Spillovers and contagion in the sovereign CDS market

Transcription

1 Bank i Kredyt 44 (6), 2013, Spillovers and contagion in the sovereign CDS market Michał Adam* Submitted: 20 December Accepted: 20 May Abstract This paper focuses on the relationship between sovereign credit default swaps (SCDS) referencing a group of selected developed and emerging economies during the recent sovereign debt crisis. Interdependence and contagion are found on the market dominated by a small number of big international participants. The results show that: (i) a strong commonality exists between global credit spreads (almost half of their variance can be attributed to a single component) with important regional resemblances, (ii) intra-regional spillovers are even more significant, as up to 80% of the forecast error variance of SCDS spreads comes from spillovers, (iii) there is a significant time-variation in spillovers, with contagion from distressed countries gradually diminishing over time as they lose access to bond markets, (iv) the impact of a country s credit spread on the system appears to be largely liquidity-driven (up to 67% is explained by various liquidity measures). Keywords: sovereign debt crisis, sovereign credit default swap, sunspot, contagion, spillover index JEL: C32, C38, F34, G01, G15 * Narodowy Bank Polski; Domestic Operations Department; michal.adam@nbp.pl.

2 572 M. Adam 1. Introduction With the transformation of the global financial crisis into a sovereign debt crisis in the Eurozone, starting from Greece in late 2009, an insight into the nature of credit risk has become crucial. A number of countries have experienced intense debt price pressures and have withdrawn from international bond markets. Furthermore, the concerns about contagion among group of countries have appeared. As suggested by Longstaff et al. (2011), the complex nature of sovereign credit risk affects the ability of market participants to diversify risk internationally. Given the large size of sovereign debt markets, the emergence and rapid development of credit derivatives as a means of insuring sovereign debt is not surprising. The most popular derivative security for managing sovereign debt exposure is a credit default swap (CDS). An in-depth description of CDS contracts has been provided e.g. by Duffie and Singleton (2004). A CDS is an OTC contract that offers insurance against credit event (in particular sovereign default). The protection buyer pays a fixed premium, called the CDS spread, to the seller until the time of the credit event or until the maturity date of the CDS, whichever is first. If the credit event occurs prior to maturity, the protection seller pays compensation to the protection buyer. The contingent amount is most often specified to be the difference between the face value of a bond and its market value, paid at the time of the credit event. It is equal to the notional principal multiplied by one less the recovery rate, where the recovery rate is equal to the ratio of the post- -default value of the bond to its face value (Hull 2009). Sovereign CDS contracts have been traded actively on emerging country and recently also on developed country debts. Longstaff et al. (2011) argue that an important advantage of using SCDS data (rather than sovereign bond data) for measuring credit risk is that the sovereign credit swap market is often more liquid than the corresponding sovereign bond market. As a result, SCDS contracts provide more accurate estimates of credit spreads. The increasing attractiveness of SCDS has also been caused by the fact that investors frequently buy protection even if they do not necessarily own the referencing bond at the time of agreement (the so-called uncovered, or naked SCDS). By that means banks can proxy hedge the exposure to a counterparty, which operates in the reference country. Global regulatory standards also have contributed to the increasing demand for such insurance. For instance, Basel rules require banks to hold capital against changes in the market price of protecting against the risk of a credit event. Such a situation takes place for example when a bank enters into the interest rate swap transaction with the sovereign. International financial markets are closely integrated worldwide. Also, a number of factors have integrated markets at the regional level. These developments include the rise of the regional initiatives like the European Union (with the emergence of the common currency itself), the liberalisation of capital flows or the tendency to diversify the portfolios of financial institutions. In this context, the focus of investors shifted from individual sovereign spreads in isolation to the co-movements across countries and regions. Ehrmann, Fratzscher and Rigobon (2011) find that asset prices react the strongest to other domestic asset price shocks, but that there is also a substantial international spillover. Therefore, they conclude that it is necessary to model international financial linkages in order to gain a better understanding of the financial transmission process across various assets. Moreover, Pan and Singleton (2008) show that during some periods a substantial portion of the co-movement among the term structures of sovereign

3 Spillovers and contagion in the sovereign spreads across countries is induced by changes in investors appetites for credit exposure at the global level. These results become of crucial importance during crises, when financial market volatility generally increases sharply and spills over across markets. In bear markets, as found by Longin and Solnik (2001), co-movements increase the most. Those periods emphasise the systemic importance of certain countries in a group. A market is systemic if it both sends and receives shocks from all other markets (ECB 2011). Hence, systemic sovereigns are those that do not only suffer most individually during a crisis, but also contribute most to overall market losses. Systemic events spread from one market to another such that the overall stability of the system may be impaired. In this paper I show that the Masson s topology (1999) of reasons why crises can occur contemporaneously over time is valid for countries credit risk. To this end I use the SCDS dataset of developed and emerging markets from the latest debt crisis period, during which the banking sector risk was transferred to sovereign borrowers. First, using a principal components analysis, I document a strong resemblance between global credit spreads and confirm that almost half of their variance can be attributed to a single component. Loadings of the first components suggest that regional factors play a significant role in determining changes in sovereign spreads and validate the subsequent regional approach, in which countries are clustered into the Eurozone, Asia, EMEA (Europe, Middle East and Africa) and Latin America groups. I then study the interdependence and contagion among pre-specified regions. I employ a recent method of spillover index (Diebold, Yilmaz 2009). In this method results are not derived from a partial equilibrium assumption, in which foreign conditions cause domestic changes. Conversely, the method fully accounts for the feedback of domestic markets to international markets. Spillover indices allow for the aggregation of spillover effects across markets, distilling a wealth of information into a single spillover measure. In this method spillovers are defined in terms of forecast error variance decompositions. The results show that intra-regional spillovers are important, as only 20% to 31% of forecast error variance is explained by domestic factors and region-specific linkages are clearly visible. In general, larger countries in terms of the size of debt markets have a more pronounced impact on the regional SCDS spread returns. 67% of the net spillovers is explained by liquidity measures. The intensity of spillovers may of course vary over time and the nature of any time-variation should be of great interest. Therefore, I use rolling estimations to detect the potential intra-regional contagion, following Forbes and Rigobon (2002) being defined as a sharp surge in spillovers across markets. I find that contagion spills from distressed countries. The influence of the countries loosing access to the bonds markets, however, gradually diminishes over time. For instance, Greek SCDS spreads appear to be systemically important only during the specific early phase of the European sovereign debt crisis. The remainder of the paper is organised as follows. In the next section the related subject literature is reviewed. The third section gives a thorough description of the employed methods. The fourth section contains details concerning the data set at hand. The results from a factor analysis are presented in the fifth section. The sixth and seventh section deal with spillover and contagion analysis, also in the time-varying framework. The last section concludes.

4 574 M. Adam 2. Studies on spillovers, contagion and the SCDS market There is a vast literature concerning co-movements across financial markets. Most of the studies, however, focus on stock markets. The early work of Hamao, Masulis and Ng (1990) finds evidence of short-run interdependence of prices across three major international stock markets located in New York, London and Tokyo. Bekaert, Hodrick and Zhang (2009) find a significant upward trend for developed stock return correlations during the period Their findings are the strongest in relation to the European countries. Regional contagion was also found for East Asia by Caporale, Cipollini and Spagnolo (2005), running from Thailand, Taiwan, Hong Kong, the Philippines and Korea. Evidence was also presented by Hashimoto and Ito (2004), where contagion was detected to be originating primarily in Hong Kong, or by Gębka and Serwa (2007), who find all intra-regional markets influence each other. The latter study confirms, also for Central and Eastern Europe and Latin America, that both intra-regional spillovers are more pronounced than the inter-regional dependencies, which underlines the importance of regional financial markets analysis. The substantial increase in the co-movement among corporate CDS spreads during the GM/Ford rating downgrade in 2005 has been documented by Acharya, Schaefer and Zhang (2008), as well as by Coudert and Gex (2010). The intra-industry information transfer effect of credit events was studied by Jorion and Zhang (2007) and illustrated by a strong co-movement across corporate CDS spreads. The authors also distinguished between contagion effects (positive correlations across credit events) and competitive effects (negative correlations). The first one occurs when the default (reorganisation) of one firm causes financial distress on other firms with which the first firm has close business ties. Assuming a fixed demand for the product, the second effect occurs because remaining firms may capture new clients from the displaced firms, or gain market power generally. The literature on SCDS contagion is rather scant. The Argentinian sovereign crisis has been studied by Chen, Wang and Tu (2011), who using copulas find a significant increase in correlation and tail dependence between Argentinian and other Latin American SCDS spreads. Arghyrou and Kontonikas (2011) observe contagion running from Greece to several EMU countries since late 2009, although their analysis concludes in early Using a similar to this study s technique, Alter and Beyer (2013) find two kinds of linkages: in a group of European sovereigns and between the sovereigns and European banks. Calani s study (2012) also uses SCDS data set and spillover index method. Kliber (2011), on the other hand, focuses on the causality between Central European sovereign spreads during the recent financial crisis. Most of the analysis is however performed on the alternative measures of sovereign risk. Caceres, Guzzo and Segoviano (2010) measure euro area spreads as spreads of sovereign bond yields to the yield on a 10-year euro swap and identify Greek, Portuguese, Spanish and Italian spreads to be main sources of euro-wide contagion. Claeys and Vasicek (2012) examine data for 10-year sovereign bond yield spreads of 16 European Union countries over the corresponding German bond yield to find the systemic importance of the Spanish bond market. Other studies focus mainly on documenting a strong commonality across sovereign credit spreads. Pan and Singleton (2008) emphasise the co-movement among the term structures of SCDS. Longstaff et al. (2011) find, that the first principal component explains 64% of the variation in global sovereign spreads during the sample period. Furthermore, this value increases

5 Spillovers and contagion in the sovereign to 75% during the crisis period in the global financial markets. Augustin and Tedongap (2011) present similar results. These studies use a large sample of both developed, high-grade countries as well as emerging economies, finding that loadings of the first principal component are roughly uniform. On this backdrop the interpretation of Bernd and Obreja (2010) would suggest, that such a phenomenon needs to be associated with the risk of economic catastrophe, as countries with the lowest credit risk usually are expected to fulfil the payments they promise to their debt holders except in the worst economic states. Augustin and Tedeongap (2011) quote macroeconomic fundamentals in the United States to be among the primary drivers of global sovereign CDS spreads and conclude that theoretical determinants are insufficient to explain sovereign risk premia (embedded in SCDS prices). They find that common variation is driven more by global events rather than by the reassessments of the fundamental strengths of sovereign entities (in particular at short-term horizons). Longstaff et al. (2011) point to global market factors, risk premia and investment flows. Upon decomposing sovereign credit risk into a systemic component and a sovereign-specific component, Ang and Longstaff (2011) argue that systemic risk represents a much smaller fraction of total credit risk for U.S. states than is the case of members of the euro area, maintaining that systemic risk is primarily an artefact of common macroeconomic fundamentals. The literature finds also non-negligible role for the liquidity of financial instruments. The explanations of contagion given by Kodres and Pritsker (2002) take into account major global financial institutions, which facing a loss in one market, turn to other markets in order to realise liquidity, so that a crisis in one market triggers crises in others. Investors specialised in a certain region who for instance face losses as a result of a crisis in one country may be forced to liquidate in a number of countries. In these explanations liquidity risk 1 plays a nontrivial role. The liquidity premium is less volatile in liquid markets. Thus, high levels of liquidity should stabilise sovereign spreads. Furthermore, the literature acknowledges that there is a high degree of co-linearity between empirical measures of liquidity and the global risk factor (Arghyrou, Kontonikas 2011). Liquidity-related explanations possibly play a significant role in the SCDS market, where the 10 largest dealers now account for 90% of trading volume by gross notional amounts. 2 Concentration is even higher in the US market, where the five biggest investment banks account for more than 99% of gross notional amounts. In such a homogenous market microstructure, it is more plausible for the agents to coordinate their decisions on market sentiment (particularly concerning pricing of the sovereign risk). Gomez-Puig (2006) finds a stabilising role for liquidity premium for the large Eurozone countries (Italy, France and Spain) after the introduction of the common currency. Eichengreen et al. (2009) argue that the rising importance of common factors from the outbreak of the subprime crisis was due to rising funding risk, while Acharya, Schaefer and Zhang (2008) notice, that the increase in the co-movement among CDS spreads during the GM/Ford rating downgrade in May occurred when dealer funding was stretched. Theoretical foundations to link an asset s market liquidity and traders funding liquidity has been provided by Brunnermeier and Pedersen (2009). Funding shocks experienced by leveraged investors may translate into declines in the market liquidity of securities. From an empirical perspective, these types of funding- 1 Liquidity risk refers to the size and depth of the sovereign bonds market. Additionally, it captures the risk of capital losses in the event of early liquidation or significant price changes resulting from a small number of transactions. 2 Since 2004 the share of 10 largest dealers in trading volume increased more than 15 percentage points (Duquerroy, Gex, Gauthier 2009).

6 576 M. Adam -induced liquidity shocks could represent a common factor driving the values of affected securities. Specifically, if the marginal investor holding sovereign debt (or being long a SCDS contract) were to be subject to these funding shocks, sovereign credit spreads might display a common pattern in liquidity. Adopting Masson s topology (1998), one may discern three reasons for which crises may occur contemporaneously in time. Firstly, they may be due to a common cause, for instance policies undertaken at a global level, e.g. by supranational initiatives, or by developed countries ( monsoonal effects ). Secondly, a crisis in one market may affect the macroeconomic fundamentals in other markets, for instance because a slowdown in growth in one country reduces the potential for other countries exports in a trade union, or because lack of liquidity in one market leads financial intermediaries to liquidate other emerging market assets. With respect to this category the term interdependence is often applied. Finally, a crisis in one country may conceivably trigger a crisis elsewhere for reasons unexplained by macroeconomic fundamentals, likely because it leads to shifts in market sentiment or changes the interpretation given to existing information. For instance, a crisis might lead investors to reassess the fundamentals of other countries, even if they have not changed, or lead to a change in the risk tolerance among investors. This category is often called contagion, as it involves changes in expectations that are not related to changes in a country s macroeconomic fundamentals. It is most natural to think of this in a context where financial markets are subject to multiple equilibria, or self-fulfilling expectations (Masson 1999). It gives a prominent role to what is commonly called market sentiment in the determination of asset prices, or, in economic literature, sunspots, that is irrelevant variables that nevertheless coordinate investors expectations. Self-fulfilling expectations can introduce extrinsic volatility that substantially exceeds the volatility generated by the macroeconomic fundamentals alone. When crisis occurs, jumps between equilibria are triggered by extraneous events. Macroeconomic linkages are simply insufficient to explain dynamic changes, as they typically take time to operate. Forbes and Rigobon (2002) also define contagion as significant cross-market co-movement increasing after a shock. Therefore, one would expect bursts in dependence measures of sovereign risk across countries as a sign of contagion, while any continued high levels of co-movement would only suggest strong linkages between sovereign entities. Vector autoregressive (VAR) models and forecast error variance decompositions (FEVD) are already well understood and widely used both by researchers and practitioners. Chen, Firth and Rui (2002) apply the variance decomposition to the stock markets of Latin America to find that a large proportion of stock market indices variance is attributable to shocks from regional markets. FEVD has been applied by Leitão and Oliveira (2007), who found Portugal to be a stock market volatility absorber. Sari and Malik (2000) use the generalised FEVD and conclude that the growth rate of the money supply contains significant information for predicting the variance of future forecast errors of stock returns. The method of Diebold and Yilmaz (2009) was originally applied to a wide range of global stock markets and generalised by Diebold and Yilmaz (2012) on a sample of US stock, bond, foreign exchange and commodities markets. Recently, this method has been employed to analyse the Central European (CE) exchange and money markets inter alia. Bubak, Kocenda and Zikes (2011) discover that along with the increasing market uncertainty, CE foreign exchange rates and US dollar volatilities co-move closely. The spillover effects increase most for the countries with troubled financial sector developments (e.g. Hungary). Kliber (2010) studies

7 Spillovers and contagion in the sovereign spillovers in a range of Central European money markets and exchange rates, finding evidence of close intra-regional relationships between the Czech Republic, Poland, Slovakia, and Hungary. The study of Calani (2012) also uses Spillover index method in assessing SCDS spreads, finding signs of contagion in the second moment (volatility) of the European SCDS spreads. 3. The spillover index method The spillover index (SI) method, first introduced by Diebold and Yilmaz (2009), is based on a VAR model. The focus is on the forecast error variance decomposition (FEVD), which allows both to aggregate spillover measures across SCDS spreads and to split forecast error variances of each variable into portions attributable to individual system shocks. Consider a covariance stationary, N-variable VAR(p) process of the following form: p x = φ x _ + ε t i = 1 i t i t (1) where ε ~ (0, ) is a vector of independently ε and identically distributed disturbances. The VAR can be rewritten in its moving average representation: xt = A i εt _ i (2) i= 0 where the N N coefficient matrices A i obey the recursion A i = φ1 Ai_ 1+ φ Ai _ φp Ai _ 2 p with A 0 being an N N identity matrix and with A i = 0 for i < 0. The error in forecasting x t H steps ahead conditional on information available at t 1 is given by: with the covariance of: H ξ ( H ) = ε + (3) t A h t h = 0 H h Cov [ ξ ( H )] = A A ' t H h = 0 h h (4) The decomposition of the forecast error variance allows to parse the variance of each variable in (4) into parts attributable to the various system shocks. The FEVD allows then to assess the fraction of the H step ahead error variance in forecasting x i that is due to shocks to x j for each i. The FEVD based on Cholesky factorisation is presented in Diebold and Yilmaz (2009). As the variance decomposition using Cholesky factorisation depends on the ordering of variables in the VAR, Diebold and Yilmaz (2012) adopt the generalised VAR framework of Pesaran and Shin (1998). In this framework the H step ahead forecast error variance decompositions, for H = 1, 2, are denoted by:

8 578 M. Adam g θ ( H ) = ij _ H 1 _ 1 jj h = 0 _ H 1 σ h = 0 ( e ' A e ) ( e ' A A' e ) i i h h h j i 2 (5) where Σ is the variance matrix for the error vector ε, σ jj is the standard deviation of the error term for the j-th equation, and e i is the selection vector, with one as the i-th element and zeros otherwise. The larger is the fraction of the H step ahead forecast error variance in forecasting the asset i due to shocks to market j, relative to the total forecast error variance, the larger will be the measure of spillovers. Since the historical shocks are not orthogonal, the sum of forecast error variance decompositions does not sum up to 100%. Diebold and Yilmaz (2012) normalise each entry of the variance decomposition matrix by the row sum as: ~ ( H ) = θ g ij θ g ij ( N θ g ij j = 1 H ) ( H ) (6) N ~ By construction, θ g N ~ ij ( H ) = 1 and θ g (H ) = N j = 1 i, j= 1 ij. For each asset i, the shares of its H step ahead forecast error variance coming from shocks to asset j, j i are added. Then these sums are added across all i = 1,..., N. The spillover index (SI) can be written as: SI g ~ N N θ g ij (H ) θ g ij i, j= 1 i, j= 1 i j (H ) i j ( H ) = 100 = 100 N ~ N θ g (H ) i, j= 1 ij ~ (7) when there are no spillovers, the SI equals zero. The index is therefore not a simple measure of co-movement of markets that reflects a similar response to a common shock, but it measures the importance of idiosyncratic shocks of all variables included in an unrestricted VAR on other markets (Claeys, Vasicek 2012). As mentioned before, in contrast to the Cholesky factorisation of the VAR model, the generalised approach is invariant to ordering. Instead, the analysis describes how the system behaves taking into account the historical patterns of correlations among the shocks. Elyasiani, Kocagil and Mansur (2007) show that the generalised variance decomposition framework provides a more accurate and realistic description of market linkages, also because it does not impose any a priori restrictions in the VAR, which might be difficult to defend based on economic theory. The Diebold and Yilmaz (2012) method also facilitates the identification of directional spillovers using the normalised elements of the generalised variance decomposition matrix. The directional spillovers received by market i from all other markets j are measured as:

9 Spillovers and contagion in the sovereign ~ N N θ g ij (H ) θ g ij j = 1 j = 1 (H ) SI g i j i j i (H ) = 100 = 100 N ~ N θ g (H ) i, j 1 = ij ~ (8) while the directional spillovers transmitted by market i to all other markets j are calculated as: ~ N θ g ji j = 1 (H ) (H ) SI g i j i j i (H ) = 100 = 100 N ~ N θ g (H ) i, j 1 = ji ~ N θ g ji j = 1 (9) The net contribution to a system of CDS spreads is obtained by simply subtracting, for a variable i, the shares of its H step ahead forecast error variance coming from shocks to asset j, j i, from its sum of shocks to other variables: g g ( ) = ( ) _ g SI i H SI i H SI i ( H ) (10) This group of measures (equations 8 10) is important, because it describes the degree of connectedness, and the degree to which various credit spreads are systemic. The directional measures elucidate how much of the total spillover comes from, or goes to a particular source. From an economic point of view, the statistics facilitate identifying the sources (i.e. transmitters) and the receivers of spillovers, serving as an instrument for the input-output analysis. Against this background, the statistics provide a synthetic indication on the mechanism of transmission and a possibility of identifying contagious elements of a given system. Obviously, a SCDS spread may be both a transmitter and a receiver of spillovers. It is the net measure which allows to quantify the net impact of a SCDS spread, i.e. if it is more of a source than a receiver of spillovers. For that reason the net measure helps to identify systemic countries. The intensity of spillovers may of course vary over time and the nature of any time-variation is of potentially great interest, as it may help in determining contagion. To allow for time-variation of SI, I calculate them in a rolling window of 260 observations, which corresponds to the 1-year period in the data I use. 4. Data The daily pricing data for 5-year SCDS 3 used in this study is provided by CMA datavision via Bloomberg terminal. The SCDS spreads are New York end-of-day quotes. It allows to avoid the possible problem of synchronicity of the data. To maintain uniformity in the contracts, I only use SCDS quotations for senior debt and denominated in US dollars. The sample covers the period from January 2008 to January 2012 in the case of Latin American and Asian countries and from 3 5-year spreads are the most liquid and constitute the largest part of the market.

10 580 M. Adam September 2008 to January 2012 in the case of the Eurozone and EMEA countries. The difference is driven by the availability of data and the routine of gradually adding new quotations to the Bloomberg system. An advantage of using such a sample is that only the recent financial crisis period is examined. Since the time span covers the transformation from an initial subprime crisis in the United States to the sovereign debt crisis in the Eurozone, any possible caveats stemming from the use of both crisis and non-crisis subsamples are avoided. Liquidity is widely acknowledged to be particularly difficult to measure. Two different sources of liquidity proxies for SCDS are used in this study. First, it is calculated as a standard bid-ask spread, where the data is obtained from Bloomberg. Alternatively, I also employ the Depository Trust & Clearing Corporation s (DTCC) data containing volumes of transactions with a weekly frequency. This data is provided by Reuters and is used to check the robustness of the results. DTCC data comes as net notional values, which are the sum of the net protection bought/sold by net buyers/sellers. Net notional positions generally represent the maximum possible net funds transfer between net sellers of protection and net buyers of protection that could be required under the occurrence of a credit event. By contrast, gross notional values are the sum of all SCDS contracts, do not take into account the offsetting positions and are not very reflective of a true market size. It is the net notional amount, which reflects the actual size of the SCDS market, even if it can be meaningfully smaller than the gross notional amount outstanding (constituting around 10% of the gross amounts). The series used as inputs for VAR models are SCDS spread daily log-returns for selected developed and emerging markets in Asia (China, Indonesia, South Korea, Malaysia, Philippines, Thailand, Vietnam), EMEA (Bulgaria, Croatia, Czech Republic, Hungary, Israel, Latvia, Lithuania, Poland, Romania, Russia, South Africa, Turkey, Ukraine), the Eurozone (Austria, Belgium, Cyprus, Estonia, Finland, France, Germany, Greece, Ireland, Italy, Netherlands, Portugal, Slovakia, Slovenia, Spain) and Latin America (Argentina, Brazil, Chile, Colombia, Mexico, Panama, Peru, Venezuela). The dimension of the models can grow large for such a number of markets, but the use of daily data and the sample of 260 observations (in a rolling window analysis) allows to avoid estimation problems occurring in small samples. Three criteria were used to include a country s SCDS spread in the sample. First, data availability. The availability of SCDS data is usually high for the big corporates, but much narrower for sovereign entities. Also, time-series data for developed economies is shorter than for the majority of emerging markets. Second, only liquid markets are analysed, as indicated by DTCC s top 1,000 most liquid entities. Therefore, for instance, Malta and Luxembourg, which are members of the Eurozone, are not included in the sample. Moreover, quotation precision is closely associated to liquidity of the SCDS market. The regional division follows a standard market convention, according to which investors group countries according to their geographical location and economic development (see also Section 5 for more arguments for the regional approach). Additionally, regional similarities, due to cultural links and both economic as well as political integration, make these countries particularly suitable for a comparative study. Usually, studies regarding sovereign risk (e.g. Longstaff et al. 2011; Augustin, Tedongap 2011) use data samples containing a considerably large number of pre-crisis observations. This approach facilitates a comparison of crisis and non-crisis periods. On the other hand, one needs to deal with structural breaks in the data, which pose obvious estimation problems. Using a sample from January 2008 onwards, this study avoids the aforementioned problems. The period I cover includes the onset of the financial crisis, as well as its transformation into the sovereign risk

11 Spillovers and contagion in the sovereign phase. Table 1 provides a summary of sovereign CDS premia. The average values of the spreads range widely across countries. The lowest is 33.8 basis points for Finland. The highest average is 1,293.7 basis points for Argentina. Both the standard deviations and the minimum/maximum values indicate that there is also a significant time-series variation in sovereign CDS spreads. For example, the spread for Greece ranges from 17.7 basis points to 11,310.0 basis points during the sample period. Table 2 provides similar information for the liquidity measures used in the study. The largest average transaction costs (bid-ask spread) are observed in Cyprus (20.1%), Finland (14.8%) and Chile (12.5%) with the lowest in Brazil (2.3%), Mexico (2.8%), South Africa and Turkey (both 2.9%). The largest weekly SCDS volume is observed in Italy (USD 23.2 billion), Spain (USD 14.9 billion), Germany (USD 14.1 billion) and Brazil (USD 13.6 billion), while the countries with the lowest volumes are Estonia, Chile (both USD 0.5 billion) and Vietnam (USD 0.6 billion). Some distortions in the data are observed, for example, bid-ask spreads are sometimes negative, but these observations are not frequent in the sample. 4 Standard deviations are relatively lower in the credit swap volumes sample. 5. Similarities between sovereign CDS spreads In this section, a principal components analysis (PCA) is applied to determine whether SCDSs exhibit a similar degree of commonality. As an input for the analysis, 43 times series of logarithmic rates of return of the SCDS premia are used. The sample is balanced in that each series consists of 886 observations beginning in September As indicated by the correlation matrix, 5 pairwise correlations vary across countries, which suggests that different sovereigns may have similar credit risk profiles conditional on their geographical, economic, cultural or historical similarities. The lowest pairwise correlation is observed between Israel and Vietnam (11%), while the highest is between Brazil and Colombia (96%). The strongest correlation patterns are found within the regions: Bulgaria and Romania (84%), Korea and Philippines (80%) or Hungary and Poland (79%). It must be stressed, however, that the signs of all pairwise correlations are positive and the average value is 43%. Hence, sovereign spreads tend to display similar dynamics. Table 3 contains the results of the PCA for the period September 2008 to January Again, a strong resemblance in the performance of sovereign CDS spreads is found. In particular, the first component explains 46% of the variation in premia. In addition, the first three components explain nearly 60% of the variation over the entire sample period. Taking into account the fact that the sample consists of both low and high risk countries, 6 it is consistent with the hypothesis of economic catastrophe risk embedded in the CDS premia, advocated by Bernd and Obreja (2010). Alternatively, it may exacerbate a feature of the CDS market, according to which the market is driven by a limited number of traders, as noted by Duquerroy, Gex and Gauthier (2009). 4 In fact, they constitute half a percent of the bid-ask spread sample. Also, only average measures are used, controlling for the possible influence of negative bid-ask spreads. The calculations using adjusted samples (without negative bid- -ask spreads) do not change the results of the study. 5 The correlation matrix for 43 sovereign CDS spreads is very large and therefore available upon request. 6 E.g. Germany is a low risk, high investment grade sovereign with AAA rating, while Greece is a high risk, low or even non-investment grade sovereign. Following a series of downgrades by major rating agencies, Greece gradually lost its investment grade rating in late 2010.

12 582 M. Adam Loadings of the first component support these results. Figure 1 shows that the first component consists of a roughly uniform weighting of credit spreads for most of the sovereigns. It is also positive for all of the countries in the sample. On the other hand, the second component places substantial positive weight on members of the euro area. The third component is heavily positively weighted towards Asian countries and heavily negatively weighted towards Latin American countries at the same time. The PCA therefore clearly suggests that geographical and economic factors might play a significant role in determining changes in sovereign spreads. Strong regional linkages justify clustering the sovereign entities in the subsequent analysis into the following regions: the Eurozone, Asia, EMEA and Latin America with the three last regions being usual market convention for grouping emerging market economies. This clustering is consistent with the financial market agents approach to specifying regional divisions. The agents often analyse a market from the perspective of its regional peers, which display similar economic patterns (e.g. similar growth perspective, economy structure or market organisation). Also, many investment decisions, including fund allocations, are made solely on the basis of geographical distance of two or more markets. To sum up, a significant intra-regional spillovers are expected upon employing this approach to clustering subjects of the study. 6. Spillovers and contagion between SCDS spreads: the role of liquidity In this section VAR models are estimated on the four sets of SCDS premia referencing countries constituting pre-defined regions. Next, FEVDs are performed to obtain estimates of the SI. The chosen models for the SCDS spread returns are first-order VAR models. 7 However, as shown by Diebold and Yilmaz (2009), the proper order of the VAR models is of secondary importance. Different orders do not change the overall results in terms of spillover patterns and dynamics, which is also true for the rolling window analysis. I compute FEVDs at a 5-day horizon, which corresponds to one business week. It is sufficient to capture the horizon at which spillover across markets occurs. Spillover indices for the Eurozone, Asia, EMEA and Latin America regions are presented in Tables 4 7. The analysis is performed on the full samples of September 2008 January 2012 in the case of the Eurozone and EMEA and January 2008 January 2012 in the case of Asia and Latin America. Spillover indices are reported in the lower right corner of each of the spillover tables. For example, in Table 4 one can see that the estimated contribution to the forecast error variance of Portugal coming from changes into Spain, equals 12.6%. The off-diagonal column sums (labelled Contribution to others) and row sums (labelled Contribution from others) are the to and from directional spillovers. In this context, SI are approximately the off-diagonal column sums (or row sums) relative to the column sums including diagonals (or row sums including diagonals), expressed as a percentage. 8 Therefore, the spillover tables provide an approximate input-output decomposition of the total SI. 7 I follow the indications of information criteria (both Schwarz and Hannan-Quinn) to select the optimal lag length. 8 As noted by Diebold and Yilmaz (2012), if Cholesky ordering was used in the variance decomposition procedure, the spillover indices would be exactly the off-diagonal column sums (or row sums) relative to the column sums including diagonals (or row sums including diagonals).

13 Spillovers and contagion in the sovereign Important conclusions arise when analysing the tables. Aggregating all of the various cross- -country spillovers into a single spillover index for the full data sample, it transpires that up to 80% of total forecast error variance comes from spillovers. The SI varies between regions from 68% (Asia) to almost 80% in the EMEA region. Going into the details, from Table 4 one can see, that Spanish SCDS spreads are responsible for 12.6% of the error variance in forecasting Portuguese spreads but only 5.0% of the error variance in forecasting Estonian spreads. Clearly, Table 4 confirms existence of specific intra-regional linkages. From the economic point of view, it confirms close ties among the groups of peripheral euro area, or Club Med countries (Portugal, Italy, Ireland, Greece, Spain), as well as core countries (Germany, France, Austria, Finland or Netherlands). Interestingly, the rows of the Table 4 contain evidence that innovations to Greek SCDS premia are not the most important drivers of the majority of euro area SCDS premia in the entire sample. For any given country one can find several foreign SCDS premia which are more influential than the Greek counterpart. The analysis does not rule out, however, the possibility that in some sub-periods the results might differ. For instance, in the sample January 2009 January 2010 (Table 8) it transpires that Greek SCDS premia explain the largest proportion of the FEVD Eurozone-wide (row contribution to others ). In fact, on the 14 January 2010, Greece announced its Stability and Growth Program which was designed to reduce the country s budget deficit from 12.7% GDP in 2009 to 2.8% GDP by 2012 in order to bring the deficit into alignment with the convergence criteria outlined in the Maastricht Treaty. During that period, the Greek SCDS spreads were the most influential among the group of 16 countries. Nevertheless, Table 4 suggests that during the entire sample period, the largest spillovers in the euro area were stemming from Spain, Italy, France or Germany (row contribution to others ). In the EMEA region the most spillovers were coming from Russia, while in Latin America the spillovers were comming form Brazil. From an economic viewpoint it means that, contrary to the popular belief, the biggest countries are the most important transmitters of spillovers, while during some periods (e.g. Greek, Irish, or Portuguese bailouts) this pattern may change. During that periods specific countries may become sources of turbulence. These results extend the study of Claeys and Vasicek (2012), examining sovereign bond yield spread data. The authors find that only 44% of all movements in the EU sovereign bond prices is caused by purely domestic factors. Alter and Beyer (2013) study selected euro area credit spreads using spillover index method. Although they show that spillovers between sovereigns are indeed of crucial importance, and policy-related events play essential role in evolution of the crisis, their results are not comparable. Different samples, reference entities, and the use of exogenous variables rule out direct comparisons. From Table 5 one can conclude that innovations to the SCDS premia of the Philippines, Malaysia and Korea were mostly responsible for the intra-regional spillovers. In the EMEA region, the spreads of Russia, Turkey, Bulgaria and Romania were the main drivers of error variance when forecasting SCDS returns (Table 6), while in Latin America the main drivers were innovations to the spreads of Brazil and Colombia. The emerging markets spillover results are consistent with the group of studies concerning intra-regional financial spillovers. For instance Lee (2010) finds the closest regional co- -movements among Latin American financial markets and the weakest in the Asian region, while Gębka and Serwa (2007) identify Russia as the main source of spillovers in the EMEA region. In general, from an economic point of view, the results again suggest that larger countries in terms of the size of debt markets have a more pronounced impact on regional SCDS spread

14 584 M. Adam returns. 9 It may be connected with the liquidity of the corresponding default swap markets, as a large amount of outstanding debt requires appropriate hedging by credit swaps. Gomez-Puig (2006) presents evidence that liquidity plays stabilising role in the euro area government bond market. The author shows that liquidity induces credit spread compression of the countries with large government bond markets and low bid-ask spreads of their bond prices. Following this argument, I check whether liquidity plays stabilising role also in the SCDS market. To this end, I look at net spillovers (equation 10). For two countries, given the same level of their transmitted spillovers (contributed to the system), higher net spillover of one country means that it receives less spillovers from the system than does the other country. In this regard the country with higher net spillovers is more stable. I regress net spillovers on time-averaged liquidity measures. 10 Figure 2 presents scatterplots of net spillovers against bid-ask spreads together with fitted regression lines and determination coefficients. All regression lines have negative slopes which means that lower transaction costs (lower bid-ask spread) are connected with a higher net spillover on average. As indicated by determination coefficients, which vary from 0.15 (Asia) up to 0.52 (EMEA), liquidity is an important stabiliser of country s SCDS spreads. These results prove robust against the liquidity measure. When using net notional amounts of SCDS contracts as a gauge, the liquidity-spillover relationship turns out to be even stronger (Figure 3) with determination coefficients reaching 0.67 in the case of the Eurozone. This time around the relationship is positive, which means that the larger the market depth, the higher the net spillover on average. Importantly, in most cases the determination coefficients are also statistically significant. Hence, liquidity plays a stabilising role, as a country s CDS spread is less vulnerable to spillovers from innovations to other SCDS premia. If a country wishes to strengthen its position in a group (to not only be the receiver of shocks), it should aim for stimulating the liquidity of the SCDS market referencing its debt. Above all, it should refrain from curbing the market activity. 7. Time-variation in SI: evidence of contagion Spillover strength is likely to vary over time and the presented spillover tables for the full sample are unable to capture these changes. To address this issue, I employ rolling-window analysis using 260-day subsamples, corresponding to a one-year period in the collected data. Figure 4 presents the variation of spillover measures for four regions. All of the regional SI estimates point to an increase between the first estimate and the end of January 2012 (end of the sample period). The largest surge is visible in the case of EMEA (from 76% to 87%), while the least change is noted for Latin America (from 79% to 81%). Overall, the indices exhibit large fluctuations, for instance the EMEA SI oscillates between 63% and 88%. Spillover index fluctuations are clearly irregular over time. Around September 2009 SI start to decrease, as the initial shock induced by the Lehman Brothers bankruptcy fades. The indices reach their minima around October 2009 and then climb up again. This is the point at which the sovereign debt 9 Consistently with the findings of Beber, Brandt and Kavajecz (2009), in crisis times this relationship would imply that more indebted countries (in absolute terms) will also produce more spillovers within the region. 10 One could argue that there might be endogeneity problems involved. However, following the strand of literature represented by Gomez-Puig (2006), it is reasonable to assume that liquidity should be treated as the explanatory variable. Granger causality tests are not conclusive when applied to the data at hand.

15 Spillovers and contagion in the sovereign crisis unfolded with the first rating downgrade of Greece in December The dramatic surge in the SI (by some 10 percentage points in a month) in the euro area merits attention. According to Forbes and Rigobon (2002), such a sharp surge in dependence across markets should be interpreted as contagion. During the spring of 2010, the dependence between Eurozone sovereign CDS spreads was the highest among all of the groups studied. The unprecedented measures ratified in May 2010, including a three-year EU/IMF-financed emergency package for Greece and a 750 billion euro ring-fencing mechanism did not prove enough to ease the crisis, which this time spread to emerging markets. EMEA countries are the most exposed to the Eurozone crisis, both through real and financial linkages. Hence, these countries also experienced contagion. In November 2010 Ireland applied for EU support and until May 2011 a strong dependence in regional SCDS markets is observed. It seems that the Stability and Growth Pact enhancements finally kicked in, which curbed the excessive spillover across the CDS markets somewhat. With the Portuguese bailout in July 2011 the markets started to co-move more closely once again. The euro area SI became increasingly volatile. As argued by De Grauwe and Ji (2012), a significant mispricing of risk took place through the course of 2010 and With the December 2011 fiscal compact agreement and the intervention of the ECB on the secondary debt markets, the spillovers stopped increasing. Important information is provided by the net intra-regional spillover measures (not to be confused by the total SI). The net measures for all countries are presented on the Figures The evolution of the net spillover measures (calculated in rolling samples) allows disentangling several periods of stress in the SCDS market. For example, in the euro area the first period (crisis build-up up to late December 2009) is characterised by a rather stable or even decreasing net spillovers. In the second period (defining contagious sovereigns) the crisis breaks up until May The Greek net spillovers are initially the strongest, which confirms systemic importance of the country at that time. In May 2010, the Greek spillover starts to fade (becoming the smallest in January 2012), while Irish, Spanish and Portuguese spillovers grow in force. In the third period, characterised by a relative low variance in the spillover measure, Portuguese, Spanish, Italian and Irish SCDS premia are mostly net contributors. These countries are now systemic. Another bout of the crisis starts in May 2011, this time around led by Spain, Belgium, Italy, and France (especially in the end of the period under review), which experience debt market tensions. After an initial drop in the SI, caused by Portuguese and second Greek bailouts, the intra-regional dependency increases. 11 It means that the country s characteristic (as a systemic country) may change over time. The rolling window analysis confirms that distressed countries become systemic, affecting the spreads of countries with sound fundamentals, but the influence of the countries which loose access to bond markets (Greece, Ireland, Portugal) gradually diminishes over time. The significance of countries identified as the sources of contagion may be confirmed in an additional exercise. Using the same rolling sample of 260 observations, first-order five-variable VAR models for the Greek SCDS spreads and SCDS spread indices for the Eurozone, Asian, EMEA and Latin American reference entities 12 are estimated. I focus on the statistical significance of the first variable (Greek SCDS spread). It transpires that the previously identified period of dynamic increase in SI for the Eurozone and dynamic increase of the Greek net spillover (manifested in 11 Similar periods of distress are visible in emerging markets. 12 The indices are constructed as weighted averages of SCDS spreads for reference entities constituting the regions (see Section 5 for details on the regions). Net notional amounts are used as the weights.