Banks, Government Bonds, and Default: What do the Data Say?

Transcription

1 Banks, Government Bonds, and Default: What do the Data Say? Nicola Gennaioli, Alberto Martin, and Stefano Rossi * 18 July 2013 We use data from Bankscope to analyze the holdings of public bonds by over 18,000 banks located in 185 countries and the role of these bonds in 18 sovereign debt crises over the period We find that: (i) banks hold a sizeable share of their assets in government bonds (about 9% on average), particularly in less financially developed countries; (ii) during sovereign crises, banks on average increase their bondholdings by 1% of their assets, but this increase is concentrated among larger and more profitable banks, and; (iii) the correlation between a bank s holdings of public bonds and its future loans is positive in normal times, but turns negative during defaults. A 10% increase in bank bond-holdings during default is associated with a 3.2% reduction in future loans, and bonds bought in normal times account for 75% of this effect. Our results are consistent with the view that there is a liquidity benefit for banks to hold public bonds in normal times, which is critical for understanding bank fragility during sovereign crises. JEL classification: F34, F36, G15, H63 Keywords: Sovereign Risk, Sovereign Default, Government Bonds * Bocconi University and CREI, nicola.gennaioli@unibocconi.it; CREI, UPF and Barcelona GSE, amartin@crei.cat; and Purdue University, CEPR, and ECGI, stefanorossi@purdue.edu. We are grateful for helpful suggestions from seminar participants at the Norwegian School of Economics, the Stockholm School of Economics, Banque de France/Sciences Po/CEPR conference on The Economics of Sovereign Debt and Default, and the Barcelona GSE Summer Forum. We have received helpful comments from Mariassunta Giannetti, Sebnem Kalemli-Ozcan (discussant), Paolo Pasquariello, and Hélène Rey (discussant). Jacopo Ponticelli provided excellent research assistantship. Gennaioli thanks the European Research Council for financial support and the Barcelona GSE Research Network. Martin acknowledges support from the Spanish Ministry of Science and Innovation (grant Ramon y Cajal RYC ), the Spanish Ministry of Economy and Competitivity (grant ECO ), the Generalitat de Catalunya-AGAUR (grant 2009SGR1157) and the Barcelona GSE Research Network. Martin and Gennaioli acknowledge support from the International Growth Center, project RA

2 1. Introduction Current events in Europe vividly illustrate how government default can endanger domestic bank stability. Growing concerns of public insolvency have caused severe stress in the European banking sector, which is loaded with Euro-area debt (Andritzky 2012). Problems are very severe for banks in troubled countries, which entered the crisis holding a sizeable share of their assets in their governments bonds: roughly 5% in Portugal and Spain, 7% in Italy and an astounding 16% in Greece (2010 Stress Test). Furthermore, since the onset of the crisis these banks appear to have substantially increased their exposure to the bonds of their financially distressed governments (Brutti and Sauré, 2012), leading to even greater fragility. As The Economist recently put it, Europe s troubled banks and broke governments are in a dangerous embrace. 1 These events are not unique to Europe: a similar relationship between sovereign defaults and the banking system has been at play also in other sovereign crises (IMF 2002). This paper provides the first systematic empirical assessment of the link between sovereign default, bank bondholdings, and bank loans. We investigate two key questions: What are the patterns of banks holdings of government bonds? How do they vary across banks and countries? Do banks accumulate public bonds during sovereign defaults or outside of these events? Do bondholdings affect banks lending behavior in normal times and during sovereign defaults? Do the banks that hold more bonds cut their loans more during these crises? If so, is this due to bonds purchased outside of or during the crises themselves? 1 The Economist, December 17 th

3 To shed light on these questions, we study the patterns of bondholdings and the lending behavior of individual banks outside of and during sovereign defaults. Our aim is to uncover broad stylized facts regarding bondholdings and their relation with lending, not to identify specific causal mechanisms. We do however conduct extensive tests of alternative mechanisms to evaluate the robustness of the stylized facts that we document. We use the BANKSCOPE dataset, which provides the main source of information on the bondholdings and characteristics of banks, covering over 18,000 banks in 185 countries over the period These observations span 18 sovereign default episodes (we often refer to these as crises ) involving 15 countries, 352 banks and 863 bank-years observations. To study the role of country- versus bank-level factors in determining the demand for bonds and the unfolding of crises, we combine the BANKSCOPE data with information from IMF and World Bank sources on the macroeconomic conditions faced by banks in each of the countries in sample. We organize our empirical exercise around the three main hypotheses concerning banks holdings of government bonds and their effects during sovereign defaults. The first hypothesis, which we denote as the liquidity view, sustains that banks hold bonds on a regular basis to store liquidity and to post them as collateral in borrowing arrangements (Bolton and Jeanne 2012, Gennaioli, Martin, and Rossi 2012). 2 In this view, government bonds allow banks to operate more effectively in normal times but they become costly during crises, when public default destroys the asset base of banks and thus their ability to borrow and lend. The second hypothesis, which we label the risk-taking view, holds that banks demand public bonds in 2 Of course, the general notion that government bonds provide liquidity to the private sector is not original to these papers (see, for instance, Holmstrom and Tirole 1998). 3

4 anticipation of, as well as during, sovereign crises in order to chase high returns, perhaps without fully internalizing the systemic consequences of doing so (Acharya and Steffen 2013). Finally, the government intervention view sustains that banks hold public bonds because the government induces them to do so, either through capital regulation in normal times or through creditor discrimination or moral suasion during crises (Livshits and Schoors 2009, Basu 2010, Broner et al. 2013). According to the latter two views, banks derive few benefits from holding bonds in normal times, and sovereign crises reduce bank lending both because they hurt bank assets and because they provide incentives for other banks to buy additional bonds, which crowds out new loans. To evaluate the empirical content of these hypotheses, we split a bank s holdings of public bonds into a time-invariant and a time-varying component. The time-varying component includes changes in bondholdings before as well as during crises, while the time-invariant component captures a bank s stable demand for government bonds. To isolate the role of common factors, we further split each component into a country-level determinant of bonds common to all banks, and an idiosyncratic bank-level factor. We then document the following main facts: Bondholdings by banks are large. In countries that experience at least one sovereign default, banks hold on average 14.4% of their assets in public bonds. Most of the variation in bondholdings (80%) is cross sectional (i.e., time-invariant). This component of the demand for bonds is larger in less financially developed countries and for banks that fund fewer loans, take less risk, and are more levered. 4

5 On average, during default events banks increase their holdings of public bonds from 14% to 15% of their assets. This increase is very heterogeneous, occurring primarily in larger and more profitable banks and in more financially developed countries. Bondholdings matter for lending: higher current bondholdings by a bank are associated to an increase in future loans during normal times, and to a decrease in future loans during defaults. A 10% increase in bondholdings during default is associated with a 3.2% reduction in a bank s loans. Approximately 75% of this drop is attributable to the stable, time invariant, component. Bonds acquired during crises reduce loans only after controlling for the characteristics of banks buying them. These facts have two interesting implications. First, the variation of bondholdings within the sample is largely driven by the behavior of their stable (cross-sectional) component. Moreover, variation in this cross-sectional component appears consistent with the liquidity view, whereby banks should hold the liquid and safe government bonds particularly when they have few investment opportunities and in less financially developed countries. The behavior of bank lending is also consistent with this view: government defaults hurt the liquidity of banks (via their stable bondholdings), leading to a reduction in future loans. This is true also when adjusting for any time-varying country-wide shocks: within the same defaulting country, it is the banks most loaded with public bonds that subsequently cut their lending the most. Second, the time-varying component also matters for understanding bondholdings and loans. It rises during sovereign defaults and is associated to a further decrease in bank lending in these episodes. This is in line with the risk-taking and government intervention views, which 5

6 suggests that they may play a role as well. Interestingly, these time-varying effects are very heterogeneous, being stronger for banks that are more profitable and are located in more financially developed countries. As we will discuss in the paper, this feature has useful implications in the interpretation of our results. This paper is related to two bodies of work. The first one studies the costs of sovereign default. In the conventional approach, these costs consist of market exclusion and external sanctions (see Eaton and Fernandez 1995). The problem is that, in reality, market exclusion is typically short-lived (Gelos et al. 2011) and sanctions are seldom observed (Tomsz, 2007). 3 In these theories, moreover, domestic agents hold no bonds or if they do they are perfectly shielded from the default (e.g., the government engineers a perfect bailout). Recent theories build on the idea that sovereign default is costly because it inflicts a collateral damage to the domestic economy. In these models, such damage arises because default is assumed to be nondiscriminatory, so that it hurts domestic bondholders as well as foreign ones and this has consequences for domestic financial markets. 4 In previous work, for instance, we built a model where public defaults destroy the net worth of banks that hold public bonds, hindering financial intermediation (Gennaioli, Martin, and Rossi 2012). We also provided country-level evidence showing that, after a public default, private credit falls more in those banking systems that are on aggregate more exposed to government bonds. 3 Arellano (2008) shows that in a standard infinite horizon model an output costs of default is needed (on top of market exclusion) in order to rationalize the observed low frequency of default events. 4 An exception to this is Sandleris (2012), who builds a model where defaults signal bad news about the economy and are thus followed by lower output growth. Broner and Ventura (2011) study a model where the government strategically choosing not to enforce private contracts cannot discriminate among domestics and foreigners. Here non-enforcement destroys both international and domestic risk sharing. Brutti (2011), Basu (2010) and Mengus (2012) build models where default is nondiscriminatory and reduces entrepreneurial wealth and investment. 6

7 Other papers have recently examined whether the data are consistent with the collateral damage view of sovereign defaults, either by using aggregate data or by focusing on a specific crisis or country. Brutti and Saure (2013) study the demand for government bonds by European banks during the recent sovereign debt crisis. Acharya and Steffen (2013) examine German banks stock returns to infer the demand for bonds during the recent crisis. Both papers find support for the risk-taking or government intervention views of bondholdings. Acharya, Drechsler, and Schnabl (2013) study the effects of public bailouts of banks, and of sovereign spreads spikes, on banks stock returns. Reinhart and Sbrancia (2011) study the increase in aggregate bondholdings around defaults, and attribute it to financial repression. 5 By using bank-level data for many countries, default episodes and time periods, we provide the first systematic analysis of the patterns of bondholdings at the bank level, and the effects of bondholdings on lending by individual banks outside of and during default events. The second body of work studies the demand for government bonds. Krishnamurthy and Vissing-Jorgensen (2012) analyze the impact of changes in the supply of U.S. treasury bonds on their yields between 1926 and 2008, identifying a liquidity and a safety component of the demand for government bonds (not necessarily by banks, though). Greenwood and Vayanos (2012) analyze the yields of U.S. public bonds of different maturities in light of a preferred habitat theory of demand. Rather than looking at yields, we study the holding of public bonds by banks and the association between these holdings and lending behavior during defaults. 5 Other work documents the association between public defaults and private credit. Arteta and Hale (2008) show that sovereign defaults are accompanied by a decline in syndicated foreign credit to domestic firms. Using aggregate data, Borensztein and Panizza (2008) show that sovereign defaults are accompanied by larger contractions in GDP when they happen in tandem with banking crises. 7

8 The paper proceeds as follows. Section 2 describes the data. Section 3 explains the basic methodology used to decompose banks bondholdings into a time-invariant and a time-varying component, lays out our empirical strategy, and discusses alternative hypotheses. Sections 4 and 5 contain the main results of the paper. Finally, Section 6 concludes. 2. The Data We obtain bank-level data from the BANKSCOPE dataset, which contains information on the holdings of public bonds for 18,776 banks in 185 countries over the period (84,900 bank-year observations). This dataset, which is provided by Bureau van Dijk Electronic Publishing (BvD), provides information on a broad range of bank characteristics, such as size, leverage, risk taking, profitability, amount of loans outstanding, balances with the Central Bank and other interbank ratios. Most important for our purposes is that BANKSCOPE also reports banks holdings of public bonds. Bonds are reported at book value, and the nationality of the bonds is not reported. We shall return to this last issue later on. The information in BANKSCOPE is suitable for international comparisons because BvD harmonizes the data. We start with the full sample of banks in BANKSCOPE. We filter out duplicate records, banks with negative values of all types of assets, banks with total assets smaller than $100,000, and years prior to 1997 when coverage is less systematic. This procedure results in 84,900 observations of the bondholdings variable at the bank-year level over We impose two additional requirements on remaining banks: first, that we observe at least two consecutive years of data so that we can perform our decomposition; and second, that data is available on 8

9 all of the other main variables such as leverage, profitability, cash and short term securities, exposure to Central Banks, and interbank balances. Our main sample then consists of 7,094 banks in 159 countries for a total 32,068 bank-year observations. We take the location of banks to be the one reported in Bankscope, which coincides with the location of the Bank s headquarters. Commercial banks account for 33.2% of our sample; cooperative banks for 38.2%; savings banks for 20.6%; investment banks for 1.6%; the rest includes holdings, real estate banks, and other credit institutions. Data on the macroeconomic conditions of the different countries is obtained from the IMF s International Financial Statistics (IFS) and the World Bank s World Development Indicators (WDI). 6 To measure the size of financial markets we use the ratio of private credit provided by money deposit banks and other financial institutions to GDP, which is drawn from Beck et al. (2000). This widely used measure is an objective, continuous proxy for the size of the domestic credit markets. Finally, we follow the existing literature and proxy for sovereign default with a dummy variable based on Standard & Poor s, which defines default as the failure of a debtor (government) to meet a principal or interest payment on the due date (or within the specified grace period) contained in the original terms of the debt issue. According to this definition, a debt restructuring under which the new debt contains less favorable terms to the creditors is coded as a default. The Greek bond swap that was launched in February of 2012, for instance, is identified as a default by Standard & Poor s because the retroactive insertion of collective action clauses was deemed to materially change the original contract terms. 6 See Table AI in the Appendix for a description of variables. 9

10 In our robustness tests, we complement our analysis by using two alternative measures of sovereign default, namely: (i) a monetary measure of creditors losses given default, i.e., haircuts, from the work of Cruces and Trebesch (2013) and Zettelmeyer, Trebesch, and Gulati (2012), and; (ii) a market-based definition based on whether sovereign bond spreads exceed a threshold identified with extreme value theory (approximately 1000 basis points), following the methodology of Pescatori and Sy (2007). These measures capture dimensions of sovereign risk that are not captured by the S&P default dummy, such as spikes in credit spreads and the economic magnitude of creditors losses. As we show in Tables AVI and AVII in the Appendix, and as we later discuss in the text, our results are very robust to these alternative measures. In our main analysis, however, we stick to the S&P default dummy. Indeed, measures of haircuts depend heavily on the assumptions one makes about counterfactuals (e.g., Sturzenegger and Zettelmeyer 2008); and measures based on sovereign bond spreads require observing reliable data on secondary market trading, which limits our sample size. Using the S&P dummy, we have 18 sovereign defaults of different duration in 15 countries, which are listed in the Table AII of the Appendix. Using the haircut measure we can only consider 12 defaults in 11 countries. Using the credit spreads measure we can only consider 12 defaults in 10 countries. Table AII shows that there is a large variation in the size of defaulting countries and in the extent of bank involvement. A few countries such as Argentina, Russia, Nigeria, Kenya and Ecuador have the lion s share of banks, while in other six defaulting countries our data contains fewer than five banks in each country. These features raise the concern that countries that are small and have few banks might drive our results. In our robustness tests (Tables AIV and AV in 10

11 the Appendix) we re-estimate our regressions focusing on the set of large defaulting countries and discarding countries with fewer than five banks during a default episode, respectively. 2.1 Data Comparability The BANKSCOPE dataset is widely used and has an established track record, 7 but there is one important dimension along which its reliability has not been scrutinized: its measure of government bondholdings. To check the quality of this measure, we compare it to other data sources on bondholdings: the country-level measure of banks net claims on the government from the IMF, and the bank-level data from the recent European Stress Test. [Table I here] Table I compares the BANKSCOPE data on bondholdings with the IMF measure. Panel A contains the mean, the median, and the standard deviation of bondholdings (as a share of total assets) in BANKSCOPE. Mean bondholdings are at 9% of assets, while median bondholdings are approximately half as high. The standard deviation of bondholdings in the sample is also high. 8 Panel B reports the same information, but only for the subset of countries for which the IMF also reports banks bondholdings. Panel C displays the IMF measure of financial institutions net claims to the government, computed as a share to total assets. 9 Mean, median and 7 See, for instance, Classens and Laeven (2004), and Kalemli-Ozcan et al. (2011). 8 The highest bondholdings in the sample are above 65% for selected banks in Argentina, Nigeria, India, Jamaica and Venezuela in 2003; the lowest bondholdings are 0% (e.g., several U.S. banks). 9 This variable reports the net positions of commercial banks, defined as holdings of securities plus direct lending minus government deposits, and it can be interpreted as a proxy for the bondholdings of banks. Other papers using this measure are Gennaioli et al. (2012) and Kumhof and Tanner (2008). 11

12 standard deviation of the IMF measure are very close to the BANKSCOPE data. The IMF data gives a slightly higher mean bondholdings, but measurement in the two datasets tends to converge towards the end of the sample. This discrepancy between IMF and BANKSCOPE data could be due to the fact that the former might also capture non-bond finance and to the fact that the banks used to compute the IMF measure may differ from those in BANKSCOPE. The IMF data cannot address the quality of the BANKSCOPE data on a bank-by-bank basis. We thus compare our measure of bondholdings to the one reported by the European stress test of This also allows us to evaluate the mismeasurement stemming from the fact that, contrary to the stress test, BANKSCOPE does not break down bonds by nationality. Table II reports bondholdings from the European stress tests of 2010 and Panel A of the table reports bondholdings for the full sample contained in the stress test, whereas Panel B reports bondholdings for the subset of the banks in the stress test sample that is contained in BANKSCOPE. The bondholdings reported by BANKSCOPE are showed in Panel C. Clearly, the data from both sources are highly comparable. The bank-by-bank correlation between the bondholdings reported by BANKSCOPE and by the stress test is 80%. The small discrepancies between our measure and the stress test measure are thus most likely due to differences in the time at which the measurement itself took place. 10 [Table II here] 10 While BANKSCOPE also counts non-eu bonds, the bondholdings of European banks consist almost exclusively of EU bonds the very reason of the stress test in the first place. 12

13 The evidence is reassuring. Even in highly integrated European markets, where domestic and foreign bonds are in many cases treated symmetrically by the regulatory framework, more than 75% of bank bondholdings correspond to domestic bonds. This share is in all likelihood much larger in the subset of developing countries that provide most of our observations on sovereign defaults. In sum, the BANKSCOPE measure is a good proxy for the domestic public bonds held by banks around the world, and we use it as such in the rest of the paper. 2.2 Summary Statistics We consider the distribution of bank characteristics in BANKSCOPE, focusing on: (i) bank size as measured by total assets, (ii) risk taking as measured by the investment in assets other than cash and other liquid securities, (iii) leverage as measured by one minus shareholders equity as a share of assets, (iv) loans outstanding as a share of assets, (v) profitability as measured by operating income over assets, (vi) exposure to the Central Bank as measured by deposits in the Central Bank over assets, (vii) balances in the interbank market, and (viii) government owned, a dummy that equals one if the government owns more than 50% of the bank s equity. 11 To neutralize the impact of outliers, all variables are winsorized at the 1 st and 99 th percentile. Table III provides descriptive statistics for these variables in our sample. [Table III here] 11 For robustness, we adopt two different definitions for risk taking. The first definition measures the share of assets other than cash and bonds. This is our preferred definition, as it reflects the status of bonds as safe assets in normal times, not just from the viewpoint of banks preferences but also of risk weighting in capital regulation. The only problem of this definition is that it mechanically decreases with bondholdings, potentially generating spurious correlations. As a result, we perform robustness tests using a definition of risk taking equal to the share of non-bond assets other than cash. Our main results do not change when we use this alternative measure. 13

14 Panel A shows that there is a fairly large variation in bank characteristics within the BANKSCOPE sample. The average bank invests roughly 90% of its resources in risky assets (60% of which are loans, and the rest includes debentures and other non-government securities), obtains 90% of its financing in the form of debt, which includes deposits (for an average leverage ratio assets/equity of about 10), and holds 3% of its assets in central bank reserves. 12 Table IV reports the correlations between different bank characteristics in our sample. [Table IV here] All correlations, except for the one between size and loans, are statistically significant. Bank profitability is positively correlated with size, exposure to the central bank and interbank balances, while it is negatively correlated with risk taking, leverage, and loans outstanding. 3. Competing Hypotheses and Methodology We organize our exercise around the three main hypotheses regarding the bank bondholdingsdefault link. The first hypothesis, which we call the liquidity view, holds that banks use public bonds as a way to store their funds in the short run, in order to finance future investments and outlays. According to this view, bondholdings are part of a bank s normal business activity. They are accumulated when the bank has few current investment opportunities relative to those in the future, but they may be costly when a sovereign crisis breaks out (Bolton and Jeanne 2012, Gennaioli et al. 2012). The risk-taking view holds instead that banks buy public 12 Panel B of Table III shows the characteristics of banks involved in the stress test. These banks are much larger and extend more loans than the median BANKSCOPE bank. They also have lower exposure to the Central Bank and to other banks. Leverage and risk taking are instead of similar magnitude to those observed in BANKSCOPE. 14

15 bonds when risk premia on sovereign bonds become large in anticipation of, as well as during, sovereign debt crises. In this view, banks choose to take on sovereign risk in order to chase high returns (Acharya and Steffen 2013). Banks risk seeking thus reduces loans not only because of the losses they suffer in the event of a sovereign default, as in the liquidity view, but also because investment in high yield bonds crowds out loans to households and business firms. Finally, the government intervention view sustains that banks hold public bonds because they are induced to do so by the government, through capital regulation in normal times, through creditor discrimination (e.g., promises of bank bailouts) and moral suasion during defaults (see Livshits and Schoors 2009, Hannoun 2011, and Broner et al. 2013). As in the risk-taking view, bondholdings and bond purchases crowd out loans. Before moving on, it is useful to remark that in principle a different form of intervention is possible, according to which the government uses its bonds to compensate troubled banks during crises. This also predicts that banks pile up bonds during default but, in contrast with the previous hypotheses, bondholdings should, if anything, boost banks loans during default. Before proceeding, we must acknowledge that our data set is ill-suited to perfectly distinguish among these hypotheses. To do so, we would need data on the returns of bonds and of alternative assets, as well as a precise knowledge of bank regulation around the world. Our dataset cannot provide this depth of information, as it has been constructed instead to cover the largest possible sample of banks and countries on and off default. Even though our data makes it difficult to univocally identify the merits of alternative views, however, it still enables us to analyze some broad implications along which they differ. 15

16 The first set of implications concerns the timing of bondholdings. According to the liquidity view, banks should if anything reduce their bondholdings during crises because risky bonds do not constitute a good store of liquidity. According to the risk-taking and government intervention views, by contrast, banks should buy government bonds particularly at certain specific times, namely right before as well as during sovereign defaults. 13 Evaluating the time variation of bondholdings thus provides a way to assess the role played by different views. The second set of implications comes from variation across banks. According to the liquidity view, bondholdings should be larger: (i) in financially underdeveloped countries (where private liquid assets are few) and (ii) among banks that currently have few alternative investment opportunities. In line with this notion, we study the determinants of bondholdings by evaluating the role of country-level measures of financial/economic development and of bank-level measures of investment opportunities such as outstanding loans, profitability, risk taking, and central bank deposits. The risk-taking and government intervention views also have cross sectional predictions that can be tested in the data. The risk-taking view predicts, among other things, that during sovereign crises bonds should be demanded by banks that are more risk tolerant (e.g. because they are larger and thus more diversified), or by banks that have more resources to invest (e.g. because they are more profitable). According to the government intervention view, large banks that are too big to fail or banks that are particularly close to the government (e.g., because they are government owned) should be expected to buy more bonds during crises. 13 Governments may induce banks to hold bonds in normal times by using capital regulation. We later discuss this possibility, but note that unlike in the liquidity view this motive becomes stronger during crises. 16

17 Guided by these considerations, we will formally test some of the (sometimes exclusive, sometimes non-exclusive) implications of these different hypotheses. To do so, we employ the econometric methodology described below. 3.1 Decomposition of Bondholdings As a first step, we decompose bondholdings into a time-invariant and a time-varying component. Let b i,c,t denote the ratio of government bonds over assets held at time t by bank i located in country c. We regress b i,c,t on a set of bank dummies and obtain: b i,c,t = b i + b i,c,t. (1) Here b i is the predicted time-invariant component of bank i s bonds, while b i,c,t is the time varying residual. In this decomposition, b i is the average bonds/asset ratio of bank i over time; all time variation in bondholdings, including the one that takes place across normal and crisis times, is accounted for by b i,c,t. We can now go back to our different hypotheses. Because the risk-taking and government intervention views especially rely on time variation, they should be captured by the time varying component b i,c,t. In particular, if we regress a bank s lending behavior during default on the two components of bonds, the time varying part b i,c,t would absorb the role of increases in bonds around the default window (be they due to risk taking or moral suasion) while b i would capture long term, stable determinants of the demand for bonds. As a result, we refer to the time invariant component b i as a bank s stable bondholdings. 17

18 To isolate the determinants of stable bondholdings that are common to all banks in a country (e.g., financial development) from those that are idiosyncratic to individual banks, we regress b i on a set of country dummies, obtaining: b i = b c + b i,c, (2) where b c is the predicted country-specific component of stable bondholdings, while b i,c is the residual from the same regression. By construction, b i,c absorbs all bank-specific factors. To assess whether changes in bondholdings over time reflect common country-level shocks (e.g., business cycle fluctuations) or bank level factors, we regress the time-varying residual b i,c,t on the interaction of time and country dummies, obtaining: b i,c,t = b c,t + b i,c,t. (3) Component b c,t is the part of time-varying bondholdings at time t that is common to all banks in a country. The residual b i,c,t captures bank-specific time-varying bondholdings at time t. Before performing our regressions, it is useful to quantify the share of the total variation of bondholding that is accounted for by each of these components. Table V performs this calculation in the full set of countries (Panel A) and in the sample of defaulters (Panel B). [Table V here] Three intriguing features stand out. First, in the full sample (Panel A), variation in stable bondholdings explains the lion s share (80%) of the total variation of bonds. This is almost equally split between country- and bank-specific subcomponents. Understanding why banks 18

19 regularly hold bonds seems therefore important. Second, the picture changes once we focus on the subset of countries that experienced at least one sovereign default during the sample period. Now (see Panel B), the explanatory power of the time-varying component increases from 20% to 40%, becoming roughly similar to variation in stable bondholdings. 14 A plausible explanation is that default events are associated with significant time variation of bonds, perhaps consistent with the risk taking and government intervention views. Figure 1 illustrates this point by showing country-specific time-varying bondholdings for selected countries in default. Third, Panel C shows that the correlation between the stable and time varying bondholdings is negative at the bank-level, suggesting that banks having large stable bondholdings may be different from those that load up more on public bonds during crises. Insofar as stable bondholdings reflect average bondholdings in normal times, this suggests that the demand for bonds may behave very differently during crises and outside of crisis periods. 15 [Figure 1 here] The importance of bank-level variation in Table V indicates that bank heterogeneity is critical. Empirical analyses based only on country-level aggregate data, which are the only ones performed so far, neglect the bulk of the variation in bonds, particularly during defaults. 14 This change may partly reflect lower cross country variation: any two countries picked from the subsample of defaulters are likely to be more similar than any two countries picked randomly from the total sample. At the same time, the relative importance of country-level factors in explaining time variation in bondholdings is almost unaffected, being equal to 24% in Panel A and to 22% in Panel B. Thus, there is not a drastic drop of country-level heterogeneity, but rather this heterogeneity plays out in the time variation, presumably during default. 15 Panel C also confirms that the correlations between the orthogonal components of time varying and time invariant bondholding demand are indeed 0. 19

20 3.2 Regression Strategy We now ask: (i) what factors shape bank bondholdings?, and (ii) how do bondholdings shape the extent to which bank lending responds to sovereign defaults? In the first part of the analysis we study the determinants of a bank s holdings of public bonds, starting with stable bondholdings. We run the following regressions: b i,c = γ + β X i,t 1 + ε i,c, (4) b c = γ + β c X c,t 1 + β i X i,t 1 + ε i,c, (5) In light of previous definitions, Equation (4) estimates the bank-level determinants of stable bondholdings, while Equation (5) estimates their country level determinants. In line with our previous discussion, X i,t 1 is a vector of bank characteristics such as risk taking, loans outstanding, exposure to central bank, interbank balances, profitability, size, and whether the bank is owned by the government. Accordingly, vector X c,t 1 includes country level factors such as a country s financial development (as measured by Private Credit to GDP and banking crises), inflation and other macro variables that affect demand for public bonds. We employ the same methodology to analyze the determinants of time variation in bondholdings. To do so, we run the following regressions: b i,c,t = γ + α default c,t 1 X i,t 1 + γ X i,t 1 + δ default c,t 1 + ε i,c,t., (6) b c,t = γ + α c default c,t 1 X c,t 1 + γ c X c,t α i default c,t 1 X i,t 1 + γ i X i,t 1 + δ c default c,t 1 + ε i,c,t, (7) 20

21 where default c,t 1 is a dummy variable taking value 1 if at t 1 country c was in default and value zero otherwise. Equation (6) estimates bank-specific time variation while Equation (7) estimates the country-level time variation, with specific reference to default events. An increase in bondholdings during defaults (i.e., positive coefficients δ c and δ), for instance, would be consistent with the risk-taking and government intervention views, particularly if this increase is concentrated in certain banks and countries as captured by the interaction between default and controls X i,t 1 and X c,t 1. In the second part of our regression analysis, we study how the association between sovereign defaults and a bank s lending behavior is affected by the bank s bondholdings. Let Λ i,c,t denote the change in loans made by bank i in country c between time t 1 and t. We then run the following regression: Λ i,c,t = γ + ψ n default c,t 1 b i + ψ cr default c,t 1 b i,c,t 1 + ζ n b i + ζ cr b i,c,t 1 + δ default c,t 1 + ε i,c,t. (8) The regression separately estimates the impact, outside of and during default crises, of stable bondholdings b i and time-varying bondholdings b i,c,t on the change in bank loans. Coefficients ζ n and ζ cr capture the role of bonds outside of default. If ζ cr > 0, higher bondholdings in normal times associate with more future loans. This is consistent with the liquidity view, whereby banks accumulates bonds at times in which investment opportunities are few in order to tap investment opportunities arising in the future. If ζ n > 0, banks that on average hold more bonds on average make more loans in normal times. The sign of this coefficient is not fully pinned down by the liquidity view. If the banks holding many bonds on 21

22 average are those having few sporadic investment opportunities, they should be expected to also make fewer new loans, i.e. ζ n < 0. If instead the banks holding many bonds on average are those expecting a steady growth of future investment opportunities, they should also expect to make more new loans, i.e. ζ n > 0. Our empirical tests allow us to distinguish between these two possibilities. On the other hand, coefficients ψ n and ψ cr in Equation (8) capture the differential effect of bonds on loans during sovereign defaults. If a high stable demand for bonds is associated with fewer loans during defaults, we should observe ψ n < 0. If instead a high time-varying component of bondholdings during default (due for instance to risk taking or moral suasion) is associated with fewer new loans during default, we should observe ψ cr < 0. In light of our previous analysis on the demand for bonds, we then estimate Equation (8) also by controlling for bank- and country- level variables X i,t 1 and X c,t 1 that may affect the demand for public bonds. We also include country dummies to control for unobserved timeinvariant country-level factors and time dummies to control for country-level shocks. Finally, to tease out the role of country level versus bank specific factors, we estimate: Λ i,c,t = ψ n,i default c,t 1 b i,c + ψ cr,i default c,t 1 b i,c,t + + ψ n,c default c,t 1 b c + ψ cr,c default c,t 1 b c,t + (9) This regression has an intuitive interpretation. If the reduction in lending during crises is due to bank- rather than to country-level variation in bonds, then coefficients ψ n,i and ψ cr,i and should be negative while ψ n,c and ψ cr,c should be zero. If instead it is country-level common 22

23 factors that drive the effects of sovereign crises on loans, the coefficient attached to bank-level variation in bonds should be zero. Separating country- versus bank-level factors allows us to: (i) address the concern that some omitted country-level factors drive our results, and (ii) evaluate the role of the distribution of individual bank characteristics in a country. Once again, the coefficients on non-interacted bondholdings capture the role of bonds outside of default: also with respect to this latter dimension, Equation (9) separates country and bank level variation. 3.3 Interpretation and Causality Regressions (4) to (7) assess the correlates of government bonds, allowing us to take a first look at whether observed bondholdings are more consistent with a liquidity, risk-taking or government intervention view. The key step in our analysis, however, is the interpretation of Equations (8) and (9), which estimate the association between bondholdings and loans both in normal times and during default episodes. As previously stressed, our goal here is not to identify specific causal mechanisms. To interpret our results, however, it is useful to consider how our estimates may be affected by the most pressing endogeneity concerns. Even if higher bondholdings are associated with more severe drops in loans after a default, it may not be government default itself that causes the drop in lending though the bank-bondholding channel. It may simply be that government defaults occur precisely when lending is weak (e.g., in recessions), which also happens to be when banks hold many government bonds. In this case, the correlation documented by the regression would be spurious. Although we do not claim that our estimation strategy fully solves these causality 23

24 concerns, it is important to stress two important properties of our decomposition. First, the stable component of bonds is by construction uncorrelated with time variation around the default event. As a result, the value of stable bondholdings is predetermined relative to the timing of these events. This fact significantly ameliorates the concern that the interactive coefficients on stable bondholdings are spurious. It also ameliorates selection at the treatment type of concern, namely, concerns that banks may change their lending behavior in anticipation of a sovereign default: by construction, not only stable bondholdings are predetermined but also do not trend in the vicinity of default. Second, and related, our exercise allows us to separate in Equation (9) the effect of country-level variation in bonds from that of idiosyncratic bank-level variation in bonds. As long as common country-level shocks affect banks in the country in a uniform way, the coefficient on the idiosyncratic bank-level component (e.g., coefficient ψ cr,c in regression (9)) will reflect the impact of a bank s bonds on its loans and not the general effect of common shocks. Of course, there may be reasons to believe that not all banks are affected in the same way by country-level shocks. It could be, for instance, that bonds are purchased precisely by those banks that anticipate the largest drop in credit demand during a sovereign debt crisis. In this case, the data would depict a negative correlation between bondholdings and loans but there would be no causal effect from the former to the latter. We attempt to deal with this concern by analyzing the characteristics of banks. To the extent that banks purchasing government bonds during default are the small, low-profitability banks, it would seem plausible that their bondholdings are driven by an anticipation of a fall in demand. By contrast, this hypothesis seems less likely if during default bondholdings are accumulated among large and 24

25 profitable banks. Presumably, these banks have access to the best investment opportunities and are thus less subject to a drop in demand than their less profitable counterparts. Finally, in different versions of our regressions we control for time dummies, country dummies, and for an extensive number of country- and bank-specific time-varying controls. We also control for country-year interacted dummies. Our results are very robust to these tests. 4. Determinants of Banks Bondholdings We now use our decomposition to study the determinants of bank bondholdings. Section 4.1 studies stable bondholdings, whereas Section 4.2 deals with time-varying bondholdings. 4.1 Stable bondholdings Table VI reports the estimation of Equation (4) in Panel A, and of Equation (5) in Panel B. To show the explanatory power of individual bank characteristics, we report a set of univariate regressions. In the last columns, we include all controls together. [Table VI here] Consider Panel A. The variables with most explanatory power are risk taking and outstanding loans, which in columns (2) and (4) respectively account for 23% and 15% of the variation of the dependent variable. Both variables have a negative impact on bonds. These results are consistent with the liquidity view: banks that are risk-seeking and that on average 25

26 access many good investment opportunities do not need safe and liquid public bonds to store their funds. Note that here a high level of outstanding loans means, ceteris paribus, that the bank has a plentiful of investment opportunities. As a result, it does not need to store many of its funds in order to tap investment opportunities in the future. This message is confirmed when all variables are considered jointly, as in column (9) of Table VI Panel A. Other interesting observations are that bondholdings increase with bank leverage, consistent with the fact that public bonds are used as collateral to lever up, and decrease with banks use of substitute liquid assets, such as central bank reserves and interbank deposits. Bondholdings also increase with bank size, consistent with the finding that larger industrial firms use relatively less cash to manage their liquidity (Opler et al. 1999). One concern with these results is that they might partially capture the effect of bank characteristics during sovereign crises. Indeed, the dependent variable averages also bonds held by banks during sovereign defaults in the sample. To avoid this problem, we re-estimate the specification in column (9) within the subsample of countries that never experience a default, and we report the results in column (10). In this column, the stable demand for bonds can be interpreted as the average demand for bonds in normal times. Most coefficients remain of the predicted sign and of similar magnitude as in column (9) with the exception of leverage, which is now insignificant and profitability, which is negatively associated with bonds, most likely because profitable banks have more investment opportunities. Finally, bondholdings are 26

27 lower for government-owned banks. 16 Overall, the economic magnitude of these effects is large and our model can explain up to 34% of the idiosyncratic variation in stable bondholdings. 17 Panel B studies the country level determinants of stable bondholdings. Crucially, these bondholdings decrease sharply in Private Credit to GDP. This is by far the most important variable, as it explains a staggering 67% of the variation alone. Financial underdevelopment seems to be a key driver of bondholdings, consistent with the liquidity view. Accordingly, a higher frequency of banking crises another proxy for financial underdevelopment is associated with higher bondholdings. 18 The positive coefficient of sovereign default in Column (1) seems prima facie inconsistent with the liquidity view. The effect is however driven by the fact that sovereign defaults are more frequent in less financially developed countries. Once we include Private Credit to GDP in Column 6, the coefficient on sovereign default turns negative. 19 Overall, Table VI is consistent with the notion that stable bondholdings are well explained by the liquidity view: banks holdings of public bonds are higher in countries that ceteris paribus experience fewer public defaults (so that public bonds are indeed a relatively 16 In univariate regressions the coefficient on profitability is positive while that on the government owned dummy is positive. This is intuitive: profitability is positively correlated with size. However, after we control for the fact that large banks hold more bonds, the sign of profitability turns negative. Accordingly, government owned banks take fewer risks. After we control for risk taking, we find that government owned banks own fewer bonds. 17 A one-standard deviation increase in risk-taking is associated with 2.7% fewer bondholdings; a one-standard deviation increase in bank profitability is associated with a 0.28% increase in bondholdings; a one-standard deviation increase in size is associated with a 0.2% increase in bondholdings. Another way to see this is to note that a bank in the bottom quartile of size, risk taking, and profitability has 4.2% of public bonds, while a bank in the top quartile of size, risk taking, leverage and profitability has 1.6%. The economic consequences of these differences are further amplified by the possibility that banks use public bonds to lever up and raise more funds. 18 As we are dealing with aggregate country-level bondholdings, also supply side forces may play a role here. Indeed, it could be that governments in countries prone to banking crises may engage more heavily in intermediation and in bailouts, and thus exhibit higher public debt levels. 19 Bank characteristics do not matter in explaining country-level bondholdings, with the exception of average risktaking in the banking sector (see columns (7) and (8)). One interpretation is that higher average risk taking may capture laxer capital regulation, which reduces banks willingness to hold public bonds. Alternatively, average risk taking may reflect the existence of private insurance mechanisms which lead banks to hold less cash and bonds. 27