* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download Too Big To Fail Or To Save? Evidence from the
Survey
Document related concepts
Transcript
Too Big To Fail Or To Save? Evidence from the CDS Market∗ Andreas Barth† Johannes Gutenberg University Mainz and GSEFM Isabel Schnabel‡ Johannes Gutenberg University Mainz, CEPR, and MPI Bonn October 1, 2012 Abstract This paper argues that bank size is not a satisfactory measure of systemic risk because it neglects important aspects such as interconnectedness, correlation, and the economic context. We show that, when controlling for systemic risk using the CoVaR measure introduced by Adrian and Brunnermeier (2011), a bank’s size has no or even a positive effect on banks’ CDS spreads, especially if the bank is based in a highly indebted country. In contrast, a bank’s contribution to systemic risk has a significant negative effect on banks’ CDS spreads. The effect of systemic risk rises sharply at the onset of the financial crisis in August 2007. A country’s ∗ This is a preliminary version of a paper prepared for the 56th Panel Meeting of Economic Policy, October 2012. We thank Valeriya Dinger, Charles Goodhart, Florian Hett, and Deyan Radev for valuable comments and suggestions. We also benefited from comments by participants of the Brown Bag Seminar at JGU Mainz and the 1st Research Workshop in Financial Economics at JGU Mainz. † Gutenberg School of Management and Economics, Johannes Gutenberg University Mainz, 55099 Mainz, Germany, telephone +49-6131-39-20746, fax +49-6131-39-25588, e-mail [email protected]. ‡ Gutenberg School of Management and Economics, Johannes Gutenberg University Mainz, 55099 Mainz, Germany, telephone +49-6131-39-24191, fax +49-6131-39-25588, e-mail [email protected]. indebtedness starts to matter only after the large-scale bail-outs in response to the Lehman default in September 2008. Hence, our results suggest that banks are not too big to fail, but they may be too systemic to fail and too big to be saved. Keywords: Too big to fail; too systemic to fail; too big to save; credit default swaps. JEL-Classification: G21, G28. 1 Introduction The too-big-to-fail doctrine is a widely accepted hypothesis. It is argued that large banks benefit from implicit bail-out guarantees because no government is willing to hazard the consequences of a large bank’s default on the financial system and on the economy as a whole. Such guarantees effectively constitute a government subsidy, which should be reflected in market prices. We argue in this paper that bail-out probabilities depend on a bank’s systemic importance rather than its size, and that the two concepts do not necessarily coincide. It is the systemic risk emanating from a bank, which justifies government intervention. Therefore, banks are not too big to fail (TBTF), but too systemic to fail (TSTF). Quite on the contrary, size may actually reduce bail-out expectations, as the events in Iceland in the fall of 2008 have shown. Being a small country, Iceland had a banking sector consisting mainly of three banks, which had vast balance sheets relative to Icelandic GDP. When these institutions got in distress, the Icelandic government was simply not able to bail them out. Consequently, the tremendous size made those banks too big to save (TBTS) rather than too big to fail. While the TBTF problem has been discussed extensively in the literature (see, e. g. Boyd and Gertler, 1994; Kaufman, 2002; Stern and Feldman, 2004), the TBTS problem has received attention only recently. An early discussion of the TBTS problem is by Hellwig (1998) who argues that bank mergers in response to the too-big-to-fail problem may make some institutions “too big to be rescued.” Hüpkes (2005) stresses that the TBTS problem depends on a country’s size. Large complex financial institutions may be too large to be saved especially in “small economies such as Belgium, the Netherlands, Sweden and Switzerland” (Hüpkes, 2005). Rime (2005) empirically analyzes the effect of bank size (as a proxy of too-big-to-fail expectations) on issuer ratings. Beside a measure of bank size (total assets and a ratio of total assets to total assets of the banking sector), he includes a proxy of the TBTS problem, namely the ratio of total bank liabilities to GDP. However, he does not find any evidence that rating agencies “incorporate some TBTBR [too-big-to-be-rescued] considerations in issuer ratings” (Rime, 2005) . Our paper is most closely related to the work by Völz and Wedow (2011) and DemirgüçKunt and Huizinga (2010), who both study the effect of bank size on CDS prices. Völz and Wedow (2011) focus mainly on the too-big-to-fail phenomenon and refer to the TBTS problem just as an aside. Size is measured by the ratio of a bank’s market capitalization 1 over the home country’s GDP. In order to allow for TBTS effects, they also include a quadratic term of this variable. They confirm that larger banks generally exhibit lower CDS spreads, supporting the presence of a TBTF problem. However, the relationship between bank size and CDS spreads is shown to revert at some point, suggesting that some banks have already reached a size that makes them too big to rescue (Völz and Wedow, 2011). In a similar vein, Demirgüç-Kunt and Huizinga (2010) find evidence of TBTF and TBTS problems in market valuations, but less so in CDS spreads. This finding is somewhat surprising as the theoretical effect on CDS prices is more straightforward than that on equity prices. To identify TBTF, the authors consider the effect of a bank’s total assets on CDS prices. In contrast, TBTS is measured by the relative bank size, measured as bank liabilities over GDP. Demirgüç-Kunt and Huizinga (2010) also consider interaction terms between relative bank size and a country’s debt over GDP ratio or its fiscal balance. Indeed, in some (though not all) regressions, the effect of relative bank size appears to vary depending on the home country’s debt level and fiscal balance, supporting the idea of a TBTS problem. Although both papers claim to provide empirical evidence for banks being too big to fail and too big to save, their analyses are in our view not able to clearly distinguish between these phenomena. Both TBTF and TBTS are measured on the basis of banks’ size, and the findings can therefore not easily be attributed to one of the two phenomena. Using non-linear regression functions or interaction terms does not adequately solve this problem. The too-big-to-save phenomenon cannot fully be captured by the convexity of the regression function because this does not take into account the fiscal situation of the home country. The too-big-to-fail problem – or too-systemic-to-fail problem, as we prefer to call it – is not adequately captured by a bank’s total assets because absolute bank size is only a crude proxy of systemic relevance. Whether a bank is too systemic to fail is not only determined by its size, but also by its interconnectedness and its correlation with the remaining banking sector, as has also been argued recently by Zhou (2010). Moreover, the economic context matters. For example, if a small bank is hit by a shock in non-crisis times, it would hardly be regarded as being systemic. But in the middle of a financial crisis, even such a small bank may be rescued due to the fear of contagion effects. The rescue of the German bank IKB is a case in point. Given that total assets change only slowly over time and that changes are only vaguely related to changes in systemic significance, bank size is not able to capture the importance of the economic context. Therefore, one has to consider broader measures of systemic relevance than just size. 2 We propose to identify the TSTF problem by employing a broad measure of systemic risk, namely the contribution of an individual bank to the system’s risk, measured by ∆CoV aR as proposed by Adrian and Brunnermeier (2011). Once we control for the TSTF phenomenon, we can identify the effect of being too big to save by a bank’s size relative to GDP. In addition to measuring the overall importance of TSTF and TBTS, we analyze how the importance changed over the financial crisis. For that purpose, we divide our sample into three sub-periods. The first period starts in 2005 and ends in July 2007, just before the beginning of the financial crisis. The second period comprises the first crisis phase until August 2008. Finally, the third period starts with the failure of the US bank Lehman Brothers in September 2008. Before the financial crisis, neither TSTF nor TBTS are expected to play a large role. CDS spreads were very low even for non-systemic banks. This is likely to change with the onset of the crisis when the TSTF problem is expected to become much more pronounced. In contrast, the TBTS problem is likely to be affected by the huge bail-outs taking place after the Lehman default and by the failure of the banking system and the emergence of sovereign debt problems in Iceland, which occurred at about the same time as the default of Lehman Brothers. These events may have been a wake-up call for investors that the credibility of bank bail-out guarantees is linked to the home country’s fiscal situation and that such guarantees may not be fulfilled under all circumstances by ailing governments. Our results suggest that markets indeed reflect banks’ contribution to systemic risk. A higher contribution to systemic risk (corresponding to a lower ∆CoV aR) translates into lower CDS spreads, supporting the existence of TSTF. Bank size does not matter when the bank’s home country has low debt levels. However, the effect of bank size increases in the debt ratio of the home country and becomes positive at moderate debt levels. These results are robust across many different specifications. Splitting the sample in sub-periods is also instructive. We find that neither TSTF nor TBTS were priced in the market before the financial crisis erupted in 2007. However, the importance of TSTF rises sharply after August 2007, whereas the TBTS problem remains insignificant. TBTS intensifies sharply after the collapse of Icelandic banks in September 2008. The paper is organized as follows. In Section 2, we introduce the concepts of banks being too systemic to fail or too big to be saved, and discuss why bank size is only a crude proxy of systemic risk. Section 3 introduces the empirical model, states the main hypotheses, provides data sources and describes the major variables used in the empirical analysis. The empirical results from our baseline regressions are shown in Section 4. In Section 5, 3 we show empirical results for the three sub-periods of our sample – pre-crisis, beginning of crisis, and post-Lehman. Section 6 concludes. 2 2.1 TSTF versus TBTS Systemic Risk and Bank Size A financial institution contributes the more to overall systemic risk, the larger are the repercussions of its failure on the financial system and the real economy. In order to provide a proper definition of the concept of systemic importance, we have to analyze the driving forces of different channels of contagion. First, there is an information contagion channel, as introduced by Chen (1999) and Acharya and Yorulmazer (2008). They show that contagion may arise due to new information revealed to depositors after a bank failure. Observing difficulties at one bank, depositors of other banks fear that their bank subsequently experiences financial troubles as well and withdraw their deposits, which can result in a bank run. The amplitude of this contagion channel depends on the similarity of financial institutions. Second, the failure of a financial institution may infect other institutions with significant direct credit exposures at this bank. This form of contagion via the interbank market, as shown by Allen and Gale (2000), arises when a default of one bank results in significant write-offs of claims at the failed bank. This can be sufficient to generate trouble in the whole financial system. The dominant determinant of the magnitude of repercussion effects through this channel is the interconnectedness of banks. The third contagion channel works via macro-economic feedbacks. Brunnermeier and Pedersen (2009) model this form of contagion in terms of liquidity spirals. They show that, if banks hold similar assets and one bank has to sell assets at a fire-sale price due to short liquidity, these fire sales of the illiquid bank depress asset prices, which also affects other banks. Such a domino effect through asset prices can happen even in the absence of bank defaults infecting other financial institutions. A decline in prices is already sufficient. The severity of this channel is mostly driven by the correlation among financial institutions, but also by their size. Furthermore, the impact of fire sales is driven to a large extent by the context. In times of financial crises, many banks suffer simultaneously from short liquidity and must sell assets at fire sale prices. Thus, in distress periods, the behavior of many small banks 4 with correlated portfolios has the same impact on prices as the action of one single huge bank. Fourth, contagion may arise from self-fulfilling expectations. The argument of this channel works similar to the bad equilibrium of the seminal work of Diamond and Dybvig (1983). The failure of one financial institution coordinates expectations of all investors and depositors, such that investors and depositors start a run. Since the crucial element of this channel is expectations, the probability of contagion effects are mostly driven by two factors: size and context. Investors are more likely to coordinate on the failure of a large bank. However, if the financial sector is in a critical condition, the collapse of a small bank can be sufficient to give rise to bad expectations of all investors and depositors. In calm periods, the collapse of the same bank would not bother investors or depositors at all. Thus, we can conclude that it is not only the size of a bank which drives the systemic importance. There are many other factors playing a vital role, especially interconnectedness, correlation, and the economic context. Hence, the preceding discussion suggests that measuring the systemic importance of a bank by its size is too narrow. Nevertheless, most empirical work so far has focussed on the too-big-to-fail problem. We argue that one should take a broader perspective and call banks benefiting from implicit bail-out guarantees too systemic to fail (TSTF) rather than too big to fail. 2.2 Measuring Systemic Risk by Conditional Value at Risk In this paper, we measure systemic importance using the ∆CoV aRt measure introduced by Adrian and Brunnermeier (2011).This measure tries to capture an individual institution’s contribution to overall systemic risk. Due to the comprehensive nature of this variable, it captures not only a bank’s size, but also other factors, such as its correlation with other banks, as well as the economic context. A bank’s CoV aR relative to the system is defined as the conditional value at risk, i. e. the q% − V aR of the whole financial sector conditional on the fact that an institution i is at its V aR level: system|X i =V aRiq CoV aRq := V aRqsystem |V aRqi (1) It thus gives the maximum dollar loss of the financial system at probability q conditional on institution i being in distress. 5 Following the procedure proposed by Adrian and Brunnermeier (2011), we estimate CoVaR by using quantile regression. The CoVaR is the predicted value of a quantile regression of the financial system on an individual institution for the q-quantile, i. e. CoV aRqsystem = α̂qi + β̂qi · V aRqi . (2) To allow for time variation of the estimated CoV aR, we model the conditional distributions as a function of a number of state variables. Following Adrian and Brunnermeier (2011), we get CoV aR as the predicted value from the following quantile regressions: CoV aRti (q) = α̂system|i + β̂ system|i V aRti (q) + γ̂ system|i Mt−1 (3) where Mt−1 is a vector of lagged state variables. For our calculations, we focus on the CoV aRti (q) of the same variable as Adrian and Brunnermeier (2011), i. e. the growth rates of market-valued total financial assets, where we get the market equity from Thomson Reuters Datastream and the leverage from Bureau van Dijk’s Bankscope. Furthermore, we use use the same vector of lagged state variables as Adrian and Brunnermeier, but exclude the real estate sector return in excess of the market return. We run the quantile regressions on a 1% stress level. ∆CoV aRt is defined as the difference between CoV aRti (q) conditioned on distress times and normal times of bank i, where normal times are characterized by the median, i. e. ∆CoV aRti (1%) = CoV aRti (1%) − CoV aRti (50%). (4) From these regressions, we obtain a panel of daily ∆CoV aRt , measuring the risk contribution of individual institutions to the system. It should capture all the determinants of the contagion channels described in section 2, i. e. the systemic importance deriving from correlation, size, interconnectedness, and from the economic context.1 Note that ∆CoV aRt is typically negative, with a more negative value indicating a greater contribution to systemic risk. For the ease of interpretation, we use −∆CoV aRt throughout the paper, implying that an increase in this variable is to be interpreted as an increase in the contribution to systemic risk. 1 For a detailed explanation of CoVaR and its estimation, see Adrian and Brunnermeier (2011). 6 8 6 4 2 0 0 1 2 3 Liabilities / GDP −∆ CoVaR 4 5 Fitted values Figure 1: Scatter plot of −∆CoV aR and the ratio of bank liabilities to GDP. The figure is based on individual bank data between 2005 and 2011. Sources: Own calculations, Bankscope, WDI. Our data illustrate the difference between size and systemic risk.2 We find a low correlation of 0.241 of the ratio of bank liabilities to GDP (which is used in the literature to measure the TBTF and TBTS phenomena) and −∆CoV aR (as our measure of TSTF). The scatter plot in Figure 1 confirms our doubts about the suitability of size as a proxy of systemic relevance. In a regression of −∆CoV aR on the size ratio, only 5.80% of the variation of systemic risk contributions are explained. The most prominent examples of banks that are small relative to home country’s GDP but that contribute to a large extent to systemic risk are given by US investment banks. For example, Merrill Lynch has a ratio of liabilities to GDP of only 5% in October 2008 (15% quantile of our sample), but a contribution to systemic risk, −∆CoV aR, of 5.80 (99% quantile). Similarly, Goldman Sachs had in November 2008 a −∆CoV aR of 6.30, with a size ratio of only 5.92% of GDP. In contrast, there also are banks with huge relative size, but only a limited contribution to overall systemic risk. For example, the Japanese bank Mitsubishi UFJ had in December 2010 a large relative size of 41.11% of its home country’s GDP, but a −∆CoV aR of only 0.36. Moreover, a systemic risk measure should reflect the economic context, i. e. it should be higher in times of crisis. As is shown in Table 1, −∆CoV aR suggests a sharp increase 2 See Section 3.3 for a description of data sources. 7 in systemic risk during the financial crisis by more than one standard deviation of the pre-crisis −∆CoV aR and a further rise after the default of Lehman Brothers. The ratio of bank liabilities to GDP, however, does hardly mirror the crisis: the maximum drops, and the mean increases only slightly in the crisis period and even drops in the post-Lehman period. Table 1: Evolution of the ratio of bank liabilities to GDP (TBTS) and −∆CoV aR during the crisis Variable Mean Std. Dev. Min. Max. N 0.005 0.004 2.669 4.848 1472 1472 0.015 0.009 4.297 4.789 835 835 8.590 4.031 1956 1956 pre-crisis period −∆CoV aR T BT S 0.854 0.556 0.503 0.827 crisis period −∆CoV aR T BT S 1.509 0.606 0.860 0.816 post-Lehman period −∆CoV aR T BT S 1.886 0.558 1.398 0.597 0.011 0.007 Descriptive statistics for −∆CoV aR and T BT S. The pre-crisis period denotes the time until July 2007. The crisis period is the period from August 2007 to August 2008. The post-Lehman period starts in September 2008. 2.3 Too Big to Save The recent crisis has shown that banks cannot just be too big to fail, but also too big to save (TBTS). In particular, if the size of a financial institution exceeds a country’s ability of a bail-out, the bank simply cannot be rescued by the safety net. Therefore, we do not expect that banks will benefit the more from a country’s safety net the bigger they are, once we control for systemic risk. While the too-big-to-fail doctrine argues that banks’ size increases the bail-out probability, we argue that it is the systemic importance of banks that drives bail-out expectations. Hence, a bank’s size should decrease the bailout probability conditional on systemic risk. In this sense, TBTS can be seen as the antagonist of the too-systemic-to-fail problem. 8 The most prominent recent example of banks being too big to save was the case of Iceland, which experienced considerable economic trouble due to its banking sector. The financial turmoil resulted in strong losses of its internationally active banks, which had vast balance sheets relative to the size of the economy.3 Consequently, the Icelandic Financial Supervisory Authority took control of these banks within one week in October 2008. It is instructive to compare the CDS prices, as indicators of the market’s default expectation, of the Icelandic bank Kaupthing and the German bank IKB, two institutions of similar size who both faced serious troubles during the recent financial crisis. Figure 2 suggests that markets perceived a sharp disparity in the bail-out probabilities of the two banks. 4500.00 4000.00 3500.00 3000.00 2500.00 2000.00 Kaupthing 1500.00 IKB 1000.00 500.00 0.00 Figure 2: CDS spreads of the Icelandic bank Kaupthing and the German bank IKB (in basis points). Source: Thomson Reuters Datastream. 3 Empirical Analysis 3.1 Hypotheses Our empirical analysis tries to explain the evolution of bank CDS spreads. A credit default swap (CDS) is an insurance contract against default or another type of credit event. More precisely, the protection buyer pays a default swap premium for receiving 3 Total assets of the three largest banks relative to GDP skyrocketed from less than 200% in 2003 to almost 1000% in 2008, cf. The Central Bank of Iceland (2009). 9 the guarantee that the protection seller covers the incurred losses if a default is triggered. Therefore, the price of a bank CDS is a function of the market’s expectations of a bank’s actual probability of default (PD): CDS = f (market expectation of actual PD) (5) The market expectation of banks’ actual PD is a function of the bank-specific fundamental PD and the probability of a government bail-out in case of distress (cf. Gropp, Hakenes, and Schnabel, 2010): actual PD = (1 − bail-out probability | fundamental default ) · fundamental PD. (6) We now formulate four hypotheses that will be tested in the empirical analysis. The first hypothesis refers to the too-systemic-to-fail problem. Since the consequences of the failure of a systemic institution for the rest of the financial system can be enormous, a bank is the more likely to be rescued by the government, the higher is its systemic risk. We therefore expect that systemic banks have a lower market expectation of default, which should show up in lower CDS spreads, as postulated by Hypothesis 1. Hypothesis 1 (Too systemic to fail) Ceteris paribus, CDS spreads are smaller for banks with a higher contribution to systemic risk. For a given level of systemic risk, there are no incentives for a government to bail out a bank just because of its size. To the contrary, if a financial institution has reached a particular size relative to its home country’s GDP, the country may not be able to bail out this institution. Therefore, we expect a non-linear relationship between bank size and CDS spreads. For relatively small bank sizes, CDS spreads should not be affected by bank size when controlling for systemic risk. For large banks, the effect of bank size should be positive. Note that this prediction differs from the prediction by Völz and Wedow (2011) who use bank size as a proxy for systemic risk. We can now establish Hypothesis 2. 10 Hypothesis 2 (Too big to save: Nonlinear size effect) Ceteris paribus, a bank’s size does not matter for CDS spreads if the bank is small. The effect of a bank’s size on CDS spreads increases in the bank’s size and becomes positive for large banks, controlling for banks’ contributions to systemic risk. However, the ability of a country to bail out banks does not only depend on the banks’ size, but also on the country’s fiscal situation (cf. Buiter and Sibert, 2008; DemirgüçKunt and Huizinga, 2010). Even if financial institutions are large relative to GDP, a government may still be able to bail them out if the country has a high debt capacity. In this situation, CDS spreads may not respond much to banks’ size. In contrast, a bail-out may not be feasible for countries with already high debt levels. In this case, the effect of bank size on CDS spreads is expected to be much stronger. By controlling for systemic risk, we can isolate a country’s willingness to conduct a bail-out (Hypothesis 1) from its ability to do so. This leads us to Hypothesis 3. Hypothesis 3 (Too big to save: Debt capacity) Ceteris paribus, the effect of a bank’s size on CDS spreads increases in a country’s debt level. Finally, banks are more likely to be bailed out if their services are hardly substitutable by other parts of the financial system. For example, in a bank-based system, banks play a more crucial role than in a market-based system. Therefore, if the real economy in a country is strongly reliant upon the banking sector (showing up in a high ratio of domestic credit relative to GDP), a government may be more likely to bail out financial institutions, implying lower CDS spreads. This yields the prediction of Hypothesis 4. Hypothesis 4 (Substitutability) Ceteris paribus, CDS spreads are larger for banks in countries where banks are more easily substitutable. Summing up, our main hypotheses are that (i) a bank’s systemic importance lowers CDS spreads, (ii) the size of a financial institution relative to its home country’s GDP raises CDS spreads when banks are large (controlling for its contribution to systemic risk), (iii) the size effect is stronger for banks in highly indebted countries, and (iv) substitutability raises CDS spreads. 11 3.2 Empirical Model In our empirical analysis, we explain the market indicator of banks’ probability of default as a function of bank-specific and country-specific characteristics. We model the CDS price of bank i in country j at time t as CDSi,j,t =β0 + β1 · T ST Fi,j,t−1 + β2 · T BT Si,j,t−1 + β3 · T BT Si,j,t−1 · debtratioj,t−1 + δ1 · Xi,j,t−1 + µi + γt + ui,t . (7) In order to take the unobserved heterogeneity of banks and over time into account, we include bank-specific fixed effects µi , as well as time fixed effects γt . TSTF measures a bank’s contribution to systemic risk, −∆CoV aR (see Section 2.2).4 TBTS, defined as bank liabilities over GDP, captures the ability of governments to bail out banks: The larger a bank is relative to its home country’s GDP, the less likely it is to be bailed out, controlling for systemic risk.5 In some specifications, this variable also enters in squared form to capture nonlinear effects. We further include an interaction term of TBTS and the ratio of government debt to GDP, as the effect of a bank’s size may depend on the debt level of its home country. Xi,j,t are control variables determining the bail-out probability and the fundamental probability of default. They are explained below. All explanatory variables are lagged by one period to avoid endogeneity problems. Moreover, since CDS prices can hardly be regarded as being stationary (especially in times of financial turmoil), we estimate our model in first differences. Standard errors are clustered at the country level to account for the correlation of banks within the same country. 3.3 Data Sources We collect daily CDS data from Thomson Reuters Datastream. We focus on senior CDS with a maturity of five years, since it has been shown that trading liquidity is highest at this maturity (see European Central Bank, 2008). We also collect the yields on home countries’ government bonds with a maturity of ten years from Thomson Reuters. Our 4 5 ∆CoV aR enters with a negative sign to facilitate interpretation. In a robustness check, we use total bank assets instead of liabilities to construct the TBTS measure. 12 second main data source is Bureau van Dijk’s Bankscope, which contains balance sheet information for a large number of banks from a broad set of countries. We use banks’ consolidated statements for two reasons. On the one hand, CDS spreads refer to the entire financial institution, not only for the parent company. On the other hand, the events in Iceland have shown that the home country of the parent company might also be responsible for its branches abroad.6 Moreover, we use national accounts data from the World Bank’s World Development Indicators database (WDI). Since most of our exogenous variables are only available at yearly frequency, we use the monthly average of all variables with daily frequency and use cubic spline interpolation for variables that are available only at yearly frequency. Our analysis is based on the top 100 largest banks in the world, measured by total assets at the end of 2008, for which data was available. We expand the dataset by several banks, which already failed in 2008 over the course of the financial crisis. However, we do not restrict our sample regarding bank specialization. We start our sample in 2005 since before that time, trading activity in the CDS market was limited to only some banks and pricing information incomplete, and we collect the data until the end of 2011.7 Our final dataset includes 76 banks from 23 countries and spans seven years. In the following subsection, we will describe the included control variables and present descriptive statistics of the data used in the analysis.8 3.4 Control Variables The systemic relevance of a bank is measured by ∆CoV aR. In order to facilitate interpretation, we multiply ∆CoV aR by −1 such that higher values of the variable indicate higher contributions to systemic risk. We follow Buiter and Sibert (2008) in defining the determinants of the too-big-to-save problem (what they call the “vulnerable quartet”). According to these authors, TBTS is particularly considerable in “(i) small countries with (ii) a large, internationally exposed 6 A letter of Iceland’s Minister of Business Affairs to the British Treasury from 2008 says that “if needed the Icelandic Government will support the Depositors’ and Investors’ Guarantee Fund in raising the necessary funds, so that the Fund would be able to meet the minimum compensation limits in the event of a failure of Landsbanki and its UK branch.” 7 Note that data coverage is relatively low in 2011 due to missing data. 8 All details on the preparation of the dataset are listed in Table A1 in Appendix 1. 13 banking sector, (iii) its own currency, and (iv) limited fiscal spare capacity relative to the possible size of the banking sector solvency gap.” In line with this characterization, we measure bank size relative to a country’s GDP. To allow for nonlinear effects, we enter this variable as a second order polynomial in some of the regressions. To control for the limited fiscal spare capacity, we include the amount of government debt relative to GDP and government’s foreign exchange reserves including gold relative to GDP from the World Bank’s World Development Indicators (WDI), as well as the yield on a ten-year government bond from Thomson Reuters Datastream. All three variables are essential for the question whether a country has enough resources or whether it can raise enough money for bailing out its banking sector. The higher the ratio of government debt to GDP, the more likely a bank is to be TBTS and the lower is the probability of a bank bail-out. Higher reserves indicate a higher ability of a sovereign to bail out its banks and thus makes a bank less likely to be TBTS. Similarly, a high government bond yield is associated with a country facing more difficulties in raising money. Therefore, we expect that it is more likely for a bank to be TBTS, the higher the government’s bond yield. Finally, the substitutability of banks is measured by the amount of domestic credit provided by the banking sector in relation to the GDP from the World Bank’s World Development Indicators (WDI) database. We also control for bank-specific characteristics affecting a bank’s fundamental probability of default. In order to capture the bank-specific risk of an individual institution, we use two alternative approaches. First, we include the value at risk (VaR) at the 1% level, the most common measure of risk used by financial institutions.9 Since our measure for systemic importance, ∆CoV aR, is by construction correlated with the VaR, we alternatively control for the bank-specific risk by additionally considering different balance sheet ratios of risk in banking. First, we control for banks’ leverage ratio (total assets over equity). Second, we include a bank’s return on average assets to capture bank performance and profitability. Finally, we include the bank’s recurring earning power, which measures after tax profits adding back provisions for bad debts as a percentage of total assets. At the country level, we use the growth rate of GDP in order to control for business cycle effects. 9 Just as ∆CoV aR, we multiply this variable by −1, such that a higher VaR indicates higher individual risk. 14 3.5 Descriptive Statistics We present descriptive statistics of bank characteristics in Table 2 and of country characteristics in Table 3. Table 2: Descriptive statistics of bank characteristics Variable CDS CDS (winsorized) −∆CoV aR −∆CoV aR (winsorized) TBTS (Total liabilities/GDP) TBTS (Total assets/GDP) −V aR −V aR (winsorized) Leverage ROAA Earning Power ln(Total assets) Mean Std. Dev. 110.002 146.655 100.910 100.224 1.456 1.158 1.367 0.901 0.568 0.728 0.592 0.751 6.607 4.492 6.231 3.146 22.654 12.816 0.359 2.363 1.098 1.676 19.88 1.184 Min. 3.296 7.500 0.005 0.087 0.004 0.005 0.705 1.885 3.971 -71.733 -20.349 15.674 Max. 2217.381 406.670 8.590 3.303 4.848 4.961 42.271 13.902 115.879 2.942 21.639 22.103 N 4263 4263 4263 4263 4263 4263 4263 4263 4263 4211 4211 4263 All variables with daily frequency are winzorized at the 5%-level before calculating monthly averages. The average CDS price is 110.00 basis points, the maximum value being more than 2200 basis points, realized by the Irish bank Anglo Irish Bank in December 2010. Furthermore, the mean of −∆CoV aR is 1.46, which is of the same order of magnitude as in Adrian and Brunnermeier (2011). In order to deal with the extreme outliers of some banks in this data with daily frequency, we use a 5% winsorization before calculating monthly averages. This has the advantage of maintaining the information that some observations are extreme without allowing for an unduly strong impact of individual observations. The average ratio of liabilities to home country’s GDP is 56.80%, while the largest institution in relative terms exceeds its home country’s GDP by more than 380%, which was attained by the Swiss bank UBS in 2007. The bank-specific risk of an individual institution, −V aR, has a sample mean of 6.61, with the Irish bank Bank of Ireland showing the highest level of individual risk in November 2008. In order to deal with those outliers, we also winsorize this variable at the 5% level. The lowest return on average assets of -71.73 comes from the Icelandic bank Glitnir in the aftermaths of its failure. Moreover, the huge leverage ratio of 115.88 refers to the Belgian bank Dexia in 2008. The average ratio of domestic credit provided by the banking sector in relation to GDP is 15 Table 3: Descriptive statistics of country characteristics Variable Mean Std. Dev. Domestic credit / GDP 172.556 64.221 Debt ratio 63.091 32.453 Reserves / GDP 11.836 19.305 Bond yield 3.945 1.099 GDP growth 1.847 3.551 Min. Max. N 57.590 331.766 4232 18.029 174.979 3824 0.315 109.378 4263 0.930 8.449 3850 -7.241 14.763 4263 172.56%, with the maximum value of 331.77% in Japan in 2010 and the minimum value of 57.59% in India in 2005. Moreover, we find in our dataset an average debt ratio of 63.09%, with Japan being the most indebted country in 2010 with a proportion of central government debt to GDP of 174.98%. Australia has the minimum debt ratio of 18.03% in 2008. The average yield for a government bond with a maturity of ten years amounts to about 4% with a maximum of 8.45%, the yield of Ireland in December 2010. 4 Estimation Results in Baseline Specification Table 4 presents the regression results from our baseline specification. The regression results show that a higher systemic importance, as measured by an increase in −∆CoV aR, significantly decreases CDS spreads in all specifications. In contrast, the effect of bank size is always insignificant when controlling for systemic risk. There is no evidence of a nonlinear size effect either: the coefficient of the squared term is insignificant and has an unexpected sign (see column 2 of Table 4). An explanation of this finding is that it is not bank size as such, but its interaction with the home country’s debt capacity that determines CDS spreads. When we additionally include an interaction term of T BT S and a country’s debt ratio, we find that bank size has no significant effect on bank CDS spreads in countries with a zero debt level (see the coefficient of T BT S). However, the effect increases in a country’s debt ratio, consistent with the idea that countries with higher debt ratios are less able to bail out financial institutions. The overall effect of relative bank size becomes positive already at moderate debt ratios (between 9.5% in specification (3) up to 18.8% in specification (5)). The significant coefficient of the debt ratio in specification (3) indicates that the government debt ratio positively affects the CDS price for mean-sized banks, with its effect rising in a bank’s ratio of liabilities to 16 GDP.10 The first result is not robust across specifications, but the increasing effect of debt (which is just the counterpart of the increasing effect of bank size) is found in all specifications. Summing up, these results strongly support Hypothesis 1: The more a financial institution contributes to systemic risk, the more likely is a bail-out in order to prevent negative repercussion effects on the rest of the financial system, and the lower are the bank’s CDS spreads. However, a bail-out becomes less likely for large banks even at moderate sovereign debt levels, and the effect strengthens when the debt ratio increases, confirming Hypothesis 3. In contrast, there is no evidence of a nonlinear size effect when controlling for systemic risk, as predicted by Hypothesis 2. There is no bank size above which banks become too big to be saved, but this level depends on a country’s debt capacity. Another interesting result concerns substitutability. In countries with a higher ratio of domestic credit over GDP, bank CDS spreads tend to be lower. This is consistent with Hypothesis 4 that bail-out probabilities are higher in countries where the banking sector is more important. The remaining coefficients at the country level are in line with our expectations. We find that the CDS price is significantly higher in countries with worse refinancing conditions, indicated by a higher government bond yield. GDP growth and reserves (foreign exchange plus gold) have the expected negative sign, but are insignificant. At the bank level, we find no effect of the value at risk (although it has the expected sign), which may stem from the high correlation with ∆CoV aR. The coefficients of the balance sheet ratios are also insignificant. We checked the robustness of our results in various ways. First, we used a milder winsorization for our risk variables ∆CoV aR and V aR (2.5% instead of 5%, see Table A3 in the Appendix). The results are virtually unchanged, with the same coefficients being significant and similar orders of magnitude of all coefficients. Second, we substitute for our TBTS measure by using the ratio of total assets (rather than liabilities) to GDP as a measure of bank size (see Table A4 in the Appendix). The results are again very similar. As before, we find a highly significant effect of banks’ systemic importance and similar effects of the TBTS variable and its interaction with the home country’s debt ratio. The only difference is that, at least in column (1), bank size has a significantly positive effect 10 In order to facilitate the interpretation, we demeaned the variable TBTS in our regressions. 17 Table 4: Baseline regression with bank fixed effects and time fixed effects VARIABLES (1) CDS (2) CDS (3) CDS (4) CDS (5) CDS coef. p-value coef. p-value coef. p-value coef. p-value coef. p-value -10.82*** (2.48e-07) -10.83*** (2.53e-07) -10.34*** (1.24e-06) -11.47*** (4.24e-06) -11.42*** (5.41e-06) 37.64 (0.115) 40.94 (0.294) -13.47 (0.717) -28.89 (0.448) -34.16 (0.383) -0.851 (0.925) T BT S · Debtratio 1.422* (0.0803) 1.625* (0.0850) 1.813* (0.0654) Debtratio 2.030** (0.0411) -0.565 (0.702) -1.009 (0.545) Bondyield 7.968** (0.0142) 7.677** (0.0197) Domesticcredit/GDP -1.179* (0.0896) -1.307* (0.0720) GDP growth -4.500 (0.181) -4.657 (0.170) Reserves/GDP -0.306 (0.882) -0.573 (0.766) −V aR 0.246 (0.634) 0.230 (0.657) ROAA -1.209 (0.717) Leverage 0.864 (0.165) EarningP ower 1.983 (0.730) -0.268 (0.653) −∆CoV aR T BT S T BT S 2 18 Constant -0.436 Observations 4,263 4,263 3,839 3,473 3,417 R-squared 0.446 0.446 0.439 0.469 0.470 76 76 72 66 65 YES YES YES YES YES YES YES YES YES YES Number of Banks Bank FE Time FE (0.395) -0.434 (0.397) -0.629 (0.243) -0.409 (0.502) OLS regressions in first differences for equation 7 with bank fixed effects and time fixed effects. Standard errors are clustered at the country level throughout. p-value are given in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels. TBTS is demeaned before taking first differences. Sample sizes of different specifications differ due to data availability. on CDS spreads, which is in line with a TBTS story. However, once we control for the home country’s debt level and the TBTS interaction, we find the same interplay between bank size and sovereign debt as before. The results regarding the other coefficients are qualitatively identical. Finally, we additionally include the logarithm of total bank assets in the regressions in order to capture the effect of absolute size (rather than just relative size). Independent of a bank’s size relative to GDP, market participants may consider a bank as being too big to fail due to its absolute size. The results presented in Table A5 in the Appendix do not support that view. Absolute size has no significant effect on CDS spreads. The remaining results are again qualitatively similar to the earlier regressions. 5 Double Wake-Up Call Within our sample period, one can identify two key events that fundamentally changed the financial system: the onset of the financial crisis in 2007 and the failure of the US bank Lehman Brothers in 2008. In order to check for potential structural breaks, we divide our sample period into three sub-periods: first, a pre-crisis period from 2005 until July 2007; second, the crisis period from August 2007 11 until shortly before the meltdown of Lehman Brothers, i. e., until August 2008; finally, the post-Lehman period from September 2008 till the end of our sample. Note that the Lehman failure coincides with the trouble in the Icelandic banking system. We run two of our baseline regressions (namely those in columns (4) and (5) of Table 4), interacting all variables with the dummy variables for each sub-period. The results are shown in Tables 5a, 5b, and 5c.12 In the pre-crisis period (see Table 5a), no bank-specific variable has a significant effect on bank CDS spreads. Only domestic credit and GDP growth have the expected negative signs and are statistically significant. This indicates that individual bank risk was not priced before the crisis, implying that CDS spreads were largely flat. However, with the beginning of the financial crisis, we find highly significant and large negative coefficients of banks’ returns (ROAA) and of our measure of systemic importance (−∆CoV aR), indicating that markets now became aware of banks’ risks and priced in a 11 Most observers date the beginning of the financial crisis on August 9, 2007. To facilitate comparisons of coefficients, Table A8 in the Appendix shows the same results in one table. Tables 5b and 5c additionally present test results for a change in coefficients across subperiods. 12 19 Table 5a: Pre-crisis, crisis, and post-Lehman analysis (1) CDS VARIABLES (2) CDS Panel A: Pre-crisis period coef. p-value coef. p-value (pre-crisis)−∆CoV aR -1.387 (0.425) -0.885 (0.574) (pre-crisis)T BT S -3.859 (0.949) 17.98 (0.718) (pre-crisis)T BT S · Debtratio -0.941 (0.198) -0.876 (0.105) (pre-crisis)Debtratio -0.544 (0.709) -0.350 (0.769) (pre-crisis)Bondyield 0.995 (0.710) 0.707 (0.794) (pre-crisis)Domesticcredit/GDP -1.567*** (0.00491) -1.333*** (0.00738) (pre-crisis)GDP growth -8.771*** (0.00594) -7.788*** (0.00246) 20 (pre-crisis)Reserves/GDP -3.077 (0.132) -3.083* (0.0953) (pre-crisis)−V aR -0.0260 (0.866) 0.0412 (0.805) (pre-crisis)ROAA 14.59 (0.125) (pre-crisis)Leverage 0.731 (0.263) (pre-crisis)EarningP ower -11.99 (0.180) OLS regressions in first differences for equation 7 with bank fixed effects and time fixed effects and interacting all variables with a time dummy. (pre-crisis) indicates a Dummy, which equals 1 before August 2007. (crisis) indicates a Dummy, which equals 1 for the period August 2007 to August 2008. (post-Lehman) indicates a Dummy, which equals 1 after September 2008. Panel A shows the pre-crisis period, Table 5b and Table 5c show the results of the same regression for the crisis and the post-Lehman period. Table A8 summarizes Tables 5a, 5b, and 5c without showing differences to previous periods. Standard errors are clustered at the country level throughout. p-value in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels. TBTS is demeaned before taking first differences. Sample sizes for different specifications differ due to data availability. lower default probability of systemically relevant institutions (see Table 5b). In contrast, sovereign debt problems do not seem to have mattered in this subperiod. In the last subperiod (see Table 5c), CDS spreads are significantly increasing in banks’ size for countries above some (higher than before, but still moderate) threshold of the debt ratio. The effect of bank size increases in the home country’s debt ratio. Since this effect appears only after the large-scale bank bail-outs and the Icelandic events, it seems that markets only now recognized that banks may not only be TSTF, but also TBTS. We also find a significant effect of the government bond yield, which further supports the TBTS hypothesis. Banks in countries with lower funding costs supposedly are better able to bail out a bank in distress. Moreover, −V aR now has a significantly positive impact on CDS spreads, and −∆CoV aR has a reduced impact on CDS spreads compared to the crisis period (although this difference is not statistically significant, as can be seen from the second and fourth column of Table 5c). This may reflect vanishing bail-out expectations in the light of rising sovereign debt problems. It may also partly be due to the non-bailout of Lehman Brothers, which put in question the 100% bail-out guarantee for systemic institutions and raised expectations of a tightening of future banking regulation (such as the introduction of bank resolution procedures), especially for systemic institutions. The events can thus be seen as a double wake-up call. While the onset of the crisis in 2007 reminded investors of the risks in banking and of the fact that some banks were more likely to be bailed out than others, the Icelandic crisis showed that government resources may not be sufficient to bail out even systemic institutions. We again checked the robustness of results by using a different level of winsorization and by defining TBTS on the basis of bank total assets rather than liabilities. The results are shown in Tables A6 and A7. The results are almost identical to those presented above. In particular, bank risk was not priced before the crisis, systemic risk and ROAA become significant in the crisis, and debt problems become important only in the postLehman period. One notable difference is found in Table A6, where −∆CoV aR becomes insignificant in the post-Lehman period, indicating a reduction of the TSTF problem with the emergence of sovereign debt problems. The overall message remains: The two crisis events in August 2007 and September 2008 fundamentally changed the relationship between banks’ CDS spreads and bank- and country-specific variables, especially those related to the TSTF and TBTS problems. 21 Table 5b: Pre-crisis, crisis, and post-Lehman analysis Panel B: Crisis period VARIABLES (1) CDS (2) CDS (3) CDS (4) CDS Difference to pre-crisis period coef. (crisis)−∆CoV aR p-value coef. -17.61*** (0.000302) -16.22*** p-value (0.00114) Difference to pre-crisis period coef. p-value coef. -17.28*** (0.000237) -16.39*** p-value (0.000592) (crisis)T BT S 95.78 (0.181) 99.64 (0.429) 97.71 (0.131) 79.73 (0.449) (crisis)T BT S · Debtratio -0.361 (0.760) 0.580 (0.705) 0.592 (0.666) 1.468 (0.362) (crisis)Debtratio 3.228 (0.185) 3.772 (0.135) 2.730 (0.210) 3.080 (0.174) (crisis)Bondyield -1.567 (0.857) -2.561 (0.757) -1.731 (0.852) -2.438 (0.782) -1.632*** (0.00163) -0.0647 (0.921) -1.657** (0.0213) -0.324 (0.674) (crisis)GDP growth 6.107 (0.389) 14.88* (0.0510) 5.029 (0.525) 12.82 (0.115) (crisis)Reserves/GDP -1.890 (0.381) 1.187 (0.723) -2.118 (0.376) 0.965 (0.761) (crisis)−V aR -1.348 (0.231) -1.322 (0.240) -1.434 (0.188) -1.475 (0.183) -34.58* (0.0531) -49.17** (0.0219) (crisis)Leverage 1.166 (0.181) 0.434 (0.749) (crisis)EarningP ower 13.11 (0.248) 25.11* (0.0967) (crisis)Domesticcredit/GDP 22 (crisis)ROAA OLS regressions in first differences for equation 7 with bank fixed effects and time fixed effects and interacting all variables with a time dummy. (pre-crisis) indicates a Dummy, which equals 1 before August 2007. (crisis) indicates a Dummy, which equals 1 for the period August 2007 to August 2008. (post-Lehman) indicates a Dummy, which equals 1 after September 2008. Column (1) and (3) in Panel B shows the crisis period, Table 5a and Table 5c show the results of the same regression for the pre-crisis and the post-Lehman period. Column (2) and (4) indicate the difference to the pre-crisis period. Table A8 summarizes Tables 5a, 5b, and 5c without showing differences to previous periods. Standard errors are clustered at the country level throughout. p-value in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels. TBTS is demeaned before taking first differences. Sample sizes for different specifications differ due to data availability. Table 5c: Pre-crisis, crisis, and post-Lehman analysis Panel C: Post-Lehman period VARIABLES (1) CDS (2) CDS (3) CDS (4) CDS Difference to crisis period (post-Lehman)−∆CoV aR (post-Lehman)T BT S (post-Lehman)T BT S · Debtratio (post-Lehman)Debtratio (post-Lehman)Bondyield Difference to crisis period coef. p-value coef. p-value coef. p-value coef. p-value -8.398** (0.0353) 9.209 (0.171) -8.341** (0.0390) 8.934 (0.177) -99.49 (0.204) -195.3** (0.0226) -112.2 (0.168) -209.9** (0.0155) 3.499** (0.0161) 3.860** (0.0129) 3.441** (0.0180) 2.849* (0.0778) -0.892 (0.598) -4.120 (0.182) -1.242 (0.527) -3.972 (0.187) 12.21 (0.204) 12.17 (0.238) 10.64*** (0.00524) 10.44*** (0.00669) 23 (post-Lehman)Domesticcredit/GDP 0.323 (0.753) 1.954** (0.0266) 0.225 (0.809) 1.882** (0.0174) (post-Lehman)GDP growth -3.432 (0.367) -9.539 (0.224) -3.575 (0.340) -8.604 (0.273) (post-Lehman)Reserves/GDP 1.183 (0.681) 3.074 (0.331) 1.625 (0.600) 3.743 (0.306) 2.173** (0.0186) 3.521** (0.0221) 2.135** (0.0170) 3.569** (0.0156) (post-Lehman)ROAA -0.436 (0.927) 34.14* (0.0839) (post-Lehman)Leverage 0.188 (0.838) -0.977 (0.432) (post-Lehman)EarningP ower 0.794 (0.913) -12.32 (0.392) (post-Lehman)−V aR OLS regressions in first differences for equation 7 with bank fixed effects and time fixed effects and interacting all variables with a time dummy. (pre-crisis) indicates a Dummy, which equals 1 before August 2007. (crisis) indicates a Dummy, which equals 1 for the period August 2007 to August 2008. (post-Lehman) indicates a Dummy, which equals 1 after September 2008. Column (1) and (3) in Panel C shows the post-Lehman period, Table 5a and Table 5b show the results of the same regression for the pre-crisis and the crisis period. Column (2) and (4) indicate the difference to the crisis period. Table A8 summarizes Tables 5a, 5b, and 5c without showing differences to previous periods. Standard errors are clustered at the country level throughout. p-value in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels. TBTS is subtracted by its mean before taking the first difference. Sample sizes for different specifications differ due to data availability. 6 Conclusion This paper has analyzed the effect of systemic relevance as well as size on banks’ CDS spreads. We argue that bank size is not a satisfactory measure of systemic risk and differentiate the effect associated with bank size from the effect arising from the institution’s systemic importance. To this end, we control for systemic risk using the ∆CoV aR measure introduced by Adrian and Brunnermeier (2011). Controlling for systemic risk, we then identify the too-big-to-save phenomenon by the effect of a bank’s size relative to GDP. We further test whether the too-big-to-save effect depends on the home country’s fiscal situation. We find a significant decrease in CDS spreads for more systemically relevant institutions, which is consistent with the too-systemic-to-fail hypothesis. Markets expect that banks that are highly systemically relevant are more likely to receive governmental support in case of financial distress. Moreover, we find no significant effect of bank size in countries with a zero debt level, once we control for systemic risk. However, the effect of bank size increases in the home country’s debt ratio, and the overall effect of relative bank size becomes positive already at moderate debt ratios. This result is consistent with the too-big-to-save hypothesis. Market participants appear to expect that, in countries with limited debt capacity, a bail-out is less likely the more funding is needed for it. We further show that the relevance of TSTF and TBTS changes over time. Before the financial crisis started in August 2007, we find neither a significant effect of TSTF nor of TBTS. However, we find a strong and significant effect of systemic relevance after the onset of the financial crisis, supporting the relevance of the TSTF problem. The increased awareness of the TSTF problem may have resulted from the large number of bail-outs (before Lehman), which were justified with the prevention of contagion effects. In contrast, TBTS mattered only in the post-Lehman period when we find a significant increase in CDS spreads for larger banks in countries above some moderate threshold of the home country’s debt ratio. We interpret this sequence of events as a double wake-up call for investors. With the advent of the financial crisis in 2007, investors were reminded of the risks in banking and of the fact that some banks were more likely to be bailed out than others. The Icelandic crisis then showed that government resources are not always sufficient to bail out systemic institutions. The recognition of the TBTS problem mitigated the TSTF problem. 24 Hence, our results suggest that banks are not too big to fail, but they may be too systemic to fail and too big to save. Rather than being constant over time, the relative importance of these problems depends on economic circumstances. Even small banks may become too systemic to fail in a financial crisis. And even a small bank may become too big to save if the bank’s home country is burdened with high levels of debt. 25 References Acharya, V. V., and T. Yorulmazer (2008): “Information Contagion and Bank Herding,” Journal of Money, Credit and Banking, 40(1), pp. 215–231. Adrian, T., and M. K. Brunnermeier (2011): “CoVaR,” Discussion paper. Allen, F., and D. Gale (2000): “Financial Contagion,” Journal of Political Economy, 108(1), pp. 1–33. Boyd, J. H., and M. Gertler (1994): “The Role of Large Banks in the Recent U.S. Banking Crisis,” Federal Reserve Bank of Minneapolis Quarterly Review, 18(1), 2–21. Brunnermeier, M. K., and L. H. Pedersen (2009): “Market Liquidity and Funding Liquidity,” The Review of Financial Studies, 22(6), pp. 2201–2238. Buiter, W., and A. Sibert (2008): “The Icelandic banking crisis and what to do about it,” CEPR Policy Insight, (26). Chen, Y. (1999): “Banking Panics: The Role of the First-Come, First-Served Rule and Information Externalities,” Journal of Political Economy, 107(5), pp. 946–968. Demirgüç-Kunt, A., and H. Huizinga (2010): “Are banks too big to fail or too big to save? International evidence from equity prices and CDS spreads,” Discussion paper. Diamond, D. W., and P. H. Dybvig (1983): “Bank Runs, Deposit Insurance, and Liquidity,” Journal of Political Economy, 91(3), pp. 401–419. European Central Bank (2008): Recommendations of the Governing Council of the European Central Bank on government guarantees for bank debt. Gropp, R., H. Hakenes, and I. Schnabel (2010): “Competition, Risk-shifting, and Public Bail-out Policies,” Review of Financial Studies, 24(6), pp. 2084–2120. Hellwig, M. (1998): “Too big to rescued,” cited in: Schweizer Bank Nr. 11 vom November 1998. Hüpkes, E. H. (2005): ““Too Big to Save” - Towards a Functional Approach to Resolving Crises in Global Financial Institutions,” in Systemic Financial Crisis: Resolving large bank insolvencies, pp. 193–215. Douglas Evanoff and George Kaufman, eds. 26 Kaufman, G. G. (2002): “Too big to fail in banking: What remains?,” The Quarterly Review of Economics and Finance, 42(3), 423–436. Rime, B. (2005): “Do “too big to fail” expectations boost large banks issuer ratings?,” Swiss National Bank Working Paper. Stern, G., and R. Feldman (2004): Too big to fail: the hazards of bank bailouts, G - Reference, Information and Interdisciplinary Subjects Series. Brookings Institution Press. Swiss National Bank (2008): Financial Stability Report. The Central Bank of Iceland (2009): Financial Stability Report. Völz, M., and M. Wedow (2011): “Market discipline and too-big-to-fail in the CDS market: Does banks’ size reduce market discipline?,” Journal of Empirical Finance, 18(2), pp. 195 – 210. Zhou, C. (2010): “Are Banks Too Big to Fail? Measuring Systemic Importance of Financial Institutions,” International Journal of Central Banking, 6(34), 205–250. 27 A Appendix A Table A1: Description of variable construction and data sources Variable name Description Data source CDS Single name 5-year senior CDS, winsorized at 5/95% Datastream ∆CoV aR Conditional VaR of market valued total financial assets, see definition in text and Adrian and Brunnermeier (2011), winsorized at 5/95%, multiplied by −1 throughout the paper Own calculation TBTS Total liabilities / current GDP (in %) BankScope & WDI TBTS Assets Total assets / current GDP (in %) BankScope & WDI Debt ratio Central government debt / GDP (in %) WDI Bondyield Yield on a government bond with a maturity of 10 years (in %) Datastream Domestic credit / GDP Domestic credit provided by banking sector / GDP (in %) WDI GDP growth Annual growth rate of GDP (in %) WDI Reserves / GDP Total reserves (includes gold) (in % of GDP $) WDI VaR Unconditional VaR of market valued total financial assets, winsorized at 5/95%, multiplied by −1 throughout the paper Own calculation ROAA Return on average assets BankScope Leverage Total assets / equity BankScope Earning Power Return on asset performance without deducting provisions for bad debt as percentage of total assets BankScope ln(Total assets) Log of total assets BankScope for the VaR and CoVaR calculation: VIX Implied volatility index Chicago Board Options Exchange Repospread Difference between the 3-month repo rate and the 3-month bill rate Bloomberg & Federal Reserve Board H.15 release Termchange Change in 3-month term Treasury bill Federal Reserve Board H.15 release Yieldchange Change in the difference of the 10-year Treasury rate and the 3-month bill rate Federal Reserve Board H.15 release Creditchange Change in the credit spread between BAA rated bonds and Treasury rate (with same maturity of 10 years) Federal Reserve Board H.15 release 28 Details on data construction All data with daily frequency were winsorized at 5/95% (or 2.5/97.5% in the robustness check) before calculating monthly averages. For all data with yearly frequency, we use cubic spline interpolation to get monthly observations. The risk measures ∆CoV aR and V aR are multiplied by −1, such that a larger value indicates a higher contribution to systemic risk and a larger individual risk, respectively. For one bank in our sample, Glitnir, we observe a negative value of equity. Since this would correspond to an infinite leverage, we set the leverage for this bank to the maximum observed leverage in the remaining sample. For the same bank, we further observe implausibly large negative values of the return on average asset. Thus, we dropped eight observations with a return on average assets smaller than −75. In all regression tables, we use the demeaned form of the TBTS variables (total liabilities / GDP or total assets / GDP). 29 Table A2: List of countries and banks in our sample Australia 1 2 3 4 5 Austria 6 7 Belgium 8 9 10 China 11 Denmark 12 France 13 14 15 16 Germany 17 18 19 20 Iceland 21 22 23 India 24 25 26 Ireland 27 28 29 Italy 30 31 32 33 Japan 34 35 36 37 38 Commonwealth Macquarie National Australia Bank St. George Westpac Bank Erste Group Bank Raiffeisenbank Ageas Fortis Dexia KBC Bank Bank of China Danske Bank BNP Paribas Credit Agricole Natixis Societe Generale Commerzbank Deutsche Bank IKB Bayerische Hypovereinsbank Glitnir Kaupthing Landsbanki Bank of India Icici Bank Statebank of India Allied Irish Banks Anglo Irish Bank Bank of Ireland Banca Italease Intesa Sao Paolo Banca Monte dei Paschi Unicredito Mitsubishi UFJ Mizuho Bank Resona Shinsei Bank Sumitomo Mitsui Malaysia 39 Netherlands 40 41 42 Norway 43 Portugal 44 45 Republic of Korea 46 47 48 Singapore 49 50 Spain 51 52 53 Sweden 54 55 56 Switzerland 57 58 United Kingdom 59 60 61 62 63 64 65 USA 66 67 68 69 70 71 72 73 74 75 76 30 Malayan Bank ING Bank Rabobank SNS Bank DNB Bank Banco Comercial de Portugues Banco Espirito Santo Industrial Bank of Korea Shinhan Bank Woori Bank Oversea Chinese Bank United Overseas Bank Banco Sabadell Banco Santander Bankinter Nordea Svenska Swedbank Credit Suisse UBS Abbey Bank Alliance Leicester Barclays HBOS Lloyds Royal Bank of Scotland Standard Chartered Bank Bank of America Bear Stearns Capital One Bank Citigroup Goldman Sachs JP Morgan Chase Merill Lynch Morgan Stanley Wachovia Washington Mutual Wells Fargo Table A3: Baseline regression with bank fixed effects and time fixed effects using a 2.5% winsorization for the risk measures ∆CoV aR and V aR VARIABLES (1) CDS (2) CDS (3) CDS (4) CDS (5) CDS coef. p-value coef. p-value coef. p-value coef. p-value coef. p-value -9.568*** (3.87e-05) -9.569*** (3.90e-05) -8.530*** (8.86e-05) -9.101*** (0.000419) -9.088*** (0.000474) 36.88 (0.116) 40.14 (0.296) -15.35 (0.680) -30.85 (0.414) -36.15 (0.352) -0.840 (0.926) T BT S · Debtratio 1.424* (0.0786) 1.629* (0.0812) 1.819* (0.0629) Debtratio 2.021** (0.0418) -0.602 (0.683) -1.065 (0.521) Bondyield 8.084** (0.0138) 7.793** (0.0191) Domesticcredit/GDP -1.190* (0.0870) -1.316* (0.0695) GDP growth -4.455 (0.182) -4.604 (0.170) Reserves/GDP -0.294 (0.887) -0.555 (0.774) −V aR 0.348 (0.558) 0.360 (0.557) ROAA -1.252 (0.705) Leverage 0.858 (0.169) EarningP ower 2.049 (0.720) 0.274 (0.706) −∆CoV aR T BT S T BT S 2 31 Constant -0.252 Observations 4,263 4,263 3,839 3,473 3,417 R-squared 0.446 0.446 0.439 0.468 0.469 76 76 72 66 65 YES YES YES YES YES YES YES YES YES YES Number of Banks Bank FE Time FE (0.584) -0.250 (0.587) -0.324 (0.560) 0.117 (0.872) OLS regressions in first differences for equation 7 with bank fixed effects and time fixed effects. Standard errors are clustered at the country level throughout. p-value are given in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels. TBTS is demeaned before taking first differences. Sample sizes of different specifications differ due to data availability. Table A4: Baseline regression with bank fixed effects and time fixed effects using the ratio of bank total assets to GDP as a measure of bank’s size (1) CDS (2) CDS (3) CDS (4) CDS (5) CDS coef. p-value coef. p-value coef. p-value coef. p-value coef. p-value -10.82*** (2.45e-07) -10.82*** (2.48e-07) -10.32*** (1.24e-06) -11.46*** (4.32e-06) -11.41*** (5.42e-06) 38.38* (0.0918) 45.34 (0.254) -12.71 (0.729) -26.36 (0.476) -31.76 (0.400) -1.784 (0.839) T BT S asset · Debtratio 1.351* (0.0920) 1.462 (0.104) 1.696* (0.0611) Debtratio 1.995** (0.0435) -0.520 (0.724) -0.997 (0.549) Bondyield 8.004** (0.0136) 7.691** (0.0195) Domesticcredit/GDP -1.168* (0.0910) -1.318* (0.0706) GDP growth -4.460 (0.185) -4.608 (0.174) Reserves/GDP -0.326 (0.875) -0.597 (0.758) −V aR 0.250 (0.628) 0.231 (0.656) ROAA -0.497 (0.882) Leverage 0.921 (0.141) EarningP ower 1.835 (0.750) -0.261 (0.659) −∆CoV aR T BT S assets T BT S assets2 32 Constant -0.431 Observations 4,263 4,263 3,839 3,473 3,417 R-squared 0.446 0.446 0.439 0.469 0.470 76 76 72 66 65 YES YES YES YES YES YES YES YES YES YES Number of Banks Bank FE Time FE (0.401) -0.427 (0.408) -0.609 (0.259) -0.394 (0.517) OLS regressions in first differences for equation 7 with bank fixed effects and time fixed effects. Standard errors are clustered at the country level throughout. p-value in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels. TBTS is subtracted by its mean before taking the first difference. Sample sizes for different specifications differ due to data availability. Table A5: Baseline regression with bank fixed effects and time fixed effects including total assets VARIABLES (1) CDS coef. −∆CoV aR (2) CDS p-value coef. (3) CDS p-value coef. (4) CDS p-value coef. -10.77*** (2.66e-07) -10.31*** (1.40e-06) -11.48*** (4.50e-06) -11.46*** T BT S 125.2 (0.116) T BT S 2 -13.02 (0.323) p-value (5.70e-06) 4.531 (0.910) -8.702 (0.811) -9.020 (0.806) T BT S · Debtratio 1.646* (0.0845) 1.876* (0.0807) 2.179* (0.0534) Debtratio 1.943* (0.0500) -0.588 (0.685) -0.959 (0.556) -46.57 (0.267) -50.95 (0.232) -65.25 (0.133) Bondyield 8.007** (0.0132) 7.706** (0.0184) Domesticcredit/GDP -1.192* (0.0850) -1.338* (0.0654) GDP growth -4.376 (0.166) -4.545 (0.155) Reserves/GDP -0.885 (0.653) -1.385 (0.456) −V aR 0.262 (0.612) 0.250 (0.628) ROAA -1.708 (0.606) Leverage 1.010* (0.0693) EarningP ower 2.229 (0.701) 0.0858 (0.904) ln(T otalassets) -71.89 (0.149) 33 Constant 0.0955 Observations 4,263 3,839 3,473 3,417 R-squared 0.447 0.440 0.469 0.471 76 72 66 65 Number of Banks (0.870) -0.316 (0.539) -0.138 (0.853) Bank FE YES YES YES YES Time FE YES YES YES YES OLS regressions in first differences for equation 7 with bank fixed effects and time fixed effects. Standard errors are clustered at the country level throughout. p-value in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels. TBTS is subtracted by its mean before taking the first difference. Total assets equals the first difference of the log of total assets. Sample sizes for different specifications differ due to data availability. Table A6: Pre-crisis, in-crisis, and post-Lehman analysis using 2.5% winsorization for the risk measures ∆CoV aR and V aR (1) ∆CDS (2) ∆CDS (pre-crisis)−∆CoV aR (pre-crisis)T BT S (pre-crisis)T BT S · Debtratio (pre-crisis)Debtratio (pre-crisis)Bondyield (pre-crisis)Domesticcredit/GDP (pre-crisis)GDP growth (pre-crisis)Reserves/GDP (pre-crisis)−V aR (pre-crisis)ROAA (pre-crisis)Leverage (pre-crisis)EarningP ower -1.664 -4.278 -0.905 -0.653 0.964 -1.587*** -8.684*** -3.018 0.0637 (0.330) (0.943) (0.219) (0.658) (0.717) (0.00482) (0.00668) (0.148) (0.618) -1.120 17.57 -0.835 -0.460 0.679 -1.349*** -7.659*** -3.022 0.117 14.66 0.726 -11.95 (0.463) (0.725) (0.127) (0.704) (0.800) (0.00722) (0.00315) (0.110) (0.391) (0.122) (0.269) (0.182) (crisis)−∆CoV aR (crisis)T BT S (crisis)T BT S · Debtratio (crisis)Debtratio (crisis)Bondyield (crisis)Domesticcredit/GDP (crisis)GDP growth (crisis)Reserves/GDP (crisis)−V aR (crisis)ROAA (crisis)Leverage (crisis)EarningP ower -16.05*** 96.07 -0.388 3.199 -1.832 -1.635*** 6.116 -1.879 -1.422 (0.000401) (0.182) (0.741) (0.187) (0.835) (0.00166) (0.387) (0.378) (0.124) -16.00*** 98.09 0.567 2.696 -1.952 -1.660** 4.991 -2.110 -1.401 -34.61* 1.182 13.10 (0.000258) (0.130) (0.678) (0.213) (0.835) (0.0222) (0.526) (0.375) (0.107) (0.0521) (0.177) (0.249) (post-Lehman)−∆CoV aR (post-Lehman)T BT S (post-Lehman)T BT S · Debtratio (post-Lehman)Debtratio (post-Lehman)Bondyield (post-Lehman)Domesticcredit/GDP (post-Lehman)GDP growth (post-Lehman)Reserves/GDP (post-Lehman)−V aR (post-Lehman)ROAA (post-Lehman)Leverage (post-Lehman)EarningP ower -4.984 -107.0 3.578** -1.039 10.64*** 0.345 -3.190 0.975 2.037* (0.210) (0.165) (0.0128) (0.541) (0.00512) (0.733) (0.396) (0.731) (0.0913) -4.942 -119.0 3.519** -1.384 10.46*** 0.265 -3.324 1.415 2.008* -0.449 0.125 0.668 (0.222) (0.138) (0.0153) (0.484) (0.00656) (0.773) (0.369) (0.641) (0.0942) (0.925) (0.892) (0.927) Constant 0.611 (0.412) 0.732 (0.256) Observations R-squared Number of Banks Bank FE Time FE 3,473 0.476 66 YES YES 3,417 0.479 65 YES YES OLS regressions in first differences for equation 7 with bank fixed effects and time fixed effects. (pre-crisis) indicates a Dummy, which equals 1 before August 2007. (crisis) indicates a Dummy, which equals 1 for the period August 2007 to August 2008. (post-Lehman) indicates a Dummy, which equals 1 after September 2008. Standard errors are clustered at the country level throughout. p-value in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels. Sample sizes for different specifications differ due to data availability. 34 Table A7: Pre-crisis, in-crisis, and post-Lehman analysis using the ratio of bank total assets to GDP as a measure of bank’s size (1) ∆CDS (2) ∆CDS (pre-crisis)−∆CoV aR (pre-crisis)T BT S assets (pre-crisis)T BT S assets · Debtratio (pre-crisis)Debtratio (pre-crisis)Bondyield (pre-crisis)Domesticcredit/GDP (pre-crisis)GDP growth (pre-crisis)Reserves/GDP (pre-crisis)−V aR (pre-crisis)ROAA (pre-crisis)Leverage (pre-crisis)EarningP ower -1.394 -6.174 -0.801 -0.544 0.966 -1.575*** -8.728*** -3.055 -0.0277 (0.423) (0.918) (0.265) (0.712) (0.721) (0.00500) (0.00549) (0.130) (0.858) -0.891 17.05 -0.785 -0.343 0.667 -1.335*** -7.790*** -3.057* 0.0404 14.50 0.712 -11.95 (0.572) (0.731) (0.144) (0.775) (0.806) (0.00706) (0.00256) (0.0925) (0.809) (0.130) (0.276) (0.182) (crisis)−∆CoV aR (crisis)T BT S assets (crisis)T BT S assets · Debtratio (crisis)Debtratio (crisis)Bondyield (crisis)Domesticcredit/GDP (crisis)GDP growth (crisis)Reserves/GDP (crisis)−V aR (crisis)ROAA (crisis)Leverage (crisis)EarningP ower -17.61*** 92.84 -0.408 3.321 -1.410 -1.615*** 6.163 -1.895 -1.346 (0.000301) (0.171) (0.714) (0.173) (0.871) (0.00171) (0.384) (0.380) (0.232) -17.26*** 94.32 0.590 2.724 -1.604 -1.640** 5.300 -2.108 -1.435 -34.89** 1.217 12.95 (0.000240) (0.121) (0.633) (0.205) (0.863) (0.0214) (0.499) (0.377) (0.187) (0.0500) (0.156) (0.255) (post-Lehman)−∆CoV aR (post-Lehman)T BT S assets (post-Lehman)T BT S assets · Debtratio (post-Lehman)Debtratio (post-Lehman)Bondyield (post-Lehman)Domesticcredit/GDP (post-Lehman)GDP growth (post-Lehman)Reserves/GDP (post-Lehman)−V aR (post-Lehman)ROAA (post-Lehman)Leverage (post-Lehman)EarningP ower -8.303** -95.44 3.474** -0.856 10.79*** 0.397 -3.311 0.907 2.186** (0.0372) (0.221) (0.0177) (0.615) (0.00454) (0.705) (0.387) (0.748) (0.0177) -8.306** -109.3 3.469** -1.224 10.52*** 0.219 -3.510 1.488 2.129** 0.845 0.288 0.478 (0.0394) (0.176) (0.0146) (0.532) (0.00631) (0.814) (0.352) (0.629) (0.0174) (0.857) (0.754) (0.947) Constant 0.608 (0.424) 0.715 (0.269) Observations R-squared Number of Banks Bank FE Time FE 3,473 0.476 66 YES YES 3,417 0.479 65 YES YES OLS regressions in first differences for equation 7 with bank fixed effects and time fixed effects. (pre-crisis) indicates a Dummy, which equals 1 before August 2007. (crisis) indicates a Dummy, which equals 1 for the period August 2007 to August 2008. (post-Lehman) indicates a Dummy, which equals 1 after September 2008. Standard errors are clustered at the country level throughout. p-value in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels. Sample sizes for different specifications differ due to data availability. 35 Table A8: Pre-crisis, in-crisis, and post-Lehman analysis (summary of Tables 5a, 5b, and 5c) (1) ∆CDS (2) ∆CDS (pre-crisis)−∆CoV aR (pre-crisis)T BT S (pre-crisis)T BT S · Debtratio (pre-crisis)Debtratio (pre-crisis)Bondyield (pre-crisis)Domesticcredit/GDP (pre-crisis)GDP growth (pre-crisis)Reserves/GDP (pre-crisis)−V aR (pre-crisis)ROAA (pre-crisis)Leverage (pre-crisis)EarningP ower -1.387 -3.859 -0.941 -0.544 0.995 -1.567*** -8.771*** -3.077 -0.0260 (0.425) (0.949) (0.198) (0.709) (0.710) (0.00491) (0.00594) (0.132) (0.866) -0.885 17.98 -0.876 -0.350 0.707 -1.333*** -7.788*** -3.083* 0.0412 14.59 0.731 -11.99 (0.574) (0.718) (0.105) (0.769) (0.794) (0.00738) (0.00246) (0.0953) (0.805) (0.125) (0.263) (0.180) (crisis)−∆CoV aR (crisis)T BT S (crisis)T BT S · Debtratio (crisis)Debtratio (crisis)Bondyield (crisis)Domesticcredit/GDP (crisis)GDP growth (crisis)Reserves/GDP (crisis)−V aR (crisis)ROAA (crisis)Leverage (crisis)EarningP ower -17.61*** 95.78 -0.361 3.228 -1.567 -1.632*** 6.107 -1.890 -1.348 (0.000302) (0.181) (0.760) (0.185) (0.857) (0.00163) (0.389) (0.381) (0.231) -17.28*** 97.71 0.592 2.730 -1.731 -1.657** 5.029 -2.118 -1.434 -34.58* 1.166 13.11 (0.000237) (0.131) (0.666) (0.210) (0.852) (0.0213) (0.525) (0.376) (0.188) (0.0531) (0.181) (0.248) (post-Lehman)−∆CoV aR (post-Lehman)T BT S (post-Lehman)T BT S · Debtratio (post-Lehman)Debtratio (post-Lehman)Bondyield (post-Lehman)Domesticcredit/GDP (post-Lehman)GDP growth (post-Lehman)Reserves/GDP (post-Lehman)−V aR (post-Lehman)ROAA (post-Lehman)Leverage (post-Lehman)EarningP ower -8.398** -99.49 3.499** -0.892 10.64*** 0.323 -3.432 1.183 2.173** (0.0353) (0.204) (0.0161) (0.598) (0.00524) (0.753) (0.367) (0.681) (0.0186) -8.341** -112.2 3.441** -1.242 10.44*** 0.225 -3.575 1.625 2.135** -0.436 0.188 0.794 (0.0390) (0.168) (0.0180) (0.527) (0.00669) (0.809) (0.340) (0.600) (0.0170) (0.927) (0.838) (0.913) Constant 0.573 (0.448) 0.700 (0.282) Observations R-squared Number of Banks Bank FE Time FE 3,473 0.476 66 YES YES 3,417 0.479 65 YES YES OLS regressions in first differences for equation 7 with bank fixed effects and time fixed effects. (pre-crisis) indicates a Dummy, which equals 1 before August 2007. (crisis) indicates a Dummy, which equals 1 for the period August 2007 to August 2008. (post-Lehman) indicates a Dummy, which equals 1 after September 2008. Standard errors are clustered at the country level throughout. p-value in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels. Sample sizes for different specifications differ due to data availability. 36