Bank business models and systemic risk in Europe

(1)

Bank business models and systemic risk in Europe

Abstract:

Using the ΔCoVaR systemic risk measure, this thesis investigates the relationship between certain bank characteristics and systemic risk for a sample of European banks. Fee &

commission income, trading income, asset size and the loans-to-assets ratio are found to have a positive relationship with systemic risk: an increase in one of those variables increases the contribution of a bank to systemic risk. Furthermore, by grouping banks into different bank business models, the findings in this thesis indicate that the group of banks that contribute most to systemic risk are the international diversified lenders, with large universal banks and fee-focused banks coming in second and third, respectively.

MSc thesis Cherine Botros, 10592164

Supervisor: dhr. prof.dr. S.J.G. van Wijnbergen MSc Economics

Faculty of Economics and Business University of Amsterdam

(2)

Statement of originality

This document is written by Student Cherine Botros 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.

(3)

1. Introduction

Before the financial crisis of 2007-2009, financial regulation was mostly focused on limiting banks’ idiosyncratic risk. However, one of the lessons of the financial crisis is that policy makers and financial regulators also need to take into account risk spillovers between banks and, therefore, that policy aimed at reducing systemic risk is also needed (Brunnermeier et al. 2012).

One possible contributor to systemic risk is non-interest income. Brunnermeier et al. (2012) found that a higher share of non-interest income increases a bank’s contribution to systemic risk in the United States (US). This thesis will investigate whether the same relationship holds for European banks. According to Albert & Alexandre (2012) European banks’ non-interest income grew from 20% in 1988 to 47% in 2006. During the crisis there was a drop in non-interest income, largely due to trading losses, but it has been gradually rising again. Currently the share of non-interest income is almost back at pre-crisis levels (ECB 2016).

Furthermore, using other bank characteristics, this thesis will expand the analysis and will also look at which bank business models contribute the most to systemic risk. Knowing how certain bank business models and bank characteristics contribute to systemic risk could have important policy implications, especially for designing macro-prudential policy.

Summarizing, this thesis will try to answer two, related, research questions:

- Do banks with a higher fee & commission income or a higher trading income contribute more to systemic risk in Europe?

- Do certain bank business models contribute more to systemic risk in Europe? It contributes to the existing systemic risk literature in the following ways: first, relatively little research has been done on the effect of bank characteristics on systemic risk in Europe, especially using the ΔCoVaR systemic risk measure (the measure that is used in this thesis). Second, these papers have focused only on the large European banks while the sample in this thesis also includes smaller banks. Third, while there has been research on which bank characteristics affect a bank’s contribution to systemic risk, very few papers have used actual bank business models.

(5)

This thesis is organized as follows: section 2 gives an overview of the different definitions of systemic risk within the literature and sets out the theory behind the ΔCoVaR systemic risk measure. Section 3 describes the methodology used to estimate ΔCoVaR as well as the bank business model classifications by Lucas et al. (2017) that will be used in this thesis. Section 4 gives an overview of both the data and how it was collected. In the first part of section 5 the characteristics of the banks in this sample are compared to the business model categorization of Lucas et al. (2017) to see how well they match. In the second part an

overview is given of the ΔCoVaR estimates per bank business model and, lastly, two regressions are run in order to answer the research questions presented above and two robustness checks are conducted. Section 6 presents a discussion on the results and, finally, section 7 concludes.

2. Theory

2.1 The definition of systemic risk

Systemic risk is often seen as a ‘we-know-it-when-we-see-it' concept and there are several definitions of systemic risk within the literature. According to Caruana (2010), systemic risk can be defined as “a risk of disruption to financial services that is caused by an impairment of all or parts of the financial system and has the potential to have serious negative consequences for the real economy". Gerlach (2009) identifies 3 characteristics of systemic risk: it must impact a substantial portion of the financial system, it involves spillovers of risk from one institution to many others and episodes in which systemic risk materializes are associated with highly adverse macro economy effects. Benoit et al. (2013) give a more minimalist definition of systemic risk: “the risk that many market participants are simultaneously affected by severe losses, which then spread through the system". In their paper they give an overview of existing systemic risk measures, and according to them this is the definition that applies to most of these measures. The systemic risk measure that will be used in this thesis is the ΔCoVaR of Adrian & Brunnermeier (2016).

2.2 Measuring systemic risk: the theory behind the ΔCoVaR

The ∆CoVaR is based on the Value-at-Risk (VaR) measure, which captures the financial risk of an individual bank. It is implicitly defined as the q-quantile of the probability distribution:

(6)

(1) Pr 𝑋_!! _{≤ 𝑉𝑎𝑅}

!!,! = 𝑞

So that it describes the worst expected loss over a certain time period t with probability q. In the context of ∆CoVaR 𝑋_!!_{stands for the percentage of asset value of bank i.}

Because VaR only looks at a bank in isolation it is not an appropriate measure for systemic risk. Addressing this particular shortcoming of VaR, Adrian & Brunnermeier (2016) proposed a measure for systemic risk they call CoVaR. The prefix ‘Co’ stands for conditional,

contagion or comovement, emphasizing its systemic nature.

𝐶𝑜𝑉𝑎𝑅_!!"!#$%|! is the VaR of the financial system conditional on institution i being in distress (as measured by its own VaR).1 Again 𝐶𝑜𝑉𝑎𝑅_!!"!#$%|! is implicitly defined as the q-quantile of the conditional probability distribution:

(2) Pr 𝑋!"!#$% _{≤ 𝐶𝑜𝑉𝑎𝑅} !

!"!#$%|!!!!"#!!_|𝑋

! = 𝑉𝑎𝑅!! = 𝑞

It should be noted that one of the properties of CoVaR is that it is directional, which means that the CoVaR of the financial system conditional on a particular institution does not equal the CoVaR of that institution conditional on the financial system. What is measured here is the contribution of an individual institution i to systemic risk. It is also possible to calculate

𝐶𝑜𝑉𝑎𝑅_!!|!"!#$%, which measures which financial institution would be most at risk in case of a financial crisis. Adrian & Brunnermeier (2016) call this the ‘exposure CoVaR’. However this thesis only looks at the contribution of an institution to systemic risk.

Finally, in order to get a measure for the marginal contribution of institution i to systemic risk, Adrian & Brunnermeier (2016) introduce the ∆CoVaR:

(3) Δ𝐶𝑜𝑉𝑎𝑅_!!"!#$%|! = 𝐶𝑜𝑉𝑎𝑅_!!"!#$%|!!!!"#!!_{− 𝐶𝑜𝑉𝑎𝑅}

!!"!#$%|!!!!"#$%&

!

Which is the difference between the VaR of the financial system conditional on institution i in distress and the VaR of the financial system conditional on institution i in ‘normal’ times (the median state).

1_{More generally, CoVaR can be calculated between institution j and institution i: 𝐶𝑜𝑉𝑎𝑅}

!!|!. In this thesis

(7)

2.3 Bank business models and systemic risk

Several studies have used the ΔCoVaR, or other systemic risk measures, to look at the relationship between bank characteristics and systemic risk. Brunnermeier et al. (2012)

studied the effect of banks’ non-interest income on systemic risk in the US using the ΔCoVaR and the Marginal Expected Shortfall (MES), a measure of systemic risk developed by

Acharya et al. 2010.2 Unlike the ΔCoVaR, which measures how much a bank contributes to systemic risk, MES measures how much a bank is exposed to the financial system being in distress and uses the same conditioning logic as the ‘exposure ΔCoVaR’. They find that banks with a higher non-interest income to interest income ratio contribute more to systemic risk. More specifically, a one standard deviation increase to a banks’ non-interest income to interest income ratio increases its systemic risk contribution by 11.6% according to the ΔCoVaR measure and 5.4% according to MES. When decomposing non-interest income into trading income and investment banking/venture capital income, they find that a one standard deviation increase to both types of income increases a bank’s contribution by approximately the same amount: 5% according to ΔCoVaR and 3% according to MES (Brunnermeier et al. 2012).

Using their own measure of systemic risk, Van Oordt & Zhou (2014) also looked at US banks in their article on systemic risk and bank business models. Similar to the MES measure, systemic risk of a bank is defined in their article as its sensitivity to severe shocks in the system. However they decompose systemic risk into two dimensions: the level of a bank’s tail risk and the linkage between the bank’s tail risk and shocks in the financial system. The tail risk of a bank contains both shocks in the financial system and other shocks. According to Van Oordt & Zhou (2014), if two banks have an equal amount of tail risk, then the one with the higher sensitivity to shocks in the system should be considered more risky, because this bank is expected to suffer larger losses compared to the other bank.

They find that non-interest income and bank size contribute positively to systemic risk. An increase of 1% in the non-interest income of banks, such as trading income, investment banking, venture capital revenues and net gains on loan sale, corresponds to an increase in the sensitivity to shocks by around 0.4%. There is no significant relation between banks’ tail risks and non-interest income, indicating that non-interest income is strongly

2_{Brunnermeier et al. (2012) call it SES (Systemic Expected Shortfall) in their paper, but in the literature it is}

(8)

related to shocks in the financial system. Besides non-interest income, bank size is also positively related to shocks in the system, with larger banks being significantly more sensitive to shocks than smaller banks. This is not due to a positive relation between size and bank tail risk, but due to the high dependence between large banks and the financial system in case of a crisis.

Finally, Van Oordt & Zhou (2014) also looked at the traditionality of the banks’ balance sheets. Banks that focus on lending have significantly higher levels of tail risk, but a lower sensitivity to shocks in the financial system. Banks with a higher deposit-to-asset ratio also exhibit a lower level of sensitivity to shocks. This suggests that banks with a more traditional balance sheet are generally less sensitive to severe shocks in the financial system (Van Oordt & Zhou 2014).

Laeven et al. (2016) use an international sample of banks to assess the effect of bank size and capital on systemic risk in 2007-2008. Employing both the ΔCoVaR measure and SRISK, a measure by Brownlees & Engle (2012), which makes use of an institution’s MES, leverage and size, they find strong evidence that bank size contributes to systemic risk. Total assets increase a bank’s contribution to systemic risk by approximately one-third its standard deviation according to the ΔCoVaR measure and half its standard deviation by the MES measure. Moreover, they find that better capitalized banks contribute less to systemic risk, particularly when banks are large (Laeven et al. 2016).

Their findings are somewhat corroborated by López-Espinosa et al. (2012), who also take a global approach to look at the relation between short-term wholesale funding and systemic risk. Focusing on large international banks and using the ΔCoVaR, they find that short-term wholesale funding is highly significantly positively related to systemic risk. They also find strong evidence for a positive relation between bank size and systemic risk, but, unlike Laeven et al. (2012), no evidence for a relationship between leverage and systemic risk.

Karimalis & Nomikas (2017) computed the ΔCoVaR for a sample of 42 large European banks over the time period 2002-2012, using a new methodology based on copula functions. They find that for European banks size and leverage are the biggest determinants of systemic risk. Again these finding are only somewhat corroborated by another study done on European banks. Wosser (2017) used the ΔCoVaR and the MES to analyze whether current balance sheet characteristics affect future values of these systemic risk measures. His sample comprises of 30 large European banks. He finds that institution size, maturity mismatch, non-performing loans and the non-interest-to-interest ratio are the biggest factors influencing

(9)

future systemic risk realizations. Using the MES measure he also finds a significant relationship between leverage and systemic risk, but not when using ΔCoVaR.

Summarizing, there seems to be a consensus on the effect of asset size and non-interest income on systemic risk within the empirical literature. Furthermore, most studies also find a significant effect for leverage.

3 Methodology

3.1 Estimation ΔCoVaR

In order to compute the ∆CoVaR, Adrian & Brunnermeier (2016) use the weekly growth rate of market-valued financial assets, 𝑋_!!_{. 𝑋}

!! is defined by: (4

) 𝑋

_!!

=

𝑀𝐸𝑡 𝑖_{∙ 𝐿𝐸𝑉𝑡}𝑖_−𝑀𝐸 𝑡−1 𝑖 _{∙ 𝐿𝐸𝑉} 𝑡−1 𝑖 𝑀𝐸𝑡−1 𝑖 ∙ 𝐿𝐸𝑉𝑡−1𝑖

where 𝑀𝐸_!!_{is the market value of equity and 𝐿𝐸𝑉}

!! is the ratio of total assets to book equity. The weekly growth rate of the market-valued assets of all N banks, 𝑋_!!"!#$%, is defined by:

(5)

𝑋

_!!"!#$%

=

∑

_𝑖=1𝑁 𝑀𝑉𝑡−1

𝑖 _{∙ 𝐿𝐸𝑉}

𝑡−1 𝑖 _{∙ 𝑋𝑡}𝑖 ∑_𝑗=1𝑁 𝑀𝑉_𝑡−1𝑖 ∙ 𝐿𝐸𝑉_𝑡−1𝑖

As can be seen from equation (5), in order to compute financial asset growth for the system, a system portfolio is constructed using the lag of the market-valued total assets as a weighing variable.

3.1.1 Estimation time-variant ΔCoVaR: state variables

According to Adrian & Brunnermeier (2016), because the ΔCoVaR is time-variant, lagged state variables are needed to capture the time variation in the joint distribution of 𝑋_!!_and 𝑋_!!"!#$%. The state variables used in this thesis are:

1. The VSTOXX Indices, the volatility of the EUROSTOXX 50, which is a proxy for volatility in the European market.

(10)

2. A short-term liquidity spread, measures by the difference between the three-month UK repo rate and the three-month LIBOR rate.

3. The change in the Euribor rate.

4. The change in the slope of the yield curve, measured by the spread between the French government 10-year bond yield and the French 3-month Treasury bill yield.

5. The FTSEurofirst 300 Index, which features the largest 300 European companies ranked by market capitalization, as a proxy for the weekly equity market returns. These state variables are European counterparts to the state variables used in Adrian & Brunnermeier (2016).3

3.1.2 Estimation time-variant ΔCoVaR: quantile regressions

Quantile regressions are used as the estimation method of ΔCoVaR. In order to estimate the ΔCoVaR, two conditional VaRs for each bank need to be calculated: 𝐶𝑜𝑉𝑎𝑅_!!"!#$%|!!!!"#!!

and 𝐶𝑜𝑉𝑎𝑅_!!"!#$%|!!!!"#$%&!_{. For the first CoVaR expression, setting q=1%, the following}

1% quantile regressions are run in order to estimate the coefficients 𝛼!_{, 𝛾}!_,

_𝛼

𝑠𝑦𝑠𝑡𝑒𝑚|𝑖_{, 𝛽}!"!#$%|!_{, 𝛾}!"!#$%|!_:

(6) 𝑋_!! _{= 𝛼}! _{+ 𝛾}!_𝑀

!!!+ 𝜀!!

(7) 𝑋_!!"!#$% = 𝛼!"!#$%|!_{+ 𝛽}!"!#$%|!_𝑋

!!+ 𝛾!"!#$%|!𝑀!!!+ 𝜀!!"!#$%|!

where 𝑀_!!! represents the lagged state variables and 𝑋_!!_{and 𝑋}

!!"!#$% the weekly growth rate of assets of institution i and the financial system, respectively.

Similarly, for the second CoVaR expression, the one conditional on institution i being in the median state, the following 50%-quantile regression is run to estimate 𝛼!,!"#$%&_and 𝛾!,!"#$%&_:

(8) 𝑋_!! _{= 𝛼}!,!"!"#$_{+ 𝛾}!,!"#$%&_𝑀

!!!+ 𝜀!!,!"#$%&

3_{The state variables are a combination of the state variables used in López-Espinosa et al. (2012) and Karimalis}

(11)

The coefficients from the quantile regressions of equation (6) and (8) are then used to predict an individual institution’s VaR and median asset return:

(9) 𝑉𝑎𝑅_!%,!! _{= 𝛼}! _{+ 𝛾}!_𝑀 !!!

(10) 𝑋_!!,!"#$%& = 𝛼!,!"#$%& _{+ 𝛾}!,!"#$%&_𝑀 !!!

After obtaining the VaR of an individual institution and its asset return in the median state from equation (9) and (10) it is possible to predict 𝐶𝑜𝑉𝑎𝑅_!%!"!#$%|!!!!"#!!%_and

𝐶𝑜𝑉𝑎𝑅_!%!"!#$%|!!!!"#$%&!_{, the VaR of the financial system conditional on institution i being in}

distress and the VaR of the financial system conditional on institution i in the median state. Using the coefficients from equation (7):

(11) 𝐶𝑜𝑉𝑎𝑅_!%,!!"!#$%|! = 𝛼!"!#$%|!_{+ 𝛾}!"!#$%|!_𝑀

!!!+ 𝛽!"!#$%|!𝑉𝑎𝑅!%,!! (12) 𝐶𝑜𝑉𝑎𝑅_!%,!!"!#$%|!,!"#$%& = 𝛼!"!#$%|!_{+ 𝛾}!"!#$%|!_𝑀

!!!+ 𝛽!"!#$%|!𝑋!!,!"#$%&

Finally the ΔCoVaR can be calculated using equations (11) and (12): (13) ∆𝐶𝑜𝑉𝑎𝑅_!%,!! _{= 𝐶𝑜𝑉𝑎𝑅}

!%,!!"!#$%|!− 𝐶𝑜𝑉𝑎𝑅!%,!!"!"#$|!,!"#$%&

By definition the value of VaR is negative. However a large part of the ∆CoVaR literature uses positive values of VaR, and, correspondingly, positive values of ∆CoVaR because this makes more intuitive sense when analyzing the empirical results. Following this convention, positive values will also be used in this thesis so that a higher (∆Co)VaR value indicates higher (systemic) risk.

3.2 Bank business models

A classification of bank business models is needed in order to study the effect of these models on systemic risk. This thesis will use the classification by Lucas et al. (2017). In their study on bank business models at zero interest rates, Lucas et al. (2017) identified the following six business models within their sample of 208 European banks:

(12)

A. Large universal banks (including G-SIBS):

These banks are the largest, with assets up to €2 trillion per bank. Net fee &

commission income and trading income are significant sources of income for these institutions. They are the most leveraged banks in the sample and around 60% of their operating revenue comes from interest-bearing assets such as loans and securities holdings.

B. International diversified lenders:

This group is comprised of banks with a high share of non-domestic loans to total loans. They are the largest banks after the first category, with total assets between €100 – 500 billion per bank. Like the large universal banks, trading income is a large part of their income. They tend to be non-deposit funded and therefore have a high loans-to-deposit ratio.

C. Fee-focused banks:

Banks in this group focus on fee-based banking activities “such as transaction banking services, trade finance, credit lines, advisory services and guarantees” (Lucas et al. 2017, p.28). Their total asset size is below €100 billion per bank. A high share of their income, approximately 30%, comes from net fees & commissions. According to Lucas et al. (2017), this particular category also contains ‘weak’ banks whose income from traditional banking activities are low. Net interest income appears to be low, which raises the share of net fee & commission income. Banks in this category have a relatively high loans-to-assets ratio of approximately 70%.

D. Domestic diversified lenders:

Banks in this category are small, with total assets typically below €50 billion. They have relatively low leverage ratios and small trading and derivatives

books. According to Lucas et al. (2017) this business model is widely used in Europe: a quarter of all the banks in their sample fit into this category.

E. Domestic retail lenders

With less than €25 billion in assets these banks are the smallest institutions, together with the small international banks. As the name suggests, they are focused on (domestic) loans and therefore have a high loans-to-assets ratio. They are well

(13)

capitalized and their trading income share is low. Most of their income comes from interest-bearing assets.

F. Small international banks

Like the domestic retail lenders, small international banks typically have less than €25 billion in assets, are well capitalized and have small trading and derivatives books. Their loans-to-deposit ratio, however, is much lower and their clients are not only domestically located.

4. Data

4.1 Data collection

There are 30 banks in the dataset, five for each bank business model. The banks were chosen using the list of banks from the article of Lucas et al. (2017). The decision to only include 30 banks in the dataset is partly based on time constraints due to data collection and partly due to data availability issues (as explained below). The data spans a time period from 2006Q1 to 2017Q2. Stock returns, market equity and the data for the ΔCoVaR state variables are from Datastream. The following quarterly income and balance sheet data have been collected from reports from the banks’ websites: net interest income, net fee & commission income, net trading income, total income, assets, loans, deposits, equity.

There were three requirements in order for a bank to be chosen from the list. The first is that the bank had to be listed during that time period, otherwise the ΔCoVaR could not be

calculated. The second and third concern data availability: income and balance sheet data had to available for a certain time period and on a quarterly basis. The last requirement,

especially, limited the amount of banks that could be chosen. For very small banks and for certain countries, France and the UK, only semi-annual reports are available and those do not always include quarterly figures. Non-availability of quarterly data is a known issue within the ΔCoVaR literature and is dealt with by converting the data to quarterly data. In this thesis the semi-annual balance sheet data is converted to quarterly data by taking the average of the amount the period before and the period after and semi-annual income data is converted to quarterly data by splitting it in half.

This does not significantly affect most of the data, with the exception of trading income, which can be quite volatile. Four of the banks in the sample have semi-annual

(14)

balance sheet data and three of those also have semi-annual income data. However, as these three banks are small international banks with a very low trading income share, the effect of the conversion of semi-annual data to quarterly data on the results of the analysis is likely to be very low.

Finally, another issue concerning data collection is that the use of quarterly reports limited the kind of data that could be collected. Because every bank has its own ways of reporting data, it was only possible to collect basic bank characteristics data (such as asset size, loans, deposits, etc). This is because not every bank (especially if it is very small) reports very detailed

breakdowns of, for example, the kind of loans they give out and to which kind of customer.

4.2 Description of the data

The following banks are in the sample:

Bank Category Country

1 ING A NL 2 HSBC A UK 3 Deutsche Bank A DE 4 Société Générale A FR 5 Credit Suisse A CH 6 BBVA B ES 7 DNB B NO 8 Erste Bank B AT 9 Svenska Handelsbanken B SE 10 KBC B BE 11 Intesa Sanpaolo C IT 12 Mediobanca C IT 13 Raiffeisen Bank C AT 14 Santander C ES 15 Swedbank C SE

(15)

16 Aareal D DE

17 Alpha Bank D GR

18 Banco BPI D PT

19 Credito Emiliano D IT

20 Bank Polska Kasa Opieki D PL

21 Alandsbanken E FI

22 Helgeland Sparebank E NO

23 Høland og Setskog Sparebank E NO

24 Melhus Sparebank E NO

25 Sandnes Sparebank E NO

26 BRD F RO

27 Liechtensteinische Landesbank F LI

28 Privedna Banka Zagreb F HR

29 Permanent TSB F IE

30 VP Bank F LI

Table 1. List of banks in the sample

A: large universal banks, B: international diversified lenders; C: fee-focused banks, D: domestic diversified lenders; E: domestic retail lenders, F: small international banks.

Within the constraints of data availability, an effort was made for the sample itself and the categories to contain as many different countries as possible. As can be seen from table 1, this was not possible for category E, the domestic retail lenders. For this category 4 out of 5 banks are Norwegian and all of the banks are from Northern Europe. In Lucas et al. (2017) this category also included banks from Switzerland, the United Kingdom (UK), Germany, Spain, France and the Netherlands. Unfortunately most of those banks were not listed (or were not listed during the entire timeframe of the sample) and none of them met the data availability requirements. However, Norwegian banks were well represented within the category, with approximately 40% of all the banks in the sample being Norwegian.

In total, 19 countries are represented within the sample. Table 2 shows how many banks are from each country:

(16)

Countries Banks Countries Banks Norway 5 UK 1 Italy 3 Netherlands 1 Austria 2 France 1 Germany 2 Finland 1 Spain 2 Portugal 1 Sweden 2 Switzerland 1 Liechtenstein 2 Greece 1 Romania 1 Belgium 1 Poland 1 Croatia 1 Ireland 1

Table 2. Countries and the number of banks from each country

This is obviously not a representative sample of Europe. Smaller countries such as Norway and Liechtenstein appear more than once, while there is only one British bank and one French bank. Part of this is due to data availability issues (particularly the amount of non-listed banks) and also, of course, due to the small sample size. A lot of banks in the larger countries, the UK, Germany, France, etc., are not listed and therefore cannot be used in the sample. This is especially the case for their smaller banks (in categories E and F), which causes a ‘strange’ breakdown of countries, with four banks being located in Norway in category E and two banks being located in Liechtenstein in F.

Due to the small sample size (5 banks per business bank models) and the decision to diversify the categories in terms of countries, most countries only appear once or twice. For example, there are 23 German banks in the dataset of Lucas et al. (2017), 12 of which are in group A. It would have been possible to have five German banks in that particular category. However, in that case, there would have been no banks from France and the UK in the

sample. A footnote to this is that this decision was only possible in categories A and B, where there are relatively more banks that meet all of the data requirements. In the categories with the smaller banks, C, D, E and F, data availability was the only concern.

Another thing that should be noted is that this thesis is based off the categorization of the sample of 208 European banks of Lucas et al. (2017), which in itself is already a small sample of all the banks situated in Europe. Like in this thesis, Norway has the most banks in the sample. See Appendix A for a comparison of their dataset and the one in this thesis.

(17)

Summarizing, there seems to be a trade-off between representing the business models equally and representing the countries within Europe in an appropriate manner. This has

repercussions for the validity of the analysis going forward: results found in this thesis cannot be generalized to Europe as a whole.

5. Empirical analysis

5.1 Descriptive statistics

5.1.1 Bank characteristics per bank business model: comparison to Lucas et al. (2017)

Considering the small sample size, it is necessary to look at the characteristics of the banks in the sample to see how well they match the bank business models by Lucas et al. (2017). This section uses asset size, the loans-to-assets ratio, leverage, the loans-to-deposits ratio and income shares to compare the groups in the sample to the descriptions of the business models as given in section 3.2.

Graph 1. Average asset size(in € mln) per business model per quarter for the period 2006Q1-2017Q2 A: large universal banks, B: international diversified lenders;

C: fee-focused banks, D: domestic diversified lenders; E: domestic retail lenders, F: small international banks.

0 200000 400000 600000 800000 1000000 1200000 1400000 1600000 1800000 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 Average asset size (in € mln) per business model per quarter (2006Q1-2017Q2) A B C D E F

(18)

As can be seen from graph 1, there is a huge difference between asset sizes between the different models.4_{However B and C, the international diversified lenders and the fee-focused}

banks, have roughly the same asset size. According to the categorization by Lucas et al. (2017), group B was supposed to have a much larger asset size than group C. Both groups also have a larger average asset size than is reported in Lucas et al. (2017), especially C. Asset size in this category was supposed to be below €100 billion, but in this sample it is almost ten times as large.5 According to the literature asset size seems to have a significant effect on systemic risk, so the results in this thesis will likely overestimate the effect of these groups.

Graph 2. Average loans-to-assets ratio per business model for the period 2006Q1-2017Q2 A: large universal banks, B: international diversified lenders;

The loans-to-assets ratios of the different bank business models seem to fit with the

categorization of Lucas et al. (2017). The domestic retail lenders, E, have the highest ratio, of around 80%, and the large universal banks, A, the lowest. For the fee-focused banks, group C, the ratio is a little off: they were supposed to have a loans-to-assets ratio of around 70%, but

4_{See appendix B for a graph that gives a better look at the asset sizes for the smallest groups (D, E, F).}

5_{This is because there are two very large banks in category C: Santander and Intesa Sanpaolo. Unfortunately due}

to data constraints it was not possible to use other banks. However it should be noted that this is technically not incorrect as they are both categorized under group C by Lucas et al. (2017).

0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 Average loans-to-assets ratio per business model per quarter (2006Q1-2017Q2) A B C D E F

(19)

in this sample the average is around 60%. It is interesting that that the fee-focused banks have such a high loans-to-assets ratio relative to the other groups, both in this thesis and in Lucas et al. (2017), considering that the loans-to-assets ratio is often used as an indicator of a

traditional bank business model.

Graph 3. Average leverage ratio (assets/book equity) per business model for the period 2006Q1-2017Q2

In graph 3 the effects of stricter financial regulation are noticeable, as leverage is slowly converging. Especially for the large universal banks and the international diversified lenders, groups A and B respectively, leverage has gone down considerably since 2006. It is

interesting that for group C, the fee-focused banks, leverage was the most stable and it is also relatively low. This group has roughly the same asset size as B, but there is a large difference in leverage for most of the time period, especially in the beginning. This could indicate that the banks in group B, the international diversified lenders, are riskier than the banks in group C, the fee-focused banks.

0 5 10 15 20 25 30 35 40 45 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 Average leverage ratio (assets/book equity) per business model per quarter (2006Q1-2017Q2) A B C D E F

(20)

Graph 4. Average loans-to-deposits ratio per business model per quarter for the period 2006Q1-2017Q2

Although it is not included in the regressions, the loans-to-deposits ratio was mentioned by Lucas et al. (2017) in their classification of the business models. According to Lucas et al. (2017) group B should have a high loans-to-deposits ratio and group F should have a low loans-to-deposits ratio. From graph 5 it is clear that is not the case. F and C, the small

international banks and the fee-focused banks have the highest ratios. For F in particular there is one bank that has a very high loans-to-deposits ratio, influencing the average for that group.

The following table shows the average income shares in % over the entire time period (2006Q1-2017Q2):6

6_{A table is used instead of a graph due to the high volatility of the income shares.} 0 0,5 1 1,5 2 2,5 3 3,5 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 Average loans-to-deposits ratio per business model per quarter (2006Q1-2017Q2) A B C D E F

(21)

A B C D E F Net interest income 48 58 59 62 67 55 Fee & commission income 21 25 29 27 20 30 Trading income 13,8 9,3 6,7 6,6 4,9 7,4

Table 3. Average income shares in % over the period 2006Q1-2017Q2 A: large universal banks, B: international diversified lenders;

The income shares mostly seem to match the classification of the bank business models by Lucas et al. (2017). Fee & commission and trading income is relatively high for the large universal banks and the international diversified lenders, A and B. The fee & commission income share for the fee-focused banks is around 30%, which matches Lucas et al. (2017) as well. Likewise, net interest income for the domestic diversified lenders and the domestic retail lenders is high. The banks in group F, however, do deviate from their classification as trading and fee & commission income is relatively high, with the fee & commission income share being even higher than that of the fee-focused banks.

Summarizing, the groups in this sample do not correspond perfectly to the bank business models by Lucas et al. (2017), particularly when it comes to asset size. However on average they seem to match and there is a clear difference between the groups with respect to their bank characteristics. Therefore, going forward, it is assumed that the groups in the sample are representative of the bank business model classification by Lucas et al. (2017). The discussion of the results will, amongst other things, take a look at possible issues with this assumption.

5.1.2. Results of the ΔCoVaR estimates

(22)

Graph 5. Average ΔCoVaR estimation per business model per quarter in the period 2006Q2-2017Q2 A: large universal banks, B: international diversified lenders;

Graph 5 shows the average level of the ΔCoVaR estimated per business model per quarter. Although ΔCoVaR estimates are calculated on a weekly basis, they are aggregated to a quarterly basis for the analysis. Interestingly, the large universal banks, group A, are not the business model with the highest values; it’s actually the international diversified lenders that contribute the most to systemic risk. The large universal banks, group B, and the fee-focused banks, group C, are in second and third place, switching places around 2010 and again in 2016. The domestic retail lenders, category E, contribute the least to systemic risk by far, about half the average contribution of the small international banks, category F. This does seem to be consistent with asset size, as the domestic retail lenders are about half the size of the small international banks. This graph indicates that although the average ΔCoVaR values per bank business model seem to be somewhat consistent with average asset size, it does not determine a bank’s contribution to systemic risk alone, especially when it concerns the big banks (groups A, B and C). Otherwise we would expect the large universal banks to contribute most to systemic risk.

0 0,1 0,2 0,3 0,4 0,5 0,6 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 Average ΔCoVaR per business model per quarter (2006Q1-2017Q2) A B C D E F

(23)

However it is not enough to look at quarterly averages. The summary statistics of the weekly values per bank business model give a better idea of the volatility of the ΔCoVaR:7

A B C D E F

Mean 4.67 5.05 3.13 2.73 0.56 1.95

Std. Dev. 2.21 1.88 0.88 1.18 0.28 0.77

Min 1.96 2.44 1.28 1.63 0.05 1.11

Max 18.47 17.72 8.82 11.43 2.89 7.03

Table 4. Weekly summary statistics for (average) ΔCoVaR per group over the period 2006Q1-2017Q2 A: large universal banks, B: international diversified lenders;

Table 4 shows that there are big differences in the standard deviations of the groups, with the standard deviation of group A being the highest, 2.21, and the standard deviation of group E, 0.28, the lowest. Interesting is the standard deviation and the mean of group C, the fee-focused banks, compared to A, the large universal banks. In the graph with the aggregated quarterly values these two groups were shown to have around the same level of contribution to systemic risk, but the table shows that this is probably due to the low standard deviation of C combined with a lower mean and the high standard deviation of A combined with a higher mean. As stated before, it is unfortunately not possible to analyze weekly ΔCoVaR estimates due to the nature of the regression used in the analysis but from the table we can see that the large universal banks are more volatile than the fee-focused banks.

The differences between the groups are even better illustrated by looking at the changes in value over time. Below are the graphs for group A and group E8:

7_{Summary statistics of the VaR and X (the market-valued weekly asset change) can be found in appendix C.} 8_{Graphs for the other groups can be found in appendix D.}

(24)

Graph 6. Weekly average ΔCoVaR and VaR estimation for the large universal banks (A) over the period 2006Q2-2017Q2

(25)

Graph 7. Weekly average ΔCoVaR and VaR estimation for the domestic retail lenders (E) over the period 2006Q2-2017Q2

From the graphs it is clear that the contribution of group A to systemic risk is much higher than group E. This makes intuitive sense; larger banks being in distress will have more of an impact on the financial system than smaller banks. The graphs also show that the values for the VaR for these two groups are actually quite similar and that, as Adrian & Brunnermeier (2016) posit, VaR is not a good indicator for systemic risk.

5.2 Regression models

Two regressions are run in order to be able to answer the research questions. The first regression looks at the effects of trading and fee & commission income and the second regression is used to analyze the contribution of the different bank business models to

systemic risk. For the regressions the weekly ΔCoVaR data is aggregated to a quarterly level. The first regression for the effects of fee & commission income and trading income on

systemic risk uses the following regression model9: (14) ΔCoVaR𝒊_𝒕 _{= 𝛽}

!+ 𝛽!𝑇𝑟𝑎𝑑𝑖𝑛𝑔𝐼𝑛𝑐𝑜𝑚𝑒𝑆ℎ𝑎𝑟𝑒!,!!!+ 𝛽!𝐹𝐶𝐼𝑛𝑐𝑜𝑚𝑒𝑆ℎ𝑎𝑟𝑒!,!!!+ 𝛽!𝐿𝑜𝑔(𝑎𝑠𝑠𝑒𝑡𝑠)!,!!!+ 𝛽!𝐷𝑒𝑝𝑜𝑠𝑖𝑡𝑠 − 𝑡𝑜 − 𝑎𝑠𝑠𝑒𝑡𝑠!,!!!+ 𝛽!𝐿𝑜𝑎𝑛𝑠 − 𝑡𝑜 − 𝑎𝑠𝑠𝑒𝑡𝑠_!,!!!+ 𝛽_!𝐿𝑒𝑣𝑒𝑟𝑎𝑔𝑒_!,!!!+ 𝑢_!,!

where t = a quarter.

ΔCoVaR is the dependent variable and the following variables are on the right-hand side: - TradingIncomeShare is the trading income share of total income. This variable is

expected to have a positive coefficient considering that banks with a higher trading income share should contribute more to systemic risk.

- FCIncomeShare is the fee & commission income share of total income. This variable is also expected to have a positive coefficient.

- Log(assets) is the log of total assets (in € millions). Like the previous two variables this one is expected to have a positive coefficient.

(26)

- Deposits-to-assets is deposits divided by assets. This variable is expected to have a negative coefficient, as a higher deposits-to-assets ratio indicates that a bank has a more traditional business model and therefore contributes less to systemic risk. - Loans-to-assets is loans divided by assets. Again this variable is expected to have a

negative coefficient as a higher loans-to-assets ratio also indicates a more traditional business model.

- Leverage is assets divided by book equity. This variable is a proxy for the solvency of a bank and, because higher leverage indicates a lower capacity to withstand shocks, is expected have a positive coefficient.

The second regression is an OLS regression which uses dummy variables for the different bank business models with the business model ‘small international banks’ (category F) as the baseline:

(15) ΔCoVaR!_! _{= 𝛽}

!+ 𝛽!𝐴 + 𝛽!𝐵 + 𝛽!𝐶 + 𝛽!𝐷 + 𝛽!𝐸 + 𝑢!,!

Based on the literature, asset size has a significant effect on a bank’s contribution to systemic risk so the first three dummies are expected to have the largest coefficients. With respect to the sample, graph 5 already showed the average ΔCoVaR values per quarter. Based on the graph, the international diversified lenders, group B, are expected to be the highest

contributors to systemic risk.

5.3 Regression 1: the effect of trading income and fee & commission income on systemic risk

Following the literature, bank and time fixed-effects panel data methodology is used to estimate the regression.10 The dependent variable, ΔCoVaR, is expressed in percentages and most of the independent variables are ratios. The Levin-Lin-Chu unit root test, using the AIC method to determine the number of lags and using the demeaned version of the test, indicates that for all these variables the null hypothesis that the panels contain unit roots can be rejected at the 5%-significance level.11_{For the log of assets variable a trend is included in the test,}

which also results in a rejection of the null hypothesis (at a 1%-level). Therefore the

10_{A Hausman test was done to make sure this kind of model fits the data (instead of a random effects model).} 11_{For all variables except the loans-to-assets ratio the null hypothesis can also be rejected at the 1%-level.}

(27)

regression is done under the assumption that no variables contain a unit root. Table 5 shows the regression output for the estimation of the regression model in equation (14):

Dependent variable: ΔCoVaR

Variables (1) (2) Log(Assets) 0.0528*** (0.0092) Deposits-to-assets -0.2016* (0.1362) Leverage -0.0004 (0.0028) Loans-to-assets 0.2600** (0.1022) Fee & commission

income share

0.3616** (0.1627)

0.3359** (0.1319) Trading income share 0.1972**

(0.0787) 0.1504** (0.0576) Constant 0.1746** (0.0655) -0.4333** (0.1888) Bank fixed effects

Time fixed effects Adjusted R! Yes Yes 0.1255 Yes Yes 0.4966 N 1350 1350

Table 5. Regression output for the contribution of a higher trading income and fee & commission output on systemic risk for the period 2006Q2-2017Q2. Standard errors are clustered at the bank level. *,**,*** denote significance at the 10%, 5% and 1% level, respectively.

Column (2) shows that the trading income share and the fee & commission income share are statistically significant at the 5%-level. Their positive coefficients denote a positive

contribution to systemic risk. Fee & commission income has the largest effect: an increase of the fee & commission income share by 1%-point corresponds to an increase in the systemic risk contribution by 0.3359%. For the trading income share this is 0.1504%. These findings are in line with the findings of Brunnermeier et al. (2012) and Van Oordt & Zhou (2014). The

(28)

variable for asset size is highly statistically significant, which is also in line with the literature (e.g. Van Oordt & Zhou 2014, Laeven et al. (2016) and López-Espinosa et al. (2012)).

Finally, the deposits-to-assets ratio is significant at the 10%-level with the expected negative coefficient.

What is not in line with other empirical studies are the signs of the coefficients of loans-to-assets and leverage and the non-significance of leverage. Loans-to-assets is

statistically significant at the 5%-level with a positive coefficient of 0.2393. This suggests that an increase in the loans-to-assets ratio of 1%-point increases a bank’s contribution to systemic risk by 0.2600%. According to theory, a higher loans-to-assets ratio means a more traditional bank focused on lending and deposit taking. Banks with a more traditional business model are expected to have a lower contribution to systemic risk because these kind of activities are more isolated from risk in the financial system (Van Oordt & Zhou 2014). However, research by Wosser (2017) and Van Oordt & Zhou (2014) shows that non-performing loans have a significant effect on systemic risk. Since there is no variable for non-performing loans in the regression, the effect of non-performing loans on systemic risk is most likely captured by the loans-to-assets ratio. Considering the difficulties with non-performing loans in Europe the past few years, it is reasonable to assume that the loans-to-assets ratio could indeed have a positive relationship with systemic risk and therefore that there is no issue with the data.

Additionally, according to Lucas et al. (2017), fee-focused banks have a high loans-to-assets ratio. Also included in this category are the ‘weak’ banks that do not generate much income from traditional banking activities. Besides the issue of non-performing loans it is also possible that the positive sign is in part due to banks in this category in particular having a high loans-to-assets ratio.

Leverage also deviates from what was expected based on the literature. The coefficient of leverage is -0.0004, which means that a 1%- point increase in leverage decreases systemic risk by 0.0004%. This result indicates that a higher leveraged bank contributes less to

systemic risk. This goes against economic theory; banks that are well capitalized have a higher buffer in case of losses. It also goes against the empirical literature, for example, leverage is positively related to systemic risk and highly significant in the studies done by Van Oordt & Zhou (2014) and Brunnermeier et al. (2012). It is unclear whether this result is due to omitted variable bias, due to the characteristics of the sample or some other issue. In section 5.5, using ΔCoVaR estimates in €, the coefficient of leverage is positive, so omitted variable bias can probably be ruled out.

(29)

The fact that the variable is non-significant, whatever the sign of the coefficient, combined with the significance of the loans-to-assets ratio is a possible indicator that an unweighted leverage ratio requirement is not as effective in limiting systemic risk compared to Risk Weighted Assets (RWA) based capital requirements. If the loans-to-assets ratio, or presumably non-performing loans, has such a significant effect on systemic risk then banks with a high number of non-performing loans are much riskier than banks with a low number of non-performing loans and the leverage ratio does not reflect this. The RWA approach to capital requirements, in this case, would be preferable (assuming risks are accurately reflected in the risk weights).

5.4 Regression 2: the contribution of the different bank business models to systemic risk

Table 6 shows the results of the regression using the dummy variables for the bank business models. Time-fixed effects are included and standard errors are clustered at the bank level.

Dependent variable: ΔCoVaR Variables

Large universal banks (A) 0.1788*** (0.0709) International diversified lenders (B) 0.2461*** (0.0665) Fee-focused banks (C) 0.1715** (0.0854) Domestic diversified lenders (D) 0.0843 (0.0878) Domestic retail lenders (E) -0.1483***

(-0.0633) Constant 0.2081*** (0.0577) Time-fixed effects R! Yes 0.5590 N 1380

(30)

Table 6. Regression output for the contribution of bank business models to systemic risk for the period 2006Q2-2017Q2, with small international banks as the baseline. Standard errors are clustered at the bank level.*,**,*** denote significance at the 10%, 5% and 1% level, respectively.

With the dummy for small international banks as the baseline, this regression shows the extent to which the intercepts of the other groups deviate from this particular group. The dummies for the large universal banks, the international diversified banks and the domestic retail lenders are statistically significant at the 1%-level, and the dummy for fee-focused banks statistically significant at the 5%-level. This shows that the intercepts of these bank business models differ significantly from the intercept of the small international banks.

The values of the coefficients for the dummies correspond to the average ΔCoVaR values per quarter shown in graph 5. The dummy for the international diversified banks has the highest coefficient, suggesting that this kind of business model contributes the most to systemic risk. The fee-focused banks and large universal banks are, respectively, the second and third largest contributors although their coefficients differ very little. The domestic diversified lenders come in fourth, but their contribution to systemic risk does not differ significantly from the small international banks, the fifth largest contributor. Finally, the domestic retail lenders contribute the least to systemic risk.

The fact that the large universal banks are not the highest contributor to systemic risk could be due to better income diversification. Table 5 shows that for the large universal banks, 82.8% of total income consists of net interest, fee & commission and trading income, compared to 92.3% of the international diversified banks. For the fee-focused banks, C, this is 94.7%, indicating that better income diversification is not the reason for a lower contribution to systemic risk for this group compared to B, the international diversified banks. The

ΔCoVaR estimates for group C are less volatile than group B, especially during the financial crisis. However, it is unclear why this is the case. A possible explanation could be the difference in leverage, which was very large during that period. However leverage is not a significant contributor to systemic risk according to the previous regression.

5.5 Robustness checks

Following the literature, robustness checks are done in order to gauge the soundness of the results.

(31)

Dependent variable ΔCoVaR ΔCoVaR in € Variables (1) (2) Log(Assets) 0.0522*** (0.0088) 0.0467*** (0.0094) Deposits-to-assets -0.2116 (0.1285) -0.1952 (0.1195) Leverage -0.0004 (0.0028) 0.0002 (0.0006) Loans-to-assets 0.2592** (0.1035) 0.2864** (0.1168) Fee & commission income share 0.3359**

(0.1275)

0.2116** (0.0989)

Trading income share 0.1496***

(0.0545) 0.0886** (0.0467) Constant -0.4244** (0.1849) -0.3576** (0.1578) Bank fixed effects

Time fixed effects Adjusted R! Yes No 0.4847 Yes Yes 0.4448 N 1350 1350

Table 7. Regression output for the contribution of bank characteristics to systemic risk for the period 2006Q2-2017Q2. In column (1) there are no time-fixed effects and in column (2) the CoVaR in euro is used as the dependent variable. Standard errors are clustered at the bank level. *,**,*** denote significance at the 10%, 5% and 1% level, respectively.

Column (1) reports the results of the regression without time-fixed effects. Though previous research on bank characteristics and systemic risk over a longer time period usually includes both bank-fixed and time-fixed effects (see e.g. Karimalis & Nomikos (2014) and López-Espinosa et al. (2012)), including time-fixed effects results in such a high amount of degrees of freedom that an F-test cannot be done. The regression below in column (1) gives a P-value of 0.000 for the F-test.

Column (2) reports the regression output for the ΔCoVaR values in euro. Following López-Espinosa et al. (2012), ΔCoVaR estimates were calculated in a country’s local

(32)

largely the same results, with the biggest change being that the value of the coefficients of the trading and fee & commission income shares are lower. Leverage is positive now, however its statistical significance remains the same.

The following table reports the same checks for the dummy regression:

Dependent variable: ΔCoVaR ΔCoVaR in €

Variables (1) (2)

Large universal banks (A) 0.1788**

(0.0709)

0.1312** (0.0640) International diversified lenders (B) 0.2461***

(0.0664) 0.2019*** (0.0381) Fee-focused banks (C) 0.1715** (0.0840) 0.1242* (0.0645) Domestic diversified lenders (D) 0.0843 (0.0878) 0.0040 (0.0710) Domestic retail lenders (E) -0.1483**

(-0.0623) -0.1670*** (0.0366) Constant 0.2154*** (0.0557) 0.2614*** (0.0302) Time-fixed effects R! No 0.5553 Yes 0.5776 N 1380 1380

Table 8. Regression output for the contribution of bank business models to systemic risk for the period 2006Q2-2017Q2. In column (1) there are no time-fixed effects and in column (2) the CoVaR in euro is used as the dependent variable. Standard errors are clustered at the bank level. *,**,*** denote significance at the 10%, 5% and 1% level, respectively.

Column (1) reports the output of a regression without time-fixed effects. Again the results remain largely unaltered and the P-value of the F-test is 0.000. Column (2) reports the regression output for the ΔCoVaR estimates in euro’s. Just like in column (1) the results remain largely unaltered. Because of a higher constant (representing the intercept for group F), the output looks different from what is reported in the previous section but the intercepts

(33)

are approximately the same, with the exception of groups D and F, which are a little lower and higher, respectively.

6. Discussion

There are two data issues in this thesis: the small sample size of the entire sample and the small sample size of the different groups. The first has consequences for the external validity of the findings in this thesis. The studies that have been done on European banks, using the ΔCoVaR measure of systemic risk, have focused on a sample of around 30-50 large European banks. To my knowledge, this thesis is the first to include smaller banks in the sample,

making it possible to differentiate larger from smaller banks. But, of course, using 30-50 large European banks provides a much better representative sample of the large banks in Europe than using 30 banks to represent practically the entire European banking system. Therefore a larger sample size is needed to validate the results found in this thesis. A footnote to this is that it might be preferable to focus solely on large banks, considering the small contribution of the smaller banks (groups D, E, F) to systemic risk and possible data availability issues.

The findings in this thesis are, for the most part, in line with previous empirical research on systemic risk: the results indicate that asset size, trading income and fee & commission income are significant contributors to systemic risk. However a positive

relationship is also found for the loans-to-assets ratio, which contradicts earlier research. This positive relationship might be due to this variable capturing the effect of non-performing loans. Using a larger sample size, it would be interesting to investigate this relationship more in further research to see whether it holds or if it is just a feature of this particular sample. No study has been done yet on European banks that includes both non-performing loans and the loans-to-assets ratio in the analysis, so including both in the analysis would be a logical next step.

The classification of Lucas et al. (2017) was used to look at the systemic risk contribution of specific bank business models. Comparing the groups in the sample to their description of the business models showed that asset size for groups B, international

diversified lenders, and C, fee-focused banks, is much larger in this thesis then the average in their sample. This means that the real systemic risk contribution of these groups is probably lower than the results in this thesis indicate. Especially for group C, for which assets are typically below €100 billion according to Lucas et al. (2017) but on average almost reach €1 trillion in 2017 in this sample. Therefore it is probably safe to conclude that the large

(34)

universal banks, group A, have a higher systemic risk contribution than the fee-focused banks, unlike what is found in this thesis. Whether the international diversified lenders, group B, would still have a larger systemic risk contribution than the large universal banks with a lower average asset size is unclear. According to Lucas et al. (2017) asset size for this group is typically between €100-€500 billion and in this sample it is between €400-€900 billion. Further research, using a larger sample of only banks from the categories A, B and possibly C, would be helpful to better investigate their respective systemic risk contributions.

Finally, this thesis only uses the ΔCoVaR measure of systemic risk. Most papers on systemic risk employ at least one other measure as a robustness check. Unfortunately due to time constraints it was not possible to calculate and use another measure. It is possible that using another systemic risk measure yields different results.12

7. Conclusion

This thesis has tried to answer two, related, research questions regarding systemic risk in Europe. The first is whether banks with a higher fee & commission income or a higher trading income contribute more to systemic risk in Europe. The findings in this thesis indicate that both trading income and fee & commission income have a positive relationship with systemic risk. This corroborates the research done on bank characteristics and systemic risk in the US. The analysis on bank characteristics also included other variables such as the loans-to-assets ratio, which, surprisingly, was found to also have a positive relationship with systemic risk. The loans-to-assets ratio is usually used as an indicator of the traditionality of a bank’s balance sheet. It is expected to have a negative relationship with systemic risk due to the expectation that banks with a traditional business model, focused on lending and deposit taking, are less connected to other banks in the financial system. However one of the key risks facing Europe is heightened levels of non-performing loans and it is possible that the loans-to-assets ratio, in this case, captured some of that risk. Further research needs to be done to confirm whether this is true, for example by including both the loans-to-assets ratio and non-performing loans in the analysis.

Another surprising result is that leverage is not a statistically significant contributor to systemic risk. This somewhat contradicts earlier research on bank characteristics and systemic

12_{However it should be noted that other systemic risk measures do not necessarily measure the same relationship}

between a financial institution and systemic risk as the ΔCoVaR. This is the case for the other often used measures of systemic risk, MES and SRISK, as is discussed in section 2.3.

(35)

risk. However, if this is indeed the case then imposing limits for leverage will not be very effective in limiting a bank’s systemic risk contribution. Because of the positive relationship between the loans-to-assets ratio and systemic risk, RWA-based capital requirements would do a better job. This is not just because leverage is not significant but also because banks with a higher amount of non-performing loans should have higher capital requirements and this would not be reflected in a leverage ratio requirement.

The second question is which bank business models contribute the most to systemic risk in Europe. Using the classification of European banks into business models by Lucas et al. (2017) the following bank business models were examined in this thesis: large universal banks, international diversified lenders, fee-focused banks, domestic diversified lenders, domestic retail lenders and small international banks. The highest contributor was found to be the international diversified banks, with the large universal banks and the fee-focused banks coming in second and third, respectively. The last three business models, with a much lower relative asset size compared to the first three, also have a much lower relative systemic risk contribution. A possible explanation for the large universal banks not contributing the most to systemic risk, despite having the largest asset size, is that better income diversification makes them relatively safer than the international diversified banks.

(36)

8. References

Acharya, V. V., Pedersen, L. H., Philippon, T., & Richardson, M. (2017). Measuring systemic risk. The Review of Financial Studies, 30(1), 2-47.

Adrian, T., and Brunnermeier, M. K. (2016). CoVaR. The American Economic Review, 106(7), 1705-1741.

Albert, S., & Alexandre, H. (2017). Banks' earnings: Empirical evidence of the influence of economic and financial market factors. Review of Financial Economics.

Benoit, S., Colletaz, G., Hurlin, C., & Pérignon, C. (2013). A theoretical and empirical comparison of systemic risk measures. Available at: https://halshs.archives-ouvertes.fr/halshs-00746272/file/Systemic_Risk_June_2013.pdf

Brownlees, C. T., & Engle, R. F. (2012). Volatility, correlation and tails for systemic risk measurement. Available at SSRN 1611229.

Brunnermeier, M. K., Dong, G. N., & Palia, D. (2012). Banks’ non-interest income and systemic risk. Available at:

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1786738

Caruana, J. (2010). Systemic risk: how to deal with it. Bank for International Settle- ments. Available at: http://www.bis.org/publ/othp08.htm

ECB. (2016). Financial Stability Review, May 2016. Available at:

https://www.ecb.europa.eu/pub/pdf/other/sfcfinancialstabilityreview201605.en.pdf

Gerlach, S. (2009). Defining and measuring systemic risk. ECB. Available at:

http://www.europarl.europa.eu/document/activities/cont/200911/20091124ATT65154/ 20091124ATT65154EN.pdf

Karimalis, E. N., & Nomikos, N. K. (2017). Measuring systemic risk in the European banking sector: A Copula CoVaR approach. The European Journal of Finance, 1-38.

(37)

Laeven, M. L., Ratnovski, L., & Tong, H. (2014). Bank size and systemic risk (No. 14). International Monetary Fund.

López-Espinosa, G., Moreno, A., Rubia, A., & Valderrama, L. (2012). Short-term wholesale funding and systemic risk: A global CoVaR approach. Journal of Banking & Finance,

36(12), 3150-3162.

Lucas, A., Schaumburg, J., & Schwaab, B. (2017). Bank business models at zero interest rates. Available at:

https://www.ecb.europa.eu//pub/pdf/scpwps/ecb.wp2084.en.pdf

Van Oordt, M. R., & Zhou, C. (2014). Systemic risk and bank business models. DNB Working Paper. Available at:

https://www.dnb.nl/en/binaries/Working%20Paper%20442_tcm47-313480.pdf Wosser, M. (2017). What drives systemic risk in Europe: the balance sheet effect. Central

Bank of Ireland. Available at:

https://www.centralbank.ie/docs/default- source/publications/research-technical-papers/08rt17---what-drives-systemic-bank-risk-in-europe.pdf

(38)

Appendix A: Comparison countries in the dataset used in this thesis and the

dataset used by Lucas et al. (2017)

Countries Banks % Norway 5 16⅔ Italy 3 10 Austria 2 6 ⅔ Germany 2 6 ⅔ Spain 2 6 ⅔ Sweden 2 6 ⅔ Liechtenstein 2 6 ⅔ Romania 1 3 ⅓ Poland 1 3 ⅓ Ireland 1 3 ⅓ UK 1 3 ⅓ Netherlands 1 3 ⅓ France 1 3 ⅓ Finland 1 3 ⅓ Portugal 1 3 ⅓ Switzerland 1 3 ⅓ Greece 1 3 ⅓ Belgium 1 3 ⅓ Croatia 1 3 ⅓ Total 30 100 Countries Banks % Norway 25 12 Germany 23 11 UK 14 6.7 Italy 14 6.7 Spain 11 5.3 Austria 11 5.3 France 10 4.8 Switzerland 9 4.3 Netherlands 7 3.4 Croatia 6 2.9 Portugal 6 2.9 Russia 5 2.4 Turkey 5 2.4 Sweden 5 2.4 Macedonia 4 1.9 Greece 4 1.9 Czech Republic 4 1.9 Slovakia 4 1.9 Slovenia 3 1.4 Latvia 3 1.4 Poland 3 1.4 Finland 3 1.4 Belgium 3 1.4 Ireland 3 1.4 Malta 2 1 Liechtenstein 2 1 Cyprus 2 1 Denmark 2 1 Estonia 2 1 Romania 2 1 Hungary 2 1 Iceland 2 1 Lithuania 1 0.5 Georgia 1 0.5 Bulgaria 1 0.5 Faroe Islands 1 0.5 Bosnia & 1 0.5

Breakdown countries/banks this thesis:

Breakdown countries/banks Lucas et al. (2017):

(39)

Table 9. Breakdown countries/banks this thesis vs. Lucas et al. (2017) Herzegovina

Luxembourg 1 0.5

Serbia 1 0.5

Bank business models and systemic risk in Europe