Was there contagion during the Greek sovereign debt crisis towards Italy and Spain?

(1)

1

Was there Contagion during the

Greek Sovereign Debt Crisis

towards Italy and Spain?

Master’s thesis

Author: Jelle Blom

Student number: 10059253

Supervisor: dr. D.J.M. Veestraeten

Second supervisor: Prof. dr. F.J.G.M. Klaassen

University of Amsterdam

Faculty Economics and Business

International Economics and Globalization

June 2015

(2)

2 Statement of Originality

This document is written by Student Jelle Blom who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

3

1. Introduction

The bankruptcy of Lehman Brothers in September 2008 shocked the financial sector and ushered in a crisis not seen since the nineteen thirties. The interbank lending market dried up and banks all around the world suffered from funding problems. All major central banks eased lending conditions and conducted enormous operations to provide banks with liquidity. For the euro area, the financial crisis was also accompanied by a debt crisis after an announcement of the Greek government on 5 November 2009. The new Greek government announced that the public deficit was 12.7% compared to the previously believed 6%, and that the earlier deficits were larger than previously stated (Lane, 2012).

Graph1. Monthly data on 10-year government bond yields for Greece, Ireland, Spain, Italy, Portugal and Germany. Source: Thomson Reuters DataStream.

Following the worsening situation in Greece, other euro area countries also suffered from increasing interest rates on their debt, see graph 1. These countries were together, mockingly, called the PIIGS, acronym for Portugal, Ireland, Italy, Greece and Spain. The fundamentals in these countries were deteriorating, especially because of high public deficits and lower economic growth that followed the financial crisis. However, the deterioration in the fundamentals due to the credit crisis might not have been the only reason for the increased interest rates in these euro area countries. Another influence on the interest rates in these countries could have been a possible contagious effect from Greece to the other PIIGS countries. If contagion occurs in the sovereign bond market this means that, during a crisis,

0% 5% 10% 15% 20% 25% 30% 1 -1 -2006 1 -6 -2006 1 -11 -2006 1 -4 -2007 1 -9 -2007 1 -2 -2008 1 -7 -2008 1 -12 -2008 1 -5 -2009 1 -10 -2009 1 -3 -2010 1 -8 -2010 1 -1 -2011 1 -6 -2011 1 -11 -2011 1 -4 -2012 1 -9 -2012 1 -2 -2013 1 -7 -2013 1 -12 -2013 1 -5 -2014 1 -10 -2014 1 -3 -2015 Greece Ireland Spain Italy Portugal Germany

(5)

5

there can be extra transmission of higher interest rates from one country to other countries than during tranquil periods, because the correlation between those interest rates has increased. As stated by Pericoli & Sbracia (2003) this extra transmission of a negative shock can lead to a decline in welfare in the other countries even though the state of the economy is fine and is thus an undesirable phenomenon.

In this thesis the definition of contagion is an increase in cross-market linkages, during a crisis in one of the countries. This is the general definition used in the literature, see for example Forbes & Rigobon (2002). This definition makes the clear distinction between cross-market linkages during ‘tranquil’ times, which are called interdependencies and the increase in these cross-market linkages during ‘crisis’ periods, which is then called contagion. Several papers examined the contagion effects emanating from the Greek sovereign debt crisis. For example, Andenmatten & Brill (2011) use data on credit defaults swap premiums (CDS-premiums) and compare the correlations between those CDS-premiums for European countries. The authors examine the crisis up to July 2010 and they find that during some parts of the Greek sovereign debt crisis only interdependencies were present, but during other periods there was contagion. Constancio (2012) shows that in 2011, between 30% and 40% of the Spanish and Italian government bond yields could be explained via contagious effects from Ireland, Portugal and Greece combined. The author gives the following example. On 18 July 2011 the Italian and Spanish 10-year government bond yields increased by respectively 100 and 80 basis points, without any worsening of the Italian or Spanish fundamentals. However at that time, there were talks about a second rescue package for Greece which would include a large private sector involvement1_{. The increased interest}

rates for Spain and Italy could thus be caused by contagion. Opposing these findings are the results found by the paper of Caporin et al. (2015). They use an approach based on quantile regression and they find hardly any evidence for contagion during the Greek sovereign debt crisis.

To contribute to the existing literature, this thesis examines the correlation between 10-year government bond yields for the country pairs Greece & Spain and Greece & Italy. Spain and Italy are chosen for the following reasons. First, when plotting the 10-year government bond yields for these countries one can see an increase in the yields for Italy and Spain after an increase in the yields of Greece, especially during 2011 and 2012. This could just be the continuation of high interdependencies that already existed before the crisis, but it could also be caused by contagion. High interdependencies between countries occur for example because the two countries trade amongst each other (Moser, 2003). Moreover, since the end of 2014 the Greek interest rates are increasing, but the interest rates for Spain and

1_{For details on the Greek debt restructuring and the private sector involvement see: Zettelmeyer, Trebesch and}

(6)

6

Italy are decreasing2_{. We want to compare cross-market linkages during this latest eruption}

of the Greek crisis to the cross-market linkages during the beginning of the crisis. Lower cross-market linkages can indicate an isolation of the Greek crisis. Second, Italy and Spain are large economies compared to Greece. These large countries have large sovereign debt markets and in this thesis we want to find out whether contagion from Greece can affect these large debt markets3_.

This thesis will try to answer the question: ‘Was there contagion during the Greek sovereign debt crisis towards Italy and Spain?’ by examining the evolution of cross-market linkages from Greece to Italy and Spain. We will thus complement to the existing literature by examining the entire crisis period up till March 2015 and try to find out whether there has been contagion from the Greek sovereign debt crisis.

To answer the research question, the methodology of Forbes & Rigobon (2002) will be used. They use the method of comparing correlations between markets before and during a crisis. This methodology is chosen because it provides the possibility to test for contagion directly. The reason for this is that a significant increase in correlation would indicate increased cross-market linkages, and thus the existence of contagion. Another reason for choosing this methodology is that it avoids having to differentiate between different channels that could cause contagion. The data that are used for this research are the 10-year government bond yields for the countries involved.

The remainder of this thesis is structured as follows. The second part gives an overview of the relevant literature and discusses different methods and results found by other authors. The third section will describe the methodology and the data. In the fourth part the results are presented and discussed. The fifth part concludes.

2_{The increase in yields was caused by increased tensions caused by disagreement over the third bail-out package}

for Greece and the fear and implementation of new elections.

3

During the crisis period 2009-2015 the size of Spanish government debt was around 2 to 3 times the size of that

of Greece and the Italian government debt was between 6 to 7 times the size of that ofGreece. These estimations

are made by the author on the basis of data from the world economic outlook.

(7)

7

2. Literature review

This literature review is structured as follows. First, there will be some elaboration about the possible channels through which contagion can take place. Second, there will be a review of methodologies used in the literature.

Dornbusch, Park & Claessens (2000) differentiate between two types of contagion, namely fundamental based and non-fundamental based contagion. There are three types of fundamental based causes of contagion. First there can be a common shock to both markets, that increases the correlation between the markets. For example, an increase in the oil price can affect both countries negatively. A second cause is that of trade links and devaluation. If a country that is hit by a crisis devaluates its currency to improve its competitiveness, the other country can suffer from a decrease in competitiveness. This channel of contagion via the exchange rate, will not be relevant in this thesis because the countries that are involved all have a common currency, thus competitive devaluation is not an option. A third channel is that of financial linkages. This channel is used in the model of Bolton & Jeanne (2011). They have created a theoretical model about how contagion can spread via the possession of international government debt by banks. A crisis in country ‘A’ can spread to country ‘B’ through the change in price of the bonds of country ‘A’ held by banks in country ‘B’. During a crisis, in which the value of those bonds is affected, the linkages between countries ‘A’ and ‘B’ can thus grow stronger than normal.

The second group of contagion channels is non-fundamental based (Dornbusch et al. 2000). These channels work via the behavior of investors during a crisis. The first one is the liquidity channel. During a crisis in a country, asset prices in that country decrease and the liquidity in the asset markets of the country can decrease. This means that selling assets from this country becomes more difficult. The following example will illustrate the possible contagious effect. Investors diversify their investments by investing in different countries. When a crisis hits one of the markets in which the investors invest, they might want to change their portfolios or increase the cash part of their portfolios. But because the liquidity in the market that is hit by the crisis is lower, they will find it difficult to sell their assets in that market without making large losses. The investors can then choose to sell assets from other markets that are more liquid and through that lowering the prices of those assets. Because of these actions there can thus be extra transmission between the countries during a crisis and the liquidity channel can thus be contagious.

Another channel in the group of non-fundamentally based contagion is that of information asymmetries. If the information in the market is not complete then investors can, possible incorrectly, assume that a crisis in one country will also lead to a crisis in

(8)

8

another country. Summers (2000) states that investors might, irrationally, withdraw from several markets at the same time without looking at the underlying fundamentals, because of panic, herding or positive feedback trading. If the countries are thus seen as both being in a crisis or both to end up in a crisis, the cross-market linkages between those countries will increase.

A large part of the literature on contagion uses the method of comparing the correlation between markets before a crisis and during a crisis. A significant increase in the correlation then suggests that cross-market linkages have increased and that there was contagion. Forbes & Rigobon (2002) use this method to determine whether the 1997 Asian crisis, the 1994 Mexican devaluation and the 1987 US stock market crash were contagious. Compared to the earlier literature that employed correlations between markets, the authors make an adjustment for the increase in volatility that often comes along with a crisis. In their paper they prove that the correlation between markets is increasing in the variance of the market that is hit by a negative shock (the source country of the contagion). To clarify this, we present the following intuitive example. Suppose that there are two stock markets, market ‘A’ and market ‘B’. Assume that part of the return in market ‘A’ is based on the return in market ‘B’. From this it follows that also part of the variance of the return in market ‘A’ is determined by market ‘B’. Now suppose that market ‘B’ is hit by a crisis and that its volatility increases significantly. Because part of the variance in market ‘A’ is determined by market ‘B’, this also leads to an increase in the variance in market ‘A’. This increase in variance is due to market ‘B’, thus during the crisis, a larger part of the variance of market ‘A’ is determined by market ‘B’. This will also lead to an increase in the covariance between markets ‘A’ and ‘B’, because they now move more in the same directions. This increase in covariance will lead to an increase in correlations. From this example one can see that the correlation between markets can increase during a crisis (which increases the variance in market ‘B’), without an increase in cross-market linkages between both markets4_{. A direct comparison of correlations}

between markets to determine whether contagion has taken place would thus lead to biased results if the variance in market ‘B’ has increased during the crisis.

We now show formally that the correlation between two bond markets is an increasing function of the variance in the source country if the two bond markets are linearly related. This proof is from Corsetti et al. (2005), with slightly adjusted notation and showing some more steps. Assume the returns in the bond markets ‘A’ and ‘B’ that are linearly related:

(9)

9

Where and are the returns in the bond markets of countries ‘A’ and ‘B’. The error term is denoted by and is independent of . and are constants. indicates the strength of the relation between the markets.

The variance (var) of , covariance (cov) and the correlation (corr) between the markets are as follows:

In the last step it can be seen that if the variance in market ‘B’ increases, the denominator becomes smaller and thus the correlation between returns in markets ‘A’ and ‘B’ will automatically increase.

In their paper Forbes & Rigobon show a method to correct for this increase in volatility, which is done with the following formula and which this thesis will apply as well:

Where is the unconditional correlation coefficient between markets, is the

correlation coefficient between markets, conditional on the market volatility in the source country and is the relative increase in variance in the source country. The following hypotheses are tested:

The null hypothesis states that there is no significant increase of the unconditional correlation coefficient. This means that the cross-market linkages during the crisis compared

(10)

10

to the cross-market linkages during the tranquil period have not increased. The null hypothesis corresponds to ‘no-contagion, only interdependencies’. In other words, the cross-market linkages during the crisis are a continuation of the cross-cross-market linkages from before the crisis. These linkages during tranquil times, interdependencies, are due to the fundamental links between the countries, for example the trade and financial linkages. If the null hypothesis is rejected this means that the cross-market linkages have increased, the transmission between the markets was stronger during the crisis period and thus contagion has taken place. In the remainder of this thesis ‘crisis’ will be abbreviated by ‘c’ and ‘tranquil period’ will be abbreviated by ‘t’ in order to avoid notational clutter.

Forbes & Rigobon (2002) state that their methodology might have some weaknesses. First, the results are only valid if there are no large exogenous shocks to the countries involved. Second, it must be clear which of the two countries is the source country. With their method Forbes & Rigobon find almost no evidence of contagion in the markets that they researched. They conclude that the high level of cross-market linkages that exists between the markets are due to high interdependencies that are present during both tranquil and crisis periods.

Corsetti, Pericoli & Sbracia (2005) criticize the methodology of Forbes & Rigobon (2002) and state that it is biased towards the null hypothesis. They show that if the increase in variance in the source country is due to increase in country specific risk, the correction proposed by Forbes & Rigobon is too large, such that the null hypothesis is accepted too often. This is because Forbes & Rigobon assume that the entire increase in variance is due to a common factor and thus also correct for the entire increase in variance. Corsetti et al. (2005) argue that the correction for the increased variance in the source country should be limited to the amount of increase in variance that is common to both countries, because only this part of the variance can influence the size of the correlation coefficient that is measured. By implementing this correction the authors find some evidence of contagion during the Hong Kong stock market crisis. Their total sample consisted of 17 countries. The authors found evidence for contagion from Hong Kong to Singapore, the Philippines, France, Italy and the United Kingdom. This result contradicts the result from Forbes & Rigobon (2002), who only found evidence for contagion during the Hong Kong stock market crisis from Hong Kong to Italy.

Andenmatten & Brill (2011) use the methodology of Forbes & Rigobon (2002) to examine possible contagion from Greece during the recent crisis and find that there is some evidence of contagion from Greece to Spain and Italy. The authors correctly state that when finding proof of contagion with a conservative method, this is relatively strong proof. They conclude that the Greek sovereign debt crisis was characterized by ‘both contagion and interdependence’. A recent paper of Caporin et al. (2015) finds results that contradict the

(11)

11

results from Andenmatten & Brill (2011). The authors employ quantile regression, using data on sovereign bond yields with a maturity of 5 year, to examine contagion during the Greek sovereign debt crisis. Caporin et al. (2015) do explicitly not define Greece as the sole source of contagion. Instead they compare country pairs consisting of the PIIGS-countries, Germany, France and the United Kingdom. They define a crisis period that starts in November 2008, shortly after the collapse of Lehman Brothers and this period does thus include the revelation of the Greek sovereign debt crisis. Their results show that the cross-country linkages are constant during the crisis period. Since the revelation of the Greek sovereign debt crisis is included in the crisis period, the constant cross-country linkages imply that there was no contagion during the Greek sovereign debt crisis. The authors explain this by stating that after the collapse of Lehman Brothers, investors already expected that the sovereign bond yields for euro area countries would start to diverge. This led to lower cross-market linkages. In addition the authors conclude that the crisis in the peripheral countries in the euro area has only increased the volatility in the sovereign debt markets and has not changed the cross-market linkages. The increased variance during the crisis period does, however, underline the importance of correcting for the increased variance when employing the methodology of Forbes & Rigobon (2002), as this thesis does.

Gorea & Radev (2014) study the joint probability of default after a shock to one country. An increase in the joint probability of default can be seen as an increase of cross-market linkages and thus is quite similar to the definition of contagion used in this thesis. The joint probabilities of default are calculated via CDS-premiums. In their paper the authors divide the euro area countries into country pairs. They make pairs containing either 0,1 or 2 countries from the PIIGS-group. Interestingly, their results show that there only is evidence for contagion in pairs that consisted of 2 countries from the PIIGS-group.

Beirne & Fratzscher (2013) use a regression based approach to identify contagion from the Greek sovereign debt crisis. However in contrast to the findings of Andenmatten & Brill (2011), their findings do not show increased cross-market linkages and thus the authors conclude that there was no contagion during the Greek sovereign debt crisis. They find that when controlling for fundamentals, such as public debt level, the increase in yields for the PIIGS countries is not caused by increased cross-market linkages, but rather by a change in the way that markets price the fundamentals of those countries. During the crisis, the quality of a country’s fundamentals became more important for market participants. This led to a change in pricing of sovereign bonds. Countries with relatively weak fundamentals paid higher interest rates during the crisis than could be expected from the way fundamentals were priced before the crisis.

(12)

12

To summarize, this section has presented an overview of the channels through which contagion takes place. We also presented different methodologies used in the literature to examine contagion and we have shown the results that other papers found when examining contagion during the Greek sovereign debt crisis. In the next section we will describe our methodology and data.

(13)

13

3. Methodology and data

This section will start by describing the methodology that is applied in this thesis, followed by an explanation about how we will examine the crisis period by splitting the latter in different sub-periods. This is followed by a paragraph describing a practical manner to combine the methodology with the critique expressed by Corsetti et al. (2005). Then the assumptions underlying the methodology will be discussed. This is followed by a description of 10-year government bond yield data from Greece, Spain and Italy.

3.1 Methodology

In this thesis the methodology from Forbes & Rigobon (2002) is used to examine the possible contagion during the Greek sovereign debt crisis from Greece to Italy and Spain. This methodology compares the correlation between markets during ‘tranquil’ times and the correlation between markets after a crisis has occurred in one of the countries. If an increase in correlation between the markets is shown, this indicates contagion. A direct comparison of correlations can lead to misleading results. As Forbes & Rigobon (2002) have shown, if the volatility in the source country of the contagion increases after a crisis has occurred this will automatically lead to an increase in the correlation between markets. If one does not correct for this increased volatility, one might incorrectly conclude that contagion has taken place. The correlation that is measured directly and that thus is conditional on the market volatility is called the conditional correlation. After correcting for this increase in volatility one can obtain the unconditional correlation. The methodology can be summarized in the following formula:

Where is the unconditional correlation coefficient, is the conditional correlation. is the relative increase in variance of the source country. Such that:

In which is the variance in the market of the source country during the crisis period and is the variance in the market of the source country during tranquil times.

The following hypotheses are tested:

(14)

14

If the null hypothesis is rejected this means that the correlation between the markets has increased and thus there is evidence of contagion. To test whether the correlation between markets has increased we can conduct a t-test.

Where =

Where ‘var’ stands for variance and ‘cov’ stands for the covariance. If we assume independence, then the covariance between the crisis period and the tranquil period will be zero and the variance simplifies to5_:

If we then assume normality and use an asymptotic approximation we can rewrite the variance to6_:

Where is the number of observations during the crisis period and is the number of observations during the tranquil period7_{. The final formula for the t-test then becomes:}

It is important to state that this thesis defines the period that is named ‘tranquil times’ as the period before the crisis, which is in agreement with other authors8_{, but opposing Forbes &}

Rigobon (2002). The latter paper compares the correlation of the crisis period with the correlation of the entire period. The authors do not give an explanation for doing so, but they state that if they used only the tranquil period to compare with the crisis period instead of the entire period, their results do not change. We are in favor of applying a clear separation between the tranquil period and the crisis period. We want to test for increased cross-market linkages during a crisis, so it is more logical to compare the crisis period with a tranquil period, thus excluding the crisis period in the latter. Then the assumption of independence between the two periods, which is made to go from equation (4) to equation (5), becomes more plausible. We follow Forbes & Rigobon (2002) in the use of a normal t-test, described in equation (7). Other papers have proposed a Fisher transformation, we do not implement

5_{The covariance of two independent samples is zero. See for example Stock & Watson (2011).}

6_{For asymptotic approximations see Kendall & Stuart (1969).}

7_{This way of re-writing the standard error is taken from Andenmatten & Brill (2011), only some notational}

changes have been made.

(15)

15

this because it only improves the results when correlations are relatively low. Our results show some high correlation and thus a normal t-test is preferred9_{. The critical value of our}

one-sided test will be 1,65. With a t-value higher than that we can reject the null hypothesis of no contagion at a 5% significance level.

3.2 Subdivision of the crisis period

The aim of this thesis is to examine the development of possible contagion during the Greek sovereign debt crisis and specifically to examine whether there was contagion during and after the period of the Greek elections of January 2015. To examine the evolution of contagion during the Greek debt crisis, we follow the approach of Andenmatten & Brill (2011) and divide the crisis period into sub-periods. These sub-periods consist of respectively 20, 40 and 60 observations. Because only trading days are included in the dataset, these periods are largely comparable to 1, 2 or 3 months respectively. With this subdivision of the crisis period we can examine during which months there might have been contagion from Greece to Spain and Italy. We choose 5 November 2009 as the starting point of the crisis, because on that date the Greek government announced that the budget deficit was way higher than previously stated. After this event the Greek crisis really took off (Lane, 2012). To give more insight in how this subdivision works we describe the first crisis periods, consisting of 20 trading days. Our crisis period starts at 5-11-2009, so now we look in our data and search for the date that is 20 trading days later, which is in this case 1-12-2009, so this is our first crisis period. Our second crisis period then starts at 2-12-2009 and lasts until 4-1-2010. This procedure is applied to the entire crisis period and is similar for the 40 and 60 day periods. This results in 68 periods of 20 observations, 34 periods of 40 observations and 22 periods of 60 observations. The last period of 60 observations ends on 2-2-2015. This is because after this date there are not enough data point in the sample to form another period of 60 observations.

The period after the starting point of the crisis will be divided into sub-periods as stated above. The correlation between the country pairs during these periods will be conditional on the variance in the Greek yields. The adjustment of Forbes & Rigobon (2002) will be applied to make this correlation unconditional. We will thus correct for the increase in variance of the Greek sovereign bond yields between the chosen crisis period and the tranquil period. The tranquil period will be defined as the period starting in mid-February 2001 lasting until the 4th_{of November 2009}10_.

9_{The Fisher transformation is used in Andenmatten and Brill (2011) and Dungey et al. (2005).}

10_{This date is chosen because of data availability and because it is the earliest data point available for which}

(16)

16 3.3 Application of the critique of Corsetti et al. (2005)

The main point in the critique of Corsetti et al. (2005) on the methodology of Forbes & Rigobon (2002) is that the correction that they propose overcorrects the conditional correlation coefficient. The authors state that the increase in variance in the markets, that often comes along with the crisis, consists of a country specific part and a common part. Only the common part is transmitted to the other country and can thus bias the correlation coefficient. However, Forbes & Rigobon (2002) correct for the entire increase in variance in the source country and thus assume that the increase in variance is completely due to increase in the common part of the variance. In this thesis we propose a simple way to capture the critique of Corsetti et al. (2005) and we will do the same analysis as stated in the previous paragraphs, but with this adjustment.

We propose the following adjustment. Instead of correcting for the entire increase in variance of the source country, we do only correct for the increase in the variance of the source country with a maximum that is set at the increase in variance in the other country. In doing so we capture the maximum amount of common variance that is possibly transmitted from the source country to the other country. This method will thus exclude the country specific increase in variance from the correction factor and thus results in a test that is less biased to the null hypothesis. In the remainder of this thesis this approach is called ‘Corsetti’ approach.

3.4 Underlying assumptions in the methodology

Forbes & Rigobon (2002) state that using their methodology requires that two important assumptions are made. Before using the methodology it is useful to have a look at these assumptions and to discuss whether or not these assumptions are reasonable. The first assumption is that there are no large exogenous shocks to the countries during the crisis period. It is difficult to measure the extent to which this assumption is true. For example, right before the Greek sovereign debt crisis began, the world was hit by a major financial crisis and during the crisis periods large changes in oil prices were present. These could account as large exogenous shocks. Forbes & Rigobon (2002) show in their appendix that when countries are well interlinked, their methodology remains ‘fairly’ accurate even when exogenous shocks are present. A possible way to quantify the degree of inter linkages would be to look at the correlation between the sovereign bond yields during the tranquil period. This is done in section 4, which shows that the correlation between the sovereign bond yields of the examined countries is already high during the tranquil time.

(17)

17

It thus is unlikely, that exogenous shocks will have a large effect on our results. In addition, the countries in the sample are all member of the euro area and thus have the same currency. This factor will contribute to a similar reaction to exogenous shocks. Thus we can probably assume the effects of exogenous shocks to be small.

The second assumption that must be made is that there may not be a strong feedback effect from the affected country to the source country of the contagion. This would lead to an endogeneity problem of simultaneous causality which would bias the results. It appears to be evident that Greece is the source country of the crisis. The initial announcement of higher government deficits that started the Greek debt crisis is a clear starting point of the crisis. Furthermore, when looking at the yields of government debt, as shown in graph 2, the yields of Greece increased earlier and more than those of Spain and Italy. So it appears to be safe to assume that Greece is the source country. Forbes & Rigobon (2002) show that when the feedback effect is expected to be small, their methodology remains accurate. Thus, because there is evidence that Greece is the source country and the expected feedback effects from Spain and Italy are small, we can safely ignore the feedback effect in our research.

3.5 Sovereign bond yields data

This section will give a description of the data of the 10-year government bond yields of Greece, Italy and Spain. The data are obtained from the global yield curve section on quandl.com. We start by graphing the data. We show two graphs both containing data on the crisis period. Graph 2 shows the 10-year government bond yields of Spain and Italy compared to that of Greece. Graph 3 is drawn to show the dynamics of the government bond yields from Spain and Italy, since the dynamics are hidden in graph 2 due to the large values of the Greek yields. In this graph it is shown that the 10-year government bond yields of Spain and Italy follow the same trend, but that there are differences at certain periods.

(18)

18

Graph 2. Daily 10-year government bond yields of Italy, Spain and Greece. Source quandl.com

Graph 3. Daily 10-year government bond yields of Italy and Spain. Source quandl.com

From graph 1 in the introduction we can see that in the days before the revelation of the Greek sovereign debt crisis, the interest rates of Italy, Greece and Spain were generally around the same level. This is in line with the phenomenon that the sovereign bond yields of the euro area countries converged to the level of the German yields (Ehrmann et al., 2011). This could be an indication of high interdependencies between the countries during tranquil times, but it can also be because of mispricing of fundamental risk as stated by Beirne &

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 1-1-2009 1-1-2010 1-1-2011 1-1-2012 1-1-2013 1-1-2014 1-1-2015 Italy Spain Greece 0% 1% 2% 3% 4% 5% 6% 7% 8% 1-1-2009 1-1-2010 1-1-2011 1-1-2012 1-1-2013 1-1-2014 1-1-2015 Italy Spain

(19)

19

Fratzscher (2013). A second observation, seen in graph 2, is that the interest rates on the Greek government debt started to increase from late 2009 onwards, which is in agreement with our definition of the crisis period, that also starts in late 2009. A third observation is that the interest rates for Spain and Italy seem to follow the trend from Greece and also reach their highest values during 2011 and 2012. Another observation is that when looking at graph 2 we can expect the data to be non-stationary, the mean and the variance appear to be changing over time. This observation will be formally tested later in the thesis. A last observation from graph 2 is that at the end of 2014 the yields on Greek debt increase again, but that the yields on the Spanish and Italian bonds keep decreasing. This could mean that the transmission from Greece to Spain and Italy has decreased and that the sovereign debt crisis has been contained for Spain and Italy. This last observation is an important motivation for writing this thesis and it will be formally tested later in the thesis.

The following tables summarize the 10-year sovereign bond yield data for Italy, Spain and Greece.

Table 1: Total period descriptive data (16-02-2001 until 31-03-2015)

Country #observations Mean Std. Dev. Minimum Maximum

Italy 3564 4.37% 0.86 1.10% 7.64%

Spain 3564 4.29% 0.93 1.05% 7.50%

Greece 3564 7.97% 6.82 3.21% 48.60%

Table 2: Tranquil period descriptive data (16-02-2001 until 04-11-2009)

Italy 2204 4.41% 0.52 3.21% 5.56%

Spain 2204 4.23% 0.56 3.03% 5.47%

Greece 2204 4.51% 0.61 3.21% 6.18%

Table 3: Total crisis period descriptive data (05-11-2009 until 31-03-2015)

Italy 1360 4.30% 1.22 1.10% 7.64%

Spain 1360 4.40% 1.32 1.05% 7.50%

Greece 1360 13.59% 8.39 3.21% 48.60%

Comparing the data for the tranquil period and the data for the crisis period, we can make some important observations. We see that the mean 10-year interest rate for Spain and Italy over the periods remains almost the same, but that the volatility in the interest rates more than doubles, this observation matches the conclusion of Caporin et al. (2015). For Greece we see that the mean interest rate triples and that the volatility increases more than 10 times. This stresses the importance of correcting for this increased volatility in the source country, as proposed by Forbes & Rigobon (2002).

(20)

20

As stated in the previous section, the data appear to be trending and thus the data are expected to be non-stationary. This can be a problem in our research. This problem is called ‘the spurious regression problem’. This is a situation in which two time-series both have a trend. When one looks at the correlation between these time-series they can appear to be strongly related, because they happen to move in the same direction because of the trend. This could thus lead to a bias towards too high correlation coefficients between the sovereign bond yields, that we examine. In the following section we will formally test for non-stationarity with an augmented Dickey-Fuller test.

3.6 Augmented Dickey-Fuller test results

This section shows the results of the augmented Dickey-Fuller tests on the data for sovereign bond yields. The tests are done with inclusion of a trend and no lags. The results show that under these specifications none of the original time series is stationary.

Dickey-Fuller test for unit root Italy Number of observations = 2802 Interpolated Dickey-Fuller

Test statistic 1% critical value 5% critical value 10% critical value

Z(t) -0.660 -3.960 -3.410 -3.120

MacKinnon approximate p-value for Z(t) = 0.9757

Dickey-Fuller test for unit root Spain Number of observations = 2802 Interpolated Dickey-Fuller

Z(t) -2.767 -3.960 -3.410 -3.120

Dickey-Fuller test for unit root Greece Number of observations = 2802 Interpolated Dickey-Fuller

Z(t) 2.973 -3.960 -3.410 -3.120

Choosing other specifications for the augmented Dickey-Fuller test only changes the result for Spain. In the case of not including a trend, the test rejects non-stationarity at a 10% level. In the case of including a drift non-stationarity is rejected at a 1% level.

(21)

21

Now that we know that we cannot reject non-stationarity in the data, we proceed by taking first differences of our data, as is standard procedure when confronted with non-stationary data. The following three tables show the results for our countries, using first differences.

Dickey-Fuller test for unit root Italy Number of observations = 2801 Interpolated Dickey-Fuller

Z(t) -47.778 -3.960 -3.410 -3.120

Dickey-Fuller test for unit root Spain Number of observations = 2801 Interpolated Dickey-Fuller

Z(t) -60.089 -3.960 -3.410 -3.120

Dickey-Fuller test for unit root Greece Number of observations = 2801 Interpolated Dickey-Fuller

Z(t) -114.184 -3.960 -3.410 -3.120

Choosing other specifications for the test on first differences does not change the outcome shown in the tables above. So we can conclude that using first-differences solves the problem of non-stationarity in the data and thus first differences are used in our analysis. In addition to this technical reason, the use of changes in yields is also more closely related to our research question in a theoretical point of view. This is because we want to examine whether the changes in sovereign yields in Greece have a stronger effect on the changes in sovereign yields in Spain and Italy during the Greek sovereign debt crisis than during the tranquil period. So we will use first differences in our analysis. This is the same approach as taken by Andenmatten & Brill (2011).

(22)

22

4. Results

This chapter describes the results from our analysis of contagion during the Greek sovereign debt crisis towards Spain and Italy. As stated in the section on the methodology we start by taking first differences of our data, to make the data stationary. Then we test whether the correlation between the sovereign bond yields increased during the crisis period compared to the defined tranquil period, taking into account the possible increased variance in the source country, Greece.

We will present our results for 3 different specifications of our methodology. The first one will show the results when not correcting for the possibly increased variance in the source country. The second will give the results for the test when adjusting for the entire increase in variance of Greece during the crisis period. The last specification will show the results when we correct for the increase in variance up to the level of increase in variance in the other country. With this last method we try to correct the methodology for the critique of Corsetti et al. (2005).

We will also show our results for 3 different timings of the tranquil period, to better compare our results with the results of other authors and to examine the effect of choosing different tranquil periods on the results. The first specification of the tranquil period is called ‘the long tranquil period’, which is the defined tranquil period by this thesis. Our tranquil period lasts from mid-February 2001 until the 4th_{of November 2009. The second}

specification is called ‘the short tranquil period’, which lasts from the first of January 2008, until the starting point of the crisis, this is the tranquil period chosen by Andenmatten & Brill (2011). The authors chose this specification to match the duration of their tranquil period to that of Forbes & Rigobon (2002). The third specification is called the ‘pre-Lehman tranquil period’ and it lasts from the 1st_{of January 2003 until the 29}th_{of December 2006. This period}

is the tranquil period chosen by Caporin et al. (2015). The authors chose this period because they wanted to exclude the fluctuations caused by the financial crisis in the United States and elsewhere. This will result in 3 tables for each country pair. The results for Greece & Spain are shown first, followed by the results for Greece & Italy.

(23)

23 4.1 Results for the country pair Greece & Spain

Hereafter follow the results of our examination of contagion from Greece towards Spain. It is important to remember here that the maximum number of indications for the 20, 40 and 60 trading days crisis periods are 68, 34 and 22 respectively. These are thus the maximum amount of indications of contagion that can be obtained for each durations of a crisis period.

Table 4. Number of indications of contagion when applying the long tranquil period Greece & Spain

‘long tranquil period’ Adjustment

technique for corr.

Unadjusted Forbes & Rigobon Corsetti approach

Correlation tranquil 0.4315 0.4315 0.4315

20 trading days 1* 1*** 0 0

*=10% significance level, **=5% significance level, ***=1% significance level.

Table 5. Number of indications of contagion when applying the short tranquil period Greece & Spain

‘short tranquil period’ Adjustment

technique for corr.

20 trading days 6* 1** 1*** 0 3* 1**

40 trading days 4* 1** 2*** 0 3* 1** 1***

60 trading days 3** 2*** 0 2* 2**

(24)

24

Table 6. Number of indications of contagion when applying the pre-Lehman tranquil period Greece & Spain

‘pre-Lehman tranquil period’ Adjustment

technique for corr.

20 trading days 1** 0 0

We start by describing the results shown in table 4. This table compares the tranquil period lasting from mid-February 2001 to the 4th_{of November 2009, with the crisis period. The}

crisis period is the same in all cases and lasts from the 5th_{of November 2009 until the end of}

our sample on 31st_{of March 2015. The correlation between the changes in 10-year}

government bond yields of Greece and Spain during this tranquil period is 0.4315. The results for the unadjusted methodology show very little evidence of contagion. For each duration of the crisis sub-periods there is one period in which contagion is found11_{. When}

applying the methodology of Forbes and Rigobon or the ‘Corsetti’ approach we do not find any evidence of contagion, when using the long tranquil period. We do conclude that there is no evidence of contagion from Greece to Spain when comparing the crisis period with the long tranquil period.

We will now look at the results in table 5. Here we use the tranquil period that was proposed by Andenmatten & Brill (2011) and which lasts from the 1st_{of January 2008 until}

the 4th_{of November 2009. The correlation between the changes in 10-year government bond}

yields of Greece and Spain is then 0.3398. The unadjusted methodology finds some evidence of contagion namely 2 indications for the 20 trading day periods, 3 indications for the 40 trading day periods and 5 indications during the 60 trading day periods12_{. The methodology}

of Forbes & Rigobon does, also with this specification of the tranquil period, not find any evidence of contagion. The ‘Corsetti’ approach finds results that are between the results of the other approaches. With this approach we find 1 indication for the 20 trading day periods, 2 indications for the 40 trading day periods and 2 indications during the 60 trading day

11_{We want at least a significance level of 5% because this is the required level used in the literature, see Forbes and}

Rigobon (2002).

12_{This difference in the number of indications of contagions for the different durations of the crisis period is}

caused by the fact that the longer the crisis period is, the smaller the denominator in equation (7) becomes :

.

This results in a higher t-value for the larger crisis periods than for smaller crisis periods when testing the same

(25)

25

periods. The results differ somewhat from the results obtained with the ‘long tranquil period’. This may find its origin in the fact that the correlation between first differences in yields during the tranquil period defined by Andenmatten & Brill (2011) is lower. This stresses the importance of the chosen tranquil period in relation to finding evidence of contagion.

In table 6 we compared the crisis period to the tranquil period that was proposed by Caporin et al. (2015) and which lasts from the 1st_{of January 2003 until 29}th_{of December}

2006. The correlation between first differences in 10-year government bond yields during this period is the highest of the three periods examined, namely 0.4926. The results in table 6 are similar to that of table 4. Only with the unadjusted methodology we find little evidence of contagion from Greece towards Spain and for the other two methodologies we do not find any evidence for contagion during the Greek sovereign debt crisis from Greece towards Spain. Comparing the results for the country pair Spain & Greece, we observe that the choice of the tranquil period is important when testing for contagion because it can influence the results. A low correlation during the tranquil period results in finding evidence for contagion more often, as can be seen from the results in table 5.

4.2 Results for the country pair Greece & Italy

This section will describe our results for the examination of contagion from Greece towards Italy.

Table 7. Number of indications of contagion when applying the long tranquil period Greece & Italy

‘long tranquil period’ Adjustment

technique for corr.

20 trading days 1* 0 0

40 trading days 0 0 0

(26)

26

Table 8. Number of indications of contagion when applying the short tranquil period Greece & Italy

‘short tranquil period’ Adjustment

technique for corr.

20 trading days 2* 0 1*

40 trading days 2* 0 0

60 trading days 2* 0 1*

Table 9. Number of indications of contagion when applying the pre-Lehman tranquil period Greece & Italy

‘pre-Lehman tranquil period’ Adjustment

technique for corr.

We again start by looking at the number of indications of contagion when comparing the crisis period to the long tranquil period. For the country pair Italy & Greece this results in zero indication of contagion on a significance level of 5% or higher for each of the three specifications of our methodology. For this specification of the tranquil period we do not find any evidence of contagion.

Next we compare the correlations between changes in the Italian and Greek 10-year government bond yields during the crisis to the correlation between those yields during the short tranquil period, as proposed by Andenmatten & Brill (2011). Although the correlation during the tranquil period is now lower than under the long tranquil period (0.5075 compared to 0.5641) we still find zero indications of contagion on a 5% significance level or higher.

In table 9, we show the results for comparing the correlation between Italian and Greek 10-year government bond yields during the crisis compared to the tranquil period lasting from the 1st_{of January 2003 until 29}th_{of December 2006. This tranquil period results}

(27)

27

in a correlation of 0.6014, which is the highest of the three specification of the tranquil period. The table shows zero indications of contagion from Greece towards Italy.

Combining the results from our three specifications of the methodology leads to the conclusion that we find no evidence of contagion from Greece towards Italy using data on 10-year sovereign bond yields.

4.3 Conclusions from the results

To summarize our results. We found very little evidence of contagion from Greece towards Spain and we found no evidence of contagion from Greece towards Italy. Comparing the results of Italy & Greece to those of Spain & Greece, we can see that the correlation during the tranquil periods is higher for the country pair Italy & Greece. This leads us to conclude that the cross-market linkages between Italy and Greece during tranquil times (interdependencies) are higher than that between Spain and Greece. This may also be a reason why we did find zero evidence of contagion from Greece to Italy and some evidence for contagion for Greece to Spain. Finding a significant increased correlation requires that the difference between correlations during the crisis and during the tranquil time is large enough to conclude a significant increase. If the correlation between the changes in yields of the countries was already high during the tranquil period, then an even stronger correlation is required, which is less likely to be found.

When comparing the three different methodologies that we used to test for contagion, we can draw some conclusions. Looking at the results of the unadjusted correlation approach we can see the importance of implementing the correction of Forbes & Rigobon (2002). If we had not implemented the correction for the increased variance in the source country of the contagion we could have concluded that the cross-market linkages had increased and that contagion had taken place in the case of country pair Greece & Spain.

The reason that we did not find evidence for contagion after correcting for the entire increase in variance in Greece (Forbes & Rigobon method) has certainly to do with the large increase in variance of the 10-year sovereign bond yields during the crisis. The large increase in variance led to a large correction of the conditional correlation coefficient, which led to a lower unconditional correlation coefficient. The test showed that none of the unconditional correlation coefficients increased significantly compared to the correlation coefficient during the different tranquil periods.

The Corsetti approach limited the correction for the increase in variance in Greece up to the level of increase in the variance of the other country in the country pair. The goal of this limit on the correction was to correct only for the increase in variance that was common between the two countries. This led to a lower correction in most cases, but still only little

(28)

28

evidence was found for contagion, only for the country pair Greece & Spain, using the short tranquil period.

We want to stress the importance of defining the tranquil period in doing research on contagion. The strength of the correlation between markets during the tranquil period had a major influence on finding evidence for contagion when using the methodology of comparing correlation. This also directly follows from our definition of contagion which is an increase in cross-market linkages, during a crisis in one of the countries. This increase can only be correctly verified when the tranquil period gives a good reflection of cross-market linkages that are present during tranquil times, also called interdependencies.

In our introduction we stated that we were specifically interested in examining whether there was contagion towards Spain and Italy during the period of the last Greek elections in January 2015 or that Greece has become ‘isolated’. With isolation we meant that there is no extra transmission from Greece towards Italy and Spain during this period. This question was sparked by the observation that, around the time of the election, the Greek 10-year government bond yields increased, when at the same time the yields for Italy and Spain decreased. By examining the entire crisis we thus also examined the period around and after the elections of January 2015 and we found no evidence of contagion during this period. This finding supports the claim that Greece has become ‘isolated’.

Our conclusion of no contagion is in agreement with the conclusion of Forbes & Rigobon (2002), who examined contagion during different crises, namely the 1997 Asian crisis, the 1994 Mexican devaluation and the 1987 US stock market crash and also found no evidence for contagion. The paper of Caporin et al. (2015) also examines contagion during the Greek sovereign debt crisis, but the authors use an approach based on quantile regression instead of comparing correlations, as is done in this thesis. Caporin et al. (2015) find that there was no contagion during the Greek sovereign debt crisis. In addition, they state that the divergence in sovereign bond yields was already expected by investors after the bankruptcy of Lehman Brothers in September 2008. Their conclusion about the existence of contagion during the Greek sovereign debt crisis is the same as the conclusion of thesis. Another paper that does not find evidence of contagion during Greek sovereign debt crisis is that of Beirne & Fratzscher (2013). These authors state that the divergence of sovereign yields in the euro area, after the revelation of the Greek debt crisis, is caused by mispricing of fundamental risk in the period before the crisis and that the increase in yields for Italy and Spain is thus not caused by contagion. The increases in yields for Greece, Italy and Spain could thus be caused by simultaneous worsening of fundamentals in these countries and the changed way of pricing the risks arising from these worsening fundamentals.

(29)

29

The paper of Andenmatten & Brill (2011), however, finds evidence of contagion from Greece to Italy and Spain during the first half of 201013_{, with a similar methodology as employed in}

this thesis. The reason that Andenmatten & Brill (2011) found evidence of contagion is probably related to their choice of the tranquil period, which lasts from the first of January 2008, until the starting point of the crisis. This tranquil period leads in our results to the lowest correlation, thus compared with the other tranquil periods this specification will find evidence of contagion more often. The chosen tranquil period of Andenmatten & Brill might not be optimal, if one applies the argumentation of Caporin et al. (2015). These authors state that after the collapse of Lehman Brothers, the cross-market linkages in the euro area decreased. Meaning that it probably does not reflect the correct correlation, between changes in sovereign bond yields, that belongs to the tranquil period. In our opinion it is better to choose a longer tranquil period, thus applying one of the other two specifications of the tranquil period, because they better reflect the correlation between markets during truly tranquil times14

13_{Note: their sample also ends after July 2010.}

14_{We are aware that in our ‘long tranquil period’ the period after the Lehman bankruptcy is also included. But}

because of the long duration of this tranquil period the effect of the Lehman bankruptcy is averaged out and the

(30)

30

5. Conclusion

In this thesis we examined the question whether there was contagion during the Greek sovereign debt crisis towards Italy and Spain. We defined the 5th_{of November 2009 as the}

start of the Greek sovereign debt crisis. The motivation for this thesis came from the observations that during and after the last elections in Greece, in January 2015, the 10-year sovereign bond yields for Greece increased, while the yields of Italy and Spain decreased. This observation combined with the apparent co-movements during the years before raised our interest in examining whether there was contagion during the period between the revelation of the Greek crisis and the period after the elections in Greece in January 2015. We defined contagion as an increase in cross-market linkages, during a crisis in one of the countries. As our methodology we choose to compare correlation coefficients, using the methodology of Forbes & Rigobon (2002). We also implemented an adjustment to this method to capture the critique of Corsetti et al. (2005).

Our results show there is almost no evidence of contagion from Greece to Spain and that there is no evidence of contagion from Greece to Italy, using data of 10-year government bond yields. The small differences in the results for the country pairs is probably caused by the fact that the cross-market linkages between Italy and Greece were already higher during the period before the crisis than the cross-market linkages between Spain and Greece. We examined the entire period of the Greek sovereign debt crisis until March 2015 and found no evidence of contagion. We also wanted to examine the recent up rise of the Greek sovereign debt crisis around the elections of January 2015. Our results also show no evidence of contagion around this period, this could support the statement that sovereign debt problems of Greece have become isolated. But because there was no evidence found for contagion, this isolation might already have occurred before the crisis started in late 2009. This statement is in agreement with the conclusion of Caporin et al. (2015). The authors state that after the collapse of Lehman brothers, investors expected a divergence of yields for the euro area countries and that this divergence led to lower cross market correlations. Another paper that does not find evidence of contagion during Greek crisis is that of Beirne & Fratzscher (2013). These authors state that the divergence of sovereign yields, after the revelation of the Greek debt crisis, is caused by mispricing of fundamental risk in the period before the crisis and that the increase in yields for Italy and Spain is thus not caused by contagion. The argument in both papers stated above could be the reason why we found no evidence of contagion during the Greek sovereign debt crisis towards Italy and Spain.

(31)

31

6. References

Andenmatten, S., & Brill, F. (2011). Measuring co-movements of CDS premia during the Greek debt crisis (No. 11-04). Discussion Papers, Department of Economics, Universität Bern.

Beirne, J., & Fratzscher, M. (2013). The pricing of sovereign risk and contagion during the European sovereign debt crisis. Journal of International Money and Finance, 34, 60 -82.

Bolton, P., & Jeanne, O. (2011). Sovereign default risk and bank fragility in financially integrated economies. IMF Economic Review, 59(2), 162-194.

Caporin, M., Pelizzon, L., Ravazzolo, F., & Rigobon, R. (2015). Measuring sovereign contagion in Europe. SAFE Working Paper No. 103.

Constancio, V. (2012). Contagion and the European debt crisis. Financial Stability Review, 16, 109-121.

Corsetti, G., Pericoli, M., & Sbracia, M. (2005). ‘Some contagion, some interdependence’: More pitfalls in tests of financial contagion. Journal of International Money and Finance, 24(8), 1177-1199.

Dornbusch, R., Park, Y. C., & Claessens, S. (2000). Contagion: understanding how it spreads. The World Bank Research Observer, 15(2), 177-197.

Dungey, M., Fry, R., González-Hermosillo, B., & Martin, V. L. (2005). Empirical modeling of contagion: a review of methodologies. Quantitative Finance, 5(1), 9-24.

Ehrmann, M., Fratzscher, M., Gürkaynak, R. S., & Swanson, E. T. (2011). Convergence and anchoring of yield curves in the euro area. The Review of Economics and Statistics, 93(1), 350-364.

Forbes, K. J., & Rigobon, R. (2002). No contagion, only interdependence: measuring stock market co-movements. The Journal of Finance, 57(5), 2223-2261

Gorea, D., & Radev, D. (2014). The euro area sovereign debt crisis: Can contagion spread from the periphery to the core?. International Review of Economics & Finance, 30, 78-100.

Gulati, M., Trebesch, C., & Zettelmeyer J. (2013). The Greek debt restructuring: an autopsy. Economic Policy, 28(75), 513-563.

Kendall, M., & Stuart, A. (1969). The advanced theory of statistics (Vol. 1). Charles Griffin and Co.

Lane, P. R. (2012). The European sovereign debt crisis. The Journal of Economic Perspectives, 26(3), 49-67.

Moser, T. (2003). What is international financial contagion? International Finance, 6(2), 157-178.

Pericoli, M., & Sbracia, M. (2003). A primer on financial contagion. Journal of Economic Surveys, 17(4), 571-608.

(32)

32

Stock, J. H., & Watson, M. W. (2011). Introduction to Econometrics. 3rd_{edition. Pearson.}

Summers, L. H. (2000). International financial crises: causes, prevention, and cures. American Economic Review, 1-16.

Was there contagion during the Greek sovereign debt crisis towards Italy and Spain?