Accepted Manuscript

Assessing the predictive ability of sovereign default risk on exchange rate re-
turns

Claudia Foroni, Francesco Ravazzolo, Barbara Sadaba

PII: S0261-5606(17)30239-5
DOI: https://doi.org/10.1016/j.jimonfin.2017.12.001
Reference: JIMF 1867

To appear in: Journal of International Money and Finance

Received Date: 17 March 2017

Accepted Date: 4 December 2017

Please cite this article as: C. Foroni, F. Ravazzolo, B. Sadaba, Assessing the predictive ability of sovereign default
risk on exchange rate returns, Journal of International Money and Finance (2017), doi: https://doi.org/10.1016/
j.jimonfin.2017.12.001

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
 Assessing the predictive ability of sovereign default risk on
exchange rate returns∗

Claudia Foroni1 , Francesco Ravazzolo2 , and Barbara Sadaba3

1
Deutsche Bundesbank
2
Free University of Bozen/Bolzano and BI Norwegian Business School
3
Bank of Canada

October 17, 2017

∗
The views expressed in this paper are those of the authors. No responsibility for them should be attributed
to the Bank of Canada or the Deutsche Bundesbank. We thank for their useful comments on a previous version
the Editor, two Referees, Michiel de Pooter, Barbara Rossi, Marcin Jaskowski, Gerdie Everaert, Job Swank,
Louis Morel and conference and seminar participants at Bank of Canada, the Center for Operations Research and
Econometrics (CORE) at Universite’ catholique de Louvain, Norges Bank, the 23rd Symposium of the Society
for Nonlinear Dynamics and Econometrics, the 3rd Annual Conference of International Association for Applied
Econometrics, the 21st International Conference on Computing in Economics and Finance and the 69th European
Meeting of the Econometric Society, for very useful comments.

1
Assessing the predictive ability of sovereign default risk on
exchange rate returns

Abstract

Increased sovereign credit risk is often associated with sharp currency movements. There-
fore, expectations of the probability of a sovereign default event can convey important in-
formation regarding future movements of exchange rates. In this paper, we investigate the
possible pass-through of risk in the sovereign debt markets to currency markets by proposing
a new risk premium factor for predicting exchange rate returns based on sovereign default
risk. We compute it from the term structure at different maturities of sovereign credit de-
fault swaps and conduct an out-of-sample forecasting exercise to test whether we can improve
upon the benchmark random walk model. Our results show that the inclusion of the default
risk factor improves the forecasting accuracy upon the random walk model and alternative
models provide less accurate predictions.
JEL Classification: C22, C52, C53, F31.
Keywords: Exchange Rates, Forecasting, Sovereign Risk, CDS, Term Structure Models

1
1 Introduction

Since the seminal work of Meese and Rogoff (1983a,b, 1988), exchange rates have been known

to be nearly impossible to predict. Later attempts to beat the random walk (RW) benchmark

using economic variables derived from theoretical exchange rate models find a general failure of

these models, demonstrating that the so-called “Meese and Rogoff” puzzle is not easily solved

(see Rossi (2013) for a review of the literature). However, in the same studies, Meese and Rogoff

(1983a,b) provide a potential explanation for their results: time-varying risk premia could be an

important determinant of their findings. This view receives further support in subsequent works.

Hodrick (1989) argues that risk premia should be included in rational expectations models that

study exchange rates as an asset market equilibrium. More recently, Alvarez et al. (2009) build

on the well-known observation by Cochrane (2001) that most returns and price variations come

from variations in risk premia, and develop a model of exchange rates that shows compelling

evidence that variation in risk premia is a prime mover behind variation in asset prices. All

this evidence directs our attention to the critical role played by risk premia in exchange rate

determination.

In this paper, we investigate the role of risk premia from a new angle. We focus on the

predictive relationship between exchange rates and expected sovereign default risk, and we

construct a new forecasting model based on that.

Expected sovereign default risk constitutes an example of markets’ expectations regarding

the probability of rare but extreme economic disasters, such as a default event. These extreme

events correspond to bad economic times and therefore they can matter disproportionately for

the determination of asset prices despite their low probability of occurrence. This is formalized

in the Rietz-Barro hypothesis (Rietz, 1988; Barro, 2006), which states that the possibility of rare

disasters such as an economic depression is a major source of risk premia in asset prices.1 In a

recent paper, Farhi and Gabaix (2016) develop a theoretical model of exchange rates in which

they allow for extreme negative economic downturns to occur at any point in time. They show

that countries differ in terms of their riskiness and this translates into the extent to which their
1
It is important to note that this literature differs from the so-called peso problem literature (Lewis, 2008),
which assumes no risk premia.

2
exchange rates would depreciate should a default occur. In particular, they claim that because

the exchange rate is an asset price whose risk affects its value, relatively riskier countries have

more depreciated exchange rates. Building on this hypothesis, we contribute to the existing

literature by investigating whether the inclusion of expected sovereign default risk as a new

source of risk premium can explain future movements of exchange rates.

In our empirical assessment, we evaluate whether the inclusion of expected sovereign default

risk for 10 economies—7 advanced and 3 emerging economies—in relation to that of the U.S.

helps improve the out-of-sample (OOS) forecast of exchange rate returns. We obtain our measure

of expected sovereign default risk from the term structure of a cross-section of sovereign credit

default swap (SCDS) spreads at four different maturities over the period from 2012 to 2015.

SCDS contracts act as a form of insurance to hedge against investors’ risk-neutral probabilities

of extreme credit events impinging on sovereign issuers. Therefore, by exploiting the term struc-

ture of SCDS spreads at different maturities, one can infer credit market default expectations

embedded in market data (Pan and Singleton, 2008). Moreover, we stress that what matters

is the perception of the risk in one country relative to the other benchmark countries (i.e., the

U.S.), and in our empirical approach we compute the default risk factor using a one-factor term

structure model of SCDS spreads market data for each country in real time and derive a relative

measure by taking the difference across countries. Our empirical approach assumes that one

factor is enough to account for the time-series dynamics of the term structure of SCDS spreads.

This modeling choice is motivated by the fact that the series of SCDS contracts at different ma-

turities present a high level of co-movement across the series, which is confirmed by a principal

components analysis.

Our findings suggest that relative sovereign default risk perceptions are important for pre-

dicting exchange rate movements, especially at shorter horizons. To be precise, we find strong

evidence of predictability of exchange rates by the default risk model at the horizon of one day

ahead. In 9 out of 10 cases considered in our analysis, our model based on the default risk factor

obtains a better forecasting performance with respect to the RW benchmark. In 7 cases, the

improvement is statistically significant. For weekly and monthly forecast horizons, results are

more mixed, with the default risk factor model still economically improving upon the benchmark

3
for most currencies, but with statistical significance in only 4 cases. Finally, when we account

for the forecasting performance of our models over time, we find that for the default risk model,

for several countries most of the gains take place at the beginning of our OOS period—fourth

quarter of 2013 and first quarter of 2014—and at the end of it—fourth quarter of 2014 and first

quarter of 2015. At the time of the first period, the U.S. economy showed the first signs of

recovery, which changed the relative risk perception of other countries, and this helps to predict

the beginning of the appreciation of the dollar. The second period coincides with the beginning

of the European Central Bank (ECB) quantitative easing (QE) program.

Our empirical approach is further supported by another strand of literature that explores the

implications of a sovereign default for exchange rate determination. In an early work, Krugman

(1979) shows how large, unsustainable budget deficits can lead to currency attacks. Reinhart

(2002) presents a comprehensive study on the relationship between actual sovereign defaults

and currency crisis. She finds a strong association between the two events, and that sovereign

default events often lead to currency crises. Recently, Lukkezen and Bonam (2014) developed

an extended New Keynesian model to allow for sovereign default risk and found that a rise in

default expectations triggers a depreciation of the nominal exchange rate.

Additionally, our work contributes to extending the literature that explores the role of risk

factors as non-traditional drivers of exchange rate movements. This literature has, so far, focused

mainly on the in-sample analysis. An example of this is Lustig et al. (2011), who are the first to

show that large co-movements among exchange rates of different currencies support a risk-based

view of exchange rate determination. Furthermore, Verdelhan (2015) finds that two risk factors

account for a substantial share of individual exchange rate time-series dynamics in developed

countries. He also presents some economic interpretation for these factors by naming them the

dollar and the carry factor. The dollar factor captures the average change in the exchange rate

of all other currencies relative to the U.S. dollar, while the carry factor refers to the change

between portfolios of high and low interest rate currencies. Our work also adds to the existing

literature in a similar fashion by identifying a specific class of risk in the risk premia of exchange

rates. Finally, another study closely related to our present work, Della Corte et al. (2015)

conducts an in-sample analysis on the explanatory power of SCDS spreads in deviations from

4
the corresponding U.S. spreads on exchange rate returns. They find a strong link between

exchange rate changes and SCDS spreads changes and, in particular, that global credit risk is

an important determinant of exchange rates.

Our results also contribute to the analysis of the importance of sovereign default risk for

developed countries’ currencies. Historically, the analysis of the relationship between a country’s

creditworthiness and currency crashes was centered on developing countries. However, in recent

years, markets have started paying attention to measures of default risk perceptions also for

developed countries that are a long way from a default event and for which the economic outlook

has remained stable. Our results present evidence to support the notion that expectations

regarding sovereign default risk also matter for developed economies.

The outline of the paper is as follows. Section 2 describes the SCDS spreads term structure

model. Section 3 presents the data and the sample specification. The empirical specification of

our forecasting model is discussed in Section 4 and results are presented in Section 5. Section 6

describes some robustness checks and Section 7 concludes.

2 A standard model of the default risk factor

In this section, we briefly outline a standard reduced-form pricing model of credit default swap

(CDS) spreads that we use to obtain the implied default risk factor following the work of Pan

and Singleton (2008). The full details on the model, including its full state-space representation,

are provided in Appendix A.

We extract the risk factor by way of the term structure of a cross-section of CDS spreads at

four different maturities. The underlying assumption is that CDS spreads at different maturities

are driven by one latent factor that is specific to the sovereign default risk of each country. The

one-factor modeling choice is motivated by a principal components analysis conducted on the

term structure of CDS spreads. In Table 1, we show that it is indeed the first principal component

that presents the highest share explained of the total variance in the data for all regions included

in our sample.

The price of a CDS contract of maturity τ at time t is defined by the ratio of the so-called

5
default leg and the fixed leg, such that

Ldef
τt
ault
cdsτ t = (1)
Lfτ tixed

where τ refers to the maturity of the contract and t is the time period.

The default leg is equal to the expected payment that the buyer of the CDS contract will

receive from the issuer in case of a default event. The fixed leg, in contrast, is the sum of

discounted CDS premium payments, taking into account the probability that default never

occurs, plus a one-time accrued premium payment made at the time of default. In line with Pan

and Singleton (2008), who argue that a constant recovery rate assumption is not unreasonable

for the sovereign debt market, we keep it constant and fixed at 40%.

Following the steps outlined in Appendix A, it is possible to write the model in a state-space

form. In particular, the CDS spreads yτ t for maturity τ (where τ = 1, 3, 5, 10) in period t can

be expressed by the following latent factor model

yτ t = cds(λt , Θ) + ετ t (2)

where cds(λt , Θ) is the price of the CDS. The price is therefore a nonlinear function of a latent

factor λt . The vector of hyper parameters is given by Θ and ετ t is the measurement error term

specified as white noise Gaussian processes ετ t ∼ W N (0, σε2τ ). The latent factor λt follows an

AR(1) process,

λt = φ0t + φ1t λt−1 + ut (3)

where φ0t and φ1t are defined as φ0t = θ(1 − e−κ∆t ) and φ1t = e−κ∆t , with ∆t being the time step

between observations. The error term ut is a white noise Gaussian process ut ∼ W N (0, ση2 ). We

estimate the Gaussian nonlinear model given by eqs.(2) and (3) and obtain the factors using the

extended Kalman filter. In Table 2, we present the estimated maximum likelihood parameters

of the model.

3 Data and sample description

In this section we describe the data, present some summary statistics and discuss the sample

choice used in our empirical analysis.

6
 We use data for seven developed economies: Norway, Sweden, Denmark, Japan, United

Kingdom, the euro area and Australia; and three emerging economies: South Africa, Mexico

and Indonesia. We take the U.S. as benchmark country. These ten economies constitute an

interesting sample for at least two reasons. First, the seven developed economies constitute the

main economic regions of the developed world and they are open economies that are perfectly

financially and economically integrated in the global economy. The emerging economies are

less integrated in global financial markets, but they represent Africa, Latin America and Asia,

respectively. Second, all ten countries have largely independent currencies that are among the

most traded in the foreign exchange market. All data are obtained from Thomson Reuters via

Datastream.

3.1 Sample choice

We use daily data over the period from 2 January 2012 to 11 March 2015. This gives a total

of 833 daily observations. Our sample period is determined by the timing of regulatory reforms

affecting the functioning of the SCDS market and the start of monetary stimulus policy programs

in most of the economies in our sample.

In an attempt to reduce the destabilizing effect of excessive speculative use of SCDS in-

struments, European authorities introduced a ban on the uncovered or “naked” purchases of

SCDS. In other words, holding SCDS contracts with no offsetting position in the underlying

debt was no longer allowed. The timeline for the policy goes as follows: on 15 November 2011,

the European Parliament formally adopted the proposed regulation, the final version of which

was passed and announced on 14 March 2012 and finally went into effect on 1 November 2012.

Changes in the trend of net notional amounts outstanding in the SCDS market since the end of

2011 show that investors started incorporating the effects of such a ban long before it officially

went into effect the following year (IMF, 2013). This presents evidence that the new regulatory

setting forced much of the speculative trading in SCDS out of the market, so that the prices

allow for a cleaner identification of the fundamentals. A stable regulatory timeline is key to our

present analysis, and thus we restrict our sample to the period after markets started pricing in

the new regulatory framework and no further legal reforms were introduced.

7
 Although the new regulation applies to trading in the European Economic Area sovereign

debt obligations, new regulatory changes were indeed a more global trend. For instance, over the

course of 2012, Japan introduced new regulations and amendments to existing legislation related

to the over-the-counter (OTC) derivatives trade to which SCDS belong. Furthermore, most of

the economies included in our sample are heavily linked to the European financial markets.

The end of 2014 and the first quarter of 2015 was the time when most of the countries in our

sample started, extended or sustained their monetary stimulus policy programs. In the E.U.,

QE started by March 2015 while Denmark introduced negative rates (a sort of undercover QE)

in January 2015 and Sweden in February 2015. Furthermore, in October 2014, the Bank of

Japan (BoJ) decided to increase the amount of asset purchases to U80 trillion per year. During

this time, the Bank of England (BoE) decided to keep its benchmark rates at record lows and,

by the end of March, to put on hold any move to increase the rates. All these events took place

while the U.S. was moving in the opposite direction by ending its QE program and therefore

reinforcing disparities with the rest of the largest developed economies around the world.2

3.2 Sovereign credit default swaps

We use SCDS spreads to obtain our measure of the default risk factor. SCDS contracts are used

to hedge against the default risk in the underlying sovereign bonds and therefore convey reliable

information on the market-implied expectations of sovereign default; see, for example, Pan and

Singleton (2008); Duffie et al. (2003); Christensen (2007); Jaskowski and McAleer (2013). Over

time, this market has become highly liquid and thus provides a better measure of sovereign

default risk than the corresponding bonds market, which also prices other risks such as inflation

risk, or credit agencies rankings, which are often slow to react to economic changes and are

inaccurate.

We use daily SCDS spreads at four different maturities: 1-year, 3-year, 5-year and 10-year,
2
We also tried to run our analysis on a completely different sample, going from 2008 to 2011, which corresponds
to the period of the financial and the initial part of the European debt crisis (compatible with data availability).
The results show that the CDS alone are not able to improve predictability, most likely because of too many
confounding factors driving the exchange rates, and the economy in general, in those years, especially for the
European countries. However, some predictability is found for Japan, Australia and South Africa, especially at
one day forecast horizon. Results are available in Appendix B.

8
for all 10 economies included in our sample. The choice of the maturities is tailored to capture

the entire range of the yield curve while ensuring the inclusion of the most liquid maturities.

Contrary to the corporate case, where most of the trading is concentrated in the 5-year segment

only, for sovereign reference entities, trading is more spread out across different maturities.

Augustin (2014) documents that shorter maturity contracts (1-year or less) and those in the

more than 1- to 5-year segment account for roughly 87% of total volumes.

Our spreads correspond to SCDS contracts with full restructure clause, senior unsecured and

denominated in U.S. dollars. Similar to Pu and Zhang (2012) and Hui and Chung (2011), for

the case of the Euro area, we construct two separate representative Euro aggregate measures,

one for the core countries3 —namely, France, Germany, The Netherlands and Belgium—and a

second one for the peripheral countries—Italy, Spain, Portugal and Ireland.4

Summary statistics are presented in Table 3. Figure 1 shows the evolution of the SCDS

throughout time at the different maturities. In general, the particularly turbulent period af-

fecting sovereigns between 2012 and 2013 is clearly noticeable in the higher rates. It is also

noticeable in the graphs the increase in the US SCDS during the taper tantrum that took place

around two main events: Ben Bernanke’s—then president of the Fed—press conferences on the

22 May and 19 June 2013. In those occasions, Bernanke talked about the possibility of the

Fed cutting back on their asset purchasing program, which caused a market turmoil with funds

flowing rapidly into the US. This caused depreciations in all currencies in our sample with the

larger effect seen in the non-European countries—Australia, South Africa, Mexico and Indone-

sia. Finally, one thing to note is that the jump did not cause an homogeneous increase across

the whole term structure but rather it concentrated in certain maturities. This is an example of

the importance of our term structure approach to estimating the default risk factor which does
3
Our idea on splitting the Euro area in Core vs Periphery is based on the consideration that the movements
in the Euro can be driven by the risk perception regarding certain countries, given that the European Debt crisis
hit hard especially the Periphery countries. However, we replicated the results also for the whole Euro area, and
the results are in line with those of the Periphery, hinting to the fact that these countries had a strong influence
on the movements of the Euro currency. Results are available upon request.
4
As customary in the sovereign debt market literature, Greece is not part of our peripheral countries sample.
The explosive behavior of the Greek CDS spreads may affect our estimations. Furthermore, it underwent an
actual default event in 2013.

9
not rely on one maturity in particular but rather exploit the full information content embedded

in the yield curve.

3.3 Foreign exchange rates

The exchange rate data are daily nominal spot exchange rates expressed as the price in units

of the domestic currency relative to the U.S. dollar (meaning that an increase in the exchange

rate is a depreciation of the domestic currency). Summary statistics are presented in Table 4.

The 10 economies included in our data and their corresponding currency names and codes are

as follows: Norway—Norwegian krone (NOK); Sweden—Swedish krona (SEK); Denmark—Danish

krone (DKK); Japan—Japanese yen (JPY); United Kingdom—British pound (GBP); Euro

area—Euro (EUR); Australia—Australian dollar (AUD); South Africa—South African rand

(ZAR); Mexico—Mexican peso (MXN); Indonesia—Indonesian rupiah (IDR). In Table 5 we

show the sample correlations among all the different currencies in our sample. Figure 2 shows

the changes in exchange rates throughout our sample period.

In our empirical analysis, we calculate log exchange rate changes at the daily frequency for

the different forecasting horizons. We provide more details on the choice of forecasting horizons

in Section 4.

4 Forecasting exchange rate returns

For the objective of our paper, we exploit the factors described in Sections 2 and 3, to see

whether they help improve the predictability of exchange rates. We employ a linear regression

framework where we regress the exchange rates on the default risk factor differentials. In this

way, we explore the role of the default risk factor as a non-traditional driver of exchange rate

movements.

4.1 Design of the forecasting exercise

We employ a recursive forecasting scheme, generating forecasts 1, 5 and 20 days ahead.5 Our

sample spans from 2 January 2012 to 11 March 2015. Our first estimation sample stops on
5
Five days correspond to one week and 20 days to one month.

10
6 November 2013, and we recursively expand it to produce OOS forecasts over the period 7

November 2013 to 10 March 2015. The starting date for our OOS is set one week after the

bilateral liquidity swap arrangement (central bank currency swap (CBCS)) between the main

six central banks was made standing. In financial markets, the CBCS involves two currencies

and allows for the possibility of making the currency issued by one central bank available in

the constituency of the other central bank(s) involved in the swap arrangement. Therefore, the

CBCS alleviates monetary and financial instability by reducing foreign currency liquidity risk.

This regulatory change is key to our analysis for at least two reasons: first, the new permanent

status of CBCS is central to protecting the free movement of capital between countries and

investments in different currencies; second, as for the choice of the end of our total sample

period, here we choose a stable regulatory time span.

We compute our h steps-ahead forecast using the factors in deviations from those corre-

sponding to our benchmark country, the U.S. In particular, forecasts based on our default risk

factor model are obtained as:

 
∆st+h = γ + β ∆xDR
t
d
− ∆x DRU S
t + εt+h (4)

where ∆st+h is the log difference of the h periods ahead (where h = 1, 5, 20), exchange rate
DRj
returns are defined as ln(St+h ) − ln(St ), ∆xt is the first difference of the default risk factor

for country j, where j = d, U S with d being the domestic country, γ is the drift parameter and

εt+h ∼ W N (0, σ 2 ).

A standard benchmark to forecast exchange rate returns h steps ahead is a driftless RW

model (see Rossi (2013) for a review of the literature):

st+h = st + εt+h (5)

4.2 Forecasting evaluation

To evaluate the accuracy of the forecasts produced by our models, we look at the mean squared

predictive errors (MSPEs) and compare them to those produced by a RW benchmark. To

evaluate whether the differences in the MSPEs between our models and the RW benchmark are

significant we use the Clark and West (CW) test (see Clark and West (2007) for details on the

11
test). The CW test is a test of OOS population-level predictability, in which under the null

hypothesis, the extra parameters in the extended model are jointly equal to 0.

While the CW test allows us to evaluate the global OOS performance of our models, we

are also interested in examining the local relative performance over the OOS period. Following

Welch and Goyal (2008), we investigate whether and how the squared prediction varies over

time by a graphical inspection of the cumulative squared prediction error difference (CSPED):
t
X
CSP EDk,t+h = fbk,s+h ,
s=t

where fbk,t+h = e2RW,t+h − e2k,t+h , with e2RW,t+h the squared error of the RW model to predict yt+h

and e2k,t+h the squared error of one of the k alternative models (where k = 1, 2, 3). Increases in

CSP EDk,t+h indicate that the model alternative to the RW benchmark predicts better at the

OOS observation t + h.

5 Results

This section presents the empirical results of the forecasting performance of our models based

on default risks using the framework described in Section 4. We start by testing the evidence of

the in-sample (IS) predictability in Section 5.1, then we move to the core of our results, that is

to the evaluation of the out-of-sample forecasting performance of our model in Section 5.2. In

Section 5.3 we finally look at the forecasting performance of our model over time.

5.1 In-Sample evidence of predictive content

Before moving to our core results on out-of-sample forecasting performance, we start with an

in-sample predictability analysis for our relative default risk measure to exchange rate returns.

In Figure 3 we present evidence using a sequence of recursive estimation windows starting at the

2 January 2012 to 6 November 2013 sample, same as in the forecasting exercise for the case of

h = 1. For all countries in our sample but Japan and for all estimation window considered, the

estimated β is positive. A positive estimated coefficient goes in line with our a priori expectations

that countries with higher relative expected default risk will have a more depreciated currency.

Also, all countries but Norway, Sweden and the UK estimated coefficients are significant at the

12
90% confidence interval for all time samples.

Furthermore, we tested for causality between exchange rate returns and our default risk

measure. In particular, we tested the null hypothesis that exchange rate returns do not granger

cause the changes in our relative sovereign default risk measure for any of the economies in our

sample. Our results in Table 6 show that for all but three countries in our sample—Denmark,

UK and Indonesia—we do not reject the null of non Granger Causality.

5.2 Forecasting performance: out-of-sample results

In Table 7, we present the MSPEs of our proposed default risk model with SCDS factors relative

to the RW benchmark for different currencies and different horizons: h = 1, 5 and 20 days ahead

(which correspond to forecasting the exchange rate one day, one week and one month ahead).

A ratio smaller than 1 implies that the model under consideration beats the RW benchmark at

that specific horizon. Moreover, we report the p-value for the CW test to detect whether the

differences in the MSPEs are statistically significant (for further details on the test, see Clark

and West (2007)).

As a general remark, we find strong evidence of predictability of the exchange rates. Our

default risk model economically outperforms the RW for all the currencies, except for the British

pound, at h = 1 forecast horizon. Moreover, in 7 out of 10 cases, the improvement in predictabil-

ity is statistically significant at 10% level. Even for h = 5 and h = 20, our results are comforting:

in most cases we beat the RW, although for the longer forecast horizon the improvement is sta-

tistically significant for fewer currencies than for h = 1. One possible explanation for the lack

of predictability of the British pound could be the similar behaviour of the US and the UK

in terms of fundamentals over our sample. The two countries were the first hit by the global

financial crisis, but they also reacted more promptly than others: for example, both the Federal

Reserve Bank and the Bank of England implemented the quantitative easing policy already in

late 2008, early 2009. Given that our default risk measure is relative to the U.S., this similarity

between the countries results in no significant effect of the relative default risk measure.

The larger statistical gains with p-values lower than 5% are for the Scandinavian countries.

For these countries, the risk of default is small and stable over the sample; however, we stress

13
that what matters is the perception of the risk in one country relative to the other benchmark

countries (i.e., the U.S.). Compared with the U.S., for example, the Scandinavian countries were

considered much safer in 2012 than at the period starting from the second half of 2013 and during

the following year, when the U.S. economy showed the first signs of resilience. The perception

of risk in an economy can therefore have changed over time even if economic fundamentals

of the country have remained stable. Our model exploits exactly the changes in the relative

risk perceptions to provide more accurate forecasts. Similar reasoning can also be applied for

Australia and Euro core countries.

A different story applies for the Euro peripheral countries: the most acute period of the

Euro debt crisis was in the second semester of 2011 and in 2012, before a series of actions by

the ECB from the second part of 2012 to 2014 and the quantitative easing plan from the end

of 2014 stabilized and improved the sustainability of the Euro peripheral countries, providing

evidence of exchange rate predictability.

Mexico and Indonesia experienced a period of relative political calm during our sample,

allowing risk perception to be related to economic performance and reforms more than political

turmoil and providing evidence of predictability.

Finally, Japan and South Africa were respectively dominated by two major events: Abe-

nomics for Japan and gold dynamics for South Africa as the largest exporter in the world. The

short-term effects of these two phenomena on the domestic economies are not clear and our

model provides economic gains but no statistical improvements.

5.3 Forecasting performance over time

Our sample period is characterized by several shocks, and especially by the European debt crisis

and all the subsequent actions that followed. Such shocks might have changed the predictability

of our model over time. In Figure 4, we present the CSPED to shed light on possible changes

in the predictability over the OOS period. Increases in CSPED indicate that our default risk

model predicts better than the benchmark.

Looking at the default factor performance, we see that most of the gains for several countries

take place at the beginning of our OOS period—fourth quarter of 2013 and first quarter of

14
2014—and at the end of it—fourth quarter of 2014 and first quarter of 2015. For the middle

part of our OOS period, the relative perception of the other countries with respect to the U.S.

weakened and predictability is lost for all but Mexico and Indonesia. During the later period,

default conditions relative to the U.S. did not change because of a convergence of policies of

each country with those of the U.S., and therefore our model cannot exploit enough information

on the default factor.

For the Scandinavian region, Norway and Sweden present the largest and only gains toward

the end of the OOS period. For both, the breaking point was the end of the asset purchase

program of the Federal Reserve (the Fed) in the U.S. by the end of October 2014 (point o);

at this point, the CSPED changes direction and, shortly afterward, we obtain very large gains

from our SCDS model. This could be related to a change in perceptions as the ECB was moving

toward more stimulus and the U.S. was coming out of it. Furthermore, the persistent low oil

prices were a heavy drag on Norway. In the case of Denmark, there is predictability at the

beginning of the OOS period when the U.S. started to show signs of resilience, and problems

in Euro zone countries began to influence the Danish outlook. However, gains are lost after the

Danish National Bank (DNB) raised deposit rates by the end of April 2014 (point g). Given

the low inflation rates at the time, this could have been taken as a positive signal that reflected

good underlying fundamentals. By mid-January 2015, when the ECB announced its quantitative

easing program (point s), predictability returns as a result of fears that Denmark wouldn’t be

able to support its currency peg with the Euro. In early February 2015, the DNB decreased

its benchmark rates into negative territory (point t) and fears decreased, only to increase again

after the ECB quantitative easing effectively started (point v).

In the case of Japan, the picture is very similar, although the reasons behind it are quite

different. At the start of the OOS period, Japan was taken as a safe haven in relation to the

U.S.; however, this was lost after the value added tax (VAT) rate hike in April 2014 (point f),

which resulted in a large contraction in consumption and output. At that point, all gains in

our SCDS model are lost. By the third quarter of 2014, while the U.S. was showing signs of

resilience, Japan fell back into recession (point r) and the perception of relative default risk went

up, allowing our model to exploit some of this information. This was reinforced after the Fed

15
ended its assets-buying program (point o) and the BoJ extended its bond-buying program in

October 2014 (point p).

For the U.K., all predictability is found at the beginning of the OOS period and reaches

its peak by early July 2014 when gross domestic product (GDP) growth reached its highest

level since the crisis (point k). After this point, the CSPED changes direction and begins its

downward trend. Shortly afterwards, all gains are lost because there was no variation in the

relative perception of risk between the U.K. and the U.S.

In the case of the Euro, both core and periphery factors loose the initial predictability gains

after the ECB cut rates into negative territory (point j). The relative perception of default risk

started to matter again after the Fed stopped its asset purchases (point o) and it continued to

increase after the ECB announcement and implementation of its quantitative easing program

(points s and v).

Australia presents a similar pattern to the rest of the countries, with a clear increase in

predictability after the change in policy by the Fed. As the Australian dollar serves as an

investing currency in carry trade strategies, most of the observed changes can likely be explained

by changes in international capital flows.

South Africa shows gains at the start of the sample, which start to fade out after the South

African Reserve Bank increased the repo rates by the end January 2014 (point d), only to be

regained after the Fed exited its quantitative easing program (point o).

Finally, for both Mexico and Indonesia, predictability is present through the entire sample

period. Some of the relative risk perceptions for Mexico seem to decrease after energy reforms

were made law, supporting the incumbent president’s reputation (point m). As expected, after

the Fed ended QE, Mexico’s perceived relative vulnerability increased as capital flew back to

the U.S., and our model’s performance improves. For Indonesia, predictability gains are overall

quite high and volatile, with one large reduction in predictability in early July 2014, when a

new president was elected and political uncertainty was reduced (point l).

16
6 Robustness

In this section, we verify the robustness of our results in different ways. First, we look at

a different measure of our default risk factor. Second, we consider a measure of global risk

factor. Finally, we compare our results with those obtained with a battery of other models used

for predicting exchange rates with data at daily frequency: a standard uncovered interest rate

parity (UIP) model with short-term interest rates, a model with the oil price as predictor, a

model with the VIX as a predictor, and a model with the TED spread, defined as the difference

between the LIBOR interbank market interest rate and the risk-free Tbill rate. While the VIX

represents a measure of uncertainty, the TED can be considered a proxy for funding liquidity.

These models cover a broad sprectrum of alternative benchmarks in the exchange rate literature.

6.1 Comparison with raw data factors

In the present paper, we chose to model the default factor based on a term structure model.

This approach offers multiple advantages such as a more complete mapping of the cross-maturity

relationship over time. However, it is also possible to obtain a quicker and simpler approach

using the raw data of SCDS spreads. Tables 8 and 9 present the forecasting results of the default

factor model versus the RW model using the 5-year SCDS rates and the across maturities simple

average, respectively. The results show that even considering a less sophisticated measure of

the default factor, the improvement over the RW model still persists. For a visual idea on the

default risk factors for credit spread and the underlying SCDS raw data see Figure 5.

6.2 Using a measure of global sovereign risk

Della Corte et al. (2015) note that the association between currency returns and sovereign

credit risk of individual countries can be driven also by global sovereign risk. As a robustness

check of our default risk measure, we redo our OOS forecasting estimations using an alternative

measure based on a global SCDS factor computed as an equally weighted average of our default

risk measure across all countries except the country being forecasted in each regression. For

instance, for the forecasting regression of the Norwegian krone (NOK), the global SCDS measure

is constructed with Norway excluded. This is done to allow for global influences only and rule

17
out potential country-specific effects. Overall, the OOS predictability results shown in Table 10

for the global measure are good but significance is rather weak, especially at the weekly and

monthly horizons. These results are in line with the evidence presented by Augustin (2014),

who argues that during times of stress and for those countries undergoing financial turmoil,

country-specific factors tend to be more relevant, while global risk factors have been found to

be important for samples of more stable countries and during non-crisis periods.

6.3 Comparison with alternative exchange rate models

As a further robustness, we consider four competing models to predict exchange rates: a standard

uncovered interest rate parity (UIP) model with short-term interest rates, a model with the oil

price as predictor, a model with the VIX as a predictor, and a model with the TED spread.

Results about their forecasting performance relative to the RW benchmark are provided in Table

11.

Further, we also compare our model to the alternative specifications using the Model Con-

fidence Set (MCS) in Hansen et al. (2011). The MCS is a set of models constructed in such a

way that it will contain the best model with a given level of confidence.

In the rest of the subsection, we briefly present the different models and their performance.

6.3.1 Comparison with the UIP model

The UIP model represents the most examined model to explain exchange rates with macroe-

conomic fundamentals. Therefore, it provides a good fundamentals-based benchmark to under-

stand the predictive ability of our default risk factor model. To estimate our interest rate factor

model, we use data on 3-month interest rates, for which summary statistics are presented in

Table 3.

Forecasts based on the interest rates are computed according to the following equation:

 
∆st+h = γ + α ∆xIR
t
d
− ∆x IRU S
t + εt+h (6)

IRj
where ∆xt is the interest rate for the j country, where j = d, U S, with d being the domestic

country, and the rest is defined as above.

18
 The results in Table 11 show that the interest rate model provides comparable results to

our default risk model only for some currencies, in particular the Norwegian krone, the Swedish

krone and the Euro. In general, the results are less satisfactory, in line with previous results in

the literature (see Rossi (2013) for a review of the literature).

6.3.2 Comparison with the oil price model

Ferraro et al. (2015) underline the importance of oil prices as predictors of exchange rates. Al-

though the importance of oil prices for currencies of oil exporting countries—Norway, South

Africa and Mexico—is straightforward, the relevance of this predictor for the other coun-

tries in our sample may need further explanation. For commodities exporters—Australia and

Indonesia—, oil price fluctuations can largely affect the prices of their exports due to the

large correlation between other commodity prices and oil prices. Finally, for commodities

importers—Sweden, Denmark, Japan, UK and EU—the relevance of oil prices is very high.

Therefore, oil prices provides a good benchmark for comparison with our default risk factor

model. We use data on WTI or Brent oil prices, for non-European and European economies

respectively. In Table 11 we present the forecasting results of the oil prices model versus the RW

model, which show that the oil prices model is comparable to our default risk model only for the

three Scandinavian currencies and the Euro for h = 1. But evidence of predictability vanishes

for the Japanese yen, Australian dollar, South African rand and Mexican pesos. Moreover, the

forecasts with the oil price models are never statistically better than the RW at longer horizons.

6.3.3 Comparison with a model with the VIX as explanatory variable

Brunnermeier et al. (2008), in their attempt to provide evidence of a strong link between currency

carry and currency crash risk, document that currency crashes are positively correlated with

increases of stock market volatility VIX. In other words, the VIX can be seen as a risk factor

affecting exchange rates. Measures of global uncertainty similar to the VIX are used as carry

trade predictors also in Bakshi and Panayotov (2013). As a robustness check, we therefore try

to predict the our exchange rates with changes in the daily VIX. Results are shown in Table 11.

While at short horizon the model obtains a predictability comparable to our model with a SCDS

19
factor (although for fewer countries), for longer horizon the VIX does not show much predictive

content. This is consistent with the results presented by Brunnermeier et al. (2008), which in the

context of carry trade analysis show that changes in the VIX have mainly a contemporaneous

correlation with carry trade portfolio, but not too much effect in the following week.

6.3.4 Comparison with a model with the TED spread as explanatory variable

In the same paper, Brunnermeier et al. (2008) show that another common risk factor affecting

exchange rates is the TED spread, the difference between the LIBOR interbank market interest

rate and the risk-free Tbill rate. This measure can be considered a proxy for funding liquidity,

see also Asness et al. (2013)6 . In the paper by Brunnermeier et al. (2008), this is considered as

alternative measure to the stock market volatility represented by the VIX. The results in Table

11 show generally no predictive ability of the TED spread at short horizons. Some predictability

is obtained only for h = 20.

6.3.5 Model Confidence Set

We finally compare our model to the alternative exchange rate models using the Model Con-

fidence Set in Hansen et al. (2011). We use the MCS R package developed by Bernardi and

Catania (2014). In particular, we compute the T R, M statistic proposed in Hansen et al. (2011)

(see also equation (8) in Bernardi and Catania (2014)).7 We include in the test our model and

the other four alternative specifications described in Sections 6.3.1 to 6.3.4. We choose a con-

fidence level α = 10% and 5000 bootstrap replications. We find that for all currencies and

all horizons, our default risk model cannot be statistically excluded from the set of the best

models. Therefore we do not report stars for our model. When looking to other specifications

in Table 11, the evidence is different: all models are rejected for some currencies. However, the

number of rejections is relative low, indicating that all risk premium variables provide evidence

of predictability.
6
Bakshi and Panayotov (2013) construct a measure of global funding liquidity similar to the TED spread.
Given the data limitations encountered already by the authors, we limit ourself to the TED spread measure as
defined in Brunnermeier et al. (2008).
7
Results are qualitative similar when using the other T max, M statistic proposed in Hansen et al. (2011).

20
7 Conclusions

In this paper, we study the predictive ability of sovereign default risk to forecast exchange

rate returns. By doing this, we shed new light on the importance of risk factors as drivers of

international asset prices. Furthermore, we identify one specific type of risk present in currency

markets: expected probabilities of rare but extreme economic disasters such as a sovereign

default event. The idea is that the possibility of rare disasters such as an economic depression

is a major source of risk premia in asset prices. Building on this hypothesis, we contribute to

the existing literature by investigating whether the inclusion of expected sovereign default risk

as a new source of risk premium can explain future movements of exchange rates.

We compute our sovereign default risk factor using a term structure model of sovereign credit

default swaps assuming a fixed market-implied recovery rate in case of default. For almost all

countries in our sample, our results show, first, that the inclusion of the default risk factor

improves the forecasting accuracy over the RW model. We find strong evidence of predictability

of the exchange rates by the default risk model at the one day ahead horizon. Second, when

looking at the forecasting performance of our models over time, we find that for our default

risk model the largest gains coincide with domestic and global events that affect the relative

perception of a country’s outlook. We conclude, then, that default risk expectations do provide

accurate information regarding exchange rate movements.

References

Alvarez, F., Atkeson, A., and Kehoe, P. J. (2009). Time-varying risk, interest rates, and exchange

rates in general equilibrium. The Review of Economic Studies, 76(3):851–878.

Asness, C. S., Moskowitz, T. J., and Pedersen, L. H. (2013). Value and Momentum Everywhere.

Journal of Finance, 68(3):929–985.

Augustin, P. (2014). Sovereign credit default swap premia. Journal of Investment Management,

12(2):65–102.

21
Bakshi, G. and Panayotov, G. (2013). Predictability of currency carry trades and asset pricing

implications. Journal of Financial Economics, 110(1):139–163.

Barro, R. J. (2006). Rare disasters and asset markets in the twentieth century. The Quarterly

Journal of Economics, pages 823–866.

Bernardi, M. and Catania, L. (2014). The model confidence set package for r. Technical report.

Brunnermeier, M. K., Nagel, S., and Pedersen, L. H. (2008). Carry trades and currency crashes.

NBER Macroeconomics Annual, Daron Acemoglu, Kenneth S. Rogoff and Michael Woodford,

eds. (Cambridge, MA: MIT Press), 23.

Christensen, J. H. E. (2007). Joint default and recovery risk estimation: an application to CDS

data. mimeo.

Clark, T. E. and West, K. D. (2007). Approximately normal tests for equal predictive accuracy

in nested models. Journal of Econometrics, 138(1):291–311.

Cochrane, J. (2001). Asset Pricing. Princeton University Press, Princeton, NJ.

de Jong, F. (2000). Time series and cross-section information in affine term-structure models.

Journal of Business and Economic Statistics, 18(3):300–314.

Della Corte, P., Sarno, L., Schmeling, M., and Wagner, C. (2015). Exchange rates and sovereign

risk. mimeo.

Duffie, D. and Garleanu, N. (2001). Risk and valuation of collateralized debt obligations. Fi-

nancial Analysts Journal, 57(1):41–59.

Duffie, D., Pedersen, L. H., and Singleton, K. J. (2003). Modeling sovereign yield spreads: a

case study of Russian debt. The Journal of Finance, 58(1):119–159.

Farhi, E. and Gabaix, X. (2016). Rare disasters and exchange rates. Quarterly Journal of

Economics, 131(1), 1-52.

22
Ferraro, D., Rogoff, K., and Rossi, B. (2015). Can oil prices forecast exchange rates? An

empirical analysis of the relationship between commodity prices and exchange rates. Journal

of International Money and Finance, 54(C):116–141.

Hansen, P., Lunde, A., and Nason, J. M. (2011). The model confidence set. Econometrica,

2(79):453–497.

Hodrick, R. J. (1989). Risk, uncertainty, and exchange rates. Journal of Monetary economics,

23(3):433–459.

Hui, C.-H. and Chung, T.-K. (2011). Crash risk of the euro in the sovereign debt crisis of

2009–2010. Journal of Banking and Finance, 35(11):2945–2955.

IMF (2013). A new look at the role of sovereign credit default swaps. Global Financial Stability

Report.

Jaskowski, M. and McAleer, M. (2013). Estimating implied recovery rates from the term struc-

ture of CDS spreads. Tinbergen Institute Discussion Paper.

Krugman, P. (1979). A model of balance-of-payments crises. Journal of Money, Credit and

Banking, 11:311–325.

Lewis, K. (2008). Peso problem. The New Palgrave Dictionary of Economics 2. McMillan,

Hampshire, England.

Lukkezen, J. and Bonam, D. (2014). Government spending shocks, sovereign risk and the

exchange rate regime. CPB Discussion Paper, No. 263.

Lustig, H., Roussanov, N., and Verdelhan, A. (2011). Common risk factors in currency markets.

Review of Financial Studies, 24:3731–3777.

Meese, R. and Rogoff, K. (1983a). The out-of-sample failure of empirical exchange rate models:

sampling error or misspecification? Exchange Rates and International Macroeconomics, 67-

105.

Meese, R. and Rogoff, K. (1988). Was it real? the exchange rate-interest differential relation

over the modern floating-rate period. The Journal of Finance, 43(4):933–948.

23
Meese, R. A. and Rogoff, K. (1983b). Empirical exchange rate models of the seventies: Do they

fit out of sample? Journal of International Economics, 14(1):3–24.

Pan, J. and Singleton, K. J. (2008). Default and recovery implicit in the term structure of

sovereign CDS spreads. The Journal of Finance, 63(5):2345–2384.

Pu, X. and Zhang, J. (2012). Can dual-currency sovereign CDS predict exchange rate returns?

Finance Research Letters, 9(3):157–166.

Reinhart, C. M. (2002). Default, currency crises and sovereign credit ratings. World Bank

Economic Review, 16(2):151–170.

Rietz, T. A. (1988). The equity risk premium: a solution. Journal of Monetary Economics,

22(1):117–131.

Rossi, B. (2013). Exchange rate predictability. Journal of Economic Literature, 51(4):1063–1119.

Vasicek, O. (1977). An equilibrium characterization of the term structure. Journal of Financial

Economics, 5(2):177–188.

Verdelhan, A. (2015). The share of systematic variation in bilateral exchange rates. Journal of

Finance (forthcoming).

Welch, I. and Goyal, A. (2008). A comprehensive look at the empirical performance of equity

premium prediction. Review of Financial Studies, 21(4):253–303.

24
 Table 1: Principal components analysis

Share variance explained

First PC Second PC Third PC

Norway 0.905 0.082 0.011

Sweden 0.965 0.029 0.006
Denmark 0.955 0.039 0.005
Japan 0.866 0.116 0.016
U.K. 0.940 0.051 0.008
Euro Core 0.939 0.056 0.005
Euro Periphery 0.981 0.018 0.001
Australia 0.848 0.138 0.012
South Africa 0.885 0.072 0.022
Mexico 0.959 0.037 0.004
Indonesia 0.969 0.030 0.001
U.S. 0.902 0.081 0.016

Notes: Principal components computed on SCDS contracts at four maturities: 1, 3, 5 and 10 year. Daily data
over the period from January 2012 to March 2015.

25
 Table 2: Maximum likelihood parameters of the default risk term structure model

Norway Sweden Denmark Japan U.K. E.U. Core

κ 0.154 0.002 0.002 0.002 0.002 0.002

(0.105) (0.010) (0.008) (0.007) (0.006) (0.007)

α 0.066 0.201 0.271 0.204 0.181 0.279

(0.007) (0.018) (0.023) (0.025) (0.018) (0.026)

σε 0.144 0.207 0.262 0.426 0.289 0.350

(0.001) (0.001) (0.001) (0.002) (0.002) (0.002)

Log lik 4.70 4.75 4.78 4.83 4.79 4.81

E.U. South
Australia Mexico Indonesia U.S.
Periphery Africa

κ 0.002 0.185 0.169 0.204 0.065 0.002

(0.007) (0.027) (0.115) (0.050) (0.006) (0.057)

α 0.888 0.382 0.136 0.410 0.107 0.345

(0.060) (0.037) (0.012) (0.040) (0.010) (0.041)

σε 0.587 0.735 0.144 0.590 0.802 0.599

(0.003) (0.004) (0.001) (0.003) (0.005) (0.003)

Log lik 4.87 4.89 4.71 4.87 4.90 4.87

Notes: Standard errors are in parentheses.

26
 Table 3: Summary statistics on SCDS premia and interest rates

SCDS premium (in bps) Interest rates (in %)

mean max min std. dev. mean max min std. dev.

Norway 1-year 4.3 19.7 1.1 2.9 3-month 1.71 2.89 1.00 0.40

3-year 9.2 33.8 4.0 4.8

5-year 17.7 44.9 10.6 6.3
10-year 31.9 55.0 24.0 5.5

Sweden 1-year 5.2 45.5 1.5 6.1 3-month 0.46 1.59 -0.69 0.64

3-year 11.9 62.4 5.0 10.5

5-year 22.6 80.6 12.6 14.4
10-year 41.6 94.7 27.0 14.5

Denmark 1-year 10.6 92.4 2.1 17.2 3-month 0.25 1.01 -0.44 0.32

3-year 20.9 117.5 4.0 26.4

5-year 36.2 138.3 13.1 33.0
10-year 59.9 155.4 34.1 32.5

Japan 1-year 14.2 70.0 3.7 8.9 3-month -0.01 0.11 -0.40 0.12

3-year 33.8 118.3 13.5 17.8

5-year 58.9 152.6 30.2 24.2
10-year 99.5 183.1 58.2 26.5

United Kingdom 1-year 7.9 45.5 2.1 6.4 3-month 0.67 1.09 0.46 0.17

3-year 18.2 68.4 6.1 10.5

5-year 34.6 99.6 15.4 17.6
10-year 60.4 119.1 35.2 18.3

Euro Core 1-year 15.2 139.1 3.5 22.5 3-month -0.07 0.17 -0.50 0.16

3-year 32.2 175.0 8.0 34.2

5-year 54.8 200.5 20.0 42.3
10-year 85.9 212.9 47.3 39.6

Euro Periphery 1-year 146.4 849.8 23.3 197.3 3-month 0.69 3.72 -0.25 0.86

3-year 207.7 837.2 53.1 211.7

5-year 235.5 778.6 78.2 188.1
10-year 260.9 684.4 121.9 147.1

Australia 1-year 22.9 88.0 2.8 16.5 3-month 3.03 4.91 2.31 0.65

3-year 56.3 150.6 20.4 28.5

5-year 133.3 293.8 71.4 54.5
10-year 131.5 220.5 88.3 30.3

South Africa 1-year 78.9 240.8 15.7 41.1 3-month 6.02 7.97 5.02 0.68

3-year 146.4 252.3 67.9 38.6

5-year 208.9 391.6 121.7 54.2
10-year 246.1 335.7 171.9 28.0

Mexico 1-year 40.9 92.8 15.1 14.9 3-month 3.61 4.50 2.85 0.55

3-year 79.0 164.5 31.5 26.2

5-year 118.9 229.5 64.2 31.5
10-year 166.5 289.9 109.3 36.4

Indonesia 1-year 49.8 130.9 10.8 22.9 3-month 6.58 8.84 4.19 1.46

3-year 108.3 206.2 59.3 30.9

5-year 177.8 288.9 118.1 36.8
10-year 243.6 375.4 175.2 39.8

United States 1-year 10.8 60.7 2.0 7.0 3-month 0.07 0.34 -0.01 0.07

3-year 13.1 36.2 6.5 5.3

5-year 19.3 41.2 10.0 6.8
10-year 30.6 62.5 20.1 7.6

Notes: Premia on sovereign credit default swaps for 1, 3, 5 and 10 year maturities (left panel) and 3-month

interest rates (right panel) at daily frequency from January 2012 to March 2015.
 Table 4: Summary statistics on foreign exchange rates

mean max min std. dev.

NOK 6.00 7.48 5.46 0.35
SEK 6.72 7.86 6.29 0.30
DKK 5.68 6.16 5.36 0.19
JPY 94.46 121.46 76.19 11.68
GBP 0.63 0.67 0.58 0.02
EUR 0.76 0.83 0.72 0.03
AUD 1.04 1.23 0.93 0.07
ZAR 9.57 11.76 7.44 1.16
MXN 13.22 15.63 11.98 0.67
IDR 10725.8 13242.0 8870.0 1266.3

Notes: Daily nominal spot exchange rate data from January 2012 to March 2015 from Datastream. All currencies
are expressed in levels and in local currencies.

Table 5: Correlation matrix

NOK SEK DKK JPY GBP EUR AUD ZAR MXN IDR
NOK 1.00
SEK 0.91 1.00
DKK 0.71 0.87 1.00
JPY 0.74 0.48 0.20 1.00
GBP 0.21 0.32 0.61 -0.04 1.00
EUR 0.71 0.87 0.99 0.19 0.62 1.00
AUD 0.87 0.66 0.38 0.89 -0.01 0.37 1.00
ZAR 0.71 0.43 0.13 0.94 -0.21 0.12 0.91 1.00
MXN 0.86 0.88 0.73 0.43 0.18 0.73 0.70 0.47 1.00
IDR 0.73 0.45 0.12 0.90 -0.34 0.11 0.89 0.95 0.51 1.00

28
 Table 6: Granger causality test

E.U. E.U. South

Norway Sweden Denmark Japan U.K. Australia Mexico Indonesia
Core Peri Africa

F-stat 0.15 1.50 4.93 1.07 6.00 2.16 1.33 1.02 0.25 1.75 3.94
p-value (0.860) (0.223) (0.007) (0.342) (0.003) (0.116) (0.266) (0.361) (0.780) (0.174) (0.020)

Notes: We use a two lag specification to perform the test so that the final number of observations is 831.

Table 7: Out-of-sample predictability for daily exchange rates returns: default risk model relative to
RW benchmark

h=1 p-value h=5 p-value h = 20 p-value

Norway 0.993 (0.035) 0.993 (0.034) 0.992 (0.013)
Sweden 0.996 (0.048) 0.996 (0.024) 0.996 (0.023)
Denmark 0.995 (0.039) 0.996 (0.049) 0.996 (0.018)
Japan 0.987 (0.121) 0.989 (0.176) 0.994 (0.351)
U.K. 1.001 (0.752) 1.003 (0.887) 1.002 (0.898)
Euro Core 0.995 (0.057) 0.997 (0.101) 0.996 (0.030)
Euro Periphery 0.995 (0.032) 0.998 (0.190) 1.001 (0.643)
Australia 0.994 (0.086) 0.994 (0.079) 0.992 (0.130)
South Africa 0.997 (0.294) 0.996 (0.259) 0.997 (0.346)
Mexico 0.996 (0.150) 0.999 (0.443) 0.991 (0.108)
Indonesia 0.976 (0.053) 1.000 (0.509) 1.005 (0.636)

Notes: The table reports the ratio of the MSPE of the default risk model relative to a RW benchmark, and the
p-values for the CW test (see Clark and West (2007)) for h = 1, h = 5 and h = 20 days ahead. The models are
estimated using recursive windows of data; the first in-sample window is 2 January 2012 to 6 November 2013.
The OOS evaluation period is 7 November 2013 to 11 March 2015. Ratios less than 1 are displayed in bold and
indicate that the default risk model outperforms the benchmark model. P-values for the CW test of less than
10% are also displayed in bold and indicate that the difference is statistically significant.

29
Table 8: Out-of-Sample Predictability for Daily Exchange Rates returns: 5-year CDS relative to RW
benchmark.

h=1 p-value h=5 p-value h = 20 p-value

Norway 0.992 (0.015) 0.993 (0.023) 1.003 (0.662)
Sweden 0.995 (0.064) 0.998 (0.314) 1.004 (0.763)
Denmark 0.993 (0.035) 0.996 (0.108) 1.002 (0.625)
Japan 0.986 (0.104) 0.992 (0.226) 0.992 (0.240)
U.K. 1.001 (0.605) 1.003 (0.742) 1.004 (0.777)
Euro Core 0.993 (0.028) 0.997 (0.275) 0.997 (0.129)
Euro Periphery 0.993 (0.008) 0.996 (0.070) 0.999 (0.359)
Australia 0.994 (0.070) 0.999 (0.399) 0.995 (0.138)
South Africa 0.986 (0.022) 0.997 (0.326) 0.998 (0.394)
Mexico 1.032 (1.000) 0.999 (0.356) 1.003 (0.660)
Indonesia 0.971 (0.100) 0.996 (0.356) 1.002 (0.547)

Notes: the table reports the ratio of the MSPE of the 5-year CDS spreads model relative to a RW benchmark, and
the p-values for the CW test (see Clark and West (2007)) for h = 1, h = 5 and h = 20 days ahead. The models
are estimated using recursive windows of data; the first in-sample window is 2 January 2012 to 6 November 2013.
The OOS evaluation period is 7 November 2013 to 11 March 2015. Ratios below 1 are displayed in bold and
indicate that the default risk model outperforms the benchmark model. P-values for the CW test below 10% are
also displayed in bold and indicate that the difference is statically significant.

30
Table 9: Out-of-Sample Predictability for Daily Exchange Rates returns: average CDS relative to RW
benchmark.

h=1 p-value h=5 p-value h = 20 p-value

Norway 0.992 (0.016) 0.993 (0.038) 1.001 (0.550)
Sweden 0.995 (0.061) 0.997 (0.209) 1.002 (0.682)
Denmark 0.993 (0.020) 0.995 (0.063) 1.000 (0.499)
Japan 0.987 (0.113) 0.991 (0.224) 0.992 (0.255)
U.K. 1.002 (0.712) 1.002 (0.657) 1.003 (0.717)
Euro Core 0.993 (0.021) 0.996 (0.210) 0.996 (0.037)
Euro Periphery 0.992 (0.005) 0.995 (0.035) 0.999 (0.390)
Australia 0.994 (0.087) 0.998 (0.390) 0.993 (0.097)
South Africa 0.992 (0.104) 0.998 (0.382) 0.994 (0.265)
Mexico 1.024 (1.000) 1.000 (0.513) 0.999 (0.398)
Indonesia 0.973 (0.107) 0.996 (0.352) 1.002 (0.551)

Notes: the table reports the ratio of the MSPE of the average model (across maturities simple average) relative
to a RW benchmark, and the p-values for the CW test (see Clark and West (2007)) for h = 1, h = 5 and h = 20
days ahead. The models are estimated using recursive windows of data; the first in-sample window is 2 January
2012 to 6 November 2013. The OOS evaluation period is 7 November 2013 to 11 March 2015. Ratios below 1 are
displayed in bold and indicate that the default risk model outperforms the benchmark model. P-values for the
CW test below 10% are also displayed in bold and indicate that the difference is statically significant.

31
Table 10: Out-of-sample predictability for daily exchange rates returns: alternative default risk models
relative to RW benchmark

Global SCDS
h=1 p-value h=5 p-value h = 20 p-value
Norway 0.992 (0.036) 0.993 (0.035) 0.992 (0.025)
Sweden 0.992 (0.011) 0.997 (0.131) 0.997 (0.091)
Denmark 0.994 (0.021) 0.997 (0.114) 0.996 (0.036)
Japan 0.988 (0.154) 0.989 (0.157) 0.996 (0.353)
U.K. 0.999 (0.395) 0.999 (0.346) 1.004 (0.952)
Euro Core 0.994 (0.027) 0.997 (0.134) 0.996 (0.045)
Euro Periphery 0.994 (0.029) 0.997 (0.094) 0.995 (0.028)
Australia 0.994 (0.089) 0.994 (0.096) 0.996 (0.159)
South Africa 0.994 (0.144) 0.994 (0.162) 0.996 (0.245)
Mexico 0.998 (0.116) 0.999 (0.394) 0.998 (0.172)
Indonesia 0.993 (0.281) 0.996 (0.365) 1.001 (0.529)

Notes: The alternative measure of default risk is a measure of global SCDS data. The table reports the ratio of
the MSPE relative to a RW benchmark, and the p-values for the CW test (see Clark and West (2007)) for h = 1,
h = 5 and h = 20 days ahead. The models are estimated using recursive windows of data; the first in-sample
window is 2 January 2012 to 6 November 2013. The OOS evaluation period is 7 November 2013 to 11 March
2015. Ratios less than 1 are displayed in bold and indicate that the default risk model outperforms the benchmark
model. P-values for the CW test of less than 10% are also displayed in bold and indicate that the difference is
statistically significant.

32
 Table 11: Out-of-Sample Predictability for Daily Exchange Rates returns: alternative models relative to RW benchmark.

h=1 h=5 h = 20
UIP Oil price VIX TED Spread UIP Oil price VIX TED Spread UIP Oil price VIX TED Spread
Norway 0.985 0.992 0.992 0.996 0.986 0.989 1.002 0.996 0.986 0.995∗∗∗ 0.990 0.992
(0.059) (0.031) (0.051) (0.119) (0.069) (0.084) (0.657) (0.203) (0.054) (0.155) (0.009) (0.016)
Sweden 0.985 0.993 0.995 0.994 0.975 1.001 1.001∗∗ 1.004 0.977 0.999∗∗∗ 0.997∗∗∗ 0.993∗∗∗
(0.075) (0.044) (0.083) (0.177) (0.017) (0.542) (0.596) (0.757) (0.020) (0.305) (0.117) (0.031)
Denmark 0.999 0.994 0.993 0.992 1.000 0.995 0.997 1.014 1.008 0.997 1.000 0.993
(0.371) (0.087) (0.017) (0.174) (0.489) (0.213) (0.273) (0.934) (0.969) (0.198) (0.494) (0.030)
Japan 0.997 0.990 0.991 1.003 1.002 0.990 0.998 1.009 1.018 1.005 0.993 0.995
(0.434) (0.176) (0.262) (0.593) (0.527) (0.175) (0.451) (0.676) (0.855) (0.642) (0.272) (0.358)
U.K. 1.002 1.012 1.007 1.003 1.002 1.002 1.003 1.014 1.002 1.003 1.002 1.001
(0.846) (0.967) (0.777) (0.842) (0.879) (0.679) (0.706) (0.959) (0.785) (0.914) (0.779) (0.618)
Euro Core 0.984 0.994 0.994 0.993 0.989 0.996 0.997 1.019 1.004 0.998 1.003 0.995
(0.039) (0.086) (0.022) (0.175) (0.040) (0.319) (0.202) (0.964) (0.922) (0.227) (0.737) (0.046)
Euro Periphery 0.986 0.994 0.994 0.993 0.986 0.996 0.997 1.019 0.986 0.998∗ 1.003 0.995∗

33
(0.020) (0.086) (0.022) (0.175) (0.013) (0.319) (0.202) (0.964) (0.010) (0.227) (0.737) (0.046)
Australia 0.994 1.054 0.995 1.048 0.994 0.995 0.996 0.993 0.995 0.995 0.995 0.994
(0.160) (0.979) (0.123) (0.964) (0.141) (0.222) (0.298) (0.166) (0.179) (0.191) (0.157) (0.294)
South Africa 1.006 1.032 1.000∗∗∗ 1.108∗∗∗ 1.010 1.000 0.999 1.000 1.002 0.997 1.006 0.995
(0.819) (0.997) (0.514) (1.000) (0.881) (0.499) (0.462) (0.505) (0.645) (0.328) (0.727) (0.284)
Mexico 0.996 1.015∗∗∗ 1.005∗∗∗ 1.098 0.992 0.999 0.999 1.002 0.991 0.998 1.008 0.998∗∗∗
(0.366) (0.998) (0.825) (1.000) (0.247) (0.410) (0.312) (0.868) (0.183) (0.231) (0.983) (0.334)
Indonesia 1.006 0.993∗∗∗ 0.993 0.984 1.012 0.992 0.997 0.996 1.040∗∗∗ 1.001 1.001 1.003
(0.615) (0.276) (0.264) (0.137) (0.717) (0.227) (0.388) (0.348) (0.970) (0.540) (0.542) (0.592)

Notes: the table reports the ratio of the MSPE of different models relative to a RW benchmark, and the p-values for the CW test (see Clark and West (2007)) for h = 1, h = 5
and h = 20 days ahead (reported in parentheses). Stars refer to the Model Confidence Set in Hansen et al. (2011). One star (∗ ), two stars (∗∗ ), three stars (∗∗∗ ) indicate that
the null hypothesis that the model is included in the set of best models is rejected, respectively, at 10%, 5%, 1% confidence levels. We repeat that the null for our model is
never rejected. The models are estimated using recursive windows of data; the first in-sample window is 2 January 2012 to 6 November 2013. The OOS evaluation period is
7 November 2013 to 11 March 2015. Ratios below 1 are displayed in bold and indicate that the default risk model outperforms the benchmark model. P-values for the CW
test below 10% are also displayed in bold and indicate that the difference is statically significant. For the oil price model, we used Brent prices for european countries and the
WTI for the rest of the countries in our sample.
 Figure 1: Sovereign credit default swap premia per country

60 100
200
50
80
CDS premium (in bps)

CDS premium (in bps)

CDS premium (in bps)

40 150
60
30
100
40
20

20 50
10

0 0 0
11−2015 11−2015 11−2015
10 10 10
02−2014 02−2014 02−2014
5 5 5
01−2013 01−2013 01−2013
3 3 Maturity (in years) 3
Date Date
Maturity (in years) Date 02−2012 1
02−2012 1 02−2012 1 Maturity (in years)

(a) Norway (b) Sweden (c) Denmark

120
200
250
100
CDS premium (in bps)

200
CDS premium (in bps)

150

CDS premium (in bps)

80

150
60
100

40 100

50
20 50

0 0
0
11−2015 11−2015
11−2015
10 10 10
02−2014 02−2014 02−2014
5 5 5
01−2013 Date 01−2013 01−2013
Date 3 3 3
Date Maturity (in years)
Maturity (in years)
02−2012 Maturity (in years) 02−2012 1 02−2012 1
1

(d) Japan (e) United Kingdom (f ) E.U. Core

300
350
1000
250 300
CDS premium (in bps)

800
CDS premium (in bps)
CDS premium (in bps)

200 250

600 200
150
150
400 100
100
200 50
50

0 0 0
11−2015 11−2015 11−2015
10 10 10
02−2014 02−2014 02−2014
5 5 5
01−2013 01−2013 01−2013
Date 3 Date 3 Date 3
Maturity (in years) Maturity (in years)
02−2012 1 02−2012 Maturity (in years)
1 02−2012 1

(g) E.U. Periphery (h) Australia (i) South Africa

80
250 400
CDS premium (in bps)

200 60
CDS premium (in bps)
CDS premium (in bps)

300

150
40
200
100
20
100
50

0 0
0
11−2015 11−2015
11−2015
10 10 10
02−2014 02−2014 02−2014
5 5 5
01−2013 01−2013 01−2013
3 3 Date 3
Date Maturity (in years) Date Maturity (in years)
02−2012 1 02−2012 Maturity (in years) 02−2012 1
1

(j) Mexico (k) Indonesia (l) United States

34
 Figure 2: Nominal Exchange Rates per country

8.5 9

8 7
8.5

7.5
8 6.5
7
7.5
6.5 6

7
6
5.5
5.5 6.5

5 6 5
2 2 12 13 3 4 4 15 2 2 2 13 3 14 4 15 2 2 12 3 13 4 4 5
01 01 20 20 01 01 01 20 01 01 01 0 01 20 01 20 01 01 0 01 0 01 01 01
.2 .2 . . .2 .2 .2 . .2 .2 .2 .2 .2 . .2 . .2 .2 .2 .2 .2 .2 .2 .2
01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03

(a) NOK (b) SEK (c) DKK

130 0.68

0.95
120 0.66

0.9
110
0.64

0.85
100
0.62

90 0.8
0.6

80 0.75
0.58
70 0.7
12 01
2 12 13 13 14 14 15 12 01
2 12 13 13 14 14 15 12 12 01
2
01
3
01
3
01
4
01
4
01
5
. 20 .2 .20 . 20 .20 . 20 .20 .20 . 20 .2 . 20 . 20 . 20 . 20 . 20 . 20 .2
0
.20 .2 .2 .2 .2 .2 .2
01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03

(d) JPY (e) GBP (f) EUR

13 16

1.3
12
15

1.2 11

14
10
1.1

9 13

1
8
12

0.9 7
12 01
2 12 13 01
3
01
4
01
4
01
5
01
2
01
2
01
2 13 01
3
01
4
01
4
01
5
01
2
01
2
01
2
01
3
01
3
01
4
01
4 15
.20 .2 . 20 . 20 .2 .2 .2 .2 .2 .2 .2 .2
0
.2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .20
01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03

(g) AUD (h) ZAR (i) MXN

·104

1.3

1.2

1.1

0.9

2 12 2 3 3 4 4 5
01 20 01 01 01 01 01 01
.2 . .2 .2 .2 .2 .2 .2
01 06 12 05 11 04 09 03

(j) IDR

35
 Figure 3: In-sample regression coefficients β

40 100

60
80
20
40 60

0 40
20

20
−20
0
0

−40 −20 −20

13 14 14 14 14 14 15 13 14 01
4 14 14 14 15
01
3 14 14 01
4
01
4
01
4 15
. 20 .20 . 20 . 20 . 20 . 20 . 20 . 20 . 20 .2 .2
0
. 20 .2
0
. 20 .2 . 20 . 20 .2 .2 .2 .2
0
11 01 04 07 09 12 03 11 01 04 07 09 12 03 11 01 04 07 09 12 03

(a) Norway (b) Sweden (c) Denmark

20 180
60
160
0
40 140

−20 120
20
100
−40
0
80

−60 −20 60

40
−80 −40
20
13 4 4 4 4 4 15 3 4 4 14 4 4 15 13 4 4 14 4 4 15
20 01 01 01 01 01 20 01 01 01 20 01 01 20 20 01 01 0 01 01 0
. .2 .2 .2 .2 .2 . .2 .2 .2 . .2 .2 . . .2 .2 .2 .2 .2 .2
11 01 04 07 09 12 03 11 01 04 07 09 12 03 11 01 04 07 09 12 03

(d) Japan (e) United Kingdom (f) Euro Area Core

700 250
280

600 260
200
500 240

220
400
150
200

300 180

100 160
200

140
100
3 4 4 14 4 4 5 3 4 14 14 4 4 5 3 14 14 4 4 4 5
01 01 01 20 01 01 01 01 01 20 20 01 01 01 01 20 20 01 01 01 01
.2 .2 .2 . .2 .2 .2 .2 .2 . . .2 .2 .2 .2 . . .2 .2 .2 .2
11 01 04 07 09 12 03 11 01 04 07 09 12 03 11 01 04 07 09 12 03

(g) Euro Area Periphery (h) Australia (i) South Africa

300
400

280

300
260

240
200

220

100
200

180
0

160
3 4 4 4 4 4 15 3 14 4 14 4 14 15
01 01 01 01 01 01 0 01 0 01 0 01 0 0
.2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2
11 01 04 07 09 12 03 11 01 04 07 09 12 03

(j) Mexico (k) Indonesia

Notes: The line represents the estimated β in the progressively expanding estimation sample. The shaded area
represents the 90% confidence interval.
36
 Figure 4: Cumulative squared prediction error difference for h = 1

0.5
1 n o s tuv g o s uv 0.4 g n o s tv
0.4
0.8 0.3
0.3

0.6
0.2 0.2

0.4
0.1
0.1
0.2 0
0
0 −0.1

−0.2 −0.2 −0.1

13 14 01
4
01
4
01
4
01
4
01
5
01
3
01
4 14 01
4 14 01
4 15 01
3
01
4
01
4 14 01
4
01
4
01
5
20 20 .2 2 2 2 .2 2 2 20 .2 20 .2 20 .2 .2 .2 .2
0
.2 .2 .2
11. 01. 04 07. 09. 12. 03 11. 01. 04. 07 09. 12 03. 11 01 04 07 09 12 03

(a) Norway (b) Sweden (c) Denmark

0.1 0.4
f r op v a i k op s v a j op s v
1.5
0.3
5 · 10−2
1 0.2

0
0.1
0.5

0
−5 · 10−2

0
−0.1

−0.1
3 4 14 4 4 4 15 3 14 14 4 14 14 5 13 14 14 14 14 14 15
01 01 0 01 01 01 0 01 0 0 01 0 0 01 20 20 20 0 0 20 20
.2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 . . . .2 .2 . .
11 01 04 07 09 12 03 11 01 04 07 09 12 03 11 01 04 07 09 12 03

(d) Japan (e) United Kingdom (f) Euro Core

domestic events foreign events

The figure reports the CSPED for forecasting at horizon h = 1:

t
X
CSP EDk,t+1 = fbk,s+1 ,
s=t

where fbk,t+1 = e2RW,t+1 − e2k,t+1 with e2RW,t+1 the squared error of the RW model to predict yt+1 and e2k,t+1 the

squared error of one of the k alternative models. Increases in CSP EDk,t+1 indicate that the alternative model

predicts better at the OOS observation t + 1. The alternative model is the linear regression with the predictor

based on CDS data. Timeline legends: (a) 18/12/13, Fed announcement of tapering by steps; (b) 01/01/14, tax

reform is implemented in Mexico; (c) 23/01/14, Spain exits bailout; (d) 29/01/14, South African Reserve Bank

raises repo rate; (e) 02/14, budget speech in South Africa with good economic forecast; (f) 01/04/14, Japan VAT

rate hike; (g) 25/04/14, Danish Central Bank (DCB) raises deposit rates to 0.05; (h) 18/05/14, Portugal exits

bailout mechanism; (i) 23/06/14, sharp drop in oil prices; (j) 11/06/14, ECB cuts rates to negative; (k) 01/07/14,

U.K. GDP growth reaches its peak since the crisis; (l) 09/07/14, presidential elections in Indonesia; (m) 07/08/14,

Mexico’s energy reforms become law; (n) 05/09/14, DCB cuts rates to negative; (o) 28/10/14, end of Fed asset

purchases and Sveriges Riksbank (SR) hits the ZLB; (p) 30/10/14, BoJ extends QE program; (q) 18/11/14, Bank

of Indonesia raises benchmark rate; (r) 2014Q3, Japan falls back into recession; (s) 22/01/15, ECB announcement

of start of QE; (t) 06/02/15, DCB undercover QE by cutting rates deeper into negative territory; (u) 12/02/15,

SR announces government bond purchases; (v) 09/03/15, ECB starts QE program.

37
 Figure 4 (cont.): Cumulative squared prediction error difference for h = 1

a c hj op s v 0.8 op s v 0.8 a d e o
0.3
0.6
0.6
0.2 0.4

0.4 0.2

0.1 0
0.2
−0.2

0 0
−0.4

−0.2 −0.6
3 4 4 4 4 4 5 3 4 14 4 4 4 5 3 4 14 14 14 14 15
01 01 01 01 01 01 01 01 01 20 01 01 01 01 01 01 0 20 0 0 0
.2 .2 .2 .2 .2 .2 .2 .2 .2 . .2 .2 .2 .2 .2 .2 .2 . .2 .2 .2
11 01 04 07 09 12 03 11 01 04 07 09 12 03 11 01 04 07 09 12 03

(a) Euro Periphery (b) Australia (c) South Africa

0.5 a b i m o a l o q
1.5
0.4

0.3 1

0.2

0.5
0.1

0
0

13 14 14 14
01
4 14 01
5 13 14 01
4
01
4
01
4
01
4 15
. 20 . 20 .20 . 20 .2 20 .2 20 .2
0
.2 .2 .2 .2 .2
0
11 01 04 07 09 12. 03 11
.
01 04 07 09 12 03

(d) Mexico (e) Indonesia

domestic events foreign events

38
 Figure 5: SCDS series and estimated default risk factors

60 0.8 100 1.4 170 2.2

50 0.7 140 1.9

80 1.2

40 0.6 110 1.5

60 1

30 0.5 85 1.2

40 0.7
20 0.4 60 0.9

20 0.4
10 0.3 30 0.5

0 0.2 0 0.2 0 0.2

2 12 2 3 3 4 4 5 2 2 2 3 3 4 4 5 2 2 2 3 3 4 4 5
01 0 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01
.2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2
01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03

(a) Norway (b) Sweden (c) Denmark

200 2 140 1.6 250 3

120 1.4
200 2.3
150 1.9
100 1.2

150 1.8
80 1
100 1.5
60 0.9
100 1.4

40 0.7
50 1
50 1
20 0.5

0 0.6 0 0.3 0 0.5

2 2 2 3 3 4 4 5 2 2 2 3 3 4 4 5 2 2 2 3 3 4 4 5
01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01
.2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2
01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03

(d) Japan (e) UK (f) EU Core

900 6.5 320 3.3 400 0.9

750 6 270 3
300 0.7
600 5 210 3

450 4 160 2.5 200 0.6

300 3 100 2.3

100 0.4
150 2 50 2

0 1.4 0 1.8 0 0.2

2 2 2 3 3 4 4 5 2 2 2 3 3 4 4 5 2 2 2 3 3 4 4 5
01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01
.2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2
01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03

(g) EU Periphery (h) Australia (i) South Africa

250 2.3 380 3.2 70 3

3 60 2.7
320
200 2
50 2.4
250 2.7
150 1.8
40 2.1
190 2.4
30 1.8
100 1.6
130 2
20 1.5
50 1.3
60 1.9
10 1.2

0 1.1 0 1.6 0 0.9

2 2 2 3 3 4 4 5 2 2 2 3 3 4 4 5 2 12 2 3 3 4 4 5
01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 20 01 01 01 20
1 01 01
.2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 .2 . .2 .2 .2 . .2 .2
01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03 01 06 12 05 11 04 09 03

(j) Mexico (k) Indonesia (l) US

1-year SCDS 3-year SCDS

5-year SCDS 10-year SCDS
TS factor

39
 Appendix

A The model

The price of a CDS contract of maturity τ at time t is defined by the ratio of the so-called
default leg and the fixed leg, such that
Ldef
τt
ault
cdsτ t = (A.1)
Lfτ tixed
The default leg is equal to the expected payment that the buyer of the CDS contract will
receive from the issuer in case of a default event. The fixed leg, in contrast, is the sum of
discounted CDS premium payments, taking into account the probability that default never
occurs, plus a one-time accrued premium payment made at the time of default. This yields

Z τ h Rv i
Ldef
τt
ault
= (1 − π) EtQ e− t λs ds λv dv
t
4τ  R 1j  Z τ (A.2)
1X − t4 λs ds
h Rv i
Lfτ tixed = EtQ e + EtQ −
e t λs ds
λv (v − τI(t) )dv
4 t
j=1

so that

Rτ h Rv i
(1 − π) t EtQ e− t λs ds λv dv
cdsτ t = 4τ  Z τ (A.3)
X Q  R 14 j h Rv i
1 − t λs ds Q − t λs ds
4 Et e + Et e λ v (v − τI(t) )dv
j=1 t

where λt is the default risk factor and π the recovery rate upon default. τI(v) represents the
date at which the last premium payment was paid before the default occurred. It is possible
that default takes place between two payment dates; for instance, between times τj and τj+1
where j = I(v). In that case, v − τI(v) is the period over which the protection buyer has to pay
an accrued premium, given that default happened at v. In line with Pan and Singleton (2008),
who argue that a constant recovery rate assumption is not unreasonable for the sovereign debt
market, we keep it constant and fixed at 40%. To solve eq.(A.3), expectations under the risk-
neutral probability measure Q have to be computed. Using the moment-generating functions
given in Duffie and Garleanu (2001), the expectations can be expressed as
h Rv i
EtQ e− t λs ds λv = eΦ(v−t,0)+Ψ(v−t,0)λt [Φu (v−t,0)+Ψu (v−t,0)λt ]
h Rv i (A.4)
EtQ e− t λs ds = eΦ(v−t,0)+Ψ(v−t,0)λt

40
where the functions Φ(v − t, u), Ψ(v − t, u) are the solutions of a set of Riccati equations and
Φu (v − t, u) and Ψu (v − t, u) are the corresponding derivatives, which we describe in detail in
Section A.1. Finally, the default risk factor λt follows a Vasicek (1977) model with a constant
volatility structure. The diffusion process for the factor is as follows:

dλt = Λ(λt − θ)dt + σdWt (A.5)

where Wt is an independent standard Brownian motion under the empirical probability measure
P. To price a derivative, we need to obtain the stochastic process for the factor under the
risk-neutral probability measure Q (to ensure the non-arbitrage condition). To this end, we
transform the diffusion process such that it is equal to

dλ = Λ∗ (λ − θ∗ )dt + σdWt∗ (A.6)

where Λ∗ = Λ and Λ∗ θ∗ = Λθ + σ.
To implement the model, we cast it in state-space form. The CDS spreads yτ t for maturity
τ (where τ = 1, 3, 5, 10) in period t can be expressed by the following latent factor model

yτ t = cds(λt , Θ) + ετ t (A.7)

where cds(λt , Θ) is the price of the CDS defined in eq.(A.3) and it is a nonlinear function of the
latent factor λt . The vector of hyper parameters is given by Θ and ετ t is the measurement error
term specified as white noise Gaussian processes ετ t ∼ W N (0, σε2τ ). The latent factor λt follows
an AR(1) process,
λt = φ0t + φ1t λt−1 + ut (A.8)

where φ0t and φ1t are defined as φ0t = θ(1 − e−κ∆t ) and φ1t = e−κ∆t , with ∆t being the time step
between observations. The error term ut is a white noise Gaussian process ut ∼ W N (0, ση2 ).
We estimate the Gaussian nonlinear model given by eqs.(A.7) and (A.8) and obtain the factors
using the extended Kalman filter. In Table 2, we present the estimated maximum likelihood
parameters of the model. The state-space representation of the model is shown in Section A.2.

A.1 Solutions to the Riccati equations

In this section, we describe the solutions to the functions Φ(t, u), Ψ(t, u), Φu (t, u) and Ψu (t, u)
in (A.4) that solve the expectations in the CDS pricing equation, eq.(A.3), given by Duffie and
Garleanu (2001):

Rt
E[e 0 qλ(s)∂s
λt euλt ] = eΦ(t,u)+Ψ(t,u)λt [Φu (t, u) + Ψu (t, u)λt ] (A.9)

41
 m(a1 c1 − d1 ) c1 + d1 eb1 t m
Φ(t, u) = log + t (A.10)
b1 c1 d1 c1 + d1 c1
1 + a1 e b 1 t
Ψ(t, u) = (A.11)
c1 + d1 eb1 t
∂Φ(t, u)
Φu (t, u) = (A.12)
∂u
∂Ψ(t, u)
Ψu (t, u) = (A.13)
∂u

where

p
−n + n2 − 2pq
c1 =
2q
p
n + pu + (n + pu)2 − p(pu2 + 2nu + 2q)
d1 = (1 − c1 u)
2nu + pu2 + 2q
a1 = (d1 + c1 )u − 1
d1 (n + 2qc1 ) + a1 (nc1 + p)
b1 =
a1 c1 − d1

The solutions of the derivative functions Φu (t, u) and Ψu (t, u) are obtained from the Symbolic
Toolbox from MATLAB.

A.2 State-space representation of the CDS model

For the term structure model of the CDS spreads, we estimate a single latent factor model
with a nonlinear observation equation which requires the use of the extended Kalman filter for
linearization before estimation. Here we outline the state-space representation of our discrete-
time reduce form model given by eqs.(A.7) and (A.8).
The state-space system with state vector Ft is given by

yt = B(Ft ) + εt εt ∼ N (0, H) (A.14)

Ft = Φ0 + Φ1 Ft−1 + ηt ηt ∼ N (0, Q) (A.15)

where

   
Yτ1 t ε τ1 t
   
 Yτ3 t   ετ3 t 
   
yt = 

 , F t = λt , ε t = 
   , η t = ηt
 Yτ5 t   ετ5 t 

   
Yτ10 t ετ10 t

Φ0 = θ(1 − e−κ∆t ), Φ1 = e−κ∆t

42
  
σ2 0 0 0
 ετ1 t 
 0 σε2τ t 0 0 
 
Q = V ar[Ft |F̂t−1 ], H =  3 
 0 0 σε2τ 0 
 
 5t 
0 0 0 σε2τ
10 t

The state vector is initialized with the unconditional moments given in Section A.3.
For the implementation of the extended Kalman filter, the observation equation, eq.(A.7),
is linearized using a first-order Taylor expansion around the predicted state λt|t−1 , as follows:

yt = cds(λt , Θ) + εt (A.16)

∂cds(λ)
yt = cds(λt|t−1 , Θ) + ∂λ + εt (A.17)
λ=λt|t−1

Re-writing cds(λ, Θ) in function notation, we get:

f (λ)
cds(λ, Θ) = (1 − π) (A.18)
g(λ) + h(λ)

∂f (λ)
∂cds(λ) ∂λ [g(λ) + h(λ)] + f (λ)[ ∂g(λ)
∂λ +
∂h(λ)
∂λ ]
= (1 − π) 2
(A.19)
∂λ [g(λ) + h(λ)]

where,
Z τ
∂f (λ) ∂ Rv
= (1 − π) { EtQ [e− t λs ∂s λv ]}∂v
∂λ t ∂λ
∂ Q − v λs ∂s
R ∂ Φλ (v−t,0)+Ψλ (v−t,0)λt λ
Et [e t λv ] = {e [Φu (v − t, 0) + Ψλu (v − t, 0)λt ]}
∂λ ∂λ
Rv λ (v−t,0)+Ψλ (v−t,0)λ
= Ψλ (v − t, 0)EtQ [e− t λs ∂s
λv ] + Ψλu (v − t, 0)eΦ t

λ (v−t,0)+Ψλ (v−t,0)λ
= Ψλ (v − t, 0)eΦ t
[Φλu (v − t, 0) + Ψλu (v − t, 0)λt ]
λ (v−t,0)+Ψλ (v−t,0)λ
+ Ψλu (v − t, 0)eΦ t

4τ
∂g(λ) 1 X ∂ Q − Rt14 j λs ∂s
= E [e ]
∂λ 4 ∂λ t
j=1
1j
∂ Q − R 4 λs ∂s ∂ Φλ (v−t,0)+Ψλ (v−t,0)λt
E [e t ]= {e }
∂λ t ∂λ
Rv
= Ψλ (v − t, 0)EtQ [e− t λs ∂s
]
λ (v−t,0)+Ψλ (v−t,0)λ
= Ψλ (v − t, 0)[eΦ t
]

43
 Z τ Rv
f (λ) = (1 − π) {EtQ [e− t λs ∂s
λv ]}∂v
t
Rv λ (v−t,0)+Ψλ (v−t,0)λ
EtQ [e− t λs ∂s
λv ] = eΦ t
[Φλu (v − t, 0) + Ψλu (v − t, 0)λt ]

4τ
1 X Q − Rt14 j λs ∂s
g(λ) = Et [e ]
4
j=1
Rv λ (v−t,0)+Ψλ (v−t,0)λ
EtQ [e− t λs ∂s
] = eΦ t

Z τ 1j
4
R
h(λ) = {EtQ [e− t λs ∂s
λv ](v − TI(v) )}∂v
t
Rv λ (v−t,0)+Ψλ (v−t,0)λ
EtQ [e− t λs ∂s
λv ] = eΦ t
[Φλu (v − t, 0) + Ψλu (v − t, 0)λt ]

τ
∂h(λ) ∂ Q − R v λs ∂s
Z
= { E [e t λv ](v − TI(v) )}∂v
∂λ t ∂λ t
∂ Q − R v λs ∂s λ λ
Et [e t λv ] = Ψλ (v − t, 0)eΦ (v−t,0)+Ψ (v−t,0)λt [Φλu (v − t, 0) + Ψλu (v − t, 0)λt ]
∂λ
λ (v−t,0)+Ψλ (v−t,0)λ
+ Ψλu (v − t, 0)eΦ t

A.3 Moments of the factors

Here we show the conditional and unconditional mean and variance used in the CDS spreads
term structure model. The conditional moments of the factors are used in the construction of the
quasi-likelihood function that is maximized in the Kalman filter estimation. The unconditional
moments serve to initialize the process.
The conditional mean and variance are as follows:

E(Ft+h |Ft ) = θ + eΛh (Ft − θ) (A.20)

1−e−(2κ)h (A.21)
V ar(Ft+h |Ft ) = 2κ a

In turn, the unconditional mean and variance are as follows:

E(Ft+s ) = θ (A.22)

a (A.23)
V ar(Ft ) = 2κ

For a derivation, see de Jong (2000).

44
B Analysis on the sample 2008-2011

The choice of the sample under investigation (2012-2015) is driven by a major change in CDS
regulation. We try here to conduct our analysis also on the sample corresponding to the Financial
and the initial years of the European Crisis. For problems relating to the availability and
reliability of data, we considered the sample 2008-2011 (data on 2007 are not available for most
of the countries in our sample). The results show that the SCDS alone are not able to improve
predictability, most likely because of too many confounding factors driving the exchange rates,
and the economy in general, in those years, especially for the European countries. Moreover,
regulatory issues as we discussed in Section 4.1 could have played an important role. However,
some predictability is found for Japan, Australia and South Africa, especially at one day forecast
horizon. Results are reported in Table B.1, here below.

Table B.1: Out-of-Sample Predictability for Daily Exchange Rates returns: default risk model relative
to RW benchmark 2008-2011 sample.

h=1 p-value h=5 p-value h = 20 p-value

Norway 1.005 (0.863) 1.003 (0.870) 1.006 (0.890)
Sweden 1.005 (0.958) 1.008 (0.953) 1.009 (0.938)
Denmark 1.005 (0.949) 1.002 (0.681) 1.005 (0.904)
Japan 0.998 (0.387) 0.998 (0.335) 0.998 (0.382)
U.K. 1.004 (0.707) 1.010 (0.905) 1.003 (0.723)
Euro Core 1.006 (0.979) 1.003 (0.928) 1.007 (0.970)
Euro Periphery 1.004 (0.960) 1.002 (0.748) 1.032 (0.999)
Australia 0.999 (0.420) 1.001 (0.549) 1.001 (0.592)
South Africa 0.999 (0.418) 0.999 (0.367) 1.000 (0.690)
Mexico 1.005 (0.850) 1.004 (0.947) 1.001 (0.689)
Indonesia 1.000 (0.494) 1.003 (0.926) 1.011 (0.961)

Notes: The table reports the ratio of the MSPE of the default risk model relative to a RW benchmark for the
crisis aftermath period, and the p-values for the CW test (see Clark and West (2007)) for h = 1, h = 5 and
h = 20 days ahead. The models are estimated using recursive windows of data; the first in-sample window is 2
January 2008 to 29 May 2009. The OOS evaluation period is 1 June 2009 to 31 December 2011. Ratios below
1 are displayed in bold and indicate that the default risk model outperforms the benchmark model. P-values for
the CW test below 10% are also displayed in bold and indicate that the difference is statically significant.

45
Highlights

HIGHLIGHTS

1. Identify new source of currency markets risk premia: expected sovereign default risk

2. Expected sovereign default risk can explain future exchange rate movements

3. For most economies the default risk model improves the RW model forecasting accuracy

4. Strong evidence of predictability is found for developed economies at daily frequency

5. Our model forecasting gains over time coincide with important domestic/global events