The causal linkages between sovereign CDS prices for the BRICS and major European economies

The article examines causal relationships between sovereign credit default swaps (CDS) prices for the BRICS and most important EU economies (Germany, France, the UK, Italy, Spain) during the European debt crisis. The cross-correlation function (CCF) approach that distinguishes between causality-in-mean and causality-in-variance and the Breitung– Candelon causality test in the frequency domain are used in the research. Both tests reveal limited dependence of the BRICS CDS (especially, in the case of Brazil, China and South Africa) on the EU CDS prices. Thus, the paper underscores the signs of decoupling in the sovereign CDS market and also supports the view that the European debt crisis has so far had a limited non-EU impact in this market. JEL C50 G10 G15


Introduction
The European debt crisis has reignited the public and scientific debate on financial contagion and spillovers. In this context, one of the central issues is whether the affected EU countries have been contagious for emerging market (EM) economies, in particular, for the BRICS. The BRICS have fared relatively well during the EU financial turmoil but, obviously, have not been insulated from the negative shocks generated within the EU. For example, Ahmad et al. (2013) find that the BRICS stock markets have been hit strongly during the Eurozone crisis period, with Italy, Spain and Ireland being the most contagious for the BRICS. However, the overall degree of the BRICS exposure to the EU shocks remains insufficiently examined as other potentially important venues of instability propagation, e.g. sovereign debt market or interbank lending linkages have not received necessary attention.
This study attempts to fill in this gap by exploring causal linkages between the most important EU economies (Germany, France, the UK, Italy and Spain) and the BRICS in the sovereign credit default swaps (CDS) market. Studying CDS prices appears to be instrumental in analyzing sovereign credit risks as CDS markets tend to be more liquid than those of the referenced sovereign bonds and disseminate marketwide information more rapidly (Forte and Peña 2009;Delis and Mylonidis 2011). These features of sovereign CDS are valid for advanced and EM economies (Longstaff et al. 2011;Li and Huang 2011;Dieсkmann and Plank 2012).
The novelty of the paper lies in the use of the cross-correlation function (CCF) approach which allows to examine the presence of two types of causality (causalityin-mean and causality-in-variance) between the major EU economies and the BRICS on country-to-country basis. I find that the major EU economies' CDS prices are largely dependent on the performance of the BRICS CDS with the exception of Germany. Italian and Spanish CDS experience the greatest number of incoming causal linkages from the BRICS. Meanwhile, India is the only BRICS sovereign to have a clear-cut negative balance of outgoing and incoming causality among the BRICS. The results make case for the decoupling hypothesis in the sovereign CDS market and for a limited magnitude of the non-EU contagion in sovereign bond markets triggered by the developments within the EU.
The remainder of the paper is organized as follows. Section 2 reviews relevant literature, Section 3 presents the data, Section 4 describes econometric methodology, Section 5 discusses the results, and Section 6 concludes.
2. Relevant literature review The paper is related to several strands of literature. Its bulkiest body focuses on intra-EU linkages in the sovereign CDS market 2 . The analysis of the papers enables to distill several stylized facts.
2 It is also worth mentioning the literature on the credit risk transfer from the banking sector to sovereigns prior to the European debt crisis and mutually reinforcing linkages between banks and sovereigns during the subsequent crisis period. See, for example, Acharya et al. (2011), Ejsing and Lemke (2011), Alter and Schuler (2012).
First, there has been substantial co-movement of the EU sovereign CDS prices. The leading EU economies, such as Germany, have also witnessed a notable rise in their CDS spreads. However, this has been a reflection of interdependence rather than contagion. It was first fueled by Greece and then by Ireland and Portugal in a relayrace manner (Caporin et al. 2013;Broto and Perez-Quiros 2013).
Second, in regard to contagion, the EU sovereign CDS market has been split into two segments -peripheral (Greece, Portugal, Ireland) and core economies (Germany, the UK, France). There was certain risk transmission between them but prior to 2010, and by now it has almost vanished (Groba et al. 2013). The core economies have more capacity to trigger contagion internationally (Kalbaska and Gatkowski 2012), whilst within the EU the intensity of risk spillovers is higher for the peripheral economies which have a long-run volatility memory, but their overall impact is relatively low (Gunduz and Kaya 2013).
Third, though often referred to as peripheral, Italy and Spain should rather be considered core economies. They play a pivotal role for the dynamics of German CDS prices, and vice versa. Consequently, any credit risk event on Italian or Spanish CDS will have catastrophic implications for the entire EU (Gonsalez-Hermosillo and Johnson 2012).
Fourth, the global non-EU factors (the VIX index or TED spread) do not influence significantly the EU core economies' CDS prices but these economies are sensitive to changes in intra-EU financial market variables such as the dynamics of the DAX index (Ang and Longstaff 2013;Zoli 2013).
Unlike their EU counterparts, EM sovereign CDS spreads are linked to global indicators more tightly. For example, the VIX index and TED spread are important predictors for Latin American CDS prices, including those of Brazil and Mexico (Wang et al. 2013). Similarly, based on a wider sample of EM sovereign CDS, Fender et al. (2012) assert that global and regional risk premia contribute to EM sovereign CDS dynamics more than country-specific determinants like credit ratings or macroeconomic variables. They enlarge the list of useful international predictors by adding S&P 500 index and US Treasury 3-month bill rate. China CDS prices are also heavily dependent on global indicators and this dependence has become more pronounced over the past years relative to the role of domestic factors such as the China stock market index and the real interest rate (Eyssell et al. 2013).
The importance attached to China CDS as a potential predictor of other countries' credit risk has been on the rise as well. Analyzing linkages among 11 Asian sovereign CDS spreads, Wong and Fong (2011) emphasize the systemic role of China and South Korea. Ang and Longstaff (2013) and Kalotychou et al. (2013) find that China CDS price dynamics is a reliable predictor of the EU credit risk. However, they do not focus on country-to-country linkages, considering the EU as an integration block.
The literature studying the EU impact on EM CDS prices is very scarce. To the best of my knowledge, the only paper that addresses this issue in regard to the BRICS is Sujithan and Avouyi-Dovi (2013). They find that EU financial indicators exerted major influence over the BRICS sovereign CDS prices in the long-run -from 2002 to 2012. Nevertheless, their analysis was not carried out on country-to-country basis either, as aggregate indicators (Dow Jones Eurostoxx 50 and Eurozone corporate benchmark 5-year yield for AAA issuers) were used as predictors of the BRICS CDS prices. Sujithan and Avouyi-Dovi also find that the EU financial factors remain robust when global indicators (the VIX and S&P 500 indices) are taken into account.
Neither of the papers, however, explicitly tackles causality issues as risk spillovers were quantified based on VAR/VECM models and econometric techniques derived from them. The paper which directly addresses the issue via the crosscorrelation function (CCF) approach and, thus, is closest in methodology to my research, is Yoshizaki et al. (2013). They study causal linkages between major EU economies' CDS (Germany, France, the UK, Italy, Spain, Greece and Portugal) and Japan in two sub-samples -before the start of the European debt crisis (January 2009-April 2010) and afterwards (May 2010-March 2012. They conclude that the causal linkages strengthened in terms of causality-in-mean after the beginning of the European debt crisis. Their direction also experienced a reversal: Japan began to trigger transmission to all the EU economies but the UK, whereas before the crisis it had been subject to incoming linkages from them.
3. Data Daily 5-year sovereign CDS prices 3 are used to conduct the research. The data are sourced from Bloomberg and contain 975 observations from January 2010 to September 2013. Thus, the time span encompasses the developments related to the European debt crisis.
CDS prices for the BRICS exhibit a high degree of co-movement. In particular, the first principal component accounts for 79 percent of the variation in the BRICS CDS prices. On pairwise basis, correlations between the BRICS range from 0.60 (between India and Brazil) to 0.92 (between China and Russia) (Fig. 1 Eurobonds. This is the largest commercial bank in India and the only one featuring in Global Fortune 500 among Indian financial institutions. The indicators of SBI economic performance are often referred to as proxies for the Indian economy by international investors. There is also a high level of commonality in the major EU countries' CDS prices. The first principal component explains 77 percent of the variation in these countries' CDS prices, though the disparity in pairwise correlations is more significant than for the BRICS. The lowest correlation is observed between the UK and Spain (0.14), whereas the highest is between Germany and France (0.96) (Fig. 2  The first principal component for the joint BRICS-EU sample accounts for 67 percent of the variation in CDS prices, also unveiling a strong co-movement of credit risk between the BRICS countries and leading EU economies. The degree of commonality appears to be only slightly lower than for the BRICS and the five EU economies examined separately. This preliminary result additionally motivates search for possible causality between the BRICS and major EU sovereigns' CDS prices. Pairwise correlations are reported below (Table 1).  For ADF, PP and DF-GLS tests the null hypothesis is that the series has a unit root, for KPSS it is that the series is stationary. *, **, *** indicate that the null hypothesis is rejected at 10, 5 and 1% significance level respectively.
In case of India and China unit root tests suggest that the series are not stationary, whilst for Brazil, South Africa and especially Russia they yield conflicting results 4 . To ensure stationarity of the series, they are first-differenced. The baseline (levels) and first-differenced series and their squares (both levels and 1 st differences) exhibit serial correlation (up to 20 lags) and ARCH effects (up to 5 lags) judging by Ljung-Box Q-statistic and LM conditional variance test. The descriptive statistics for the leading EU sovereigns' CDS series are presented in Table 3.

LEVELS 1st DIFFERENCES
Note: The figures in square brackets show the probability (p-values) of rejecting the null hypothesis. For ADF, PP and DF-GLS tests the null hypothesis is that the series has a unit root, for KPSS it is that the series is stationary. *, **, *** indicate that the null hypothesis is rejected at 10, 5 and 1% significance level respectively.
The daily mean ranges from 53.79 for Germany to 305.44 for Spain. Thus, mean Spanish and Italian CDS prices exceed the corresponding indicators for the BRICS.
The Jarque-Bera test shows that the EU CDS price series are non-normal. Unlike the BRICS series, unit root is present in all the EU sovereigns' series, so, they are firstdifferenced. Ljung-Box Q-statistic and LM conditional variance test indicate the 4 Most widespread unit root tests (Augmented Dickey-Fuller test (ADF) and Phillips-Perron test (PP) and Dickey-Fuller-GLS (DF-GLS)) reject the null hypothesis that the CDS prices of Brazil, Russia and South Africa have a unit root at least at 10% level. However, the result contradicts Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test that has more power in comparison with the above-mentioned tests. existence of serial correlation and ARCH effect in levels and 1 st differences and in the respective squared series.

Econometric methodology
The two-stage cross-correlation function (CCF) approach proposed by Cheung and Ng (1996) and modified by Hong (2001) is used. It has become widespread in analyzing causality between stock market returns (Xu and Hamori 2012; Korkmaz et al. 2012), sovereign bond yields (Tamakoshi 2011) and different segments of the financial sector (Tamakoshi and Hamori 2012).
At the first stage GARCH models should be fitted to univariate series in question. Usually Autoregressive, Generalized Autoregressive Conditional Heterodoskedasticity (AR-GARCH) or Autoregressive, Exponential Generalized Autoregressive Conditional Heterodoskedasticity (AR-EGARCH) specifications are considered. Autoregressive-moving-average (ARMA) models for mean equations can also be applied.
At the second stage special statistics to study causality-in-mean and causalityin-variance are computed on the basis of standardized residuals and squared standardized residuals derived from the fitted GARCH model. The standardized residuals ν and squared standardized residuals u are represented respectively as follows: where t ϕ are residuals of the GARCH model, µ -mean of the residuals and t hconditional variance. Let η and ρ be standardized residuals and squared standardized residuals for another GARCH model fitted to the series that presumably has causal linkages with the series in question. In order to test the null hypothesis of no causality-in-mean between the two series during the first k lags, an S-statistic proposed by Cheung and Ng (1996) following a null asymptotic where T is the sample size of the residual series, k -the number of lags and ( ) i r 2 νη squared cross-correlation ratio between the standardized residuals ν and η at lag i .
In case of causality-in-variance this statistic is calculated in the same way, the standardized residuals being replaced with squared standardized residuals u and ρ : The shortcoming of this S-statistic is that each lag is weighted uniformly, making no difference between recent and distant cross-correlations. It is inconsistent with an intuitive expectation that more recent information should play a primary role, with cross-correlations decreasing to 0 as the lag order increases. Hong (2001) proposed a new Q-statistic to overcome this weakness of the S-statistic. The Q-statistics to test causality-in-mean and causality-in-variance are given as follows: Q-statistic is designed to test one-sided causality; upper-tailed standard normal distribution critical values must be used. If the Q-statistic is larger than the critical value of the normal distribution, the null hypothesis of no causality during the first k lags is rejected. The correct application of the CCF approach depends on the adequate specification of GARCH models and the unbiased estimation of the GARCH parameters. The key problem that might arise and seriously affect the results of the causality tests based on the CCF approach is the presence of structural breaks in the variances of the series. Van Dijk et al. (2005) and Rodriguez and Rubia (2007) find that severe size distortions in causality-in-variance tests occur when these breaks are observed. Thus, prior to testing for causality-in-variance, the presence of the structural breaks in the variances of the series should be examined.
The "iterative cumulative sum of squares (ICSS)" test procedure proposed in Inclan and Tiao (1994) and its modified versions (for example, Sanso et al. (2004)) have been used to detect structural breaks in the variances. Nevertheless, the usefulness of the procedures has been questioned as they are not well suited to identify multiple structural breaks 5 . This paper proposes an approach based on Bai-Perron (2003a, b) structural break tests which may be a viable alternative to the methods enlisted above. Like the competing procedures, Bai-Perron tests are aimed at determining structural breaks endogenously, without any a priori information on their dates, but they are more flexible and instrumental in case of multiple breaks. This approach has been applied to detect volatility in sovereign bond markets (Tamakoshi and Hamori 2013), but to the best of my knowledge, this is the first attempt to use Bai-Perron tests to achieve superior GARCH estimation results as a prerequisite for further multiple causality analysis in the sovereign CDS market. After GARCH models have been estimated, GARCH variances are subject to Bai-Perron structural break test. To this end, the variances are regressed on a constant and then Bai-Perron sequential subset testing procedure is implemented. If a structural break is identified, a dummy variable corresponding to its date is constructed and considered as a variance regressor. Then the variance equation of the initial GARCH model is re-estimated to account for shifts in volatility. Based on the re-estimated GARCH model, the S-and Q-statistics are computed.

Results and discussion
At the first stage of the CCF approach adequate GARCH models have been fitted to the first-differenced CDS price series of the BRICS and leading EU economies. The baseline model specification for all the series is ARMA(k,m)-EGARCH(p,q) which is represented as follows: Bayesian information criterion (SBIC) whilst conducting residual diagnostics to avoid autocorrelation. The EGARCH (1,1) model has been selected for all variance equations 7 whereas the order of AR/ARMA models fitted to mean equations differs.
After the preliminary estimation of the ARMA(k,m)-EGARCH(p,q) models, GARCH variances have been generated and examined for potential structural breaks. The results of Bai-Perron test indicate that the number of potential breaks does not exceed four dates, except for Russia and South Africa (Table 4). GED distribution is argued to have a certain advantage over normal and Student's-t distributions in modeling time series with heavy tails and, thus can be considered a generalization of both. However, in case of the UK and India EGARCH(1,1)-normal distributional assumption is found to fit the model better than GED.
These countries did not experience any shifts in their CDS prices within the period in question. For the rest of the BRICS and EU countries break dates constitute four "time clusters" -late July 2010, summer 2011, January-February 2012 and early November 2012. The break dates that occurred in summer 2011 and early 2012 can be treated as common for the BRICS and major EU economies. In late July 2010 they largely concentrate in the BRICS and, on the contrary, in January-February 2012 the potential break dates refer to the EU countries.
Then, as stated in Section 4, dummy variables corresponding to the potential break dates have been constructed. The initial ARMA(k,m)-EGARCH(p,q) specifications are re-estimated, with the dummy variables entering variance equations. Tables 5 and 6 present the parameter estimates for the models of the BRICS and major EU sovereigns' CDS prices.  It appears that the use of the dummy variables has improved the quality of only two models -those of Germany and the UK. The initial and re-estimated GARCH specifications have been compared based on the values of maximum likelihood estimators of the equations and Schwartz Bayesian information criterion. Thus, structural shifts in volatility over January 2010 -September 2013 mattered more for the EU economies than for the BRICS. All ARCH ( 1 α ) and GARCH ( 1 β ) coefficients of the equations presented in Tables 5 and 6 are statistically significant at least at the 5% level. The Ljung-Box statistics, Q-stat (20) and , show that the null hypothesis of no autocorrelation up to lag 20 for the standardized and squared standardized residuals holds at the 1% percent significance. It enables to argue that the overall quality of the suggested model specifications is reasonably good and they can be used at the second stage of the CCF approach. The appendix reports empirical results of the CCF analysis to test for the null hypothesis of no causality up to lag k (1, 2,.., 15), measured in days, for each combination of the BRICS-EU series. To generalize the results in a convenient way, a causality table is filled in (Table 7). Table 7. Causality-in-mean and causality-in-variance between the BRICS and major EU sovereigns' CDS prices, January 2010-September 2013. Note: Only causal linkages significant at least at the 5% level are taken in account and denoted as "+".

EU country→BRICS country BRICS country→EU country EU country→BRICS country BRICS country→EU country
The density of causal linkages between the BRICS and major EU sovereigns is quite moderate. It is equal to 24 and 30 per cent of potential linkages in regard to causality-in-variance and causality-in-mean respectively. Moreover, unidirectional causality is predominant. The result is in line with Peltonen et al. (2013) who study bilateral exposures within a big CDS network encompassing 642 sovereigns and financial institutions and find that it is heterogeneous and of low concentration. These properties make it resemble big interbank lending and payment system networks.
In regard to causality directions, Germany is the only sovereign to have a positive balance of outgoing and incoming causal linkages with the BRICS. Within the time span in question, German CDS prices Granger caused those of Brazil and India with respect to causality-in-variance and that of India regarding causality-inmean. It experienced no feedback from the BRICS. The rest of the EU sovereigns are mostly Granger caused by the BRICS counterparts. This direction of causality becomes more pronounced from the UK and France to Italy and Spain.
Among the BRICS, Brazil has the greatest number of outgoing linkages, both in terms of causality-in-variance and causality-in-mean. China, Russia and South Africa are completely decoupled from the EU influence in any type of causality. Their influence is largely concentrated on Italy and Spain and to a less extent on France. On the contrary, India appears to be the most vulnerable to the EU influence, though it is entirely channeled via causality-in-mean.
Overall, the findings indicate that there was no significant dependence of the BRCS sovereign credit risk on the developments in the EU. Rather, the major EU economies were affected by the changes in the credit risk of the BRCS. In case of Italy and Spain this conclusion is especially true.
The result meshes well with the studies that emphasize a quite satisfactory performance of EM sovereign bonds during 2009-2012 when many of them were reckoned as a safe haven by international investors. EM sovereign bonds denominated in local currencies and US dollars fared best of all (Miyajima et al. 2012). This interest in EM public debt was motivated by the relative shortage of global safe assets and resulted in generally stable EM bond yields during the European debt crisis. The revealed causality-in-mean from the major EU economies to Indian CDS may reflect the difficulties with external financing which the State Bank of India faced in the international interbank market rather than any dramatic deterioration of the Indian macro-fundamentals.
Given the substantial and still growing importance of the BRICS in international economics and finance, the paper provides supportive evidence for a limited magnitude of the non-EU contagion in sovereign bond markets triggered by the developments in the EU. This finding is consonant with Beirne and Fratzscher (2013) who argue that intra-and cross-regional contagion and spillovers in the sovereign CDS market have not increased systematically during the crisis but have become much more dependent on the countries' fundamentals. More fragile economies have experienced stronger and more durable shocks than resilient countries which only faced transitory rises in CDS prices. This type of financial relationships between countries has been dubbed "wake-up call" contagion (Giordano et al. 2013) which is to be contrasted with the "pure" contagion when interrelations are not based on the indicators of economic performance.
However, the contagion associated with the EU sovereign credit risk can be transmitted to the BRICS via other financial markets (e.g. stock exchange panics) and/or cumulative effect of global risk aversion. Being more sensitive to global indicators in comparison with the major EU sovereigns' CDS, the BRICS CDS prices may also cause a pass-through effect on them. These two hypotheses are to be examined in the course of future research.

Conclusions
In this paper, based on the CCF approach I study causal linkages between the BRICS and major EU economies in the sovereign CDS market after the outbreak of the European debt crisis. As these linkages approximate the transmission of sovereign credit risks, this analysis is intended to empirically assess the bilateral impact of the most important EU and EM economies.
Although ordinary correlations and principal component analysis indicate a high degree of co-movement in the EU and BRICS CDS prices during January 2010 -September 2013, the density of linkages in terms of causality-in-mean and causalityin-variance is quite moderate. The balance of outgoing and incoming causalities is in favor of the BRICS economies. The only exception is India but this result may be due to the idiosyncratic deterioration of the financial conditions of the State Bank of India which is a conventional proxy for India in the CDS market rather than reflect a dramatic worsening of its macro-fundamentals. Of the major EU economies, Germany is the only sovereign to have a positive balance of outgoing and incoming causal linkages with the BRICS, whilst Italian and Spanish CDS prices are strongly driven by their BRICS counterparts.
Thus, the paper underscores the signs of decoupling effects in the sovereign CDS market and also supports the view that the European debt crisis has so far had a limited non-EU impact in this market.   Table A1. China-Germany causal linkages     Table A5. China-Spain causal linkages               Table A21. India-Germany causal linkages