Systemic Risk and the European Project
Master Thesis Finance – University of Amsterdam
Program code: MSc FIN
Track: Quantitative Finance
By: Koen Oosterhek LLM (10547916)
Under the supervision of: Prof. dr. T. (Tanju) Yorulmazer
An empirical application of MES and ΔCoVaR
2
Statement of Originality
This document is written by Student Koen Oosterhek who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document are 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.
Acknowledgements
Special thanks go to Lotte Oosterhek and Marcel Oosterhek for the helpful comments, as well as to Professor Yorulmazer for the excellent supervision.
3
Table of Contents
Table of Contents ... 3
1. Introduction... 6
2. Literature review ... 10
2.1 What is systemic risk exactly? ... 10
2.2 Measuring and estimating systemic risk ... 12
2.3 The economic and financial effects of European integration ... 15
3. Methodology ... 17
3.1 Estimating systemic risk contribution ... 18
3.1.1 Conditional Value at Risk ... 18
3.1.2 Marginal Expected Shortfall ... 20
3.2 IV and FE ... 21
3.2.1 Fixed effects regression ... 21
3.2.2 Instrumental Variable regression ... 22
3.2.3 Control variables ... 22
4. Data and descriptive statistics ... 23
5. Effects of European integration on systemic risk contribution ... 27
5.1 Systemic Risk in Europe ... 27
5.1.1 Measuring Systemic Risk ... 27
5.1.2 Did the risk contribution change after EU or Eurozone membership? ... 28
5.2 European integration and systemic risk ... 30
5.2.1 Fixed effects ... 30
5.2.2 IV regressions ... 32
6. How did the systemic risk contribution change? ... 34
7. Robustness ... 39
8. Conclusion ... 42
Appendix ... 45
5
Systemic Risk and the European project. An
empirical application of MES and ΔCoVaR.
Koen Oosterhek
1July 2018
Abstract
I study the effects of becoming a member of the European Union (EU) and the Eurozone on systemic risk in Europe, hypothesizing that flourishing European financial integration disproportionally affected the systemic risk contribution of major European financial institutions. Systemic risk contribution is measured by using Marginal Expected Shortfall (MES) and Delta Conditional Value at Risk (ΔCoVaR). To solve for endogeneity issues a fixed effects regression and an instrumental variable regression are applied, utilizing the fact that EU and Eurozone countries entered at different times and not all European countries entered the EU or Eurozone. My main finding is that there is a significant shift of systemic risk contribution after an EU or Eurozone membership to major financial institutions, implying an increased role of larger financial institutions during systemic events after the memberships.
JEL classifications: F36, G01, G21, G32
Keywords: [Systemic Risk, Systemic contribution, European integration, EU, Euro, MES, ΔCoVaR, DCC-GARCH]
6
1. Introduction
The history of the European Union (EU) has been a compelling one. Founded initially as a coal and steel collaboration in the wake of the Second World War, it quickly redeveloped to become an integrated economic powerhouse with a deeply intertwined financial system (Boltho and Eichengreen 2008; Kim et al. 2005). The literature on the consequences of European integration is extensive. The general consensus being that European countries saw significant financial integration after they joined the EU or the Eurozone (e.g. Casu et al. 2016; Bekaert et al. 2013; Gębka and Karoglou 2013; Kalemli-Ozcan et al. 2010). Furthermore, several authors have studied a variety of aspects of systemic risk and systemic risk contribution in Europe (e.g. Acharya and Steffen 2012; Black et al. 2013; Engle et al. 2015). However, there is currently very little written about the direct effects of European integration on the contribution of major financial institutions to systemic risk. In this thesis I delve into this relationship. More specifically, I study if an EU or Eurozone membership increases the contribution of major financial institutions to systemic risk. To the best of my knowledge is this the first paper that analyses the direct causal relation between European integration and systemic risk contribution, and my thesis fills this gap in the literature.
There are several reasons to suspect a relationship between systemic risk contribution and European integration exists. First, there is evidence that EU membership caused major European institutions and markets to be more intertangled, depended and connected (e.g. Rughoo and Sarantis 2014; Bartman et al. 2007), both to other financial institutions as to the real economy. This occurred due to several phenomenon, including reduced capital movement constrains, which allowed for easier access to foreign capital (Bris et al. 2014), as well as becoming part of a global network of systemically important financial institutions (Paltalidis et al. 2015). These effects were further exacerbated with the introduction of the Euro (Kalemli-Ozcan et al. 2010). Furthermore, there is substantial evidence that financial entanglement increases systemic risk (e.g. Billio et al. 2012; Giglio et al. 2016). I therefore want to investigate if the entanglement caused by the integration of the European countries increased the contribution of major EU and Eurozone banks to European systemic risk, more so than it did in non-EU and non-Eurozone countries and – most importantly – more so than it did to smaller institutions. I believe larger institutions to be disproportionally affected by an EU or Eurozone membership, most importantly due to larger institutions generally benefitting more from
7 increased integration (Bris et al. 2014; Gozzi et al. 2010) and considering the relative ease to which access to foreign markets is acquired as well as the fact that the size of the financial institution is related to their engagement in foreign activities (Niepmann 2013) and to foreign markets access (Cetorelli and Goldberg 2009). Furthermore, recent evidence has shown that foreign financial institutions tend to take more risk (Chen et al. 2017), making the disproportional effect likely. My research will therefore specifically focus on the shift in systemic risk contribution to major financial institutions and does not necessarily entail or imply a general increase in systemic risk. The fact that European countries entered the EU and adopted the Euro at different times as well as the fact that not all countries in Europe entered the EU or adopted the Euro, provides me the opportunity to investigate this relationship.
The possibility of a shift in systemic risk contribution is a very relevant and critical topic to study. The relevance derives from the possibility to pre-emptively determine major actors in systemic events, as well as to determine factors that might increase the systemic risk contribution of financial institutions. The ability to monitor potential actors before and during systemic events might decrease and prevent possible costs as well as risks related to systemic events and is therefore invaluable. Furthermore, being aware of the factors that shift systemic risk to certain institutions might lead to a better allocation of resources and might reduce the severity of certain systemic events. Earlier research has, inter alia, already defined relevant sectors that contribute to systemic risk (Bernal et al. 2014), explored differences between to systemic risk related individual and contagion risks of institutions (Zedda and Cannas 2017) and has determined several, mostly idiosyncratic, factors related to changes in risk contribution (e.g. Weiβ et al. 2014; Laeven et al. 2016). My research would add to the current literature by determining if two European integration events – i.e. EU and Eurozone membership – have significantly shifted the contribution of major financial institutions to European systemic risk, providing helpful insights for relevant European institutions and actors as well as for countries and financial institutions in the midst of the ongoing process of European integration.
The proposed causal relationship I want to study is a difficult relation to study. This stems from two main reasons. First, it requires a proper method to estimate the systemic risk contribution of a financial institution, since the risk contribution cannot be directly measured. Secondly, there are some glaring endogeneity issues that might bias the results. For a lot of Eurozone
8 countries, the post-Euro period for example contained the 2008 Financial Crisis, during which systemic risk was at extraordinarily high levels (Giglio et al. 2016). Furthermore, there has been a more general increase in global financial integration, liberalization, technological innovations and faster communications that might bias the results (e.g. Bekaert et al. 2005).
I propose two methods to overcome these difficulties. The first problem I will try to solve by utilizing the literature regarding systemic risk that mostly developed in the wake of the 2008 Financial Crisis. In the past decade extensive research has be done to determine systemically important banks, systemic risk of the system and systemic risk of individual banks. I will utilize this burgeoning literature to determine systemic risk in the EU and Eurozone by using Marginal Expected Shortfall (MES) developed by Acharya et al. (2010) and Delta Conditional Value at Risk (ΔCoVaR) developed by Adrian and Brunnermeier (2016). These methods require return data as input and their implementation is relatively straightforward. The methods furthermore allow me to meticulously measure the often oscillating contribution to systemic risk of an individual institution over time. The second problem I will try to solve by applying methodology used by Bekaert, Harvey, Lundblad and Siegel (2013). The authors use several methods to study the effect an EU membership and the adoption of the Euro have on integration by using several proxies. By using fixed effects regressions and an instrumental variable regression the authors find that the Euro did not lead to more integration, but an EU membership did. The methodology applied by Bekaert et al. (2013) solved possible endogeneity in their research setup but is also perfectly apt for studying the effect the membership and the adoption had on
systemic risk (see section 3). I therefore will use both instrumental variable regressions and
fixed effects regressions, singling out the contribution of the membership and the Euro on systemic risk. My focus will be on large financial institutions and I take several steps to select the systemically most important ones.
My results show that there is an increase in the contribution of major systemically important institutions after both memberships. The relation is observed in fixed effects regressions and instrumental variable regressions. Furthermore, the relation is robust to sample bias (excluding the 2008 Financial Crisis and the Sovereign Debt crisis, as well as the exclusion of the “GIPSI” countries). Part of the results do not appear, however, when I use a different method of estimating ΔCoVaR, which highlights the complexity of estimating contributions to systemic
9 risk. I also test the relation separating the sample in several quantiles. Even when controlling for size and other factors an obvious pattern appears of an increasing contribution of larger institutions and a decreasing contribution of smaller institutions. The latter point supports the main conjecture that becoming a member of the EU or Eurozone shifts the systemic risk contribution to larger financial institutions. When including both dummies in a regression the results indicate that primarily the Eurozone membership causes the effect. This is somewhat hard to align with results found by Bekaert et al. (2013) who argue that the EU led to more integration but not the Euro. If the effect is indeed primarily driven by the Eurozone than it might be more probable that factors more closely related to the Euro, for example the fixing of exchange rates, causes the shift in contribution. There seems to be no observable variation in the effect when differentiating between stronger Euro countries and weaker countries as defined by Bris et al. (2014).
The results indicate that these two particular integration events made my sample of large systemically important European institutions more central in the systemic risk dynamic. The regressions imply a clear shift from other parts of the system towards these large institutions, which has evident policy consequences. Most importantly, it should encourage regulators to shift the monitoring and scrutinization of financial institutions in countries that want to enter or recently have entered the EU or the Eurozone to larger financial institutions, given the possibility of their increased role in current and future systemic events. Vice-versa, although the results do not necessarily imply this, there could also be a conversion of risk contribution to smaller institutions during disintegration processes, possibly due to the partial departure of foreign institutions and the increased role of smaller institutions in local financial activities. However, my results do not allow for such an interpretation and further research would have to properly establish this relation.
The remainder of my thesis is structured as follows. In section 2 I will analyse the current literature on systemic risk, systemic risk measures and the effects of an EU membership and the introduction of the Euro on several related factors. In section 3 I will discuss the applied methodology. Section 4 delves into the data and sample selection. Section 5 discusses the empirical results. Section 6 will delve deeper into the causes of these results. In section 7 I will discuss some robustness checks and I will conclude this thesis in section 8.
10
2. Literature review
There are three main bodies of literature related to the topic and hypothesis of my thesis. First, I explore the literature on systemic risk and examine a variety of definitions and explanations of systemic risk. Also, I will refer to the complexity of defining systemic risk or a systemic event. Second, I will holistically discuss methods proposed in the literature that measure systemic risk. Third, the literature on the effects of European integration and systemic risk in Europe, the EU and the Eurozone is of indispensable importance. I will succinctly discuss a variety of papers in this area and relate them to the topic of my thesis.
2.1 What is systemic risk exactly?
Since the Financial Crisis (2008) and the ensuing Great Recession, systemic risk has become significantly more present in the academic finance literature. Despite the renewed interest there seems to be no widely accepted comprehensive definition of systemic risk (Smaga 2014). Therefore, before delving into contemporary systemic risk measures, I want to shortly explore the semantics of systemic risk.
In 2009 the then president of the ECB, Jean-Claude Trichet, defined systemic risk as follows (Trichet 2009): “[Systemic risk] is the threat that developments in the financial system can cause a seizing-up or breakdown of this system and trigger massive damages to the real economy.” According to this definition systemic risk has two distinct features. First, it is related to and starts with developments in the financial system. Second, these developments can lead to “massive damages to the real economy”. Finding a measurement of systemic risk would then, using this definition, require measuring threatening developments in the financial system that could cause substantial problems for the real economy. The possibility of damaging the real economy demonstrates that systemic risk is more than just the joined idiosyncratic risk of all individual actors in the financial system. There is an added cost that goes beyond the damage done solely to the financial system. It would therefore be insufficient to, for example, aggregate Value at Risk (VaR) data (see section 2.2 as well) to measure the risk that the system poses. A proper measurement of systemic risk would require measuring damage done to the financial sector and measuring the damage done to the real economy. A further question would be what kind of developments in the financial industry could cause this system to breakdown. However, measuring the causes of systemic events requires measuring systemic
11 risk or at least defining a systemic event, which is at least to some extend arbitrary. Danielsson et al. (2016) for example found that during times of high financial volatility, different risk models tend to provide contrasting results.
One approach to circumvent this issue has been to identify the main actors during systemic events. The Basel Comity defines five areas that determine if an institution is such an actor, i.e. if it is a Global Systemically Important Bank (G-SIB). The variables that determine if a bank is a G-SIB is the bank its size, its interconnectedness, its substitutability2, its cross-border activity
and its complexity. The five variables are equally weighted (each 20%) and a score is determined for a sample of banks. Banks above a certain threshold will be qualified as G-SIB’s. G-SIB’s are subject to tougher and additional capital requirements, as well as extra supervision.
The variables that Basel III determines as relevant for determining which banks are systemically important, are similar to relevant idiosyncratic variables found in the literature. Zhang et al. (2015) study a variety of variables and emphasizes that predominately size is relevant for determining the contribution of institutions to systemic risk. Other authors find that the riskiness of the bank’s asset, leverage and again size, are relevant determinants (Hovakimian et al. 2012). It should also be noted that some authors have found outcomes that contradict these results. Most notably Weiß et al. (2014) who study changes in systemic risk, proxied by Marginal Expected Shortfall (MES), and conclude that among others size and leverage are not appropriate variables for determining systemic risk. This goes to show that different methods of estimation can provide widely different results and it implicates the inherent complexity of measuring systemic risk and its determinants.
However complex a phenomenon to measure, I want to note that it is in all respects important for an applied (e.g. by regulators) risk measure to get it right. A risk measure that would underweight possible systemic risk has obvious downsides, mostly related to the severe damage a systemic event can do to the real economy. On the other hand, a risk measure that would overweight certain systemic risks might have negative consequences as well. As the measure might cause regulators to impose redundant capital requirements which could unnecessarily reduce financial efficiency (see for example Barth and Seckinger 2018).
12 2.2 Measuring and estimating systemic risk
Especially since the 2008 Financial Crisis a wide variety of measures have been proposed in papers that try to understand and measure systemic risk. In this section I will discuss the literature on understanding and measuring systemic risk, as well as relate the measures to the definition in section 2.1 and analyse if the proposed methods measure systemic risk in line with this definition. I will refrain from exploring all the methods proposed in the literature and limit myself to a quick holistic summary.3
Even well before the 2008 Financial Crisis financial institutions engaged in measuring their financial risk. Most well-known and often applied measurement instruments were Value at Risk (VaR) and Expected Shortfall (ES), which could be used to measure the idiosyncratic risk of an institution. According to Christoffersen (2012: 12) the VaR simply measures the loss “that will only be exceeded p*100% of the time in the next K trading days”. Expected Shortfall is the expected return (loss) if such a VaR exceedance takes place. Although VaR and ES are widely used, the methods are not without criticism. As already mentioned in section 2.1 VaR is not sub-additive (Frey and McNeil 2002). It would be erroneous to add VaR results to measure the VaR of a portfolio. Other comments come from Berkowitz and O’Brien (2002), who analysed commercial banking data and argue that conventional ARMA or GARCH methods provide similar (or better) results as VaR methods. Furthermore, Danielsson et al. (2016) note that VaR can only be estimated, not measured. A lot of different methods have been proposed to estimate VaR (and ES), but results tend to vary.
After the Financial Crisis methods were proposed that extended upon these initial VaR and ES measures (as well as on other measures). Adrian and Brunnermeier (2016) for example introduced the, now renowned, ΔCoVaR. CoVaR measures the contribution of an institution to the risk of the system by measuring the extent to which the system is at its VaR level if the individual institution is at his VaR level. Acharya et al. (2017) propose a similar method, Systemic Expected Shortfall (SES), that measures the contribution of an individual firm to systemic risk. As part of the SES measure, Acharya et al. (2017) also developed Marginal
3 For a (comprehensive) survey of systemic risk measures, I would like to refer readers to Bisias et al. (2012), Kimura et al. (2017) and Giglio et al.
13 Expected Shortfall (MES), a now widely used method that simply measures the expected return of a financial firm conditional on their being a systemic event.
The ΔCoVaR and the MES method are applied in this thesis to determine the contribution of major banks to systemic risk. I apply these methods due to the relative convenience with which they can be implemented as well as their intuitive interpretation. Furthermore, the methods allow for estimating the systemic risk contribution of an institution over time, which provide the opportunity to gain an understanding of the time-variation in the contribution and ΔCoVaR is one of the few methods that has, ostensibly, a certain predictive power (Zhang et al. 2015). Additionally, the methods allow for estimating the risk contribution to the financial system as well as to the real economy, simply by using a sector wide index as a proxy for the system (see section 3), which aligns with the previously given definition of systemic risk.
ΔCoVaR and MES have often been used in the literature to measure systemic risk contribution in a similar context as my thesis. Weiß et al. (2014) for example use ΔMES – that is, the change in MES – to investigate changes in systemic risk after a variety of crises, while also paying attention to the difference between global and local systemic risk. The authors find that especially regulatory systems are main determinants of global systemic risk. Anginer and Demirgüç-Kunt (2014) study the role of various types of capital on – among others – MES and ΔCoVaR. According to the authors the type of capital matters with regard to systemic risk and most importantly they underscore the importance of quality capital. Bernal et al. (2014) use ΔCoVaR to investigate and determine which financial sectors are most relevant with regard to systemic risk contribution in the Eurozone. The authors conclude that the other financial
services are comparatively the biggest contributor. It has to be noted that, similar as for the
VaR and ES methods, the MES and ΔCoVaR methods are not without critique. Idier et al. (2013), for example, find that there are better methods (mainly balance-sheet variables) to predict severe equity costs than Marginal Expected Shortfall. Danielsson et al. (2016) find that CoVaR in general has a higher model risk than simple VaR or MES methods.4
Besides measuring systemic risk some authors studied important facets of systemic risk and events to get a better grip on the phenomenon. Notable is a study done by Yorulmazer and
14 Goldsmith-Pinkham (2010). The authors study the bank run on Northern Rock in the wake of the Financial Crisis and find, among other things, that the bank run caused substantial negative abnormal returns for other, related, UK banks. Other authors have also looked at spillover effects.5 Alter and Beyer (2014) use CDS spreads and develop an index that estimates
contagion. The authors find that government action (e.g. via policy) reduced the extent to which financial distress spilled over.
Related to papers that analyse and try to understand contagion and spillover effects are papers that analyse and construct so called network models. Most authors build upon mathematical topology for determining these networks. Paltalidis et al. (2015) apply a so called Maximum Entropy method to study contagion effects within financial institutions in countries that have the Euro as currency. Based on their model, the authors conclude that especially the southern part of the Eurozone is more vulnerable to contagion. Castén and Rancan (2015), who also use CDS spreads, construct a network model which relates the financial sector to the real economy, as well as linking sectors of different countries with each other. By applying simulations, they are able to get a better grasp of linkages between institutions and sectors in their model.
And finally, I want to specifically mention some literature on systemic risk in Europe. There are a couple of papers that use the methods discussed above to study and analyse the development of systemic risk in the European Union and Eurozone. Black et al. (2013) study systemic risk during the Sovereign Debt Crisis and find that it peaked in November 2011. Furthermore, they split out systemic risk in different EU countries and allude to the importance of regulating larger EU banks. Another important paper in this field is a paper by Engle et al. (2015). The authors study a sample of European financial institutions in the period 2002-2012. They conclude that in their sample mainly banks (82%) are responsible for European systemic risk and they estimate that France and the UK (50%) contribute the most to European systemic risk. Acharya and Steffen (2012) use the SES method to rank both European banks and countries. They also estimate the capital necessary to allow investors to regain confidence in the European system.
5 Meaning the impact or shift of financial distress from one sector or institution to another institution or other sectors of the economy. The terms
spillover and contagion are often used in the same context in a similar manner. For this thesis I will follow Rigobon (2016) and use the terms interchangeably.
15 2.3 The economic and financial effects of European integration
In this section I will delve into papers that are most closely related to my thesis. The section will be split into three respective parts. First, the literature on the effects of the EU and European integration on several factors will be discussed. Second, the literature on the effects of the introduction of the Euro will be analysed. Third, I will succinctly summarize and highlight the main hypothesis.
Several authors have explored the extent to which the European project caused further integration and convergence of the European countries. Christiansen (2014) studies bond markets in the context of European integration. The author finds that countries that entered the EU later have less integrated bond markets than the older EU countries. Furthermore, the author finds that for a segment of the countries that are part of the Economic and Monetary Union (EMU) integration is stronger than for countries that are not part of the EMU. Another paper that studies the effects of European integration is a paper by Rughoo and Sarantis (2014). The authors delve into the specifics and developments in the retail banking sector, finding that integration and convergence (in retail banking) increased in the period before the 2008 Financial Crisis, but not during the crisis period. These results are somewhat contradicted by Degl’Innocenti et al. (2017) who mainly study developments in productivity during the Financial Crisis, specifying several phases of the crisis. Their main conclusion is that there is substantial convergence during the most recent crisis.
A more extensive part of the literature focusses on the effects of the introduction of the Euro on several factors. The Euro is a relatively recent invention of the EU countries, which further cemented the European integration. After its official introduction in 1999 by eleven EU countries, the European markets, businesses and people were quick in adopting the new currency. Initially the Euro dropped against the dollar, but the Euro, and more specifically the Eurozone, experienced for the remainder of the early 2000’s a generally stable and sound growth (Bordo and James 2013). With the 2008 Financial Crisis there was a temporary change. The Financial Crisis caused an economic turbulence not experienced since the Great Depression and it unveiled the extent to which the European systems were exposed to a prodigious amount of systemic risk (see e.g. Acharya and Steffen 2012). The Financial Crisis also
16 engendered the sovereign debt crisis, which in turn exposed some structural flaws in the construction of the Eurozone (Orphanides 2014; O'Rourke and Alan 2013).
A lot of authors have studied the introduction of the Euro in a similar way as the consequences of an EU membership. Abad et al. (2010), for example, study the effects of the Euro on bond markets. The authors find evidence of an incomplete integration, arguing that local factors particularly determine debt returns in the Monetary Union. Beine et al. (2010) study the comovement of returns and argue that since the introduction of the Euro returns have been comoving more. The results are relevant for my thesis, in particular the indication by the authors of increased collective reliance of financial institutions in relation to financial openness during periods when stocks perform poorly. Another relevant paper in this regard is the paper by Bartram et al. (2007) who investigate the consequences of the Euro on stock dependence. The authors find that mostly in the bigger European countries markets became more dependent after the Euro introduction.
In a somewhat similar analysis Bartram and Wang (2015) study the dependence between European markets, using a Gaussian copula approach. Their results are extensive but could be summarized as a relation between the introduction of the Euro and a boost in dependence. They also find that dependence increased quickly up until the Euro crisis, after which the increase subdued markedly. Other research is done by Bris et al. (2009), who emphasize the increasing role of debt for Eurozone countries before 2008, relative to non-Eurozone countries. A very relevant finding is that the size of the firm is related to the extent to which an institution increased its financing by means of debt, where the authors highlight the disproportioned benefits of integration for a select group of large firms in this context.
My thesis is also closely related to a paper by Morana and Beltratti (2002), who analyse the role of the Euro in the context of stock volatility in a sample that ends in 2000. The authors apply both a GARCH model as well as some Markov simulations. The models provide contrasting results, but the Markov model indicates that the Euro initially increased volatility, but later reduced volatility in especially the southern countries, Spain and Italy. Also related is a study done by Kalemli-Ozcan et al. (2010). The authors try to find an answer to what caused increased financial integration due to the introduction of the Euro. They find that the
17 eradication of currency risk is the main driver. Finally, the paper by Bartman and Karolyi (2006) is also relevant. The authors study and discuss the effects of the Euro introduction on a variety of factors over a small period. The for my thesis most relevant factor the authors study is the effect on market risk. The authors find a falling off in market risk after the Euro introduction for nonfinancial entities. This is ostensibly true for entities inside and outside Eurozone.
To summarize, substantial research has been devoted to understanding systemic risk, to determine the sectors that contribute most to systemic risk and to analyse how risk contribution develops over time. Furthermore, quite a few papers have been written delving into the extend of European, EU and Eurozone integration as well as into the role of European integration on a variety of factors. However, there is currently not much written about the effect of European integration on the systemic risk contribution of financial institutions. In this thesis I will try to shed some light on this relationship. Given previous literature I would suspect there to be an increase in the contribution of major financial institutions relative to smaller institutions. Earlier evidence has pointed to the extent to which larger institutions generally benefit more from increased integration in a similar context (Bris et al. 2014; Gozzi et al. 2010). Additionally, authors have emphasized the relative ease to which access to foreign markets is acquired and that the size of the financial institution is related to their engagement in foreign activities (Niepmann 2013) as well as to foreign markets access (Cetorelli and Goldberg 2009). Given a, with regard to size, comparatively constant systemic risk contribution otherwise, or maybe even a tendency to take on more risk by larger institutions (Chen et al. 2017), I would suspect the relation to be positive. My research therefore specifically hypothesises that there is a positive shift in systemic risk contribution to major financial institutions, with larger institutions contributing more after the membership. A positive shift does not necessarily entail or imply a general increase in systemic risk.
3. Methodology
Studying the proposed relation requires two main steps. First, I will estimate systemic risk contribution by major banks by using Delta Conditional Value at Risk (ΔCoVaR) developed by Adrian and Brunnermeier (2016) and Marginal Expected Shortfall (MES) developed by Acharya et al. (2010). In section 3.1 I will delve into these methods. The second part of my methodology consists of applying some steps Bekaert, Harvey, Lundblad and Siegel (2013) used. I will utilize
18 their methodology to solve for endogeneity issues and to examine the possible causal relationship between EU integration and systemic risk. In section 3.2 I will explore these methods.
3.1 Estimating systemic risk contribution
There are a variety of methods proposed in the literature to estimate the systemic risk contribution of financial institutions. I will be using two very common and intuitive ones. First, Conditional Value at Risk based on the RiskMetrics method and second, Marginal Expected Shortfall. Both methods allow for estimating the time-varying systemic risk contribution of individual institutions and can be implemented using solely return data.
3.1.1 Conditional Value at Risk
As explained in the introduction CoVaR builds upon the initial Value at Risk (VaR) method. Recall that the VaR at t + 1 (e.g. tomorrow) for institution i which has return 𝑋𝑡𝑖 can be
described as
Pr(Xt+1i ≤ VaRiq,t+1) = 𝑞 (1)
This implies that the VaR tomorrow, conditional on the information until today, is equal to the p*100% left-tail observation. Given this VaR formula, Adrian and Brunnermeier (2016) define CoVaR as
Pr (Xs≤ CoVaRs|C(Xq i)|C(Xi)) = 𝑞 (2) This implies that the CoVaR is equal to the VaR of system s given event 𝐶(𝑋𝑖). System s could also be another institution or a group of institutions.6 The event 𝐶(𝑋𝑖) could for example be
the institution being at its VaR level or a specific systemic event or events could be defined. If we assume that 𝐶(𝑋𝑖) = 𝑉𝑎𝑅𝑞,𝑡𝑖 , Adrian and Brunnermeier (2016) than subsequently define
the change in CoVaR (ΔCoVaR) as
∆𝐶𝑜𝑉𝑎𝑅𝑞,𝑡𝑠|𝑖 = 𝐶𝑜𝑉𝑎𝑅𝑞,𝑡𝑠|𝑋𝑖=𝑉𝑎𝑅𝑞𝑖 − 𝐶𝑜𝑉𝑎𝑅𝑞,𝑡𝑠|𝑋𝑖=𝑉𝑎𝑅0.5𝑖 (3)
This implies that ∆𝐶𝑜𝑉𝑎𝑅𝑞,𝑡𝑠|𝑖 is equal to the difference of the system’s (Co)VaR when institution i is at its VaR q% level (e.g. its 5% VaR level) compared to being at its median VaR level. The final formula estimates the contribution of a financial institution to the systemic risk of another
19 institution or the whole system. More specifically the effect to the system (or another institution) of a change from a “normal” situation to a “distressed” situation of a financial institution is estimated. For my research question I require measuring the extent to which European institutions contribute to European systemic risk. Furthermore, as explained in section 2.1, I argue that systemic risk is more than just the damage done to the financial system. I will therefore use an European sector wide index as a proxy for the system (see section 4 as well).
ΔCoVaR can be estimated in several ways. Adrian and Brunnermeier (2016) propose a quantile regression approach, I will be using a GARCH method (see e.g. Girardi and Ergün 2013). This relies on the further assumption that the returns of system s and institution i are jointly normally distributed, which allows ∆𝐶𝑜𝑉𝑎𝑅𝑞,𝑡𝑠|𝑖 to have the following construct
∆𝐶𝑜𝑉𝑎𝑅𝑞,𝑡𝑠|𝑖= −𝜎𝑡𝑠𝑝𝑡𝑠𝑖Φ𝑞−1 (4)
See appendix A for the derivations. In this particular formula Φ𝑞−1 is the inverse normal
cumulative distribution, 𝜎𝑡𝑠 the volatility series of the system and 𝑝𝑡𝑠𝑖 the conditional
correlation of the system and the institution. There are a variety of ways to estimate ΔCoVaR. I will be using both the RiskMetrics approach and the DCC-GARCH approach. The latter will be applied in section 7 for robustness. The RiskMetrics approach consist of three simple steps. The first step consists of applying a GARCH model to the returns of the system and of the institutions. I will be applying an univariate GJR-GARCH method. The GJR-GARCH method allows for asymmetric effects of the returns on the variance (so called leverage effect), where negative returns might possibly lead to higher volatility. The model has the following structure (Christoffersen 2012) 𝜎𝑖,𝑡+12 = 𝜔 + 𝛼𝑟𝑖,𝑡2 + 𝛼𝜃𝐼𝑡𝑟𝑖,𝑡2 + 𝛽𝜎𝑖,𝑡2 (5) With 𝐼𝑡 = { 1, 𝑖𝑓 𝑟𝑖,𝑡2 < 0 0, 𝑖𝑓 𝑟𝑖,𝑡2 ≥ 0 (6) Where in this case i can be equal to s. The parameters of the model and the subsequent volatility series can be estimated by maximum likelihood. With the univariate series I can now also estimate the conditional correlation between the system and institution i. For this second step I first compute the standardized returns, i.e. 𝜎𝑟̂𝑖,𝑡
𝑖,𝑡 and
𝑟𝑠,𝑡
𝜎
20 smoothers (EWMA) to the standardized returns (using the RiskMetrics λ = 0.94) to compute 𝑞𝑡𝑖𝑠, 𝑞𝑡𝑖𝑖 and 𝑞𝑡𝑠𝑠 (see e.g. Engle 2002). Where 𝑞𝑡𝑖𝑖 has the following structure
𝑞𝑡+1𝑖𝑠 = (1 − 𝜆)𝑧𝑡+1𝑖 𝑧𝑡+1𝑠 + 𝜆𝑞𝑡𝑖𝑠 (7)
for all pairs of i and s, including 𝑖 = 𝑠. With the conditional correlation being equal to 𝑝̂𝑡𝑠𝑖 =
𝑞̂𝑡𝑠𝑖 √𝑞̂𝑡𝑖𝑖𝑞̂𝑡𝑠𝑠
(8)
3.1.2 Marginal Expected Shortfall
Due to the inherent complexity of estimating the systemic risk contribution as well as the possibility of measurement errors, I will also estimate systemic risk using Marginal Expected Shortfall (MES) as developed by Acharya et al. (2010). A second method allows me to compare both results and provides also some robustness. MES builds from Expected Shortfall (ES), which has the following structure
𝐸𝑆𝑡+1𝑝 = −𝐸𝑡[𝑋𝑖|𝑋𝑠 ≤ 𝑉𝑎𝑅𝑝,𝑡+1𝑠 ] (9)
MES can, in a similar way, be defined as the average return of institution i, given that there is a systemic event. I will be using a similar notation as Weiß et al. (2013), defining MES as
𝑀𝐸𝑆𝑖𝑝= −𝐸[𝑋𝑖|𝐼𝑝𝑠] (10)
Where, again similar as in Weiß et al. (2013), 𝐼𝑝𝑠 is equal to the p*100% worst days of system s
(return wise) and 𝑋𝑖 is the equity return of institution i. I will calculate MES as follows
𝑀𝐸𝑆𝑖,𝑡𝑝 =1𝑇 ∑𝑇𝑞=1𝑋𝑞𝑖|𝐼𝑝𝑠 (11)
Where q are all observations where 𝐼𝑝𝑠 (e.g. the system) is in its p tail. I will be using p=5%.
To create a time-series of an institutions systemic risk contribution a rolling MES method can be estimated (Brownlees and Engle 2012; Idier et al. 2013). I will be using a one year, or rather 252 business days rolling period, defining the risk contribution today as the average contribution during the past year. In section 7 I will also apply a longer rolling period, which gives similar results. The rolling MES has the following structure
𝑀𝐸𝑆𝑖,𝑡𝑝 = ∑252𝑞=0𝑋𝑡−1−𝑞𝑖 (12) Where 𝑋𝑡𝑖 = { 𝑋𝑡𝑖, 𝑖𝑓 𝑋𝑡𝑠 ≤ 𝑉𝑎𝑅𝑝𝑠 0, 𝑖𝑓 𝑋𝑡𝑠> 𝑉𝑎𝑅𝑝𝑠 (13)
21 In other words, the sum of the returns given that the system is in its tail. For 𝑋𝑡𝑖 log returns will
be used. The system will, again, be proxied by an European sector wide index.
3.2 IV and FE
Simply looking at the change in systemic risk since the introduction of the Euro is prone to endogeneity mistakes. The period after 2000 contained some (deep) crises, which might influence the results and, as Cannas and Zedda (2017) showed, contributions to systemic risk are influenced by the significance of the crisis. Furthermore, there are several factors that might influence the result but are consistent over time or across countries. The inherent nature of European systemic risk and the fact that I utilize an index wide sector means that the systemic risk contribution might fluctuate significantly over time influenced by macro-economic factors and events that do not vary across countries. Furthermore, my sample contains different European countries with different geography, financial culture and business practices. I will therefore apply methodology used by Bekaert et al. (2013), utilizing both a fixed effects regression controlling for country and year fixed effects and an instrumental variable regression.
3.2.1 Fixed effects regression
With the fixed effects regression I try to estimate the contribution of integration to systemic risk by controlling for time and country specific fixed effects. Furthermore, I control for several institution specific variables. Controls will be explained in section 3.2.3.
First, I will estimate the effects of becoming an EU member using both systemic risk time-series separately. The linear regression model is as follows
𝑆𝑦𝑠𝑡𝑒𝑚𝑖𝑐 𝑟𝑖𝑠𝑘 𝑐𝑜𝑛𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛𝑡𝑖 = 𝑎 + 𝑏 ∗ 𝐸𝑈𝑡𝑖+ 𝑐 ∗ 𝐶𝑡𝑖+ 𝑑𝑖+ 𝑒𝑡+ 𝜖𝑡𝑖 (14)
Where 𝑆𝑦𝑠𝑡𝑒𝑚𝑖𝑐 𝑟𝑖𝑠𝑘 𝑐𝑜𝑛𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛𝑡𝑖 is equal to the Marginal Expected Shortfall at time t for
institution i or the delta Conditional Value at Risk at time t for institution i. Both will be separately specified in regression tables. Furthermore, 𝐸𝑈𝑡𝑖 is a dummy variable equal to 1 if
the firm is in an EU country, 𝐶𝑡𝑖 is a vector of control variables, 𝑑𝑖 are country fixed-effects and
𝑒𝑡 are time fixed-effects. Standard errors will be clustered at institution level. Second, I will
estimate the effects of the introduction of the Euro, which has the following, similar, structure 𝑆𝑦𝑠𝑡𝑒𝑚𝑖𝑐 𝑟𝑖𝑠𝑘 𝑐𝑜𝑛𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛𝑡𝑖 = 𝑎 + 𝑏 ∗ 𝐸𝑈𝑅𝑂𝑡𝑖+ 𝑐 ∗ 𝐶𝑡𝑖+ 𝑑𝑖+ 𝑒𝑡+ 𝜖𝑡𝑖 (15)
22 Where we now have an Euro dummy 𝐸𝑈𝑅𝑂𝑡𝑖 equal to 1 if the institution is in a country that
has adopted the Euro. As explained in previous sections I expect the dummies to be positively related to the systemic risk contribution, meaning an increased contribution for this particular sample of large systemically important institutions. A negative beta would imply a decreasing contribution for this sample after the integration events, which invariably means an increase somewhere else in the system.
3.2.2 Instrumental Variable regression
To further test the robustness of the results I also apply an IV regression. For my thesis to remain comparable to the paper by Bekaert et al. (2013) I will use the same instrument as the authors use, the distance to Brussel, in my IV. The authors believe the instrument to be relevant given that many of the activities of the Eurozone and EU take place in Brussel as well as a more general institutional growth materializing from that area of Europe. The regression has the following structure
𝐸𝑈𝑡𝑖 = 𝑎 + 𝑏 ∗ 𝐵𝑟𝑢𝑠𝑠𝑒𝑙𝑖+ 𝑐 ∗ 𝐶𝑡𝑖+ 𝑑𝑡+ 𝜐𝑡𝑖 (16)
Where 𝐵𝑟𝑢𝑠𝑠𝑒𝑙𝑖 is a variable containing the distance to Brussel for institution i established in
an EU country and where 𝐶𝑡𝑖 and 𝑑𝑡 are control variables and time fixed-effects respectively.
Although Bekaert et al. (2013) use the instrument only with regard to the EU, I believe the instrument to be equally relevant for the Eurozone countries given that similar arguments apply. I therefore also instrument the Euro dummy in the following regression
𝐸𝑈𝑅𝑂𝑡𝑖 = 𝑎 + 𝑏 ∗ 𝐵𝑟𝑢𝑠𝑠𝑒𝑙
𝑖+ 𝑐 ∗ 𝐶𝑡𝑖+ 𝑑𝑡+ 𝜐𝑡𝑖 (17)
Where, again, 𝐶𝑡𝑖 and 𝑑𝑡 are control variables and time fixed-effects respectively. The
instrument will be tested for relevance in section 5.
3.2.3 Control variables
Both in the fixed effects regression and the IV regression I use several control variables that are considered to be relevant determinants of systemic risk contribution. As idiosyncratic variables I will first and foremost use size, which will be proxied by the natural logarithm of assets. Based on the literature I expect size to be positively related to the systemic risk contribution since it is generally assumed that larger institutions contribute more (Hovakimian et al. 2012; Zhang et al. 2015). Furthermore, leverage (Acharya and Thakor 2016) will be used as well as type of capital (Anginer and Demirgüç-Kunt 2014), for which I will be using separately
23 the log of the tier 1 and tier 2 capital ratios as specified by the Basel accords, and finally I will use loans divided by assets, as utilized by Anginer and Demirgüç-Kunt (2014). Controls are further explained in appendix B. Leverage is generally assumed to be positively related to the systemic risk contribution with more leveraged firms contributing more and being more susceptible to systemic events. Capital tiers should have different effects depending on the capital tier (see Anginer and Demirgüç-Kunt 2014) with capital tier 1 being negatively related to systemic risk, i.e. reducing the systemic risk contribution, and tier 2 being positively related, supposedly increasing the systemic risk contribution. The final variable is expected to be negatively related to systemic risk as it is a proxy of the relative amount of activities an institution spends on conventional financial activities (Anginer and Demirgüç-Kunt 2014).
For the IV regression I will also use several country specific controls. These include a dummy equalling 1 if the country is located in Eastern Europe as used by Bekaert et al. (2013) and a dummy indicating if the particular country in my sample has more than forty different institutions. I also use the Hirschman Herfindahl Index (HHI) as used by Weiβ et al. (2014) as a measure of competitiveness. I furthermore control for two country specific factors related to credit. Credit info is a variable between 0 and 8 indicating the available credit information (and more specifically the depth of the information) and credit to private is a ratio that determines that ratio of domestic credit to the private sector. I expect the Eastern Europe and more than forty dummy to be respectively negatively and positively related to systemic risk. Eastern European countries generally contain systemic institutions of a relatively smaller size. More than forty mostly applies to the core countries and are therefore mostly closer to the global financial system and additionally have larger financial institution. With regard to the credit variables similar arguments as for the more than forty apply, where I except more indebted countries to be more affected by systemic events. Finally, for the HHI I suspect a positive correlation, given the argument that more concentrated economies rely stronger on a small group of financial intermediaries and are possibly more prone to systemic events.
4. Data and descriptive statistics
To analyse my hypothesis a data sample is constructed of European countries. My sample consists of 35 European countries. 28 of these are currently EU countries of which 17 entered the EU during the sample period. 19 of these are currently Eurozone countries, all of which
24 entered during the sample period. This implies that for the EU membership I have partial pre and post sample data and for the conversion to the Euro all countries have pre and post data. The countries that didn’t enter the EU or the Eurozone are used as control countries. Table 1 displays the countries in my sample, their date of entry, their date of conversion and the sample period.
Table 1
Sample selection. This table presents countries in Europe (plus ISO code, their date of ascension to the EU, date of conversion to the Euro, and their current currency (plus ISO code). The final column shows the length of the sample per country. All countries entered the Euro on the first of January. Countries that entered before 1993, entered a predecessor of the EU, e.g. the European Coal and Steel Community or the European Economic Community. The following countries are founders of (the predecessor of) the EU: Belgium, France, (West-)Germany, Italy, Luxembourg and the Netherlands. Countries are mentioned in alphabetical order. Source: https://ec.europa.eu/info/business-economy-euro/euro-area/euro/eu-countries-and-euro_en.
Country Entered EU Current currency Conversion to Euro Sample period
Austria (AUT) 1995 Euro (EUR) 1999 1985-2018
Belgium (BEL) 1957 Euro (EUR) 1999 1985-2018
Bulgaria (BGR) 2007 Bulgarian Lev (BGN) - 1995-2018
Croatia (HRV) 2013 Croatian Kuna (HRK) - 1998-2018
Cyprus (CYP) 2004 Euro (EUR) 2008 1997-2018
Czech Republic (CZE) 2004 Czech Koruna (CZK) - 1995-2018
Denmark (DNK) 1973 Danish Krone (DKK) - 1985-2018
Estonia (EST) 2004 Euro (EUR) 2011 1996-2018
Finland (FIN) 1995 Euro (EUR) 1999 1985-2018
France (FRA) 1957 Euro (EUR) 1999 1985-2018
Germany (DEU) 1957 Euro (EUR) 1999 1985-2018
Greece (GRC) 1981 Euro (EUR) 2001 1988-2018
Hungary (HUN) 2004 Hungarian Forint (HUF) - 1993-2018
Iceland (ISL) - Icelandic Krona (ISK) - 1995-2018
Ireland (IRL) 1973 Euro (EUR) 1999 1986-2018
Italy (ITA) 1957 Euro (EUR) 1999 1985-2018
Latvia (LVA) 2004 Euro (EUR) 2014 2000-2018
Liechtenstein (LIE) - Swiss Franc (CHF) - 1989-2018
Lithuania (LTU) 2004 Euro (EUR) 2015 1997-2018
Luxembourg (LUX) 1957 Euro (EUR) 1999 1985-2018
Macedonia (MKD) - Macedonian Denar (MKD) - 2007-2018
Malta (MLT) 2004 Euro (EUR) 2008 1995-2018
Netherlands (NLD) 1957 Euro (EUR) 1999 1985-2018
Norway (NOR) - Norwegian Krone (NOK) 1986-2018
Poland (POL) 2004 Polish Złoty (PLN) - 1993-2018
Portugal (PRT) 1986 Euro (EUR) 1999 1988-2018
Romania (ROU) 2007 Romanian Leu (RON) - 1997-2018
Russia (RUS) - Russian Ruble (RUB) - 1995-2018
Serbia (SRB) - Serbian Dinar (RSD) - 2009-2018
Slovakia (SVK) 2004 Euro (EUR) 2009 1995-2018
Slovenia (SVN) 2004 Euro (EUR) 2007 1995-2018
Spain (ESP) 1986 Euro (EUR) 1999 1985-2018
Sweden (SWE) 1995 Swedish Krona (SEK) - 1985-2018
Switzerland (CHE) - Swiss Franc (CHF) - 1985-2018
Turkey (TUR) - Turkish Lira (TRY) - 1990-2018
Ukraine (UKR) - Ukrainian Hryvnia (UAH) - 2008-2016
25 The European Union has known several phases of integration starting in particular in the wake of the Second World War. The core countries that founded the initial organisations were Belgium, France, (West-)Germany, Luxembourg, the Netherlands and Italy. The actual European Union was established in 1993. For this thesis I will use the ascension to the EU or one of its predecessors as the main integration event. Of the 28 EU countries, 6 are founders and 6 entered a predecessor. The other countries entered the actual European Union. With regard to the Eurozone membership, a group of 11 countries entered in 1999 after which another 6 countries entered over a period of sixteen years.
The data required for this thesis is mainly return data and some accounting data for controls. Daily closing price equity data is derived from Compustat – Capital IQ. For most countries I have data starting in 1985, however not for all countries data is available for the full period. My sample is therefore unbalanced. For the index data, which will represent the system in my thesis, STOXX Europe 600 will be used. As previously explained I believe a European sector wide index to be appropriate as a proxy for the system given that I want to measure the contribution to European systemic risk, were systemic risk is more than just the damage done to the financial system. The STOXX Europe 600 contains a variety of companies in Europe of differing sizes.7 Furthermore, as intended, the STOXX Europe 600 index is an European index, so I will be
measuring the contribution to European systemic risk. The STOXX Europe 600 data is retrieved from Datastream, with sample data available from 1986 until 2018. Finally, for the descriptive statistics and for control variables accounting data is retrieved from Compustat.
After dropping all the sic codes smaller than 6000 and larger than 6799 I start off with a sample of 20,066 financial institutions. I drop institutions with substantially anomalies in the data and then take several steps to select the largest systemically most important institutions in a country. Given my hypothesis of increased contribution for major institutions this step is vital. First, institutions are dropped with a market cap of 10bn or lower and or with assets of 5bn or less. If no institution in the country is of that size (this is only true for Lithuania and Latvia) the 95percentile of lowest market cap banks are dropped. Furthermore, I drop institutions that have a sample size of less than a year (or rather 252 business days) to prevent computational
26 issues with MES. My final sample consists of 740 large systemically important financial institutions.
Table 2
Descriptive statistics. Table depicting descriptive statistics of the institutions in my sample both before and after the respective countries entered the EU or the Eurozone. Variable definitions are given in Appendix B. ROA, leverage and MTB, credit to private are ratios. More than forty, Eastern Europe, HHI, credit info and credit to private are country specific variables. More than forty is a dummy indicating if the country in my sample has more than forty institutions. Credit info is a score between 0 and 8 determining how much information with regard to credit is available in each country. Employers in thousands. Variables are winsorized at the 1% level.
Non-EU (N=241) EU (N=561) Non-Euro (N=668) Euro (N=318)
Mean Sd Mean Sd Mean Sd Mean Sd
ln(assets) 12.356 1.295 12.538 1.533 12.428 1.399 12.656 1.651 ROA 3.055 2.483 3.873 3.701 3.902 3.627 3.216 3.092 Leverage 21.732 15.862 24.042 21.319 23.682 19.871 23.136 21.041 MTB 1.239 1.701 1.814 2.931 1.973 2.864 1.045 2.186 ln(Cap1) 2.609 0.357 2.442 0.342 2.494 0.346 2.489 0.375 ln(Cap2) 2.001 1.839 0.809 1.042 0.547 1.553 0.801 0.649 ln(depos) 7.801 1.919 8.592 2.455 8.149 2.252 8.874 2.477 ln(cash) 7.127 1.508 6.596 2.047 6.795 1.827 6.532 2.199 Employers 9.592 8.601 13.577 12.074 12.419 12.031 13.527 9.949
More than forty 0.359 0.48 0.72 0.449 0.624 0.484 0.66 0.474
Eastern Europe 0.138 0.345 0.051 0.221 0.102 0.302 0.077 0.088
HHI 0.065 0.021 0.536 0.017 0.058 0.019 0.053 0.019
Credit info 6.445 0.638 6.709 1.126 6.613 1.033 6.72 1.048
Credit to private 1.138 0.507 0.995 0.256 1.083 0.377 0.913 0.186
Table 2 displays summary statistics for the institutions in the sample. Variables are winsorized
at the 1% level. To present the data two dummy variables have been created equalling 1 if the country is in the EU during that year or in the Eurozone respectively. My sample consists of 241 institutions in EU countries, 561 institutions in EU countries, 668 institutions in non-Eurozone countries and 318 institutions in non-Eurozone countries. The data in table 2 shows that the institutions differ quite a lot over the different sample periods. For example, the leverage of banks increases by about 10% from non-EU to EU and total assets is on average about 40bn greater for the EU sample than the non-EU sample. This further emphasizes the need for proper control variables. The country control variables used have very similar values as in the literature, this is specifically true for the HHI index (Weiβ et al. 2014) and Eastern Europe (Bekaert et al. 2013). Note that Eastern countries are only a small part of the data, which might bias the results. Note furthermore that credit info is relatively high, indicating the generally high availability of credit information in Europe.
27
5. Effects of European integration on systemic risk contribution
5.1 Systemic Risk in Europe
Before measuring the extent to which an EU or Eurozone membership influences the systemic risk contribution in Europe, it would be informative to succinctly display and analyse the pattern of the risk contribution time-series. I will therefore graph the time-series of four major banks and relate them to the literature in section 5.1.1. In section 5.1.2 I will give a first arithmetic indication of the developments related to the memberships.
5.1.1 Measuring Systemic Risk
As explained in section 3, systemic risk contribution is measured by using ΔCoVaR and MES. The estimates are a time-series of daily observations across the institutions sample period. To get an idea about the structure of the measures in my sample, figure 1 displays both ΔCoVaR and MES for four systemically important institutions. Although the lines follow similar patterns they don’t necessarily estimate the same effect. Where MES estimates the contribution of an individual firm to the systemic risk in the system, ΔCoVaR estimates the change in the systems VaR when the institution shifts from its VaR level to its median level. Both however can be – and will be – used as proxies for systemic risk contribution. The risk contribution of the institutions depicted in figure 1 clearly fluctuate over time, with in some periods the risk contribution being significantly higher than in other periods. During the Financial Crisis risk contribution by these four major banks was at its highest, which aligns with the general consensus about the Financial Crisis (also) being a systemic banking crisis (e.g. Black et al. 2013). Notice also the apparent “ghost” feature of the MES measure, which can be attributed to the fact that a historical simulation method (rolling) is used. The depicted values on the left axis can be interpreted as percentage. For example, 0.04 MES value implies a 4% contribution to the systemic risk of the system. These particular firms have on average a relatively high contribution which emphasizes their more central role in the systemic risk dynamic. The average ΔCoVaR value is about 0.5% and the average MES value about 1%, but substantially fluctuating over time. The percentage contribution as well as the observed pattern of my estimations are very similar to percentages and patterns found in the literature (see e.g. Acharya et al. 20178; Adrian and Brunnermeier 2016; Banulescu and Dumitrescu 2014).
28 Figure 1. Systemic Risk contribution of four major institutions estimated by MES (252 rolling) and ΔCoVaR
(RiskMetrics).
5.1.2 Did the risk contribution change after EU or Eurozone membership?
A simple method to better understand the possible differences in risk contribution before and after the introduction of the Euro or becoming a member of the EU is to measure the mean ΔCoVaR and MES before and after the event and subtract the pre from the post means. This would obviously not say much about causality given the many factors that might play a role, but it does provide an early insight into the possible changes after the memberships. I therefore compute simple pre- and post-event arithmetic and weighted means for institutions for which I have observations available both before and after the event (i.e. EU or Eurozone membership). A weighting ratio is created by dividing the institutions mean assets by the total assets of institutions that are in my sample and for which pre and post event data is available.
Table 3 displays these results. Furthermore, in appendix C the arithmetically summed
-0.05 0 0.05 0.1 1991 1994 1997 2000 2003 2006 2009 2012 2015
ING
ΔCoVaR MES -0.05 0 0.05 0.1 1987 1990 1993 1996 1999 2002 2005 2008 2011 2014 2017Deutsche Bank
ΔCoVaR MES -0.02 0 0.02 0.04 0.06 0.08 0.1 1987 1990 1993 1996 1999 2002 2005 2008 2011 2014 2017HSBC
ΔCoVaR MES -0.05 0 0.05 0.1 1987 1990 1993 1996 1999 2002 2005 2008 2011 2014 2017Société Générale
ΔCoVaR MES29 differences per location are displayed where all countries are included that have pre and post event data.
Table 3
Arithmetic and weighted differences. This table displays the change in risk contribution after an EU membership and the introduction of the Euro. ΔCoVaR is estimated by the RiskMetrics approach and MES is calculated as the mean of the log returns during the systems worst 5% days (see methodology). For the first two columns observations per country are equally aggregated. For the second column weights are constructed as the institutions assets divided by the total assets of all institutions. Countries that did not switch currency or entered the EU are not included in this particular calculation and table. For the EU countries the event is becoming a member of the EU, for the Eurozone countries the introduction of the Euro. ∗∗∗ significance at 1%; ∗∗ significance at 5%; ∗ significance at 10%.
Change MES (1) Change ΔCoVaR (2) Weighted change MES (3) Weighted change ΔCoVaR (4) Becoming an EU member 0.0119* (0.00722) 0.00261*** (0.00051) 0.00163* (0.00093) 0.00529** (0.00302)
Entering the Eurozone 0.00503***
(0.00197) 0.00121*** (0.00025) 0.00251*** (0.00053) 0.00747*** (0.00191)
As can be noted in table 3, the systemic risk contribution of the major financial institutions did indeed change after both events for all four methods of estimation. The changes are quite sizeable. Looking for example at the second row of the first column (1), the unweighted change in MES before the event ‘Entering the Eurozone’, it shows that the MES increased with 0.00503. This implies on average more than a half percent increase in systemic risk contribution after becoming a member of the EU for this sample. If we take a look at the significance of the change it is notable that the results in the first row, becoming a member of the EU, are a bit less significant than the results in the second row, becoming a member of the Eurozone. This might indicate that there is more variability in the change in systemic risk for the first event and might be an illustration that becoming a member of the Eurozone is more directly related to an increase in systemic risk contribution.
A further thing to note is that the results seem to vary quite a lot. The percentage change depicted seems to be different for ΔCoVaR and MES, where the change in arithmetically added MES overall seems to be somewhat higher than the change in ΔCoVaR and the weighted change in ΔCoVaR higher than weighted MES. Furthermore, the systemic risk contribution after the Eurozone membership seems to be smaller, but more significant, than the EU membership. This particular observation is in contrast with the main conclusion of Bekaert et al. (2013) who’s
30 results imply that becoming a member of the EU increased integration, whereas becoming a member of the Eurozone didn’t.
To be sure, subtracting pre and post means is solely for demonstrative purposes and has little to do with causality. As mentioned before, the pre-Euro and to a lesser extend the pre-EU sample period contained arguably less systemic events or risk than the post period, mostly due to the 2008 Financial Crisis. I also included solely institutions that have pre and post treatment systemic risk observations so not all countries and institutions are added to this calculation. Furthermore, other factors could also be at play, such as a more general increase in systemic risk by major financial institutions or the possibility that only countries that are financially integrated to a certain extent and therefore happen to have more systemic financial institutions are allowed to enter the EU or Eurozone. The next section deals with these issues.
5.2 European integration and systemic risk
The calculation in the previous section was prone to obvious endogeneity issues. This section and the next attempts to resolve these, first by applying a fixed effects regression and second by using an instrumental variable regression.
5.2.1 Fixed effects
One of the main problems with the ΔCoVaR and MES data is that the estimates fluctuate substantially over time influenced, among others, by global and regional macro-economic factors and events. Due to the inherent nature and computation of the measures, however, the effects don’t tend vary across countries. Using time fixed effects allows me therefore to considerably reduce the omitted variable bias. Furthermore, my sample contains different European countries with different geography, financial culture and business practices. I will therefore also control for country fixed effects.
I apply the fixed effects regression, using two dummy variables: an EU dummy equalling 1 if institution i is located in country j that is a member of the EU in year t and an Eurozone dummy equalling 1 if institution i is located in country j that has adopted the Euro in year t. I run separate regressions for the effect on MES and on ΔCoVaR and use heteroskedasticity robust (clustered at entity level) standard errors. Furthermore, I control for idiosyncratic variables that
