP-USD exchange rate, since results are symmetric. This is because it does not matter what country
is selected as domestic and what country is selected as foreign; the information content in the factors is still
the same. Hence the regression results for the U.S. dollar in terms of the British pound from table 11 are
the same as the results for the British pound in terms of the U.S. dollar from table 8.
For the other countries results are similar, although not the same, since instead of the U.S. term structure
we are now modeling exchange rates with U.K. interest rate factors, which exhibit different shapes. In general,
results are similar whether one is using the U.S. or the U.K. as a base: The coefficients are significant at the
99% confidence level after 3 months (except for Norway under N.W. standard errors, which is significant at
a 95% confidence level, just as in the base model).
For exchange rate fluctuations in the Australian dollar, Canadian dollar, Japanese yen, and Swiss franc,
results are either slightly better or significantly better when using the U.K. as a country base. For example,
the Swiss franc, which has the worst fit at most maturities when modeled with respect to the U.S. dollar,
significantly improves the fit with a 47% R2 at the one-year horizon and 92% at the five-year horizon, in
contrast with 41% and 55% under the USD, respectively. In the base model, 45% of one-year CAD-USD
exchange rate fluctuations are explained by the factors, whereas the model with U.K. factors accounts for
more than 64% of the CAD-GBP one-year exchange rate. Moreover, except for Sweden, all other currencies
experience a better fit at the five-year horizon when modeled in terms of the British pound, with R2 from
80% (NOK) to 92% (JPY and CHF). Estimates for the Nordic countries explain less of the exchange rate
movements when the GBP is used as a base. For Norway, results are slightly worse, whereas for Sweden, a lot
worse. This may suggest that some countries have more information than others about expected exchange
rate movements, although we can say with certainty that the U.S. is not particularly more informative than
other countries, such as the U.K.
20

4.2

Implications in the context of the UIP Puzzle

In this section, I evaluate my model in the context of the Uncovered Interest Parity (UIP) puzzle. UIP
claims that any excess return on foreign-denominated deposits is offset by the expected depreciation against
the domestic currency. That is, if we run the following regression



f
d
sobs
t+k = + U IP rt+k1 rt+k1 + t+k

(24)

we would expect that = 0, U IP = 1.8 However, the data do not hold against this proposition, since
the coefficient of the interest rate differential is not only different from one, but also negative, and the
explanatory power of the interest rate differential is very low. Several papers have focused on studying this
phenomenon.9 In my model, the UIP condition does not hold, since the expected log exchange rate change
is not only a function of the interest rate differential but also a function of a risk premium term implied by
my framework:

 h


i
1 d0
f
f0
f
d
d
sobs
=

r
+

+ t+k .
IRD
RP
t+k
t+k1
t+k1
t+k1 t+k1
2 t+k1 t+k1

(25)

This is consistent with the idea proposed by Fama (1984) and Hansen and Hodrick (1983), that attribute
the anomalous findings to a time-varying risk premium. I evaluate the ability of my model to reproduce the
UIP puzzle when equation (24) is estimated and I examine to what extent, the augmented UIP regression
implied by my model, equation (25), can help solve the puzzle.
Table 12 shows the results for the UIP and augmented-UIP regressions. The coefficient of the UIP
regression, U IP , is sometimes positive, sometimes negative, but never statistically significant in the short
run. At longer horizons (after two years), some coefficients are significant but always large and negative.
Moreover, the adjusted R2 is below 5% for one-year or lower frequency changes for every country. This is
consistent with the UIP puzzle observed in the data.
When incorporating the model-implied risk premium, the adjusted R2 always increases, the coefficient
of the interest rate differential, IRD is always positive (except for Switzerland at the one-month and threemonth horizon) although only statistically significant at longer horizons. For the three-year horizon, the
interest rate differential is close to one for every currency and the coefficients are statistically significant at
8 Notice that this regression is a slight variation from the standard UIP regression, since the interest rate differential is
lagged one period with respect to the log exchange rate changes. This is because the interest rate differential component comes
from the model and it is constructed as a linear function of the factors. Standard UIP regressions use the actual interest rate
differential from the data with yields of the corresponding maturity k, such as, for example, Flood and Taylor (1996) or Dong
(2006). The idea in this case is to evaluate the interest rate differential implied by the model.
9 Surveys on the UIP literature can be found in Hodrick (1987), Froot and Thaler (1990), Lewis (1995), and Engel (1996).

21

the 99% confidence level, with the lowest coefficient being 0.60 for the Swiss franc. The coefficient of the risk
premium, RP , is always positive, approaches one, and becomes statistically significant at the 99% confidence
level for every country after the three-month horizon (except for Norway, which is statistically significant at
the 95% level). These results suggest that the risk premium term is clearly important in modeling exchange
rates, even when only interest rate factors are considered, and contributes to closing the gap with respect to
the UIP puzzle.
Besides attributing the failure of UIP to the presence of a time-varying risk premium, Fama (1984)
identified two conditions necessary for the risk premium to solve the UIP puzzle. First, the risk premium has
to be more volatile than the interest rate differential, and second, the model-implied risk premium and the
interest rate differential need to be negatively correlated. Table 13 shows that the Fama conditions hold for
all currencies and for all horizons considered. The variance of the risk premium is higher than the variance of
the interest rate differential by a magnitude of two or higher. Also, the correlation between the interest rate
differential and the risk premium is always negative and large as it increases in horizon, as expected. This
suggests that the risk premium implied by the model with interest rate factors is a key driver of exchange
rates, which is consistent with Fama (1984) and Hansen and Hodrick (1983) explanation of the UIP puzzle,
and does not rely solely on large shocks to drive movements in the exchange rate.

4.3

Implications for the Euro Dynamics

In the baseline model, the factors extracted from two countries help determine their exchange rate fluctuations through each countriess stochastic discount factor. However, when multiple countries share the same
currency, what are the factors that determine movements in, for example, the euro-U.S. dollar exchange
rate?
I explore the determination of the path of the euro by estimating the stochastic discount factor of ten
countries under the euro regime: Austria, Belgium, Finland, France, Germany, Ireland, Italy, the Netherlands, Portugal, and Spain. That is, I estimate pricing kernels for those countries individually to explore
whether their interest rate factors alone contribute to explaining changes in the euro (EUR). My model
allows one to observe whether any particular country in the Euro area has more useful information on the
EUR-USD exchange rate fluctuations than others.
Figure 6 shows the evolution of yield curves for each country. Although some term structures exhibit
similarities, such as Germanys, Finlands, and the Netherlands, other countries behave quite differently,
such as Ireland, Portugal, Italy, or Spain; these countries exhibit different yield curve movements, particularly
during the 2011-2012 European crisis. Although the short rate increased for every European country during

22

that time, some were more susceptible to the Greek crisis than others, with their short rates reaching
significantly higher levels. For example, Portugals interest rates rose to 17% in January 2012. Because
these countries exhibit different yield curves, interest rate factors extracted from each term structure will
also be quite different.
Which country can explain euro fluctuations the most? As expected, the yield curve factors for a country
such as Germany are important in determining changes in the euro. However, risk factors from countries
with more volatile interest rates, such as Italy and Spain, also have high explanatory power on the exchange
rate dynamics. For example, Italy has higher explanatory power over euro-USD fluctuations than Germany
at every horizon. The results suggest that the risk factors that govern fluctuations in the euro are priced to
a similar extent in the yield curves of different European countries.
Table 14 reports the results of the linear projection of euro movements on the model-implied exchange
rates for ten European countries in terms of the U.S. dollar. All countries seem to have similar degrees of
explanatory power on the euro. In general, the same conclusion holds for this set of countries: exchange
rate fit improves at longer horizons and the addition of a curvature factor generally improves the fit of the
model. At the one-year exchange rate horizon, all the countries are able to explain around 50% of exchange
rate movements. Ireland and Finland exhibit the worst fit (39% and 40%), while Italy, Spain, Portugal, and
Belgium show the best fit (58%, 51%, 50%, and 50%). However, at the five-year horizon, every country is
able to explain 83% or more of the total variation in the EUR-USD exchange rate, with the lowest being
Austria (83.8%) and Spain (88.6%) the highest. After the three-month horizon, every country is statistically
significant at the 99% confidence level, except for Finland and Portugal, which are significant at the 95%
level with N.W. standard errors. Results are quite similar regardless of what country is the base, the GBP
or the USD. These results suggest that currency risk in the euro is priced in every European country, and
term structure cross-sectional variation captures those expectations really well.

Conclusion

This paper examines whether interest rate factors alone can help determine fluctuations in the exchange
rate using a no-arbitrage term structure model. I propose a framework that models exchange rate changes
at different horizons as the ratio of two countries stochastic discount factors in which most of the variation
is driven by a nonlinear risk premium. The results suggest that interest rate factors are indeed helpful in
modeling exchange rates, particularly at longer horizons. These results are in line with findings by Mark
(1995), who identifies a predictable component in exchange rate fluctuations as the horizon increases. Key
to my results is the fact that interest rate factors extracted from the yield curve are observable, which allows

23

for a straight forward estimation of the model by Maximum Likelihood as in Joslin, Singleton, and Zhu
(2011).
Interest rate factors, level, slope, and curvature, can explain half of the variation in one-year exchange
rates for countries with free-floating currencies from 1999 to 2014: Australia, Canada, Japan, Norway,
Sweden, Switzerland, and the U.K., with respect to the U.S. dollar. Moreover, the fit of the model improves
as the horizon increases, suggesting that long-term expected fluctuations in the exchange rate are captured
by yield curve factors. Results are robust to accounting for small-sample bias, potential heteroskedasticity
in the error term, and across multiple currencies relative to a different base. A drawback in the literature
is that there is no clear link between interest rate factors and the macroeconomy, particularly the curvature
factor. In this paper, I present evidence that despite of the curvature factor explaining a small fraction of
yield curve movements, it provides a significant contribution to explaining exchange rates. When using fewer
factors, either the level or the level and slope, the explanatory power of the model decreases, suggesting that
the curvature factor is important in explaining exchange rate fluctuations for most currencies in the sample.
A key implication of the model is that a risk premium, nonlinear in the interest rate factors, drives
most of the variation in exchange rates. Further analysis in the context of the uncovered interest parity
puzzle suggests that the model-implied risk premium satisfies the Fama (1984) conditions: is negatively
correlated with the interest rate differential and exhibits significantly larger variance. This two-country term
structure model with observable interest rate factors is quite successful at deriving stochastic discount factors
that imply a theory-consistent risk premium. Moreover, this paper explores the ability of ten European
countries to account for fluctuations in the euro by only using their own interest rate factors. Although
European countries exhibit different yield curves, they all capture exchange rate movements to a similar
extent, suggesting that exchange rate risk is implicit in yield curves of every European country in the sample.
Further work on this area includes exploring whether these results are time-independent, or if during crises,
certain European countries capture exchange rate information better than others.
There remain multiple avenues to be explored within this framework. Given that this work presents
evidence of interest rate factors explanatory power over exchange rate movements at different horizons, a
possible extension is to investigate the factors ability to forecast exchange rates out of sample. Despite the
large body of literature devoted to modeling exchange rates, there is a high degree of uncertainty about how
exchange rates react to changes in expectations, particularly out of sample.
Motivated by the idea that a unique stochastic discount factor should be able to account for prices of
different assets, another extension of this paper is to incorporate equities. Given that the same stochastic
discount factor prices securities in the bond and foreign exchange markets, it should price securities in
the equity market as well. Similarly to exchange rates, the equity returns literature also finds a long-run
24

predictable component in stock returns: the price-dividend ratio, e.g., Fama and French (1988), which is a
characteristic observed in other assets as well.
One of the present challenges the literature faces is that the stochastic discount factor is not derived
within a general equilibrium framework in terms of consumption preferences and technology parameters.
Integrating term structure models into a preference-based framework is of interest not only to academics
seeking to understand the theoretical connections of asset pricing dynamics, but also to monetary policy
makers, bond investors, and other financial market participants concerned with fluctuations in asset prices
from a general equilibrium perspective.

25

References
[1] Alexius A. and P. Sellin (2012), Exchange Rates and Long-Term Bonds, Scandinavian Journal of
Economics 114(3), 974-990.
[2] Ahn, D. (2004), Common Factors and Local Factors: Implications for Term Structures and Exchange
Rates, Journal of Financial and Quantitative Analysis 39(1), 69-102.
[3] Andrews, D. (1991), Heteroskedasticity and Autocorrelation Consistent Covariance Matrix Estimation, Econometrica 59(3), 817-58.
[4] Ang, A. and J. Chen (2010), Yield Curve Predictors of Foreign Exchange Returns, Working Paper,
March 2010 version.
[5] Backus, D. K., S. Foresi, A. Mozumdar, and L. Wu (2001), Predictable Changes in Yields and Forward
Rates, Journal of Financial Economics 59(3), 281-311.
[6] Backus, D. K., S. Foresi, and C. I. Telmer (2001), Affine Term Structure Models and the Forward
Premium Anomaly, Journal of Finance 56(1), 279-304.
[7] Bauer, G. H. and A. Diez de Los Rios (2012), An International Dynamic Term Structure Model with
Economic Restrictions and Unspanned Risks, Bank of Canada Working Paper 12-5, January 2012
version.
[8] Bauer, M. D., G. D. Rudebusch, and J. C. Wu (2012), Correcting Estimation Bias in Dynamic Term
Structure Models, Journal of Business and Economic Statistics 30(3), 454-467.
[9] Bauer, M. D., G. D. Rudebusch, and J. C. Wu (2014), Term Premia and Inflation Uncertainty: Empirical Evidence from an International Panel Dataset: Comment, American Economic Review 104(1),
323-337.
[10] Bekaert, G. (1996), The Time Variation of Risk and Return in Foreign Exchange Markets: A General
Equilibrium Perspective, Review of Financial Studies 9(2), 427-70.
[11] Benati, L. (2006), Affine term structure models for the foreign exchange risk premium, Bank of
England Working Paper 291.
[12] Campbell, J. Y. and R. Shiller (1991), Yield Spreads and Interest Rate Movements: A Birds Eye
View, Review of Economic Studies 58, 495514.

26

[13] Chabi-Yo, F. and J. Yang (2007), A No-Arbitrage Analysis of Macroeconomic Determinants of Term
Structures and the Exchange Rate, Working Paper 2007-21.
[14] Clarida, R. H. and M. P. Taylor (1997), The Term Structure of Forward Exchange Premiums and
the Forecastability of Spot Exchange Rates: Correcting the Errors, The Review of Economics and
Statistics 79(3), 353-361.
[15] Dai, Q. and K. Singleton (2000), Specification Analysis of Affine Term Structure Models, Journal of
Finance 55, 1943-78.
[16] Dai, Q. and K. Singleton (2002), Expectation Puzzles, Time-varying Risk Premia, and Affine Models
of the Term Structure, Journal of Financial Economics 63, 415-441.
[17] Dai, Q. and K. Singleton (2003), Term Structure Dynamics in Theory and Reality, Review of Financial
Studies 16, 631-678.
[18] Dewachter H. and K. Maes (2001), An Affine Model for International Bond Markets, Working Paper,
February 2001 version.
[19] Diebold, F. X., M. Piazzesi, and G. D. Rudebusch (2005), Modeling Bond Yields in Finance and
Macroeconomics, American Economic Review P&P 95(2), 415-520.
[20] Diebold, F.X., G. D. Rudebusch, B. and Aruoba (2006), The Macroeconomy and the Yield Curve: A
Dynamic Latent Factor Approach, Journal of Econometrics 131, 309-338.
[21] Diez de Los Rios, A. (2009), Can Affine Term Structure Models Help us Predict Exchange Rates?,
Journal of Money, Credit and Banking 41 (4), 755-766.
[22] Dong, S. (2006), Macro Variables Do Drive Exchange Rate Movements: Evidence from a No-Arbitrage
Model, Working Paper, July 2006 version.
[23] Duffee, G. (2002), Term Premia and Interest Rate Forecasts in Affine Models, Journal of Finance 57,
405-443.
[24] Duffie, D. and R. Kan (1996), A Yield-Factor Model of Interest Rates, Mathematical Finance 6,
379-406.
[25] Egorov, A. V., H. Li, and D. Ng (2011), A tale of two yield curves: Modeling the joint term structure
of dollar and euro interest rates, Journal of Econometrics 162(1), 55-70.

27

[26] Engel, C. (1996), The Forward Discount Anomaly and the Risk Premium: A Survey of Recent Evidence, Journal of Empirical Finance 3(2), 123-192.
[27] Evans, C. L. and D. A. Marshall (1998), Monetary Policy and the Term Structure of Nominal Interest
Rates: Evidence and Theory, Carnegie-Rochester Conference Series on Public Policy 49(1), 53-111.
[28] Evans, C. L. and D. A. Marshall (2007), Economic Determinants of the Nominal Treasury Yield Curve,
Journal of Monetary Economics 54(7), 1986-2003.
[29] Fama, E. (1984), Forward and Spot Exchange Rates, Journal of Monetary Economics 14, 319-338.
[30] Fama, E. F. and K. R. French (1988), Dividend Yields and Expected Stock Returns, Journal of
Financial Economics 33, 3-25.
[31] Flood, R. P. and M. P. Taylor (1996), Exchange Rate Economics: Whats Wrong with the Conventional
Macro Approach?, NBER Chapters, in: The Microstructure of Foreign Exchange Markets, 261-302.
[32] Frachot, A. (1996), A reexamination of the uncovered interest parity hypothesis, Journal of International Money and Finance 15(3), 419-437.
[33] Froot, K. and R. Thaler (1990), Anomalies: Foreign Exchange, Journal of Economic Perspectives
4(3), 179-192.
[34] Girsanov, I. (1958), On Transforming a Certain Class of Stochastic Processes by Absolutely Continuous
Substitution of Measures, Theory Probability and Its Applications 5, 285-301.
[35] Grkaynak, R. and J. Wright (2012), Macroeconomics and the Term Structure, Journal of Economic
Literature 50(2), 331-67.
[36] Hansen, L. and R. Hodrick (1983), Risk Averse Speculation in the Forward Exchange Markets: An
Econometric Analysis of Linear Models, In J. A. Frenkel (Ed.), Exchange Rates and International
Economics. Chicago: University of Chicago Press.
[37] Hodrick, R. (1987), The Empirical Evidence on the Efficiency of Forward and Futures Foreign Exchange
Markets, Chur, Switzerland: Harwood Academic Publishers.
[38] Inci, A. and B. Lu (2004), Exchange Rates and Interest Rates: Can Term Structure Models Explain
Currency Movements?, Journal of Economic Dynamics & Control 28, 1595-1624.
[39] Joslin, S., A. Le, and K. Singleton (2013), Why Gaussian Macro-Finance Term Structure Models Are
(Nearly) Unconstrained Factor-VARs, Journal of Financial Economics 109, 604-622.
28

[40] Joslin, S., K. Singleton, and H. Zhu (2011), A New Perspective on Gaussian Dynamic Term Structure
Models, Review of Financial Studies 24, 926-970.
[41] Lewis, K. (1995), Puzzles in International Financial Markets, in G. Grossman and K. Rogoff, eds.
Handbook of International Economics (North Holland: Amsterdam).
[42] Li, J. and W. Yin (2014), Macroeconomic Fundamentals and the Exchange Rate Dynamics: A NoArbitrage Macro-Finance Approach, Journal of International Money and Finance 41, 46-64.
[43] Litterman, R. and J. Scheinkman (1991), Common factors affecting bond returns, Journal of Fixed
Income 1, 54-61.
[44] Lustig, H., N. Roussanov, and A. Verdelhan (2011), Common Risk Factors in Currency Markets,
Review of Financial Studies 24 (11), 3731-3777.
[45] Mark, N. C. (1995), Exchange Rates and Fundamentals: Evidence on Long-Horizon Predictability,
American Economic Review 85(1), 201-218.
[46] Meese, R. A. and K. Rogoff (1983), Empirical Exchange Rate Models of the Seventies: Do they Fit
Out of Sample?, Journal of International Economics 14, 3-24.
[47] Pericoli, M. and M. Taboga (2011), Bond risk premia, macroeconomic fundamentals and the exchange
rate, International Review of Economics & Finance 22(1), 42-65.
[48] Piazzesi, M. (2010), Affine Term Structure Models, Handbook of Financial Econometrics 12, 691-766,
Elsevier.
[49] Piazzesi, M. and M. Schneider (2007), Equilibrium Yield Curves, in Daron Acemoglu, Kenneth Rogoff,
and Michael Woodford, NBER Macroeconomics Annual 2006, Cambridge MA: MIT Press p. 389-442.
[50] Singleton, K. J. (2006), Empirical Dynamic Asset Pricing: Model Specification and Econometric Assessment, Princeton University Press 2006.
[51] Vasicek, O. (1977), An Equilibrium Characterization of the Term Structure, Journal of Financial
Economics 5, 177-188.
[52] Wu, T. (2006), Macro Factors and the Affine Term Structure of Interest Rates, Journal of Money,
Credit, and Banking 38(7), 1847-1875.
[53] Zellner, A. (1962), An Efficient Method of Estimating Seemingly Unrelated Regressions and Tests for
Aggregation, Journal of the American Statistical Association 57(298), 348-368.

29

Tables and Figures

Table 1: Summary Statistics of One-Year Yields

Country

Australia
Canada
Japan
Norway
Sweden
Switzerland
United Kingdom
United States

Mean

Median

Min.

Max.

Std. dev.

Skewness

Kurtosis

(12)

4.82
2.85
0.20
3.80
2.61
1.43
3.30
2.38

4.90
2.82
0.12
2.88
2.61
0.96
4.25
1.88

2.28
0.49
0.01
1.26
0.32
0.01
0.12
0.10

7.07
6.34
0.81
7.59
4.87
3.97
6.52
6.30

1.21
1.64
0.21
2.04
1.41
1.13
2.12
2.05

-0.42
0.30
1.34
0.43
-0.05
0.63
-0.37
0.44

2.48
1.99
3.54
1.64
1.61
2.08
1.51
1.76

0.35
0.66
0.56
0.66
0.58
0.56
0.74
0.67

Note: Monthly interest rate data for one-year yields (in %) from January 1999 to January 2014, data
source: Bloomberg, L.P. 2014. (12) refers to the autocorrelation function at the 12-month lag.

Table 2: Correlation of One-Year Yields

Country

AUL

CAN

JAP

NOR

SWE

SWI

UK

US

Australia
Canada
Japan
Norway
Sweden
Switzerland
United Kingdom
United States

1
0.71
0.50
0.53
0.70
0.71
0.81
0.70

1
0.29
0.76
0.82
0.77
0.92
0.95

1
0.26
0.30
0.61
0.30
0.36

1
0.92
0.80
0.72
0.64

1
0.81
0.85
0.69

1
0.77
0.77

1
0.89

Note: Correlation of monthly interest rate data for one-year yields from January
1999 to January 2014, data source: Bloomberg, L.P. 2014.

30

Table 3: Correlation of 3-Month Yields With Other Maturities

Country

AUL
CAN
JAP
NOR
SWE
SWI
UK
USA

6 months

1 year

2 years

3 years

4 years

5 years

6 years

7 years

8 years

9 years

10 years

1.00
1.00
0.99
1.00
1.00
1.00
1.00
1.00

0.98
0.99
0.97
0.98
0.99
0.99
0.99
0.99

0.93
0.96
0.89
0.96
0.95
0.94
0.98
0.98

0.88
0.93
0.81
0.93
0.91
0.92
0.96
0.96

0.85
0.90
0.73
0.91
0.88
0.89
0.95
0.94

0.82
0.88
0.64
0.90
0.86
0.87
0.93
0.91

0.80
0.85
0.56
0.88
0.84
0.84
0.92
0.90

0.77
0.83
0.46
0.87
0.83
0.82
0.91
0.88

0.75
0.80
0.40
0.86
0.81
0.80
0.89
0.86

0.73
0.78
0.39
0.84
0.80
0.79
0.86
0.84

0.72
0.77
0.37
0.84
0.79
0.78
0.84
0.81

3m
3m
3m
3m
3m
3m
3m
3m

Note: Correlation of 3-month yields with interest rates at different maturities from January 1999 to January 2014,
data source: Bloomberg, L.P. 2014.

Table 4: Summary Statistics of Exchange Rates Expressed in USD

Currency

AUD
CAD
JPY
NOK
SEK
CHF
GBP

Mean

Median

Min.

Max.

Std. dev.

Skewness

Kurtosis

(12)

0.78
0.84
0.01
0.15
0.13
0.84
1.66

0.76
0.86
0.01
0.16
0.14
0.82
1.61

0.49
0.62
0.01
0.11
0.09
0.56
1.41

1.10
1.06
0.01
0.20
0.17
1.27
2.08

0.17
0.14
0.00
0.02
0.02
0.18
0.17

0.11
-0.15
0.83
-0.31
-0.39
0.20
0.65

1.92
1.53
2.57
2.16
2.25
2.07
2.37

0.75
0.79
0.76
0.68
0.62
0.78
0.57

Note: Exchange rate data from January 1999 to January 2014, source: Bloomberg, L.P. 2014. All
currencies for Australia, Canada, Japan, Norway, Sweden, Switzerland, and the United Kingdom
are expressed in terms of the U.S. dollar. (12) refers to the autocorrelation function at lag 12.

Table 5: Summary Statistics of Interest Rate Factors

Median
Factors:
Australia
Canada
Japan
Norway
Sweden
Switzerland
United Kingdom
United States

Minimum

Maximum

(12)

Std. dev.

% Variance Explained

Total

0.6
0.7
-0.8
-0.5
0.6
-0.3
1.4
-0.1

0.1
0.1
0.2
0.1
0.0
0.0
0.1
-0.0

-0.0
0.0
0.1
0.0
0.0
-0.0
0.0
-0.0

-8.3
-5.4
-4.9
-5.5
-5.9
-5.5
-6.8
-5.4

-2.4
-2.1
-2.9
-1.8
-1.9
-1.8
-1.7
-1.6

-0.7
-0.5
-1.4
-0.5
-0.5
-0.5
-0.6
-0.4

5.7
6.9
7.9
6.2
5.8
7.0
5.4
7.1

1.7
1.4
3.1
1.5
1.6
1.9
1.4
1.3

0.6
0.4
0.8
0.7
0.5
1.3
0.4
0.4

3.4
3.4
3.1
3.4
3.4
3.4
3.4
3.4

0.9
0.9
1.4
0.7
0.8
0.8
0.7
0.7

0.2
0.2
0.4
0.2
0.2
0.3
0.2
0.2

0.4
0.8
0.5
0.7
0.7
0.7
0.7
0.7

0.1
0.4
0.4
0.1
0.1
0.3
0.3
0.3

-0.1
-0.1
-0.1
0.1
0.1
0.2
-0.3
-0.2

92.9
93.8
81.7
95.5
94.3
93.4
96.2
95.3

6.7
6.0
16.9
4.0
5.3
5.8
3.5
4.5

0.3
0.2
1.1
0.4
0.4
0.6
0.2
0.2

99.9
100
99.7
99.9
99.9
99.8
99.9
100

Note: Interest rate factors are the principal components extracted from each individual country. Principal component analysis is computed on the covariance matrix of standardized monthly interest rate data from January
1999 to January 2014. There are 12 yields per country (maturities 3 months, 6 months, and 1 to 10 years) and
181 observations per yield. 12 components are calculated but only the first 3 are reported. Mean of each factor is
0 by construction. (12) refers to the autocorrelation function at 12 months. % variance explained refers to the
percentage of variance explained by each factor, whereas Total accounts for the cumulative variation of the first
three factors combined.

31

Table 6: Parameters of Term Structure Model

Parameters:

Australia

Canada

Japan

Norway

Sweden

Switzerland

United Kingdom

United States

P
K0P
(S.E.)


P
K1P
+ I (S.E.)

P (S.E.)

0.69
-0.07
0.35

(0.61)
(0.23)
(0.09)

0.980
0.02
0.00

(0.02)
(0.01)
(0.00)

-0.12
0.94
0.01

(0.06)
(0.03)
(0.01)

-0.37
-0.22
0.78

(0.31)
(0.12)
(0.05)

0.86
-0.11
-0.02

(0.25)
(0.04)
(0.04)

0.31
-0.02

(0.07)
(0.02)

0.13

(0.01)

0.14
0.14
0.14

(0.22)
(0.10)
(0.05)

0.992
0.00
0.00

(0.01)
(0.00)
(0.00)

0.01
0.97
0.01

(0.04)
(0.02)
(0.01)

-0.14
-0.35
0.78

(0.20)
(0.09)
(0.05)

0.69
-0.11
-0.07

(0.07)
(0.03)
(0.02)

0.28
0.01

(0.01)
(0.01)

0.14

(0.01)

0.45
-0.09
0.04

(0.14)
(0.05)
(0.02)

0.943
0.01
0.01

(0.02)
(0.01)
(0.00)

0.01
0.95
0.00

(0.06)
(0.02)
(0.01)

-0.65
0.05
0.87

(0.23)
(0.08)
(0.04)

0.32
-0.04
0.00

(0.04)
(0.05)
(0.03)

0.10
-0.02

(0.01)
(0.01)

0.05

(0.00)

0.36
0.01
0.18

(0.23)
(0.15)
(0.07)

0.995
0.00
0.00

(0.01)
(0.01)
(0.00)

-0.02
0.91
0.01

(0.04)
(0.03)
(0.01)

-0.43
-0.32
0.83

(0.13)
(0.08)
(0.04)

0.69
-0.12
-0.05

(0.06)
(0.03)
(0.01)

0.42
0.01

(0.02)
(0.02)

0.20

(0.01)

0.15
0.02
0.10

(0.21)
(0.10)
(0.05)

0.993
0.01
0.00

(0.01)
(0.01)
(0.00)

-0.09
0.94
0.01

(0.05)
(0.02)
(0.01)

-0.42
-0.33
0.86

(0.17)
(0.08)
(0.04)

0.69
-0.11
-0.03

(0.06)
(0.04)
(0.03)

0.32
0.02

(0.01)
(0.01)

0.15

(0.00)

0.22
0.01
-0.08

(0.13)
(0.07)
(0.04)

0.996
0.00
-0.00

(0.01)
(0.01)
(0.00)

-0.02
0.96
0.00

(0.04)
(0.02)
(0.01)

0.41
0.14
0.84

(0.13)
(0.07)
(0.04)

0.48
-0.08
0.04

(0.03)
(0.02)
(0.01)

0.24
-0.04

(0.03)
(0.01)

0.14

(0.00)

0.89
-0.31
0.42

(0.37)
(0.21)
(0.10)

0.997
0.00
0.00

(0.01)
(0.00)
(0.00)

0.06
0.92
-0.00

(0.05)
(0.03)
(0.01)

-0.43
-0.02
0.74

(0.18)
(0.10)
(0.05)

0.66
-0.18
0.02

(0.14)
(0.07)
(0.05)

0.33
0.01

(0.02)
(0.01)

0.18

(0.00)

0.94
-0.15
0.18

(0.35)
(0.21)
(0.06)

0.992
0.00
0.00

(0.01)
(0.01)
(0.00)

0.05
0.92
0.00

(0.05)
(0.03)
(0.01)

-0.62
-0.19
0.84

(0.22)
(0.13)
(0.04)

0.78
-0.31
-0.03

(0.07)
(0.03)
(0.01)

0.35
-0.00

(0.05)
(0.03)

0.14

(0.02)

P,i
P,i
Note: Estimates and asymptotic standard errors in parentheses (S.E.) for country i from Pti = K0P
+ K1P

i
i P,i
Pt1 + P t , where Pt are the interest rate observable factors and P represents the physical distribution. All
P
P
+ I.
100 and P 100 are reported, as well as K1P
parameters are annualized, K0P

Table 7: Parameters of Term Structure Model and Fit of Yields

Parameters:

Australia
Canada
Japan
Norway
Sweden
Switzerland
United Kingdom

0P

-0.002
-0.013
-0.002
-0.008
-0.010
-0.009
-0.017



Q
eig K1P
+I

1P

0.265
0.346
0.053
0.383
0.323
0.343
0.346

0.579
0.344
0.402
0.475
0.458
0.443
0.338

0.825
1.126
0.821
0.831
1.002
-1.054
1.183
0

0.998
0.966
0.993
0.959
0.969
0.956
0.975

0.928
0.963
0.990
0.952
0.964
0.951
0.921

0.927
0.870
0.933
0.946
0.899
0.946
0.907

RMSE (yields)

0.049
0.109
0.047
0.104
0.110
0.118
0.094

Q,i
Q,i
i
Note: Estimates for country i from equations rti = i0P + i1P Pti and Pti = K0P
+ K1P
Pt1
+ iP Q,i
t , where rt is the
risk-free rate, Pt are the interest rate observable factors, and Q represents the risk-neutral distribution. All parameters
are annualized. Root mean square errors (RMSE) are in basis points.

32

Table 8: Model Fit for Exchange Rate Changes (in USD)

Horizon:

Australia

Canada

Japan

Norway

Sweden

Switzerland

United Kingdom

1-month

3-month

6-month

1-year

1.5-year

2-year

3-year

4-year

5-year

Robust
N.W.

0.492
(0.197)**
(0.228)**

0.668
(0.150)***
(0.204)***

0.827
(0.115)***
(0.199)***

0.977
(0.070)***
(0.131)***

1.018
(0.041)***
(0.066)***

1.023
(0.040)***
(0.046)***

1.024
(0.047)***
(0.062)***

1.033
(0.074)***
(0.072)***

1.099
(0.072)***
(0.056)***

R2

0.023

0.113

0.259

0.567

0.788

0.770

0.774

0.593

0.718

Robust
N.W.

0.466
(0.208)**
(0.197)**

0.528
(0.136)***
(0.181)***

0.646
(0.108)***
(0.186)***

0.961
(0.100)***
(0.216)***

1.080
(0.062)***
(0.125)***

1.055
(0.057)***
(0.091)***

1.029
(0.039)***
(0.066)***

0.990
(0.042)***
(0.065)***

1.038
(0.064)***
(0.062)***

R2

0.019

0.074

0.158

0.451

0.667

0.698

0.771

0.725

0.767

Robust
N.W.

0.890
(0.211)***
(0.224)***

0.859
(0.142)***
(0.213)***

0.875
(0.109)***
(0.183)***

1.004
(0.066)***
(0.147)***

1.033
(0.052)***
(0.114)***

1.040
(0.045)***
(0.097)***

1.005
(0.030)***
(0.060)***

0.958
(0.035)***
(0.051)***

1.026
(0.040)***
(0.076)***

R2

0.071

0.159

0.275

0.614

0.728

0.805

0.878

0.854

0.822

Robust
N.W.

0.515
(0.208)**
(0.201)**

0.644
(0.163)***
(0.248)**

0.851
(0.157)***
(0.267)***

1.147
(0.089)***
(0.161)***

1.098
(0.050)***
(0.090)***

1.090
(0.051)***
(0.047)***

1.029
(0.046)***
(0.038)***

1.054
(0.058)***
(0.051)***

1.007
(0.046)***
(0.027)***

R2

0.026

0.094

0.204

0.494

0.748

0.726

0.729

0.646

0.786

Robust
N.W.

0.523
(0.212)**
(0.212)**

0.874
(0.129)***
(0.190)***

0.960
(0.096)***
(0.181)***

1.003
(0.052)***
(0.106)***

1.025
(0.039)***
(0.090)***

1.096
(0.039)***
(0.081)***

1.123
(0.032)***
(0.047)***

0.985
(0.042)***
(0.036)***

0.994
(0.039)***
(0.062)***

R2

0.045

0.295

0.490

0.703

0.825

0.827

0.871

0.793

0.832

Robust
N.W.

0.413
(0.262)
(0.268)

0.682
(0.179)***
(0.233)***

0.793
(0.117)***
(0.202)***

0.923
(0.077)***
(0.137)***

0.994
(0.056)***
(0.092)***

1.089
(0.062)***
(0.090)***

1.008
(0.061)***
(0.084)***

1.087
(0.070)***
(0.062)***

0.921
(0.076)***
(0.066)***

R2

0.006

0.078

0.195

0.418

0.689

0.737

0.682

0.622

0.550

Robust
N.W.

0.577
(0.190)***
(0.174)***

0.821
(0.134)***
(0.217)***

0.862
(0.122)***
(0.226)***

0.925
(0.074)***
(0.144)***

0.997
(0.042)***
(0.090)***

0.985
(0.039)***
(0.058)***

1.008
(0.024)***
(0.033)***

1.007
(0.029)***
(0.033)***

1.039
(0.022)***
(0.019)***

R2

0.045

0.194

0.300

0.606

0.820

0.797

0.910

0.865

0.916

obs
Note: Estimates for sobs
t+k = + st+k + t+k , where st+k is the observed annualized log exchange rate change in
USD between t and t + k and st+k is the model-implied exchange rate change, modeled as the ratio of two countries
(log) pricing kernels, where the domestic country is the U.S. Robust and N.W. refer to robust (White) standard errors
and Newey-West standard errors, where the lag is equal to the horizon k + 1 (in months). The significance level of
the coefficient is determined by (*) given the corresponding method to estimate standard errors, where *** p<0.01,
** p<0.05, * p<0.1. R2 is adjusted for degrees of freedom.

33

Table 9: Model Fit for Exchange Rate Changes (Bias-Corrected Estimates)

Horizon:
Bias-Correction:

Australia

Canada

Japan

Norway

Sweden

Switzerland

United Kingdom

1-month
mean
median

6-month
mean
median

1-year
mean
median

3-year
mean
median

5-year
mean
median

N.W.

0.569**
(0.244)

0.566**
(0.243)

0.860***
(0.212)

0.858***
(0.212)

0.983***
(0.147)

0.983***
(0.146)

1.060***
(0.072)

1.058***
(0.072)

1.116***
(0.091)

1.115***
(0.089)

R2

0.022

0.022

0.229

0.230

0.510

0.513

0.773

0.773

0.670

0.672

N.W.

0.476**
(0.217)

0.461**
(0.199)

0.678***
(0.205)

0.628***
(0.188)

0.978***
(0.213)

0.956***
(0.218)

1.057***
(0.071)

1.038***
(0.073)

1.054***
(0.081)

1.042***
(0.079)

R2

0.014

0.017

0.153

0.147

0.438

0.438

0.752

0.756

0.763

0.759

N.W.

0.764***
(0.203)

0.813***
(0.213)

0.802***
(0.174)

0.826***
(0.178)

0.955***
(0.130)

0.964***
(0.133)

1.013***
(0.063)

1.014***
(0.063)

1.037***
(0.069)

1.038***
(0.069)

R2

0.062

0.065

0.261

0.266

0.608

0.608

0.871

0.874

0.824

0.825

N.W.

0.649**
(0.307)

0.652**
(0.308)

1.021***
(0.259)

1.021***
(0.259)

1.109***
(0.157)

1.109***
(0.156)

1.031***
(0.030)

1.032***
(0.029)

0.972***
(0.037)

0.972***
(0.037)

R2

0.048

0.048

0.356

0.357

0.571

0.572

0.824

0.824

0.809

0.810

N.W.

0.490***
(0.163)

0.497***
(0.166)

0.763***
(0.216)

0.785***
(0.214)

0.951***
(0.142)

0.958***
(0.139)

1.185***
(0.105)

1.178***
(0.093)

1.060***
(0.108)

1.052***
(0.101)

R2

0.043

0.044

0.315

0.333

0.591

0.606

0.803

0.814

0.753

0.762

N.W.

0.225
(0.250)

0.228
(0.248)

0.663***
(0.216)

0.660***
(0.215)

0.967***
(0.116)

0.966***
(0.114)

1.029***
(0.090)

1.028***
(0.089)

0.920***
(0.092)

0.920***
(0.094)

R2

0.000

0.000

0.162

0.162

0.436

0.438

0.719

0.719

0.574

0.575

N.W.

0.574***
(0.171)

0.576***
(0.171)

0.836***
(0.227)

0.836***
(0.226)

0.916***
(0.147)

0.917***
(0.146)

1.015***
(0.033)

1.015***
(0.033)

1.043***
(0.018)

1.042***
(0.018)

R2

0.042

0.043

0.279

0.279

0.584

0.586

0.910

0.909

0.912

0.912

obs
Note: Estimates for sobs
t+k = + st+k + t+k , where st+k is the observed annualized log exchange rate change in
USD between t and t + k and st+k is the model-implied exchange rate change, modeled as the ratio of two countries
(log) pricing kernels, where the domestic country is the U.S. N.W. refers to Newey-West standard errors, where the
lag is equal to the horizon k + 1 (in months). The significance level of the coefficient is determined by (*) where ***
p<0.01, ** p<0.05, * p<0.1. R2 is adjusted for degrees of freedom. Model is either mean or median-corrected to
account for small-sample bias, as suggested by Bauer, Rudebusch, and Wu (2012, 2014).

34

Table 10: Model Fit for Exchange Rate Changes (One- and Two-Factor Model)

Horizon:
Number of Factors:

Australia

Canada

Japan

Norway

Sweden

Switzerland

United Kingdom

1-month
1-factor
2-factor

6-month
1-factor
2-factor

1-year
1-factor
2-factor

3-year
1-factor
2-factor

5-year
1-factor
2-factor

Robust
N.W.

0.543
(0.394)
(0.398)

0.679
(0.291)**
(0.271)**

0.799
(0.139)***
(0.257)***

0.893
(0.125)***
(0.218)***

0.947
(0.099)***
(0.245)***

1.017
(0.084)***
(0.166)***

1.082
(0.087)***
(0.270)***

1.258
(0.061)***
(0.133)***

1.024
(0.127)***
(0.183)***

1.145
(0.098)***
(0.193)***

R2

0.005

0.015

0.094

0.142

0.254

0.339

0.537

0.713

0.316

0.474

Robust
N.W.

0.173
(0.462)
(0.477)

0.539
(0.240)**
(0.214)**

0.486
(0.177)***
(0.366)

0.846
(0.135)***
(0.237)***

0.663
(0.118)***
(0.274)**

1.181
(0.116)***
(0.262)***

1.032
(0.060)***
(0.121)***

1.120
(0.053)***
(0.128)***

1.249
(0.098)***
(0.126)***

1.155
(0.094)***
(0.143)***

R2

0.000

0.013

0.021

0.159

0.099

0.417

0.473

0.684

0.640

0.706

Robust
N.W.

0.740
(0.277)***
(0.303)**

0.560
(0.207)***
(0.220)**

0.700
(0.159)***
(0.308)**

0.558
(0.121)***
(0.227)**

0.771
(0.128)***
(0.360)**

0.690
(0.080)***
(0.207)***

0.987
(0.068)***
(0.238)***

1.037
(0.042)***
(0.057)***

1.057
(0.067)***
(0.180)***

1.149
(0.070)***
(0.196)***

R2

0.028

0.027

0.110

0.115

0.218

0.283

0.471

0.624

0.668

0.658

Robust
N.W.

0.130
(1.448)
(1.336)

0.965
(0.374)**
(0.433)**

0.118
(0.551)
(1.144)

1.125
(0.236)***
(0.418)***

0.344
(0.397)
(0.978)

1.269
(0.118)***
(0.196)***

0.619
(0.195)***
(0.360)*

0.990
(0.070)***
(0.123)***

1.606
(0.258)***
(0.231)***

0.985
(0.049)***
(0.067)***

R2

0.000

0.044

0.000

0.202

0.000

0.388

0.033

0.568

0.245

0.732

Robust
N.W.

0.023
(0.595)
(0.669)

0.815
(0.299)***
(0.314)**

0.519
(0.295)*
(0.608)

0.882
(0.151)***
(0.287)***

0.695
(0.216)***
(0.569)

1.033
(0.098)***
(0.175)***

1.166
(0.178)***
(0.693)*

1.186
(0.074)***
(0.227)***

1.473
(0.355)***
(0.811)*

1.073
(0.053)***
(0.098)***

R2

0.000

0.042

0.012

0.194

0.050

0.397

0.244

0.627

0.211

0.703

Robust
N.W.

1.197
(1.006)
(0.751)

0.528
(0.377)
(0.401)

1.296
(0.308)***
(0.613)**

0.596
(0.185)***
(0.327)*

1.332
(0.239)***
(0.571)**

0.930
(0.128)***
(0.264)***

1.038
(0.142)***
(0.406)**

1.025
(0.110)***
(0.348)***

0.437
(0.177)**
(0.325)

1.163
(0.111)***
(0.116)***

R2

0.007

0.006

0.091

0.063

0.182

0.245

0.296

0.444

0.033

0.451

Robust
N.W.

-0.726
(0.718)
(0.810)

0.644
(0.252)**
(0.222)***

-0.258
(0.356)
(0.774)

0.711
(0.150)***
(0.254)***

0.165
(0.238)
(0.681)

0.950
(0.105)***
(0.228)***

0.785
(0.163)***
(0.648)

1.093
(0.044)***
(0.103)***

1.700
(0.255)***
(0.528)***

1.114
(0.034)***
(0.053)***

R2

0.000

0.027

0.000

0.118

0.000

0.347

0.105

0.780

0.451

0.841

obs
Note: Estimates for sobs
t+k = + st+k + t+k , where st+k is the observed annualized log exchange rate change in
USD between t and t + k and st+k is the model-implied exchange rate change, modeled as the ratio of two countries
(log) pricing kernels, where the domestic country is the U.S. Robust and N.W. refer to robust (White) standard errors
and Newey-West standard errors, where the lag is equal to the horizon k + 1 (in months). The significance level of
the coefficient is determined by (*) given the corresponding method to estimate standard errors, where *** p<0.01,
** p<0.05, * p<0.1. R2 is adjusted for degrees of freedom. The 1-factor model is estimated with the first factor
extracted from interest rate data in each country (level), and the 2-factor model, with the first two factors (level and
slope).

35

Table 11: Model Fit for Exchange Rate Changes (in GBP)

Horizon:

Australia

Canada

Japan

Norway

Sweden

Switzerland

United States

1-month

3-month

6-month

1-year

1.5-year

2-year

3-year

4-year

5-year

Robust
N.W.

0.720
(0.391)*
(0.291)**

0.823
(0.154)***
(0.183)***

0.882
(0.114)***
(0.169)***

0.989
(0.060)***
(0.070)***

0.998
(0.039)***
(0.042)***

1.003
(0.047)***
(0.056)***

1.003
(0.035)***
(0.043)***

0.993
(0.031)***
(0.034)***

1.014
(0.031)***
(0.022)***

R2

0.038

0.159

0.332

0.619

0.761

0.791

0.854

0.883

0.892

Robust
N.W.

1.130
(0.282)***
(0.239)***

1.110
(0.169)***
(0.212)***

1.092
(0.114)***
(0.172)***

1.096
(0.061)***
(0.099)***

1.033
(0.052)***
(0.078)***

0.971
(0.050)***
(0.051)***

1.021
(0.039)***
(0.031)***

1.001
(0.036)***
(0.034)***

1.028
(0.035)***
(0.022)***

R2

0.083

0.219

0.327

0.644

0.722

0.752

0.853

0.877

0.867

Robust
N.W.

0.361
(0.158)**
(0.147)**

0.631
(0.120)***
(0.153)***

0.728
(0.108)***
(0.169)***

0.920
(0.069)***
(0.130)***

1.038
(0.054)***
(0.105)***

1.037
(0.040)***
(0.061)***

1.020
(0.022)***
(0.029)***

1.014
(0.020)***
(0.024)***

1.016
(0.019)***
(0.027)***

R2

0.027

0.173

0.304

0.623

0.794

0.852

0.899

0.928

0.926

Robust
N.W.

0.558
(0.226)**
(0.238)**

0.490
(0.142)***
(0.215)**

0.625
(0.123)***
(0.241)**

1.090
(0.083)***
(0.145)***

1.093
(0.058)***
(0.081)***

1.051
(0.050)***
(0.060)***

0.965
(0.053)***
(0.061)***

1.056
(0.049)***
(0.063)***

0.998
(0.044)***
(0.047)***

R2

0.034

0.072

0.159

0.497

0.694

0.707

0.704

0.770

0.807

Robust
N.W.

0.438
(0.268)
(0.239)*

0.563
(0.137)***
(0.151)***

0.582
(0.091)***
(0.142)***

0.912
(0.094)***
(0.175)***

1.078
(0.095)***
(0.212)***

1.082
(0.099)***
(0.234)***

1.052
(0.064)***
(0.143)***

0.991
(0.059)***
(0.067)***

1.056
(0.044)***
(0.049)***

R2

0.009

0.071

0.129

0.315

0.435

0.465

0.616

0.676

0.827

Robust
N.W.

0.468
(0.269)*
(0.206)**

0.713
(0.157)***
(0.228)***

0.778
(0.112)***
(0.206)***

0.865
(0.066)***
(0.151)***

0.967
(0.065)***
(0.122)***

0.999
(0.039)***
(0.063)***

0.965
(0.033)***
(0.053)***

1.015
(0.031)***
(0.034)***

1.016
(0.022)***
(0.026)***

R2

0.019

0.142

0.287

0.471

0.694

0.814

0.844

0.912

0.929

Robust
N.W.

0.576
(0.188)***
(0.173)***

0.821
(0.134)***
(0.218)***

0.862
(0.122)***
(0.227)***

0.925
(0.074)***
(0.144)***

0.997
(0.042)***
(0.090)***

0.986
(0.039)***
(0.058)***

1.009
(0.024)***
(0.033)***

1.007
(0.029)***
(0.033)***

1.039
(0.022)***
(0.019)***

R2

0.045

0.194

0.300

0.606

0.819

0.796

0.910

0.865

0.916

obs
Note: Estimates for sobs
t+k = + st+k + t+k , where st+k is the observed annualized log exchange rate change in
GBP between t and t + k and st+k is the model-implied exchange rate change, modeled as the ratio of two countries
(log) pricing kernels, where the domestic country is the U.K. Robust and N.W. refer to robust (White) standard errors
and Newey-West standard errors, where the lag is equal to the horizon k + 1 (in months). The significance level of
the coefficient is determined by (*) given the corresponding method to estimate standard errors, where *** p<0.01,
** p<0.05, * p<0.1. R2 is adjusted for degrees of freedom.

36

Table 12: Uncovered Interest Parity Regressions

Horizon:

1-month
U IP

3-month

IRD

RP

U IP

IRD

6-month
RP

U IP

IRD

RP

Australia

N.W.
R2

-0.770
(2.289)
0.000

0.727
0.502**
(2.469)
(0.249)
0.018

-0.942
(2.147)
0.000

1.218
0.686***
(2.181)
(0.233)
0.111

-1.351
(2.091)
0.004

1.530
0.853***
(1.812)
(0.226)
0.260

Canada

N.W.
R2

0.733
(2.514)
0.000

1.444
0.475**
(2.479)
(0.201)
0.014

1.150
(2.635)
0.000

1.932
0.538***
(2.486)
(0.184)
0.075

0.842
(2.750)
0.000

1.751
0.654***
(2.268)
(0.191)
0.160

Japan

N.W.
R2

-0.526
(1.279)
0.000

1.128
0.876***
(1.264)
(0.235)
0.062

-0.492
(1.215)
0.000

1.136
0.860***
(1.164)
(0.224)
0.152

-0.588
(1.180)
0.001

1.048
0.877***
(1.000)
(0.194)
0.268

Norway

N.W.
R2

0.177
(1.717)
0.000

0.640
0.514**
(1.588)
(0.203)
0.020

0.309
(1.737)
0.000

1.047
0.643**
(1.573)
(0.253)
0.089

0.088
(1.636)
0.000

1.265
0.855***
(1.345)
(0.276)
0.201

Sweden

N.W.
R2

-0.162
(1.987)
0.000

0.201
0.524**
(1.687)
(0.212)
0.039

-0.309
(1.974)
0.000

0.508
0.870***
(1.292)
(0.193)
0.289

-0.515
(1.989)
0.000

0.575
0.954***
(1.082)
(0.186)
0.485

Switzerland

N.W.
R2

-1.216
(1.802)
0.000

-0.597
0.388
(1.694)
(0.274)
0.001

-1.452
(1.628)
0.006

-0.288
0.651***
(1.361)
(0.235)
0.075

-1.360
(1.553)
0.015

0.145
0.770***
(1.190)
(0.210)
0.194

United Kingdom

N.W.
R2

1.827
(2.414)
0.000

2.109
0.566***
(2.168)
(0.169)
0.042

1.836
(2.509)
0.007

2.322
0.810***
(2.078)
(0.208)
0.197

1.341
(2.552)
0.005

2.042
0.855***
(1.854)
(0.218)
0.303

Horizon:

1-year
U IP

2-year

IRD

RP

U IP

IRD

3-year
RP

U IP

IRD

RP

Australia

N.W.
R2

-2.264
(1.936)
0.049

1.289
0.990***
(1.138)
(0.157)
0.566

-3.206**
(1.511)
0.184

0.698*
1.000***
(0.382)
(0.041)
0.772

-2.921***
(1.112)
0.251

0.768*** 0.996***
(0.269)
(0.048)
0.774

Canada

N.W.
R2

0.134
(2.477)
0.000

1.858
0.969***
(1.378)
(0.215)
0.455

-1.095
(2.205)
0.014

1.290** 1.058***
(0.495)
(0.093)
0.699

-1.523
(1.693)
0.042

1.193*** 1.032***
(0.358)
(0.073)
0.771

Japan

N.W.
R2

-0.917
(1.167)
0.026

0.982
1.002***
(0.640)
(0.154)
0.611

-1.424
(1.005)
0.112

0.834*** 1.021***
(0.282)
(0.096)
0.804

-1.492*
(0.835)
0.174

0.811*** 0.984***
(0.192)
(0.059)
0.878

Norway

N.W.
R2

-0.337
(1.403)
0.000

1.393* 1.153***
(0.728)
(0.171)
0.491

-0.527
(1.093)
0.014

1.056*** 1.084***
(0.210)
(0.053)
0.724

-0.356
(0.801)
0.011

0.963*** 1.017***
(0.087)
(0.038)
0.726

Sweden

N.W.
R2

-1.091
(1.948)
0.014

0.381
0.992***
(0.687)
(0.112)
0.705

-1.657
(1.940)
0.070

0.518
1.069***
(0.393)
(0.069)
0.832

-1.752
(1.481)
0.141

0.870*** 1.095***
(0.243)
(0.049)
0.869

Switzerland

N.W.
R2

-1.545
(1.322)
0.045

0.367
0.896***
(0.677)
(0.142)
0.420

-1.442
(1.017)
0.083

0.792*** 1.065***
(0.245)
(0.081)
0.737

-1.437**
(0.600)
0.169

0.597*** 0.954***
(0.150)
(0.078)
0.688

United Kingdom

N.W.
R2

-0.061
(2.538)
0.000

1.179
0.923***
(1.202)
(0.144)
0.604

-1.768
(2.125)
0.043

1.037*
0.984***
(0.592)
(0.068)
0.795

-3.613**
(1.425)
0.232

1.029*** 1.007***
(0.246)
(0.036)
0.910

Note:
The first column reports estimates for:
sobs
t+k
The second and third columns report the estimates for:
RP

 h
1
2

dt+k1 dt+k1 ft+k1 ft+k1

i

f
d
=
+ U IP rt+k1
rt+k1
+ t+k .


f
d
sobs
= + IRD rt+k1
rt+k1
+
t+k

+ t+k . N.W. refers to Newey-West standard errors where the lag is

equal to the horizon k + 1 (in months). Significance level: *** p<0.01, ** p<0.05, * p<0.1. R2 adjusted for degrees
of freedom.

37

Table 13: Fama Conditions for the Risk Premium

Condition I
var (IRD)

Country

<

Condition II

var (RP )

corr (IRD, RP ) < 0

Australia

1-month
6-month
1-year
2-year
3-year

0.02
0.83
3.03
9.36
15.90

2.62
61.54
176.33
396.59
448.62

-0.29
-0.38
-0.45
-0.57
-0.67

Canada

1-month
6-month
1-year
2-year
3-year

0.01
0.23
0.80
2.29
3.29

1.12
24.28
65.28
152.03
247.74

-0.12
-0.13
-0.19
-0.27
-0.30

Japan

1-month
6-month
1-year
2-year
3-year

0.04
1.47
5.60
18.27
30.94

1.22
31.86
100.85
270.38
386.22

-0.35
-0.39
-0.41
-0.41
-0.30

Norway

1-month
6-month
1-year
2-year
3-year

0.04
1.45
5.34
16.61
27.13

1.90
32.92
81.76
188.84
240.73

-0.13
-0.29
-0.40
-0.48
-0.53

Sweden

1-month
6-month
1-year
2-year
3-year

0.03
1.00
3.53
9.52
12.45

3.11
64.99
167.28
286.97
279.28

-0.07
-0.14
-0.21
-0.37
-0.47

Switzerland

1-month
6-month
1-year
2-year
3-year

0.02
0.79
2.79
7.83
11.86

1.02
25.06
76.19
162.28
202.23

-0.24
-0.33
-0.37
-0.35
-0.27

United Kingdom

1-month
6-month
1-year
2-year
3-year

0.01
0.43
1.53
4.46
7.49

1.37
30.19
92.62
215.58
288.39

-0.05
-0.09
-0.16
-0.36
-0.56

Note: Fama Conditions for the risk premium: (I) variance of IRD < variance of RP , and (II) correlation
f
d
between IRD and RP < 0, where IRDk = rt+k1
rt+k1
is the interest rate differential and RPk =

dt+k1 dt+k1 ft+k1 ft+k1

100.
1
2

is the risk premium at horizon k in months. Variance reported

38

Table 14: Model Fit for Exchange Rate Changes (Countries Under the Euro)

Horizon:

Austria

Belgium

Finland

France

Germany

Ireland

Italy

Netherlands

Portugal

Spain

1-month

3-month

6-month

1-year

1.5-year

2-year

3-year

4-year

5-year

N.W.

0.495**
(0.242)

0.687***
(0.198)

0.753***
(0.176)

0.907***
(0.138)

0.982***
(0.112)

1.089***
(0.083)

1.083***
(0.060)

1.023***
(0.047)

0.974***
(0.037)

R2

0.014

0.094

0.197

0.444

0.701

0.741

0.807

0.816

0.838

N.W.

0.433*
(0.225)

0.689***
(0.194)

0.846***
(0.190)

1.012***
(0.136)

1.076***
(0.093)

1.117***
(0.083)

1.049***
(0.067)

0.986***
(0.044)

0.953***
(0.023)

R2

0.010

0.091

0.230

0.504

0.782

0.779

0.810

0.818

0.870

N.W.

0.531**
(0.223)

0.467**
(0.211)

0.531***
(0.176)

0.845***
(0.124)

1.047***
(0.096)

1.024***
(0.115)

1.108***
(0.075)

1.016***
(0.069)

1.014***
(0.030)

R2

0.022

0.048

0.112

0.403

0.724

0.648

0.777

0.750

0.853

N.W.

0.523**
(0.257)

0.691***
(0.239)

0.777***
(0.215)

0.983***
(0.148)

1.041***
(0.102)

1.060***
(0.106)

1.071***
(0.091)

1.009***
(0.061)

0.960***
(0.025)

R2

0.015

0.085

0.189

0.474

0.746

0.727

0.798

0.796

0.863

N.W.

0.524**
(0.265)

0.739***
(0.236)

0.818***
(0.208)

0.994***
(0.143)

1.044***
(0.108)

1.085***
(0.105)

1.071***
(0.090)

1.001***
(0.061)

0.960***
(0.025)

R2

0.014

0.091

0.194

0.458

0.724

0.727

0.796

0.793

0.861

N.W.

0.484*
(0.246)

0.673***
(0.206)

0.737***
(0.195)

0.826***
(0.155)

0.952***
(0.115)

1.084***
(0.089)

1.053***
(0.110)

1.062***
(0.063)

0.972***
(0.030)

R2

0.019

0.104

0.206

0.392

0.687

0.792

0.758

0.811

0.856

N.W.

0.505***
(0.158)

0.662***
(0.187)

0.830***
(0.162)

0.969***
(0.109)

1.004***
(0.080)

1.046***
(0.058)

1.078***
(0.055)

1.015***
(0.028)

0.979***
(0.028)

R2

0.028

0.123

0.299

0.583

0.817

0.808

0.857

0.833

0.883

N.W.

0.492*
(0.254)

0.725***
(0.227)

0.810***
(0.208)

0.988***
(0.147)

1.029***
(0.108)

1.061***
(0.104)

1.061***
(0.085)

0.999***
(0.060)

0.948***
(0.026)

R2

0.013

0.095

0.203

0.475

0.738

0.735

0.802

0.799

0.858

N.W.

0.223
(0.209)

0.488**
(0.210)

0.655***
(0.168)

0.874***
(0.112)

0.970***
(0.065)

0.981***
(0.084)

1.091***
(0.043)

1.026***
(0.032)

1.007***
(0.034)

R2

0.002

0.069

0.195

0.505

0.827

0.759

0.873

0.794

0.871

N.W.

0.371
(0.250)

0.675***
(0.217)

0.770***
(0.191)

0.996***
(0.113)

1.050***
(0.068)

1.080***
(0.049)

1.082***
(0.050)

0.997***
(0.030)

0.998***
(0.017)

R2

0.009

0.101

0.210

0.512

0.817

0.810

0.865

0.796

0.886

obs
Note: Estimates for sobs
t+k = + st+k + t+k , where st+k is the observed annualized log exchange rate change
in USD between t and t + k and st+k is the model-implied exchange rate change, modeled as the ratio of two
countries (log) pricing kernels, where the domestic country is the U.S. N.W. refers to Newey-West standard errors,
where the lag is equal to the horizon k + 1 (in months). The significance level of the coefficient is determined by
(*), where *** p<0.01, ** p<0.05, * p<0.1. R2 is adjusted for degrees of freedom.

39

Figure 1: Evolution of Yield Curves (1999-2014)

(a) Australia

(b) Canada

(c) Japan

(d) Norway

(e) Sweden

(f) Switzerland

(g) United Kingdom

(h) United States

Note: Monthly zero coupon yields from January 1999 to January 2014, data source: Bloomberg, L.P. 2014.
Yields are in percentage for maturities of 3 months, 6 months, and 1 to 10 years for each country.

40

Figure 2: Interest Rate Factor Loadings

(a) Australia

(b) Canada

(c) Japan

(d) Norway

(e) Sweden

(f) Switzerland

(g) United Kingdom

(h) United States

Note: Interest Rate Factor Loadings are the weights that principal components extracted from each
individual country have on the corresponding maturity. On the vertical axis, the loading scores; on the
horizontal axis, each maturity. The red solid line represents the first factor loading, the blue circles
represent the second factor loading, and the green dashes represent the third factor loading. Principal
component analysis is computed on the covariance matrix of standardized monthly interest rate data from
January 1999 to January 2014. There are 12 yields per country (maturities 3 months, 6 months, and 1 to
10 years) and 181 observations per yield. 12 components are calculated but only the first 3 are reported.

41

Figure 3: One-Month Exchange Rate Changes (in USD)

(a) Australian Dollar

(b) Canadian Dollar

(c) Japanese Yen

(d) Norwegian Krone

(e) Swedish Krona

(f) Swiss Franc

(g) British Pound

Note: The blue line represents the annualized changes in the log exchange
rate in terms of the USD, the red circles represent the model-implied exf
d
pected exchange rate changes, where Et st+k = rt+k1
rt+k1
+
1
2

h 0
i
0
dt+k1 dt+k1 ft+k1 ft+k1 .

42

Figure 4: One-Year Exchange Rate Changes (in USD)

(a) Australian Dollar

(b) Canadian Dollar

(c) Japanese Yen

(d) Norwegian Krone

(e) Swedish Krona

(f) Swiss Franc

(g) British Pound

Note: The blue line represents the annualized changes in the log exchange
rate in terms of the USD, the red circles represent the model-implied exf
d
pected exchange rate changes, where Et st+k = rt+k1
rt+k1
+
1
2

h 0
i
0
dt+k1 dt+k1 ft+k1 ft+k1 .

43

Figure 5: Five-Year Exchange Rate Changes (in USD)

(a) Australian Dollar

(b) Canadian Dollar

(c) Japanese Yen

(d) Norwegian Krone

(e) Swedish Krona

(f) Swiss Franc

(g) British Pound

Note: The blue line represents the annualized changes in the log exchange
rate in terms of the USD, the red circles represent the model-implied exf
d
pected exchange rate changes, where Et st+k = rt+k1
rt+k1
+
1
2

h 0
i
0
dt+k1 dt+k1 ft+k1 ft+k1 .

44

Figure 6: Evolution of Yield Curves for Countries Under the Euro (1999-2014)

(a) Austria

(b) Belgium

(c) Finland

(d) France

(e) Germany

(f) Ireland

(g) Italy

(h) Netherlands

(i) Portugal

(j) Spain

Note: Monthly zero coupon yields from January 1999 to January 2014, data
source: Bloomberg, L.P. 2014. Yields are in percentage for maturities of 3
months, 6 months, and 1 to 10 years for each country.

45

Appendix: Details of the Term Structure Model

Observable factors can be linked to an N -dimensional latent state vector Xt through invariant transformations as in Joslin, Singleton, and Zhu (2011), where the risk-free rate is given by

rt = 0X + 01X Xt ,

(26)

where 0X is a constant and 1X is an (N 1) vector. Given equation (26), the entire term structure
of zero coupon bonds with maturity m (in years) can also be expressed in terms of Xt . The linearity in
the factors allows for closed-form solutions to the prices of bonds, in order to avoid solution methods for
partial differential equations or Monte Carlo methods to estimate yields, which are undesirable given their
computational cost and limitations. This context assumes that no-arbitrage holds, which makes the time
evolution of yields and the cross-sectional shape of the yield curve consistent at any point in time. By
the fundamental theorem of asset pricing, the condition of no-arbitrage is equivalent to the existence of a
risk-neutral measure, Q. Therefore, the price of an m-maturity zero coupon bond at time t, Ptm , can be
expressed as the expected value of the future payoffs (in this case = 1) discounted at the risk-free rate, under
 
risk-neutrality EQ
t :
"
!#
m1
X
Q
m
m
Pt = Et exp
rt+i
= exp [am
(27)
X + bX Xt ] ,
i=0
0
0
m
where am
X and bX solve first-order Riccati difference equations with initial conditions aX = 0 and bX = 0 as

in Duffie and Kan (1996). This allows one to model the yields at different horizons as a linear function of
m
m
m
the state variable Xt , where the yield loadings are given by Am
X = aX /m and BX = bX /m,

m
ytm = Am
X + BX Xt .

(28)

The evolution of the unobserved factors is expressed under two equivalent probability measures: P and
Q. P is the probability measure of the random process X, referred to as the actual or physical distribution
of the state vector. This process can be represented, for example, by an unrestricted vector auto-regressive
process that describes the evolution of the yields. Q is an equivalent martingale measure,10 referred to as
the risk-neutral distribution of X. Under this distribution, the process has already been adjusted for risk,
allowing for a risk-neutral representation of the movement of factors over time. The state unobserved vector
then evolves as an auto-regressive process under the two distributions:
10 Defined

as a finite stochastic process in which each realization is independent from the previous one.

46

P
P
Xt = K0X
+ K1X
Xt1 + X Pt ,

(29)

Q
Q
Xt = K0X
+ K1X
Xt1 + X Q
t ,

(30)





Q
Q
P
P
where K0X
, K0X
are (N 1) constant vectors, K1X
, K1X
are (N N ) matrices, and errors are i.i.d
0
Gaussian, Pt , Q
t N (0, IN ). X is a lower triangular matrix such that X X is the (N N ) variance-

covariance matrix of Xt .
The estimations can be simplified by imposing identification conditions that yield an observationally
equivalent Gaussian dynamic term structure model. Let 0X = 0 and 1X = [1

1]0 so that the

risk-free rate is the sum of the factors: rt = Xt , where is a vector of ones, and let X be lower triangular
Q
is then parameterized as a diagonal matrix with ordered Q eigenvalues. Finally,
with positive diagonal. K1X
Q
Q
Q
K0X
can be normalized by K0X
= [ k

0
Q
0 ] , where k is proportional to the long run mean

Q
of the short rate whenever the model is stationary under Q.11 Given Q and k
, there exists a unique

set of parameters consistent with no-arbitrage. This invariant transformation of the model proposed by
Joslin, Singleton, and Zhu (2011) simplifies the estimation process and significantly improves computational
performance. As defined by Dai and Singleton (2000), this invariant property is an arbitrary combination of
transformations that are possible because the factors are unobserved, the model is linear in the factors, and
the admissibility condition is preserved (i.e., the model leads to well defined solutions to the bond pricing
equation). Hence the identification conditions preserve the short rate and the bonds unchanged.
Let the observable factors be defined as Pt = W yt . From equation (28), Pt can be expressed as a
function of the latent vector Xt

m
m
m
Pt = W [Am
X + BX Xt ] = W AX + W BX Xt .

(31)

Equivalently,

m
Xt = (W BX
)

(Pt W Am
X) .

(32)

Now, the yields can be expressed in terms of observable factors by substituting (32) into (28):

h
i
m
m 1
y t = Am
(Pt W Am
X + BX (W BX )
X) ,

(33)

11 Note that this identification assumes that the factors are stationary under the risk-neutral distribution, which may not be
true. It is not uncommon for the estimated roots to be close to 0, given the possibility of a unit root in the pricing factors. JSZ
show that an equivalent re-normalization of K0Q and 0 allows for the preservation of the canonical form in the presence of zero
or negative Q roots as well. Refer to Joslin, Singleton, and Zhu (2011) and their online appendix for more details.

47

h
i
m
m
m 1
m
m
m 1
m
where Am
W and BP
= BX
(W BX
) , such that yt = Am
P = AX I BX (W BX )
P + BP P t .
The risk-free rate can also be expressed in terms of observable factors:

rt = Xt

(34)

h
i
m 1
rt = (W BX
) (Pt W Am
X) ,

(35)

h
i
m 1
m 1
where 0P = (W BX
) W Am
, such that rt = 0P + 01P Pt .
X and 1P = (W BX )
P
P
Finally, equation (29) can be expressed as Pt = K0P
+ K1P
Pt1 + P Pt , where

P
P
m
P
K0P
= K0X
(W BX
) K1P
(W Am
X) ,

P
m
P
m
K1P
= (W BX
) K1X
(W BX
)

m
m
P 0P = (W BX
) X 0X (W BX
).

(36)

(37)

(38)

Similarly for the evolution of factors under the risk-neutral distribution, equation (30).
Given the introduction of observable factors, the model can be now summarized by the following equations:
rt = 0P + 01P Pt ,

(39)

m
ytm = Am
P + BP P t ,

(40)

P
P
Pt = K0P
+ K1P
Pt1 + P Pt ,

(41)

Q
Q
Pt = K0P
+ K1P
Pt1 + P Q
t ,

(42)

m
where internal consistency conditions are satisfied, i.e., W Am
P = 0 and W BP = IN . The yield coefficients

are rotated so that

h
i
m
m
m 1
Am
W ,
P = AX I BX (W BX )
1

m
m
m
BP
= BX
(W BX
)

(43)
(44)

n
o
Q
Q
The parameters under the risk-neutral distribution, Q = K0P
, K1P
, P , 0P , 1P , can be expressed
Q
in terms of Q and k
given the mapping between Xt and Pt , and the yield coefficients determined by

48

m
recursive rules {Am
X , BX } along with the weighting matrix W :

Q
m
P
m
K1P
= (W BX
) K1X
(W BX
)

Q
m
K0P
= (W BX
)

0


Q
k

m
1P = (W BX
)

(45)

Q
K1P
(W Am
X) ,

(46)

(47)

0P = 1P (W Am
X) .

(48)

The Joslin, Singleton, and Zhu (2011) transformation allows for a significant reduction in the number of
parameters to be estimated by Maximum Likelihood, since the risk-neutral distribution can then be fully


Q
specified by Q = Q , k
, P . This separation in the parameters estimated under the different probability
measures is a key property of the model, which I preserve when extending to a two-country framework and
integrating exchange rates.

49