Sovereign credit risk and market discipline in the financial sector

(1)

Sovereign Credit Risk and Market Discipline in the Financial

Sector

Master thesis - by: Melle J. Meertens

July 7, 2014

University of Amsterdam, Amsterdam Business School

MSc Business Economics - Finance track

Thesis supervisor: Rafael Almeida da Matta

Abstract

This thesis empirically shows that sovereign credit risk increases market discipline exerted by bank creditors. The interaction between sovereign credit risk and bank capital ratios, reinforces the sen-sitivity of funding costs to capital ratios. However, this eﬀect could also indicate better capitalized banks to be less prone to sovereign risk spillovers. Both explanations are not mutually exclusive. The suggested increase in market discipline is presumably caused by values of implicit and explicit gov-ernment guarantees deteriorating in value, in reaction to higher sovereign risk. It is also found that market discipline based on withdrawal of funds, is not reinforced. Some explanations are proposed for this unexpected result, which do not contradict the ﬁndings regarding funding cost based discipline per se.

(2)

Introduction

In recent years, sovereign credit risk has found its place on the financial agenda. Not only the Euro crisis, but also the recent example of imminent Argentinian sovereign default, reminds us that sovereigns are just another type of entities borrowing money. This thesis tries to add up to knowledge about the real effects of sovereign credit risk. A recent stand in literature has explicated the spillover effects of sovereign credit risk on the financial sector. Another important stand in literature has indicated that government guarantees to the financial sector lead to moral hazard. Moral hazard arises because market discipline exerted by bank creditors, is undermined by these guarantees. Aggregating these two stands, a potential amplifying effect on sovereign risk becomes clear. Due to lower likeliness of government support, increased risk sensitivity of bank creditors may reinforce the sovereign risk spillover.

The specific question to be answered, is whether sovereign credit risk affects market discipline. The channel through which is in fact less important. Therefore the effect is directly measured, without quan-tifying values of government guarantees. The results could constitute new knowledge to academics, as to our knowledge the effect was not studied before. The results are even more relevant to regulators, as it is (again) emphasized that decisions influencing the sovereign debt position should be viewed with caution. To the same regulators however, the findings also tell that regulatory attention often emerges in situations where it is not needed, as the market does the work.

The dataset consists 6524 banks from 20 countries, over the period 2008-2013. The implicit cost of debt of banks, and changes in term deposits are the dependent variables used to gauge market discipline. These variables are separately regressed on a bank risk variable, sovereign CDS spreads, and the interaction between the two. The interaction and the bank risk coefficient are expected to have the same sign. This is interpreted as sovereign credit risk having a reinforcing effect on the sensitivity of debt costs to bank specific risk. In other words, market discipline is reinforced by sovereign credit risk. For implicit cost of debt as the dependent variable, the results are affirmative. The effect persists after inclusion of 2-way fixed effects plus an extended set of control variables. The effect is robust against using different underlying maturities of sovereign CDS spreads. The effect is also largely robust against usage of different proxies for bank default risk, and against lags in all control variables. Two of the risk proxies used are capitalization ratios. Accordingly, an alternative explanation for the observed interaction coefficients, is that higher capitalization ratios mitigate the spillover of sovereign risk. However, this alternative explanation does not exclude the original hypothesized effect. The one-way effect of sovereign risk on market discipline, is verified in an additional test.

(5)

As empirical research was performed on a completely new issue, no direct comparison can be made with existing studies. The results are strongly in line with literature on the spillover of sovereign risk to the financial sector (such as Acharya, Drechsler & Schnabl, 2011). The results are also in line with the sizable stand in literature explaining the adverse effects of guarantees to the financial sector. Nier & Baumann (2006) can additionally be mentioned as a strong empirical fundamental, as they explicated the key factors determining market discipline. Other important fundamentals on which the tested hypothesis was built are market discipline studies by Cubillas, Fernandez-Alvarez & González (2013) and Martinez-Peria & Schmukler (2001).

Besides the effect studied being largely novel, this thesis offers some important theoretical contributions. One is that domestic sovereign CDS spreads appear to be significant and consistent negative predictors of changes in term deposits. This effect is well-founded in literature but to our knowledge, rarely tested specifically. Another contribution, is the suggestion that sovereign CDS spreads with 5-year underlying maturity are the better predictors of bank funding conditions, compared to 1-year CDS. This could be indicative of liquidity being concentrated around 5-year CDS contracts, as is the case with corporate CDS. A third contribution is that regulatory capital ratios are used as (inverse) risk proxies. Regulatory capital ratios have risk-weighted assets in their denominator, making them more sensitive to asset risk than the regular equity ratio. Lastly, a highly recent sample is used. The financial sector is highly complex and continuously evolving, not in the last place due to ever-changing regulation. It is therefore necessary to keep studying the same effects over time, as important trends can be shed light upon doing so.

Section I provides a conceptual background, which takes the place of a theoretical model explaining the underlying mechanisms of the empirics. It also highlights all relevant concepts used throughout the thesis, and ends with the formulation of the main hypotheses. Section II presents the dataset and the methodology used to test the hypotheses, as well as a clear elaboration on all variables used in the regressions. Section III presents and discusses the results, and concludes with theoretical and practical implications. Section IV encompasses some additional tests, and section V sums up the conclusions.

(6)

I Conceptual background

The effect of sovereign credit risk on market discipline in the financial sector has no precedent in theoretical literature. For this reason it is essential that it is first explicated why market discipline would be influenced by sovereign credit risk, which is the purpose of this section. All relevant concepts are presented and their roles in relation to the thesis subject are highlighted.

I.A Determinants of bank default risk

Banks control their level of default risk in three ways, according to Froot & Stein (1998). First, banks influence their amount of asset risk by hedging risk exposures. This is called risk management. In reality, not all risks are hedgeable. As a result, default risk can only be influenced by risk management to a certain extent. Second, banks can alter their investment in securities carrying non-hedgeable risks. If a bank wants to decrease its asset risk while this is not possible by risk management, it can decrease its investment in assets carrying unhedgeable risk. Third, the management of a bank can alter leverage. Higher leverage indicates higher default risk. This is illustrated best by imagining a firm without leverage. No financial obligations implies no default on debt when a negative outcome occurs on the asset side (e.g. a drop in earnings due to a default of some debtors). The more levered the firm, the higher the chance a drop in earnings translates in to a default on debt. The same logic applies to banks.

Leverage influences default risk in a second way. It does not just amplify adverse shocks to the balance sheet, but it also influences the bank’s chosen level of asset risk. When leverage is high, it will be the debtholders who bear the losses from additional non-performing investments. The residual claim the owners have, already reached its bottom line anyway (as the operational proceeds are flowing to the debtholders). Owners can therefore “gamble” on upside outcomes by increasing asset risk, using creditor’s money. This agency problem between owners and debtholders is called the asset substitution problem (Jensen & Meckling, 1976). Because of the high leverage ratios banks and other financial institutions usually have, the asset substitution problem especially applies to these types of firms.

The asset substitution problem means a tendency in banks to increase risk at the expense of debtholders (throughout the thesis, for simplicity it is assumed that shareholder’s and manager’s interests are perfectly aligned). The logic of the asset substitution problem also applies to high leverage resulting from low implicit market values of equity. This implicit market value is referred to as bank charter value. Low charter value means a higher tendency to bank owners to gamble on upside outcomes (Hellmann et al. 2000).

(7)

I.B Determinants of market discipline

In a perfect capital market, the asymmetry of interests between banks and debtholders as described above, straightens itself. The reason for this, is that creditors expect to be compensated for increased bank risk by increased default risk premia on the bank’s debt. Doing so, credit markets restrain the risky activity of banks as elaborated on in the previous subsection. If banks themselves bear the costs of increased risk taking, they are unable to pass on default risk to creditors. Although this disciplining mechanism is an important force controlling bank risk (e.g. Flannery & Sorescu, 1996), it does not work perfectly in reality. As a result, today regulators have put in place a wide array of measures to deal with socially undesirable levels of bank risk.

The notion of credit markets controlling bank risk taking, is called market discipline. It can not always be said if a risk increase is due to a conscious decision by the bank. For this reason, in the remainder of the thesis the term “market discipline” is used for the reaction of credit markets to bank risk in general. Market discipline translates both in to higher borrowing rates and withdrawal of deposits, in reaction to changes in bank risk (Martinez-Peria & Schmukler, 2001). Creditors monitor bank risk mainly by looking at superﬁcial risk measures such as return on assets, credit ratings and leverage. However, there are indications that they also look at the asset mix of banks (Morgan & Stiroh, 2001).

The reason market discipline does not function perfectly in reality, is that its effectiveness depends on several factors. The three most important are explicated by Nier & Baumann (2006). The first is the visibility of the bank’s risk choice. If the asset mix of a bank is not fully disclosed, creditors are unable to sufficiently monitor the level of asset risk the bank has chosen for. Accordingly, credit markets will be less able to discipline the bank appropriately by increased rates or withdrawing deposits. Unfortunately, the asset mixes of banks and other financial institutions are regularly to a large extent unobservable (Morgan, 2002).

The second factor determining the effectiveness of market discipline according to Nier & Baumann (2006), is the ratio of uninsured deposits to total liabilities. In most countries, specific classes of bank deposits are explicitly insured by a specialized institution. An important reason for this is to avoid bank runs when a drop in depositor confidence occurs (Diamond & Dybvig, 1983).1_{If depositors know they will}

be reimbursed anyway when a bank fails, they won’t engage in a bank run. The maximum insured amount per depositors strongly varies across countries. The negative consequence is that the incentive to depositors

1_{Banks have a “liquidity mismatch” between assets and liabilities. Many of their assets are long-term ﬁxed investments,}

while many of their liabilities can be withdrawn overnight. In a world without deposit insurance, this would cause the banking sector to rely completely on depositor’s conﬁdence of getting disbursed when they wish to.

(8)

to monitor the risk choices of the bank also decreases, if their reimbursement is no longer at stake (as in Demirgüc-Kunt & Huizinga, 2004). The higher the ratio of uninsured deposits to total liabilities, the larger the fraction of creditors which actually do have an incentive to monitor the risk taking decisions of the bank. Both the ratio of uninsured liabilities and the country-speciﬁc level of deposit insurance coverage are therefore important determinants of market discipline.

A third determinant of market discipline is the extent of the safety net (Nier & Baumann, 2006). Governments sometimes consider it necessary to intervene in the financial sector by bailing out a financial institution or providing extensive financial aid (Allen et al. 2011). Such interventions are to avoid the detrimental effects on the real economy bank failures can have, through contagion and liquidity dry-up (e.g. Bernanke, 1983 & Diamond & Rajan, 2005). The knowledge that a bailout is a real possibility, means an implicit guarantee to banks. The presence of such a guarantee is especially evident when a bank’s relative importance to the financial system is high (in other words, when the bank is considered “too big to fail”; e.g. Tsesmelidakis & Merton, 2013). In the context of financial crises such as in the early 21th century, governments may even decide to proclaim explicitly stated guarantees on bank debt (Panetta et al. 2011). When any kind of government guarantees are in place, bank liabilities are no longer at risk (or, at least, to a lesser extent). Similar as with deposit insurance, the incentive to creditors to monitor and react to default risk gets distorted.2

I.C Sovereign credit risk and default risk of banks

Besides being a determinant of market discipline, a guarantee on bank debt also means a direct reduction of default risk. Tsesmelidakis & Merton (2013) show that discounts in bank funding costs due to being “too big to fail” can be substantial. These discounts indicate the market to perceive default risk on bank debt to be substantially lower, when guarantees are in place. The value of an implicit or explicit guarantee can be thought of as the present value of all these future discounts (as in Lucas & Mcdonald, 2010).3

Because the discounts depend on the market’s expectation of support likeliness, the value of guarantees is positively related to this expectation (Schich & Kim, 2012). When government support becomes less likely, the default risk premium on bank debt increases, reﬂecting lower guarantee values. For example, in recent years Eurozone countries tended to waive regulatory policies involving capital-intensive intervention.

2_{It should also be noted that the existence of deposit insurance and safety nets demonstrates the inability of market}

disci-pline to withhold banks from taking on socially undesirable levels of default risk (Flannery & Sorescu, 1996). With perfectly functioning market discipline, the measures won’t be necessary. Whether these measures are the cause of malfunctioning market discipline or an implication of it, leaves plenty of space for further research.

3_{Correspondingly, the value of the implicit guarantees can be thought of as the diﬀerential in the market value of bank}

(9)

Because of the resulting reduction in intervention likeliness, the value of the guarantees decreased. However, the main determinant of support likeliness is the financial capacity of the sovereign. In recent years, financial constraints of sovereigns increasingly affected financial sector risk through this channel (Panetta et al. 2011). Acharya, Drechsler & Schnabl (2011) explain this risk spillover nicely in the context of bailouts. When a bank bailout is considered necessary, a government can fund the bailout by either raising taxes or issuing debt. Raising taxes is often impossible, so new sovereign debt is issued. This in turn leads to an increase in sovereign credit risk, as more debt increases the financial obligations of the government, while income does not change. Spreads on new sovereign debt increase, reflecting the increase in sovereign credit risk. The increased spreads on sovereign debt make additional intervention more expensive. Implicit and explicit guarantees to the banking sector become less valuable as a result, increasing default risk of banks. One should notice that the same logic applies when no bailout is done, i.e. sovereign credit risk has increased due to “normal” causes (such as a “regular” increase in the public deficit; public deficits are a main determinant of sovereign credit risk, Aizenman, Hutchinson, & Jinjarak, 2013).

Sovereign risk spilling over to the financial sector by guarantees declining in value, is a finding well represented in literature. Correa et al. (2014) show empirically that sovereign credit rating downgrades have a negative effect on bank stock returns. The effect is stronger for banks receiving a higher Fitch support probability rating (A rating assigned to banks by rating agency Fitch, which is indicative of the likeliness of the bank to receive financial support from the government). This is indicative for guarantees in place being affected by a deterioration of the sovereign credit position. Demirgüc-Kunt & Huizinga (2013) find a similar relationship between the market value of banks and the country’s fiscal balance. Whelan (2010) explains how during the Irish sovereign debt crisis of 2010, doubts about the solvency of the sovereign undermined the credibility of explicit guarantees. Less credibility means less intervention likeliness, implying lower guarantee values. Estrella & Schich (2012) model the value of guarantees to banks as a function of the value of sovereign debt and bank debt, respectively. In their model, the value of guarantees is positively related to the value of sovereign debt, while it is negatively related to the value of bank debt.4

The value of guarantees is not the only channel through which sovereign credit risk spills over to the ﬁnancial sector. The spillover also ﬂows through (I) the value sovereign debt on the balance sheets of banks; (II) a decrease in collateral capacity of banks, leading to funding constraints, increasing default

4_{This is the case because a guarantee has the most value to banks of which liabilities are the most risky when guarantees}

are absent: recall that the value of guarantees depends on the market-perceived likeliness of intervention (Estrella & Schich, 2012).

(10)

risk of the bank even further; and (III) a downgrade of the domestic sovereign credit rating, translating in lower ratings for the banks in the country in question, in turn leading to funding constraints (Angeloni & Wolff, 2012). The increase in bank default risk resulting from the risk spillover, is reflected in the funding costs of banks (Panetta et al. 2011). Empirical evidence is provided by Albertazzi, Ropele & Sene (2014) which show spreads paid by Italian banks to be positively affected by sovereign debt spreads. Acharya et al. (2013) indicate that non-financial firm’s costs of borrowing increase in reaction to sovereign credit risk. Banks may increase their lending rents to maintain their profitability targets, in reaction to sovereign credit risk inflating their costs of funding.

The risk spillover also means that it is becomes less attractive for creditors to provide funding to banks in countries being “at risk”. Gennaioli, Martin & Rossi (2014) provide a model where banks under ﬁnancially strong sovereigns more easily attract funding. Correspondingly, Panetta et al. (2011) explain that banks in Portugal and Greece had substantial diﬃculties in raising wholesale debt during the recent sovereign debt crises in those countries.

I.D Sovereign credit risk and market discipline

Recall that the presence of guarantees is an important determinant of credit markets disciplining bank risk (Nier & Baumann, 2006, as explicated in section I.B). In addition, in the previous subsection it was explained how sovereign credit risk influences the value of implicit and explicit guarantees. It follows naturally that market discipline is likely to be influenced by sovereign credit risk. Guarantees may drop in value, due to government support becoming less likely. As a result, bank creditors will regain their incentive to engage in monitoring, as well as to react to changes in bank risk. Correspondingly, the level of market discipline increases. This effect is in addition to the risk spillover from sovereigns to the financial sector. The increase in funding costs associated with the risk spillover, could therefore be amplified by the parallel increase in market discipline. I.e. the increase in funding costs will be larger if market discipline is higher. This is the case for market discipline “exerted” by increased rates, as well as by withdrawal of deposits. In the latter case it is reduced funding supply leading to higher funding costs.

When a bailout is the cause of an increase in sovereign credit risk (as in Acharya, Drechsler & Schnabl, 2011), an associated increase in market discipline could be a negative externality to banks, due to the amplifying effect. A bank may be saved at first, but the sovereign credit position has worsened. Increased sovereign risk spills over to banks. The resulting adverse effects of increased funding costs, are amplified by increased market discipline.

(11)

The increase in market discipline can also be considered a positive externality when the society as a whole is considered. When a bank is taken over by the government or it received an explicit rescue package of some sort, excessive risk taking by the bank will be considered increasingly undesirable. Increased monitoring and disciplining by creditors can therefore be welcome as such. The same applies to banks in the general context of sovereign debt crises.

It should be noted that market discipline is also be reﬂected by decreased funding costs: less risky banks will attract more creditors, leading to lower funding costs. Accordingly, market discipline can be measured by the sensitivity of funding costs to risk (also see section II.B.1).

The relationship between sovereign credit risk and market discipline is - to our knowledge - not modeled before and is also not yet tested empirically. Only Cubillas, Fernandez-Alvarez & González (2013) test whether market discipline is influenced by the existence of a domestic public deficit. They find a negative and significant coefficient on the triple interaction between risk and size of banks plus the presence of a public deficit. This indicates budget deficits to have a diminishing effect on the difference in observed market discipline between big and small banks.5_{This finding is presented as evidence of the}

“too-big-to-save” hypothesis postponed by Demirgüc-Kunt & Huizinga (2013) and Völz & Wedow (2011). This hypothesis states that some banks have become too large to have a credible claim on the safety net. According to Cubillas, Fernandez-Alvarez & González, a public deﬁcit brings the level of market discipline exerted upon big banks, back to the level for small banks, due to the “too-big-to-save”-eﬀect.

However, Cubillas, Fernandez-Alvarez & González (2013) do not find a significant coefficient the inter-action between risk and the presence of a public deficit (i.e. without the size variable). They interpret this as a deficit not being influential to market discipline, when a bank has insufficient size to hold a credible claim on the safety net. Recall, however, that bailouts often occur in the context of systemic crises. Dur-ing a systemic crisis, many banks may need to be bailed out to avoid aggravation of the crisis, includDur-ing small ones (Acharya & Yorulmazer, 2007). Even if the scope of such a crisis is not system-wide, failure of a mediocre-sized bank can lead to contagion, in turn instigating a systemic crisis. As a result, implicit guarantees have a wider application than to banks deemed too-big-to-fail only. The conclusions of Cu-billas, Fernandez-Alvarez & González regarding the effect of a public deficit on general market discipline, are therefore contested. The most important takeaway from Cubillas, Fernandez-Alvarez & González is that a deficit is indeed somehow influential to the relationship between bank risk and funding costs. This knowledge is a good starting point in further exploring the relationship between sovereign credit risk and

5_{The main ﬁnding of Cubillas, Fernandez-Alvarez & González (2013) is that lower market discipline is observed in big}

(12)

market discipline.

As Forssbaeck (2011) justly mentions, it is the expected reimbursement that matters to creditors, when they “set” their level of market discipline. This makes the difference between implicit and explicit government guarantees less important. Implicit guarantees may be directly affected by sovereign credit risk, but the credibility of explicit guarantees could also be undermined (as in Whelan, 2010). Note that proclamation of explicit guarantee schemes often occurs in the context of system-wide financial crises. The scope of such a crisis often makes reimbursement of all explicitly guaranteed debt questionable (as in Allen et al. 2011).

Even for private deposit insurance institutions, sovereign credit risk could be of influence. Sovereign credit risk may affect the balance sheets of these institutions, undermining credibility of the private in-surance schemes. However, the goal of deposit inin-surance institutions is to provide coverage for all risk regarding bank deposits; the influence of sovereign debt problems should be far less influential as a result.

I.E Hypothesis development

To summarize, sovereign credit risk has two important effects on the financial sector. The first is the direct spillover of sovereign credit risk on default risk of banks. The second is a hypothesized effect on market discipline: bank creditors increase their monitoring and reaction to bank risk, when government guarantees on bank liabilities drop in value. The increase in default risk premia on bank debt due to the risk spillover, may be amplified due to this increased market discipline. The following main hypothesis can be formulated:

H1. Domestic sovereign credit risk positively affects market discipline in the financial sector Market discipline is reflected both by increased costs on bank debt, and by depositors withdrawing in reaction to unfavorable levels of bank risk (Martinez-Peria & Schmukler, 2001). Correspondingly, H1 can be split in to two sub-hypotheses, which are more specifically testable than H1:

H1.1: Domestic sovereign credit risk reinforces the sensitivity of a bank’s implicit cost of debt

to default risk of the bank

H1.2: Domestic sovereign credit risk reinforces the sensitivity of increases in term deposits at

a bank to default risk of the bank

In the methodology section, risk proxies are used which resemble inverse bank risk measures. Therefore one should state H1.1 and H1.2 diﬀerently if the speciﬁc proxy variables used were included, instead of the

(13)

general notion of “bank default risk”. Also note that the preliminary tests performed (see section III.A) are not meant to confirm or reject any hypothesis, but have methodological reasons (i.e. showing the main effects of sovereign credit risk and bank risk to have the expected effect of the dependent variables. These effects generally have a solid foundation in literature, so no hypothesis statement is needed).

(14)

II

Dataset & methodology

In this section, the dataset and methodology are presented which are used in studying the hypothesized eﬀect. Subsection II.A elaborates on the dataset, subsection II.B presents the methodology, and subsection II.C introduces all variables used in the regressions in a structured fashion.

II.A

Data

The dataset consists of an unbalanced annual panel of 6524 banks from 20 countries measured over a period spanning from 2008 to 2013. Banks are included which are under the same holding company, resulting in relatively large numbers of banks per country. The relatively short time period is because sovereign CDS6 _{spreads are used as proxy for sovereign credit risk. Sovereign CDS spreads are only available on}

Datastream from 2008 up until today. Both locally listed and non-listed banks are included in the panel. The choice is made consciously to include a relatively large cross-section of banks, against the loss of the possibility to use certain variables based on market data (i.e. data only available for listed banks). Bank accounting data comes from BvD BankScope through WRDS. Macroeconomic and sovereign CDS data comes from Datastream or the World Bank (WDI). National levels of deposit insurance coverage are hand-collected (this is included as a control variable, see section II.C.5). Hand-collecting is done to assure this data to be as up-to-date as possible. In the variables list included in the appendix, the data source is stated for each variable. Table 1 shows the countries included in the sample, the total observation frequency per country, and the number of banks per country.

As a measure of data richness and quality, the availability of COMPUSTAT bank accounting data is used. Although COMPUSTAT data is not used in the regressions, it is assumed that when domestic bank data from a country is included in COMPUSTAT this is indicative for high country-level data quality and data richness. This procedure led to the relatively limited set of 20 countries.

The US are excluded from the sample because the sample would consist of too many US banks if it was included. This would have biased the results too much towards the population of US banks only.

6_{Credit Default Swap: a credit derivative providing insurance to its buyer against default of the underlying security. A}

(15)

Table 1: Frequencies and relative proportions of observations per country + number of banks per country

The high number of banks per country is due to diﬀerent banks under the same holding company being included in the sample. The strongly varying ratio of banks to observations per country is due to the panel being unbalanced; i.e. not all banks in the sample have data for all six years of the sample period (2008-2013).

Country Observations % of total Banks % of total

Argentina 600 1.99 125 1.92 Australia 526 1.74 121 1.85 Brazil 1042 3.46 234 3.59 Chile 302 1.00 75 1.15 Colombia 529 1.75 115 1.76 Germany 9732 32.28 2082 31.91 Denmark 906 3.01 192 2.94 Spain 1232 4.09 323 4.95 France 2921 9.69 620 9.50 United Kingdom 3284 10.89 698 10.70 Greece 189 0.63 44 0.67 Ireland 302 1.00 79 1.21 Japan 5004 16.60 975 14.94 South Korea 575 1.91 169 2.59 Mexico 697 2.31 177 2.71 the Netherlands 610 2.02 135 2.07 Peru 244 0.81 49 0.75 Singapore 254 0.84 61 0.94 Sweden 734 2.43 150 2.30 South Africa 466 1.55 100 1.53 Total: 30149 100.0 6524 100.0

(16)

II.B

Methodology

II.B.1 Methodology outline & regression model

A method to measure market discipline was ﬁrst to be selected. Cubillas, Fernandez-Alvarez & González (2013), Demirgüc-Kunt & Huizinga (2004) and Martinez-Peria & Schmukler (2001) are followed in how to measure market discipline. These papers gauge market discipline as the sensitivity of the implicit cost of debt of banks, to proxies for bank default risk. Accordingly, market discipline can be measured by regressing the implicit cost of debt on measures of bank risk. Market discipline is also exerted by withdrawal of deposits in reaction to unfavorable levels of bank risk. Sensitivity of withdrawals to default risk is measured by regressing the change in term deposits on proxies of bank risk.

For the specific purpose of measuring the effect of sovereign credit risk on market discipline, sovereign credit risk is added as a covariate. Sovereign credit risk is also interacted with bank risk. The coefficient on this interaction term will enable us to draw conclusions on the effect sovereign credit risk has on market discipline. If the interaction coefficient has the same sign as the population coefficient on bank risk, this indicates a joint increase in sovereign credit risk and bank risk works in the same direction as a stand-alone change in bank risk. This would mean that sovereign credit risk strengthens the effect bank risk has on the dependent variable. In other words, sovereign credit risk strengthens market discipline. The baseline model looks as follows:

DebtCosti,t = β0Riski,t−1+ β1SovCDS(t)i,t+ β2(Riski,t−1∗ SovCDS[t]i,t) +

γ0Xi,t+ γ1Zi,t+ αi+ λt+ ei,t (1)

∆Depositsi,t = β0Riski,t−1+ β1SovCDS(t)i,t+ β2(Riski,t−1∗ SovCDS[t]i,t) +

γ0Xi,t+ γ1Zi,t+ αi+ λt+ ei,t (2)

Equation (1) has the implicit cost of debt as the dependent variable, to measure market discipline by increased funding costs. Funding costs (as measured by the implicit cost of debt) should increase with bank risk. Equation (2) measures market discipline exerted by withdrawing deposits: the dependent variable here is the periodic increase in deposits, which is expected to be negatively aﬀected by bank risk. Note that the expected sign on the bank risk variable depends on the speciﬁc risk proxy used (see section II.C.2). Sovereign credit risk is indicated by SovCDS(t)i,t, because sovereign CDS spreads are used as

(17)

proxy for sovereign credit risk. (t) indicates the underlying maturity of the sovereign CDS spreads used in the specific regression. The coefficient of interest is the interaction coefficient β2 which resembles the

joint eﬀect sovereign CDS and the risk proxy have on the dependent variable. Xi,t is a vector of bank-level

control variables. Zi,t is a vector of country-level control variables. αi constitutes bank ﬁxed eﬀects, and

λtcomprises year ﬁxed eﬀects (also see next section II.B.2).

Because both main effects are also included, the coefficient estimate on the interaction term does not suffer from omitted variable bias due to correlation between bank risk and sovereign credit risk. It is important to note that including the interaction term may cause multicollinearity of the bank risk and sovereign credit risk variables. The reason for this is that the interaction term is inherently correlated with the main effects (Balli & Sørensen, 2013). Presence of multicollinearity could adversely affect the magnitude and significance of the coefficient estimates ˆβ0and ˆβ1in the baseline regressions. The focus in

terms of interpretation will therefore be on the coefficient of the interaction term. However, the sign on the main effects is necessary to know if one desires to know whether the interaction effect is reinforcing or weakening. In order to establish the direction of the main effects, a preliminary test is performed using the same model as in equations (1) and (2), but without the interaction term:

DebtCosti,t= β0Riski,t−1+ β1SovCDS(t)i,t+ γ0Xi,t+ γ1Zi,t+ αi+ λt+ ei,t (3)

∆Depositsi,t = β0Riski,t₋₁+ β1SovCDS(t)i,t+ γ0Xi,t+ γ1Zi,t+ αi+ λt+ ei,t (4)

The model above will additionally provide information on the magnitude of the direct eﬀects of sovereign credit risk on bank funding conditions. However, this relationship is not the main focus of this thesis (as no hypotheses were formulated around it).

Standard errors in all panel regressions are corrected for bank-level autocorrelation of standard errors (clustering). Using clustered standard errors is not problematic as a relatively large cross section of entities (banks) is considered.

II.B.2 Endogeneity issues

Note that the risk proxy is lagged one year in the baseline model (equations [1] and [2]) as well as in the preliminary test model (equations [3] and [4]). This is done to avoid simultaneous causality between the dependent variable and the risk proxy. For example, a risk proxy such as a capitalization ratio, is likely to

(18)

be influenced by a change in deposits in the same period. This makes it difficult to estimate the one-way effect of the capital ratio on deposit withdrawals. Affecting a variable backwards in time is not possible, so including a lag offers a solution to the simultaneity problem. Sovereign credit risk is not lagged, as simultaneity between either one of the dependent variables and sovereign credit risk is less likely.

It must be noted that using lags also introduces the threat of weak results, as annual time intervals are considered. Creditors possibly need much less than a year to adjust their required rates in reaction to increased risk (or to withdraw deposits in reaction to increased risk). Annual time intervals may lead to insigniﬁcant results while there in fact exists a relationship between the variables of interest. Using shorter time intervals, however, is not possible when BankScope data is used (BankScope only provides annual time intervals).

To deal with omitted variable bias, the regression model is tooled with control variables. The aim of these control variables is to make sure that the variables of interest are uncorrelated with the error term. Conditional mean independence is therefore likely to be satisfied, but conditional mean zero not. This means that one should be careful with interpreting coefficient estimates on control variables. Sign and significance can provide some information, however. To further deal with possible estimation bias resulting from omitted variables, bank and year fixed effects are included. Omitted variables which are either invariant over time (bank fixed effects) or over the cross section of banks (year fixed effects) are controlled for by doing so.

II.C

Variables

II.C.1 Dependent variables

The implicit cost of debt is calculated as total interest expense on interest-bearing liabilities, divided by year-average total interest bearing liabilities. Cubillas, Fernandez-Alvarez & González (2013), Demirgüc-Kunt & Huizinga (2004) and Martinez-Peria & Schmukler (2001) are followed in using this speciﬁc measure of bank funding costs.

For calculating the change in deposits, customer term deposits are used, similar as in Martinez-Peria & Schmukler (2001). Using all deposits instead would introduce too much variation in deposits changes which has nothing to do with depositor choices regarding bank risk. Because market discipline is exerted not just by customers but also by other banks, deposits from banks is added up to customer term deposits. After adding up these two deposit classes, the natural logarithm is taken. The ﬁrst diﬀerence of the logarithmically transformed variable is used in the regressions. This variable resembles the relative change

(19)

in customer term plus bank deposits over the last ﬁscal year. All dependent variables are measured in decimals.

II.C.2 Risk proxies

Ideally, the risk proxies used should measure the same variation in bank default risk, as what creditors react to by altering rates or withdrawing deposits. Three diﬀerent proxies for bank default risk are used. This is done in favor of robustness, but also to have a good balance between proxy quality and data availability.

For the ﬁrst risk proxy, direction is provided by Demirgüc-Kunt & Huizinga (2004, 2011 & 2013) and Martinez-Peria & Schmukler (2001), among others. These papers use the equity to assets ratio as a measure for bank risk. The equity to assets ratio, being the inverse of leverage, is an important factor determining default risk of banks (as explained in section I.A). A high equity ratio means a high capital buﬀer, which indicates lower default risk (Nier & Baumann, 2006). For this reason it is, that a higher sensitivity of the implicit cost of debt to the equity ratio indicates higher market discipline (the same applies to sensitivity of changes in deposits to the equity ratio). However, as Morgan & Stiroh (2001) rightly mention, default risk can still be quite high despite relatively low leverage. To deal with this problem, instead of the basic equity to assets ratio, the tier 1 capital ratio is used. This is a measure used by regulators to assess capital adequacy of banks.7 _{The tier 1 capital ratio is calculated as total tier 1 capital over risk-weighted assets.}

Tier 1 capital is calculated as equity plus disclosed reserves plus non-redeemable preferred stock. Risk-weighted assets enjoy a value discount compared to the original value of book assets. This value discount is based on an assumed risk level per asset class: the less risky asset types the bank has invested in, the higher the discount. The higher the discount, the lower the value of risk-weighted assets. Accordingly, the tier 1 capital ratio will become lower when a bank has a greater portion of its assets invested in risky asset classes. However, it must be noted that risk-weighted assets does not perfectly measure asset risk. Banks can set their de facto level of asset risk independently of the assigned level of risk-weighted assets, by making use of risk management (Blundell & Atkinson, 2010). In addition, since the Basel II accord banks themselves can influence the risk weight of each asset (although supervisory ratification is required). This leads to the level of risk-weighted assets often to be set in a way beneficial to the bank (for instance, to control funding costs). Note, however, that we are not looking for the ideal measure of bank risk, but only for a measure reflecting both leverage and asset risk - i.e., a risk measure credit markets act upon,

7_{Capital adequacy can also be thought of as the “buﬀer capacity” banks have, meant to deal with adverse shocks to their}

balance sheets (e.g. Hellmann et al, 2000). Minimum regulatory ratios for capital adequacy where set ﬁrst internationally in the Basel I accord of 1988.

(20)

Table 2: Cross-correlation table of risk proxies used

Non-transformed, winsorized values are used to calculate correlations.

Overlapping obs: 8377 Tier 1 Total regulatory Z-score capital ratio capital ratio

Tier 1 capital ratio 1

Total regulatory capital ratio 0.9341 1

Z-score -0.0221 -0.0012 1

as the aim is to measure market discipline. Another problem with the tier 1 capital ratio is that it has just 9,300 observations over the entire sample. However, 19 of 20 countries are still represented by the sampled data on the tier 1 capital ratio. The other risk proxies both have more observations and have all 20 countries represented.

As a second risk proxy, the total regulatory capital ratio is used which is highly related to the tier 1 capital ratio. It is measured as total regulatory capital over risk-weighted assets. Total regulatory capital includes tier 1 capital, plus certain long-term, low risk debt classes, as well as reserves and hybrid instruments (such as convertibles). An advantage of the total regulatory capital ratio over the tier 1 capital ratio is that it has more observations (15,983).

As a third measure of bank risk the Z-score is used, which is not dependent on risk-weighted assets. A large stand in banking literature is followed doing so, including Boyd, de Nicoló & Jalal (2006), Cubillas, Fonseca & González (2012), and Hadad et al. (2011). The Z-score is in fact a measure of bank solvency; a higher Z-score means lower default risk (similar as with the capital ratio). To be precise, the Z-score measures how many standard deviations return on assets can drop, before book equity reaches a value of 0. More standard deviations means a higher eﬀective equity buﬀer. The Z-score is calculated as

ROA+(equity/assets)

σROA with σROA being the standard deviation of ROA. The standard deviation is regularly

measured as a rolling standard deviation over a speciﬁed time window. Due to the relatively short time series, in this thesis the standard deviation is calculated over the entire time series of data of the bank in question. The Z-score has a relatively high number of observations compared to the other two risk proxies (27,451).

Table 2 shows the correlations between the risk proxies used. Almost no correlation between the capital ratios and the Z-score can be observed, indicating these risk measures to be substantially diﬀerent. The capital ratios are highly correlated, as expected.

(21)

Table 3: Cross-correlation table of sovereign CDS spreads of diﬀerent maturities

Both correlation tables are based on all 30149 observations.

Sovereign CDS spreads Sovereign CDS spreads (log)

Maturity: 1-year 2-year 5-year 10-year Maturity: 1-year 2-year 5-year 10-year

1-year 1 1-year 1

2-year 0.9981 1 2-year 0.9861 1

5-year 0.0513 0.0529 1 5-year 0.5198 0.5516 1

10-year 0.9799 0.9871 0.0683 1 10-year 0.8708 0.9201 0.5878 1

ROA can drop before equity reaches a value of 0 (as previously explained).

II.C.3 Sovereign credit risk

Spreads on sovereign credit default swaps are used as proxies for sovereign credit risk. A high spread means that a high premium must be paid for insurance against default of the underlying security (the underlying security is usually a sovereign bond or another type of sovereign debt). Acharya, Drechsler & Schnabl (2011) as well as Schweikhard & Tsesmelidakis (2012), among others, use sovereign CDS spreads as proxy for sovereign credit risk.

Table 3 shows the sample cross-correlations of sovereign CDS spreads with diﬀerent underlying matu-rities. According to Pan & Singleton (2008), liquidity of sovereign CDS securities is not as concentrated around the 5-year maturity as corporate CDS are. The reason for this, is that sovereign CDS contracts of most underlying maturities all are actively traded. As a result, the choice which underlying CDS maturity to use is less important. However, as can be seen in the cross-correlation table it appears that the spreads of the 5-year maturity CDS are to a lesser extent correlated with the spreads of the other three maturities. This could indicate that liquidity also plays a role to sovereign CDS. Because 1-year and 5-year sovereign CDS spreads are the least correlated, those are both used in the panel regressions in favor of robustness. Daily sovereign CDS spreads were averaged to come at annual data intervals.

Sovereign CDS spreads of each maturity are measured in basis points. In the summary statistics table however, CDS spreads are reported in decimals (see table 4).

II.C.4 Bank-level controls

An important factor inﬂuencing bank risk taking is charter value. The charter value of a bank is equal to discounted value of all its future earnings (e.g. Cordella & Yeyati, 2003). The lower the charter value, the more inclined the owners of the bank are to gamble for upside proﬁt. It is very likely that charter value is correlated with the dependent variables and therefore endogenous. To proxy for charter value, return on

(22)

assets is used.8 _{Controlling for ROA also serves the goal that the inﬂuence of sovereign credit risk can be}

measured regardless of the proﬁtability of the bank.

Absolute size measured as book value of assets is a control variable standard in banking literature and therefore included.

In measuring market discipline, size is especially important. A large bank is more important to the system, leading to higher implicit guarantees. However, when the bank becomes too large, it becomes “too big to save” so the guarantees on its liabilities decline in value (Demirgüc-Kunt & Huizinga, 2013 and Völz & Wedow, 2011). This interplay between size and implicit guarantee value has its eﬀect on market discipline, according to the same papers. Size to GDP is therefore also included as a bank-level control (in addition to absolute size). Including this variable has the secondary beneﬁt to control for market power of the bank. Market power is likely to be correlated with at least two variables of interest (implicit cost of debt and bank risk, Fonseca & González, 2010). Size to GDP is calculated as book assets divided by domestic nominal GDP.

Asset substitution is the transfer of value from debtholders to shareholders by substituting high-risk assets for low-risk ones (Leland, 1998, also see section I.A). Credit markets take this tendency in to account by increasing required rates. This is a manifestation of market discipline. The impact of the asset substitution problem is inﬂuenced by the amount of long-term debt among liabilities (Johnson, 2003). Long-term debt has more time to maturity so there exists a greater opportunity to bank owners to engage in gambling at the expense of debtholders. As a result, long-term debt to assets must be included as a control variable, as it is a determinant of market discipline (i.e. it could be correlated both with bank risk and with either one of the dependent variables). Long-term debt is calculated as assets minus equity minus deposits, money market & short term funding.

When a bank is government owned for an extended period, this often implies an implicit guarantee to the creditors of the bank (Borisova & Megginson, 2011). The same applies to banks which are bailed out. The value of a guarantee of this kind will also be related to sovereign credit risk, probably also negatively. However, the occurrence of a bailout implies an initial divergence of bank and sovereign risk (right after the bailout) followed by comovement (Acharya, Drechsler & Schnabl, 2011 & Ejsing & Lemke, 2011). This could make a bank bailout within the sample period, distort estimation of the relationship between sovereign risk and market discipline. To deal with this problem, a dummy is included for banks which were delisted during the sample period. The dummy equals 1 in the year of delisting and each year

8_{Gropp & Vesala (2004), among others, use tobin’s Q as a proxy for charter value. However, tobin’s Q is not available in}

(23)

thereafter. Bailouts are often accompanied by a delisting, so including this rough proxy could take away some of the bias caused by banks in the sample being bailed out.

Nier & Baumann (2006) and Martinez-Peria & Schmukler (2001) use the ratio of uninsured deposits to total liabilities to account for the eﬀect of deposit insurance on risk taking (see section I.B). National deposit insurance institutions diﬀer in the classes of deposits they insure. For this reason, it is hard to distinguish between insured and uninsured deposits, when an international sample of banks is considered. However, it is assumed that it will in general be customer deposits which are insured. For this reason, the customer deposits to total assets ratio is included as a control variable. The national level of deposit insurance coverage is also controlled for as it is an important determinant of market discipline, as further elaborated on in the next subsection.

All bank-level control variables are measured in decimals, except the bailout dummy and absolute size (the latter is resembled by total book assets measured in US dollars).

II.C.5 Country-level controls

Angkinand & Wihlborg (2010), Cubillas, Fernández-Alvarez & González (2013), Demirgüc-Kunt & Huizinga (2004), and Hadad et al. (2011) were helpful in the selection of macroeconomic controls. Each of these market discipline studies at least uses the real interest rate, inﬂation rate, real GDP per capita, and real GDP growth as macroeconomic control variables. Macroeconomic factors also are related to sovereign credit risk (see Aizenman, Hutchinson & Jinjarak, 2013) and are therefore indispensable to include as a control variable.

Boyd, de Nicoló & Jalal (2006), as well as Hadad et al. (2011) indicate that country-level competition among banks is an important factor inﬂuencing risk taking by banks. Because competition is also likely to be correlated with interest expenses on bank deposits (which is a “competitive price”), country-level competition should be included as a control variable in any market discipline regression. The Panzar-Rosse H-statistic9 _{is used as a proxy for country-level competition in the banking sector. The H-statistic has a}

negative values for monopoly-like levels of banking competition, and is equal to 1 for perfect competition. A negative relationship between the value of implicit and explicit guarantees appears to exist. High

9_{The Panzar-Rosse H-statistic is calculated in two steps: ﬁrst cross-sectional regressions are run of log total revenues}

on log input prices of banks. Input prices are the implicit cost of debt, personnel expenses to assets and other expenses to assets, respectively. Cross-sectional subset regressions are run within each country-year. Within each country-year, the resulting coeﬃcient estimates on the three independent variables are added together. The resulting panel of added coeﬃcients resembles country-year level sensitivity of revenues to input prices. Under a monopoly, an increase in input prices will lead to a decrease in revenues due to rising marginal costs (rising marginal costs lead to a lower optimal level of sales, Panzar & Rosse, 1982). As a result, the Panzar-Rosse H-statistic has a negative or zero value for monopoly-like levels of banking competition, and is equal to 1 for perfect competition.

(24)

explicit guarantees or deposit insurance coverage implies no large implicit safety net is necessary (Gropp & Vesala, 2004). No explicitly stated safety net is simply not credible, suggesting higher implicit safety nets. This relationship was modelled by Angkinand & Wihlborg (2010). They suggest a U-shaped relationship between market discipline and deposit insurance coverage: low deposit insurance means high implicit guarantees and therefore low market discipline. High deposit insurance again leads to low market discipline. The level of deposit insurance maximizing market discipline lies in between. To control for this U-shaped relationship, deposit insurance coverage is entered in to the regression model with both a regular and a quadratic term. Deposit insurance coverage is calculated as the national deposit insurance limit divided by deposits per capita. One country (Chile) has no deposit insurance limit; its deposit insurance limit was therefore set to the maximum deposit insurance limit in the sample. Another country (South Africa) has zero deposit insurance coverage, leading to missing values after logarithmic transformation. To avoid exclusion of South Africa from the sample, these missing values were replaced by the minimum value of the transformed variable.

Other macroeconomic factors, such as governance-related factors and market structure, are also reg-ularly included in market discipline studies. However, here these factors are excluded because these are assumed to be either captured by the bank fixed effects, or sufficiently correlated with the country-level controls included in the regressions (recall that the time-series of the panel is relatively short).

The sample covers country-years which are subject to systemic financial crises and sovereign debt crises (such as the 2012 euro crisis). The presence of such a crisis could constitute an omitted variable correlated with each of the variables of interest, causing endogeneity. However, it is reasonably assumed that country-level macroeconomic controls together with bank-level profitability (ROA) and year fixed effects fully capture the effect these financial crises have on the variables of interest.

The real 3-month interest rate, inﬂation, and real GDP growth are measured in decimals. Real GDP per capita is measured in US dollars. The Panzar-Rosse H-statistic as well as deposit insurance coverage have variable-speciﬁc units of measurement.

II.C.6 Summary statistics & preparation of variables

Table 4 reports summary statistics on all variables used in the panel regressions. When a variable had outliers of a magnitude characteristic for bad data (such as typos, etc.), or outliers of a magnitude impeding the validity of regression estimators, the variable was winsorized at the top and bottom 0.5%.10 _Whether

10_{Some variables still show outliers, though. The winsorization procedure was standardized to 0.5% winsorization, to}

(25)

winsorization was applied can be seen in the second-last column. Values in the summary statistics are post-winsorization, if applicable.

We see that the average implicit cost of debt across the sample is 0.03 with a standard deviation of 0.043. This seems reasonable. The relative change in time deposits is on average -0.024. However, it has a high standard deviation relative to the mean of 0.379. The conclusion of total deposits having decreased over the sample should be drawn with caution, as the data shows outliers for this variable, even after winsorization. The capital ratios have reasonable means and standard deviations (0.154 with σ = 0.179 for the tier 1 ratio, and 0.194 with σ = 0.313 for the total regulatory ratio). These figures reflect the total regulatory capital ratio to be a broader definition of capital than the tier 1 ratio is. The Z-score appears to be positively skewed (which is not due to a small number of outliers). This distributional shape translates in a standard deviation of as much as 269.623, compared to a median of 30.729. 1-year sovereign CDS has a mean of 70 basis points, compared to a mean of 520 basis points for 5-year CDS. 5-year CDS spreads reflect higher default risk premia, due to longer times to maturity of the underlying. The standard deviation of 5-year CDS is also much higher than that of 1-year CDS (0.229 vs. 0.026).

Interpretation of control variable statistics is skipped for brevity, but one needs to be highlighted. The mean customer deposits to assets ratio is 0.623, against a median of 0.713. Note that this ratio excludes deposits from banks and some other institutions. It may not be new knowledge, but deposits are clearly the single most important way of funding to banks, i.e. far more important than term loans, bonds, and other types of debt.

Before being included in the regressions, some variables are transformed to their natural logarithm. The natural logarithm of a variable has a distribution which is more close to a normal distribution, which is beneﬁcial to the validity of the regression estimators. It are the original, non-transformed values which are reported in table 4. This is to allow for more straightforward interpretation of the summary statistics. The last column shows which variables enter in to the regressions as their natural logarithm. For bank-level variables, logarithms are only taken when less than 2.5% of the observations of a variable have a zero or negative value (after winsorization, if applicable). This is done because logarithmic transformation of a negative or zero value results in a missing observation. As it is not desired to exclude full country-year observations, country-level variables were treated more carefully in this sense. From all country-level variables, a natural logarithm is only taken of sovereign CDS, real GDP per capita and deposit insurance coverage. The ﬁrst two have no zero or negative values; for deposit insurance coverage, resulting missing values were set to the minimum value of the variable after taking the natural logarithm (as explained in

(26)

(27)

T able 4: Summary statistics of all v ariables The non-transformed v alues of the v ariables are presen ted to allo w more straigh tforw ard in terpretation. ∆ Log time dep osits is the only exception, as this v ariable resem bles the first difference of a logarithmically transformed v ariable. “Obs” indicates the n um b er of observ ations. σ indicates the standard devia tion of a v ariable o v er the full sample. The column “Win. ” indicates whether a v ariable w as winsorized at the top and b ottom 0.5%. The last column (T rans.) indicates whic h transformation is used to normalize the distribution of a v ariable, b efore inclusion in the regressions. All v ariables as presen ted in the table are measured in decim als, except Z-score, total assets, real GDP/capita, H-statistic and dep osit insurance co v erage. T otal assets is measured in millions of US dollars, real GDP/capita in US dollars, and Z-score, H-statistic and dep osit insurance co v erage eac h ha v e v a riable-sp ecific units of measuremen t. In the app endix a description of all v ariables is pro vided as w ell as the source of the data. V ariable Obs Mean σ Min Q1 Median Q3 Max Win. T rans. Implicit cost of debt 26,788 0.030 0.043 0.000 0.011 0.020 0.033 0.399 Y es Log ∆ Log time dep osits 14,583 -0.024 0.379 -6.767 -0.143 -0.012 0.095 9.086 Y es (Log) Tier 1 capital ratio 9,300 0.154 0.179 0.000 0.092 0.117 0.155 1.936 Y es Log T otal regulatory capital ratio 15,983 0.196 0.313 0.014 0.120 0.148 0.186 3.598 Y es Log Z-score 27,451 105.46 269.623 -1.577 11.927 30.729 74.263 2,342.468 Y es Log So v. CDS spread (1y) 30,149 0.007 0.026 0.000 a 0.001 0.002 0.005 0.256 Y es Log So v. CDS spread (5y) 30,149 0.052 0.229 0.001 0.003 0.005 0.010 1.49 Y es Log Return on assets 29,757 0.006 0.040 -0.266 0.000 0.003 0.008 0.248 Y es -T otal assets 30,149 44,641.740 215,719.5 0.234 151.262 932.169 5,643.680 2,233,245 Y es Log Assets/nominal GDP 30,146 0.129 0.596 0.000 0.000 0.002 0.015 5.831 Y es Log Long-term debt to assets 28,463 0.135 0.198 0.000 0.016 0.044 0.162 0.949 Y es -Cust. dep osits to assets 25,658 0.623 0.280 0.000 0.464 0.713 0.835 1.191 -“Bailout” dumm y 30,149 0.026 0.159 0 0 0 0 1 -Real 3-mon th in terest rate 30,149 0.002 0.018 -0.055 -0.009 -0.001 0.014 0.078 -Inflation 30,149 0.020 0.019 -0.035 0.009 0.021 0.028 0.116 -Real GDP/capita 30,143 36,668.070 14,284.940 2,085.276 25,142.640 41,314.700 45,812.510 68,344.800 -Log Real GDP gro wth 30,146 -0.016 0.087 -0.256 -0.085 -0.023 0.063 0.201 -P anzar-Rosse H-statistic 29,853 0.550 0.278 -1.257 0.421 0.503 0.635 1.711 -Dep osit insurance co v erage 30,149 5.153 13.275 0.000 1.671 2.772 2.799 75.395 -Log b aRounded. bF or the dep osit insurance co v erage v ariable D epI nsC ov i,t , missing v alues w ere set to the minim um non-missing v alue after logarithmic transformation. This w as necessary b ecause one coun try (South Africa) has no dep osit insurance: the en tire coun try w ould ha v e b een excluded with only taking the natural logarithm of dep osit insurance co v erage (as the natural logarithm of 0 do es not exist).

(28)

III

Results

In this section, the results from the baseline panel regressions (equations [1] and [2]) are explicated and interpreted. Subsection III.A presents the results of the preliminary test regressions. The preliminary tests were necessary because in the baseline regressions, the coefficient estimates on the main effects may be biased due to multicollinearity with the interaction term. In the preliminary tests the same regression model is used as in the baseline model, except for the interaction terms being excluded (see equations [3] and [4]). Doing so, it can be determined in which direction the dependent variables are affected by bank risk and sovereign risk, respectively. This knowledge is taken further to the baseline regressions, so that we are not inferentially dependent on main effect estimates which may be biased due to multicollinearity. Subsection III.B.1 elaborates on the regressions with the implicit cost of debt as the dependent variable; subsection III.B.2 elaborates on the regressions with the change in term deposits as the dependent variable; in subsection III.C the coefficients on the control variables are interpreted; final thoughts and results implications are presented in subsection III.D.

III.A

Preliminary test of main eﬀects

Tables 5 and 6 show the preliminary tests of the main effects. Market discipline is evident when bank funding costs are positively affected by bank risk (Martinez-Peria & Schmukler, 2001). Therefore we expect negative and significant signs on the capital ratios and the Z-score, as these are an inverse measures of bank risk. We expect a positive sign on the sovereign CDS variables, as sovereign risk spills over to the financial sector (e.g. Angeloni & Wolff, 2012). When changes in term deposits is the dependent variable, we expect a positive effect of the capital ratio and the Z-score. This is because creditors also react to bank risk by withdrawing deposits. For the sovereign CDS variables a negative sign is expected. Sovereign credit risk may deter investors to deposit in banks in “risky countries” (as happened in the recent Euro crisis, see Panetta et al. 2011).

Table 5 has the implicit cost of debt as the dependent variable and table 6 has the change in term deposits as the dependent variable. In each regression, the risk proxy is lagged in order to avoid simul-taneous causality bias. In both tables, columns (1) to (3) use 1-year sovereign CDS spreads as proxy for sovereign credit risk, while columns (4) to (6) use 5-year sovereign CDS as proxy. In both tables, columns (1) and (4) have the tier 1 capital ratio as bank risk proxy, columns (2) and (5) have the total regulatory capital ratio as bank risk proxy, and columns (3) and (6) have the Z-score as bank risk proxy. Bank and year ﬁxed eﬀects are included in all regressions, their indication is omitted to save space.

(29)

With the implicit cost of debt as the dependent variable, the tier 1 capital ratio has a highly significant coefficient of -0.150 in both regressions where it was included. This means that an increase in the capital ratio leads to a decrease in the implicit cost of debt. This is in line with expectation, as better capitalized banks have less default risk what should be reflected in their cost of debt. Note that both the tier 1 capital ratio and the dependent variable did undergo logarithmic transformation. Correspondingly, the coefficient of -0.150 means that a 1% increase in the tier 1 capital ratio is associated with a 0.15% decrease in the implicit cost of debt. Note that a percentage decrease is considered which is applied to a variable measured in decimals (DebtCosti,t).11 For instance, a one standard deviation increase in the tier 1 ratio corresponds

to a (0.179 / 0.154) * 100% = 116.2% increase relative to the mean (see table 4 for the summary statistics). Correspondingly, a one standard deviation increase of the tier 1 ratio relative to the mean, would cause a 116.2 * -0.15 = 17.43% decrease in the implicit cost of debt. An old cost of debt of 0.05 would decrease to 0.05 * (1 - 0.1743) = 0.0412.

The coefficients on the total regulatory capital ratio both have the right sign, but are only significant at the 20% level. Apparently this ratio only has a minimal impact on the cost of debt, if any impact at all. The total regulatory capital ratio probably encompasses a too broad definition of buffer capital in the eyes of creditors; an inferior measure of capital buffer, must also be a worse predictor of funding costs.

The Z-score does have predictive power for the implicit cost of debt. Its coeﬃcient estimates are signiﬁ-cant at the 5% level, although they have a lower magnitude than those on the tier 1 ratio. Both regressions (3) and (6) from table 5 show that a 1% increase in the Z-score leads to a decrease of approximately 0.044% in the cost of debt.

To summarize, the tier 1 ratio and the Z-score have signiﬁcant coeﬃcient estimates in each regression using the implicit cost of debt. All risk proxies have the expected sign.

Next, we look at the coefficient estimates on sovereign CDS spreads with implicit cost of debt as the dependent variable (which can also be found in table 5). All coefficients on sovereign CDS spreads are positive and significant on the 1% level. This indicates that bank creditors require to be compensated for increases in sovereign credit risk, as sovereign credit risk usually spills over to the financial sector (see section I.C). There does not seem to be a structural difference between 1-year or 5-year sovereign CDS spreads regarding their effect on the implicit cost of debt at first. The six coefficients on the sovereign CDS variables in columns (1) to (6) in table 5, indicate that a 1% change in sovereign CDS spreads leads to an increase in the implicit cost of debt of in between 0.030% and 0.104%. For example, if CDS spreads would

11_{Regression coeﬃcients where both the dependent as well as the considered independent variable were logarithmically}

transformed, have the interpretation of a elasticity; i.e. the percentage change resulting from a 1% change in the independent variable. Such types of regressions are also referred to as “log-log” models.

(30)

double (i.e. a 100% increase), and the old cost of debt would be 0.05, the new cost of debt would be of 1.104 * 0.05 = 0.0552 (using the interaction coefficient from column [5]). This effect at first appears to be small; however, this effect should be thought of in the context of banks generally being assumed to hedge away specific risks, such as exposure to domestic sovereign credit risk (see Froot & Stein, 1998). Clearly, banks within the sample fail to do so. If they didn’t, creditors wouldn’t require to be compensated. This finding sheds new light on the risk management ability of banks. Furthermore, the fact that creditors clearly require to be compensated for sovereign credit risk, can be interpreted as evidence against the risk-free status of sovereign debt a fortiori.

The results from the preliminary test regressions using the change in log time deposits as the dependent variable, are also largely in line with expectations (see table 6). Both capital ratios and the Z-score are positively related to changes in term deposits. Decreasing deposits in reaction to decreased capital adequacy or solvency (measured by the Z-score), indicates depositors to discipline the bank for risk by withdrawing.

Contrary to the the regressions with the implicit cost of debt as the dependent variable, here the total regulatory capital ratio is a 5%-level significant predictor of the dependent variable as well. Interpretation of the coefficient estimates is different compared to the cost of debt regressions. Here, a linear-log model is considered as the dependent variable resembles the “linear” percentage change in term deposits (logs were only used in calculating it). Recall that a one-standard deviation increase in the tier 1 ratio corresponds to a 116.2% increase relative to the mean. This means that, for regression (1), a one standard deviation increase in the tier 1 capital ratio increases the change in term deposits with 116.2% * 0.01 * 0.227 = 0.264. The risk proxy coefficient estimates in regressions (2) to (6) have similar interpretations. All are significant at the 1% level.

Sovereign credit risk and market discipline in the financial sector