Russian and Eastern European Dynamic Conditional Correlations in light of the Russian crisis of 2014-2015

School of Business Finance Department

Russian and Eastern European Dynamic Conditional Correlations in light of the Russian crisis of 2014-2015

Master’s thesis

Tobias Burton 0329201 Instructors:

Dr. Kashif Saleem Professor Eero Pätäri

ABSTRACT

Author: Burton, Tobias

Title: Russian and Eastern European Dynamic Conditional Correlations in light of the Russian crisis of 2014-2015

Faculty: Lappeenranta University of Technology, School of Business

Year: 2015

Major: Finance

Master’s thesis:

Examiners: Dr Kashif Saleem Professor Eero Pätäri

Keywords: Dynamic Conditional Correlations, GARCH, Russia, Central and Eastern Europe, cointegration This thesis studies the impact of the latest Russian crisis on global markets, and especially Central and Eastern Europe. The results are compared to other shocks and crises over the last twenty years to see how significant they have been. The cointegration process of Central and Eastern European financial markets is also reviewed and updated. Using three separate conditional correlation GARCH models, the latest crisis is not found to have initiated similar surges in conditional correlations to previous crises over the last two decades. Market cointegration for Central and Eastern Europe is found to have stalled somewhat after initial correlation increases post EU accession.

TIIVISTELMÄ

TEKIJÄ: Burton, Tobias

Tutkielman nimi: Russian and Eastern European Dynamic Conditional Correlations in light of the Russian crisis of 2014-2015

Tiedekunta: Lappeenrannan teknillinen yliopisto, Kauppatieteellinen tiedekunta

Pääaine: Rahoitus

Vuosi: 2015

Pro gradu -tutkielma:

Tarkastajat: KTT Kashif Saleem Professori Eero Pätäri

Avainsanat: Dynaamiset ehdolliset korrelaatiot, GARCH, Venäjä, Keski- ja Itä-Eurooppa, kointegraatio Tämä tutkielmaa selvittää Venäjän viimeisimmän talouskriisin vaikutusta kansainvälisiin rahoitusmarkkinoihin, erityisesti Keski- ja Itä-Eurooppaan.

Saatuja tuloksia vertaillaan muihin, viimeisen kahdenkymmenen vuoden aikana tapahtuneisiin finanssikriiseihin suhteellisen merkitsevyyden arvioimiseksi. Työ myös päivittää ja analysoi Keski- ja Itä-Euroopan rahoitusmarkkinoiden kointegraatio prosessin tilaa. Kolmen erillisen ehdollisen korrelaation mallin antamat tulokset ilmentävät, ettei viimeisin kriisi ole aiheuttanut samanlaisia ehdollisten korrelaatioiden nousupyrähdyksiä kuin muut kriisit viimeisen kahden vuosikymmenen aikana. Keski- ja Itä-Euroopan markkinoiden kointegroituminen on myös hiipunut parhaista kasvuajoista Euroopan Unioniin liittymisen jälkeen.

ACKNOWLEDGEMENTS

Thank you to everyone I met in Lappeenranta and afar for making my student days such a pleasure, sprinkled with some unforgettable moments. I would like to thank my instructors Kashif Saleem and Eero Pätäri who patiently spared their time to give me valuable help and advice that aided the completion of this thesis. Finally, a special ‘cheers’ to all my family and friends for providing both aid and entertainment throughout the years.

Helsinki 3.5.2015 Tobias Burton

Table of contents

1. Introduction ... 1

2. Literature review ... 6

3. Methodology ... 15

3.1. Estimating the variance equation ... 23

4. Data ... 25

4.1. The market background of Central and Eastern European states ... 26

4.2. Descriptive statistics ... 28

5. Results ... 32

5.1. Regression coefficients ... 32

5.2. Variance development ... 44

5.3. Dynamic Conditional Correlations ... 51

6. Conclusions ... 60

7. References ... 62

7.1. Online references ... 66

8. Appendices ... 67

1

1. Introduction

Central and Eastern Europe encompasses many countries that differ ethnically, culturally and economically from each other. With varying histories and climates, there is little to find in common between many of these states, except for the likelihood to be influenced by events generating in Russia.

This tendency also extends into the economic sphere, with Russia playing an extensive trade and investment role due to regionality and historical factors. Russia remains the leading economic and political force in the Central and Eastern European region, and its stock exchange enjoys the highest market capitalisation within the area (Lucey and Voronkova 2008, p. 1304).

Recently, economic sanctions imposed by the European Union and the USA over the Russian role in the ongoing Ukrainian crisis have damaged trade and investment ties between regions. A second economic factor hurting Russian economy is the large drop in global oil prices. Russia is the world’s biggest exporter of natural gas and second biggest exporter of oil (Dülger et al. 2013, p. 605). Indeed, a study by Benedictow et al. (2013) finds that Russia has become increasingly dependent on oil sales since the start of the 21^st century, with some evidence of the so called Dutch disease, whereby the export of oil has crowded out other production and export items.

All in all, between early 2014 and 16 January 2015, the Russian rouble lost well over 50 percent of its value against the dollar, while the dollar denominated RTS Index of the Moscow exchange plummeted by over 40 percent (Jevacott 2015).

The developing Ukrainian crisis, which started in the spring of 2014, and a slide in oil prices, commencing in the fall of 2014, have proven that dependency on Russia as a trading partner carries significant risks, and countries that share extensive geographical, historical and trade links are the ones most likely to be influenced by disturbances in the Russian economy.

2

Consequently, trade and financial markets are invariably linked. A significant drop in trade should show up with a corresponding slide in the stock market, since financial and trade links tend to be correlated (Gelos and Sahay 2001, p. 56).

The last financial crisis generated in Russia in 1998 lead to the value of the rouble collapsing by 70 percent and sent the value of Russian stock exchange tumbling (Rautava 2004, p. 315). The Russian crisis of 1998 had global ramifications with significant volatility increases experienced in many parts of the world (Saleem 2009, p. 244). It is feasible that the present Russian market turbulence may generate notable contagion effects abroad, which will manifest through an inexplicable increase in contemporary correlation. These effects might be all the more severe within the local area of countries with notable economic ties with Russia.

It is of interest to see whether Central and Eastern European states have experienced these regime changes during the current crisis period, and if they have, whether this is the result of a period of natural increase in cointegration between Central and Eastern European states and Russia. However, prior research into recent links between Russian and European stock markets is thin on the ground, and the author is not aware of any paper that takes this most recent turn of events into account. An update is required.

Using a dataset that includes the 2014-2015 period of turmoil in Russia, this paper uncovers what sort of significant impact the economic underperformance of Russia in 2014-2015 has had on its European neighbours and whether the increase in the volatility regime has spread. Since any volatility interlinkage will become evident through the the size of the comovement of indices, the focus of this paper is on studying these. The list of countries from Europe to be examined include; Finland, Estonia, Poland, Hungary and the Czech Republic. Each of these countries will be individually pitted against Russia to see what sort of reaction the recent Russian crisis generates in their comovement. Under examination is whether there is a unified response from the group or whether each country responds in an individual way. In addition, using a proxy Eastern European index, the group cointegration with the regional Eurozone market and global US and Asian markets will be measured for comparison. To see whether Russia is becoming more cointegrated with global

3

markets and whether, once again, correlations have risen on a global scale as a result of a Russian crisis, the Russian market will be analysed against the US, Eurozone and Asian markets.

A timespan of twenty years was chosen to give a sense of perspective and to see the long-term picture of the cointegration process. In view of Russia’s status as a former superpower and a member of the so called BRIC countries, it is important to compare the impact of the current Russian crisis to that of the 1998 rouble crisis, as this will indicate whether the role of the Russian economy in world has significantly increased or diminished compared to 15 years ago.

Measuring the long-term cointegration of Central and Eastern European states to regional and global markets as a whole is also important, since many states in Central and Eastern Europe have actively attempted to cointegrate with European and US markets. The most evident sign of this is how many states have joined the European Union –including all the Central and Eastern European countries measured in this study. The correlation span of 20 years will show how successful these attempts have been.

Finally, to check the robustness of results and to measure whether time-varying models significantly outperform more constant equivalents, three separate models will be run in this paper.

Correlations between markets are not static, but tend to drift through time. For developed markets especially, the comovement in Europe has increased markedly from decades before. Taking this into account, separate drivers of comovements in series will be looked at. First, conditional correlations that fluctuate only due to variance shocks from individual markets, while correlation coefficient is constant will be measured. Then, conditional correlations that fluctuate not only due to variance impulses, but also due to shocks to the actual correlation relationship will be used. If there are significant changes in the correlation regime, the latter should outperform the former in analysing them. It is a well held belief that economic spillovers and uncertainty increase in times of crisis. Therefore, an asymmetric model will be

4

employed to see whether negative shocks induce stronger correlation effects than positive shocks. Should financial distress cause significantly higher correlations than financial boom, this model should be better able to capture it.

Specifically, this paper take as its null hypotheses that:

1. No long-term cointegration process is evident and no conditional correlation is noticeably higher now than 20 years ago.

2. The most recent Russian index and currency collapse an subsequent loss of trade has not caused significant jumps in the respective conditional correlations between Russia and global regions or Russia and Central and Eastern European states

3. Correlation increases generated by the current Russian crisis do not significantly differ from the ripple effects of Russian crisis of 1998.

4. Any possible response to the Russian collapse in the volatility regimes of the Central and Eastern European markets has been similar for all states measured.

5. The current Russian crisis has led to no change in the respective significance of regional markets, represented by the Eurozone, or Global markets, represented by the US and Asian markets on Central and Eastern European states.

6. The three models employed will not significantly differ in accuracy of predictions offered.

Since the crisis is ongoing, it is hard to say whether any potential correlation increase is a temporary contagion spike before a return to previous levels, or a more permanent acceleration in the cointegration process, with any potential increase in conditional correlations being a manifestation of this. Thus a more permanent picture of whether the Russian crisis has had any effect; whether it has caused a sudden contagious volatility spike, or whether it has ushered in a new normal level for cointegration may be better envisioned at a later date. Currently the level of trade and investment with Russia is largely linked to Western sanctions and the price of oil, two things that are notoriously hard to predict. However, this paper should provide some very useful insight into the progress of events so far. Also, there is

5

always a level of ambiguity involved with country correlation, with a deep underlying factor having the ability to create a spurious relationship between two markets with the illusion of integration. Studying the level of cointegration is not an exact science, but more a process to indicate which countries generally tend to affect one another and whether these mutual effects are becoming more or less severe.

The structure of this paper is as follows; Section 2 gives a review of previous literature on economic cointegration with Russia or Central and Eastern European states, and also clarifies Russia’s economic structure and relationship with energy exports. A discussion on the nature of cointegration and contagion is also included.

Section 3 gives a thorough introduction to the methodology used in the paper and justifications for its use. Section 4 takes a look at the data used for this study. The country indices, their general nature and descriptive statistics are displayed and analysed. Section 5 displays the results that are acquired and draws from these the answers to the hypotheses posed in the introduction. Reasoning behind the results is also proposed. Section 6 offers the final conclusions drawn and concludes the paper.

6

2. Literature review

Though the influences of the spillovers of the Russian crisis of 1998 are covered to some extent, more recent reviews on how Russian shocks affect the global market are harder to come by. The fact that Russian linkages with European states have not been under closer scrutiny seems an oversight, since as pointed out by Lucey and Voronkova (2008), Russia remains arguably the most important market in Central and Eastern Europe.

Lucey and Voronkova (2008) themselves view stock market cointegration with Russia between 1995 and 2004 using a number of different cointegration approaches. They also employ a DCC-GARCH model in their paper to investigate whether the nature of conditional correlations between Central and Eastern European states and Russia has changed since the aftermath of the Russian crisis in 1998. The results drawn from the cointegration tests indicate that Russia is still relatively isolated, and long term cointegration has not happened since the 1998 crisis. With the DCC-GARCH model, Lucey and Voronkova (2008) find short-term increases in conditional correlations around the crisis period, but rises in the conditional correlations post crisis are not significantly larger, when compared to the precrisis levels in 1998. (Lucey and Voronkova 2008, p. 1317-1320)

Saleem (2009) studies stock market cointegration with Russia before, during and after the 1998 crisis with a GARCH-BEKK model. One of the pairs studied is shocks between Russia and Emerging Europe for which between 1998 and 2007 new information shocks are significant in both directions, but shocks from previous volatility estimates are only significant from Russia to Emerging Europe. Saleem (2009) finds that shock spillovers from Russia to Emerging Europe increased during the 1998 crisis when compared to the 1995-1998 level, but that cointegration between the two then declined to below mid-nineties levels. Spillovers also became unidirectional, after the 1998 crisis Russian spillovers affected Emerging European markets, but not vice versa. Saleem (2009) also studied spillovers between Russia and three global regions; the European Union, the United States and Asia. A similar trend to that between Russia and Emerging Europe can also be found for the other markets. The most significant shock and volatility spillovers happen during the 1998

7

crisis, while only the Russia-Asia pair sees precrisis insignificant volatility spillovers become significant post crisis. For the rest, shock and volatility spillovers return to around their respective precrisis levels –though interestingly some flip in sign– for the post crisis regime. (Saleem 2009, p.251-253)

Similarly to Saleem (2009), Dungey et al. (2007) find significant spillovers of contagion emanating from the Russian equity market during the crisis of 1998.

These shocks are primarily sourced from the Russian bond markets, with the equity market acting as a conductor. Dungey et al. (2006) argue that most of the Russian crisis effects on equity markets were channelled through the US equity markets, thus backing up the claim of Kaminsky and Reinhart (2003) that contagion that spreads beyond a regional sphere spreads to the periphery through global financial hubs – such as the US market (Kaminsky and Reinhart 2003, p. 20-21). Possibly as a result of this, when using a multi-regime factor model on daily returns for the year 1998, Dungey et al. (2007) intriguingly find contagion from Russia to be more significant in developed markets than developing economies. The authors find that individually, while European Union countries Germany and the UK had the severest reaction to the 1998 crisis, the effects of Russian contagion on Poland are mild. (Dungey et al.

2007, p. 156-157,169, 171-172)

Gelos and Sahay (2001) look at stock market and foreign exchange spillovers between Russia, the Czech Republic, Hungary, Poland, Lithuania and Estonia amongst others. They look for periods of pressure for exchange rates to readjust, and then analyse the period’s stock and exchange market data closely to see the size of regional comovements and whether they fluctuate between crisis and calm periods. They also look to see if volatility in some country systematically leads that of another and whether Central and Eastern European states behave similarly to other countries within the region. Like other studies, Gelos and Sahay (2001) find the Russian market leading the Czech, Hungarian and Polish market in the stock market during the Russian crisis of 1998, but not in tranquil periods. They also found the responses to be asymmetric (Gelos and Sahay 2001, p.72-73).

On the other hand Fedorova and Saleem (2010) find significant unidirectional shock and volatility spillovers from Russian stocks into Hungarian, Polish and Czech markets at five percent significance. Using a pairwise GARCH-BEKK model, they

8

analyse weekly stock market data between 1995 and 2008. (Fedorova and Saleem 2010, p. 528-529)

A research by Caporale and Spagnolo (2011) uses a VAR-GARCH-in-mean model to analyse the interdependence between Central and Eastern European countries:

the Czech Republic, Hungary, Poland, and larger economies: Russia and the UK.

Their sample is between 1996 and 2008, and incorporates the period when all three Central and Eastern European states joined the European Union. Caporale and Spagnolo (2011) find that conditional correlations from Russia to the Czech Republic, Hungary and Poland become more significant after European Union membership is gained. Correlations with the UK follow a similar trend to those with Russia, with the Central and Eastern European states being more highly correlated with the British market once European Union accession is achieved. (Caporale and Spagnolo 2011, p. 162-167)

Syllignakis and Kouretas (2011) look at weekly data between 1997 and 2009 to analyse the correlations between Russia, Germany and the US and a range of Eastern European states with a DCC-GARCH model. With a group including Estonia, the Czech Republic, Poland and Hungary amongst other states, they argue that geographic proximity is not a lead cause for interdependence, as all Eastern European states exhibit stronger cointegration with Germany and the US than with Russia, despite dynamic correlations with Russia rising significantly for all Eastern European states except the Czech Republic. The authors attribute a large part of the rising correlations with Germany and the US to Eastern Europe’s ascension into the EU. (Syllignakis and Kouretas 2011, p.723-726, 731)

Syllignakis and Kouretas (2011) also probe how three separate global crises affected the correlation relationships. Firstly, they found that for the Russian and Asian crises of 1997 and 1998, Russian correlations with the Czech Republic and Estonia rose significantly, while correlations with Poland and Hungary fell significantly.

Correlations with the US fell significantly during the crisis for Estonia, Poland, and Hungary while there was no significant reaction from the Czech Republic. Against Germany, correlations fell for the Czech Republic, Estonia and Hungary, but the correlation change for Poland was insignificant. The dot-com bubble between 2000 and 2002 also yielded mixed reactions, with US correlation rising significantly for

9

Estonia and falling significantly for all other markets. Likewise, Estonia was the only country whose correlation significantly rose with Germany. The Czech Republic’s and Poland’s correlation with Germany fell, while Hungary’s correlation remained the same. Against Russia, correlations fell for the Czech Republic, Hungary and Poland, but Estonia’s correlation did not move. Finally, the current financial crisis has seen a remarkably unequivocal response, with correlations of all the Eastern European states rising significantly for each of Germany, Russia and the US. (Syllignakis and Kouretas 2011, p.727)

The former communist Central and Eastern European states’ cointegration process with Europe has received more attention. Many studies have taken advantage of the natural break point in regimes that is offered by the date of accession to the European Union for many Central and Eastern European states.

Schotman and Zalewska (2006) study whether cointegration estimates for Poland, Hungary and the Czech Republic can be improved by accounting for non- synchronous trading. Studying the spillovers between these three markets and those of Germany, the UK and the US between 1994 and 2004, Schotman and Zalewska (2006) find that adjusting for non-synchronous trading removes reductions to correlations caused by geographical factors and time differences. The authors use a time-varying parameter regression model. Overall cointegration values significantly increased for Central and Eastern European states, especially around the East Asian and Russian crises of 1997 and 1998. In their study, Hungary proves the most susceptible to global and regional shock, while the Czech Republic is the least susceptible. Schotman and Zalewska note that this is in direct relation to the amount of foreign holdings for each market, with the Hungarian market having the most foreign market participants and the Czech Republic containing the least foreign participants for the period studied. On correlations with the Russian crisis of 1998, the authors find that, following a clear spike at the time of the crisis, correlations decline, though they remain clearly higher than their precrisis levels. This indicates that the Russian crisis of 1998 would have partially accelerated the cointegration process of Central and Eastern European markets and developed economies.

(Schotman and Zalewska 2006, p. 463-464, 770, 491-492)

10

Li and Majerowska (2008) review shock and volatility spillovers between Poland, Hungary, Germany -representing regional shocks- and the US –a proxy for global shocks. Through a GARCH BEKK model, they find that global shock and volatility spillovers from the US were significant in both Hungary and Poland, while regional volatility spillovers from Germany were also significant in both Poland and Hungary.

There are no significant spillovers going in the other direction. (Li and Majerowska 2008 p. 255-256)

Savva and Aslanidis (2010) look at Czech, Hungarian and Polish cointegration with the Eurozone before and after accession into the European Union. They use a STCC-GARCH model, an extension to the CCC-GARCH model discussed in section 3 of this paper, with a VAR mean process to study the returns of the markets between 1997 and 2008 (Savva and Aslanidis 2010, p. 340-341). Results are mixed, with the authors finding that while the Czech Republic and Poland have become more cointegrated with the Eurozone, Hungary has not. The authors conclude that the lack of a significant increase in correlations on the part of Hungary may be down to Hungary already having high level of correlation prior to joining the EU (Savva and Aslanidis 2010, p. 350).

Cappiello et al. (2006a) follow the cointegration process of a group of countries that joined the European Union in 2004, including Estonia, Poland, Hungary and the Czech Republic. Using a comovement box methodology developed by Cappiello et al. (2005), they find that the equity markets of Estonia, Poland and the Czech Republic have all increased comovements with the Eurozone. Hungary’s increase in comovement is not statistically significant, but like Savva and Aslanidis (2010), Cappiello et al. (2006a) attribute this to a higher base level of comovement between Hungary and the Eurozone to begin with. (Cappiello et al. 2006a p. 12-13, 19, 38- 39)

Beckmann et al. (2011) find that changes in incomes and share prices from the Eurozone and the US clearly affect the economic and consumer sentiments of Hungary, Poland and the Czech Republic. They argue this to be a clear sign of significant international linkages between Central and Eastern European countries and global markets. The results for Beckmann et al.’s (2011) paper are acquired using the Johansen VAR cointegration model on data between 1997 and 2008, with

11

the effects of international income and share prices increasing after Hungary’s, Poland’s and the Czech Republic’s EU accession in 2004. (Beckmann et al. 2011, p.

147-152, 154-155)

Wang and Moore (2008) encounter similar results for their DCC modelling of Hungary’s, Poland’s and the Czech Republic’s correlation with a value weighted contruction of 12 main EMU indices. Between 1994 and 2006, they find that Correlations with the EMU increase for all three Central European countries, with hikes especially prominent after crises. This they attribute in part to European Union membership and to the development of the financial sector of Poland Hungary and the Czech Republic. (Wang and Moore 2008 p.124-128.131)

Focusing on Estonia, the cointegration between the Baltic states and the West is researched by Maneschiöld (2006). Johansen’s VAR cointegration tests and Granger causality tests were applied to study the long-term cointegration trend for the short term cause and effect relationship between 1996 and 2005. Maneschiöld (2006) shows that Estonia’s long-term cointegration with Europe and the US is low, but that unidirectional effects are caused on Estonia by the US, Germany, France and the UK, but not Japan (Maneschiöld 2006, p. 37-39, 41).

Results also remain mixed amongst articles that study the cointegration process of Central and Eastern European states after the global financial crisis erupted.

The pairwise correlations of Estonia, Latvia, Lithuania, the Czech Republic, Poland, Hungary, Romania and Bulgaria with the Euro Stoxx 50 index between 2002 and 2012 were studied by Harkmann (2014). Using DCC-GARCH, She found that correlations between the Eastern European states and the Euro Stoxx index temporarily rose a statistically significant amount after the bankruptcy of Lehman Brothers and the start of the Greek bailout. Since the groups amount of overall integration did not significantly increase between 2002 and 2012, Harkmann took these temporary spikes as evidence of contagion. Individually, Poland, Hungary and the Czech Republic were found to have the highest correlations. The Baltic states, Romania and Bulgaria show noticeably less correlation with the Euro Stoxx 50 index, though Romania and Bulgaria were the states to experience the sharpest rise in correlations once the crisis erupted. (Harkmann 2014 p. 63-64)

12

On the other hand, adding an asymmetric term to their DCC model, Gjika and Horvath (2013) study Poland, Hungary and the Czech Republic to find out whether their mutual interdependence and cointergration with Eurozone –represented by Euro Stoxx 50- increases between 2001 and 2011. Using pairwise correlations, they too find that correlations with the eurozone increase significantly after the financial crisis erupts in 2008. However, unlike Harkmann’s (2014) results, the crisis seems to spark deeper interdependence between the three countries themselves. Gjika and Horvath (2013) find significant asymmetry only in the conditional correlation between Hungary and Poland.

How contagion itself is defined is not quite cut and dry. Adler and Dumas (1983) began the discussion of contagion definition when they noted that the yardstick for an optimum risk return ratio –and thus the optimum portfolio composition- varies from country to country. Domestic factors, interest rates and exchange rates lead to investors being able to reduce risk through country diversification. This segmentation is at times punctured by global events that lead the returns of one market to greatly depend on the performance of another. In another ground-breaking article on contagion King and Wadhwani (1990) conclude that contagion is irreversibly linked to sharp increases in volatility in the initial country, which then causes an increase in correlations across borders.

Some define contagion as any sort of spillover effect irrespective of size, while others hold that only certain spillovers transmitted through certain mechanisms should be viewed as contagion (Forbes and Rigobon 2002, p. 2223). Forbes and Rigobon (2002) argue that contagion should constitute a significant increase in correlations as well as comovement in crisis, respective to the precrisis level (Forbes and Rigobon 2002, p. 2224). They view large crisis comovements between countries with high level of correlation only as interdependence -comovements that can be reasonably expected given the correlation regime- if the level of correlation does not simultaneously rise significantly. Using this definition, Forbes and Rigobon found that none of the three crises studied in their paper - the 1987 US stock market crash, the 1994 Mexican peso fall or the 1997 East Asian crisis - exhibited any signs of contagion, only interdependence (Forbes and Rigobon 2002, p. 2250).

13

Other definitions for contagion exist. Masson (1998) sees contagion as the spillovers from one country to another that can not be explained by changes in macroeconomic fundamentals of the country receiving the spillovers. The most obvious example of this may be investor herding behaviour, whereby a crisis in one economy may spark a flight to safety reaction from investors in another market perceived risky, even if the fundamentals of that market remain unchanged (Masson 1998, p. 4-5). Bekaert et al.

(2005) offer a similar definition, with contagion being classed as the excess correlation not supported by economic fundamentals in times of crisis. With such a definition, Bekaert et al. (2005) find contagion present in the residuals when employing a factor model to study East Asian crisis. Like Forbes and Rigobon (2002), Bekaert et al. (2005) do not locate any contagion for the Mexican peso crisis.

As well as varying definitions, contagion has also been categorised by types.

Bekaert et al. (2011) follow the definition of Bekaert et al. (2005) to break the types of contagion down into US led, global and domestic contagion. Bekaert et al. (2011) speculate that these varying types of contagion, would emanate through different channels. Contagion that spreads due to a globalisation factor would most affect countries with the most integrated trade and financial links. The contagion that Masson (1998) refers to, whereby a crisis in a particular country or market segment leads to investors reassessing the perceived riskiness of other segments or countries, is another channel for contagion. Indiscriminate investor herding and flight to safety is also a channel for contagion spreading. Bekaert et al. (2011) found domestic contagion, where investors are more receptive to shocks generated within a country which spread throughout the financial system, is the most evident type of contagion for the global crisis of 2008-2009. (Bekaert et al. 2011, p. 5-6, 8-9, 24-25) Many other definitions of contagion exist that are beyond the scope of this paper.

Dungey et al. (2005) offer a more extensive review of the definitions and forms of contagion.

The definition for contagion used in this paper is both simple and easy to measure visually. Allowing for all types of spillover spread and a change in economic fundamentals due to a decline in trade, we see whether correlations with Russia increase abruptly. As pointed out by Gelos and Sahay (2001), large increases in correlation with the Russian market in times of crisis would imply structural breaks in

14

the correlation regime, which are hard to explain as anything other than contagion (Gelos and Sahay 2001, p. 57).

15

3. Methodology

For the mean equation, a one lag bivariate Vector Autoregressive (VAR) system –as generalised by Sims (1980) – will be introduced to see how foreign markets respond to the shocks of the previous day. Although it is not common practice generally, in this paper the spillover coefficient will be reported. Since the mean equation is not the main focus of this study, deeper delving into impulse responses and variance decomposition will be left to other work. Unlike in the variance equation the instantaneous joint effects of shocks are not considered in the mean equation, leaving all explanatory variables endogenous. The benefit of this is that the order of variables need not be considered in the VAR system, nor arbitrary constraints imposed. The bivariate VAR system used is of the form

{𝑥_𝑡 = 𝑥_𝑡−1+ 𝑦_𝑡−1

𝑦_𝑡 = 𝑦_𝑡−1+ 𝑥_𝑡−1 (1)

(Christiano et al. 1999, p. 2-3, 8-12, Brooks 2008, p. 290-292)

Measuring how cointegrated markets are with one another can be seen through how markets react in tandem to shocks and changes to volatility. This relies on the assumption that the size of residuals is not constant but varies, which empirically tends to be very common for economic and financial data (Brooks, 2008 p.386, Ibragimov and Muller, 2010 p.453).

One consequence of the rejection of constant variance is that the Gauss-Markov requirements for ordinary least squares (OLS) to be the most efficient estimator fail (Brooks, 2008 p.135-136, 149-150). The presence of autocorrelation and heteroskedasticity leads to the requirement of considering models with a non-linear variance structure, which better tackle unstable volatility. Recognising that the dispersion of residuals is not constant for all observations, the autoregressive conditionally heteroskedastic (ARCH) model proposed by Engle in 1982 forms a volatility estimate for moment 𝑡 from two parts.

Assuming that the residuals 𝑢_𝑡 of a given mean equation are normally distributed around a zero mean with an ARCH structured variance ℎ_𝑡

𝑢_𝑡 ~𝑁(0, ℎ_𝑡) (2)

16

The ARCH model for the volatility of residuals takes the form

ℎ_𝑡 = 𝜔𝜎̅ + ∑^𝑞_𝑖=1𝛼_𝑖𝑢_𝑡−𝑖² (3) The first part of the model is the traditional constant variance estimate 𝜎̅, the mean of the long term volatility that is unconditional to all other observations. The second part of the ARCH structure contains past observations of the variance of residuals 𝑢_𝑡−𝑖² . The amount of weight given to each part of the equation is denoted by the 𝛼′𝑠. The model demonstrated above is a so called ARCH (𝑞) model, since the amount of past observations considered are allowed to go right up to lag 𝑞. The model is mean reverting in that the amount of weight provided for a past variance observation generally declines as the lag increases, while increasing the amount of past observations included will make the ℎ_𝑡 estimate approach 𝜎̅. Considering an ARCH (1) model where only the latest volatility observation is included

ℎ_𝑡 = 𝜔𝜎̅ + 𝛼𝑢_𝑡−1² (4)

It is obvious that the second part of the model makes ℎ_𝑡 non-constant and conditional to the most recent residual variance.

ℎ_𝑡 = 𝑣𝑎𝑟(𝑢_𝑡|𝑢_𝑡−1) (5) This allows the ARCH model to be used to investigate autocorrelation within the series. Knowing the extent to which recent shocks affect the current estimate helps forecasting how future shifts in volatility will behave. (Brooks, 2008 p. 386-388) The ARCH model does have its limitations though. Returning to an ARCH (𝑞) model

ℎ_𝑡 = 𝜔𝜎̅ + 𝛼₁𝑢_𝑡−1² + 𝛼₂𝑢_𝑡−2² + 𝛼₃𝑢_𝑡−3² + ⋯ + 𝛼_𝑞𝑢_𝑡−𝑞² (6) there are 𝑞 parameters to model for the conditional variance. Since negative variance cannot be justified by theoretical literature, it must be assumed to be positive. However, empirical observations have noted that as the number of parameters in the model is increased, the risk of violating this non-negativity constraint increases (See Brooks, 2008 p. 389, 391-392).

17

A more parsimonious version of the ARCH model, the generalised autoregressive conditionally heteroskedastic model (GARCH) was presented by Bollerslev (1986). A GARCH (𝑞, 𝑝) model is given by

ℎ_𝑡 = 𝜔𝜎̅ + ∑^𝑞_𝑖=1𝛼_𝑖𝑢_𝑡−𝑖² + ∑^𝑝_𝑗=1𝛽_𝑗ℎ_𝑡−𝑗 (7) In the GARCH model, in addition to the constant variance and the ARCH effect, the past estimate of the variance ℎ_𝑡−𝑗 –the GARCH effect– is included and assigned the weight 𝛽_𝑗. The reason the past estimate of ℎ_𝑡 is included is intuitive. In a GARCH (1, 1) model given below

ℎ_𝑡 = 𝜔𝜎̅ + 𝛼𝑢_𝑡−1² + 𝛽ℎ_𝑡−1 (8) the past estimate of volatility ℎ_𝑡−1 can be assumed to be made up from

ℎ_𝑡−1= 𝜔𝜎̅ + 𝛼𝑢_𝑡−2² + 𝛽ℎ_𝑡−2 (9) The model gives information about the previous period volatility shock and the previous volatility estimate. The past GARCH estimate from the above, is in turn

ℎ_𝑡−2= 𝜔𝜎̅ + 𝛼𝑢_𝑡−3² + 𝛽ℎ_𝑡−3 (10) Reverse installing the past estimates into the GARCH (1, 1) model gives

ℎ_𝑡 = 𝜔𝜎̅ + 𝛼𝑢_𝑡−1² + 𝛽(𝛼₀𝜎̅ + 𝛼𝑢_𝑡−2² + 𝛽{𝜔𝜎̅ + 𝛼𝑢_𝑡−3² + 𝛽ℎ_𝑡−3}) (11) The above can be rearranged as

ℎ_𝑡 = 𝜔𝜎̅ + 𝛽𝜔𝜎̅ + 𝛽²𝜔𝜎̅ + 𝛼𝑢_𝑡−1² + 𝛽𝛼𝑢_𝑡−2² + 𝛽²𝛼𝑢_𝑡−3² + 𝛽³𝛼₁ℎ_𝑡−3 (12) As the lags of the GARCH process are increased towards infinity all previous volatility shocks become evident, though with declining significance at a ratio of weight 𝛽. Thus, a GARCH (1, 1) model contains information of more distant volatility shocks through the GARCH estimate and has only three parameters, making non- negativity errors much more unlikely. The constant long-term variance estimate must be positive and the weights of the three parameters in a GARCH (1, 1) should sum up to one in a way that reflects this. To guarantee that the weight on the unconditional volatility𝜔 is larger than zero, the weights of the ARCH and GARCH effects must be less than one

18

𝛼 + 𝛽 < 1 (13)

leaving a positive difference for 𝜔

𝜔 = 1 − 𝛼 − 𝛽 (14)

This stationarity in variance means that when the GARCH (1, 1) model is used for forecasting, it will revert towards the constant mean variance at a rate of 𝜔 as opposing to tending away from the mean at a rate of −𝜔. (Brooks, 2008 p. 392-394, Hull, 2000 p.372-373, 379-380)

Extending the GARCH (1, 1) model to allow for asymmetric responses to positive and negative shocks has been suggested by Glosten et al. (1993). They include an additional term to the model that only captures negative shocks.

ℎ_𝑡 = 𝜔𝜎̅ + 𝛼𝑢_𝑡−1² + 𝛽ℎ_𝑡−1+ 𝛾𝑢_𝑡−1² 𝐼_𝑡−1 (15) where 𝐼_𝑡−1 acquires the value 1 if 𝑢_𝑡−1² is negative or 0 if 𝑢_𝑡−1² is positive. If 𝛾 were positive and significant, it would mean asymmetry in the responses to shocks.

Volatility would rise more after a negative shock than after a positive shock. (Glosten et al. 1993, 1782-1788)

To examine market linkages between the stock exchanges of separate countries, multivariate GARCH models have been introduced. These models seek beyond simply analysing fluctuations in a specific variable’s variance and go on to analysing how adding a correlation or covariance factor for different markets helps explain mutual cointegration. Two families of Multivariate models are introduced below, with the direct extensions of GARCH for multiple variables discussed first.

Variance- covariance matrices can be produced by multiplying vectors of errors with their transposed vectors, so that for moment 𝑡, the bivariate variance- covariance matrix ×_𝑡 is given by

×_𝑡= ([𝑢_1𝑡

𝑢_2𝑡] [𝑢1𝑡 𝑢_2𝑡]) (16)

×_𝑡= ( 𝑢_1𝑡² 𝑢_1𝑡𝑢_2𝑡

𝑢_2𝑡𝑢_1𝑡 𝑢_2𝑡² ) (17)

19

A one-lag multivariate GARCH model with a vector half operator (M-GARCH-VECH), specified by Bollerslev, Engle and Wooldridge (1988), can be used to see how efficiently shocks flush through the system from one stock index to the next. It takes either the upper or lower half of the conditional variance- covariance matrix from the variance diagonal. A full M-GARCH VECH is given as

𝑉𝐸𝐶𝐻(𝐻_𝑡) = 𝑊 + 𝐴𝑉𝐸𝐶𝐻(𝑈_𝑡−1𝑈_𝑡−1^′ ) + 𝐵𝑉𝐸𝐶𝐻(𝐻_𝑡−1) (18) In the model, the half vector for the current conditional variance-covariance matrix, 𝑉𝐸𝐶𝐻(𝐻_𝑡), equates to; a vector of constant variances, 𝑊, a matrix of the most recent error half vectors, 𝐴𝑉𝐸𝐶𝐻(𝑈_𝑡−1𝑈_𝑡−1^′ ), and a matrix of the most recent estimate of the top or bottom half of the variance-covariance matrix, 𝐵𝑉𝐸𝐶𝐻(𝐻_𝑡−1).

Although the M-GARCH-VECH is appealing on a theoretical level, there are a few problems that hamper its implementation on a practical level. First, even in a bivariate model, 21 parameters need to be estimated, and the inclusion of additional variables makes the amount of parameters grow exponentially. Second, the variance-covariance matrices produced by M-GARCH-VECH models do not guarantee internal consistency, known as being positive semi-definite. (Brooks, 2008 p. 432-435, Hull, 2000 p. 384)

The M-GARCH-BEKK model (Kroner and Engle 1995) differs from GARCH-VECH presentations in that it is a quadratic variation of a multivariate GARCH model. The composition of the GARCH-BEKK model, given below, guarantees that the variance- covariance matrix, 𝐻_𝑡, will be positive semi-definite (Brooks, 2008 p. 435).

𝐻_𝑡 = 𝑊𝑊^′+ 𝐴^′𝑈_𝑡−1𝑈_𝑡−1^′ 𝐴 + 𝐵^′𝐻_𝑡−1𝐵 (19) Above, 𝑊 expresses the upper triangular matrix of constant parameters, 𝐴 denotes the matrix of weight estimates given to the variance-covariance matrix of recent shocks and 𝐵 is the matrix of weights given to the matrix of past estimates of 𝐻_𝑡. While the GARCH-BEKK model, requires slightly less parameters than the VECH model to be estimated, it is still very complex and unwieldy. (Brooks, 2008 p. 435).

Instead, more indirect models based on nonlinear combinations of multiple univariate GARCH estimations -which require fewer parameters- will be used to market cointegration between index pairs. Conditional correlation models are very suitable

20

for analysing market cointegration in crisis, since they can show how correlations react to common changes in volatility, as opposed to idiosyncratic lags. Though conditional correlation models make no assumption about the direction of volatility spillovers, it is fair to assume that most recent significant drivers of cointergration in pairs including Russia, emanates from the Russian index, due to the current crisis.

Another thing to note is that for cointegration, conditional correlation models model conditional correlation directly. VECH and BEKK models produce conditional covariance estimates, meaning that correlations need to be acquired indirectly.

Bollerslev (1990) proposed a multivariate GARCH with constant conditional correlation, known as CCC-GARCH. This model requires far fewer parameters to be estimated. As Bollerslev notes, generally the correlation between the 𝑖th and 𝑗th term, 𝜌_𝑖𝑗𝑡, in a conditional variance-covariance matrix 𝐻_𝑡 at time 𝑡 would be given by the covariance term over the product of the standard deviations

𝜌_𝑖𝑗𝑡 = ^{ℎ𝑖𝑗𝑡}

(ℎ_𝑖𝑖𝑡ℎ_𝑗𝑗𝑡)^0.5 (20)

Rearranging the above gives the conditional covariances as

ℎ_𝑖𝑗𝑡 = 𝜌_𝑖𝑗𝑡(ℎ_𝑖𝑖𝑡ℎ_𝑗𝑗𝑡)^0.5 (21) In some cases it may be appropriate to consider that the covariance term may fluctuate only when the conditional volatility of each variable changes individually, leaving the conditional correlation relationship to be constant for all moments, as presented below

ℎ_𝑖𝑗𝑡 = 𝜌_𝑖𝑗(ℎ_𝑖𝑖𝑡ℎ_𝑗𝑗𝑡)^0.5 (22) The more general matrix form of this CCC-GARCH model is

𝐻_𝑡 = 𝐷_𝑡𝑃𝐷_𝑡 (23)

where 𝐷_𝑡 represents a diagonal matrix of conditional standard deviation estimates at moment 𝑡, while 𝑃 indicates a matrix of mean correlation coefficients that do not change at any point of time. The individual standard deviation estimates for each variable can be acquired through separate univariate GARCH processes, though estimates for the unconditional covariance is more difficult to compute. A CCC-

21

GARCH process with (1, 1) lags contains N(N+5)/2 parameters to estimate. Like with the GARCH-BEKK formulation, positive-semidefiniteness is easy to impose.

(Bollerslev, 1990 p. 499, Bauwens et al., 2006 p. 88-89)

Despite being easy to compute, the CCC-GARCH model assumption of constant correlations will only be realistic in a limited number of cases. Changes in Russian and Eastern European cointegration may be better served by a less static assumption that, like standard deviations, correlations change over time. Time- dependent correlations make formula 23

𝐻_𝑡 = 𝐷_𝑡𝑃_𝑡𝐷_𝑡 (24) One proposition on how to make 𝑃_𝑡 dynamic comes from Engle (2002). Similar to (20) but instead of ℎ’s, let 𝑞’s denote integrated elements through which correlations are calculated.

𝜌_𝑖𝑗𝑡 = ^{𝑞𝑖𝑗𝑡}

(𝑞_𝑖𝑖𝑡𝑞_𝑗𝑗𝑡)^0.5 (25)

The suggested scalar dynamic conditional correlations model (DCC-GARCH) has 𝑃_𝑡 = 𝑑𝑖𝑎𝑔𝑄^−0.5𝑄𝑑𝑖𝑎𝑔𝑄^−0.5 (26) Where 𝑄 is a full matrix of 𝑞’s including off-diagonal cross elements and 𝑑𝑖𝑎𝑔𝑄 is a matrix that includes only the diagonal elements of 𝑞. The values of matrix Q and all 𝑞’s in it at moment 𝑡 is estimated from

𝑄_𝑡= (1 − 𝛼 − 𝛽)Σ̅ + 𝛼𝑢_𝑡−1𝑢_𝑡−1^′ + 𝛽𝑄_𝑡−1 (27) where Σ̅ is a unconditional mean variance-covariance matrix of correlations given by 𝐸(𝑢_𝑡𝑢_𝑡^′). 𝛼 and 𝛽 are scalars for the ARCH and GARCH effect respectively, and the residuals in the ARCH effect are standardised. An individual correlation estimate in a DCC-GARCH (1, 1) case is thus obtained by

𝜌_𝑖𝑗𝑡 = ^{(1−𝛼−𝛽)σ̅̅̅̅+𝛼𝑢𝑖𝑗 𝑖𝑡−1𝑢𝑗𝑡−1}

′ +𝛽𝑞_{𝑖𝑗𝑡−1}

([(1−𝛼−𝛽)σ̅̅̅̅+𝛼𝑢𝑖𝑖 _𝑖𝑡−1² +𝛽𝑞_{𝑖𝑖𝑡−1}][(1−𝛼−𝛽)σ̅̅̅̅̅+𝛼𝑢𝑗𝑗 _𝑗𝑡−1² +𝛽𝑞_{𝑗𝑗𝑡−1}])^0.5 (28) (Bauwens et al., 2006 p. 90, Engle, 2002 p. 341)

22

ARCH and GARCH terms for the dynamic correlations behave similarly to those in a univariate GARCH model. If they are insignificant changes in correlations only depend on changes to variables’ variance, as in a CCC-GARCH model.

Since the period under supervision contains numerous crises of varying severity, it would be prudent to see whether negative shocks increase conditional correlations more than positive shocks. To incorporate asymmetric responses in shocks to dynamic correlations, Cappiello et al. (2006b) adjust the DCC-GARCH (1, 1) model by adding an asymmetric term similar to that of Glosten et al. (1993). For an asymmetric scalar DCC-GARCH (ADCC-GARCH), 𝑄_𝑡 is given as

𝑄_𝑡= (Σ̅ − 𝛼²Σ̅ − 𝛽²Σ̅ − 𝛾²𝑁̅)Σ̅ + 𝛼²𝑢_𝑡−1𝑢_𝑡−1^′ + 𝛽²𝑄_𝑡−1+ 𝛾²𝑛_𝑡−1𝑛_𝑡−1^′ (29) where 𝑛 denotes shocks that acquire their true value if negative and 0 if positive. 𝛾 is the asymmetric coefficient. It should be noted that now requirements for a stable positive semi-definite model is that

𝛼²+ 𝛽²+ 𝜏𝛾²< 1 (30) with 𝜏 having a maximum eigenvalue of (Σ̅^−0.5𝑁̅Σ̅^−0.5)

This is the model considered to be suitable to study the pairwise conditional correlations to see how Russia and Central and Eastern European states are cointegrated with each other and the rest of the world.

To see if a dynamic conditional correlation model can outperform a constant conditional correlation model, and whether the shocks are asymmetric, pairwise VAR-CCC-GARCH, VAR-DCC-GARCH and VAR-ADCC-GARCH models will be run.

For the regressions all lags will be set as 1.

Laurent et al. (2012) point out that an Asymmetric DCC-GARCH model is most valid specification for volatility forecasting. For a superior predicting ability (SPA) test; the null hypothesis that Engle’s DCC-GARCH with leverage effects cannot be outperformed in a large group of 125 models -including the GARCH-BEKK family and other DCC variants- could not be rejected. (Laurent, et al., 2012 p. 952)

23

3.1. Estimating the variance equation

Since conditional correlation models can no longer be estimated using ordinary least squares, maximum likelihood estimators need to be employed to yield consistent estimates. As opposed to minimising the sum of squares, maximum likelihood involves finding the global maximum for the probability that coefficients acquire that value. In general, the estimate for parameter vector 𝜃 can be retrieved by maximising the log likelihood function. For DCC-GARCH, Engle (2002, p. 342) gives the maximum likelihood function as

𝐿_𝑇(𝜃) = −¹₂∑^𝑇_𝑡=1(𝑛𝑙𝑜𝑔(2𝜋 + 𝑙𝑜𝑔|𝐻_𝑡| + 𝑢_𝑡′𝐻_𝑡⁻¹𝑢_𝑡) (31) The maximum likelihood function assumes normally distributed errors, while in reality, the errors are kurtotic with fat tails. However, Bollerslev and Wooldridge (1992) showed that even if the assumption of normality is violated, a quasi-maximum likelihood estimator is still asymptotically consistent for jointly parameterized conditional mean and conditional variance. Bollerslev and Wooldridge (1992) give their quasi-maximum likelihood function as

𝐿_𝑇(𝜃) = −¹₂∑^𝑇_𝑡=1𝑙𝑜𝑔|𝐻_𝑡|−¹

2∑^𝑇_𝑡=1𝑢_𝑡′𝐻_𝑡⁻¹𝑢_𝑡 (32) The quasi-maximum likelihood function will be maximised here. (Bollerslev and Wooldridge 1992, p. 148, Bauwens et al. p. 96-97)

Engle and Sheppard (2001) claim that one of the advantages of the DCC-GARCH model is that the estimation process can be done in two steps, thus allowing for more complex models and more variables (See, for instance, Bauwens et al. p. 98- 99). Conversely, Caporin and McAleer (2012) criticise the two-step approach and argue that it does not yield consistent results for DCC-GARCH (Caporin and McAleer, 2012 p. 747). The use of small pairwise models allows this paper to employ single-step estimation for the maximisation of the quasi-maximum likelihood function.

Once the models have been estimated, a test will be employed to see how well the GARCH specification of conditional volatility has accounted for autocorrelation. The amount of serial autocorrelation still remaining in the series will be jointly checked using a test formulated by Ljung and Box (1978). The Ljung-Box test is

24

𝑄^∗ = 𝑇(𝑇 + 2) ∑ ^𝜏̂𝑘2

𝑇−𝑘

𝑚𝑘=1 ~𝜒_𝑚² (33)

where 𝑇 is the number of observations, 𝜏 is the autocorrelation at lag 𝑘 and 𝑚 is the total degrees of freedom, equal to the number of lags used. The test checks the significance through a 𝜒² –distribution for 𝑚 lags. (Brooks 2008, p. 210)

In this paper, the optimisation will be carried out using the Broyden, Fletcher, Goldfarb and Shanno (BFGS) algorithm.

25

4. Data

Country and regional indices were provided by MSCI and were retrieved from Thomson-Reuters DataStream. The composition of the indices has been constructed by MSCI to generally mirror stock market performance of the respective country of region. Indices produced by MSCI have the advantage that they are easier to compare with one another than indices created by individual national stock exchanges. MSCI indices are generally considered to be reliable.

All Central and Eastern European countries’ series are euro denominated, while Eurozone, Eastern European, Asian and US indices were US dollar denominated.

For Russia a euro denominated series and a US dollar denominated series were retrieved. To ensure consistency, the Russian euro denominated series was used in regressions with the euro denominated Central and Eastern European countries, and the Russian USD denominated series were regressed against US dollar denominated series from other parts of the world. The differences caused by the separate currencies were negligible.

The data of most series is daily and spans 20 years, from January 23th 1995 to January 23rd 2015. This amounts to a total of 5220 observations for each series.

The exception is Estonia, for which a shorter time period was available. Estonia’s data starts on June 3rd 2002 and contains 3301 observations. To ensure that all series are stationary, the returns are taken as continuously compounded first differences by dividing the natural logarithms of consecutive observations

𝑅_𝑡 = log ( ^𝑃𝑡

𝑃𝑡−1) (34)

Differencing the series leads to the first observation to be lost for all markets.

Augmented Dickey-Fuller tests are run for all the differenced series and the results are shown in table 1. The null hypothesis for the augmented Dickey-Fuller test is that a unit root exists for the series, meaning shocks do not decompose and the series is not stationary. The t-statistics in table 1 clearly show that none of the series used have a unit root, instead all are stationary.

26

Table 1. Augmented Dickey-Fuller Probabilities

Russia Eastern Europe EMU ASIA USA

t-stat -38.34 -66.37 -71.52 -71.15 -77.82

probability 0.0000 0.0001 0.0001 0.0001 0.0001 Finland Estonia Poland Czech Hungary

t-stat -71.84 -53.19 -66.38 -67.82 -67.27

probability 0.0001 0.0001 0.0001 0.0001 0.0001

The use of first lags of differences in the mean equation means that there is no result for the remaining first observation in the first series. Thus, the final number of observation used is 5218 for all series except Estonia, for which it is 3299.

4.1. The market background of Central and Eastern European states

The following part offers a very brief discussion on how the Central and Eastern European states have come to be in the position they are, when this paper begins to examine them from 1995 onwards.

Following the disintegration of the Soviet Union in 1991, the former communist Central and Eastern European states began to adopt market economic principles, and open up their economies for foreign investment. However, countries had very different approaches on how market economies were to be gained. How former public companies were privatised and stock markets were set up varied, leading to mixed results. Of the Central and Eastern European economies studied, Hungary favoured slow paced privatisation and foreign ownership, meaning that after growth was slow and trading extremely thin in subsequent years following the founding of the Budapest stock exchange in 1991. Hungarian market capitalisation was only 2300 million US dollars in 1995, though this had increased to 16700 million US dollars in 2003. Allowing foreign investors free access to the Hungarian market resulted in 68,7 percent of shares in being in foreign ownership in 1997. (Schotman and Zalewska 2006, p. 467-468)

Like Hungary, Poland began with a low number of listed companies when their stock exchange was formed in 1991. However, unlike Hungary, the volume of trades was

27

extra ordinarily high for a new market economy from the onset. By the mid 1990’s trading volumes began to decline sharply as optimistically expected results failed to materialise. As the volume of trades decreased, the value of the market increased. In 1995 the market capitalisation of the Polish stock exchange had grown to just 4290 million US dollars. Market performance and the increased number of listings meant this figure rocketed to 44800 million USD by 2003. Though foreign investors were allowed to invest freely in Polish stocks foreign ownership was much lower than in Hungary, being under 20 percent in 1997. (Schotman and Zalewska 2006, p. 468- 469)

For the Czech Republic, the story of stock market development in the early years is quite different. Sparked by a policy of mass privatisation, the Prague stock exchange was quickly flooded with listings when it was founded in 1993. However, positive expectations soon proved unfounded, and midway through the 1990’s the Czech market slipped into a domestic crisis. Of the 1716 listed companies in 1995, only 65 remained in 2003. Meanwhile, the market capitalisation of the stock exchange dropped from 24500 million US dollars in 1995 to 9400 million US dollars in 2001, before recovering to 24800 million US dollars in 2003. The Czech domestic crisis in 1997 also meant that the effects of the 1997 East Asian crisis were not as evident in the Czech market as they were for Poland and Hungary. It is also notable that foreign investment was much more restricted in the Czech Republic, with just 37 percent of all shares available to foreign investment in 1997. (Schotman and Zalewska 2006, p. 468-469)

The Estonian market is much smaller than those discussed so far, with a market capitalisation of just 500 million US dollars in 1998, which had risen to 3790 million US dollars in 2003. The stock market was founded in 1996 and most stocks remain illiquid. After joining the European Union in 2004, Estonia standardised various market regulation inline with developed countries, though some restrictions on foreign investment were still in place. (Maneschiöld 2006, p. 31-34)

Finland’s economic background is different. Never a socialist country, Finland can offer an interesting case example of a peripheral stock market within a developed economy with relatively little political, legal or currency risk. Trading in smaller Finnish stock can at times be thin, while large bear sell-offs and bull buy-ups of